Idea Transcript
Studies in East Asian Linguistics
Sophia Yat Mei Lee
Emotion and Cause Linguistic Theory and Computational Implementation
Studies in East Asian Linguistics Editors-in-Chief Jong-Bok Kim, Kyung Hee University, Seoul, Korea (Republic of) Chu-Ren Huang, Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong Associate Editors Shirley N. Dita, De La Salle University, Malate, Manila, Philippines Yasunari Harada, Waseda University, Tokyo, Japan Victoria Rau, National Chung-Cheng University, Taiwan John Wakefield, Hong Kong Baptist University, Hong Kong, Hong Kong Jie Xu, University of Macau, Macao
The series will publish studies in the general and interdisciplinary area of linguistics with a particular focus on East Asian languages. The series will feature not only formally but also empirically oriented work on any aspect of the theoretical analysis or processing of the languages. The topics of the series include a variety of language studies (syntax, semantics, pragmatics, phonology, morphology, corpus linguistics, historical linguistics, discourse analysis, language acquisition, psycholinguistics, language learning) as well as computational linguistics (cognitive modeling of language, dialogue and interactive systems, information retrieval, language resources, machine translation, NLP applications, sentiment analysis, social medial, text classification, and the like). The series can be about a single East Asian language or a comparative work among East Asian languages and/or between an East Asian language and any other languages that contributes to the advancement of the study of East Asian languages. The importance of understanding the linguistic features of East Asian languages has been increasing as does the role of the East Asian countries in the era of global world. While there have been numerous research activities for the languages, few has initiated the dissemination of the research results accessible to the audiences around the world and published by a globally recognized publisher. The series would be an optimal place to report new developments in the East Asian language study of language, information, logic, and computation. The series would make new results, ideas, and approaches for the study of East Asian languages available to the public in the most timely and focused manner.
More information about this series at http://www.springer.com/series/15584
Sophia Yat Mei Lee
Emotion and Cause Linguistic Theory and Computational Implementation
123
Sophia Yat Mei Lee The Hong Kong Polytechnic University Hong Kong Hong Kong
ISSN 2522-5103 ISSN 2522-5111 (electronic) Studies in East Asian Linguistics ISBN 978-981-10-6192-9 ISBN 978-981-10-6194-3 (eBook) https://doi.org/10.1007/978-981-10-6194-3 Library of Congress Control Number: 2017955241 © Springer Nature Singapore Pte Ltd. 2019 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Contents
1 Towards a Linguistic Theory of Emotion and Expression of Emotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 What Is Emotion? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.2 Classification of Emotion . . . . . . . . . . . . . . . . . . . . . . 1.2 Linguistic Theories of Emotion . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Natural Semantic Metalanguage . . . . . . . . . . . . . . . . . 1.2.2 The Generative Lexicon . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
1 1 1 5 11 11 15
2 The Linguistic Expression of Emotion and Cause in the Chinese Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Expression of Emotion in Chinese . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Emotion Verbs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2 Emotion Adjectives and Adverbials . . . . . . . . . . . . . . . . . . 2.1.3 Emotions: SADNESS and HAPPINESS . . . . . . . . . . . . . . . . . . . . 2.1.4 Emotion Types and Keywords . . . . . . . . . . . . . . . . . . . . . . 2.1.5 Emotion Classification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Chinese Emotion Verb Classes and Taxonomy . . . . . . . . . . . . . . . 2.3 Cause and Emotion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Emotion Constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Causal Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Emotion Cause Events in Chinese . . . . . . . . . . . . . . . . . . . . . . . . .
17 17 17 19 20 20 23 25 29 29 34 35
3 Linguistic Resources for Study of Emotion . . . . . . . . . . . . 3.1 Empirical Approaches Towards Emotion Analysis . . . . 3.1.1 Corpora . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.2 Chinese Word Sketch . . . . . . . . . . . . . . . . . . . . 3.1.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37 37 37 38 38
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . . . .
. . . . .
. . . . . . .
. . . . .
. . . . . . .
. . . . .
. . . . .
v
vi
Contents
3.2 Emotion Annotated Corpora . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Emotion Annotation Scheme . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Emotion Corpus Construction . . . . . . . . . . . . . . . . . . . . . . .
39 41 42
4 Linguistic Expression of Cause Event I: Transitivity . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Concepts of Transitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Lakoff’s Prototypical Agent-Patient Sentences . . . . . . 4.2.2 Hopper and Thompson’s Transitivity Hypothesis . . . . 4.2.3 Givón’s Transitivity . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.4 Transitivity and Affectedness . . . . . . . . . . . . . . . . . . . 4.3 Cause Event Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Agentivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Kinesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Participation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
47 47 48 48 49 50 51 51 52 53 53 58 58 59 65 68
5 Linguistic Expression of Cause Event II: Epistemicity . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Concepts of Epistemicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Emotion Causes and Epistemic Markers . . . . . . . . . . . . . . . . . 5.4 A Corpus Study of Epistemic Marking of Emotion Causes . . 5.4.1 Methodology and Data Observation . . . . . . . . . . . . . . 5.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
69 69 70 73 75 75 80 82 85
6 A Linguistic Model for Emotion Expression . . . . . . . . . . . . . . . . 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 NSM Reconsidered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Linguistic Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Emotion as Events . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.2 Interaction Between Transitivity and Epistemicity . . . . 6.4 Emotion Representation Model . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
87 87 87 89 89 91 93 99
7 Implementation and Verification: Automatic Detection of Emotion Causes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 7.2 Previous Work on Automatic Emotion Detection and Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Contents
7.3
7.4
7.5
7.6 7.7
vii
7.2.1 Emotion Classification and Representation . . . . 7.2.2 Emotion Processing in Text . . . . . . . . . . . . . . . Emotion Cause Corpus . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 Cause Events. . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.2 Corpus Data and Annotation Scheme . . . . . . . . 7.3.3 Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.4 Corpus Analysis . . . . . . . . . . . . . . . . . . . . . . . . A Rule-Based System for Emotion Cause Detection . . 7.4.1 Cause Event Markers . . . . . . . . . . . . . . . . . . . . 7.4.2 Linguistic Rules for Cause Detection . . . . . . . . Experiment and Discussion . . . . . . . . . . . . . . . . . . . . . . 7.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.2 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . 7.5.3 Results and Discussion . . . . . . . . . . . . . . . . . . . 7.5.4 Error Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
102 104 106 106 108 111 112 121 121 121 128 128 128 132 134 140 140
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
List of Figures
Fig. Fig. Fig. Fig. Fig.
1.1 1.2 1.3 1.4 2.1
Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig.
2.2 3.1 4.1 5.1 5.2 5.3 5.4 6.1 6.2 6.3 6.4
Fig. 7.1 Fig. 7.2 Fig. 7.3
Types of emotion language . . . . . . . . . . . . . . . . . . . . . . . . . . . . Circumplex model of affect (Russell 1980) . . . . . . . . . . . . . . . . Desmet’s circumplex model of emotion (Desmet 2002) . . . . . . The nature of emotions (Plutchik 1994) . . . . . . . . . . . . . . . . . . The structure of the classification of emotional verb frames (Liu and Hong 2008) . . . . . . . . . . . . . . . . . . . . . . Chinese emotion classification tree (Xu et al. 2008) . . . . . . . . . The definition of emotion-related elements . . . . . . . . . . . . . . . . The transitivity continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . The search page of word sketch . . . . . . . . . . . . . . . . . . . . . . . . The word sketch for gao1xing4 . . . . . . . . . . . . . . . . . . . . . . . . Relevant epistemic markers of gao1xing4 in word sketch . . . . Epistemic marking continuum. . . . . . . . . . . . . . . . . . . . . . . . . . Event representation of an emotion. . . . . . . . . . . . . . . . . . . . . . The transitivity continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . Epistemic marking continuum. . . . . . . . . . . . . . . . . . . . . . . . . . The comparison of the transitivity and epistemic marking continuums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two examples of cause event annotation . . . . . . . . . . . . . . . . . The architecture of the rule-based system . . . . . . . . . . . . . . . . . The definitions of metrics for cause detection . . . . . . . . . . . . . .
. . . .
. . . .
3 7 7 8
. . . . . . . . . . .
. . . . . . . . . . .
19 25 41 67 76 77 80 85 91 92 92
. . . .
. 92 . 110 . 128 . 129
ix
List of Tables
Table 1.1 Table 1.2 Table 1.3 Table 2.1 Table 2.2 Table 2.3 Table 2.4 Table Table Table Table Table Table Table Table Table
2.5 2.6 3.1 3.2 3.3 4.1 4.2 4.3 4.4
Table 4.5 Table 4.6 Table 4.7 Table Table Table Table Table Table
5.1 5.2 5.3 5.4 5.5 7.1
Emotions and their derivatives (Plutchik 1994) . . . . . . . . . . Primary emotions and the mixing of primary emotions (Turner 1996) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Proposed semantic primitives (Wierzbicka 1996) . . . . . . . . . Chinese emotion keywords (Xu and Tao 2003) . . . . . . . . . . The verbs of emotion in the Sinica Corpus (Chang et al. 2000) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The dichotomy of emotion verbs (Chang et al. 2000) . . . . . Distributional syntactic differences between two types of verbs (Chang et al. 2000) . . . . . . . . . . . . . . . . . . . . Revised Turner-Plutchik emotion taxonomy . . . . . . . . . . . . . Chinese emotion taxonomy . . . . . . . . . . . . . . . . . . . . . . . . . Primary emotions and the corresponding causes . . . . . . . . . Summary of emotion corpus . . . . . . . . . . . . . . . . . . . . . . . . The accuracy of the emotion-driven corpora . . . . . . . . . . . . Common types of motion cause events . . . . . . . . . . . . . . . . Common types of non-motion cause events . . . . . . . . . . . . . Number of sentences for cause event feature analysis . . . . . Distributional tendency of cause event features of change-of-state emotion verbs . . . . . . . . . . . . . . . . . . . . . . . Distributional tendency of cause event features of homogeneous state emotion verbs . . . . . . . . . . . . . . . . . . . . Cause event features of each emotion class . . . . . . . . . . . . . Comparing cause event features between emotion verb types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . List of primary emotion verbs for analysis . . . . . . . . . . . . . . Types of epistemic markers . . . . . . . . . . . . . . . . . . . . . . . . . Examples of epistemic markers for gao1xing4 . . . . . . . . . . . Epistemic markers of change-of-state emotion verbs . . . . . . Epistemic markers of homogeneous state emotion verbs . . . Summary of cause corpus data . . . . . . . . . . . . . . . . . . . . . . .
..
9
.. .. ..
9 13 21
.. ..
24 24
. . . . . . . . .
. . . . . . . . .
25 27 30 40 44 44 54 56 58
..
65
.. ..
66 66
. . . . . . .
. 67 . 76 . 80 . 81 . 83 . 84 . 109 xi
xii
Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table
List of Tables
7.2 7.3 7.4 7.5 7.6 7.7 7.8 7.9 7.10 7.11 7.12 7.13 7.14 7.15 7.16 7.17
Cause event position of each emotion . . . . . . . . . . . . . . . . . Lists of potential linguistic cues . . . . . . . . . . . . . . . . . . . . . . Position of causative verbs relative to cause events . . . . . . . Position of reported verbs relative to cause events . . . . . . . . Position of epistemic markers relative to cause events . . . . . Position of prepositions relative to cause events. . . . . . . . . . Position of conjunctions relative to cause events . . . . . . . . . Position of other cues relative to cause events . . . . . . . . . . . Seven groups of cause event markers . . . . . . . . . . . . . . . . . . Linguistic rules for emotion cause detection . . . . . . . . . . . . The overall performances . . . . . . . . . . . . . . . . . . . . . . . . . . . The overall accuracy in Phase 1 . . . . . . . . . . . . . . . . . . . . . . The detailed performances in Phase 1 . . . . . . . . . . . . . . . . . The detailed performances in Phase 2 . . . . . . . . . . . . . . . . . The accuracy of each rule . . . . . . . . . . . . . . . . . . . . . . . . . . The contribution of each rule . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
112 114 114 115 117 118 119 120 122 124 133 133 133 134 135 135
Chapter 1
Towards a Linguistic Theory of Emotion and Expression of Emotion
1.1 1.1.1
What Is Emotion? Definition
Early modern philosophy has shown a great interest in the concept of emotions, starting with Aristotle, whose Rhetoric details an insightful theory of emotion which “in many ways anticipates contemporary theories” (Solomon 2003: 5). He argued that the nature of emotion lies in a strong moral belief about how others should behave, and involves the perception of a situation dominated by a desire. This is evidently shown in his analysis of ANGER where ANGER is characterized as the desire for revenge. Two thousand years later, Descartes (1649) defined emotions (which is named ‘passions’ in his book) as “the perceptions, feelings, or emotions of the soul which we relate specially to it, and which are caused, maintained, and fortified by some movement of the spirits”. In reaction to Descartes’s theory, Spinoza (1675: 93), another influential philosopher, suggested that an emotion is “the modifications of the body, whereby the active power of the said body is increased or diminished, aided or constrained, and also the ideas of such modifications”. In his view, an emotion implies an image or idea, and follows from the latter as from its cause. The biologist, Darwin (1859), proposed the theory of evolution in which emotions have an evolutionary history and a survival function in the life of animals. This view of emotions is a functional one in that emotions serve as signals for action. James (1884: 189–190), both a psychologist and philosopher, offered a new way of looking at emotions: “the bodily changes follow directly the perception of the exciting fact, and that our feeling of the same changes as they occur is the emotion”. Contrary to what was commonly believed, James argued that an emotion occurs when one experiences certain bodily changes. In other words, “we feel sorry because we cry, angry because we strike, afraid because we tremble, and not that we cry, strike, or tremble, because we are sorry, angry, or fearful, as the case may © Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_1
1
1 Towards a Linguistic Theory of Emotion …
2
be” (James 1884: 190). While James placed an emphasis on the automatic changes in the internal involuntary organs, Cannon (1927) challenged James’ hypotheses and focused on the changes in the brain instead, specifically the hypothalamus, with the support of some experimental results in physiology. Damasio (1994, 1999) improved on James’ proposal by arguing that an emotion is a neurological process, with accompanying feelings in our brain which play a crucial role in social cognition and decision-making. His idea is supported by the extensive work on emotional deficiencies. Plutchik (1962: 151) considered the matter from a psychological perspective when he claimed that an emotion is “a patterned bodily reaction of either destruction, reproduction, incorporation, orientation, protection, deprivation, rejection or exploration, or some combination of these, which is brought by a stimulus”. He highlighted the stimulus-to-emotion sequence. Similar to Plutchik, Ortony et al. (1988: 13) described emotions as “valenced reactions to events, agents or objects, with their particular nature being determined by the way in which the eliciting situation is construed”. From a linguistic point of view, Harkins and Wierzbicka’s (2001: 2) offered the relevant idea that emotions are “the least tangible aspects of human experience, yet they exert powerful influences upon our thoughts and actions, and even upon our physical appearance and physiological processes occurring within our bodies”. They recognize the complex nature as well as the great impact of emotions. With such varied definitions and concepts of emotion, the question whether “we feel sad because we cry” or “we cry because we feel sad” is still debatable. Despite differing viewpoints, one important issue stands out: namely, that a theory of emotion should be concerned with the relations between conscious feelings and changes inside the body which might be the result of both internal and external stimuli.
1.1.1.1
Emotions and Feelings
Recent literature has examined the distinctions between emotions and feelings. To what extent are they similar? Plutchik (1994: 139) argued that “because emotions are complex states of the organism involving feelings, behaviour, impulses, physiological changes and efforts at control, the measurement of emotions is also a complex process”. This implies that instead of being the same concept, feelings are components of emotions. In addition, emotions are said to be objective in nature as they have a biological foundation, whereas feelings are subjective. Caffi and Janney (1994) suggested that the difference between the two is that emotions have to be objectively triggered by an external stimulus, while feelings are merely subjective physiological arousal. Apart from looking into the causes of emotions or feelings, many researchers have attempted to investigate the differences between the two by their nature. Besnier (1990: 421) pointed out that feelings are “a broad category of person-centered psychophysiological sensations” and emotions, on the other hand,
1.1 What Is Emotion?
3
are “a subset of particularly “visible” and “identifiable” feelings”. Damasio (2003) further supported the idea by proposing that more attention should be paid to one’s bodily state when emotions occur. This implies that during the course of emotions, certain features become visible on one’s face, audible in the voice and identifiable in specific behaviours, whereas feelings that refer to mental representations, states or images are not available to anyone other than the experiencer. Both emotions and feelings are related to thought. The difference is that the former is associated with bodily events or processes, while the latter is not. Therefore, it is emotions, not feelings, which can be studied. Wierzbicka (1999) further explained the difference between emotions and feelings in that the concept of feeling is universal whereas the concept of emotion is culture-bound. Therefore, the concept of ANGER may differ across cultures. In human experience, it is common to use the term emotion to describe the state of feeling, but in fact, emotions are considerably more complex.
1.1.1.2
Expressive Versus Descriptive Emotions
Emotion words are basically classified into two groups, i.e., expressive and descriptive (Kövecses 2000). Figure 1.1 shows the types of emotion language. Expressive emotion words are words that express emotions directly, such as Shit, Wow, Yuck, whereas descriptive emotion words are words that describe emotions, such as ANGER, JOY, SADNESS, etc. Potts (2007), in a more general sense, divided semantic types into descriptive and expressive types. This can be extended to the understanding of emotion in that of sentences, such as “I am angry, I left my keys in the car!” is described as descriptive, while a sentence, such as “Damn, I left my keys in the car!” is considered expressive. Descriptive emotion words have mainly been the focus of most research on emotions. Among the descriptive emotion words, there are literal and figurative emotion terms. For the literal emotion terms, some are basic while others are nonbasic. Basic emotion terms include more basic ones, such as ANGER, SADNESS,
emotion language
expressive
descriptive figurative
literal basic Fig. 1.1 Types of emotion language
nonbasic
metaphor
metonymy
1 Towards a Linguistic Theory of Emotion …
4 FEAR, JOY, ANGER.
and LOVE and less basic ones, such as ANNOYANCE, WRATH, and RAGE for The sense of basicness can be presented as a vertical hierarchy shown
below: Superordinate level emotion Middle (basic level) anger Subordinate level annoyance As people’s emotional responses differ, emotions cannot be defined purely in terms of situation, context or eliciting conditions. It is the language that provides a conceptual connection between two disparate emotion experiences by giving them the same label. In addition, emotion concepts are often untranslatable. For example, Doi (2004) argued that amae in Japanese, which describes the emotional state of a baby who wants to receive love from its parents, cannot be directly translated into English. Neither the diversity nor the universal aspects of emotions can be studied without an appropriate metalanguage. Seemingly, all attempts to study emotions in terms of English are bound to lead to distortions.
1.1.1.3
Primary Human Emotions
It is assumed that a small number of emotions are primary emotions and that other emotions are secondary emotions which are the mixtures of primary emotions. Many different lists of basic emotions have been proposed; there is, nonetheless, no agreement either on the number of basic emotions or on the classes of emotions. For example, Kemper (1987) suggested that there are at least four basic emotions, i.e. FEAR, ANGER, SADNESS, and SATISFACTION. The rationale for considering them to be basic is that (1) they can be observed in most animals; (2) they are universally found in all cultures; (3) they appear early in the course of human development; (4) they are outcomes of power and status interactions; and (5) they are associated with distinct automatic patterns of physiological changes. Similar to Kemper, Ortony and Turner (1990) also gave reasons for assuming the existence of basic emotions: (1) some emotions appear to exist in all cultures; (2) some can be identified in other higher animals; (3) some have characteristic facial expressions; and (4) some seem to increase the chances of survival. Various researchers (see e.g. Sabini and Silver 2005; Keltner et al. 2014; Scheff 2015) have proposed lists of basic emotions where the list varies from two to ten items. A summary of primary emotions proposed in recent theories is presented in Appendix 1. FEAR and ANGER appear on every list, whereas HAPPINESS and SADNESS appear on most of the lists. These four emotions, i.e. FEAR, ANGER, HAPPINESS, and SADNESS, are the most common primary emotions. Other less common primary emotions are SURPRISE, DISGUST, SHAME, DISTRESS, GUILT, INTEREST, PAIN, and ACCEPTANCE.
1.1 What Is Emotion?
1.1.2
5
Classification of Emotion
Researchers have attempted to study emotion classification from a wide range of perspectives in different fields, from linguistics through neuropsychology to computer science. A consensus has yet to be reached on the classification of emotions. This section will discuss the various approaches in detail.
1.1.2.1
Linguistic-Based Approaches
Emotional meaning is linguistically characterized in different views (Kövecses 2000): the ‘label’ view, the ‘core meaning’ view, the ‘dimensional’ view, the ‘implicational’ view, the ‘prototype’ view, the ‘social-constructionist’ view, and the ‘embodied cultural prototype’ view. The ‘label’ view suggests an association between labels, e.g., ANGER and FEAR, and some real emotional phenomena, e.g., physiological processes and behaviour. In this view, emotion terms do not have much conceptual context and organization. However, this view has been questioned in many studies. An important representative of the ‘core meaning’ view is the semantic primitive approach proposed by Wierzbicka (1972). According to Wierzbicka, meanings can be core or peripheral. Core meaning is more important, whereas peripheral meaning is less important. Emotional meaning is composed of universal core meanings. This will be discussed in Chap. 2 in further detail. The ‘dimensional’ view considers emotional meaning as being constituted by values on a fixed set of dimensions of meaning. Solomon (1976) proposed the following dimensions of meaning: DIRECTION, SCOPE/FOCUS, OBJECT, CRITERIA, STATUS, EVALUATIONS, RESPONSIBILITY, INTERSUBJECTIVITY, DISTANCE, MYTHOLOGY, DESIRE, POWER, and STRATEGY. According to Solomon, the dimensions that apply to a given emotion provide a component profile that uniquely characterizes that emotion. The ‘implicational meaning’ approach focuses on connotative emotion meaning, in contrast with the core meaning, i.e. what it implies or suggests to those who understand it. The connotative meaning of emotion varies from culture to culture. For instance, the concept of ANGER (Shweder 1991) is very different in its connotative implications for the Ilongots, for example, for whom ANGER is so dangerous that it can destroy society; while for the Eskimos, it is something that only children experience, and for the Americans ANGER helps overcome fear and attain independence. The ‘prototype’ view looks for best examples of the category of emotion. For example, a prototypical example of FEAR suggested by Fehr and Russell (1984: 482) is described as: A dangerous situation occurs suddenly. You are startled, and you scream. You try to focus all your attention on the danger, try to figure a way out, but you feel your heart pounding and your limbs trembling. Thoughts race through your mind. Your palms feel cold and wet. There are butterflies in your stomach. You turn and flee.
1 Towards a Linguistic Theory of Emotion …
6
The ‘social-constructionist’ view is that the properties of emotion meaning depend on particular aspects of the society and culture. For instance, Lutz (1988) gave an account of ANGER in Ifaluk: 1. 2. 3. 4. 5.
There is a rule or value violation It is pointed out by someone This person simultaneously condemns the act The perpetrator reacts in fear to that anger The perpetrator amends his or her ways.
This is different from how the English word ANGER functions. An angry person in English speaking cultures needs not confront a perpetrator; and, even if there is confrontation, there may not be changes in a perpetrator’s behavior. For the ‘embodied cultural’ view, the concept of emotion is both motivated by the human body and produced by a particular social and cultural environment. An example would be anger being expressed by increased body heat.
1.1.2.2
Neuropsychological-Based Approaches
There has been considerable debate concerning how emotions should be classified in the field of neuroscience and psychology. Plutchik (1962) pointed out that at that time there was no single, integrating, and comprehensive theory which had relevance to all aspects of the field. Russell (1980) proposed the Circumplex Model of Affect in which emotion concepts are organized according to a circular structure in a two-dimensional space of pleasure-displeasure and degree of activation as shown in Fig. 1.2.1 In this model, there are eight primary emotions; each can be either pleasant or unpleasant with different levels of activation which is indicated by increased heartbeat and increased left-prefrontal activity in the brain. For example, unpleasant emotions with high activation are TENSE, NERVOUS, STRESSED, and UPSET; pleasant emotions with low activation are CALM, RELAXED, SERENE, and CONTENTED. Based on Russell’s model, Desmet (2002) suggested a circumplex model of 41 Product Relevant Emotions as shown in Fig. 1.3.2 The dimension of pleasantness can be classified into PLEASANT, NEUTRAL, or UNPLEASANT; while activation can be EXCITED, AVERAGE, or CALM. Instead of bi-polar dimensions, Plutchik (1994) proposed a three-dimensional circumplex model which describes the relations among emotions. This model is similar to a color wheel as shown in Fig. 1.4. It describes the relations among emotion concepts analogous to relations among the colors on a color wheel. The cone’s vertical dimension represents intensity and the circle represents degrees of similarity among the emotions. The eight sectors are designed to indicate that there
1
The clearer version of Russell’s Circumplex Model appears in Posner et al. (2005). The figure is taken from Desmet and Hekkert (2007).
2
1.1 What Is Emotion?
7
Fig. 1.2 Circumplex model of affect (Russell 1980)
Fig. 1.3 Desmet’s circumplex model of emotion (Desmet 2002)
are eight primary emotion dimensions defined by the theory which are arranged as four pairs of opposites, i.e. JOY-SADNESS, TRUST-DISGUST, FEAR-ANGER, and SURPRISEANTICIPATION. In the extended model, the emotions in the blank spaces are the primary dyads, i.e. emotions that are mixtures of two primary emotions. For example, REMORSE is a blend of DISGUST and SADNESS, whereas LOVE is a blend of JOY and TRUST. Mixtures of three primary emotions are secondary dyads. Although emotional substrates cannot always be discerned in the behavior of nonhuman animals, Plutchik (1980) observed that many stimuli are experienced by
8
1 Towards a Linguistic Theory of Emotion …
Fig. 1.4 The nature of emotions (Plutchik 1994)
people and animals alike and result in prototypical behavior followed by, generally, the re-establishment of an equilibrium state that might not have been achieved without the impulse precipitated by the inner state. The stimulus event, cognition, overt behavior, and effect of each proposed emotion are given in Table 1.1. Turner (1996) offered another possible view on how primary emotions can be combined into variants, first-order combinations, and second-order combinations, as shown in Table 1.2. Turner (1996) proposed five basic emotions, i.e. HAPPINESS, ANGER, FEAR, SADNESS, and SURPRISE (SURPRISE is considered optional in his later model published in 2000). He sees HAPPINESS as satisfaction, FEAR as aversion, ANGER as assertion, and SADNESS as disappointment. The combination of two basic emotions forms a first-order emotion, e.g., FEAR + SADNESS = WORRY, while the combination of three emotions
1.1 What Is Emotion?
9
Table 1.1 Emotions and their derivatives (Plutchik 1994) Stimulus event
Cognition
Feeling state
Overt behavior
Effect
threat obstacle gain of valued object loss of valued object member of one’s group unpalatable object new territory
“danger” “enemy” “possess”
fear anger joy
safety destroy obstacle gain resources
“abandonment”
sadness
escape attack retain or repeat cry
“friend”
acceptance
groom
“poison” “examine”
disgust expectation
vomit map
unexpected event
“what is it?”
surprise
stop
reattach to lost object mutual support eject poison knowledge of territory gain time to orient
Table 1.2 Primary emotions and the mixing of primary emotions (Turner 1996) Primary emotions
Range of emotions
First-order combinations
HAPPINESS
satisfaction contentment affection love terror anxiety apprehensiveness aversiveness annoyance distaste loathing contempt aggressiveness resignation ennui sorrow
pride, gratitude, hope, relief, wonder (FEAR) vengeance (ANGER) joy, ecstasy (SURPRISE)
FEAR
ANGER
SADNESS
SURPRISE
startlement amazement astonishment
awe (HAPPINESS) guilt, envy (ANGER) worry (SADNESS) panic, anticipation(SURPRISE) shame, hate, jealousy (FEAR) snubbing (HAPPINESS) depression (SADNESS) rage, fury (SURPRISE) yearning (HAPPINESS) hopefulness (FEAR) grief, boredom (ANGER) crestfallen (SURPRISE) delight (HAPPINESS) shock (FEAR) disgust (ANGER) disappointment (SADNESS)
forms a second-order emotion, e.g., SADNESS + FEAR + ANGER = GUILT. Details of the classification are given in Table 1.2. The main difference between this model and Plutchik’s (1980) is that the combination of emotions involves a dominant primary emotion and lesser amounts of another primary emotion. For example, the emotions of SHAME, HATE, and JEALOUSY are a combination of ANGER and FEAR, with greater amounts of ANGER and varying amounts of FEAR.
10
1.1.2.3
1 Towards a Linguistic Theory of Emotion …
Computational-Based Approaches
In the past decade, computer scientists have attempted to develop ways to recognize, interpret, and process human emotions. Emotion also plays an important role in human interaction with the computer. Why? Emotions have often been identified based on speech, facial expressions, and physiological phenomena, such as voice, gestures, and blood flow. Recent research has begun to place more emphasis on automatic emotion detection from textual input (see Chuang and Wu 2002; Mihalcea and Liu 2006; Ahmad 2008; Strapparava and Mihalcea 2008). Identifying emotions in text has become crucial with higher expectations of human-like computers. Computers are expected to be able to recognize users’ mental states through their language use. Therefore, emotionally intelligent products, such as emotional computers and emotional robots are greatly needed. In the literature, there are several approaches to automatic emotion detection, each with its strengths and weaknesses. The existing well-known approaches include keyword identification models, lexical affinity models, statistical models, and real-world knowledgebase models. I will give a brief picture of how these models work as well as their limitations for modeling emotion detection. The keyword identification model is the most commonly used method for identifying emotions from text (Ortony et al. 1988; Elliot 1992; Wiebe et al. 2005). A piece of text is classified into emotion categories based on the presence of fairly unambiguous emotion words, emotion intensity modifiers, and a handful of cue phrases. However, this approach does not accurately identify emotions as it fails to recognize the emotion correctly when negation is involved in the sentence. For instance, a sentence such as “Mary is not happy today” will be classified as a happy condition due to the keyword happy. In addition, it heavily relies on surface features. This poses a great problem when an emotion is described without the presence of a keyword, i.e. a descriptive emotion, such as “it has been a tragic waste of money!”, which is not uncommon in texts. In contrast, the lexical affinity method detects more than just obvious emotion keywords. It assigns arbitrary words a probabilistic affinity for a particular emotion and classifies the emotion based on the affinity of the arbitrary words and an affective keyword (Subasic and Huettner 2001; Stevenson at el. 2007; Mohammad and Turney 2013). Such an approach provides a higher accuracy than keyword-based models. However, it is far from perfect. First, it requires a large-scale corpus to get statistically significant results. Second, the fact that it operates solely on the word level leads to problems with negation and ambiguous sentences similar to those found in keyword identification. Third, there is bias towards texts of a particular genre. These factors make it hard to develop a reusable, domain-independent model. Unlike the previous methods, a statistical model feeds a machine learning algorithm a large training corpus of texts previously annotated for emotions (Alm et al. 2005; Mishne 2005; Danisman and Alpkocak 2008; Poria et al. 2013). It learns the valence of emotion and other arbitrary keywords, punctuation, and word occurrence frequencies.
1.1 What Is Emotion?
11
Its performance has been quite promising even though it is often criticized for being unsuccessful on smaller text units that are semantically weak. Real-world knowledgebase models inherit the affective nature of everyday situations to classify sentences into emotion categories based on large-scale real-world knowledgebases (Liu et al. 2003; Cambria et al. 2009; Poria et al. 2014). Unlike keyword-based and lexical affinity models, it can successfully detect emotions even if the keywords are absent. The main problem, however, is the limited available large-scale real-world knowledgebases. Picard’s (1995/2000, 2007, 2010) work on affective computing serves as a pioneer in the field. She examined emotions from two perspectives: philosophical/ psychological, as well as implementational. She detected emotion information with passive sensors, which in turn identified users’ physical state or behaviour without interpreting the input.
1.2
Linguistic Theories of Emotion
In Sect. 1.1, various previously proposed theories of emotions were discussed. These theories were developed based on different perspectives and thus are designed to deal with different aspects of emotions. This section discusses two of these frameworks, Natural Semantic Metalanguage and Generative Lexicon, which will be adopted with some modifications in the later analyses. Although neither framework is specifically designed for emotion analysis, they provide a comprehensive and practical way for accounting for emotions. Wierzbicka’s Natural Semantic Metalanguage model (1972) defines emotions from a cognitive perspective, in which emotions are concretely described in terms of universal, irreducible semantic primitives. Pustejovsky’s (1991, 1995) Generative Lexicon theory presents a lexical semantic analysis of emotion predicates and describes how causal relations are lexicalized in languages. The details of the two theories are presented in Sects. 1.2.1 and 1.2.2.
1.2.1
Natural Semantic Metalanguage
Devised by Wierzbicka (1972), Natural Semantic Metalanguage (NSM) is one of the most well-developed, comprehensive, and practical approaches to cross-cultural semantics. It describes complex and abstract concepts in terms of simpler and concrete terms. In particular, emotions are decomposed into complex events involving a cause and a mental state which can be further described by a set of universal, irreducible cores called semantic primitives. There are two fundamental assumptions: “[that] every language has an irreducible core in terms of which the speakers can understand all complex thoughts and utterances, and that the irreducible cores of all natural languages match, so that
1 Towards a Linguistic Theory of Emotion …
12
we can speak, effectively, of the irreducible core of all languages, reflecting the irreducible core of human thought” (Wierzbicka 1998: 113). Therefore, a complex meaning, such as emotion, is also composed of an irreducible core (i.e. universal semantic primitive). While the concepts HAPPINESS and ANGER are culture-bound and language-specific, semantic primitives, such as GOOD and BAD are not. What makes these semantic primitives universal is that they appear to have their semantic equivalents in all languages of the world. When the NSM model was first introduced in the early 1970s, only 14 semantic primitives made up its repertoire. Over the 1980s and 1990s, however, the number of proposed primes was greatly expanded, reaching its current total of 60 or so, as shown in Table 1.3. This set of semantic primitives can describe some basic themes characterizing different emotion concepts independently of a specific language. Although emotion prototypes are different cross-culturally, their semantic primitive equivalents can be universal. The semantic structure of most emotion concepts can be represented as: (1) Standard Mode of Semantic Descriptions X felt something because X thought something Sometimes a person thinks something like this: … … … X felt something like this Because X thought something like this (1) shows the standard representation that applies to all emotion descriptions. This template of emotions suggests that an emotion refers to the perception of a human towards something. Emotions, for instance HAPPY and FEAR, can be described as the semantic structures in (2) and (3): (2)
HAPPY
(a) (b) (c) (d) (e) (f) (g) (h) (3)
X felt something because X thought something Sometimes a person thinks: “something good happened to me I wanted this I don’t want other things” When this person thinks this, this person feels something good X felt something like this Because X thought something like this
FEAR
(a) (b) (c) (d)
(X was happy)
(X felt fear)
X felt something because X thought something Sometimes a person thinks: “I don’t know what will happen Some bad things can happen
1.2 Linguistic Theories of Emotion
13
Table 1.3 Proposed semantic primitives (Wierzbicka 1996) Substantives Relational substantives Determiners Quantifiers Evaluators Descriptors Mental predicates Speech Actions, events, movement, contact Location, existence, possession, specification Life and death Time Space “Logical” concepts Intensifier, augmentor Similarity
(e) (f) (g) (h)
I, YOU, SOMEONE, PEOPLE, SOMETHING/THING, BODY KIND, PART THIS, THE SAME, OTHER/ELSE ONE, TWO, SOME, ALL, MUCH/MANY GOOD, BAD BIG, SMALL THINK, KNOW, WANT, FEEL, SEE, HEAR SAY, WORDS, TRUE DO, HAPPEN, MOVE, TOUCH BE (SOMEWHERE), THERE IS, HAVE, BE (SOMEONE/ SOMETHING) LIVE, DIE WHEN/TIME, NOW, BEFORE, AFTER, A LONG TIME, A SHORT TIME, FOR SOME TIME, MOMENT WHERE/PLACE, HERE, ABOVE, BELOW, FAR, NEAR, SIDE, INSIDE NOT, MAYBE, CAN, BECAUSE, IF VERY, MORE LIKE
I don’t want these things to happen” When this person thinks this, this person feels something bad X felt something like this Because X thought something like this
The emotion HAPPY in (2) shows that lines (a), (b), (f), (g) and (h) form the template for describing any emotion, whereas (c)–(e) provide the overt and concrete description of the cause of the specific emotion. The cause requires the perception of a human being and would in turn lead to certain thoughts, i.e. (f), (g), and (h). The cause of HAPPY is described as “something good happened, I wanted this to happen”. In other words, “something happened” is perceived as good, which makes the person feel good. This representation describes how an emotion is perceived and what causes this emotion. “Something happened” is, however, not always the cause for emotions, but “something may happen” as in (3). The cause of FEAR is described as “I don’t know what will happen, some bad things can happen, I don’t want these things to happen”. It shows that FEAR is not triggered by the exact event that happened, but by the potential threat that one thinks may occur. The potential threat is perceived as something bad, which makes the person feel bad. Apart from identifying the differences between different emotion concepts, NSM can also describe or represent emotion differences across cultures. This is illustrated
1 Towards a Linguistic Theory of Emotion …
14
by different ways of describing (Wierzbicka 1992).
HAPPINESS
in the context of English and Polish
(4) Sčastliv(yj) (a) (b) (c) (d) (e) (f) (g) (h) (i)
X felt something because X thought something Sometimes a person thinks: “Something very good happened to me I wanted this everything is good I can’t want other things” When this person thinks this, this person feels something very good X felt something like this Because X thought something like this
In Polish, the emotion HAPPINESS is not used as often as in English or other languages. It is “generally reserved for rare states of profound bliss, or total satisfaction with serious things such as love, family, the meaning of life, and so on” (Barańczak 1990: 12). In contrast with (2), (4) shows a higher level of HAPPINESS as expressed by “something very good happened”, “everything is good”, and “I can’t want other things”. Such an approach to defining emotions identifies the exact differences and connections between emotion concepts in terms of human conceptualization of emotion causes both within a culture and across cultures. This provides an immediate cue for emotion classification. There are, however, some limitations regarding the NSM. Kripke (1972), from a philosophical perspective, argues that the meaning of proper names cannot be adequately paraphrased by any description or set of descriptions. These arguments are extended to terms of natural kind (e.g., human, water, etc.), and to other types of predicates. Since proper names and terms of natural kind carry a significant portion of the burden of meaning in every natural language, these arguments strongly suggest that the NSM model cannot provide a complete account of natural language meaning. In addition, NSM strictly limits the way that paraphrases are able to talk about the linguistic structure undergoing paraphrasing. It cannot provide adequate accounts of such fundamental semantic phenomena as indexicality, performatives, or presupposition, unless the concepts used for paraphrases are non-basic, i.e. systematically decomposable into more basic semantic elements. Moreover, it assumes a uniform way of describing each emotion with a small set of semantic primitives, which is, instead, semantically motivated. This highly abstract model is difficult to verify empirically. NSM also lacks a full account of causal relations, such as the question of how the causal relation is expressed linguistically. Moreover, there is no way to express presuppositional meaning as the set of semantic primitives lacks a suitable range of presupposition triggers. However, with an in-depth analysis of cause events, the NSM model could still offer a plausible framework to account for emotions.
1.2 Linguistic Theories of Emotion
1.2.2
15
The Generative Lexicon
The Generative Lexicon (GL) is a theory of lexical semantics that focuses on the nature of lexical compositionality. The theory was first proposed by Pustejovsky (1991). It emphasizes the importance of the lexicon which encodes much of the information conveyed by language use. The lexically encoded knowledge is said to be exploited in the construction of interpretations for linguistic utterance (Pustejovsky 1995). In GL theory, a lexical item consists of four levels: (5) a. b. c. d.
Lexical Typing Structure Argument Structure Event Structure Qualia Structure.
Lexical typing structure gives an explicit type for a word positioned within a type system for the language and provides an explicit link to commonsense knowledge; argument structure specifies the number and nature of the arguments to a predicate; event structure defines the event type of the expression and any subeventual structure it may have; and qualia structure identifies the structural differentiation of the predicative force for a lexical item defined as formal, constitutive, telic, and agentive. Formal refers to the basic category that distinguishes the meaning of a word within a larger domain; constitutive shows the relation between an object and its constituent parts; telic identifies the purpose or function of the object; and agentive refers to the factors involved in the object’s origins. The conventional representation of the four levels is given in (6).
(6)
α ARGSTR
EVENTSTR
QUALIA
ARG1 = x … =
E1 = e1 …
=
ONST = what x is made of FORMAL = what x is TELIC = function of x AGENTIVE = how x came into being
16
1 Towards a Linguistic Theory of Emotion …
Making use of the four-level representation of qualia structures, Pustejovsky (1995) presents various mechanisms including coercion, type shifting, co-composition, selective binding, and semantic selection. These concepts will be discussed in subsequent chapters where related issues are dealt with. Although GL is not a theory designed for emotion analysis per se, it offers a comprehensive representation for various causative constructions. It also discusses how nominals and nominalizations are represented, which links syntax and semantics to real world knowledge.
Chapter 2
The Linguistic Expression of Emotion and Cause in the Chinese Language
2.1
Expression of Emotion in Chinese
Despite the long history of emotion study in research in general, there is not much work on the study of emotions in Chinese. This section reviews the previous works from varying perspectives. The following sections also discuss various fundamental issues regarding the present research, such as the proposed emotion verb classes and taxonomy in Sect. 2.2. Furthermore, Sect. 2.3 gives clear definitions of important emotion-related concepts, such as emotional sentences and causal relations. Section 2.4 describes how emotion causes, as tangible realizations of emotion, are linguistically expressed in Chinese.
2.1.1
Emotion Verbs
Liu and colleagues (Liu 2002, 2009; Liu and Hong 2008; Liu et al. 2009) have conducted a series of studies on Mandarin emotion verbs. Liu (2002) first examined verbs of emotional activity by focusing on the two near-synonym pairs, 羨慕 xian4mu4 ‘to envy’ and 嫉妒 ji2du4 ‘to be jealous’. She categorized two groups of emotional activity verbs, Group A including verbs, such as xian4mu4 and Group B such as 嫉妒 ji2du4. Similar to Chang et al. (2000) which will be discussed in Sect. 2.1.5, the two groups of emotional activity verbs differ in their markedness, grammatical functions, cause occurrence, and verbal aspect. Group A verbs tend to be unmarked, are used more as predicates, are associated with overt cause, and are externally caused. Group B verbs, on the other hand, tend to be marked, are used more as nominalized forms, are not associated with overt cause, and are internally caused. This work is extended in Liu and Hong (2008). They classify Chinese emotion verbs into nine frames based on FrameNet’s emotion categorization (Fillmore and © Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_2
17
18
2 The Linguistic Expression of Emotion and Cause …
Atkins 1992)1 and corpus observation. The nine frames include FEELING, EMOTION_DIRECTED, EXPERIENCER_SUBJ, EMOTION_ACTIVE, CONTRITION, CAUSE_TO_EXPERIENCE, EXPERIENCER_OBJ, JUDGMENT, AND FORGIVENESS. Based on their syntactic and semantic properties, the verbs are further generalized into two classes: complement-requiring verbs (which require a complement) and emotion-predicating verbs (which do not require any complement). Complement-requiring verbs are further divided into two groups: emotion-taking verbs and complement-taking verbs. An emotion-taking verb takes an emotional state or emotion-predicating verb as its complement, such as 感覺 gan3jue2, 感到 gan3dao4, 覺得 jue2de, all meaning ‘to feel/think’. A complement-taking verb ‘indicates that a judge refrains from imposing, or demanding a punishment for an evaluee who has committed an offense’, such as 寬恕 kuan1shu4 and 原諒 yuan2liang4, both meaning ‘to forgive’. This classification of emotion verb frames is presented in Fig. 2.1. Liu and Hong (2008) argued that such a classification of emotion verb frames accounts for the syntax-to-semantic variations among emotion verbs and is more suitable for Mandarin emotion verbs. However, the rationales of the proposed nine frames are not clearly discussed. In addition, some of the frames are in fact not emotion verbs. For instance, verbs such as 感到 gan3dao4 and 覺得 jue2de in the frame FEELING are usually not considered emotion verbs in most emotion theories. By examining their morphological, semantic, and syntactic properties, Liu (2009) proposed that there are three lexicalization patterns in the lexicon of emotion verbs, namely experiencer as subject, stimulus as subject, and affector as subject. The predominant pattern decides whether a language is experiencer-prominent, stimulus-prominent, or affect-prominent. Mandarin is regarded as one of the experiencer-prominent languages, whereas English is regarded as one of the stimulus-prominent languages. In addition, Liu et al. (2009) explored, in particular, the interface between lexical semantics and pragmatics in a group of Mandarin emotion verbs, 可 ke3+V, as illustrated with the verb 可憐 ke3lian2 ‘to be pitiful’. They argued that the semantics of 可憐 ke3lian2 may shift from subject-oriented emotional state to speaker-oriented personal judgment on the proposition.
1
FrameNet is an on-line lexical resource for English based on semantic frames and supported by corpus evidence. A semantic frame can be thought of as a concept with a script. It is used to describe an object, state or event. The aim of FrameNet is to document the range of semantic and syntactic combinatory possibilities (valences) of each word in each of its senses. The FrameNet lexical database currently contains more than 11,600 lexical units, more than 6800 of which are fully annotated, in more than 960 semantic frames, exemplified in more than 150,000 annotated sentences. Official website: http://framenet.icsi.berkeley.edu/.
2.1 Expression of Emotion in Chinese
19
Fig. 2.1 The structure of the classification of emotional verb frames (Liu and Hong 2008)
2.1.2
Emotion Adjectives and Adverbials
Zhao (2007) identified five semantic features for Chinese emotion adjectives,2 including self-experience, self-evidence, volatility, causation, and preference. He also examined the relations between emotion adjectives and seven semantic roles of nominal arguments. The semantic roles include EXPERIENCER, ATTRIBUTE, PRESENTATION, ENGENDERER, CAUSER, SPACE-TIME ENTITY, and LOGIC ENTITY. He found that causers only occur with experiencers and adjacent roles do not co-occur. In addition, he proposed an argument hierarchy to account for the semantic correlations between arguments and to show the rules of co-occurrence arguments, i.e. core, basic, conditional, and cognitive. Zhao (2009) identified three types of semantic relationship between the emotional adverbials and the verbal predicates: result-cause, cause-result, and parallel. For the result-cause constructions, the emotional adverbial serves as the result, whereas the verbal predicate functions as the cause, including the three common verb types, i.e. ‘to know’, ‘to lose’, and ‘to run into’. For the cause-result constructions, there are four types of predicates, i.e. actions, bodily changes, abilities, and mental states. The parallel constructions refer to the action accompanied by the emotional state, whereas the predicates are mainly associated with speaking and seeing, the body, expectations, and head and face.
2
There is a controversy over the existence of adjectives in Chinese since most of the supposed adjectives function as verbs syntactically. In this book, adjectives are considered as verbs.
20
2.1.3
2 The Linguistic Expression of Emotion and Cause …
Emotions: SADNESS and HAPPINESS
Ye (2001) examined the differences among the three SADNESS emotions in Chinese literary texts, i.e. 悲 bei1, 哀 ai1, and 愁 chou2 which are usually claimed to be interchangeable. She showed that the three emotion words are by no means interchangeable, nor are they equivalent to Western characterizations of SADNESS. She found that there is a strong connection between SADNESS-like (e.g., 悲 bei1 and 哀 ai1) and WORRY-like emotions (e.g., 愁 chou2) in Chinese. Thus glossing 愁 chou2 as SADNESS would distort the true picture of Chinese people’s categorization of emotion domains. Hence, it is the cognitive element found in emotions that is fundamental to the universality of emotion, which is indeed shaped by culture. Ye (2006) also investigated the two supposedly interchangeable JOY-like emotion words in the text of Hongloumeng,3 namely 喜 xi3 and 樂 le4 ‘to be happy’. She noticed that 喜 xi3 and 樂 le4 are not interchangeable in that 喜 xi3 “can be triggered by mysterious external forces, and is sudden and unexpected”, whereas 樂 le4 has “an earthy, material, and secular sense that is deeply rooted in human effort and is on-going” (Ye 2006: 70). In addition, 樂 le4 specifies the source or cause of the good event which is absent from 喜 xi3. Overall, Ye’s works on Chinese emotions shed light on the basic Chinese emotional experience and offer implications for the discussion of whether there are basic emotions in all cultures.
2.1.4
Emotion Types and Keywords
Xu and Tao (2003) defined Chinese emotions as comprising basic emotions, psychological emotions, and behavioral emotions. Based on readings in Chinese literature, they argued that there are seven basic emotions in Chinese, i.e. LOVE, HAPPINESS, DISGUST, ANGER, SADNESS, FEAR, and DESIRE. From the psychological perspective, they identified certain emotion keywords. Behavioral emotions refer to the factors influencing one’s behavior, such as attitude and personality. They then classified Chinese emotions into a total of 24 groups with 390 keywords, as shown in Table 2.1. Xu and Tao’s grouping of emotions and list of emotion verbs are comprehensive; however, the definition of emotions in general is not sufficiently clear. Apart from emotions, they also include keywords of feelings, attitudes, and personality, which may not be suitable for emotion analysis.
3
Hongloumeng, Dream of the Red Chamber, composed by Cao Xueqin, is one of China’s Four Great Classic Novels. It was composed sometime in the middle of the 18th century during the Qing Dynasty. It is a masterpiece of Chinese vernacular literature and is generally acknowledged to be the pinnacle of classical Chinese novels (Wikipedia 2010).
2.1 Expression of Emotion in Chinese
21
Table 2.1 Chinese emotion keywords (Xu and Tao 2003) No.
Category
No. of words
List of words
1
喜 xi3, 樂 le4
45
稱心 chen4xin1, 痛快 tong4kuai4, 得意 de2yi4, 欣慰 xin1wei4, 高興 gao1xing4, 愉悅 yu2yue4, 欣喜 xin1xi3, 歡 欣 huan1xin1, 可意 de2yi4, 樂 le4, 可心 ke3xin1, 歡暢 huan1chang4, 開心 kai1xin1, 康樂 kang1le4, 歡快 huan1kuai4, 快慰 kuai4wei4, 歡 huan1, 舒暢 shu1chang4, 快樂kuai4le4, 快活 kuai4huo2, 歡樂 huan1le4, 暢快 chang4kuai4, 舒心 shu1xin1, 舒坦 shu1tan, 歡娛 huan1yu2, 如意 ru2yi4, 喜悅 xi3yue4, 順心 shun4xin1, 歡悅 huan1yue4, 舒服 shu1fu, 爽心 shuang3xin1, 曉暢 xiao3chang4, 鬆快 song1kuai4, 幸福 xing4fu2, 驚喜 jing1xi3, 歡愉 huan1yu2, 稱意 chen4yi4, 得志 de2zhi4, 情 願 qing2yuan4, 願意 yuan4yi4, 歡喜 huan1xi3, 振奮 zhen4fen4, 樂意 le4yi4, 留神 liu2shen2, 樂於 le4yu2 關懷 guan1huai2, 偏愛 pian1ai4, 珍愛 zhen1ai4, 珍惜 zhen1xi1, 神往 shen2wang3, 痴迷 chi1mi2, 喜愛 xi3ai4, 器 重 qi4zhong4, 嬌寵 jiao1chong3, 溺愛 ni4ai4, 珍視 zhen1shi4, 喜歡 xi3huan1, 動心dong4xin1, 掛牽 gua4qian1, 讚賞 zan4shang3, 愛好 ai4hao4, 滿意 man3yi4, 羨慕 xian4mu4, 賞識 shang3shi2, 熱愛 re4ai4, 鍾愛 zhong1ai4, 眷戀 juan4lian4, 關注 guan1zhu4, 贊同 zan4tong2, 喜歡 xi3huan1, 想 xiang3, 掛心 gua4xin1, 掛念 gua4nian4, 惦念 dian4nian4, 掛慮 gua4lü4, 懷念 huai2nian4, 關切 guan1qie4, 關心 guan1xin1, 牽掛 qian1gua4, 憐憫 lian2min3, 同情 tong2qing2, 吝嗇 lin4se4, 可惜 ke3lian2, 憐惜 lian2xi1, 感謝 gan3xie4, 感激 gan3ji1, 在乎 zai4hu, 操心 cao1xin1 窩心 wo1xin1, 沉悶chen1men4, 憋氣bie1qi4, 鬱悒yu4yi4, 不悅bu2yue4, 悲愁bei1chou2, 苦悶ku3men4, 苦惱 ku3nao3, 無聊 wu2liao2, 愁苦 chou2ku3, 困惑 kun4huo4, 愁悶 chou2men4, 窩囊 wo1nang, 鬱悶 yu4men4, 惆悵 chou2chang4, 乏味 fa2wei4, 沉鬱 chen2yu4, 憋悶 bie1men4, 憂愁 you1chou2, 哀愁 ai1chou2, 憂鬱 you1yu4, 陰鬱 yin1yu4, 不快 bu2kuai4, 發愁 fa1chou2, 煩悶 fan2men4, 悵惘 chou2chang4, 擔憂 dan1you1, 擔心 dan1xin1, 犯愁 fan4chou2, 愁 chou2, 憂慮 you1lü4 苦 ku3, 哀怨 ai1yuan4, 悲慟 bei1tong4, 悲痛 bei1tong4, 哀 傷 ai1shang1, 慘痛 can3tong4, 沉重 chen2zhong4, 感傷 ai1shang1, 悲壯 bei1zhuang4, 酸辛 suan1xin1, 傷心 shang1xin1, 辛酸 xin1suan1, 悲哀 bei1ai1, 哀痛 ai1tong4, 沉痛 chen2tong4, 痛心 tong4xin1, 悲涼 bei1liang2, 悲凄 bei1qi1, 傷感 shang1gan3, 悲切 bei1qie4, 哀戚 ai1qi1, 悲 傷 bei1shang1, 心酸 xin1suan1, 悲愴 bei1chuang4, 無奈 wu2nai4, 蒼涼 cang1liang2, 不好過 bu4hao3guo4, 抑鬱 yi4yu4 嚇人 xia4ren2, 畏怯 wei4qie4, 緊張 jin3zhang1, 惶恐 huang2kong3, 慌張 huang1zhang1, 驚駭 jing1hai4, 恐慌 kong3huang1, 慌亂 huang1luan4, 心虛 xin1xu1, 驚慌 (continued)
HAPPINESS
2
愛 ai4
43
LOVE
3
愁 chou2, 悶 men4
31
WORRY
4
悲 bei1
28
SADNESS
5
慌 huang1 FEAR
21
2 The Linguistic Expression of Emotion and Cause …
22 Table 2.1 (continued) No.
6
Category
敬 jing4
No. of words
20
AWE
7
激動 ji1dong4
17
EXCITEMENT
8
羞 xiu1, 疚 jiu4 SHYNESS, GUILT
17
9
煩 fan2
15
ANNOYANCE
10
急 ji2
14
ANXIETY
11
傲 ao4
12
PRIDE
12
吃驚 chi1jing1 SURPRISE
12
13
怒 nu4
12
ANGER
14
失望 shi1wang4
12
List of words jing1huang1, 惶惑 huang2huo4, 驚惶 jing1huang2, 驚懼 jing1ju4, 驚恐 jing1kong3, 恐懼 kong3ju4, 心慌 xin1huang1, 害怕 hai4pa4, 怕 pa4, 畏懼 wei4ju4, 發慌 fa1huang1, 發憷 fa1chu4 推崇 tui1chong2, 尊敬 zun1jing4, 擁護 yong1hu4, 倚重 yi3zhong4, 崇尚 chong2shang4, 尊崇 zun1chong2, 敬仰 jing4yang3, 敬佩 jing4pei4, 尊重 zun1zhong4, 敬慕 jing4mu4, 佩服 pei4fu, 景仰 jing3yang3, 敬重 jing4zhong4, 景慕 jing3mu4, 崇敬 chong2jing4, 瞧得起 qiao2deqi3, 崇 奉 chong2feng4, 欽佩 qin1pei4, 崇拜chong2bai4, 孝敬 xiao4jing4 來勁 lai2jin4, 熾烈 chi4lie4, 熾熱 chi4re4, 衝動 chong1dong4, 狂熱 kuang2re4, 激昂 ji1ang2, 激動 ji1dong4, 高亢 gao1kang4, 亢奮 kang4fen4, 帶勁 dai4jin4, 高漲 gao1zhang3, 高昂 gao1ang2, 投入 tou2ru4, 興奮 xing1fen4, 瘋狂 feng1kuang2, 狂亂 kuang2luan4, 感動 gan3dong4 羞澀 xiu1se4, 羞怯 xiu1qie4, 羞愧 xiu1kui4, 負疚 fu4jiu4, 窘 jiong3, 窘促 jiong3cu4, 不過意 bu2guo4yi4, 慚愧 can2kui4, 不好意思 bu4hao3yi4si, 害羞 hai4xiu1, 害臊 hai4sao4, 困窘 kun4jiong3, 抱歉 bao4qian4, 抱愧 bao4kui4, 對不起 dui4buqi3, 羞愧 xiu1kui4, 對不住 dui4buzhu4 煩躁 fan2zao4, 煩燥 fan2zao4, 煩 fan2, 熬心 ao2xin1, 糟心 zao1xin1, 煩亂 fan2luan4, 煩心 fan2xin1, 煩人 fan2ren2, 煩惱 fan2nao3, 煩雜 fan2za2, 膩煩 ni4fan2, 厭倦 yan4juan4, 厭煩 yan4fan2, 討厭 tao3yan4, 頭疼 tou2teng2 急 ji2, 浮躁 fu2zao4, 焦慮 jiao1lü4, 焦渴 jiao1ke3, 焦急 jiao1ji2, 焦躁 jiao1zao4, 焦炙 jiao1zhi4, 心浮 xin1fu2, 心 焦 xin1jiao1, 揪心 jiu1xin1, 心急 xin1ji2, 心切 xin1qie4, 急 ji2, 不安 bu4an1 自傲 zi4ao4, 驕橫 jiao1heng4, 驕慢 jiao1man4, 驕矜 jiao1jin1, 驕傲 jiao1ao4, 自負 zi4fu4, 自信 zi4xin4, 自豪 zi4hao2, 自滿 zi4man3, 自大 zi4da4, 狂 kuang2, 炫耀 xuan4yao4 詫異 cha4yi4, 吃驚 chi1jing1, 驚疑 jing1yi2, 愕然 e4ran2, 驚訝 jing1ya4, 驚奇 jing1qi2, 駭怪 hai4guai4, 駭異 hai4yi4, 驚詫 jing1cha4, 驚愕 jing1e4, 震驚 zhen4jing1, 奇 怪 qi2guai4 憤怒 fen4nu4, 忿恨 fen4hen4, 激憤 ji1fen4, 生氣 sheng1qi4, 憤懣 fen4men4, 憤慨 fen4kai3, 忿怒 fen4nu4, 悲 憤 bei1fen4, 窩火 wo1huo3, 暴怒 bao4nu4, 不平 bu4ping2, 火 huo3 失望 shi1wang4, 絕對 jue2dui4, 灰心 hui1xin1, 喪氣 sang4qi4, 低落 di1luo4, 心寒 xin1han2, 沮喪 ju3sang4, 消
DISAPPOINTMENT
(continued)
2.1 Expression of Emotion in Chinese
23
Table 2.1 (continued) No.
15
Category
安心 an1xin1
No. of words
4
沉 xiao1chen2, 頹喪 tui2sang4, 頹唐 tui2tang2, 低沉 di1chen2, 不滿 bu4man3 安寧 an1ning2, 閑雅 xian2ya3, 逍遙 xiao1yao2, 閑適 xian2shi4, 怡和 yi2he2, 沉靜 chen2jing4, 放鬆 fang4song1, 安心 an1xin1, 寬心kuan1xin1, 自在 zi4zai4, 放心 fang4xin1 看不慣 kan4buguan4, 痛恨 tong4hen4, 厭惡 yan4wu4, 惱 恨 nao3hen4, 反對 fan3dui4, 搗亂 dao3luan4, 怨恨 yuan4hen4, 憎惡 zeng1wu4, 歧視 qi2shi4, 敵視 di2shi4, 憤 恨 fen4hen4 妒嫉 du4ji2, 妒忌 du4ji4, 嫉妒 ji2du4, 嫉恨 ji2hen4, 眼紅 yan3hong2, 忌恨 ji4hen4, 忌妒 ji4du4 蔑視 mie4shi4, 瞧不起 qiao2buqi3, 怠慢 dai4man4, 輕蔑 qing1mie4, 鄙夷 bi3yi2, 鄙薄 bi3bo2, 鄙視 bi3shi4 背悔 bei4hui3, 後悔 hou4hui3, 懊惱 ao4nao3, 懊悔 ao4hui3, 悔恨 hui3hen4, 懊喪 ao4sang4 委屈 wei3qu1, 冤 yuan1, 冤枉 yuan1wang, 無辜 wu2gu1, 抱委屈 bao4wei3qu1 體諒 ti3liang4, 理解 li3jie3, 了解 liao3jie3, 體貼 ti3tie1
4
信任xin4ren4, 信賴xin4lai4, 相信xiang1xin4, 信服xin4fu2
4
過敏 guo4min3, 懷疑 huai2yi2, 疑心 yi2xin1, 疑惑 yi2huo4
15
纏綿 chan2mian2, 自卑 zi4bei1, 自愛 zi4ai4, 反感 fan3gan3, 感慨 gan3kai3, 動搖 dong4yao2, 消魂 xiao1hun2, 癢癢 yang3yang3, 為難 wei2nan2, 解恨 jie3hen4, 遲疑 chi2yi2, 多情 duo1qing2, 充實 chong1shi2, 寂寞 ji4mo4, 遺憾 yi2han4
11
RELIEF
16
恨 hen4
11
HATRED
17
嫉 ji2
7
JEALOUSY
18
蔑視 mie4shi4
7
CONTEMPT
19
悔 hui3
6
REGRET
20
委屈 wei3qu1
5
GRIEVANCE
21
諒 liang4
List of words
UNDERSTANDING
22
信 xin4 TRUST
23
疑 yi2 SUSPECT
24
其他 qi2ta1 OTHERS
2.1.5
Emotion Classification
There has been some work done on the classification of Chinese emotions. Based on a corpus study using the Academia Sinica Balanced Corpus (Sinica Corpus), which contains five million Chinese words (CKIP 1995), Chang et al. (2000) proposed seven sets of the most frequent emotion verbs, namely HAPPY, DEPRESSED, SAD, REGRETFUL, ANGRY, AFRAID, and WORRIED. They identified a total of 33 verbs in the seven types of emotion verbs, each with a frequency of over 40 in the corpus. The verbs and frequency are listed in Table 2.2. They then proposed a dichotomy of emotion verbs in the seven subtypes of Chinese emotion verbs, shown in Table 2.3. For each group, there are two types of emotion verbs: change-of-state emotion verbs (Group A) and homogeneous state emotion verbs (Group B). Chang et al.
2 The Linguistic Expression of Emotion and Cause …
24
Table 2.2 The verbs of emotion in the Sinica Corpus (Chang et al. 2000) Subtype HAPPY
DEPRESSED
SAD REGRET ANGRY AFRAID WORRIED
Verbs and the frequency in the Sinica Corpus 快樂 kuai4le4 (942), 高興 gao1xing4 (669), 愉快 yu2kuai4 (271), 樂 le4 (264), 喜悅 xi3yue4 (156), 開心 kai1xin1 (152), 歡樂 huan1le4 (141), 歡喜 huan1xi3 (107), 快活 kuai4huo2 (48), 痛快 tong4kuai4 (40) 痛苦 tong4ku3 (443), 痛 tong4 (281), 難過 nan2guo4 (232), 沉重 chen2zhong4 (83), 沮喪 ju3sang4 (62), 痛心 tong4xin1 (40) 傷心 shang1xin1 (134), 悲傷 bei1shang1 (52) 遺憾 yi2han4 (198), 後悔 hou4hui3 (102) 生氣 sheng1qi4 (295), 氣 qi4 (126), 憤怒 fen4nu4 (112), 氣憤 qi4fen4 (49) 怕 pa4 (548), 害怕 hai4pa4 (261), 恐懼 kong3ju4 (149), 畏懼 wei4ju4 (40) 擔心 dan1xin1 (609), 煩惱 fan2nao3 (199), 擔憂 dan1you1 (64), 煩 fan2 (54), 憂心 you1xin1 (46), 苦惱 ku3nao3 (45)
Table 2.3 The dichotomy of emotion verbs (Chang et al. 2000)
Subtype HAPPY
DEPRESSED
Group A
Group B
高興 gao1xing4 (669) 開心 kai1xin1 (152)
快樂 愉快 喜悅 歡樂 歡喜 快活 痛快 痛苦 沉重 (83) 沮喪 悲傷
難過 nan2guo4 (232) 痛心 tong4xin1 (48)
ANGRY
傷心 shang1xin1 (134) 後悔 hou4hui3 (102) 生氣 sheng1qi1 (307)
AFRAID
害怕 hai4pa4 (261)
SAD
REGRET
WORRIED
擔心 dan1xin1 (609) 擔憂 dan1you1 (64) 憂心 you1xin1 (46)
kuai4le4 (942) yu2kuai4 (271) xi3yue4 (156) huan1le4 (141) huan1xi3 (107) kuai4huo2 (48) tong4kuai4 (40) tong4ku3 (443) chen2zhong4 ju3sang4 (62) bei1shang1 (52)
遺憾 yi2han4 (198) 憤怒 fen4nu4 (112) qi4fen4 氣憤 (49) 恐懼 kong3ju4 (149) 畏懼 wei4ju4 (40) 煩惱 fan2nao3 (199) 苦惱 ku3nao3 (45)
(2000) found that the two groups of verbs behaved differently in terms of the following five criteria: i. ii. iii. iv. v.
distribution of grammatical functions cooccurrence restrictions when they function as an adjunct appropriateness in the imperative and evaluative constructions verbal aspect transitivity
2.1 Expression of Emotion in Chinese
25
Table 2.4 Distributional syntactic differences between two types of verbs (Chang et al. 2000) Criteria
Change-of-state verbs
Homogeneous state verbs
i
Mostly used as predicates
ii
Can only modify a very restricted set of nouns or verbs All verbs appear in imperative or evaluative constructions More often associated with inchoative state Take causes or goals as direct objects
Mostly used as nominalization or nominal modifiers Can modify a number of nouns or verbs
iii iv v
Only one verb (i.e. fannao ‘to be worried’) appears in imperative or evaluative constructions More often associated with homogeneous state Do not take causes or goals as direct objects
Emotion
neutral
neutral
happiness
happy relieved
like
anger
sadness
fear
praise trust respect like moved expectation
angry
sad frightened regret nervous yearning disappointed
dislike
surprise
shy hate suspect jealous criticize
surprise
Fig. 2.2 Chinese emotion classification tree (Xu et al. 2008)
The major distributional differences between the two groups of verbs are summarized in Table 2.4. Xu et al. (2008) also presented an emotion classification hierarchy for Chinese. In their classification, there are seven types of emotions where each type is associated with some representative keywords. The classification tree is shown in Fig. 2.2. This kind of classification, however, is unclear and unsatisfactory due to the lack of a clear rationale. The absence of a theoretical backup makes this classification less convincing.
2.2
Chinese Emotion Verb Classes and Taxonomy
As mentioned in Sect. 2.1, the existing Chinese emotion taxonomies are mostly intuition-driven and are not theoretically supported. This section attempts to present a Chinese emotion taxonomy based on cognitive theories which can act as a representative of modern Chinese language.
2 The Linguistic Expression of Emotion and Cause …
26
Although many emotion theories have been proposed in different fields, such as biology, psychology, and linguistics, most of them agree that emotion can be divided into primary emotions and complex emotions. Even though there is still controversy over the selection of primary emotions, most emotion theories consider the four emotions HAPPINESS, SADNESS, ANGER, and FEAR as primary emotions. In a classic cognitive emotion classification model, Plutchik (1980) followed the division of primary emotions and complex emotions, each to varying degrees. He also suggested a list of English emotion keywords. Extending Plutchik’s work, Turner (2000) allowed more flexible combinations of primary emotions to form complex emotions. In building a Chinese emotion taxonomy, I adapted Turner’s taxonomy emphasizing three main points: (1) Each primary emotion is divided into three levels according to its intensity: high, moderate, and low. In addition to HAPPINESS, SADNESS, ANGER, and FEAR, Turner suggests that DISGUST and SURPRISE can be primary emotions too (Turner 1996; Turner 2007). In Chinese, the character 驚 jing1 ‘surprise’ is rather productive in forming emotion words, such as 驚喜 jing1xi3 ‘surprise and happiness’, and 驚嚇 jing1xia4 ‘surprise and fear’, which is consistent with the description of the surprise emotion by Plutchik (1991): “when the stimulus has been evaluated, the surprise may quickly change to any other emotion.” Therefore, our Chinese emotion taxonomy considers HAPPINESS, SADNESS, ANGER, FEAR, and SURPRISE as primary emotions. (2) Complex emotions can be divided into first-order complex emotions (the combinations of two primary emotions), second-order complex emotions (the combinations of three primary emotions), and so on, according to the number of primary emotions involved in the complex emotion. (3) Compared to other emotion taxonomies, Turner’s classification has a more flexible structure, and more extensions can be done for different applications. For example, in a complex emotion, not only are its primary emotions listed, but also their intensity. For instance, three emotion types which belong to “ANGER + FEAR” are specified as follows: JEALOUSY: ANGER
(Moderate) + FEAR (Moderate) (Low) + FEAR (Low) ABHORRENCE: ANGER (High) + FEAR (Low) SUSPICION: ANGER
Since there are considerable differences between English and Chinese, some necessary changes regarding the classes of complex emotions have to be made. Moreover, Turner’s emotion taxonomy excludes a number of frequently used emotion keywords that appear in Plutchik’s list (1980). Hence, the emotion keywords in Plutchik’s classification model are added to the corresponding classes and levels in Turner’s taxonomy. Table 2.5 shows the revised Turner-Plutchik emotion taxonomy, listing the English primary emotion keywords as well as the complex emotions. Several emotion keywords which express similar emotion meaning are grouped into an emotion type. For example, the emotion keywords awe, reverence, veneration are
cheerful, satisfy, pleased, enjoy, interests
gloomy, dismay, sad, unhappy, disappoint
misgivings, self-conscious, scare, panic, anxious
ecstatic, eager, joy, enthusiastic, happy
deject, despondent, sorrow, anguish, despair
horror, terror
HAPPINESS
SADNESS
FEAR
Moderate
Variation in intensity High
Primary emotions
Table 2.5 Revised Turner-Plutchik emotion taxonomy
bewilder, reluct, shy, puzzles, confuse
dispirit, downcast, discourage
sanguine, serene, content, grateful
Low +FEAR Wonder: wonder, wondering, hopeful Pride: pride, boastful +ANGER Vengeance: vengeance, vengeful Calm: appeased, calmed, calm, soothed Bemused: bemused +SADNESS Yearning: nostalgia, yearning +HAPPINESS Acceptance: acceptance, tolerant Solace: moroseness, solace, melancholy +FEAR Hopeless: forlorn, lonely, hopeless, miserable Remorseful: remorseful, ashamed, humiliated +ANGER Discontent: aggrieved, discontent, dissatisfied, unfulfilled Boredom: boredom Grief: grief, sullenness +HAPPINESS Awe: awe, reverence, veneration +ANGER Antagonism: antagonism, revulse Envy: envy +SADNESS
First-order emotions
(continued)
+FEAR, ANGER Guilt: guilt
Second-order emotions
2.2 Chinese Emotion Verb Classes and Taxonomy 27
Moderate
contentious, offend, frustrate, hostile, angry
startled, amaze, surprise
Variation in intensity High
dislike, disgust, outrage, furious, hate
astonish
Primary emotions
ANGER
SURPRISE
Table 2.5 (continued)
contemptuous, agitate, irritate, annoy, impatient
Low Worry: dread, wariness, pensive, helpless, apprehension, worried +HAPPINESS Unfriendly: snubbing, mollified, rudeness, placated, apathetic, unsympathetic, unfriendly, unaffectionate Sarcastic: sarcastic +FEAR Jealousy: jealous Suspicion: suspicion, distrustful Abhorrence: abhorrence +SADNESS Depression: bitter, depression Intolerant: intolerant +HAPPINESS Delight: delight +SADNESS Embarrassed: embarrassed
First-order emotions
Second-order emotions
28 2 The Linguistic Expression of Emotion and Cause …
2.2 Chinese Emotion Verb Classes and Taxonomy
29
grouped into emotion type AWE. For a complex emotion, the order of primary emotions indicates their importance in the complex emotion. For example, PRIDE is ‘HAPPINESS + FEAR’, which contains a greater amount of HAPPINESS with a lesser amount of FEAR; whereas AWE is ‘FEAR + HAPPINESS’ containing greater amount of FEAR and lesser amount of HAPPINESS. With the revised Turner-Plutchik emotion taxonomy, I then selected some Chinese emotion keywords from the cognitive-based feeling words listed in Xu and Tao (2003), and map those emotion keywords to the revised taxonomy with certain modifications. Before the mapping takes place, I reduced the emotion keywords from 390 words to 201 words as I do not consider keywords describing personality, such as 多情 duo1qing2 ‘sentimental’, 自大 zi4da4 ‘arrogant’, and attitudes, such as 留神 liu2shen2 ‘to be careful’, and 偏愛 pian1ai4 ‘favour’ as a type of emotion. The proposed Chinese emotion taxonomy is shown in Table 2.6. The taxonomy includes only relevant emotions and eliminates words of feelings or attitudes that most existing emotion keyword lists contain. The analyses in this work are made based on this Chinese emotion taxonomy. For later analyses, ambiguous keywords are removed for the purpose of computational implementation.
2.3
Cause and Emotion
This section deals with three major questions: (1) What kinds of constructions are considered emotional? (2) What are causal relations? (3) How are causal relations expressed in Chinese?
2.3.1
Emotion Constructions
The present study deals with constructions describing explicit emotions, i.e. descriptive emotions (Kövecses 2000; Potts 2007) as mentioned in Chap. 1. An explicit emotion refers to the direct description of an emotion by means of keywords. For instance, ‘The book surprised me!’ is an example of explicit emotion as the emotion SURPRISE is overtly conveyed. On the other hand, a construction like ‘I was speechless after reading the book’, contains emotional information but is ambiguous. The speaker could be expressing one of several possible emotions, for example being speechless with ANGER, SURPRISE, SADNESS, etc. In another example, such as ‘he got first runner-up this year’, an intuitive interpretation is that ‘he is happy about getting first runner-up’, but it is not necessarily true. He may as well be sad about failing to win the championship, or be surprised or even neutral about getting the prize. Unless it is explicitly expressed in the context, constructions such as “he got first runner-up this year” that require inference or connotation are considered to be neutral sentences which will not be analyzed in the present study.
哀 ai1
SADNESS
喜 xi3
HAPPINESS/
Primary emotions 閒適 xian2shi4, 怡和 yi2he2, 放鬆 fang4song1, 自在 zi4zai4
灰心 hui1xin1, 喪氣 sang4qi4
感傷 gan3shang1, 傷心 shang1xin1, 傷感 shang1gan3, 心酸 xin1suan1, 沉悶 chen2men4, 憋氣 bie1qi4, 鬱 悒 yu4yi4, 苦悶 ku3men4, 無聊 wu2liao2, 鬱悶 yu4men4, 乏味 fa2wei4, 沉鬱 chen2yu4, 憋悶 bie1men4, 憂鬱 you1yu4, 陰鬱 yin1yu4, 悵惘 chang4wang3, 沮喪 ju3sang4, 消沉 xiao1chen2, 頹喪 tui2sang4, 頹唐 tui2tang2, 煩悶 fan2men4
悲慟 bei1tong4, 悲痛 bei1tong4, 哀 傷 ai1shang1, 悲哀 bei1ai1, 哀痛 ai1tong4, 沉痛 chen2tong4, 痛心 tong4xin1, 悲涼 bei1liang2, 悲淒 bei1qi1, 悲切 bei1qie4, 悲傷 bei1shang1, 悲愴 bei1chuang4, 哀 戚 ai1qi1, 絕望 jue2wang4
Low
欣慰 xin1wei4, 高興 gao1xing4, 愉 悅 yu2yue4, 欣喜 xin1xi3, 歡欣 huan1xin1, 樂 le4, 歡暢 huan1chang4, 開心 kai1xin1, 康樂 kang1le4, 歡快 huan1kuai4, 快慰 kuai4wei4, 歡 huan1, 舒暢 shu1chang4, 快樂 kuai4le4, 快活 kuai4huo2, 歡樂 huan1le4, 暢快 chang4kuai4, 舒心 shu1xin1, 舒坦 shu1tan, 歡娛 huan1yu2, 如意 ru2yi4, 喜悅 xi3yue4, 順心 shun4xin1, 歡悅 huan1yue4, 爽心 shuang3xin1, 曉暢 xiao3chang4, 鬆 快 song1kuai4, 歡愉 huan1yu2, 歡 喜 huan1xi3
Moderate
痛快 tong4kuai4, 振奮 zhen4fen4, 亢奮 kang4fen4, 興奮 xing1fen4
Variations in intensity High
Table 2.6 Chinese emotion taxonomy
+FEAR: Pride/傲 ao4: 自傲 zi4ao4, 驕橫 jiao1heng4, 驕慢 jiao1man4, 驕矜 jiao1jin1, 驕傲 jiao1ao4, 自負 zi4fu4, 自信 zi4xin4, 自豪 zi4hao2, 自滿 zi4man3, 自大 zi4da4, 自狂 zi4kuang2, 狂 kuang2, 炫耀 xuan4yao4, 得意 de2yi4 Relief/安心 an1xin1: 安心 an1xin1, 寬心 kuan1xin1, 放心 fang4xin1 +ANGER: Appeased/解恨 jie3hen4: 解恨 jie3hen4 +SADNESS: Moved/感動gan3dong4: 感動 gan3dong4 +FEAR: Misery/悲愁 bei1chou2: 悲愁 bei1chou2, 哀愁 ai1chou2, 愁悶 chou2men4, 惆悵 chou2chang4 Remorseful/後悔 hui4hui3: 後悔 hou4hui3, 慚愧 can2kui4, 抱歉 bao4qian4, 抱愧 bao4kui4, 對不起 dui4buqi3, 羞愧 xiu1kui4, 背悔 bei4hui3, 懊惱 ao4nao3, 懊悔 ao4hui3, 悔恨 hui3hen4, 懊喪 ao4sang4 +ANGER: Aggrieved/委屈 wei3qu1: 委屈 wei3qu1, 冤 yuan1, 冤枉 yuan1wang, 抱委屈 bao4wei3qu1, 哀怨 ai1yuan4 Dissatisfied/不滿: 不滿, 不快, 不悅
First-order emotions
(continued)
+FEAR, ANGER Guilt/疚 jiu4: 疚 jiu4, 內疚 nei4jiu4, 負 疚 fu4jiu4
Second-order emotions
30 2 The Linguistic Expression of Emotion and Cause …
nu4
ANGER/怒
kong3
FEAR/恐
Primary emotions
羞澀 xiu1se4, 羞怯 xiu1qie4,羞慚 xiu1can2, 害羞 hai4xiu1, 害臊 hai4sao4, 遲疑 chi2yi2, 為難 wei2nan2
煩 fan2, 煩躁 fan2zao4, 煩亂 fan2luan4, 煩心 fan2xin1,煩人 fan2ren1, 煩惱 fan2nao3, 煩雜 fan2za2, 浮躁 fu2zao4
生氣 sheng1qi4, 窩火 wo1huo3, 火 huo3, 厭倦 yan4juan4, 討厭 tao3yan4, 厭惡 yan4wu4, 反感 fan3gan3, 敵視 di2shi4, 衝動 chong1dong4
憤怒 fen4nu4, 忿恨 fen4hen4, 激憤 ji1fen4, 憤懣 fen4men4, 憤慨 fen4kai3, 忿怒 fen4nu4, 悲憤 bei1fen4, 暴怒 bao4nu4, 蔑視 mie4shi4, 瞧不起 qiao2buqi3, 輕蔑
Low
畏怯 wei4qie4, 心虛 xin1xu1, 心慌 xin1huang1, 害怕 hai4pa4, 怕 pa4, 畏懼 wei4ju4, 發慌 fa1huang1, 發 憷 fa1chu4
Moderate
惶恐 huang2kong3, 恐慌 kong3huang1, 恐懼 kong3ju4
Variations in intensity High
Table 2.6 (continued)
+SURPRISE Disappointment/失望 shi1wang4: 失 望 shi1wang4, 心寒 xin1han2 Embarrassed/窘 jiong3: 窘 jiong3 Panic/驚恐 jing1kong3: 驚恐 jing1kong3, 驚駭 jing1hai4, 驚惶 jing1huang2, 驚懼 jing1ju4, 嚇人 xia4ren2, 慌張 huang1zhang1, 驚慌 jing1huang1, 惶惑 huang2huo4, 慌 亂 huangluan4 +ANGER: Envy/嫉 ji2: 嫉妒 ji2du4, 妒嫉 du4ji2, 妒忌 du4ji4, 忌妒 ji4du4, 嫉 恨 ji2hen4, 眼紅 yan3hong2, 忌恨 ji4hen4 +SADNESS: Anxious/急ji2: 焦慮 jiao1lü4, 焦渴 jiao1ke3, 焦急 jiao1ji2, 焦躁 jiao1zao4, 焦炙 jiao1zhi4, 心浮 xin1fu2, 心焦 xin1jiao1, 揪心 jiu1xin1, 心急 xin1ji2, 心切 xin1qie4, 急 ji2 Worry/愁 chou2: 愁 chou2, 苦惱 ku3nao3, 愁苦 chou2ku3, 憂愁 you1chou2, 發愁 fa1chou2, 擔憂 dan1you1, 擔心 dan1xin1, 犯愁 fan1chou2, 憂慮 you1lü4, 緊張 jin3zhang1, 困惑 kun4huo4 +HAPPINESS Rudeness/瘋狂 feng1kuang2: 瘋狂 feng1kuang2 +FEAR
First-order emotions
(continued)
Second-order emotions
2.3 Cause and Emotion 31
SURPRISE/ 驚 jing1
Primary emotions
qing1mie4, 鄙夷 bi3yi2, 鄙薄 bi3bo2, 鄙視 bi3shi4, 歧視 qi2shi4, 自卑 zi4bei1, 痛恨 tong4hen4, 怨恨 yuan4hen4, 憎惡 zeng1wu4, 憤恨 fen4hen4, 厭煩 yan4fan2, 膩煩 ni4fan2, 惱恨 nao3hen4 駭怪 hai4guai4, 駭異 hai4yi4, 震驚 zhen4jing1
Variations in intensity High
Table 2.6 (continued)
詫異 cha4yi4, 吃驚 chi1jing1, 愕然 e4ren2, 驚訝 jing1ya4, 驚奇 jing1qi2, 驚詫 jing1cha4, 驚愕 jing1e4
Moderate
奇怪 qi2guai4
Low Suspicion/疑 yi2: 疑 yi2, 懷疑 huai2yi2, 疑心 yi2xin1, 疑惑 yi2huo4 +SADNESS Bitterness/辛酸xin1suan1: 辛酸 xin1suan1, 酸辛suan1xin1 +HAPPINESS Delighted/驚喜 jing1xi3: 驚喜 jing1xi3
First-order emotions
Second-order emotions
32 2 The Linguistic Expression of Emotion and Cause …
2.3 Cause and Emotion
33
While explicit emotions are usually expressed by certain emotion keywords, the presence of emotion keywords does not necessarily mean that the construction is emotional. Negation, for example, is a common phenomenon which is able to neutralize or reverse the meaning of the emotion keyword, such as ‘he is not sad’, which implies no specific emotion, and ‘he is not happy’ meaning ANGER or SADNESS. Hence, the former sentence is considered neutral and the latter emotional yet ambiguous. For the linguistic analyses, I collect emotional data by filtering out a number of constructions which are not regarded as emotional, including negatives, conditionals, general statements, and hypothetical remarks. They are illustrated in (a)–(d) below: a. negative sentences keywords: 不 bu4, 不是 bu2shi4, 沒有 mei2you3, 非 fei1, 不會 bu2hui4, 不大 bu2da4, 不再 bu2zai4, 不見得 bu2jian4de2, 說不上 shuo1bu2shang4, 談不上 tan2bu2shang4 e.g.: ta1 bu2 hai4pa4 3.SG.M no becoming frightened ‘He is not frightened.’
b. conditional sentences keywords: 如果 ru2guo3 ‘if’, 假如 jia3ru2 ‘if’, 假使 jia3shi3 ‘suppose’, 假設 jia3she4 ‘suppose’, 假想 jia3xiang3 ‘suppose’, 要是 yao4shi4 ‘if’, 萬一 wan4yi1 ‘in case’, 的話 dehua4 ‘if’ e.g.: ru2guo3 ni3 lai2 dehua4, ta1 yi2ding4 if 2.SG come if, 3.SG surely ‘If you come, s/he will surely be very happy.’
hen3 gao1xing4 very becoming happy
c. general statements
e.g.: yi2 ge4 kuai4le4 de ren2 yi1ding4 hui4 de2dao4 xing4fu2 one CL to be happy POSS person surely will receive bliss ‘A happy person will surely be blessed.’
d. hypothetical remarks keywords: 認為 ren4wei2 ‘think’, 感覺 gan3jue2 ‘feel’, 想 xiang3 ‘think’, 猜 cai1 ‘guess’,
2 The Linguistic Expression of Emotion and Cause …
34 e.g.
wo3 xiang3 ta1 ying3gai1 1.SG think 3.SG.M should ‘I think he should be very sad.’
2.3.2
hen3 nan2guo4 very becoming sad
Causal Relations
A causal relation refers to the relation in which an event or state brings about another event or state. The two events or states are generally called the causing event and the caused event (Talmy 2000). The causing event leads to the caused event and the caused event is the result of the causing event. Talmy (2000: 481) argued that “the cause of the simple event is itself also a simple event rather than, for instance, a (physical) object”. This is justified by the following sentences: (1) *The window’s breaking resulted from a ball. (2) The window’s breaking resulted from a ball’s sailing into it. The nominal causing event in (1) makes the sentence unacceptable; while a simple event in (2) works just fine. There are nominals, however, that occur after from or as a result of as in (3)–(5) (Talmy 2000): (3) The window cracked from the wind. (4) The window cracked from the rain. (5) The window cracked from the fire. According to Fillmore (1976), these constructions are derived from the conflation of a deeper clause that actually denotes an event as indicated in (6)-(8): (6) The window cracked from the air blowing on it. (7) The window cracked from the rain falling on it. (8) The window cracked from flames acting on it. In the context of emotion constructions, the causal relation is established by the link between a cause and the emotional state. Following Talmy (2000), the cause of an emotion should be an event itself. In this work, I will often use the term cause or event to refer to a linguistic expression that I am analyzing as denoting a cause or event. In addition, cause event refers to a linguistic expression that is analyzed as denoting the cause event of an emotion. By cause event, it does not necessarily mean the actual cause of the emotion or what leads to the emotion. Rather, it refers to the immediate cause of the emotion which can be the actual trigger event or the perception of the trigger event. An actual trigger event is the event which triggers the presence of an emotion, whereas the trigger event that requires a certain degree of perception is actually a potential event which may or may not happen. For example, the emotion FEAR is more likely to be linked to the perception of a trigger event. For instance, a snake is not the direct cause of FEAR in a person, but FEAR arises through the cognitive awareness of the possible danger that the snake may
2.3 Cause and Emotion
35
bring to the person. Therefore, any event that is highly associated with the presence of an emotion is considered a cause event.
2.4
Emotion Cause Events in Chinese
Adapting the TimeML annotation scheme (Saurí et al. 2004), events are situations that happen or occur. They can be punctual or last for a period of time, and they can also be denoted by predicates describing states or circumstances in which something obtains or holds true. In this study, cause events specifically refer to the denotations of explicitly expressed arguments or to events that are highly linked with the presence of the corresponding emotions. They are usually linguistically expressed by means of verbs, nominalizations, and nominals. In Chinese, there are mainly two different ways of denoting cause events: verbal events and nominal events. A verbal event refers to a linguistic expression denoting an event that involves a verb or nominalization, whereas a nominal event is simply a noun. Some examples of cause events are given in bold face in (9)–(14). (9) zhe4 tou2 niu2 de zhu3ren2, yan3 kan4 zi4ji3 de niu2 re3chu1 huo4 DET CL ox POSS owner, eye see oneself POSS ox cause trouble lai2 le, fei1chang2 hai4pa4, jiu4 ba3 zhe4 tou2 niu2 come ASP, very becoming frightened, then BA DET CL ox di1jia4 mai4chu1 low price sell ‘The owner was frightened to see that his ox had caused trouble, so he sold it at a low price.’
(10) mei2 xiang3dao4 ta1 shuo1 de dou1 shi4 zhen1hua4, rang4 not think 3.SG.F say POSS all is true word, cause zhen4jing1 bu4yi3 to be shocked very ‘Unexpectedly, what she said was the truth, which surprised him greatly.’
ta1 3.SG.M
(11) ta1 dui4 zhe4 ge4 chong1man3 nong2hou4 3.SG.M for DET CL full of dense gao1xing4 de shou3wu3zu2dao3 becoming happy DE flourish ‘He was very happy about this lovely idea.’
ai4yi4 love
de xiang3fa3 POSS idea
36
2 The Linguistic Expression of Emotion and Cause …
(12) zhe4ci4 yan3chu1 de jing1zhi4 dao4 shi4 ling4 wo3 shi2fen1 jing1ya4 this time perform POSS exquisite yet is make 1.SG very becoming surprised ‘I was very surprised by this exquisite performance.’
(13) Ni2ao4 de hua4 hen3 ling4 kai3luo4lin2 shang1xin1 Leo POSS word very make Caroline becoming sad ‘Caroline was saddened by Leo’s words.’
(14) dui4yu2 wei4lai2, lao3shi2shuo1 wo3 for future, frankly 1.SG ‘Frankly, I am very scared about the future.’
hen3 very
hai4pa4 becoming scared
The causes in (9) and (10) are events which indicate the actual events involved in causing the emotions. The ones in (11) and (12) are nominalized causes, whereas (13) and (14) involve nominal causes. Since nominalized and nominal causes are both noun phrases (NPs), they should be reinterpreted as events, the idea of which will be further discussed in Chap. 4.
Chapter 3
Linguistic Resources for Study of Emotion
3.1
Empirical Approaches Towards Emotion Analysis
The present work constitutes a corpus-based study on emotion analysis. It provides a systematic analysis of authentic examples of Chinese from a variety of genres. Instead of resorting to introspective examples, corpus data represent naturally occurring data. By observing the text in the corpus with the help of data collection tools, relevant generalizations can be made. This chapter discusses the research methodology in terms of the various existing tools and resources adopted in the current work as well as the development of the Chinese emotion corpus serving as the main data for the current work.
3.1.1
Corpora
The data used in this study are taken from two written text corpora: Academia Sinica Balanced Corpus of Modern Chinese (Sinica Corpus)1 and Chinese Gigaword Corpus.2 The Sinica Corpus is a Mandarin corpus containing a total of ten million words. The texts in the corpus are collected from different sources and are related to various topics including philosophy, science, arts, etc. Each text is segmented and tagged with its part-of-speech. The Chinese Gigaword Corpus is a news corpus which contains a total of 1.4 billion characters from Taiwan’s Central News Agency, China’s Xinhua News agency, and Singapore Zaobao. The data is segmented and fully tagged (Ma and Huang 2006). This is so far the largest Chinese corpus which provides a good source of data for quantitative research. However, as it heavily relies on news data, it may not accurately reflect the actual spoken language. The Sinica Corpus pro1
http://www.sinica.edu.tw/SinicaCorpus/. http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC2005T14.
2
© Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_3
37
38
3 Linguistic Resources for Study of Emotion
vides a more balanced source of data which can better reflect how Mandarin Chinese is actually used. The current work makes use of both corpora: the Sinica Corpus for qualitative analysis, and the Chinese Gigaword Corpus for quantitative analysis.
3.1.2
Chinese Word Sketch
The Chinese Word Sketch,3 serving as a data collection tool, is a web-based program which takes the corpus of any language as its input provided that the corpus is processed with an appropriate level of linguistic mark-up. The two corpora used by CWS are Chinese Gigaword Corpus second edition and the Sinica Corpus fifth edition. CWS generates linguistically meaningful collocations automatically from the two corpora and provides rich lexicon-based grammatical information accompanied by statistical information (Huang 2006). It extracts the most salient relations, rather than those extracted by most other search engines. The grammar rules of the Chinese Word Sketch are based on Information-based Case Grammar (ICG, Chen and Huang 1990). ICG entails that the lexical instance for each word contains both semantic and syntactic feature structures. They encode syntactic and semantic constraints on grammatical phrasal patterns in terms of thematic structures and encode the precedence relations in terms of adjunct structures. The Chinese Word Sketch has a number of language-analysis functions, the core ones being the concordance and the word sketch programs. These functions discover grammatical and collocational behaviour from the corpora. In addition, the present work also makes use of other functions. One of them, the Word Sketch Difference, compares two similar words by showing patterns and combinations that they have in common, as well as those that are more typical of, or unique to, one word rather than the other. Another function, the Thesaurus, dentifies which words occur with the same collocates as other words, and thus generates a distributional thesaurus. In addition, the Chinese Word Sketch allows us to create a random sample of the concordances.
3.1.3
Data
This study focuses on the five primary emotions in the Chinese emotion taxonomy, i.e. HAPPINESS, SADNESS, FEAR, ANGER, and SURPRISE. In each class of emotion, it follows Chang et al.’s (2000) dichotomical classification of emotion verbs, i.e. change-of-state and homogeneous state emotion verbs. To give a detailed analysis
3
The Word Sketch Engine was developed by the team led by Adam Kilgarriff, http://www. sketchengine.co.uk/. The Chinese version of the word sketch, Chinese Word Sketch, was done by Chu-Ren Huang’s research team, http://wordsketch.ling.sinica.edu.tw/.
3.1 Empirical Approaches Towards Emotion Analysis
39
of the two types of primary emotions, it specifically examines the most frequent verbs in the five classes of emotions, one from each type of emotion verb. Unlike Chang et al.’s proposal for seven emotions, this work considers DEPRESSION as a complex emotion in the subset of SADNESS, WORRY and REGRET, and SURPRISE as an additional primary emotion. Since SURPRISE was not examined in Chang et al.’s analysis, the emotion verbs of SURPRISE in Chinese was tested according to the five criteria based on the Sinica Corpus. As discussed in Sect. 2.1.5, distinction between change-of-state and homogeneous state emotion verbs in Chinese is made according to five criteria: (1) their grammatical functions; (2) their cooccurrence restrictions; (3) the appropriateness in the imperative and evaluative constructions; (4) their verbal aspect; and (5) their transitivity. It is found that the two most frequent verbs of SURPRISE are 驚訝 jing1ya4 and 震驚 zhen4jing1, and that 驚訝 jing1ya4 behaves more like a change-of-state emotion verb whereas 震驚 zhen4jing1 is more of a homogeneous state emotion verb. The ten emotion keywords discussed in this study are 高興 gao1xing4, 傷心 shang1xin1, 生氣 sheng1qi4, 害怕 hai4pa4, 驚訝 jing1ya4 (change-of-state emotion verbs), and 快樂 kuai4le4, 悲傷 bei1shang1, 憤怒 fen4nu4, 恐懼 kong3ju4, 煩惱 fan2nao4 (homogeneous state emotion verbs).4 They are shown in Table 3.1 with example sentences and the corresponding emotion causes.
3.2
Emotion Annotated Corpora
Based on the assumptions of NSM and the Chinese emotion taxonomy, my colleagues and I have designed an emotion annotation scheme and developed an annotated emotion corpus for emotion processing (Chen et al. 2009a, b, 2010; Lee et al. 2009, 2010a, b). The present study also made use of this emotion corpus for the emotion analyses discussed in the subsequent chapters. In this section, I briefly discuss the construction of the emotion corpus and the major objectives of the relevant projects. In creating an emotion corpus, it seems to be intuitively justified to extract sentences that contain an emotion keyword, such as the keyword joyful indicating the presence of the emotion HAPPINESS. Yet, as can be seen above, an emotion can be expressed with or without an emotion keyword in the text. There are also other possible challenges, such as ambiguities of some emotion keywords and context shift (Polanyi and Zaenen 2004). To deal with these problems, we have built the emotion corpus based on the Natural Semantic Metalanguage (NSM) theory. Adopting the spirit and the motivations of NSM, there are three assumptions in constructing the corpus: (1) emotions can be decomposed into semantic primitives;
4
The tones of these representative emotion keywords may be omitted in the later references for simplicity.
40
3 Linguistic Resources for Study of Emotion
Table 3.1 Primary emotions and the corresponding causes
(2) the cause event is essential to emotion classification; and (3) linguistic cues can be derived from different emotions. According to the NSM theory, an emotion is provoked by a cause. This indicates that one possible way to detect emotions in text is by detecting emotion causes which are often provided in the context. Since
3.2 Emotion Annotated Corpora
41
emotion is rather subjective, the stimulus-based approach works only when the cause is explicitly provided in the context. For example, the cause ‘the bank is going to lay off 100 workers in Asia’ may result in different emotions, such as FEAR, WORRY, or ANGER, depending on the context. It is also noted that the text containing an emotion keyword may also contain emotional stimulus and its context. Thus, the emotion corpus contains a collection of causes of explicitly described emotions.
3.2.1
Emotion Annotation Scheme
The emotion annotation scheme is designed to encode emotion information based on the previously discussed Chinese emotion taxonomy. Our emotion annotation scheme is an XML scheme and conforms to the Text Encoding Initiative (TEI) scheme with several modifications. This annotation scheme encodes emotion information for a sentence and can be compatible with any TEI-based annotated corpora as long as sentences are clearly marked. The emotion-related elements (tags) in our annotation scheme are described as follows. Figure 3.1 gives the definition of each element:
element emotion { (emotionType)+, } element emotionType { attribute name (optional), attribute keyword (optional), (primaryEmotion)+ } element primaryEmotion { attribute order (optional), attribute name (necessary), attribute intensity (optional) } element neutral { } Fig. 3.1 The definition of emotion-related elements
42
3 Linguistic Resources for Study of Emotion
Note that is a tag for a sentence-like division of a text and its attribute n gives the sentence index. The element is used only when the sentence expresses emotions. It contains a list of elements and a element. As a sentence may express several emotions, an element can contain several instances of , and each describes an emotion occurring in that sentence separately. The element is used only when the sentence does not contain any emotion expression. It contains only an element. The element describes a type of emotion in that sentence. It contains an ordered sequence of elements. The attribute name provides the name of the emotion type, such as SURPRISE, REMORSEFUL, and so on, and is optional. If an emotion type is a primary emotion, e.g., HAPPINESS, the will have only one element, e.g., , which encodes the information of this primary emotion. If the emotion is a complex emotion, e.g., REMORSEFUL, the element will have several elements (each of them describing the primary emotion involved in that complex emotion), e.g., . The attribute keyword is an optional designation if annotators want to provide the indicator of a text for that emotion. The element describes the property of a primary emotion involved in the emotion type. There are three attributes: ORDER, NAME, and INTENSITY. ORDER gives the weight of this primary emotion in the emotion type, and the weight value decreases with the ascending order value. NAME and INTENSITY provide the name and intensity of a primary emotion. To encode the information in our emotion taxonomy, the value of ORDER is {1, 2, 3, 4, 5}, the value of NAME is {HAPPINESS, SADNESS, ANGER, FEAR, SURPRISE}, and the value of INTENSITY is {high, moderate, low}. This annotation scheme has the versatility to provide emotion data for a wide variety of applications.
3.2.2
Emotion Corpus Construction
A pattern-based approach is used to compile the emotion corpus, which is similar to the one used in Tokuhisa et al. (2008); however, we do not limit our work to event-driven emotions (Kozareva et al. 2007). There are five steps regarding the construction of the corpus: (i) Extract emotion sentences: sentences containing emotion keywords and their context are extracted by keyword matching. (ii) Delete ambiguous structures: sentences which contain structures, such as negation and modals are filtered out automatically. (iii) Delete ambiguous emotion keywords: all sentences containing ambiguous emotion keywords are filtered out.
3.2 Emotion Annotated Corpora
43
(iv) Provide emotion tags: each remaining sentence is marked with its emotion tag according to the emotion type which the focus emotion word belongs to (with reference to our Chinese emotion taxonomy). (v) Ignore the focus emotion keywords: for emotion computing, the emotion word is removed from each sentence. Step (i) extracts emotional sentences in a way such that each extraction includes the target sentence that contains the emotion keywords, the preceding sentence, and the subsequent sentence. Step (ii) removes two kinds of ambiguous contextual structures: negation and modal. In Chinese, a negated emotion expression can be interpreted as one of three possible meanings, as exemplified in (1)–(3).
With the negation words 不 bu4 ‘not’ and 沒有 mei2you3 ‘no’ (1) expresses the kind of emotion that is the opposite of the target emotion HAPPINESS, and thus is marked as another emotion; (2) denies the existence of HAPPINESS, and is marked as a neutral sentence; (3) confirms the existence of HAPPINESS and is marked as a sentence conveying HAPPINESS. The modal structure often indicates that the emotion expression is based on counter-factual assumption; hence, the emotion does not exist at all, as in (4) and (5).
These sentences often do not express any emotion. Therefore, to ensure the quality of the emotion corpus, all sentences containing a negation or a modal are removed. Their removal is based on their being detected by certain rules, plus a list of keywords (negation polarity words for the negation structure and modal words for the modal structure). Step (iii) deletes some highly ambiguous emotion keywords. Five sentences of each emotion keyword are randomly selected and annotated by two annotators. The
44 Table 3.2 Summary of emotion corpus
3 Linguistic Resources for Study of Emotion Emotions
No. of instances Gigaword Sinica
Total
HAPPINESS
35,556 10,334 19,122 6218 11,115 174,698 257,043
38,100 11,235 20,386 7656 12,456 183,338 273,171
SADNESS FEAR ANGER SURPRISE COMPLEX
Total
2544 901 1264 1438 1341 8640 16,128
emotion keyword is removed from our emotion taxonomy when the accuracy of the five sentences is lower than 40%, such as 緊張 jin3zhang1 ‘to be nervous’. Finally, 226 emotion keywords remain (141 primary emotions plus 84 complex emotions). Overall, 20.2% of Sinica sentences and 22.4% of Gigaword sentences are removed in step (ii) and (iii). A summary of the emotion corpus data is given in Table 3.2. Tokuhisa et al. (2007) found that a major challenge for emotion computing, especially for emotion detection, is to collect neutral sentences. Since neutral sentences are unmarked and hard to detect, we develop a naïve yet effective algorithm to create a neutral corpus. A sentence is considered as neutral only when the sentence itself and its context (i.e. the previous sentence and the following sentence) do not contain any of the given emotion keywords. The study runs the emotion sentence extraction and neutral sentence extraction on two corpora: the Sinica Corpus and the Chinese Gigaword Corpus. We then create two emotion corpora and two neutral corpora separately. To estimate the accuracy of our emotion and neutral sentence extraction, about 1000 sentences from each corpus are extracted and are checked by two annotators. Table 3.3 lists the accuracy statistics of the emotion and neutral sentences. The high accuracy of the neutral corpus proves that our approach is effective in extracting neutral sentences from the document-based corpus which contains contextual information. Although the accuracy of emotion corpus is lower, it is much higher than the one reported by Kozareva et al. (2007), i.e. 49.4. The accuracy is significantly increased by deleting ambiguous emotion keywords in Step (iii). For the 2474 randomly selected Chinese sentences, the overall accuracy of the remaining 1751 sentences is increased by about 14% after Step (iii). Thus, a tradeoff between the coverage and the accuracy of the emotion corpus is made when deciding whether or not to remove the ambiguous emotion keywords as in Step (iii). We also explore emotions through sentences that do not contain any given emotion keyword (no-emotion-keyword sentences), because our approach extracts only part of the neutral sentences and part of the emotion sentences in reality. For Table 3.3 The accuracy of the emotion-driven corpora
Gigaword Sinica
Emotion corpus
Neutral corpus
82.17 77.56
98.61 98.39
3.2 Emotion Annotated Corpora
45
each corpus, about 1000 no-emotion-keyword sentences are randomly selected and checked by two annotators. Results show that only about 1% of those sentences express or attribute emotions. This indicates that for real emotion computing, which mainly works on formal written text, it is important to deal with the emotion expressions which contain emotion keywords that are ambiguous, such as the sentences deleted in Steps (ii) and (iii). More exploration is needed for emotion and neutral sentence distribution on other kinds of written text, such as blogs, as well as on spoken text. Such a semi-unsupervised corpus creation approach can easily be adapted for different languages and different emotion applications. Moreover, it provides a standard for emotion annotation, which avoids the controversy over emotion classes and types. Overall, the created annotated emotion corpus is of a comparatively high quality and is suitable for both empirical emotion analysis and emotion processing. As the size of the neutral corpus is much bigger than its corresponding emotion corpus, to avoid model bias, we randomly select some neutral sentences from the neutral corpus to combine them with their corresponding emotion sentences to form a complete emotion-driven corpus.
Chapter 4
Linguistic Expression of Cause Event I: Transitivity
4.1
Introduction
As defined in Chap. 2, the causal relation is established by the linking of an existing event and the emotion itself. Such an event is referred to as the cause event. Despite its integral role in emotion studies, little research focuses on the interactions between cause events and emotions. This chapter aims to present a linguistic analysis of causal relations with regard to two dimensions of emotions: emotion classes (i.e. HAPPINESS, SADNESS, FEAR, ANGER, and SURPRISE) and emotion verb types (change-of-state emotion verbs vs. homogeneous state emotion verbs). This study first examines the degree of transitivity of cause events based on three features, namely agentivity, kinesis, and participation. It argues that an in-depth linguistic analysis of cause events will have important implications for a linguistic theory of emotions where emotions, instead of being defined in an abstract way, are concretely described based on the linguistic behaviour of cause events. Such valuable information is also crucial for real world applications. Information, such as who the experiencer of an emotion is, and what triggers an emotion, are essential to economic forecasting and product design, among other functions. Section 4.2 presents the previous analyses of the concept of transitivity. Section 4.3 proposes the cause event features based on the degree of transitivity in analyzing emotion causal relations. Section 4.4 describes the methodology of the present study, and how emotion causes are linguistically expressed and analyzed in terms of cause event features in Chinese. Section 4.5 shows the results of the cause event feature analysis of the five primary emotions and discusses the correlations between the causes and the corresponding emotions. Section 4.6 gives a summary of the main points addressed in the chapter.
© Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_4
47
48
4 Linguistic Expression of Cause Event I: Transitivity
4.2
Concepts of Transitivity
The notion of transitivity has always been one of the main issues in linguistics as it is “a central phenomenon in the structure of human languages, and appears to be universal” (Næss 2007: 1). Transitivity is traditionally understood to be a property of verbs that determines whether or not a verb can take direct objects (Robins 1964; Richards et al. 1985). Transitive verbs such as kick and beat can take a direct object, whereas intransitive verbs such as cry and sleep cannot take a direct object. This attempt of making a clear-cut dichotomy of transitive and intransitive clauses is rather loose and results in a number of problems, such as failing to account for ergative constructions (Comrie 1978; Hopper and Thompson 1980). Instead of merely depending on morphosyntactic features, later proposals employ a prototype approach to define transitivity, in which transitivity is considered to be a continuum rather than a binary category. This approach takes into account the degree to which an action affects its object. For example, the verb hear is described as having “lower transitivity” than the verb hit. I will briefly discuss three contemporary prototypical approaches by Lakoff (1977), Hopper and Thompson (1980), and Givón (1993), as all three endeavours to account for transitivity in a universal sense.
4.2.1
Lakoff’s Prototypical Agent-Patient Sentences
Lakoff (1977) proposed that a prototypical transitive sentence, i.e. an agent-patient sentence, should have the following fourteen semantic properties, as shown in (1). Such semantic properties emphasize the volitionality, animacy, and control of the agent as well as the affectedness of the patient. The more semantic properties a sentence satisfies, the higher the transitivity of the sentence. (1) Lakoff’s Transitivity Properties a. there is a [sic] agent, who does something b. there is a patient, who undergoes a change to a new state (the new state is typically nonnormal or unexpected) c. the change in the patient results from the action by the agent d. the agent’s action is volitional e. the agent is in control of what he does f. the agent is primarily responsible for what happens (his action and the resulting change) g. the agent is the energy source in the action; the patient is the energy goal (that is, the agent is directing his energies toward the patient) h. there is a single event (there is spatio-temporal overlap between the agent’s action and the patient’s change)
4.2 Concepts of Transitivity
i. j. k. l. m. n.
49
there is a single, definite agent there is a single, definite patient the agent uses his hands, body, or some instrument the change in the patient is perceptible the agent perceives the change the agent is looking at the patient
(Lakoff 1977: 244)
4.2.2
Hopper and Thompson’s Transitivity Hypothesis
Similar to Lakoff, Hopper and Thompson (1980) argued that transitivity involves a number of component features. They identified transitivity as a composition notion consisting of ten interacting but basically separate parameters, as shown in (2): (2) Hopper and Thompson’s Transitivity Parameters (Hopper and Thompson 1980: 252) a. b. c. d. e. f. g. h. i. j.
Criterion
High transitivity
Low transitivity
Participants Kinesis Aspect Punctuality Volitionality Affirmation Mode Agency Affectedness of Object Individuation of Object
2 or more, agent and object Action Telic Punctual Volitional Affirmative Realis A high in potency Totally affected Highly individuated
1 participant Non-action Atelic Non-punctual Non-volitional Negative Irrealis A low in potency Not affected Non-individuated
Some of the parameters refer to morphosyntactic properties, while others refer to semantic entities. Of the latter group, some refer to the agent of the event, while others refer to the patient of the event. In the following, each parameter will be briefly delineated. A. Participants: Only the presence of two or more participants can do the transfer. B. Kinesis: Only actions, not states, can be transferred from one participant to another. Thus, something happens to John in I hit John, but not in I love John. C. Aspect: An action in which its endpoint is provided, i.e. a telic action, is more effectively transferred from an agent to a patient than one without such an endpoint, i.e. atelic action. For example, in the sentence I walked to the house, the transferral is completely carried out, whereas in I was walking, the activity is viewed as incomplete.
50
4 Linguistic Expression of Cause Event I: Transitivity
D. Punctuality: actions carried out with no obvious transitional phase between start-point and endpoint have a more marked effect on their patients than actions which are inherently on-going. E. Volitionality: The effect on the patient is typically more apparent when the agent is presented as acting purposefully. F. Affirmation: This is the affirmative or negative parameter. G. Mode: This refers to the distinction between the ‘realis’ and ‘irrealis’ encoding of events. An action which either did not occur, or which is presented as occurring in a non-real world is obviously less effective than one whose occurrence is actually asserted as corresponding directly to a real event. H. Agency: Participants high in Agency can produce a transfer of an action, whereas those low in Agency cannot. I. Affectedness: The degree to which an action is transferred to a patient is a function of how completely that patient is affected. J. Individuation: It refers both to the distinctness of the patient from the agent and its distinctness from its own background. The ten parameters determine the effectiveness with which an action is transferred from one participant to another, each of which constitutes a scale ranging from high to low. Similar to Lakoff’s proposal, the more effective the transfer, the more transitive the clause is.
4.2.3
Givón’s Transitivity
Givón (1993) defined the concept of transitivity in terms of semantic as well as syntactic terms. Compared to Lakoff (1977) and Hopper and Thompson (1980), Givón (1993) generalized a much smaller set of semantic features to characterize the transitive prototype, as can be seen in (3): (3) Semantic definition of the prototype transitive clause (Givón 1993) a. Agentivity: The subject of a prototypical transitive clause is a deliberately acting agent. b. Affectedness: The direct object of a prototypical transitive clause is a concrete, visibly affected patient. c. Perfectivity: The prototypical transitive verb codes a bounded, terminated, fast-changing event that took place in real time. The parameters summarize the relevant semantic features into three main components of transitivity. These are agent-related, patient-related, and verb-related properties. Just as in the other prototype theories, agent and patient need to possess specific properties to allow action transfer. Givón (1993) also provided a common syntactic definition of transitive clauses for English:
4.2 Concepts of Transitivity
51
(4) Syntactic definition of the transitive clause (Givón 1993: 100) Verbs (and clauses) that have a direct object will be considered transitive; verbs (and clauses) that do not have a direct object will be considered intransitive. Givón linked the two definitions by noting that there is a high statistical overlap between semantic and syntactic transitivity in English. That is, the majority of semantically transitive clauses is also syntactically transitive, and vice versa.
4.2.4
Transitivity and Affectedness
Having examined these three well-known prototype approaches to transitivity, it is apparent that the ultimate basis of transitivity lies in the volitionality of the agent as well as the affectedness of the object in the sentence. However, as Tsunoda (1999) points out, affectedness is most relevant to morphosyntactic manifestations of transitivity since it is always the affectedness of the patient, rather than the volitionality, that plays a positive role in manifesting a transitive case frame. As a result, Tsunoda (1999) argues that the parameter “affectedness” is generally sufficient in determining transitivity. This can also be supported by the fact that most of Hopper and Thompsons’s (1980) parameters can be subsumed under the property of affectedness. For instance, a patient is more likely to be affected by an action (kinesis), a sudden event (punctuality), in an affirmative sentence (affirmation), or in a realis sentence (mode), etc. Following this line of thinking, it is reasonable to define transitive prototypes based on its degree of affectedness. In other words, transitivity is the degree to which an argument (usually an agent) affects another argument (usually a patient) in an event. Nevertheless, the concept of affectedness is rather abstract and complicated. Thus, the present work will focus instead on the three parameters of agentivity, kinesis, and participation as these are properties that can be identified and extracted from the corpora.
4.3
Cause Event Features
In order to give an in-depth analysis of cause events based on their degree of transitivity, this study specifically focuses on three parameters, namely agentivity, kinesis, and participation as has been pointed out above. These parameters have the advantage of being similar to some of those proposed by Hopper and Thompson (1980), which are relatively easy to identify in the corpora and can be adapted with certain modifications in response to causal relations. They are used to examine how and to what extent the cause event affects the experiencer of the emotion. Experiencer here refers to the entity experiencing some psychological state due to
52
4 Linguistic Expression of Cause Event I: Transitivity
the occurrence of certain events. The following sections will discuss in detail how the three parameters are determined.
4.3.1
Agentivity
Semantic roles have been well studied; however, the nature, the definitions and even the number of argument roles vary in different theories. Following the pioneering work of Fillmore (1968) and Jackendoff (1972), many theorists attempted to define a set of semantic roles, such as agent, patient, theme, experiencer, and goal (Cruse 1973; van Valin 1990; Carnie 2006). However, such theories of semantic roles lead to a proliferation problem as one looks at more and more examples in different languages. Therefore, there is always a question of how fine-grained the role types should be. While role types are generally considered discrete categories, Dowty (1991) argued that arguments may have different degrees of membership in a role type. Hence, it is efficient to have only two basic prototypical role types to describe argument selection, which are the Agent Proto-role and Patient Proto-role respectively. These argument roles are characterized by a list of contributing properties, as given in (5) and (6). (5) Contributing properties for the Agent Proto-role: a. b. c. d. e.
volitional involvement in the event or state sentience (and/or perception) causing an event or change of state in another participant movement (relative to the position of another participant) (exists independently of the event named by the verb)
(6) Contributing properties for the Patient Proto-role: a. b. c. d. e.
undergoes change of state incremental theme causally affected by another participant stationary relative to movement of another participant (does not exist independently of the event, or not at all) (Dowty 1991: 572)
Arguments of the verbs will be more agent-like or patient-like according to the number of agent or patient proto-role properties they fulfill. The argument with the largest number of agent-role properties will be the subject, and the other argument will be the object. In case they fulfill the same number of proto-agent properties, either one can be the subject. This approach overcomes the problems of role fragmentation and unclear role boundary that traditional role type distinctions have. I identify the subject of the cause event as one of the two prototypical role types according to the lists of contributing properties as shown in (5) and (6). The subject of the cause event is considered more agent-like or patient-like according to the
4.3 Cause Event Features
53
number of Agent or Patient Proto-role properties they fulfill. The more agent-like the subject, the higher intention the subject of the cause event affects the experiencer. Moreover, high agentivity implies high energy level which is more likely to be carried over to the patient causing the patient to be affected.
4.3.2
Kinesis
Kinesis refers to whether the cause event is a motion or non-motion. According to Fillmore (1968), a motion event denotes an event in which the participant undergoes change of location or change of state. Such motion activity can be transferred from one participant to another; therefore, a motion cause event is more likely to directly affect the experiencer. Non-motion events, on the other hand, tend to affect the experiencer less directly. In addition, Gao (2001: 58) identified three separate states in referring to the notion of motion, as in (7): (7) Gao’s Three States of Motion 1. an independent motion action of the Agent rooted in the verb itself which entails nothing about the Figure’s moving 2. a motion event in which the Figure is caused by the Agent’s impact to move to a location 3. a motion event in which the Agent and the Figure move together by the impact of the Agent to a location In the analysis of cause events of emotions, this study takes Fillmore’s viewpoint regarding the concept of motion events and identify the types of motion events according to Gao’s classification. Some typical examples of motion and non-motion cause events are listed in Tables 4.1 and 4.2.
4.3.3
Participation
Participation indicates whether the emotion experiencer takes part in the cause event. I examine whether the experiencer is one of the participants of the cause event. An experiencer involved in the cause event may act as an agent or a patient in the event. When an experiencer takes part in the cause event, it is more likely that the experiencer is directly affected in one way or another.
54
4 Linguistic Expression of Cause Event I: Transitivity
Table 4.1 Common types of motion cause events Type
physical movements
Examples
ge1ge ba3 qi4shui3 da3kai1 elder brother BA soda drink open ‘My elder brother opened the soda drink’
chuang3jin4 wo3men de break into 1.PL POSS ‘[Someone]a broke into our house’
intention
appearing/ disappearing
fang2zi house
di4di yao4 shang4xue2 le younger brother need go to school LE ‘My younger brother plans to go to school’
Kai3luo4lin2 chu1xian4 zai4 can1guan3 Caroline appear at restaurant ‘Caroline appeared at the restaurant doorway’
shao3 le zhi4ai4 de lose ASP loved POSS ‘[Someone] lost a loved one’
men2kou3 doorway
ren2 person
mian4lin2 zhe4xie1 jie1zhong3er2zhi4 de face these follow POSS ‘[Someone] faced with the difficulties followed’
kun4nan3 difficulty
encountering zao1yu4 cuo4zhe2 encounter setback ‘[Someone] encountered the setbacks’
(continued)
4.3 Cause Event Features
55
Table 4.1 (continued)
wo3 zhong1yu1 hui4 qi2 1.SG finally can ride ‘I was finally able to ride’
le ASP
ability Ma3ying1jiu3 fu4 ou1 kao3cha2 wen2hua4 Ma Ying-jeou go Europe inspect culture chan3ye4 neng2 cheng2xing2 industry can form the trip ‘The trip for Ma Ying-jeou to inspect the cultural industry in Europe has been confirmed’
yong3duo2 guan4jun1 win champion ‘Won the championship’ receiving Dong1ni2 de2dao4 Tony get ‘Tony got the approval’
winning/ losing
le ASP
shou3ken3 approval
ge1ge da3 duo3bi4qiu2 sheng4 elder brother hit dodge-ball win ‘My elder brother won the dodge-ball game’
shi1qu4 zi4you2 huo2dong4 de lose free move POSS ‘[Someone] lost the ability to move freely’
admitting
le ASP
neng2li4 ability
cheng2ren4 zi4ji3 zhen1zheng4 de yu4wang4 admit oneself real POSS desire ‘[Someone] admitted their real desire’
[Someone/something] is glossed to indicate there is no corresponding overt expression in the Chinese sentence
56
4 Linguistic Expression of Cause Event I: Transitivity
Table 4.2 Common types of non-motion cause events Type
being
Examples
ta1 jia1 de yin1xiang3 shi4 shi4jie4shang4 3.SG.M home POSS audio equipment is world zui4hao3 de zu3he2 zhi1yi1 the best POSS combination one of ‘His home sound system is one of the best in the world’
cheng2wei2 tong3yi1shi1 de yi4yuan2 become the Lions POSS member ‘[Someone] became a member of the Lions’
laughing/ crying
tong2xue2 chao2xiao4 ta1 de xiang1xia4 classmate laugh 3.SG.M POSS rural ‘Classmates laughed at his rural accent’
gong1nü3men hai2shi4 zhi3 gu4 Maiden.PL still only concern ‘The maidens are all laughing’
death/ illness
kou3yin1 accent
xi1xiao4 laugh
Xue1ping2gui4 yi3jing1 zhan4 si3 Xue Ping Gui already war die ‘Xue Ping Gui has been killed in the war’
xiang3dao4 zhe4le3 think of here ‘When [someone] was thinking of this, …’ thinking ta1 zhong1yu2 xiang3tong1 le 3.SG.M finally figure out ASP ‘He has finally figured it out’ xiao4zhang3 xiang4 wo3men bao4gao4 zhe4ge4 hao3 xiao1xi the vice chancellor to 1.PL report this good news ‘The Vice-Chancellor reports to us the good news’ speaking ba4ba ma1ma ma4 wo3 dad mom scold 1.SG ‘Dad and mom scolded me’
4.3 Cause Event Features
meeting/ seeing
57
fu4nü3 zai4 feng1yu3 zhong1 father and daughter at rain under ‘The father and the daughter met in the rain’
xiang1jian4 meet
zi3xi4 guan1shang3 guo4 zhe4xie1 she4ying3ji2 careful watch ASP these photo collection ‘[Someone] looked at these photo collections carefully’
accompany
capability
ba4ba ma1ma cong2 kai1mu4 dao4 bi4mu4 dad mom from opening to closing dou1 yi4zhi2 pei2zhe wo3men also PROG accompany 1.PL ‘Dad and Mom have accompanyied us from the beginning to the end [of the event]’
yu3yan2 neng2li4 hai2 bu2gou4 language ability yet not enough ‘Language ability is not good enough’
hao3 good
mai4 bu2 diao4 sell no away ‘[Something] cannot be sold able.’
bu4neng2 dui4 wo3 zuo4 xie1 cannot to 1.SG do some ‘[Someone] cannot do anything to me’
shen2me something
negation ta1men bu4neng2 bai3tuo1 3.PL cannot get rid of ‘They cannot get rid of traditions’
chuan2tong3 tradition
zuo4 zai4 can1zhuo1 qian2 bu4 li2kai1 sit at dining table front not leave ‘[Someone] was sitting at the dining table without leaving’
58
4.4
4 Linguistic Expression of Cause Event I: Transitivity
Methodology
This study gives an in-depth analysis of causes based on the three features of transitivity, i.e. agentivity, kinesis, and participation. It examines whether there are relations between the classes of causes and the classes of emotions and/or between the classes of causes and the emotion verb types. This would offer some predictions as to what kind of causes tends to be more associated with certain emotions.
4.4.1
Data Collection
This work aims to analyze 100 emotional sentences with cause events from each group of primary emotion verbs. The data are mostly taken from the Sinica Corpus with a few from the Chinese Gigaword Corpus. Using Chinese Word Sketch, it randomly extracts 100 sentences of each representative verb in each emotion class (i.e. HAPPINESS, SADNESS, FEAR, ANGER, and SURPRISE) and emotion type (i.e. change-of-state emotion verb vs. homogeneous state emotion verb) from the Sinica Corpus together with their contexts in which the causes of the emotions appear, i.e. the focus sentence that contains the emotion keywords plus the sentence before it and the sentence after it. In other words, there are 100 sentences with the homogeneous state verb HAPPINESS (i.e. 快樂 kuai4le4), 100 sentences with the change-of-state verb HAPPINESS (i.e. 高興 gao1xing4), 100 sentences with the homogeneous state verb SADNESS (i.e. 悲傷 bei1shang1), 100 sentences with the change-of-state verb SADNESS (i.e. 傷心 shang1xin1), and so on. As the main objective of this study is to examine the relations between cause events and emotions, only emotion sentences with explicit cause events would be examined. Recall that sentences containing emotion verbs do not always express or ascribe emotion; i.e. they are not ‘emotional’ in this sense. Therefore, this study filters out non-emotion sentences as well as emotion sentences without explicit causes (as explained in Sect. 2.3.1 in Chap. 2). When sentences do not reach 100 in number after filtering out the non-emotional sentences and emotional sentences without explicit causes, I extract another 100 sentences and follow the analysis until there are 100 sentences for each type of emotion class. In total, 1000 sentences are analyzed, as summarized in Table 4.3.
Table 4.3 Number of sentences for cause event feature analysis
HAPPINESS SADNESS FEAR ANGER SURPRISE
Total
Change-of-state
Homogeneous state
100 100 100 100 100 1000
100 100 100 100 100
4.4 Methodology
59
In the cases of the homogeneous state verb SADNESS 悲傷 bei1shang1, homogeneous state verb ANGER 憤怒 fen4nu4, and homogeneous state verb SURPRISE 驚訝 jing1ya4 in which less than 100 sentences are found in the Sinica Corpus, I extracted the rest of the data from the Chinese Gigaword Corpus.
4.4.2
Data Analysis
As discussed in Chap. 2, cause events in Chinese are usually expressed by means of propositions, nominalizations, and nominals. Each expression has specific constructions which require different analyses with regard to the degree of transitivity.
4.4.2.1
Propositional Causes
Cause events are mostly represented by a proposition. In Chinese, a proposition is usually expressed as [subject, verb, object/complement]. For example, see (8) and (9) where the propositional cause events are underlined. (8) zhe4 tou2 niu2 de zhu3ren2, yan3kan4 zi4ji3 de niu2 re3chu1 huo4 lai2 le, DET CL ox POSS owner, see oneself POSS ox cause trouble come ASP, fei1chang2 hai4pa4, jiu4 ba3 zhe4 tou2 niu2 di1jia4 mai4 chu1 very becoming frightened, then BA DET CL ox low price sell out ‘The owner was frightened to see that his ox was causing trouble, so he sold it off at a low price.’ (9) mei2 xiang3dao4 ta1 shuo1 de dou1 shi4 zhen1 hua4, rang4 ta1 zhen4jing1 bu4yi3 not think 3SG.F say POSS all is true word, cause 3.SG.M to be shocked very ‘Unexpectedly, what she said was the truth, which surprised him greatly.’
Based on their transitivity properties, the cause events of (8) and (9) are analyzed as [+agent, +motion, −participation] and [+agent, −motion, −participation], respectively. This shows that the cause event of (8) has a higher transitivity than that of (9). Nonetheless, it is not uncommon to find cases of object-drop, complement-drop, and even subject-drop. Consider (10) for instance:
60
4 Linguistic Expression of Cause Event I: Transitivity
(10) wo3 xiang3 wo3 hai2shi4 mei2you3 yong3qi4 fang4xia4 yi2qie4 qu4 zhao3xun2 1.SG think 1.SG still no courage leave everything to find meng4jing4 li3 de na4 ge4 hu2 ba ye3xu3 PROi hai4pa4 dream in POSS DET CL lake PART perhaps PROi becoming frightened PROi zhao3bu2dao4 PROj, ye3xu3 PROi hai4pa4 PROi zhao3dao4 le PROj PROi cannot find PROj, perhaps PROi becoming frightened PROi find ASP PROj PROi hui4 shi1wang4 PROi will disappointed ‘Ii don’t think Ii have the courage to leave everything behind to find the lakej appearing in the dream! Perhaps, PROi fear that PROi will not find PROj, perhaps PROi fear that PROi will be disappointed when PROi find PROj…’
We can see that in the cases of the two FEAR keywords 害怕 hai4pa4, the subjects and the objects are not explicitly shown in the causes. The covert arguments hinder the transitivity analysis. Therefore, we need to retrieve the covert subjects and objects in order to analyze the cause event features. The missing arguments can usually be retrieved with the help of the context, as indicated in (11). With the retrieved arguments, the two cause events of the two emotion verbs of FEAR are analyzed as [+agent, −motion, +participation] and [+agent, +motion, +participation], respectively. (11) wo3 xiang3 wo3 hai2shi4 mei2you3 yong3qi4 fang4xia4 yi2qie4 qu4 zhao3xun2 1.SG think 1.SG still no courage leave everything to find ye3xu3 wo3 meng4jing4 li3 de na4 ge4 hu2 ba hai4pa4 dream in POSS DET CL lake PART perhaps 1.SG becoming frightened ye3xu3 wo3 hai4pa4 wo3 zhao3bu2dao4 na4 ge4 hu2, 1.SG cannot find DET CL lake, perhaps 1.SG becoming frightened wo3 zhao3dao4 le na4 ge4 hu2 wo3 hui4 shi1wang4 1.SG find ASP DET CL lake 1.SG will disappointed ‘I don’t think I have the courage to leave everything behind to find the lake which appeared in my dream! Perhaps, I fear that I will not find it, perhaps I fear that I will be disappointed when I find it.’
4.4.2.2
Nominalized Causes
Cause events can sometimes be expressed as nominalizations. Nominalization refers to the construction in which a verb phrase or a clause is turned into a noun phrase. In Chinese, nominalization is mostly context-dependent without morphological transformation (Yip and Rimmington 2004). In such case, the nominalized phrases have the same form as their corresponding verb predicates or clauses as in (12) and (13), respectively:
4.4 Methodology
61
(12) piao1bo2 shi4 ling4 drift is cause ‘Drifting is depressing.’
ren2 people
an4ran2shen2shang1 depressed
(13) bu4neng2 yu3 jia1ren2 chi1fan4 shi4 yi2 jian4 hen3 is one CL very cannot and family members eat ‘Not being able to eat with one’s family is a very sad thing.’
bei1ai1 de sad POSS
shi4 thing
In (12), the verb 漂泊 piao1bo2 ‘to drift’ is nominalized as a noun ‘drifting’, whereas in (13), the clause 不能與家人吃飯 bu4neng2 yu3 jia1ren2 chi1fan4 ‘(one) is not able to eat with the family’ is taken as a noun ‘not being able to eat with the family’. The nominalized phrases occur primarily as subjects or objects. In some cases, nominalization can also undergo morphological transformation by placing a clitic de after what is otherwise a verb phrase or sentence, as in (14) and (15). (14) ta1 dui4 zhe4 ge4 chong1man3 nong2hou4 3.SG.M for DET CL full of dense gao1xing4 de shou3wu3zu2dao3 becoming happy DE flourish ‘He was very happy about this lovely idea.’
ai4yi4 de love DE
xiang3fa3 idea
(15) zhe4 ci4 yan3chu1 de jing1zhi4 dao4shi4 ling4 is cause DET CL perform POSS exquisite ‘I was very surprised by this exquisite performance.’
wo3 shi2fen1 jing1ya4 1.SG very becoming surprised
In connecting nominalizations and events, Davidson (1967) argues that an event variable should be added to verbs, such as destroying and nouns, such as development based on inference patterns. Following Davidson, (14) and (15) can be reinterpreted as (16) and (17), respectively. (16) zhe4 ge4 xiang3fa3 chong1man3 nong2hou4 ai4yi4, ta1 DET CL idea full of dense love, 3.SG.M gao1xing4 de shou3wu3zu2dao3 becoming happy DE flourish ‘He was very happy about an event which was an idea full of love.’
dui4 ci3 for this
(17) zhe4 ci4 yan3chu1 hen3 jing1zhi4, ling4 wo3 shi2fen1 jing1ya4 very becoming surprised DET CL perform very exquisite, cause 1.SG ‘I was very surprised by an event which was an exquisite performance.’
62
4 Linguistic Expression of Cause Event I: Transitivity
By inferences, the pair of sentences (14) and (16) provides similar, if not the same, semantic information. This is also illustrated in the other pair of sentences (15) and (17). In order to analyze emotions in terms of cause events, nominalizations should be reinterpreted as events by inferences.
4.4.2.3
Nominal Causes
While both propositions and nominalizations involve the presence of verbs, cause events can be simply expressed as nominals, such as in (18) and (19): (18) Ni2ao4 de hua4 hen3 ling4 Kai3luo4lin2 shang1xin1 Leo POSS word very make Caroline becoming sad ‘Caroline was very saddened by Leo’s words.’ (19) dui4yu1 wei4lai2, lao3shi2shuo1 wo3 hen3 hai4pa4 for future, frankly 1.SG very becoming scared ‘Frankly, I am very scared about the future.’
4.4.2.4
Event Coercion
From the discussion above, it can be noted that unlike propositional and nominalized cause events, nominal causes are noun phrases (NPs) which lack a predicate that subcategorizes the required arguments. However, it is rather intuitive that a cause is actually an event of some sort. Regardless of the surface syntactic form of the NP causes, their semantic environment still represents an event. Since the semantic environment of nominal causes is not syntactically satisfied, coercion can be applied to reconstruct the semantics of the causes (Pustejovsky 1995). Pustejovsky (1995: 111) defines coercion as “a semantic operation that converts an argument to the type which is expected by a function, where it would otherwise result in a type error”. Following his idea, I argue that nominal causes are typed as propositions; if the syntactic form of a cause event matches the required type, i.e. a proposition, the structure is considered well-formed. If, however, the type is not propositional, i.e. a nominal cause, it should be coerced to match the type required by the typing restriction on the cause event. Undergoing coercion allows interpretations that are licensed by the reconstruction of the semantics of the nominals. Sentences with a nominal cause must ensure argument coherence so as to satisfy the causal relation. In doing so, the nominal cause should undergo a metonymic reconstruction of an experiencing event between the experiencer and the nominal cause along with pragmatic enrichment. This allows the functions of a lexical item
4.4 Methodology
63
to be expressed in a single form which requires reference to multiple type lattices through the qualia structure. Therefore, in (18), it is (Caroline’s listening to) Leo’s words which depressed her, while in (19), it is (my thinking about) the future that scared me, as shown in (20) and (21), respectively. (20) Kai3luo4lin2 ting1dao4 Ni2ao4 de hua4, ling4 Caroline hear Leo POSS word, make ‘Caroline’s listening to Leo’s words depressed her.’
Kai3luo4lin2 shang1xin1 Caroline becoming sad
(21) dui4yu1 wo3 xiang3qi3 wei4lai2, lao3shi2shuo1 wo3 hen3 for 1.SG think about future, frankly 1.SG very ‘Frankly, I am very scared when thinking about future.’
hai4pa4 becoming scared
For words such as Leo’s words and future, the TELIC quale role values of listen and think respectively are part of the semantics of the words. The experiencer verbs 傷心 shang1xin1 and 害怕 hai4pa4 are selected for direct perceptual experiences as a result of direct visual and auditory stimuli, i.e. listening and thinking. One may argue that there are different ways of experiencing the nominal causes, such as Caroline being sad by listening, seeing, or thinking about what Leo said. Yet, this is by no means a defect in the coercion mechanism as the qualia structure identifies two types of information for coercion: i. Type and sort information which the qualia must satisfy; ii. Specific qualia values which are explanatory modes in understanding a word. (Pustejovsky 1995: 209) In other words, when a nominal cause enters into a coercive environment, the qualia values determine how the semantic type is reconstructed and also explain the causal relation between the experiencer and the cause. Consider the verb 傷心 shang1xin1 ‘to be sad’, which involves someone who experiences something and as a result becomes sad. The experiencing event in (18) is indicated by the nominal Leo’s words which do not satisfy the type required by the predicate shang1xin1 as it is a proposition. Therefore, the predicate coerces the nominal cause into an event denotation whose qualia structure contributes information to the interpretation of what kind of experiencing event is involved through qualia projection. The lexical representation for the nominal Leo’s word is given in (22):
64
(22)
4 Linguistic Expression of Cause Event I: Transitivity
Leo’s words ARGSTR =
[ ARG1 = x:info ] FORMAL = x
QUALIA =
TELIC = listen (e, y, x) AGENTIVE = say (e, Leo, x)
For a nominal, such as Leo’s words, the TELIC quale role value of listen is part of the semantics of the nominal. This qualia value determines the default assignment for type reconstruction to satisfy the type environment. As a result of the coercion operation, the nominal cause is coerced to an experiencer event, i.e. listen to Leo’s words. The coerced event projected from the TELIC value of Leo’s words satisfies the argument requirements of the experiencer verb shang1xin1 on its subject, i.e. an event (e1). This is illustrated in the semantics of the verb 傷心 shang1xin1 in (23):
(23)
sad E1 = e1:process EVENTSTR
E2 = e2:state RESTR = < α ARG1 = x []
ARGSTR =
ARG1 = y [animate_ind ] experiencer_lcp
QUALIA =
FORMAL = sad (e2, y) AGENTIVE = exp_act (e1, y)
As indicated in (23), an experiencer verb such as sad, selects function (e1) for a process event in the subject position and for a state event (e2) in object position. When the nominal in the subject position does not satisfy the selectional requirement of the experiencer verb, it is coerced into an event type by the verb. This default coercion of experienced causation, however, does not necessarily apply to Chinese emotion sentences. Unlike English which is a subject-prominent language, Chinese is considered a topic-prominent language as the basic sentence structure can be more insightfully described in terms of the topic-comment relation than in terms of the subject-predicate relation (Li and Thompson 1981). In addition, it is the
4.4 Methodology
65
semantic factors rather than the syntactic ones that determine the order of major constituents of experiencer verbs. In Chinese, a cause can not only take the subject position of an experiencer verb, but can also be the topic or the object. Therefore, a cause, be it a subject, a topic, or an object, should be reconstructed as an event which is coerced by the experiencer verb through the qualia structure. This is not in any way inconsistent with the Generative Lexicon approach since “there is no one-to-one mapping from underlying semantic types to syntactic representations, a syntactic phrase is only fully interpretable within the specific semantic context within which it is embedded” in a Generative Lexicon (Pustejovsky 1995: 132).
4.5
Results and Discussion
The analysis of 1000 emotional sentences with cause events reveals the relationship between emotion classes and causes as well as between emotion verb types and causes. A summary of the cause event feature tendency of each emotion class is presented in Tables 4.4 and 4.5. Table 4.4 shows the cause event features of change-of-state emotion verbs, while Table 4.5 shows the cause event features of homogeneous state emotion verbs. As shown in Tables 4.4 and 4.5, there are tendencies of cause event features for different emotion classes. The cause events of change-of-state and homogeneous state emotion verbs are expressed similarly, except for FEAR. The preferred cause event features of each emotion class are summarized in Table 4.6. Cause events of HAPPINESS have the highest transitivity among the five primary emotions where the subjects of cause events are mostly proto-agent, the causes tend
Table 4.4 Distributional tendency of cause event features of change-of-state emotion verbs Emotions
Agentivity Proto-agent (%)
Proto-patient (%)
高興 gao1xing4 89 11 ‘becoming happy’ 傷心 shang1xin1 75 25 ‘becoming sad’ 生氣 sheng1qi4 84 16 ‘becoming angry’ 害怕 hai4pa4 79 21 ‘becoming frightened’ 驚訝 jing1ya4 67 33 ‘becoming surprised’ Based on 100 sentences for each verb
Kinesis Motion (%)
Non-motion (%)
Participation Yes No (%) (%)
52
48
75
25
36
64
41
59
55
45
49
51
41
59
57
43
62
38
23
77
66
4 Linguistic Expression of Cause Event I: Transitivity
Table 4.5 Distributional tendency of cause event features of homogeneous state emotion verbs Emotions
Agentivity Proto-agent (%)
快樂 kuai4le4 97 ‘to be happy’ 悲傷 57 bei1shang1 ‘to be sad’ 憤怒 fen4nu4 90 ‘to be angry’ 恐懼 63 kong3ju4 ‘to be frightened’ 震驚 58 zhen4jing1 ‘to be surprised’ Based on 100 sentences for each
Kinesis Motion (%)
Non-motion (%)
Participation Yes No (%) (%)
3
78
22
85
15
43
29
81
36
64
10
57
43
43
57
37
64
36
64
36
42
52
48
9
91
Proto-patient (%)
verb
Table 4.6 Cause event features of each emotion class Emotion class
Cause event features
HAPPINESS
[+proto-agent] [+motion] [+participation] [+proto-agent] [−motion] [−participation] [+proto-agent] [+motion] [−participation] Change-of-state [+proto-agent] [−motion] [+participation] Homogeneous [+proto-agent] [+motion] [+participation] [+proto-agent] [+motion] [−participation]
SADNESS ANGER FEAR
SURPRISE
to be motion events, and the experiencers are mostly involved in the cause events. This can be explained by the fact that HAPPINESS is a desirable emotion which is usually a by-product of events or active states (Izard 1977). SADNESS, in contrast, is usually triggered by disappointments about people, relationships, conditions, or unknown reasons (Johnson-Laird and Oatley 1989). Thus, the cause events tend to have lower transitivity, i.e. are less agentive, do not have motion, and include less involvement on the part of the experiencer. The two emotions being at the two ends of the transitivity continuum also explains why HAPPINESS and SADNESS are often considered opposite emotions. The cause events of ANGER and SURPRISE share similar features, i.e. the proto-agent subjects, the motion events, and the lower involvement of the experiencer. This may be attributed to the fact that both emotions often result from sudden, physical or psychological activity, such as hitting for ANGER and appearing
4.5 Results and Discussion
67
Table 4.7 Comparing cause event features between emotion verb types Emotions
Agentivity Proto-agent (%)
Kinesis Motion (%)
89 (high) 97 (high)
51 78
Non-motion (%)
Participation Yes (%) No (%)
HAPPINESS
高興 gao1xing4 快樂 kuai4le4
75 85
SADNESS
傷心 shang1xin1 悲傷 bei1shang1
75 (low) 57 (low)
64 81
59 64
ANGER
生氣 sheng1qi4 憤怒 fen4nu4
84 (low) 90 (high)
55 57
51 57
67 (low) 58 (low)
62 52
77 91
SURPRISE
驚訝 jing1ya4 震驚 zhen4jing4
high
transitivity HAPPINESS
> homogeneous FEAR > CoS-FEAR, SURPRISE, ANGER >
low SADNESS
Fig. 4.1 The transitivity continuum
for SURPRISE. As hitting and appearing suggest the degree of agentivity associated with the causes of ANGER is significantly higher than that of SURPRISE. FEAR, which differs from the other primary emotions, can reflect involvement with two types of cause event features: [+proto-agent] [−motion] [+participation] for change-of-state emotion verbs and [+proto-agent] [+motion] [+participation] for homogeneous state emotion verbs. After looking into the analyzed data, I find that change-of-state emotion verbs tend to be associated with potential causes, whereas homogeneous state emotion verbs tend to be associated with actual causes. Potential causes are usually imaginary events that may or may not happen, e.g., a fear of being alone, an image of a tiger, etc. These kinds of causes do not necessarily involve motion. Actual causes are events that happened and triggered the presence of FEAR, such as being attacked by a tiger, being kidnapped, etc. These events are usually activities that involve motion. In addition, experiencers of FEAR usually participate in the cause events regardless of the emotion verb types. They are mostly patients in the cause events. Although change-of-state and homogeneous state emotion verbs of most emotion classes share similar cause event features, they appear to differ in their semantic tendencies. Table 4.7 shows the statistical comparisons between the cause event features of change-of-state verbs and homogeneous state verbs. It can be seen that homogeneous state verbs tend to have a stronger tendency toward the found cause event features than that of change-of-state verbs, as indicated by the great number of cause event feature tendencies in the group of homogeneous state verbs. For
68
4 Linguistic Expression of Cause Event I: Transitivity
instance, while verbs of HAPPINESS tend to involve high agentivity, motion, and self-participated events, the homogeneous state verb 快樂 kuai4le4 has a stronger tendency of agentivity (97%), motion event (78%), and self-participating events (85%) compared to that of change-of-state verb 高興 gao1xing4, i.e. 89% for agentivity, 51% for kinesis, 75% for participation. Figure 4.1 summarizes the transitivity of cause events of the five primary emotions with HAPPINESS having the highest transitivity and SADNESS the lowest transitivity. The other three primary emotions fall between HAPPINESS and SADNESS, where homogeneous state verbs deriving from FEAR have higher transitivity than change-of-state (CoS) verbs of FEAR, SURPRISE and ANGER.
4.6
Summary
This chapter examines the semantic features of cause events according to their degree of transitivity. The transitivity of cause events is evaluated based on three semantic features, namely agentivity, kinesis, and participation in which the more positive the semantic features are, the more transitive the cause event is. In general, it is found that there are strong correlations between emotion verb classes (e.g., HAPPINESS vs. SADNESS) and their corresponding causes regarding transitivity. Among the emotion verb classes, verbs of HAPPINESS tend to be associated with cause events with the highest transitivity, while verbs of SADNESS are more likely to be triggered by cause events with the lowest transitivity. In addition, I note that emotion verb types (change-of-state verbs vs. homogeneous state verbs) determine the tendency of cause event features in a way such that homogeneous state verbs generally have a stronger tendency of cause event features than those of change-of-state verbs.
Chapter 5
Linguistic Expression of Cause Event II: Epistemicity
5.1
Introduction
After analyzing the cause event features of each of the 8 Chinese primary emotion verbs in Chap. 4, this chapter deals with the issue of how emotion causes are linguistically identified. I focus on what Chang et al. (2000) described as change-of-state emotion verbs since these verbs often have causes explicitly expressed in a complement attached to them. One would expect the cause to be expressed by the immediate subordinate clause as in (1). (1) ta1men hen3 gao1xing4 Wang2hui4zhen1 rong2huo4 jin1pai2 3.PL very becoming happy Wang Huizhen win gold medal ‘They are happy that Wang Huizhen won the gold medal.’
The emotion verb 高興 gao1xing4 ‘becoming happy’ in (1) takes the complement 王惠珍榮獲金牌 wang2hui4zhen1 rong2huo4 jin1pai2 ‘Wang Huizhen won the gold medal’ as its cause. The cause complement is not expressed by the presence of a complementizer as in the English equivalent ‘that Wang Huizhen won the gold medal’. However, surprisingly enough, I find that apart from being expressed by the immediate subordinate clauses, the cause events are often found to be embedded in clauses headed by some verbs in the complements, as in (2) and (3):
© Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_5
69
70
5 Linguistic Expression of Cause Event II: Epistemicity
(2) wo3 hen3 gao1xing4 kan4dao4 1.SG very becoming happy see ‘I am happy to see that they came back.’
(3)
ta1men 3.PL
wo3men hen3 gao1xing4 de2zhi1 jue2yi4 resolution 1.PL very becoming happy know ‘We are happy to know that the resolution was approved.’
hui2lai2 return
huo4de2 get
tong1guo4 approve
In (2), the cause of 高興 gao1xing4 ‘becoming happy’ is the event described by the embedded clause 他們回來 ta1men hui2lai2 ‘they came back’ which is preceded by a higher predicate 看到 kan4dao4 ‘to see’. Similarly, in (3), the immediate subordinate clause is not the cause event of HAPPINESS 決議獲得通過 jue2yi4 huo4de2 tong1guo4 ‘the resolution was approved’, but is headed by the verb 得知 de2zhi1 ‘to know’. The unexpected fact that emotion cause events are not directly subordinated but further embedded below clauses headed by certain verbs requires explanation. These verbs are what I call epistemic markers. Several questions arise with regard to epistemic marking: What are epistemic markers? What are the functions of epistemic markers? Do epistemic markers only appear in the constructions of change-of-state emotion verbs? These questions will be addressed in the subsequent sections of the chapter. I start with the discussion of epistemicity, specifically epistemic verbs and what role they play in emotion analysis in Sect. 5.2. Section 5.3 discusses some preliminary observations regarding emotion causes and epistemic verbs. To verify the hypothesis made in Sects. 5.3 and 5.4 demonstrates a corpus study of epistemic markup of cause events. Section 5.5 presents the discussion and Sect. 5.6 concludes the chapter.
5.2
Concepts of Epistemicity
The concept of epistemicity has a long history of discussion both in the philosophical and linguistic literature. There is still disagreement as to how the concept is to be explained; however, there is a general consensus that epistemicity refers to the speaker’s attitude toward the certainty of a proposition. Givón (2009: 315) defines epistemicity as ‘pertaining to how a person views the facts of the world, or how they view another person’s view of such facts’. In general, the concept of epistemicity is often centered on propositions containing epistemic verbs, such as know, see, and say, as in (4); epistemic adverbs, such as perhaps, undoubtedly, and supposedly, as in (5); and modal auxiliaries, such as could, may, might, as in (6).
5.2 Concepts of Epistemicity
71
(4) a. He knew that she had arrived. b. He saw that she had gone. c. He said that she was coming. (5) a. Perhaps he has left already. b. He has finished, undoubtedly. c. He went to London supposedly last week. (6) a. He could come in five minutes. b. He may have arrived already. c. He might be wrong. This chapter mainly deals with so-called epistemic verbs. Givón (1993) argued that epistemic verbs are usually perception-cognition-utterance (PCU) verbs. He described PCU verbs in terms of the semantic and syntactic relations between the main and complement clause as in (7): (7) Definition of PCU verbs (Givón 1993, II:4) a. The main clause codes mental or verbal activity, with a verb (or adjective) of perception, cognition, mental attitude or verbal utterance. b. The state or event coded in the complement clause is the object of the mental or verbal activity coded by the main verb. c. No coreference restrictions hold between arguments of the main clause and complement clause. Typical examples of PCU verbs that code the epistemic attitude of the speaker toward the complement clause are know, see, and say, as was shown in (4). The semantic and syntactic relations of PCU verbs defined in (7) can be illustrated by sentence (4a) and the corresponding tree diagram in (8): (8) Tree diagram for P-C-U verbs S
SUBJ [NP]
VP
V
perceiver cognizer utterer He
P-C-U verb knew
COMP [S]
perceived, cognized, uttered state/event that she had arrived
72
5 Linguistic Expression of Cause Event II: Epistemicity
Epistemic PCU verbs code different degrees of certainty: high epistemic certainty, low epistemic certainty, or negative epistemic certainty. PCU verbs of high epistemic certainty, such as know and remember are often characterized as presuppositional or factive (Givón 1993). For instance, in (4a) the speaker considers the proposition in the complement clause ‘that she had arrived’ to be true regardless of the truth value of the main-clause proposition. Other PCU verbs of high epistemic certainty include forget, see, hear, find out, discover, understand, perceive, etc. In both (9a) and (9b), the verbs presuppose the truth of the propositions, i.e. he left last Monday and no one was there. (9) a. She forgot that he left last Monday. (i.e. he left last Monday.) b. He discovered that no one was there. (i.e. no one was there.) Some epistemic PCU verbs code relatively lower epistemic certainty, such as think, guess, suspect, assume, feel, etc., as in (10): (10) a. He thinks that she will be going to the party. b. She assumed that he was there. When a PCU verb codes epistemic uncertainty the speaker does not presuppose the truth of the proposition in the complement clause. Consider the verbs hope and wonder in (11): (11) a. He hopes that she is in Hong Kong. b. He wonders if she is in Hong Kong. In (11a) and (11b), it remains an open question as to whether or not she is in Hong Kong. In other words, verbs of epistemic uncertainty are non-factive. However, they do not entail the fallacy of the proposition either. Some PCU verbs of epistemic attitude code negative certainty where the complement clause is often preceded by the subordinators ‘if’ or ‘whether’. These verbs include pretend and lie which are characterized as negative factives. That is, the speaker considers the proposition in the complement clause to be false. Consider (12) where the verb pretend presupposes that she had not left, whereas lie presupposes that she has not come.
5.2 Concepts of Epistemicity
73
(12) a. He pretended that she had left. (i.e. she had not left.) b. He lied that she had come. (i.e. she had not come.) To sum up, the epistemic verbs along the epistemic certainty continuum of complementation scale are summarized in (13): (13) Epistemic Certainty Continuum (Givón 1993) Epistemicity
Factivity
Epistemic verbs
Strong epistemic certainty Weak epistemic certainty Epistemic uncertainty Negative epistemic certainty
Factive Semi-factive Non-factive Negative factive
know, remember, forget, see think, assume, guess, suspect hope, wonder, doubt pretend, lie
5.3
Emotion Causes and Epistemic Markers
As briefly discussed in Sect. 5.1, based on the Sinica Corpus, Chang et al. (2000) find that change-of-state emotion verbs take cause events as sentential objects, while homogeneous state verbs cannot. They also find that in general only change-of-state verbs of HAPPINESS, FEAR, and WORRY take the cause events as sentential objects. This is demonstrated in (14) and (15) where the change-of-state verb of HAPPINESS 高興 gao1xing4 takes a sentential object, but the homogeneous state verb of HAPPINESS 快樂 kuai4le4 does not. (14) ta1 hen3 gao1xing4 ba4ba hui2lai2 3.SG.F very becoming happy father come back ‘She was very happy that her father had come back.’
le ASP
ta1 hen3 kuai4le4 ba4ba hui2lai2 3.SG.F very to be happy father come back ‘She was very happy that her father had come back.’
le ASP
(15)
Chang et al. (2000) also note that a few homogeneous state verbs, such as kung3ju4-FEAR and fan2nao3-WORRY do take sentential objects although only in rare cases. Since WORRY is not considered one of the primary emotions in this work, let us consider the verb of FEAR kung3ju4:
74
5 Linguistic Expression of Cause Event II: Epistemicity
(16) xu3duo1 zheng4ren2 yin1 kong3ju4 zao1 bao4fu4 er2 bu4 gan3 many witness because to be frightened encounter retaliation so not dare chu1ting2 appear in court ‘Many witnesses did not appear in court for fear of being retaliated against.’ (17) yin1wei4 kong3ju4 gong1ren2 you2xing2 yin3fa1 bao4li4, xu3duo1 shang1dian4 because to be frightened worker demonstration trigger violence, many store ye3 xuan1bu4 guan1men2 also announce close ‘For fear that the demonstration of workers would trigger violence, many stores also decided to close.’
As a homogeneous state verb of FEAR, 恐懼 kong3ju4 seems in general to allow sentential complementation as cause event in the form of a VP complement 遭報復 zao1bao4fu4 ‘being retaliated’ as in (16), or an event 工人遊行引發暴力 gong1ren2 you2xing2 yin3fa1 bao4li4 ‘fear that the demonstration of workers would trigger violence’, as in (17). A further observation is that the cause events of the emotion verbs do not appear in the immediate subordinate clause, but are often preceded by certain verbs, as in (18) and (19): (18) jia1ren2 jing1ya4 family becoming surprised
fa1xian4 ta1 jing4ran2 discover 3.SG.M unexpectedly
liu2xia4 300 duo1 wan4 leave more than 3 million
de yi2chan3 POSS estate ‘His family was surprised to find that he left an estate of more than three million dollars.’ (19) ta1 hen3 gao1xing4 jian4dao4 bei3jing1 zun1shou3 cheng2nuo4 Beijing keep promise 3.SG.M very becoming happy see ‘He was happy to see that Beijing kept its promise.’
In (18), the verb of SURPRISE 驚訝 jing1ya4 takes the cause event 他竟然留下 300 多萬的遺產 ta1 jing4ran2 liu2xia4 300 duo1wan4 de yi2chan3 ‘he left an estate of more than three million dollars’ which is headed by the verb 發現 fa1xian4 ‘discover’. What triggers the emotion is the event described by the embedded clause, i.e. the cause event rather than the action of discovery introduced by the higher predicate 發現 fa1xian4 ‘discover’. Similarly, in (19), the immediate subordinate verb 見到 jian4dao4 ‘see’ does not actually describe the cause event of HAPPINESS 北京遵守承諾 bei3jing1 zun1shou3 cheng2nuo4 ‘Beijing kept its promise’. The verbs heading the cause events are what I refer to as epistemic markers. To answer the questions raised in Sect. 5.1, by epistemic markers I mean verbs, such as discover and see, that encode the experiencer’s certainty of the assertion of
5.3 Emotion Causes and Epistemic Markers
75
certain emotional states. What I observe is when the immediate subordinate clause is headed by an epistemic marker, the direct cause is further embedded. In other words, the function of these epistemic markers is to create a transparent environment for emotion causal relations. I hypothesize that epistemic markers tend to appear in constructions of change-of-state emotion verbs rather than that of homogeneous state emotion verbs. If this hypothesis is correct, it begs another question: how can the causes of change-of-state emotion verbs be identified by means of epistemic marking?
5.4
A Corpus Study of Epistemic Marking of Emotion Causes
To verify the hypothesis that emotion causes of some change-of-state verbs, but not homogeneous state verbs, can be expressed by means of epistemic marking, I start with a preliminary quantitative study by examining the sentential complement collocations of each emotion verb. Data are taken from the Chinese Gigaword Corpus, and the statistics are generated using the Chinese Word Sketch.
5.4.1
Methodology and Data Observation
This study is based on data in the Chinese Gigaword Corpus (Gigaword Corpus). As mentioned in Chap. 2, the Gigaword Corpus contains a total of 1.4 billion characters, including written texts of news from Taiwan, China, and Singapore. Although the Gigaword Corpus lacks a variety of genres, there are two main reasons for taking it as my working corpus: (1) since Chang et al.’s (2000) work focuses on the Sinica Corpus, studying a different corpus would allow non-biased analysis as well as a comparison of the results; (2) a large sample of data is needed for any quantitative study; therefore, a gigantic corpus, such as the Gigaword Corpus is the most suitable corpus to make generalizations. Moreover, the primary emotion verbs appear to be rather infrequent in a corpus of a smaller size, such as the Sinica Corpus, which adds a further argument to using the Gigaword Corpus. The focus is, once again, on the five primary emotions, i.e. HAPPINESS, SADNESS, FEAR, ANGER, and SURPRISE, and the two types of emotion verbs in Chinese, i.e. change-of-state emotion verbs and homogeneous state emotion verbs. Similar to the study in Chap. 4, the most frequent verb in each group of the primary emotions in Chinese is analyzed for the convenience of discussion. The default contrast pair of each emotion class is repeated in Table 5.1. The Word Sketch, one of the core functions of the Chinese Word Sketch Engine (cf. Chap. 3), is a corpus-based summary of a word’s grammatical and collocational
76
5 Linguistic Expression of Cause Event II: Epistemicity
Table 5.1 List of primary emotion verbs for analysis Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Change-of-state
Homogeneous state
高興 傷心 害怕 生氣 驚訝
快樂 悲傷 恐懼 憤怒 震驚
gao1xing4 shang1xin1 hai4pa4 sheng1qi4 jing1ya4
kuai4le4 bei1shang1 kong3ju4 fen4nu4 zhen4jing1
Fig. 5.1 The search page of word sketch
behaviour. It allows one to get the collocations of the emotion verbs based on various similarity scores, including T-score, log likelihood, and salience. Let us see how it works by examining a lemma, e.g., 高興 gao1xing4 ‘becoming happy’. Figure 5.1 shows the search page of the Word Sketch function. After inputting the search word 高興 gao1xing4 in the word form box and selecting the related options, such as minimum frequency, minimum salience, etc., the system will take us to the Word Sketch results for 高興 gao1xing4, as in Fig. 5.2. First of all, Fig. 5.2 indicates that the total number of tokens for 高興 gao1xing4 occurring in the Gigaword Corpus is 41,043. Each column shows the words that typically collocate with 高興 gao1xing4 in particular grammatical relations. There are 12 significant collocational grammatical relations identified by the Word Sketch, each listed in order of their statistical significance (the numbers in black) rather than the raw frequency (the numbers in blue). For instance, the ‘subject’ list shows that 高興 gao1xing4 is most typically the subject of some pronouns 我 wo3
5.4 A Corpus Study of Epistemic Marking of Emotion Causes
Fig. 5.2 The word sketch for gao1xing4
77
78
5 Linguistic Expression of Cause Event II: Epistemicity
‘I’, 我們 wo3men ‘we’, and 他 ta1 ‘he/she/it’; various prepositional phrases headed by 同 tong2 ‘and’, 與 yu3 ‘and’, 在 zai4 ‘at’. Also, one can always click on the number next to the collcated word to look at the concordance. As seen in Fig. 5.2, 高興 gao1xing4 is likely to take sentential objects, i.e. there are 6267 tokens. It is also noted that the most salient collocates as the sentential objects of 高興 gao1xing4 are 看到 kan4dao4 ‘to see’, 有 you3 ‘to exist/have’, 見到 jian4dao4 ‘to see’, and 聽到 ting1dao4 ‘to hear’, which yielded 771, 835, 229 and 113 tokens, respectively. Consider Example (20): (20) emotion keyword epistemic marker wo3 hen3 gao1xing4 kan4dao4 1.SG very becoming happy to see 3.PL ‘I was very happy to see that they came back.’
cause event ta1men hui2lai2 come back
In (20), the emotion keyword 高興 gao1xing4 takes a sentential object which is supposedly a cause event. However, instead of occupying the immediate subordinate clause, the cause event 他們回來 ta1men hui2lai2 ‘they came back’ is further embedded in the clause headed by an epistemic marker 看到 kan4dao4 ‘to see’. There are cases wherethe seemingly epistemic verbs do not function as epistemic markers. For instance, in the case of the seeing verb 看到 kan4dao4 ‘to see’ where it can be part of a cause event rather than the epistemic verb that marks the cause event exemplified in (21) and (22): (21) a. wo3 hen3 gao1xing4 1.SG very becoming happy ‘I was very happy to see them.’
kan4dao4 to see
ta1men 3.PL
b. ta1 hen3 gao1xing4 kan4dao4 xu3duo1 ren2 re4xin1 yu1 zheng4zhi4 3.SG.M very becoming happy to see many people enthusiastic about politics ‘He was very happy that many people are enthusiastic about politics.’ (22) a ta1 hen3 gao1xing4 ting1dao4 3.SG.M very becoming happy to listen ‘He was very happy to listen to this song.’
zhe4 shou3 DET CL
ge1 song
b. ta1 hen3 gao1xing4 ting1dao4 Wu2hua2 huo4pan4 wu2zui4 3.SG.M very becoming happy to hear WuHua sentence innocent ‘He was very happy to hear that Wu-hua was acquitted.’
In (21a), the two events ‘to be happy’ and ‘to see’, are co-temporal. Such co-temporality is indeed a prerequisite for construing two events as an integrated single event (Givón 1993II:14). This means what triggers the presence of HAPPINESS
5.4 A Corpus Study of Epistemic Marking of Emotion Causes
79
is the seeing event, thus, the event 看到他們 kan4dao4 ta1men ‘to see them’ is the cause event of HAPPINESS. In (21b), on the other hand, the two events are not co-temporal. It is the perception that ‘many people are enthusiastic about politics’ which causes the HAPPINESS, rather than the physical seeing event. In such case, the verb 看到 kan4dao4 ‘to see’ in (21a) serves as the verb in the cause event, while the one in (22b) serves as an epistemic verb that marks the cause event. This analysis also applies in (22) where the cause event of (22a) is 聽到這首歌 ting1dao4 zhe4 shou3 ge1 ‘to listen to this song’, whereas the one of (22b) is 吳華獲判無罪 wu2hua2 huo4pan4 wu2zui4 ‘Wuhua was acquitted’ which is headed by the epistemic marker 聽到 ting1dao4 ‘to hear’. In addition, the two events in both (21b) and (22b) are not physical perception verbs anymore, but rather verbs of mental reflection. According to Givón (1993, II:14), “when one perceives an event, the perception is co-temporal with the perceived event”. On the other hand, ‘when one reflects upon an event, all manner of temporal gapping may exist between reflection and the reflected-upon event’. In light of this, it is possible that there is no seeing or hearing involved in (21b) and (22b), but it reflects the cognitive awareness of the event. In this study, I filter out sentences without epistemic verbs, such as those in (21a) and (22a). Furthermore, some sentences involve negation which may result in neutral sentences as in (23), or sentences that do not reflect the meaning of the emotion verbs, as in (24): (23) ta1 mei2you3bu4 gao1xing4 kan4dao4 wo3men lai2dao4 da4lu4 3.SG.M no not becoming happy to see 1.PL come the mainland ‘He was not sad (or angry) to see that we came to the mainland.’ (24) ta1 bu4 gao1xing4 kan4dao4 wo3men 3.SG.M not becoming happy to see 1.PL ‘He was sad (or angry) to see that we came to the mainland.’
lai2dao4 da4lu4 come the mainland
Sentences (23) and (24) need to be filtered out before the analysis is conducted. On the other hand, it is found that some cause events headed by an epistemic marker are missing in the CWS results. Therefore, I manually select the relevant epistemic markers from the list of sentential objects in the CWS results, such as 看 到 kan4dao4 ‘to see’ and 有 you3 ‘to exist’ (as marked in Fig. 5.3), and analyze all concordances of the emotional keywords and the epistemic markers when they occur.
80
5 Linguistic Expression of Cause Event II: Epistemicity
Fig. 5.3 Relevant epistemic markers of gao1xing4 in word sketch
Table 5.2 Types of epistemic markers
Categories SEEING
HEARING KNOWING
DISCOVERY EXISTENCE
5.4.2
Epistemic markers 看 kan4, 見 jian4, 看見 kan4jian4, 看到 kan4dao4, 見到 jian4dao4 聽 ting1, 聽到 ting1dao4, 聽說 ting1shuo1 知道 zhi1dao4, 得知 de2zhi1, 得悉 de2xi1, 獲知 huo4zhi1, 獲悉 huo4xi1 發現 fa1xian4 有 you3
Results
Based on the CWS results, I group the epistemic markers into five categories, as shown in Table 5.2. The appearing verbs are of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE. There are five verbs, i.e. 看 kan4, 見 jian4, 看見 kan4jian4, 看到 kan4dao4, and 見到 jian4dao4, functioning as epistemic markers for the SEEING type, all of them meaning ‘to see’. For verbs of HEARING there are three epistemic markers, i.e. 聽 ting1, 聽到 ting1dao4, and 聽說 ting1shuo1, meaning ‘to hear’ or ‘to hear of’. For the type of KNOWING verbs, five verbs function as epistemic markers, i.e. 知道 zhi1dao4, 得知 de2zhi1, 得悉 de2xi1, 獲知 huo4zhi1, and 獲悉 huo4xi1, which either mean ‘to know’ or ‘to learn about’.
5.4 A Corpus Study of Epistemic Marking of Emotion Causes
81
There is only one marker for the verb of DISCOVERY and EXISTENCE respectively, 發現 fa1xian4 meaning ‘to discover/find’ and 有 you3 ‘to exist/have’. Some examples of these epistemic markers are given in Table 5.3. Table 5.3 Examples of epistemic markers for gao1xing4 Categories of Epistemic Markers
Examples (i) ta1 hen3 gao1xing4 kan4dao4 zhong1guo2da4lu4 China 3.SG.M very becoming happy see zheng4ju2 wen3ding4 political situation stable ‘He was very happy to see political stability in China.’
SEEING
(ii) wo3 hen3 gao1xing4 ting1dao4 hen3duo1 guan1yu2 many about 1.SG very becoming happy hear jing3cha2 xing2dong4 de fan3ying4 shou4dao4 ren2min2 de zhu4yi4 police action POSS reaction receive people POSS attention ‘I was glad to hear that a lot about the reaction of the police action drew people's attention.’
HEARING
(iii) ta1 hen3 gao1xing4 de2zhi1 xiang1gang3 yu3 Hong Kong and 3.SG.M very becoming happy know tai2wan1 jian1 de lü3you2 neng2 you3 jin4zhan3 Taiwan between POSS travel can have progress ‘He was happy to know that the tourist industry between Hong Kong and Taiwan had made some progress.’
KNOWING
(iv) ta1 hen3 gao1xing4 fa1xian4 nü3xing4 ge1mi2 bi3 3.SG.F very becoming happy discover female fans compare yi3qian2 zeng1jia1 hen3duo1 before increase very much ‘She was very happy to find that there are more female fans now than before.’
DISCOVERY
(v)
EXISTENCE
ta1 hen3 gao1xing4 you3 hen3duo1 dui4hua4 3.SG.M very becoming happy have many conversation yi3 chu1xian4 bing4 luo4shi2 already emerge and implement ‘He is very happy that a lot of dialogue already emerged and had been implemented.’
82
5 Linguistic Expression of Cause Event II: Epistemicity
Results show that some change-of-state emotion verbs tend to allow epistemic marking of cause events in the complement clause; whereas homogeneous state emotion verbs rarely mark the cause events by epistemic verbs, as shown in Tables 5.4 and 5.5. In general, all the five primary change-of-state emotion verbs mark the cause events by means of epistemic markup; however, only a few cases of homogeneous emotion verbs of ANGER (0.03%) and SURPRISE (0.08%) allow epistemic verbs of KNOWING to mark the cause events. Among the five change-of-state emotion verbs, epistemic markup of cause events is the most active with the verbs of HAPPINESS (8.5%) making use of all five types of epistemic markers. The cause events of SURPRISE are also usually marked by epistemic verbs (3.5%) which are mostly verbs of DISCOVERY. For the verbs of FEAR, the epistemic marking is relatively lower than that of HAPPINESS and SURPRISE. FEAR’s most highly collocated epistemic verb is the verb of EXISTENCE. Verbs of ANGER (0.09%) and SADNESS (0.07%) in general do not usually mark cause events by epistemic verbs, as compared to other change-of-state emotion verbs. In fact, patterns suggest an association with some homogeneous state emotion verbs, such as ANGER (0.03%) and SURPRISE (0.08%).
5.5
Results and Discussion
This corpus-based study implies that two change-of-state emotion verbs, i.e. 高興 gao1xing4 ‘becoming happy’ and 驚訝 jing1ya4 ‘becoming surprised’, tend to mark cause events explicitly by means of epistemic verbs, while homogeneous state emotion verbs do not. This section looks at the motivation behind the favourable use of explicit epistemic markup of cause events that involve change-of-state emotion verbs, and also how they are used differently in the five primary emotion classes. Assuming that the cut-off point of the normalized total is 1%, based on the corpus results shown in Table 5.4, the cause events of HAPPINESS can be explicitly expressed in the complement introduced by the five types of epistemic markers (8.5%). The types of epistemic markers include verbs of sensory perception, i.e. SEEING and HEARING, and verbs of mental perception, i.e. KNOWING, DISCOVERY, and EXISTENCE. Yet, most of the sensory perception verbs no longer indicate the subject’s sensory perception of the cause event, but rather the mental reflection of the cause event, as shown in (21) and (22). Therefore, epistemic markers are verbs that indicate the experiencer’s cognitive awareness of the emotion causes, i.e. the perceived events. In other words, the higher motivation the experiencer has to assert the certainty of the emotion, the more explicit the epistemic marking of cause events. In the case of HAPPINESS, people are more eager to assert its presence as this is the positive emotion they want to pursue (Izard 1977). Another emotion that allows explicit epistemic marking is the change-of-state verb of SURPRISE. As shown in Table 5.4, 3.8% of the total tokens of SURPRISE verbs are marked with epistemic markers with verbs of DISCOVERY dominating (219 out of
5.5 Results and Discussion
83
Table 5.4 Epistemic markers of change-of-state emotion verbs Categories
高興 gao1xing4 ‘happiness’
傷心 shang1xin1 ‘sadness’
害怕 hai4pa4 ‘fear’
生氣 sheng1qi4 ‘anger’
驚訝 jing1ya4 ‘surprise’
SEEING
2462 11 1 16 2207 227 123 0 119 4 137 42 46 8 4 37 16 16 755 755 3493
2 0 0 0 2 0 0 0 0 0 2 0 1 0 1 0 0 0 1 1 5
8 1 0 0 6 1 4 0 4 0 0 0 0 0 0 0 1 1 43 43 56
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 1 1 3
21 1 0 1 19 0 2 0 2 0 3 1 2 0 0 0 219 219 17 17 262
41,043 8.5%
5967 0.07%
7989 0.7%
3315 0.09%
6892 3.8%
看 kan4 見 jian4 看見 kan4jian4 看到 kan4dao4 見到 jian4dao4 HEARING
聽 ting1 聽到 ting1dao4 聽說 ting1shuo1 KNOWING
知道 得知 得悉 獲知 獲悉
zhi1dao4 de2zhi1 de2xi1 huo4zhi1 huo4xi1
DISCOVERY
發現 fa1xian4 EXISTENCE
有 you3 Total # of epistemic markers Total # of tokens Normalized total
262 tokens). One will not feel surprised until they are aware of the cause event, or discover that something unexpected has happened. Therefore, experiencers of SURPRISE often assert the certainty of the emotional state by expressing what it was that made them surprised. FEAR, being a marginal case of epistemic markup (0.7%), allows epistemic marking mostly by means of verbs of HEARING and EXISTENCE. FEAR is an emotion of anticipated danger or unpleasant events which experiencers attempt to avoid (Oatley and Jenkins 1996). To escape from danger or unpleasant events, experiencers need to know to some degree the certainty of the cause event. Therefore, it is reasonable that verbs of FEAR are allowed to take some epistemic verbs. The other two emotion verbs, ANGER and SADNESS, do not usually allow epistemic marking in the complement clause. The normalized totals of epistemic marking are 0.09% and 0.07%, respectively. It is not surprising that cause events of ANGER and SADNESS only allow very limited epistemic marking since both of them have lower motivation to express the certainty of the emotion. As mentioned in Chap. 4,
84
5 Linguistic Expression of Cause Event II: Epistemicity
Table 5.5 Epistemic markers of homogeneous state emotion verbs Categories
快樂 kuai4le4 ‘happiness’
悲傷 bei1shang1 ‘sadness’
恐懼 kong3ju4 ‘fear’
憤怒 fen4nu4 ‘anger’
震驚 zhen4jing4 ‘surprise’
SEEING
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 1 1 1 1 0 0 3
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 10 0 0 10
18,479 0
2390 0
8463 0
9976 0.03%
12,494 0.08%
看 kan4 見 jian4 看見 kan4jian4 看到 kan4dao4 見到 jian4dao4 HEARING
聽 ting1 聽到 ting1dao4 聽見 ting1jian4 聽說 ting1shuo1 KNOWING
知道 得知 得悉 獲知 獲悉
zhi1dao4 de2zhi1 de2xi1 huo4zhi1 huo4xi1
DISCOVERY
發現 fa1xian4 EXISTENCE
有 you3 Total # of epistemic markers Total # of tokens Normalized total
SADNESS can often be triggered by obscure conditions, such as a rainy day or a loss of childhood (Power 1999), or even unknown reasons (Johnson-Laird and Oatley 1989). Knowing that whatever made one sad cannot be changed, experiencers of SADNESS are not motivated to affirm the certainty of what causes them to be sad. As for ANGER, quite a few studies (Oatley and Johnson-Laird 1996; Berkowitz 1999) have noted that people are not fully aware of why or how negative stimuli are affecting them before anger arises. Thus, when reacting to some negative stimuli that are outside of awareness, people may develop angry feelings before they have anyone to blame. In this case, experiencers of ANGER have less motivation to express the certainty of their state, but they are more aware of the aggression-associated reactions, such as hitting someone.
5.5 Results and Discussion more HAPPINESS
>
85 epistemic marking > FEAR
SURPRISE
less >
ANGER
>
SADNESS
Fig. 5.4 Epistemic marking continuum
The degree of epistemic marking among different emotion classes can be put on a scale, shown in Fig. 5.4. Verbs of HAPPINESS favour explicit epistemic markup of cause events the most, and is followed by SURPRISE, FEAR, ANGER, and SADNESS in that order.
5.6
Summary
This chapter examines how cause events of change-of-state emotion verbs are identified by means of epistemic marking. Epistemic markers are verbs that mark the experiencer’s cognitive awareness of the cause event. I identify five types of epistemic markers heading the cause events of change-of-state emotion verbs in the complement. The five types of epistemic markers are verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE. They provide a transparent environment for cause events by marking the experiencer’s cognitive awareness of the event. Among the five primary emotions, HAPPINESS tends to allow the most explicit epistemic markup of causes; this is followed by SURPRISE and FEAR. ANGER and SADNESS only allow very limited epistemic marking. This study argues that there is a strong relation between epistemic marking and one’s motivation to assert the certainty of the state of emotion. The higher motivation the experiencer has to assert the certainty of the emotion, the more explicit epistemic marking of cause events is allowed. Epistemic markers tend to mark the cause events of emotions when experiencers are more motivated to express the certainty of the cause events, e.g., HAPPINESS and SURPRISE. Emotions that are associated with obscure or unknown events do not usually allow epistemic marking, e.g., SADNESS.
Chapter 6
A Linguistic Model for Emotion Expression
6.1
Introduction
The lack of a comprehensive linguistic theory of emotion is surprising given the rich literature in neighbouring fields, such as cognitive psychology and computational linguistics. The current Natural Semantic Metalanguage (NSM) theory offers a comprehensive language-independent account of emotions based on a list of universal semantic primitives. It can represent emotions through three of the proposed semantic primitives: the mental predicate FEEL, and the evaluators GOOD and BAD. However, such descriptions do not go far beyond the positive and negative dichotomy commonly adopted in computational sentiment analysis. What is important to know is what the good and bad things are and how these are linguistically expressed in a language. In addition, such descriptions are rather abstract for empirical research as well as computational implementation. This chapter deals with the enrichments of the NSM model by discussing in what way the introduction of cause event features and epistemic markers discussed in Chaps. 4 and 5 can be incorporated into the current NSM emotion model. I attempt to make explicit links between emotions and cause events. I believe that the incorporation of these linguistic enrichments into the NSM model would further our understanding of emotions as events. It would also strengthen the explanatory and predicting power of the NSM theory in defining emotions.
6.2
NSM Reconsidered
As mentioned in Chap. 1, NSM provides a semantic-primitive approach to defining emotion concepts. Wierzbicka (1992) proposed a list of primitives, such as GOOD, BAD, FEEL, and WANT, which can describe some of the basic themes that characterize emotions. For example, emotions are described as something that an experiencer © Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_6
87
88
6 A Linguistic Model for Emotion Expression
feels and thinks, the experiencer or other people do, as well as something that involves good or bad things that happen to the experiencer or other people. Although a restricted set of semantic primitives is used to describe these basic themes, they seem to capture some subtle distinctions between different emotions. They are also able to define the most basic ones, i.e. the primary emotions. Following the English version (Wierzbicka 1996), the five primary emotions in Chinese can be represented as (1)–(5): (1) 喜 xi3 (X feels happy) (a) X feels something because X thinks something (b) Sometimes a person thinks something like this: (c) “Something good happened (d) I wanted this to happen” (e) Because of this, this person feels something good (f) X feels something like this (2) 悲 bei1 (X feels sad) (a) X feels something because X thinks something (b) Sometimes a person thinks something like this: (c) “Something bad happened now (d) I know that after this, good things will not happen any more (e) I don’t want things like this to happen (f) I cannot do anything (g) Because I know no one can do anything when things like this happen” (h) Because of this, this person feels something very bad (i) X felt something like this (3) 恐 kong3 (X feels frightened) (a) X feels something because X thinks something (b) Sometimes a person thinks something like this: (c) “I don’t know what will happen (d) Some bad things can happen (e) I don’t want these things to happen” (f) Because of this, this person feels something bad (g) X feels something like this (4) 怒 (a) (b) (c) (d) (e) (f) (g) (h)
nu4 (X feels angry) X feels something because X thinks something Sometimes a person thinks something like this: “This person did something bad I don’t want this Because of this, I want to do something I would want to do something bad to this person” Because of this, this person feels something bad X feels something like this
6.2 NSM Reconsidered
89
(5) 驚 jing1 (X feels surprised) (a) X feels something because X thinks something (b) Sometimes a person thinks something like this: (c) “Something happened now (d) I didn’t think before now: this will happen (e) If I thought about this, I would have said: this will not happen” (f) Because of this, this person feels something (g) X feels like this Different from classical definitions, Wierzbicka (1992) calls descriptions such as (1)– (5) “semantic explications”. A main theme of the semantic explications involves good and bad things that happen to some people or good and bad things that people do. As we can see from (1) to (5), the emotion of HAPPINESS is triggered by the happening of good things, whereas SADNESS and ANGER are triggered by the happening of bad things. FEAR, on the other hand, is triggered by the potentially bad thing that may happen. Unlike the other primary emotions, SURPRISE is triggered by undefined events which can be good, bad, or neutral. The definition of emotion concepts proposed by Wierzbicka (1992: 552) “embodies a hypothesis about a language-specific psychological script, unconsciously used by speakers of a given language in interpreting their own and other people’s emotional experience”. In other words, the definitions are expected to summarize our understanding of emotion concepts and the underlying psychological realities. While NSM sounds appealing in discriminating different emotion concepts based on a limited set of universal semantic primitives, there are some shortcomings. Although the descriptions of each emotion seem to contain information about the situations for emotions, the description does not contain clear semantic content other than its being good or bad. For example, what are the characteristics of these good things and bad things? Is the bad thing involved in SADNESS the same as that in ANGER? The current proposal aims to enrich the NSM theory in characterizing emotions based on linguistic cues in context. This will be done by incorporating the linguistic findings in Chaps. 4 and 5 so that the distinctions among emotions are supported by linguistic realizations as to how the emotion-evoking good and bad things are semantically represented and syntactically structured.
6.3 6.3.1
Linguistic Integration Emotion as Events
As suggested in NSM, there is a close relationship between emotions and events. However, the linkage between the two is not clearly described. In the literature, little research focuses on the interaction between emotions and events. Among the few previous studies, Ortony et al. (1988), viewed emotion from the psychological perspective as a paradigmatic psychological state of feeling that arises from attending to events which are appraised as being desirable or undesirable. The
90
6 A Linguistic Model for Emotion Expression
desirability of an event comprises the eliciting conditions for the emotion type, for instance, the difference between HAPPINESS and SADNESS. Similar thoughts have also been presented by Frijda (1987) in the sense that the implicated emotion of an event is denoted by its valence. Despite the little research on the interaction of event-denoting expressions and emotion, I argue that emotion is one of the most important factors involving events. According to the Generative Lexicon (GL, Pustejovsky 1995), one level of semantic description involves an event-based interpretation of a word or phrase. In other words, the event structure of a word is one level of the semantic specification of a lexical item. There are three components in this structure: the primitive event type, the focus of the event, and the rules for the event composition. For the primitive event type, one can be a state, a process, or a transition. Pustejovsky (1995) also added sub-event structures to these events. Within an event, the relation between the event and its sub-events should be represented. An event structure with structured sub-events is what Pustejovsky calls the “extended event structure”. An event with the relation of “exhaustive ordered part of”,
SADNESS
Fig. 6.2 The transitivity continuum
more HAPPINESS
>
SURPRISE
epistemic marking > FEAR >
less ANGER
>
SADNESS
Fig. 6.3 Epistemic marking continuum
HAPPINESS
> homogeneous FEAR
high HAPPINESS
>
> CoS-FEAR, SURPRISE, ANGER transitivity epistemic marking SURPRISE > FEAR > ANGER >
Fig. 6.4 The comparison of the transitivity and epistemic marking continuums
>
SADNESS
low SADNESS
6.3 Linguistic Integration
93
cause event. This idea is influenced by the insight of Wierzbicka’s (1996) good/bad dichotomy in defining cause events. According to Wierzbicka (1996), HAPPINESS is often grown out of desirable cause events. By desirable events, I mean it is desirable to the experiencer, but not desirable in a general sense. On the one hand, epistemic marking is favourable in the case of desirable cause events, as experiencers are strongly motivated to recognize the causes which are usually perceived events. On the other hand, HAPPINESS has the most potential to be linked to more-affected events, as cause events of HAPPINESS tend to be perceived events associated with strong kinesis which can be carried over to the resulted event, i.e. the experiencer is more affected. SADNESS, on the contrary, is more often triggered by undesirable events which cannot be reinstated. These undesirable events make the experiencer feel something very bad. Therefore, causes of SADNESS are often less kinetic and tend to be associated with less-affected events. It allows limited epistemic marking as experiencers can be self-motivated to be sad and mostly do not require the confirmation of causes of their emotion. Similar to SADNESS, ANGER is also triggered by undesirable events, but differs from SADNESS in the sense that some events of ANGER can be reinstated (e.g., through apologies, etc.). According to Wierzbicka (1996), the experiencer of ANGER feels something bad because of the undesirable events. Therefore, the cause events of SADNESS are in general less desirable than those of ANGER. As for FEAR, it tends to be triggered by the potentially undesirable thing that may or may not happen in the future. This undesirable event is often imaginary which may actually protect the experiencer from further unpleasantness (Oatley and Jenkins 1996). Hence, the causes of FEAR are less undesirable than those of SADNESS and ANGER. SURPRISE differs from other primary emotions in that it is often associated with undefined cause events which can be good, bad, or neutral. The cause of SURPRISE is thus uncertain in terms of desirability. To sum up, the desirability of the causes of the five primary emotions is basically in line with the transitivity and epistemic marking continuums. However, further work is needed to explore the feasibility of this proposal.
6.4
Emotion Representation Model
NSM semantic explications mainly define emotion concepts by describing the scenario in which good and bad cause events happen. They are, however, not clear enough as to how cause events are linguistically expressed. Based on the linguistic cues in context, I attempt to enrich the NSM emotion descriptions by integrating the linguistic phenomena associated with emotion-cause interaction. Some of the NSM-defined scenarios for certain emotions may not be applicable in the Chinese context. For instance, experiencers may not have had the thought process that they “wanted this to happen” before they feel happy, or that they “know that after this,
94
6 A Linguistic Model for Emotion Expression
good things will not happen anymore” before they feel sad, or that they “want to do something bad to the person who did a bad thing” before they feel angry. Therefore, to emphasize the crucial role of cause events in emotion conception, the NSM semantic explication of HAPPINESS in Chinese, for example, can be simplified such as in (10) and (11) where (10) represents the change-of-state of HAPPINESS and (11) represents the homogeneous state of HAPPINESS: (10) Change-of-State of HAPPINESS: 高興 gao1xing4 ‘becoming happy’ (X feels happy) (a) X thinks “something good happened (e1)” (b) Because of this (c) X feels good (e2) Linguistic properties of e1 of 高興 gao1xing4: (i) Semantic properties: ["transitivity: "agentivity, "motion, "participation]1 (ii) Syntactic properties: ["epistemic markup] by verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE (11) Homogeneous State of HAPPINESS: 快樂 kuai4le4 ‘to be happy’ (X feels happy) (a) X thinks “something good happened (e1)” (b) Because of this (c) X feels good (e2) Linguistic properties of e1 of 快樂 kuai4le4: (i) Semantic properties: ["transitivity: "agentivity, "motion, "participation] (ii) Syntactic properties: [-epistemic markup] The definitions of (10) and (11) clearly state the clausal relation between two events, e1 and e2, as well as the linguistic properties of the cause event e1 in Chinese. The events involved in an emotional construction are temporally ordered, except for complex emotions (which will be discussed later). For the change-of-state verb of HAPPINESS 高興 gao1xing4, its e1 tends to be semantically realized as high transitivity with features of high agentivity, motion, and event participation, i.e. ["transitivity: "agentivity, "motion, "participation]. In terms of syntactic realization, it tends to be explicitly marked by epistemic verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE in the complement, i.e. ["epistemic markup]. What distinguishes 高興 gao1xing4 from the homogeneous state of HAPPINESS 快樂 kuai4le4 is the linguistic properties of their cause events in that 快 樂 kuai4le4 does not allow epistemic markup of cause events, i.e. [-epistemic markup]. The semantic explications of the other four primary emotion classes, each with both change-of-state and homogeneous state emotion verbs, are represented as follows:
" = high; # = low; "# = mid; – = NIL.
1
6.4 Emotion Representation Model
95
(12) Change-of-State of SADNESS: 傷心 shang1xin1 ‘becoming sad’ (X feels sad) (a) X thinks “something bad happened (e1)” (b) Because of this (c) X feels very bad (e2) Linguistic Properties of e1 of 傷心 shang1xin1: (i) Semantic properties: [#transitivity: "agentivity, #motion, #participation] (ii) Syntactic properties: [#epistemic markup] (13) Homogeneous State of SADNESS: 悲傷 bei1shang1 ‘to be sad’ (X feels sad) (a) X thinks “something bad happened (e1)” (b) Because of this (c) X feels very bad (e2) Linguistic Properties of e1 of 悲傷 bei1shang1: (i) Semantic properties: [#transitivity: "agentivity, #motion, #participation] (ii) Syntactic properties: [-epistemic markup] In (12) and (13), the change-of-state verbs and homogeneous state verbs of i.e. 傷心 shang1xin1 and 悲傷 bei1shang1, share similar semantic and syntactic features in terms of their cause events (e1). Their e1 tends to be semantically realized as low transitivity with features of low agentivity, motionless events, and self-participation, i.e. [#transitivity: "agentivity, #motion, #participation]. As for its syntactic realization, e1 does not favour the epistemic marking in the complement resulting in the change-of-state verb of SADNESS allowing very limited epistemic marking, i.e. [#epistemic markup], while the homogeneous state verb of SADNESS does not allow epistemic marking at all, i.e. [-epistemic markup]. Examples (14) and (15) demonstrate the emotion representations of the change-of-state verb and homogeneous state verb of FEAR, i.e. 害怕 hai4pa4 and 恐 懼 kong3ju4, respectively. The e1 of the homogeneous state verb of FEAR shows relatively clearer semantic and syntactic properties in that it tends to be semantically realized as high transitivity ["transitivity] and does not allow epistemic marking [epistemic markup]. As for the change-of-state verb of FEAR, its e1 does not show a clear pattern of semantic properties, i.e. ["#transitivity]. In other words, it is uncertain how transitive it is. In addition, the e1 allows rather limited epistemic marking, i.e. [#epistemic markup] through the use of verbs of SEEING, HEARING, DISCOVERY, and EXISTENCE. SADNESS,
(14) Change-of-State of FEAR: 害怕 hai4pa4 ‘becoming frightened’ (X feels frightened) (a) X thinks “something bad happened (e1) or may happen (e1*)”
96
6 A Linguistic Model for Emotion Expression
(b) Because of this (c) X feels bad (e2) Linguistic Properties of e1 of 害怕 hai4pa4: (i) Semantic properties: ["#transitivity: "agentivity, #motion, "participation] (ii) Syntactic properties: [#epistemic markup] by verbs of SEEING, HEARING, DISCOVERY, and EXISTENCE (15) Homogeneous State of FEAR: 恐懼 kong3ju4 ‘to be frightened’ (X feels frightened) (a) X thinks “something bad happened (e1) or may happen (e1*)” (b) Because of this (c) X feels bad (e2) Linguistic Properties of e1 of 恐懼 kong3ju4: (i) Semantic properties: ["transitivity: "agentivity, "motion, "participation] (ii) Syntactic properties: [-epistemic markup] For the e1 of ANGER, as we can see in (16) and (17), both the change-of-state and homogeneous state verbs tend to be uncertain in terms of transitivity. They are semantically more associated with features of high agentivity, motion event, and less self-participation, i.e. ["#transitivity: "agentivity, "motion, #participation]. (16) Change-of-State of ANGER: 生氣 sheng1qi4 ‘becoming angry’ (X feels angry) (a) X thinks “something bad happened (e1)” (b) Because of this (c) X feels bad (e2) Linguistic properties of e1 of 生氣 sheng1qi4: (i) Semantic properties: ["#transitivity: "agentivity, "motion, #participation] (ii) Syntactic properties: [#epistemic markup] by verbs of DISCOVERY and EXISTENCE
(17) Homogeneous State of ANGER: 憤怒 fen4nu4 ‘to be angry’ (X feels angry) (a) X thinks “something bad happened (e1)” (b) Because of this (c) X feels bad (e2) Linguistic properties of e1 of 憤怒 fen4nu4: (i) Semantic properties: ["#transitivity: "agentivity, "motion, #participation] (ii) Syntactic properties: [-epistemic markup]
6.4 Emotion Representation Model
97
What makes the change-of-state verb differ from the homogeneous state verb of is the syntactic properties of cause events in that the change-of-state verb allows limited epistemic markup of cause events, i.e. [#epistemic markup], whereas the homogeneous state verb does not allow epistemic markup at all. Parallels between the change-of-state and homogeneous state verbs are also found in the case of SURPRISE, as shown in (18) and (19). ANGER
(18) Change-of-State of SURPRISE: 驚訝 jing1ya4 ‘becoming surprised’ (X feels surprised) (a) (b) (c) (g)
X thinks “something happened now (e1), I didn’t think this would happen” Because of this X feels like this (e2)
Linguistic properties of e1 of 驚訝 jing1ya4: (i) Semantic properties: ["#transitivity: "agentivity, "motion, #participation] (ii) Syntactic properties: ["epistemic markup] by verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE (19) Homogeneous State of SURPRISE: 震驚 zhen4jing1 ‘to be surprised’ (X feels surprised) (a) (b) (c) (g)
X thinks “something happened now (e1), I didn’t think this would happen” Because of this X feels like this (e2)
Linguistic properties of e1 of 震驚 zhen4jing1: (i) Semantic properties: ["#transitivity: "agentivity, "motion, #participation] (ii) Syntactic properties: [-epistemic markup] As Turner (2000) argues, complex emotions are combinations of primary emotions. For instance, the complex emotion DELIGHT is the combination of a greater amount of SURPRISE and a lesser amount of HAPPINESS. REGRET is the combination of a greater amount of SADNESS and a lesser amount of FEAR. Such analysis of complex emotions; however, is inadequate in that it fails to specify how two emotions are combined, and has neither falsifiable, nor verifiable consequences. On the other hand, Wierzbicka (1992) states that: … a definition (or explication) is a hypothesis about the meaning of a word. It is arrived at by examining the range of a word’s use, and it is verified by checking whether it can account for that range. For example, if we establish that two words have an overlapping range of use, overlapping definitions are needed, which would account for both the similarity and the difference (Wierzbicka 1992: 551)
98
6 A Linguistic Model for Emotion Expression
Following this line of thinking, I propose that semantic explications of primary emotions can be combined to form complex emotions with the simplified NSM representations. For instance, DELIGHT and REGRET can be described as in (20) and (21), respectively. (20) 驚喜 jing1xi3 (X feels delighted) (a) (b) (c) (g)
X thinks “something good (e1) happened now, I didn’t think this would happen” Because of this X feels good (e2)
(21) 後悔 hou4hui3 (X feels regret) (a) (b) (c) (d) (e) For
X did something (e1) Because of (e1), something bad (e2) happened Because of (e2), something bad (e3) may happen Because of this X feels bad (e4)
DELIGHT,
the frame of the explication basically corresponds to that of except that the undefined cause event is now described as good to indicate the HAPPINESS that it brings about. REGRET involves more events: The fact that X did something (e1) causes something bad (e2) which makes X sad; while e2 may lead to something bad in the future causing a fine amount of FEAR. The new approach of defining complex emotions based on primary emotion semantic explications allows one to have a better understanding of complex events. I show that the proposed emotion representations based on NSM’s emotion semantic explications are applicable to the analysis of complex events. Further analysis on the linguistic properties of complex emotions could involve mapping onto events in the explications. Another possibility for further research could be the combining of the proposed emotion representation and GL’s event structures. Consider the following example. The emotion construction in (22) can be represented as a simplified semantic explication as in (23). A preliminary emotion-event representation of (22) is shown in (24).
SURPRISE,
(23) 傷心 shang1xin1 (X feels sad) (a) X thinks “something bad happened (e1)”
6.4 Emotion Representation Model
99
(b) Because of this (c) X feels bad (e2) (24) Emotion-Event Representation
The left hand side of (24) shows a GL lexical representation of the verb 傷心 shang1xin1 ‘becoming sad’. The lexical representation includes three structures, namely event structure (EVENTSTR), argument structure (ARGSTR), and qualia structure (QUALIA). The event structure indicates that the emotion verb 傷心 shang1xin1 is associated with two events, e1 and e2, where e1 is a process and e2 is a state. Such relation can be represented in a tree structure shown on the right hand side of which indicates that the two sub-events are temporally ordered in a way that e1 precedes e2 (cf. Sect. 6.3.1). The tree structure is then integrated with the semantic and syntactic properties of e1 for discriminating different emotions, i.e. [#transitivity] and [#epistemic markup]. In order to develop a complete emotion-event representation extensive and in-depth analyses of event structures are required.
6.5
Summary
This chapter proposes an emotion representation model through enriching NSM’s emotion semantic explications based on the linguistic realizations of emotion causes as well as GL’s event structures. My proposal puts special emphasis on the close relation between emotions and events. This new model decodes emotion as an event type and assumes that an emotion construction typically comprises a series of events, including cause events, emotional state, and elicited events all of which can be represented in GL’s event structures. I focus on the link between emotion and cause events by incorporating the previous linguistic findings of emotion cause events into the emotion representation model to establish a set of linguistic criteria for emotion classification and representation. These linguistic criteria include semantic properties, such as transitivity and syntactic markup, such as epistemicity. In other words, the emotion representation model not only provides deep linguistic
100
6 A Linguistic Model for Emotion Expression
criteria for emotion classification of emotion cause events, but also offers an event-based account of emotion classification. These linguistic criteria will be attested by applications in language technology in Chap. 7 adding value to current approaches by providing a comparatively richer knowledge representation.
Chapter 7
Implementation and Verification: Automatic Detection of Emotion Causes
7.1
Introduction
Given the linguistic findings discussed in Chaps. 4, 5, and 6, we now have a better understanding of emotions in terms of their correlations with cause events based on empirical data. Following the discussion of the linguistic criteria of emotion cause events for emotion classification, this chapter aims to attest to these linguistic criteria in computational applications. Indubitably, emotion plays a crucial role in human communication. With the advances of computer-mediated communication, emotion information is widely conveyed by means of written language, such as in emails and social media. Understanding emotions is becoming increasingly important not only for speech processing but also for written text. Consequently, textual emotion analysis has begun to attract attention and has become a popular topic for research in natural language processing (NLP) recently. Up to this date, most research has focused on emotion detection and classification by identifying emotion types, for instance, HAPPINESS and SADNESS, for a given sentence or document (Wiebe et al. 2005; Mihalcea and Liu 2006; Tokuhisa et al. 2008). However, on top of this surface level information, deeper level information regarding the experiencer, cause, and result of an emotion needs to be extracted and analyzed for real-world applications (Alm 2009). This chapter aims at mining one of the crucial pieces of deep level information, i.e. emotion cause, which provides useful information for applications ranging from economic forecasting through public opinion mining to product design. Most theories of emotion treat recognition of a triggering cause event as an integral part of emotion (Descartes 1649; James 1884; Plutchik 1962; Wierzbicka 1996). As discussed in Chap. 2, cause events refer to the explicitly expressed arguments or events that evoke the presence of the corresponding emotions. They are usually expressed by means of verbs, nominalizations, and nominals. For example, “they like it” is the cause event of the emotion HAPPINESS in the sentence “I © Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3_7
101
102
7 Implementation and Verification: Automatic Detection …
was very happy that they like it”. This chapter aims to develop an automatic system for cause event detection. First, I build a Chinese emotion cause corpus annotated for the five primary emotions, i.e. HAPPINESS, SADNESS, ANGER, FEAR, and SURPRISE. I then examine various linguistic cues which help detect emotion cause events, such as the position of the cause event and the experiencer relative to the emotion keyword, causative and reported verbs, epistemic markers, conjunctions, and prepositions. With the help of these cues, a list of linguistic rules is generated. Finally, based on these linguistic rules, I develop a rule-based system for emotion cause detection. Experiments show that such a rule-based system performs promisingly well and, at the same time, verifies my proposed qualitative linguistic framework. I believe that the current study will lay the groundwork for future research on inferences and entailments of new information based on cause-event relations, such as the detection of implicit emotion or cause, as well as the prediction of public opinion based on cause events, etc. The chapter is organized as follows: Sect. 7.2 discusses the relevant work on various aspects of emotion analysis. Section 7.3 describes the construction of the emotion cause corpus and provides an analysis of the corpus. Section 7.4 presents the rule-based system for emotion cause detection. Section 7.5 describes its evaluation and performance. Finally, Sect. 7.6 highlights the main contributions of the automatic system for emotion detection as well as the current work.
7.2
Previous Work on Automatic Emotion Detection and Classification
This section will discuss previous studies on automatic emotion analysis and underline some fundamental yet unresolved issues, such as emotion classification and representation. It will also survey the previous attempts on textual emotion processing and discuss how the present study differs from previous work.
7.2.1
Emotion Classification and Representation1
Emotions classification has long been a challenge. As discussed in Chap. 1, various approaches to emotion classification were proposed in different fields each varying in their taxonomy as well as from language to language. The cognitive perspective focuses on how to define an emotion and how to discern one emotion from another emotion, for instance, SADNESS vs. DEPRESSION, ENVY vs. JEALOUSY, etc. Yet, in terms of automatic emotion processing, rather than the definition or the classification, the
1
Relevant discussions regarding emotion classification and representation have been presented in one of our previous papers (see Chen et al. 2009b).
7.2 Previous Work on Automatic Emotion Detection and Classification
103
most crucial and fundamental issue is how emotions should be represented. The representation of an emotion directly influences the choice of the classification method. Generally, there are two commonly adopted emotion representations, enumerative representation and compositional representation. I will discuss these two types along with the corresponding classification.
7.2.1.1
Enumerative and Compositional Representations
When emotions are enumeratively represented they are identified with a unique label, such as SADNESS, DEPRESSION, WORRY, REMORSE, etc. Such labelling is a straightforward way of representing emotions. There are, however, a number of drawbacks to this. One of the first major drawbacks of enumerative representation is the problem of data sparsity. For instance, studies in relation to English emotion made use of data from a blog corpus collected from LiveJournal (Mishne 2005; Mihalcea and Liu 2006). In LiveJournal, authors are allowed to describe their emotions towards each post in a way that the description words are either selected from a predefined list of 132 common emotions or entered freely by bloggers. Mishne (2005) found that 54,487 unique emotion words appeared in 624,905 blog posts and that 85.4% of the emotion words appeared only once. Mishne’s findings imply that adopting the enumerative representation requires a very fine-grained classification of emotions. In doing so, the list of emotion keywords should be comprehensive enough to cover most existing emotions. It is, nonetheless, impossible to collect all emotion words in a language. Moreover, the high percentage of non-frequent emotion keywords usually leads to a problem of data sparsity that, in turn, hinders emotion processing. Another drawback of enumerative representation is that it fails to capture the complicated relationships between individual emotions. For instance, there is a closer relationship between SADNESS and REMORSE than between SADNESS and ANGER, etc. Most emotion theories realize that except for a few prototypical emotions, an emotion often involves several other emotions, such as REMORSE involving both SADNESS and FEAR. A satisfactory emotion classification model should thus be able to detect or identify this kind of relationship. Instead of enumerating all possible emotions, some emotion theories suggest representing an emotion by small-scale fixed dimensions (Plutchik 1980; Turner 2000). These, so-called compositional representations offer a rather loose way of describing an emotion resulting in potential information loss. Kemper (1987) suggested that complex emotions usually result from various aspects of social interaction which are rather culture-specific. GUILT, for instance, apart from being decomposed into JOY and FEAR, may involve other culture-related moods which are lost in this compositional representation. Thus, one crucial factor for being an adequate compositional representation is the selection of dimensions and the decomposing of emotions to capture as much information as possible. Most applications utilize some prototypical emotions as dimensions and other
104
7 Implementation and Verification: Automatic Detection …
complement dimensions specifically designed for the applications. For example, Quan and Ren (2009) designed a scheme to annotate an emotion corpus for robots in which an emotion is expressed by eight prototypical emotions with other accessory dimensions. Comparing these two representation methods, I find that the problems resulting from the enumerative representation can be avoided in the compositional representation, even though the enumerative representation is indeed capable of containing more information of an emotion than the compositional representation. In the emotion corpus, an emotion is compositionally represented by the five primary emotions (i.e. five dimensions), namely HAPPINESS, SADNESS, FEAR, ANGER, and SURPRISE based on my proposed emotion taxonomy discussed in Chap. 2.
7.2.1.2
Single-Label and Multi-label Classifications
A common technology for classification in NLP is single-label classification. For single-label classification each instance contains, one and only one, label. The labels are pre-defined and are assumed to be mutually exclusive. This assumption is invalid in an emotion classification that adopts compositional representation. Thus, multi-label classification is what is adopted in this work. Multi-label classification has been widely applied in text categorization (e.g., McCallum 1999; Fujino et al. 2008; Sajnani et al. 2011) as more than one topic can be assigned to a document. The important difference between single-label classification and multi-label classification is that the latter requires capturing the relationships among different labels. Recently, a number of technologies have been developed to achieve this mutual information (Zhu et al. 2005; Ji et al. 2008; Tang et al. 2009). However, not much work (Trohidis et al. 2008) has been done to use this technology for emotion analysis, partially because emotion analysis is still a controversial issue and is not as well-developed as text categorization. This corpus is designed to accommodate multi-label classification, even though only primary emotions are included for the present work.
7.2.2
Emotion Processing in Text
Textual emotion processing is still in its early stages in NLP. Most of the previous research focus on emotion classification in a known emotion context, such as sentence or document using either rule-based (Masum et al. 2007; Chaumartin 2007) or statistical approaches (Mihalcea and Liu 2006; Kozareva et al. 2007). The performance reported in these studies is far from satisfactory though. Moreover, many basic issues remain unresolved; for instance, the relationships among emotions, emotion type selection, etc. Tokuhisa et al. (2008) were the first to explore both the issues of emotion detection and classification. They created a Japanese emotion-provoking event corpus for an emotion classification task using an
7.2 Previous Work on Automatic Emotion Detection and Classification
105
unsupervised approach. However, only 49.4% of the cases are correctly labeled. Chen et al. (2009b) developed two cognitive-based Chinese emotion corpora using a semi-unsupervised approach, i.e. an emotion-sentence (sentences containing emotions) corpus and a neutral-sentence (sentences containing no emotion) corpus. They concluded that studies based on the emotion-sentence corpus (*70%) outperformed previous corpora. Apart from corpora, some works have focused on other important resources for emotion processing, i.e. emotion ontologies. An emotion ontology defines the structure of emotion and identifies the interaction among emotions. Obrenovic et al. (2005) designed a predicate-based emotion ontology for English. Oltramari (2006) combined the features of two important lexical resources, i.e. WordNet2 and FrameNet. Such integration generally improved the interoperability, user-friendliness, and usability of both lexical resources. Yang et al. (2008) semi-automatically created a Chinese emotion ontology using the event hierarchy in HowNet.3 The final ontology contains 5498 verb concepts. López et al. (2008) developed an ontology for describing emotions based on OWL,4 Descriptive Ontology for Linguistic and Cognitive Engineering (DOLCE),5 and FrameNet. Although emotion processing has attracted considerable attention in recent years, little research, if any, has been done to examine the interactions between emotions and the corresponding cause events which may contribute towards an effective emotion classification model. The lack of research on cause events limits current emotion analysis to simple classificatory work without exploring the potentials of putting emotion ‘in context’. In fact, emotions are invoked by perceptions of external events and, in turn, trigger reactions. The ability to detect implicit invoking causes as well as predict actual reactions will add rich dimensions to emotion analysis and open up future vistas in event processing.
2
WordNet is a large lexical database of English. Nouns, verbs, adjectives, and adverbs are grouped into sets of cognitive synonyms (synsets), each expressing a distinct concept. Synsets are interlinked by means of conceptual-semantic and lexical relations. Its structure makes it a useful tool for computational linguistics and natural language processing. Official website: http://wordnet. princeton.edu/. 3 HowNet is a Chinese lexical dictionary which describes the inter-conceptual relations and inter-attribute relations among words and concepts as a network (Dong and Dong 2000). 4 The Web Ontology Language (OWL) is a family of knowledge representation languages for authoring ontologies. The languages are characterised by formal semantics and RDF/XML-based serializations for the Semantic Web. Official website: http://www.w3.org/TR/owl-ref/. 5 DOLCE aims at capturing the ontological categories underlying natural language and human commonsense. Events or processes constitute a sub-class of perdurant occurrences that are disjoint from the entities of endurance (i.e. objects or substances), quality, and abstract. Official website: http://www.loa.istc.cnr.it/old/DOLCE.html.
7 Implementation and Verification: Automatic Detection …
106
7.3
Emotion Cause Corpus
A state-of-the-art statistical NLP system is based on a supervised machine learning approach where labeled corpus is essential for training the system. As mentioned in the previous section, there have not been any studies focusing on emotion cause detection. In order to analyze emotion causes and build an automatic detection system with machine learning approaches, an emotion cause corpus is created. This section details how the emotion cause corpus was constructed. Before discussing the corpus, I first intend to clarify the definition of an emotion cause, discuss what is considered an emotion cause event, and how it is linguistically expressed in Chinese. I then describe the corpus data and the annotation scheme (see also Lee et al. 2010a, 2013).
7.3.1
Cause Events
Recall that in Chap. 2, in the context of emotion constructions, cause events referred to any events that are highly associated with the presence of an emotion. They may serve as an actual trigger event or a potential trigger event. In Chinese, cause events are categorized into two types: verbal events and nominal events. Verbal events refer to events that involve verbs (1) or nominalizations (2), whereas nominal events are simply nouns (3). Some examples of cause event types are repeated below: (1) mei2 xiang3dao4 ta1 shuo1 de dou1 shi4 zhen1 hua4, not think 3.SG.F say POSS all is true word, zhen4jing1 bu4yi3 to be shocked very ‘Unexpectedly, what she said was the truth, which surprised him greatly.’
rang4 ta1 lead 3.SG.M
(2) ta1 dui4 zhe4 ge4 chong1man3 3.SG.M for DET CL full of gao1xing4 de shou3wu3zu2dao3 becoming happy DE flourish ‘He was very happy about this lovely idea.’
nong2hou4 ai4yi4 dense love
de DE
xiang3fa3 idea
(3) Ni2ao4 de hua4 hen3 ling4 Leo POSS word very make ‘Caroline was saddened by Leo’s words.’
Kai3luo4lin2 Caroline
shang1xin1 becoming sad
The bolded parts in (1)–(3) exemplify cause events that do not require coreference. Yet, there are many cases requiring coreference in the corpus. For instance, in (4), 此 ci3 ‘this’ is considered the cause event of anger. It is an anaphoric
7.3 Emotion Cause Corpus
107
expression that refers to the previous clause “The Taiwan authorities refused the person involved returning to Taiwan on political grounds”. In cases such as (4), both references are marked as cause events; the former as a nominal cause event, whereas the latter as a verbal cause event. (4) Tai3wan1 dang1ju2 yi3 zheng4zhi4 li3you2 ju4jue2 dang1shi4ren2 fan3 Taiwan authority as political reason refuse the person involved retur tai2. dang1shi4ren2 dui4 ci3 fen4men4 Taiwan. the person involved for this angry ‘The Taiwan authorities refused to allow the person involved to return to Taiwan on political grounds. Because of this, the person involved was furious.’
When the antecedent is not expressed in the context, only the anaphoric expression is marked, as in (5): (5) fu1qi1liang3 zuo2tian1 dou1 dui4 ci3 xi3xun4 xing4fen4bu4yi3 husband and wife yesterday both forDET good news happy very ‘Yesterday, the couple were very excited about this good news.’
As discussed in Chap. 4, the occurrence of covert subjects and objects is a common phenomenon in Chinese syntactic structure. Therefore, it is not uncommon that a single verb alone overtly denotes the cause event. Consider the cause event in (6), 我聽了 wo3ting1le ‘after I heard’. In this case, the object which refers to the previous sentence, i.e. what mom said, is dropped. There are also cases where the actual cause events are not expressed in the context as in (7) and (8). For consistency, instead of marking the previous sentence as the cause event, I mark the trigger event that is closest to the emotion keyword as the cause. Moreover, the actual cause event is typically excessively long and complicated as it involves several events, as seen in (6).
7 Implementation and Verification: Automatic Detection …
108 (6)
ma1ma wei4le gu3li4 wo3 kao3shi4 de2 man3 fen1, bian4 gao4su4 wo3 shuo1: mom for encourage 1.SG exam get full mark, then tell 1.SG say: ru2guo3 ni3 kao3 yi4 zhang1 yi4bai3 fen1, ma1ma jiu4 gei3 ni3 if 2.SG exam one CL one hundred mark, mom then give 2.SG wu3 yuan2 zuo4 jiang3li4. wo3 ting1 le hao3 gao1xing4 five dollar as reward. 1.SG hear ASP very becoming happy ‘In order to encourage me to get full marks in the exam, Mom told me, ‘if you get 100 points, Mom will give you five dollars as an award’. I was very happy when I heard [this].’ (7) … …jiu4 le wo3 de xing4ming4. ren2yu2 gong1zhu3 ting1 le hen3 shang1xin1 …save ASP 1.SG POSS life. Mermaid princess hear ASP very becoming sad ‘… saved my life. The mermaid was very sad after hearing [this].’ (8) ta1 yue4 xiang3 yue4 sheng1qi4 3.SG.M more think more becoming angry ‘The more he thought, the angrier he became.’
In general, in terms of event annotation, emotion causes are loosely defined. I mark the shortest meaningful cause events that are closest to the emotion keywords.
7.3.2
Corpus Data and Annotation Scheme
Based on the list of the 91 primary emotion keywords identified in Chap. 2 (an earlier version of which appeared in Chen et al. 2009b), I extract 6058 instances of sentences by keyword matching from the Sinica Corpus in the emotion corpus (Lee et al. 2010a). Each instance contains the focus sentence with the emotion keyword “” plus the sentence before “” and after “” it. The extracted instances include all primary emotion keywords that occurred in the Sinica Corpus except for the emotion class HAPPINESS as the keywords of HAPPINESS exceptionally outnumber other emotion classes (there are 2544 keywords of HAPPINESS). In order to balance the number of each emotion class, I set the upper limit at about 1600 instances for each primary emotion (see Table 7.1). Note that the presence of emotion keywords does not necessarily convey emotional information due to various possible reasons such as negative polarity and sense ambiguity. Therefore, I remove instances that (1) do not involve
7.3 Emotion Cause Corpus Table 7.1 Summary of cause corpus data
109 Emotions
No. of instances Extracted Emotional
With causes
HAPPINESS
1646 902 898 1177 1341 5964
1132 468 567 629 664 3460
SADNESS FEAR ANGER SURPRISE
Total (average)
1327 616 689 847 781 4260
(81%) (68%) (77%) (72%) (58%) (71%)
(82%) (75%) (81%) (73%) (85%) (81%)
emotion-related information; (2) contain highly ambiguous emotion keywords,6 such as 如意 ru2yi4 ‘to be happy’, 害羞 hai4xiu1 ‘to be shy’, 為難 wei2nan2 ‘to feel awkward’7 etc. After the removal of eight ambiguous emotion keywords, there are 5629 remaining instances in the emotion cause corpus. Among these, I also remove the emotion keywords where the sentence does not express the same emotion as when it is negated. The total number of emotion keywords in the corpus is 5964. An example of an instance is given in (9): (9) 3015 Y NONE 還是你有別的想法吧! 楊業聽了很不 高興 , 說道「我不是 害怕 打仗,而是覺得眼 前遼軍氣勢正盛,我們取勝的機會很小,卻會犧牲很多戰士,是很划不來 的。 你們既然要我去和遼軍拼戰,我也 沒什麼不敢的! 3015 Y NONE Perhaps you have other idea! After hearing that, Yang Ye was not very happy , and said, “I am not afraid of war, but the Liao army is showing vigorous momentum. A very small chance that we will win, but at the expense of many soldiers, it doesn’t worth it at all. Now that you ask me to battle against the Liao army, I have no fear! The instance number of (9) is 3015, the two emotion keywords are marked as 高興 gao1xing4 ‘becoming happy’ and 害怕 hai4pa4 ‘becoming afraid’ . The keyword 高興 gao1xing4 in the context does not express the emotion of HAPPINESS, but another emotion 不高興 bu4 gao1xing4 ‘angry’ instead. The other keyword 害 6
Highly ambiguous emotion keywords refer to the emotion keywords which are likely to convey non-emotional information or different emotions. For the purpose of emotion processing they should be removed in order to avoid unnecessary noise, even though these emotion keywords may be disambiguated by the context. 7 The eight ambiguous emotion keywords are: 為難 wei2nan2 ‘to feel awkward’, 害羞 hai4xiu1 ‘shy’, 羞怯 xiu1qie4 ‘timid’, 苦惱 ku3nao3 ‘worried’, 困惑 kun4huo4 ‘bewildered’, 討厭 tao3yan4 ‘dislike’, 如意 ru2yi4 ‘happy’, 舒坦 shu1tan ‘to feel comfortable’.
110
7 Implementation and Verification: Automatic Detection …
怕 hai4pa4 ‘being afraid’ is modified by 不是 bu2shi4 ‘not’ meaning “not afraid of” which is non-emotional. In other words, the emotion involved in this instance is angry which is not expressed by either emotion keywords. The emotion keywords are thus replaced by “NONE” indicating none of the emotion keywords expresses the emotion that they should express. For each emotion instance, two annotators manually annotate the cause events of each keyword. Since more than one emotion can be present in an instance, the emotion keywords are tagged as , , and so on. Two examples are given in Fig. 7.1. For an emotion keyword to be tagged as in instance 573, [*01n] marks the beginning of its cause event, while [*02n] marks the end. The “0” shows which index of emotion keyword it refers to, “1” marks the beginning of the cause event, “2” marks the end, and “n” indicates that the cause is a nominal event. For an emotion keyword to be tagged as in instance 3016, [*11e] marks the beginning of the cause event, while [*12e] marks the end where “e” refers to a verbal event, i.e. either an event or a nominalization. An emotion keyword can sometimes be associated with more than one cause, in which case both causes are marked. The emotional sentences containing no
Fig. 7.1 Two examples of cause event annotation
7.3 Emotion Cause Corpus
111
explicitly expressed cause events remain as they are. Instances without causes explicitly expressed are mainly due to the following reasons: (i) there is not enough contextual information, for instance, the previous or the suffix sentence is an interjection, e.g. 嗯哼 en heng ‘aha’; (ii) the cause falling outside the context when the focus sentence is the beginning or the end of a paragraph, thus no prefix or suffix sentence being extracted as the context; or (iii) when the cause is obscure either in that it is very abstract, or even unknown. This is especially true for heterogeneous emotion verbs. Table 7.1 presents the actual number of extracted instances of each emotion class to be analyzed, the emotional instances, and the instances with cause events. The total number of extracted instances after the removal of the eight ambiguous keywords is 5964. Based on the annotators’ markup, we can see that on average 71% of the extracted instances express emotions and 81% of the emotional instances have a cause. In the corpus, HAPPINESS (1327) keywords are the most frequent and SADNESS keywords (616) are the least. For each emotion type, about 80% of the emotional sentences, on average, contain a cause event with SURPRISE yielding the highest percentage (85%) and ANGER yielding the lowest (73%). This indicates that an emotion mostly occurs with the cause event explicitly expressed in the text regardless of emotion classes, which, by extension, implies the prominent role of cause events in expressing emotion. Thus, an extensive investigation of cause events will have considerable implications for emotion analysis and processing.
7.3.3
Data Sets
The emotion cause corpus is randomly partitioned into two sets of data for different purposes, namely the development data and the test data. The development data is the data set that allows us to conduct careful analysis of emotion causes and to develop a rule-based system. The test data is the data set that is reserved for testing purposes, i.e. to test how powerful the system is. Therefore, it should not be accessed during the development of the rule-based system. In this study, 80% of the emotion cause corpus serves as development data. It is used for processing the rule-based system, including corpus analysis and rule generalization. The rule-based system is then tested by making predictions against the remaining 20% corpus data, i.e. test data. The results of the test data determine the predictive power of the system. More details concerning the mechanism of the rule-based system will be given in Sect. 7.5.1.
7 Implementation and Verification: Automatic Detection …
112
7.3.4
Corpus Analysis
Analyzing the corpus data allows better prediction of the relation between cause events and emotions as well as the linguistic cues facilitating cause event detection. Based on the development data, I examine the position of cause events as well as a list of linguistic cues, including causative verbs, reported verbs, epistemic markers, conjunctions, and prepositions, and a group of cue words called others (see Lee et al. 2010a, 2013). Each of these cues is shown to be in collocation with cause events in certain ways, which will be discussed in detail in the subsequent sections.
7.3.4.1
Position of Cause Events
I calculate the distribution of cause event types of each emotion and the position of cause events relative to emotion keywords. The total number of emotion instances regarding each emotion is given in Table 7.2: Table 7.2 suggests that emotion cause events tend to be expressed more often by verbal events (85% on average) than by nominal events, and that cause events tend to occur at the position to the left (67%) of the emotion keyword, except for FEAR, which reflects a more or less even split between left (52%) and right. This may be attributed to the fact that FEAR can be triggered by either factive or potential causes, which is rare for other primary emotions. For FEAR, factive causes tend to take the left position; whereas potential causes tend to take the right position. Based on these observations, it is assumed that most of the undetected causes should come before the emotion keyword (cf. Rules 14 and 15 in Sect. 7.4.2). Considering also the findings in Chap. 5, according to which change-of-state verbs of certain emotions display their cause events as a complement, the causes of FEAR, HAPPINESS, and SURPRISE sometimes take the right position (cf. Rules 4 and 8). The cause can sometimes be long and complicated, involving several events. In order to explore the span of a cause text, I do the following analysis. Firstly, for each emotion keyword, an instance is segmented into clauses with four punctuations (i.e. commas, periods, question marks, and exclamation marks), and thus an instance becomes a list of cause candidates. For example, when an instance has four clauses, its corresponding list of cause candidates contains five text units, i.e. Table 7.2 Cause event position of each emotion
Emotions
Cause type (%) Event
Nominal
Cause position (%) Left Right
HAPPINESS
93 89 84 76 86 85
7 11 16 24 14 15
74 80 52 71 59 67
SADNESS FEAR ANGER SURPRISE
Average
26 20 48 29 41 33
7.3 Emotion Cause Corpus
113
. When I assume the clause where an emotion keyword is located is a focus clause, [left_2] and [left_1] are the two previous clauses, and [right_1] is the subsequent one. [left_0] and [right_0] are the partial texts of the focus clause which are located in the left side and the right side of the emotion keyword, respectively. Moreover, a cause candidate must contain either a noun or a verb because a cause is either a verbal event or a nominal event; otherwise, it will be removed from the list. Then, I calculate whether a cause candidate overlaps with the real cause and find that the most possible cause span should be [left_2, right_1]. Moreover, for all causes occurring between [left_2] and [right_1], I calculate whether a cause occurs across clauses. It is observed that most causes are located within the same clause of the representation of the emotion (85.57%). This suggests that a clause may be the most appropriate unit to detect a cause. With this observation in mind, it is assumed that (1) most of the undetected causes should appear before the emotion keyword, and (2) the most possible search span for causes should be [left_2, right_1].
7.3.4.2
Linguistic Cues
In addition to the position of cause events, I also examine the correlations between emotions and cause events in terms of multiple linguistic cues: causative verbs, reported verbs, epistemic markers, conjunctions, prepositions, and others. I hypothesize that these cues mark cause events in a way which helps generate the linguistic rules presented in Sect. 7.4. First, I postulate lists of linguistic cues that are likely to be collocated with cause events, as presented in Table 7.3. Next, the frequencies of these linguistic cues collocating with cause events of a given emotion type are calculated. Based on the previous observation, I set the window size for statistics at [left_2, right_1]. Within the window size, I calculate the position of the six groups of linguistic cue words relative to the cause event, i.e. to the left or right of the cause event. Statistics indicate that the window size [left_2, right_1], i.e. the size between two clauses before the beginning of the cause event and one clause after the end of the cause event, shows a clear pattern of marking the cause event. Note that the statistics do not indicate the accurate frequencies of the keywords functioning as a cause event cue, since there are certainly occurrences of keywords in the context which do not serve as a cue. Even so, it provides a tendency indicating the reliability of the cue words as well as the position of the cue words relative to the cause events. In the following sub-sections, each group of linguistic cues is discussed and evaluated.
7.3.4.3
Causative Verbs
Causative verbs are verbs involved in expressions of causing or forcing a patient to perform certain actions or to be in certain states. In Chinese, common causatives are 使 shi3, 令 ling4, 讓 rang4, which correspond to the English ‘to make’ or ‘to
7 Implementation and Verification: Automatic Detection …
114
Table 7.3 Lists of potential linguistic cues Group
Keywords
Causative verbs Reported verbs
“to cause”: 讓 rang4, 令 ling4, 使 shi3
Epistemic markers
Prepositions
Conjunctions
Others
‘to think about’: 想到 xiang3dao4, 想起 xiang3qi3, 一想 yi4xiang3, 想來 xiang3lai2 ‘to talk about’: 說到 shuo1dao4, 說起 shuo1qi3, 一說 yi4shuo1, 講到 jiang3dao4, 講起 jiang3qi3, 一講 yi4jiang3, 談到 tan2dao4, 談起 tan2qi3, 一談 yi4tan2, 提到 ti2dao4, 提起 ti2qi3, 一提 yi4ti2 ‘to hear’: 聽 ting1, 聽到 ting1dao4, 聽說 ting1shuo1 ‘to see’: 看 kan4, 看到 kan4dao4, 看見 kan4jian4, 見到 jian4dao4, 見 jian4, 眼看 yan3kan4, 瞧見 qiao2jian4 ‘to know’: 知道 zhi1dao4, 得知 de2zhi1, 得悉 de2xi1, 獲知 huo4zhi1, 獲悉 huo4xi1 ‘to discover’: 發現 fa1xian4, 發覺 fa1jue2 ‘to exist’: 有 you3 ‘for’ as in ‘I will do this for you’: 為 wei4, 為了 wei4le ‘for’ as in ‘He is too old for the job’: 對 dui4, 對於 dui4yu2 ‘as’: 以 yi3 ‘because’: 因 yin1, 因為 yin1wei4, 由於 you2yu2 ‘so’: 於是 yu1shi4, 所以 suo3yi3, 因而 yin1er2 ‘but’: 可是 ke3shi4 ‘is’: 的是 deshi4 ‘say’: 的說 deshuo1 ‘at’: 於 yu2 ‘can’: 能 neng2
Table 7.4 Position of causative verbs relative to cause events
Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Average
Causative verbs (%) Left Focus
Right
42.1 39.5 48.2 42.7 45.6 44
47.7 55.5 44.4 42.6 47.3 47
10.2 5.0 7.4 14.7 7.1 9
Total (#) 197 119 81 129 226 752
cause’. Table 7.4 shows the position distribution of causative verbs relative to cause events. Results indicate that causative verbs usually come either before the cause event (44%) or after the cause events (47%). Examples of both constructions are given in (10) and (11), respectively.
7.3 Emotion Cause Corpus (10)
(11)
115
19 (Rule 1)20 zhi1jian1 de yin1 guo3 guan1xi4, ling4 ren2 ai1shang1 liang3zhe3 two parties between DE cause result relation, to cause people sad ‘The cause-effect relationship between the two made people sad.’ [*01n]
[*02n]
[*01e] [*02e] (Rule 6) ling4 wai4jie4 cha4yi4 deshi4 zai4 lian2he2 ji4zhe3hui4 shang4, to cause outside world surprise is that atjoint press conference up, Ye4er3xin1 ti3li4 yu3 jing1shen2 zhuang4kuang4 dou1 qian4jia1 Yeltsin physical and mental condition both bad ‘What surprised the outside world was that Yeltsin was in very poor physical and mental condition at the joint press conference.’
When looking into the development data, I find that most of the causative verbs appear to the right of the cause events indicating the end of the cause events as can be seen in (10). On the other hand, the ones appearing to the left of the cause events mark the beginning of cause events, setting the constraint that the emotion keywords should be followed by words such as 的是 deshi1 ‘say’, as in (11). Therefore, in order to differentiate the cause events at different positions and avoid incorrect detections, I set constraints to the corresponding linguistic rules, as presented in (12) and (13) (see Sect. 7.4.2 for the linguistic rules). (12) Rule 1: K(F) cannot be followed by 是 shi4 ‘is’, 的是 deshi4 ‘is that’, 莫過於 mo4guo4yu2 ‘nothing is more … than…’ (13) Rule 6: K(F) must be followed by 是 shi4 ‘is’, 的是 deshi4 ‘is that’, 莫過於 mo4guo4yu2 ‘nothing is more … than…’
7.3.4.4
Reported Verbs
Reported verbs are verbs which report what other speakers say. The reported verbs marking the cause events are thinking and talking verbs. Statistics clearly show that the reported verbs tend to appear to the left of the cause events (59% vs. 20% (focus) and 21% (right)), as can be seen in Table 7.5. In other words, reported verbs mainly indicate the beginning of cause events (cf. Rules 2 and 3), as shown in (14). Table 7.5 Position of reported verbs relative to cause events
Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Average
Reported verbs (%) Left Focus
Right
45.7 68.8 87.5 50.0 40.9 59
23.9 18.8 6.25 27.8 27.3 21
30.4 12.4 6.25 22.2 31.8 20
Total (#) 46 32 16 18 22 134
7 Implementation and Verification: Automatic Detection …
116 (14)
[*01n] [*02n] Tan2dao4 bao3lai2 guo2ji4 jin1rong2 ji1chang3, bao3lai2 zheng4quan4 speaking of Polaris international finance airport, Polaris security ji2tuan2 dong3shi4zhang3 Bai2wen2zheng4 nan2 yan3 xing1fen4 zhi1 qing2 group chairman Wayne Pai hard conceal excited POSS feeling ‘Speaking of ‘Polaris International Financial Airport’, the Chairman of Polaris Securities Group Wayne Pai could not conceal his excitement.’
It is not surprising that the reported verbs frequently collocate with cause events of FEAR (87.5%), since fear is often triggered by potential causes that require a certain degree of perception of possible events. The cause events are usually recalled from past experience by the reported verbs. SURPRISE, in contrast, is usually triggered by a sudden change of the situation instead of recalling a past experience; thus, it collocates less with reported verbs (40.9%).
7.3.4.4.1 Epistemic Markers Epistemic markers, as defined in Chap. 5, are verbs which mark the cognitive awareness of emotion cause events. Five types of epistemic markers are identified, including SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE. It is noted in Table 7.6 that epistemic markers tend to appear before the cause event marking the beginning of cause events, such as in (15)–(17). (15)
[*01e]
[*02e]
Lan2ni1 gong1zhu3 ting1dao4 zi4ji3 yong3yuan3 bu4 neng2 hui1fu4 Lanny princess to hear oneself forever not can return ren2xing2, shang1xin1 ji2 le human figure, becoming sad extreme ASP ‘Princess Lanny was very sad to hear that she would never be turned back into a human being. (16)
[*01e] [*02e] fa1xian4 guo2wang2 yi3jing1 e4 si3 zai4 na4 le, to discover king already hungry die at there ASP, dou1 shi2fen1 shang1xin1 all very becoming sad ‘Everyone was very sad to find that the King was starved to death.’
da4jia1 everybody
(17)
[*01e] [*02e] wo3 shi2zai4 tai4 gao1xing4 kan4dao4 min2jian1 zhu3dong4 1.SG indeed very becoming happy to see non-governmental circles initiate cheng2li4 zhe4 ge4 wei3yuan2hui4 form DET CL committee ‘I am very happy to see that non-governmental circles took the initiative to set up this committee.’
7.3 Emotion Cause Corpus Table 7.6 Position of epistemic markers relative to cause events
117 Emotions
Epistemic markers (%) Left Focus Right
Total (#)
HAPPINESS
42.1 54.8 47.8 50.9 42.7 48
935 347 387 446 571 2686
SADNESS FEAR ANGER SURPRISE
Average
41.8 25.0 30.5 30.0 35.9 32
16.1 20.2 21.7 19.1 21.4 20
The insignificant difference of epistemic markers appearing at the left position and the focus position, i.e. within the cause event, is attributed to the fact that the verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE actually denote parts of the cause event, instead of serving as the indicator marking the cause events as such (see Chap. 5 for argument). Consider the two sentences in (18) and (19): (18)
[*01e] [*02e] wo3 kan4dao4 zhe4 jian4 zuo4pin3 hen3 zhen4jing1 1.SG to see DET CL piece very to be surprised ‘When I saw this piece, [I] was very surprised.
(19)
[*01e] [*02e] nong2fu1 ting1dao4 zhe4 ju4 hua4, xin1li3 fei1chang2 gao1xing4 farmer to hear DET CL word, heart very becoming happy ‘When the farmer heard this, [s/he] was very happy.’
The cause events in (18) and (19) refer to the actual seeing ‘see the piece’ and hearing ‘hear this’ actions rather than functioning as cause event markers. Therefore, they are marked as part of the cause events. In addition, the comparatively low percentage of epistemic markers occurring before the cause event is partially due to the considerable noise resulting from the inclusion of the verb of existing 有 you3. Therefore, I set a constraint to exclude the verb 有 you3 in Rules 2 and 3, whereas Rule 4 is less affected. For instance, the cause event in (20) can be wrongly extracted as 看法 kan4fa3 ‘stance’ when the cause event should indeed be the complement. (20)
…
[*01e] [*02e] …ge4 you3 kan4fa3. Lin2bao3zhang1 biao3shi4, ta1 shi2fen1 tong4xin1 zai4 …each have stance. Lin Bao Zhang say, 3.SG.M very sad during zi4ji3 ren4 nei4 fa1sheng1 liu2hui4feng1bo1 himself term of office within occur disturbance ‘… each has a different stance, Lin Bao Zhang said, he was very sad that disturbances occurred in the flow of his term of office.’
7 Implementation and Verification: Automatic Detection …
118 Table 7.7 Position of prepositions relative to cause events
Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Total
Prepositions (%) Left Focus
Right
58.4 57.4 59.3 62.3 51.9 58
21.3 25.3 23.5 21.8 24.6 23
20.3 17.3 17.2 15.9 23.5 19
Total (#) 567 284 378 446 378 2053
7.3.4.4.2 Prepositions As seen in Table 7.7, most prepositions appear to the left of cause events (58% on average vs. 19% (focus) and 23% (right)), and mark the beginning of cause events across all primary emotions. (21)
[*01e] [*02e] (Rule 2) ta1 dui4 yao2yan2 san4bu4 zhi1 kuai4 gan3dao4 jing1ya4 3.SG.M for rumor spread POSS fast feel becoming surprised ‘He was surprised by how quickly the rumours spread.’
(22)
[*01e] [*02e] (Rule 3) dui4yu2 neng2 you3 ji1hui4 jing1ying2 ru2ci3 chu1se4 de ju4le4bu4, for can have opportunity run such great DE club, wo3 gan3dao4 fei1chang2 kuai4le4 1.SG feel very to be happy ‘I am very happy for having the opportunity to run such a great club.’
The cause events in (21) and (22) are marked by the indicators 對 dui4 ‘for’ and 對於 dui4yu2 ‘for’. It is noteworthy that the preposition 以 yi3 ‘as’ frequently occurs in the corpus and yet is not a reliable cue to mark cause events. It is thus removed from the list of prepositions.
7.3.4.4.3 Conjunctions Conjunctions are words that link two phrases or clauses. For cause event detection, three conjunctions are identified: because, so, and but. The overall statistics show that most conjunctions (68% on average vs. 4 (focus) and 28 (right)) appear before the cause events, as indicated in Table 7.8.
7.3 Emotion Cause Corpus Table 7.8 Position of conjunctions relative to cause events
119 Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Average
Conjunctions (%) Left Focus
Right
71.9 69.1 66.0 69.5 59.2 68
24.2 23.5 30.6 28.4 38.0 28
3.9 7.4 3.4 2.1 2.8 4
Total (#) 128 68 147 95 71 509
(23) (Rule 5) dan4 hen3 kuai4 de ta1men2 you4 gao1xing4 qi3lai2 le, yin1wei4 however very fast DE 3.PL again becoming happy begin ASP, because ta1men2 shi4 wei2yi1 liao3jie3 ri4ben3 guo4qu4 li4shi3 de ren2 3.PL is only one understand Japan past history POSS people ‘However, they became happy soon because they were the only people who understood the history of Japan.’ (24)
(Rule 2) you2yu2 ai4zi1bing4 chu1xian4 yin3qi3 da4jia1 because AIDS appear cause everybody ‘Because of the existence of AIDS, people are frightened.’
de kong3huang1 POSS fear
Based on the corpus data, the because-conjunction effectively marks the beginning of cause events, such as (23) and (24), as it is often used for stating the reason for something happening. However, I find that the so- and but-conjunctions do not show clear patterns in marking cause events. They are thus removed from the list of conjunctions.
7.3.4.4.4 Others There are some useful cues that do not fall into any of the above five groups of linguistic cues. They are grouped under “Others”, including 的是 deshi4 ‘is’, 的說 deshuo1 ‘to say’, 於 yu2 ‘at’, and 能 neng2 ‘can’. These cues are shown to be effectively marking the beginning of cause events as they tend to occur more often at the left position of cause events (58% on average). Examples are given in (25) and (26) (Table 7.9).
7 Implementation and Verification: Automatic Detection …
120 Table 7.9 Position of other cues relative to cause events
Emotions HAPPINESS SADNESS FEAR ANGER SURPRISE
Average
Others (%) Left Focus
Right
50.2 54.8 60.1 55.2 67.6 58
22.7 17.3 28.6 29.1 18.4 23
27.1 27.9 11.3 15.7 14.0 19
Total (#) 321 104 168 134 222 949
(25) (Rule 4) wo3 jing1ya4 deshi4, ju1ran2 you3 hou4xuan3ren2 yi3 zhe4 zhong3 1.SG becoming surprised is that, unexpectedly exist candidate as DET CL fang1shi4 lai2 he2li3hua4 dui4 nü3xing4 can1zheng4 de qi2shi4 way to rationalize for female participate in politics POSS discrimination ‘I was surprised that there were candidates who rationalized the discrimination against women in politics in this way.’ (26) (Rule 4) wo3 shi2zai4 tai4 gao1xing4 neng2 ying2de2 zhe4 ge4 jiang3, 1.SG indeed too becoming happy can win DET CL award, rang4 wo3 bei4gan3 guang1rong2 cause 1.SG feel honour ‘I am very happy that I could win this award, it makes me feel honoured!
In the development data, I notice two variants of the cue word 的說 deshuo1 ‘to say’: 道 dao4 ‘say’ and 說 shuo1 ‘to say’. These three cue words may occur before or after the emotion keywords as in (27) and (28), respectively; whereas other members of this group of cues can only occur after the emotion keywords. (27)
(Rule 12) wo3 xin1suan1 dao4: ke3shi4 ta1men hao3 1.SG sad say: but 3.PL very ‘I said sadly, ‘but they are pitiful!’
ke3lian2 o pityPART
(28) (Rule 13) na4 ge4 nan2ren2 shuo1: xiang4 wo3 zhe4yang4 hao3 tiao2jian4, geng4hao3 de DET CL man say: like 1.SG such good quality, better POSS nü3hai2 duo1 de shi4. zhe4 ci4 Qiao2yi1si1 ku1 de hen3 shang1xin1 girl many DE is. DET CL Joyce cry DE very becoming sad ‘The man said, ‘There are many desirable girls for a man with good qualities like me.’ At this, Joyce cried sadly.’
Due to the different marking behaviours, 的說 deshuo1 ‘to say’ is taken out of the group “Others” and instead forms another type of linguistic cue called ‘saying
7.3 Emotion Cause Corpus
121
verbs’ together with its variant 說 shuo1 and 道 dao4. Constraints are set for the rules in that the ‘say’ words should be followed by a colon in order to indicate that the cause event is in the subsequent conversation.
7.4 7.4.1
A Rule-Based System for Emotion Cause Detection8 Cause Event Markers
After investigating the corpus data, seven groups of linguistic cues that frequently collocate with cause events are obtained. They include causative verbs, reported verbs, saying verbs, epistemic markers, conjunctions, prepositions, and others. A summary is provided in Table 7.10. Group I includes three common causative verbs and Group II a list of reported verbs for thinking and talking. Group III is a list of saying verbs which should be followed by a colon. Group IV is a list of epistemic markers which are usually verbs explicitly marking the cognitive awareness of HAPPINESS, FEAR, SURPRISE in the complement position. The epistemic markers include verbs of SEEING, HEARING, KNOWING, DISCOVERY, and EXISTENCE. Group V covers some prepositions all of which roughly mean ‘for’. Group VI contains the because-conjunctions that explicitly mark the cause of the emotion. Group VII includes the other linguistic cues that indicate the cause event, yet do not fall into any of the six groups. Each group of linguistic cues, to a great extent, provides a transparent environment for emotion causes in different structures of emotional constructions, in which causative verbs specifically mark the end of the cause events while the other six groups mark the beginning of the cause events.
7.4.2
Linguistic Rules for Cause Detection
With the help of the various practical cues found in the corpus, I attempt to formulate some linguistic rules for cause event detection. To begin with, 100 emotional sentences featuring the emotion verbs under consideration are randomly extracted from the development data (i.e. 80% of the corpus data) to analyze the emotion-cause sentence constructions. As mentioned in Chap. 2, the cause is considered as a proposition. It is generally assumed that a proposition has a verb which optionally takes a noun occurring before it as the subject and a noun after it as the object. However, a cause can also be expressed as a nominal. In other words, both the predicate and the two
8
A shorter version of the rule-based system for emotion cause detection appears in Lee et al. (2013).
7 Implementation and Verification: Automatic Detection …
122
Table 7.10 Seven groups of cause event markers Group I
Cue words Causative verbs Reported verbs
II
III IV
Saying verbs Epistemic markers
V
Prepositions
VI VII
Conjunctions Others
‘to cause’: 讓 rang4, 令 ling4, 使 shi3 ‘to think about’: 想到 xiang3dao4, 想起 xiang3qi3, 一想 yi4xiang3, 想來 xiang3 lai2 ‘to talk about’: 說到 shuo1dao4, 說起 shuo1qi3, 一說 yi4shuo1, 講到 jiang3dao4, 講起 jiang3qi3, 一講 yi4jiang3, 談到 tan2dao4, 談起 tan2qi3, 一談 yi4tan2, 提到 ti2dao4, 提起 ti2qi3, 一提 yi4ti2 ‘to say’: 說 shuo1, 的說 deshuo1, 道 dao4 ‘to see’: 看 kan4, 看到 kan4dao4, 看見 kan4jian4, 見到 jian4dao4, 見 jian4, 看 yan3kan4, 瞧見 qiao2jian4 ‘to hear’: 聽 ting1, 聽到 ting1dao4, 聽說 ting1shuo1 ‘to know’: 知道 zhi1dao4, 得知 de2zhi1, 得悉 de2xi1, 獲知 huo4zhi1, 悉 huo4xi1 ‘to discover’: 發現 fa1xian4, 發覺 fa1jue2 ‘to exist’: 有 you3 ‘for’ as in ‘I will do this for you’: 為 wei4, 為了 wei4le ‘for’ as in ‘He is too old for the job’: 對 dui4, 對於 dui4yu2 ‘because’: 因 yin1, 因為 yin1wei4, 由於 you2yu2 ‘is’: 的是 deshi4 ‘at’: 於 yu2 ‘can’: 能 neng2
arguments are optional provided that at least one of them is present. The position of emotion verbs, experiencers, and linguistic cues are also taken into account. The clause containing the emotion verb is called the focus clause. All other components, i.e. cause events, experiencers, and linguistic cues may occur in the focus clause, the clause before the focus clause, or the clause after the focus clause. The abbreviations used in the rules are given in (29): (29) Abbreviations in Rules C: the Cause event E: the Experiencer K: the Keyword-emotion verb B: the clause Before the focus clause F: the Focus clause A: the clause After the focus clause
I: Causative verbs II: Reported verbs III: Saying verbs IV: Epistemic markers V: Conjunctions VI: Prepositions VII: Others
For illustration, an example of the rule description is given in Rule 1. Rule 1: (i) C(B-F) + I(F) + E(F) + K(F) (ii) E = the nearest Na-Nb-Nc-Nh after I in F (iii) C = the nearest (N) + (V) + (N) before I in F-B
7.4 A Rule-Based System for Emotion Cause Detection
123
(iv) Constraint: K(F) cannot be followed by 的 de ‘POSS’, 的是 deshi4 ‘is that’, 是 shi4 ‘is’ Rule 1 indicates that the experiencer (E) appears to be the nearest Na (common noun)/Nb (proper noun)/Nc (place noun)/Nh (pronoun) after Group I cue words in the focus clause (F), while at the same time coming before the keyword (K). In addition, the cause (C) comes before Group I cue words. I simplify the proposition as a structure of (N) + (V) + (N) which is very likely to contain the cause event. Theoretically, in identifying C, we should first look for the nearest verb occurring before List I cue words in the focus sentence (F) or the clause before the focus clause (B) and consider this verb as an anchor. From this verb, we search to the left for the nearest noun and consider it as the subject; we then search to the right for the nearest noun until the presence of a cue word and consider it as the object. The detected subject, verb, and object form the cause event. In most cases, the experiencer is allowed to be covertly expressed. In practice, however, it is difficult to detect such propositional causes as causes may contain no verbs as in the case of a nominal, and the two arguments can be optional. Therefore, I take the clause instead of the structure (N) + (V) + (N) as the cause event. In this system, a cause event is clause-based and is searched based on the keyword to the left clause or to the right clause. The search range for causes is between the two clauses before the focus clause and the clause after the focus clause, i.e. [left_2, right_1]. If the cause is detected in the focus clause, the range for cause extraction is between the beginning of the focus clause and the keyword, i.e. [left_0] or between the keyword and the end of the focus clause, i.e. [right_1]. The same mechanism is applied to the baseline. If the cause is not in the focus clause, the whole clause is considered as a cause. For instance, if the cause is detected in the clause before the focus clause, the whole clause, i.e. [left_1], is considered the cause event. To filter out incorrect causes, I set a constraint (cf. Sect. 7.3.4.4.4) to Rule 1 that the keyword in the focus sentence cannot be followed by 的 de ‘POSS’, 的是 deshi4 ‘is that’, or 是 shi4 ‘is’ since it is very likely to have the cause event occurring after such words which should be detected by another rule instead (i.e. Rule 4). Examples are given in (30) and (31). For both sentences, the clause that comes before the cue word is taken as the cause event of the emotion in question.
7 Implementation and Verification: Automatic Detection …
124 (30)
[C yi1la1ke4 xi4jun1 wu3qi4 mi4mi4 de bao4guang1], [I shi3] [C Iraq bacteria weapon secret POSS reveal], [I is] [E lian2he2guo2 da4wei2] [K zhen4jing1] [E the United Nation very] [K shock] ‘The revealing of Iraq’s secret bacteriological weapons shocked the United Nations.’ (31) [C heng2shan1 jin1tian1 ti2chu1 ci2cheng2], [I ling4] [C Yokoyama today propose resignation], [I cause] [E da4ban3 shi4min2] zhi4wei2 [K fen4nu4] [E Osaka citizen] very [K furious] ‘Yokoyama submitted his resignation today, [this] made the people of Osaka furious.’
Table 7.11 summarizes the generalized rules for detecting the cause events of the five primary emotions in Chinese. I identify two sets of rules: (1) the specific rules that apply to all emotional instances (Rules 1–11); (2) the general rules that apply to the emotional instances where no causes are found after applying the first set of rules (Rules 12–15). Constraints are set for certain rules to filter out incorrect causes.
Table 7.11 Linguistic rules for emotion cause detection Rule
Patterns and examples
1
(i) C(B/F) + I(F) + E(F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh after I in F (iii) C = the nearest (N) + (V) + (N) before I in B or F (iv) Constraint: K(F) cannot be followed by 是 shi4 ‘is’, 的是 deshi4 ‘is that’, 莫過於 mo4guo4yu1 ‘nothing is more … than…’ (30) [C伊拉克細菌武器秘密的曝光], [I使] [E聯合國大為] [K震驚] ‘The revealing of Iraq’s secret bacteriological weapons shocked the United Nations.’ (31) [C橫山今天提出辭呈], [I令] [E大阪市民] 至為 [K憤怒] ‘Yokoyama submitted his resignation today, (this) made the people of Osaka furious.’ (i) E(B/F) + II/IV/V/VI(B/F) + C(B/F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh before II/IV/V/VI in B or F (iii) C = the nearest (N) + (V) + (N) before K in B or F (iv) Constraint: the cue word有 you3 ‘to exist’ in Group III is excluded (15) [E蘭妮公主] [IV聽到] [C自己永遠不能恢復人形], [K傷心] 極了 ‘Princess Lanny was very sad to hear that she would never be turned back into a human being.’ (21) [E他] [V對] [C謠言散布之快] 感到 [K驚訝] ‘He was surprised by the fast spreading of rumors.’ (i) II/IV/V/VI (B) + C(B) + E(F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh before K in F (iii) C = the nearest (N) + (V) + (N) after II/IV/V/VI in B (iv) Constraint: the cue word有 you3 ‘to exist’ in Group III is excluded (16) [I發現] [C國王已經餓死在那了], [E大家] 都十分 [K傷心] (continued)
2
3
7.4 A Rule-Based System for Emotion Cause Detection
125
Table 7.11 (continued) Rule
4
5
6
7
8
9
Patterns and examples ‘Everyone was very sad to find that the King was starved to death.’ (22) [I對於] [C能有機會經營如此出色的俱樂部], [E我] 感到非常 [K快樂] ‘I am very happy for having the opportunity to run such a great club.’ (i) E(B/F) + K(F) + IV/VII(F) + C(F/A) (ii) E = a: the nearest Na/Nb/Nc/Nh before K in F; b: the first Na/Nb/Nc/Nh in B (iii) C = the nearest (N) + (V) + (N) after IV or VII in F or A (iv) Constraint: it only applies to change-of-state emotion verbs of HAPPINESS, SURPRISE and FEAR, such as 高興 gao1xing4 ‘becoming happy’, 驚訝 jing1ya4 ‘becoming surprised’, 害怕 hai4pa4 ‘becoming frightened’ (25) [E我] [K驚訝] [VII的是], [C居然有候選人以這種方式來合理化對女性參政的 歧視] ‘I was surprised that there were candidates who rationalized the discrimination against women in politics in this way.’ (26) [E我] 實在太 [K高興] [VII能] [C贏得這個獎] 讓我備感光榮! ‘I am very happy that I could win this award, it makes me feel honoured!’ (i) E(F) + K(F) + VI(A) + C(A) (ii) E = the nearest Na/Nb/Nc/Nh before K in F (iii) C = the nearest (N) + (V) + (N) after VI in A (27) 在這間酒吧中 [E我] 很 [K快樂], [VI因為] [C這裡有我的朋友] ‘I feel very happy in this pub, because I have friends here.’ (28) [E老鷹] 很 [K快樂], [VI因為] [C他的孩子會飛了] ‘The eagle was very happy, because his children could fly.’ (i) I(F) + E(F) + K(F) + C(F/A) (ii) E = the nearest Na/Nb/Nc/Nh after令ling4‘to cause’ in F (iii) C = the nearest N + V+N after K in F or A (iv) Constraint: K(F) must be followed by 是 shi4 ‘is’, 的是 deshi4 ‘is that’, 莫過於 mo4guo4yu3 ‘nothing is more … than…’ (29) 但 [I令] [E人] [K驚訝] 的是, [C那位老人居然不識字] ‘But what is surprising is that the old man is actually illiterate.’ (i) E(B/F) + 越 yue4 ‘the more’ C越 yue4 ‘the more’ K(F) (ii) E = the nearest Na/Nb/Nc/Nh before the first 越 yue4 ‘the more’ in B or F (iii) C = the verb in between the two 越 yue4 ‘the more’ in F (iv) Constraint: the cause cannot be 來 lai2 ‘become’ (30) [E好友] 越 [C聽] 越 [K生氣] ‘The more the friend heard, the angrier he was.’ (i) E(F) + K(F) + C(F) (ii) E = the nearest Na/Nb/Nc/Nh before K in F (iii) C = the nearest (N) + (V) + (N) after K in F (iv) Constraint: restricted to emotion verbs 高興, 興奮, 驚訝, 奇怪, 害怕, 恐懼, 恐慌, 心虛, 忿恨, 痛恨, 討厭, 厭倦, 厭惡, 厭煩, 傷感, 哀傷, 感傷; emotion verbs cannot be followed by 的, 得, 到, 般, 一樣, 下, 之下, 之際, 中, 不止, 不已, 起來, 與, 和, 又, 也, 或, 說, 道, 是 and the punctuation mark 頓號 (i.e. a mark in Chinese punctuation that sets off items in a series) (31) [E我們] 很 [K高興] [C創刊號終於發行了] ‘We are very happy that the first issue was finally released.’ (i) E(F) + IV(F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh before III in F (iii) C = IV + (an aspectual marker such as了 le, 完 wan2, 後 hou4 or 得 de) in F (continued)
7 Implementation and Verification: Automatic Detection …
126 Table 7.11 (continued) Rule
10
11
12
13
14
Patterns and examples (32) [E炸彈客] [C聽完], 露出一 臉 [K喪氣] 的表情 ‘The bomber was upset after listening (to…)’ (i) K(F) + E(F) + 的 de ‘POSS’(F) + C(F) (ii) E = the nearest Na/Nb/Nc/Nh after K in F (iii) C = the nearest N + V+N/N + 的 + N after 的 in F (iv) Constraint: it is only applied to SURPRISE (33) 一九七九年, 台灣爆發了 [K震驚] [E海內外] 的 [C「美麗島事件」] ‘In 1979, the ‘Formosa Incident’ broke out in Taiwan which was shocking within and abroad the country.’ (i) C(F) + K(F) + E(F) (ii) E = the nearest Na/Nb/Nc/Nh after K in F (iii) C = the nearest (N) + (V) + (N) before K in F (iv) Constraint: it only applies to SURPRISE (34) [C這個怪異的情景立刻] [K震驚] 了 [E全市民]… ‘The bizarre scene shocked the whole community at once…’ (i) E(B) + K(B) + III(B) + C(F) (ii) E = the nearest Na/Nb/Nc/Nh before K in B (iii) C = the nearest (N) + (V) + (N) after III in F (iv) Constraint: no verb between the keyword and III, and should be followed by a colon (35) [E我] [K心酸] 道:「可是[C他們好可憐]喔!」 ‘I said sadly, ‘But they are pitiful!’’ (i) III(B) + C(B) + E(F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh before K in F (iii) C = the nearest (N) + (V) + (N) after III in B (iv) Constraint: III should be followed by a colon (36) 那個男人 [III說]: 「[C像我這樣好條件,更好的女孩多的是]」。這次 [E喬依 絲] 哭的很[K傷心] ‘The man said, ‘There are many desirable girls for a man with good qualities like me.’ Joyce cried sadly.’ (i) C(B) + E(F) + K(F) (ii) E = the nearest Na/Nb/Nc/Nh before K in F (iii) C = the nearest (N) + (V) + (N) before K in B (iv) Constraint: it is used as the last rule for the remaining no-cause sentences (37) [C金帝完顏亮南侵], [E宋朝廷] [K震驚], 一時群情激憤 ‘Emperor Jin, Wanyan Liang, invaded the south, the Song court was shocked and enraged.’ (38) [C他們給我一條肉乾做獎品]。 [E父親] [K高興] 極了 (continued)
7.4 A Rule-Based System for Emotion Cause Detection
127
Table 7.11 (continued) Rule
Patterns and examples
‘They gave me a piece of dried meat as a prize. My father was very happy.’ (i) E(B) + C(B) + K(F) (ii) E = the first Na/Nb/Nc/Nh in B (iii) C = the nearest (N) + (V) + (N) before K in B (iv) Constraint: there should be no verb following the keyword in the target clause; used as the last rule for the remaining no-cause sentences (39) [E我] [C聽到音樂的那一剎那], 心裡很 [K快樂] ‘The moment that I heard the music, I felt very happy.’ (40) [E農夫] [C聽到這句話], 心裡非常 [K高興] ‘When the farmer heard this, he was very happy.’ Constraint: restricted to emotion verbs 高興 gao1xing4 ‘becoming happy’, 興奮 xing1fen4 ‘excited’, 驚訝 jing1ya4 ‘surprised’, 奇怪 qi2guai4 ‘surprising’, 害怕 hai4pa4 ‘frightened’, 恐懼 kong3ju4 ‘frightened’, 恐慌 kong3huang1 ‘frightened’, 心虛 xin1xu1 ‘frightened’, 忿恨 fen4hen4 ‘angry’, 痛恨 tong4hen4 ‘angry’, 討厭 tao3yan4 ‘dislike’, 厭倦 yan4juan4 ‘fed up’, 厭惡 yan4wu4 ‘detest’, 厭煩 yan4fan2 ‘fed up’, 傷感 shang1gan3 ‘sad’, 哀傷 ai1shang1 ‘sad’, 感傷 gan3shang1 ‘sad’ Emotion verbs cannot be followed by 的 de ‘to the extent that’, 得 de ‘to the extent that’, 到 dao4 ‘to the extent that’, 般 ban1 ‘as’, 一樣 yi2yang4 ‘the same’, 下 xia4 ‘under’, 之下 zhi1xia4 ‘under’, 之際 zhi1ji4 ‘when’, 中 zhong1 ‘in the course of’, 不止 bu4zhi3 ‘more than’, 不已 bu4yi3 ‘endlessly’, 起來 qi3lai2 ‘become’, 與 yu3 ‘and’, 和 he2 ‘and’, 又 you4 ‘and’, 也 ye3 ‘also’, 或 huo4 ‘or’, 說 shuo1 ‘say’, 道 dao4 ‘say’, 是 shi4 ‘is’ 15
Apart from the ones discussed in the previous section, there are other rules, such as Rule 4 which only applies to instances containing keywords of HAPPINESS, FEAR, and SURPRISE. The reason for this was given in Chap. 5, namely that only change-of-state emotion verbs of HAPPINESS, FEAR, and SURPRISE take causes as complements. The cause of Rule 7 cannot be 來 lai2 ‘come’ so as to avoid the frozen expression 越來越 yue4lai2yue4 ‘is becoming’; by data observation, Rule 9 is restricted to certain emotion verbs of HAPPINESS (高興 gao1xing4, 興奮 xing1fen4), SADNESS (傷感 shang1gan3, 哀傷 ai1shang1, 感傷 gan3shang1), FEAR (害怕 hai4pa4, 恐懼 kong3ju4, 恐慌 kong3huang1, 心虛 xin1xu1), ANGER (忿恨 fen4hen4, 痛恨 tong4hen4, 討厭 tao3yan4, 厭倦 yan4juan4, 厭惡 yan4wu4, 厭煩 yan4fan2), and SURPRISE (驚訝 jing1ya4, 奇怪 qi2guai4). Furthermore, the emotion verbs cannot be followed by words that are very likely to be followed by a cause: 的 de ‘DE’, 得 de ‘DE’, 到 dao4 ‘to the extent that’, 般 ban1 ‘as’, 一樣 yi2yang4 ‘same as’, 下 xia4 ‘under’, 之下 zhi1xia4 ‘under’, 之際 zhi1ji4 ‘while’, 中 zhong1 ‘while’, 不止 bu4zhi3 ‘more than’, 不已 bu4yi3 ‘endlessly’, 起來 qi3lai2 ‘getting’, 與 yu3 ‘and’, 和 he2 ‘and’, 又 you4 ‘again’, 也 ye3 ‘also’, 或 huo4 ‘or’, 說 shuo1 ‘say’, 道 dao4 ‘say’, 是 shi4 ‘is’ and the punctuation mark 頓號 dun4hao4 (i.e. a mark in Chinese punctuation that sets off items in a series). The details of the rules are presented and illustrated below with examples.
7 Implementation and Verification: Automatic Detection …
128
7.5
Experiment and Discussion
7.5.1
Overview
As mentioned in Sect. 7.3.3, 80% of the emotion cause corpus serves as development data and 20% as test data. The development data is the dataset the rule-based system is derived from, whereas the test data is the dataset that the system is evaluated on. The results of the test data determine the predictive power of the system. The mechanism of the experiment is illustrated in Fig. 7.2. The performance of the rule-based system is compared to the baseline. A baseline is an imaginary line used to provide a comparison for assessing the impact of the linguistic rules. The baseline is designed as follows: find a verb to the left of the keyword in question and consider the clause containing the verb as the cause of the emotion in question. The following sections describe the proposed evaluation scheme for assessing the rule-based system, report on the performance of the system, and discuss how the system can be improved.
7.5.2
Evaluation Metrics
An evaluation scheme is designed to assess the ability of the rule-based system to extract the cause of an emotion in context. Similarly to most NLP tasks, an overall evaluation is given in terms of three common measures: precision, recall, and F-score. Precision measures how accurately the system finds emotion causes and recall measures how fully the system finds emotion causes. The F-score provides the mean of the precision and recall scores.
Emotion Cause Corpus
80% Development Data
20% Test Data
Performance
feedback Rule-based System
Fig. 7.2 The architecture of the rule-based system
7.5 Experiment and Discussion
129
I specifically look into two phases of the performance of the cause recognition system. Phase 1 assesses the detection of an emotion co-occurrence with a cause; Phase 2 evaluates the recognition of the cause texts. Overall Evaluation The definitions of related metrics are presented in Fig. 7.3. For each emotion in a sentence, if neither the gold-standard file, nor the system file has a cause, both the precision and recall scores are 1; otherwise, precision and recall are calculated by the scoring method ScoreForTwoListOfCauses. As an emotion may have more than one cause, ScoreForTwoListOfCauses calculates the overlap scores between two lists of cause texts. Since emotion cause recognition is rather complicated, two relaxed string match scoring methods are selected to compare two cause texts,
Fig. 7.3 The definitions of metrics for cause detection
130
7 Implementation and Verification: Automatic Detection …
ScoreForTwoStrings: Relaxed Match 1 uses the minimal overlap between the gold-standard cause and the system cause. The system cause is considered as correct provided that there is at least one overlapping character; Relaxed Match 2 is more rigid in taking into account the overlapping text length during scoring. Phase 1: The Detection of Cause Occurrence The detection of cause occurrence is considered a preliminary task for emotion cause recognition and is compounded by the fact that neutral sentences are difficult to detect, as observed in Tokuhisa et al. (2008). For Phase 1, each emotion keyword in a sentence has a binary tag: Y (i.e. with a cause) or N (i.e. without a cause). Phase 2: The Detection of Causes The evaluation in Phase 2 is limited to the emotion keywords with a cause either in the gold-standard file or in the system file. The performance is calculated according to the Overall Evaluation scheme. The evaluation scheme above is illustrated by Example (32). There are two instances in (32), 742 and 934. Instance 742 contains one emotion keyword, i.e. 傷 感 shang1gan3 ‘sadness’ whose cause is expressed in the focus clause. Instance 934 contains two emotion keywords, one expressed with a cause i.e. 厭惡 yan4wu4 ‘anger’ and the other without it, i.e. 驚訝 jing1ya4 ‘surprise’. It is shown in (32) how the evaluation scheme calculates the performance of cause occurrence detection (i.e. Phase 1) as well as that of cause event detection (i.e. Phase 2) of each instance as well as the overall performance of the two instances using Relaxed Match 1 and 2. (32) An Example for Evaluation: Instance 742 742 Y 0/傷感/Sadness 他對我有所注意是在大學的畢業 晚會上。 我這人雖然愛湊熱鬧,不甘 寂寞,但因為 [*01e] 四年之後的分手 [*02e] 讓我挺 傷感 的。 所以就在 大家聚在教室裡聊天、玩牌的時候,我一個人蹲在門口兒,默默地為大家 煮咖啡。 742 Y 0/shang1gan3/Sadness He started to notice me at the university congregation. Although I love being in a crowd and do not like being alone, because of [*01e] the break-up four years later [*02e] it made me very sad . So when everyone gathered in the classroom, chatting and playing cards, I squatted at the door, silently making coffee for everyone.
7.5 Experiment and Discussion
131
Golden cause 1 (GC1): 36:42= ‘the breakup four years later’ System cause 1 (SC1): 36:45= ‘the breakup four years later made me quite…’ System cause 2 (SC2): 50:68= ‘so when everyone gathered in the classroom, chatting and playing cards, I squatted at the door, silently making coffee for everyone.’ Step 1: Relaxed Match 1:
Relaxed Match 2:
Step 2: Relaxed Match 2
Precision (GC1, SC1) =1; Precision (GC1, SC2) =0;
Recall (GC1, SC1) = 1; Recall (GC1, SC2) = 0;
Precision (GC1, SC1) =7/10= 0.7; Recall (GC1, SC1) = 1; Precision (GC1, SC2) =0; Recall (GC1, SC2) = 0;
ScoreForTwoListOfCauses ({GC1}, {SC1, SC2}) Precision = (0.7+0)/2 = 0.35 Recall = 1/1 = 1
Instance 934: 934 Y 0/厭惡/Anger,1/驚訝/Surprise 她有著灑脫的短髮, 卻給人一種好像水般的溫柔。 她有著 十分美麗的雙眼,卻不曾真正地去看過任何人的眼睛,怕被人看透,或者 厭惡 [*01e] 看透任何人 [*02e] 我們 驚訝 地說不出話來,那跟店長照片上 的她幾乎一模一樣,差的只是有沒有制服罷了! 店裡煞時安靜了下來,店長緩緩地回過頭,在與那位女學 生的眼神交會時停住了! 934 Y 0/yan4wu4/Anger,1/jing1ya4/Surprise She has short hair, free and easy, but it is as gentle as water. She has a pair of beautiful eyes, but she has never looked at other’s eyes, she is afraid of being understood, or detest of [*01e] understanding anyone [*02e]. We were so surprised that we were speechless, she was exactly the same as the one in the manager’s photo, the only difference is the uniforms! The store was silent all of a sudden, the manager turned his head slowly, and stopped when he saw the girl!
7 Implementation and Verification: Automatic Detection …
132 0/
yan4wu4/Anger: Golden cause 1 (GC1): 61:65= “understand anyone” System cause 1 (SC1): 41:47= “… anyone’s eyes” System cause 1 (SC2): 61:75= ‘… understand anyone, we are so surprised that we were speechless’
1/
jing1ya4/Surprise: Golden cause 1 (GC1) : N/A System cause 1 (SC1): 41:47= ‘… anyone’s eyes’ System cause 1 (SC2): 70:75= ‘… speechless’
Step 1: using Relaxed Match 2 0/ yan4wu4/Anger: Precision (GC1, SC1) =0; Recall (GC1, SC1) = 0; Precision (GC1, SC2) =5/15 = 0.33; Recall (GC1, SC2) = 1; jing1ya4/Surprise: Precision (GC1, SC1) =0; Recall (GC1, SC1) = 0; Precision (GC1, SC2) =0; Recall (GC1, SC2) = 0;
1/
Step 2: ScoreForTwoListOfCauses: 0/ yan4wu4/Fear: Precision = (0+0.33)/2 = 0.17; Recall = 1/1 = 1; 1/
jing1ya4/Surprise: Precision = 0; Recall = 0
Overall score for instances 742 and 934: 0:35 þ 0:17 þ 0 ¼ 0:17 3 1þ1þ0 RecallðGF; SFÞ ¼ ¼ 0:67 3
PrecisionðGF; SFÞ ¼
7.5.3
Results and Discussion
Table 7.12 shows the overall evaluation of performances. I find that the overall performances of the rule-based (RB) system have significantly improved using Relaxed Match 1 and Relaxed Match 2 by 19 and 18%, respectively compared to
7.5 Experiment and Discussion
133
Table 7.12 The overall performances
Baseline RB System
Relaxed Match 1 Precision Recall
F-score
Relaxed Match 2 Precision Recall
F-score
29.88 45.55
29.74 48.23
21.01 40.29
23.80 41.89
29.60 51.24
27.45 43.63
Table 7.13 The overall accuracy in Phase 1 Accuracy
Baseline
Rule-based system
79.56
79.37
Table 7.14 The detailed performances in Phase 1 Emotions
Baseline Precision
Recall
F-score
RB System Precision
Recall
F-score
With causes Without causes
99.42 4.39
79.74 66.67
88.50 8.23
96.87 13.16
80.85 52.63
88.14 21.05
the baseline. Although the overall performance of the system (48.23% for Relaxed Match 1 and 41.89% for Relaxed Match 2) is not very high, it marks a good start for emotion cause detection and extraction. Tables 7.13 and 7.14 show the performances of the baseline and the RB system in Phase 1. Table 7.13 shows the overall accuracy and Table 7.14 shows the detailed performances. In Table 7.13, I find that the RB system and the baseline have similar high accuracy scores. Yet, Table 7.14 shows that both systems achieve a high performance for emotions with a cause, but much worse for emotions without a cause. It is important to note that even though the naive baseline system has comparably higher performance with the RB system in judging whether there is a cause in context, the result is biased in two ways. First, as the corpus contains more than 80% of instances expressed with causes, a system which is biased toward detecting a cause, such as the baseline system, naturally performs well. In addition, once the actual cause is examined, I can see that the baseline actually detects a lot of false positives for emotion constructions without causes (the correctly identified causes make up no more than 4.39%). The RB system shows great promise in being able to deal with emotion constructions without explicit causes effectively and being able to detect the correct cause at least three times more frequently than the baseline. Table 7.15 shows the performances in Phase 2. Compared to the baseline, I find that the linguistic rules improve the performance of cause recognition using Relaxed Match 1 and 2 scoring by 20 and 20%, respectively. The 7% gap in F-score between Relaxed Match 1 and 2 also indicates that the rules can effectively locate the clause of a cause. The rather low performances of the baseline; on the other hand, show that most causes recognized by the baseline are wrong although the baseline effectively detects the cause occurrence, as indicated in Table 7.14.
7 Implementation and Verification: Automatic Detection …
134
Table 7.15 The detailed performances in Phase 2
Baseline RB System
7.5.4
Relaxed Match 1 Precision Recall
F-score
Relaxed Match 2 Precision Recall
F-score
29.37 45.16
32.46 52.00
20.38 39.61
25.35 44.85
36.27 61.68
33.54 51.68
Error Analysis
To evaluate the performance of each rule, I identify the accuracy (precision) and contribution (recall) of each rule. I discuss and explain the best and worst rules in terms of their precision and recall and how the set of rules can be improved overall. Table 7.16 summarizes the overall accuracy of each linguistic rule. In descending order, the top three accurate rules are Rules 7, 10, and 11; the bottom three accurate rules are Rules 13, 15, and 12. Table 7.17 shows the recall rate of the rules. The top three contributive rules are Rules 2, 15, and 14; the bottom three contributive rules are Rules 12, 10, 13. When we look at the two tables carefully, it is noticeable that rules with high precision often get low recall rate, and vice versa, except for Rules 12 and 13. As shown in Table 7.16, Rules 7 (100%), 10 (100%), and 11 (75%)9 achieve the best performance in terms of precision among the linguistic rules. These rules are all highly specific in that Rule 7 targets the specific pattern “越 yue4 ‘the more’ C 越 yue4 ‘the more’ K”; while Rules 10 and 11 are only applied to emotion constructions of SURPRISE. The specificity of these rules results in high precision yet low recall, as indicated in Table 7.17. The rules that obtained the lowest precision are Rules 13 (33.3%), 15 (30.7%), and 12 (26.7%). Rules 12 and 13 identify the first clause uttered by the experiencer as the cause event. However, in the real data, the cause events are usually not the first, but the second or third clause in which the first clause usually gives the background knowledge of the scene, as indicated in (33) and (34). (33) 736 Y 0/傷感/Sadness 頓時,所有的星星都不見了,天空 中只有一輪孤單的月兒,陪伴黑暗的夜幕。 園裡的花兒嘰嘰呱呱的談著,清香的曇花一面吐露沁人 心脾的芬芳,一面 傷感 的說:我雖然 清新脫俗,可是, [*01e] 我只能在夜晚開放,在黑暗中孤芳自賞,沒有人會留 意我 [*02e],哎! 說著,曇花就枯萎了。 736 Y 0/shang1gan3/Sadness Suddenly, all the stars disappeared, there was only a lonely moon in the sky, staying with the dark night. The flowers in the garden were talking 9
Similar results are shown in Relaxed Match 1 and 2; therefore, unless otherwise specified, only the results of Relaxed Match 1 are given for comparison.
7.5 Experiment and Discussion
135
Table 7.16 The accuracy of each rule Rank
Rule
Relaxed Match 1 (%)
Rule
Relaxed Match 2 (%)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
7 10 11 1 4 6 8 9 2 5 14 3 13 15 12
100 100 75 74.713 66.667 63.889 61.585 61.538 51.290 38.235 36.792 35.514 33.333 30.693 26.667
7 10 11 1 4 9 8 6 2 5 3 14 15 13 12
100 100 71.703 65.227 58.230 56.731 54.177 50.179 42.322 35.539 32.741 31.851 28.509 27.500 23.548
Table 7.17 The contribution of each rule Rank
Rule
Relaxed Match 1 (%)
Rule
Relaxed Match 2 (%)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
2 15 14 3 8 1 6 5 4 9 11 7 12 10 13
18.308 14.195 13.384 13.152 11.645 7.532 2.665 1.506 1.159 0.811 0.695 0.463 0.463 0.348 0.232
2 14 8 3 15 1 6 5 9 11 4 7 10 12 13
14.320 11.610 11.109 9.370 8.794 6.436 2.111 1.232 0.758 0.695 0.643 0.463 0.348 0.340 0.180
non-stop, while the delicate night-blooming cereus was showing the refreshing fragrance, it sadly said, “Although I am natural and refined, but [*01e] I can only bloom at night, and indulge in self-admiration in the dark, nobody will pay attention to me [*02e], ah! After saying that, night-blooming cereus wilted.
7 Implementation and Verification: Automatic Detection …
136 golden =
0/Sadness cause:82:106= ‘I can only bloom at night, and indulge in self-admiration in the dark, nobody will pay attention to me’
system =
rule10 0: cause=71:77=
shang1gan3 ‘sadness’ ‘Although I am natural and refined’
(34) 13622 Y 0/驚奇/Surprise 他 提心弔膽 地 去 見 考官 , 並 且 老老實實 地 承認 自己 太 粗心 了 , 把 太 字 少 寫 了 一 點 。 考官 卻 非常 < emotionword id=0> 驚 奇 地 說 : 我 看 得 很 仔細 , [*01e] 你 的 試卷 上 並 沒有 別字 [*02e] 啊 。 說 著 , 他 把 那 份 試卷 取出來 , 果然 找 不 出 什麼 錯字 。 13622 Y 0/jing1qi2/Surprise He approached the examiner with trepidation, and acknowledged honestly that he was too careless, in writing the word “tai4” incorrectly by leaving out a stroke. But the examiner said surprisingly , “I looked carefully, [*01e] there is no wrongly written word in your examination paper [*02e]. As he was saying this, he took out the examination paper, and true to his word there were no wrongly written characters.
golden =
0/Surprise cause:53:62=
system =
rule10 0: cause=46:51=
‘there is no wrongly written word in your examination paper’ jing1qi2 ‘becoming surprised’ ‘I looked carefully’
It is also found that those uttered by the experiencer accompanying emotions, rather than signifying the cause, comment on the overall situation as in (35) or the reaction of the emotion, i.e. the event resulting from the emotion, as in (36). (35) 11640 Y 0/生氣/Anger [*01e] 螃蟹寶寶把事情的經過告 訴了粟子、蜜蜂和石臼三個好朋友 [*02e]。 三個好朋友 [*01e] 聽了 [*02e] 非常非常的 生氣 說:「這隻猴子實在太可惡了! 不但狡猾而且又那麼的貪心! 11640 Y 0/sheng1qi4/Anger [*01e] The baby crab told the three good friends, millet, honeybee and stone mortar, what was happening
7.5 Experiment and Discussion
137
[*02e]. [*01e] After hearing this [*02e], the three good friends were very very angry and said, “This monkey is really hateful! [It] is cunning and greedy! golden =
0/Anger cause:33:33=
le ‘ASP’
0/Anger cause:1:26= ‘The baby crab told the three good friends, millet, honeybee and stone mortar, what was happening’ system =
rule10 0: cause=44:54=
sheng1qi4 ‘angry’ ‘This monkey is really hateful’
(36) 14335 Y 0/高興/Happiness [*01e] 大獅子跟老牛說:我要 買一頂帽子來遮陽 [*02e]。 老牛 高興 的說:我馬上找給你。 老牛把店裡所有的大帽子都搬出來,偏 偏找不出適合獅子的大帽子,他很不好意思的說:獅子先生,真不好意思,我 沒有那麼大的帽子,你訂做一頂吧! 14335 Y 0/gao1xing4/Happiness [*01e] The big lion said to the old ox, “I want to get a hat to protect against the sun [*02e]”. The old ox said happily , “I’ll find one for you immediately.” The old ox took out all the big hats, but could not find one that fitted the big lion. He said embarrassedly, “Mr. Lion, I am really sorry that I don’t have such a big hat for you, you may want to get one tailor-made?
golden =
0/Happiness cause: 0:17= ‘The big lion said to the old ox, “I want to get a hat to protect me against the sun’
system =
rule10 0: gao1xing4 ‘becoming happy’ cause= 26:31= ‘I’ll find one for you immediately’
To avoid having to extract the non-causes, Rules 12 and 13 should be reconsidered as part of the second set of rules, but before the last set of rules, i.e. Rules 14 and 15. However, when considering the low recall rate of Rules 12 (0.463%) and 13
7 Implementation and Verification: Automatic Detection …
138
(0.232%), the two rules can be removed altogether from the system in order to achieve a better performance. It is, however, not surprising that Rules 14 (36.792%) and 15 (30.693%) achieve rather low precision as they assume that the causes of all the remaining instances without causes detected by Rules 1–13 are positioned in the clause before the focus clause. Despite the low precision rate, they achieve a relatively high recall rate, i.e. 13.384% for Rule 14 and 14.195% for Rule 15. This proves that the assumption that emotion causes very often occur before the emotion keyword is feasible. Overall, Rule 2 has achieved the highest F-score performance (26.948%); whereas Rule 13 the lowest (0.460%). Rule 2 makes use of four groups of linguistic cues, including reported verbs, epistemic markers, prepositions, and conjunctions, which has yielded the highest recall and a satisfactory score of precision. It is noted that the rule performance based on Relaxed Match 1 and 2 are similar, yet that of Relaxed Match 2 is comparatively lower. For Relaxed Match 2, the precision measures how accurate the extracted cause event is when compared with the golden cause event. Since cause events are assumed to be clause-based, i.e. they take the whole clause as the cause event, the extracted cause events usually include unnecessary parts, such as the experiencer, adverbs, etc. Consider (37): (37)
shuo1dao4 [*01e] bian4hua4 [*02e] wo3men jiu4 hen3 speaking of [*01e] changes[*02e] 1.PL then very shang1gan3 sad ‘We are very sad when speaking of changes.’
The golden cause event of 傷感 shang1gan3 ‘sad’ in (37) is 變化 bian4hua4 ‘changes’, being marked by the cue word 說到 shuo1dao4 ‘speaking of’. However, the cause event extracted by the RB system using Rule 2 is longer than that of the golden one. When instance (37) comes into the system, the system first locates the emotion keyword 傷感 shang1gan3 ‘sad’ in a clause called target clause. Then Rule 2 detects the cue word 說到 shuo1dao4 ‘speaking of’ in the focus clause. The clause after the cue word is taken as the cause event minus the emotion keyword, i.e. 變化我們就很 bian4hua4 wo3men jiu4 hen3 ‘…changes then we are very…’, as shown in (38).
(38) rule2 0:
shang1gan3 ‘sadness’ cause=30:35= bian4hua4 wo3men jiu4 hen3 changes 1.PL then very ‘… changes then we are very…’
7.5 Experiment and Discussion
139
As is observable, instead of the golden cause event, the experiencer (i.e. 我們 wo3men ‘1.PL’) and adverbs (i.e. 就 jiu4 ‘then’ and 很 hen3 ‘very’) are also taken as part of the cause event. This is due to the assumption on the part of the RB system that the whole clause is loosely taken as a cause event, which lowers the precision rate of Relaxed Match 2. In addition, as cause events are assumed to be a single clause, any cause event that exceeds the length of a clause will lower the precision rate of Relaxed Match 2. This is illustrated in (39). (39) [*01e][*11e] na4xie1 gao4mi4 jian1shi4 bu3zhuo1 kao3da3 tai2wan1ren2 de, monitor capture torture Taiwanese POSS, [*01e][*11e] those inform dou1 shi4 tai2wan1ren2 zi4ji3[*02e][*12e], zen3 bu2 ling4 ren2 all is Taiwanese oneself[*02e][*12e], how not cause people fen4nu4 bei1ai1 sad angry ‘Those who inform, monitor, capture, and torture the people of Taiwan are the Taiwanese people themselves, how can people not get angry and sad!’ golden cause events 0/Sadness cause:30:54= na4xie1 gao4mi4 jian1shi4 bu3zhuo1 kao3da3 tai2wan1ren2 de, dou1 shi4 monitor capture torture Taiwanese POSS, all is those inform tai2wan1ren2 zi4ji3 Taiwanese oneself ‘Those who inform, monitor, capture, and torture the people of Taiwan are the Taiwanese people themselves’ 1/Anger cause:30:54= na4xie1 gao4mi4 jian1shi4 bu3zhuo1 kao3da3 tai2wan1ren2 de, monitor capture torture Taiwanese POSS, those inform dou1 shi4 tai2wan1ren2 zi4ji3 all is Taiwanese oneself ‘Those who inform, monitor, capture, and torture the people of Taiwan are the Taiwanese people themselves’ system cause events bei1ai1 ‘sad’ rule 1 0: cause=48:54= dou1 shi4 tai2wan1ren2 zi4ji3 all is Taiwanese oneself ‘… are the Taiwanese people themselves’ rule 1 1: cause=48:54= dou1 shi4 tai2wan1ren2 zi4ji3 all is Taiwanese oneself … are the Taiwanese people themselves’
7 Implementation and Verification: Automatic Detection …
140
7.6
Summary
Cause detection is a new research area in emotion computing, despite the fact that most theories of emotion treat recognition of a triggering cause event as an integral part of an emotion. In emotion computing, the cause event and emotion correlation is a fertile ground for extraction and entailment of new information. As a first step towards fully automatic inference of cause-emotion correlation, I propose a text-driven, rule-based approach to emotion cause detection. First of all, an annotated emotion cause corpus is constructed based on a proposed annotation scheme. By analyzing the corpus data, I identify seven groups of linguistic cues and generate two sets of linguistic rules for detecting emotion causes. With the help of these linguistic rules, I then develop a rule-based system for emotion cause detection. In addition, I propose an evaluation scheme with two phases for performance assessment. Experiments show that the rule-based system achieves a promising performance for cause occurrence detection as well as cause event detection. The proposed rule-based system not only effectively detects emotion causes, but also verifies the significant contribution of linguistic analysis and modeling. Taking the above into account, the rule-based system was further implemented in Chen et al. (2010). In this paper, we have developed a multi-label approach to automatically detect emotion causes based on the rule-based cause detection model proposed in the previous chapter. The multi-label model has succeeded in identifying multi-clause cause events, as well as in capturing long-distance information to facilitate emotion cause detection. With the help of the linguistic analysis proposed in this book, we created a set of automatically generalized patterns in the course of feature extraction. Incorporating my rule-based analyses into Chen et al. (2010) automatically generalized patterns, general cause expressions or specific constructions for emotion causes can effectively be extracted. We show that the multi-label system based on linguistic rules achieves a better performance than a baseline model. The implementation of the rule-based heuristic system as well as the multi-label learning system attests to the validity of the model proposed in this study for emotion detection and classification. I believe that the current work can be applied to many real-world applications and future research based on cause-event relation, such as detection of implicit emotions or causes, as well as prediction of public opinion based on cause events.
7.7
Conclusion
The purpose of this book is to investigate Chinese emotion detection and classification with specific focus on causal relations based on the assumption that cause events are the most concrete components of emotions. From Aristotle’s emotion theory in the Rhetoric, Descartes’ The Passion of the Souls (1649), Spinoza’s Ethics (1675), and James’ What is an Emotion (1884), to Plutchik’s Circumplex Models of
7.7 Conclusion
141
Emotions (1980), Wierzbicka’s Natural Semantic Metalanguage Model (1992), and Damasio’s The Feeling of What Happens (1999), attempts were made to define the concept of emotion. These models are by no means competing to replace one another, but provide different perspectives in describing emotions. However, they neither offer a concrete and comprehensive definition, nor can they be used as the representational foundation for emotion analysis and processing. While it is often concluded that emotion concepts cannot be defined at all, this book argues that cause events, being an indispensable part of emotions, provide a new dimension of how emotions should be defined and classified. Blending the insights from two prominent theories, the Natural Semantic Metalanguage (Wierzbicka 1992) and the Generative Lexicon (Pustejovsky 1995), I proposed a linguistic model of emotion combining event representation and emotion classification. I see emotion as a type of event which is triggered by actual or perceived events, i.e. cause events. With this assumption, cause events were examined in terms of two dimensions, namely transitivity and epistemicity. Finally, all the threads of the linguistic analyses come together in Chap. 5 where an integrated NSM model is developed to further our understanding of emotions as events. The integrated NSM model emphasizes the inseparable relation between emotions and events where emotion is decoded as an event type and is triggered by an event (a cause event) and elicits another event (an elicited event). In other words, an emotion construction typically comprises a series of events, including cause events, an emotional state, and elicited events. I specifically focused on the link between emotion and cause events, incorporating previous linguistic findings of emotion cause events into the proposed integrated NSM representation. By doing so, a set of linguistic criteria for emotion classification and representation was established. This not only provides deep linguistic criteria of emotion cause events for emotion classification, but also offers an event-based account of emotion classification. This account of emotion-cause interaction enriches the existing emotion frameworks, contributing particularly to the development of emotion detection and classification. It also lays the groundwork for further research with reference to emotion analysis. The theoretical account of emotion and cause formed the basis for creating an emotion cause corpus. The emotion cause corpus was annotated in terms of five primary emotions and their corresponding cause events. Based on the corpus analysis, seven groups of cause event markers were identified. Such a cause corpus offers a valuable resource for both linguistic analysis as well as natural language processing of emotion and causes. For linguistic analysis, it provides an empirical basis for the development of linguistic accounts of emotion; for natural language processing, it serves as the training and the test data for an automatic system of emotion classification. Furthermore, since manual detection of cause events is labour-intensive and time-consuming, the current emotion cause corpus can be used to produce an automatic extraction system for emotion cause events with machine learning methods. With the help of the emotion cause corpus, the proposed linguistic framework of emotion was then modeled for computational implementation. As a first step toward
142
7 Implementation and Verification: Automatic Detection …
a fully automatic inference of cause-emotion correlation, a text-driven, rule-based approach to emotion cause detection was proposed earlier in this chapter. Two sets of linguistic rules were generalized to automatically detect emotion causes. Experiments show that the system achieves a promising performance for cause occurrence detection as well as cause event detection. This proves that a rule-based system not only effectively detects emotion causes, but also reinforces the significant contribution of linguistic analysis and modeling. The linguistic-driven rule-based heuristic system can be viewed as a first step towards inferences and entailments of new information based on cause-event relations. It can be applied to many real world applications. For instance, the detected emotion cause events can serve as clues for the identification of implicit emotions which are not indicated by emotion keywords. A stochastic model can be developed for identifying and classifying emotions and events automatically. The possibility of linking this line of research to economic forecasting and the design of actual products can also be explored. I believe that the current study will have implications not only for the linguistic theory of emotions, but also for the linguistic account of events as well as automatic detection and classification of emotion in language technology.
Appendix
Representative examples of statements on primary emotions (Turner 2000) JohnsonLaird/ Oatley (1992)
Emde (1980)
Happiness
Joy
Fear
Fear
Anger
Sadness
Panksepp (1982)
Sroufe (1979)
Turner (1996)
Trevarthen (1984)
Pleasure
Happiness
Happiness
Fear Panic
Fear
Fear
Fear
Anger
Rage
Anger
Anger
Anger
Sadness
Sorrow Loneliness Grief
Sadness
Sadness
Surprise Disgust
Arnold (1960)
Osgood (1966)
Darwin (1872)
Izard (1977, 1992)
Joy Quiet pleasure
Pleasure Joy affection
Enjoyment
Fight
Fear Anxiety
Terror
Fear
Fight Defensive Aggressive
Anger
Anger
Anger Contempt
Sorrow
Surprise
Amazement
Disgust
Astonishment
Disgust
Surprise Disgust
Shame Shyness
Shame Shyness
Distress
Distress
Guilt Interest
Guilt Expectancy
Approach
Interest Expectancy
Inhibition
Boredom
Interest Pain
Ekman (1984)
Epstein (1984)
Arieti (1970)
Fromme/ O’Brien (1982)
Plutchik (1980)
Scott (1980)
Fehr/ Russell (1984)
Gray (1982)
Kemper (1987)
Malatesta/ Haviland (1982)
Happiness
Joy Love
Satisfaction
Joy Elation Satisfaction
Joy
Pleasure Love
Happiness Love
Hope
Satisfaction
Joy
Fear
Fear
Fear Tension
Fear
Fear
Fear Anxiety
Fear
Anxiety
Fear
Fear
Anger
Anger
Rage
Anger
Anger
Anger
Anger
Anger
Anger
Anger
Sadness
Sadness
Unpleasure
Grief Resignation
Sadness
Loneliness
Sadness
Sadness
Depression
Sadness
Surprise
Shock
Disgust
Surprise Disgust Anticipation
Curiosity
Interest Pain
Appetite
Acceptance
© Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3
Brownflash Knitbrow
143
References
Ahmad, K. (ed.). 2008. In Proceedings of the Workshop on Sentiment Analysis: Emotion, Metaphor, Ontology and Terminology (EMOT-08). In Association with LREC-08, Marrakech, Morocco, May 27. Alm, C.O., D. Roth, and R. Sproat. 2005. Emotions from Text: Machine Learning for Text based Emotion Prediction. In Proceedings of HLT/EMNLP. Vancouver, October. Alm, C.O. 2009. Affect in Text and Speech. Saarbrucken: VDM Verlag. Arieti, S. 1970. Cognition and Feeling. In Feelings and Emotions: The Loyola Symposium on Feelings and Emotions, ed. M.B. Arnold. New York: Academic Press. Arnold, M.B. 1960. Emotion and Personality. New York: Columbia University Press. Barańczak, S. 1990. Breathing under Water and other East European Essays. Cambridge, MA: Harvard University Press. Berkowitz, L. Anger. 1999. In Handbook of Cognition and Emotion, ed. T. Dalgleish and M. J. Power. New York: Wiley, 411–428. Besnier, N. 1990. Language and Affect. Annual Review of Anthropology 19: 419–451. Caffi, C., and R.W. Janney. 1994. Toward a Pragmatics of Emotive Communication. Journal of pragmatics 22 (3): 325–373. Cambria, E., A. Hussain, C. Havasi, and C. Eckl. 2009. Affective Space: Blending Common Sense and Affective Knowledge to Perform Emotive Reasoning. In Proceedings of CAEPIA, 32–41. Cannon, W.B. 1927. The James-Lange Theory of Emotions: A Critical Examination and an Alternative Theory. American Journal of Psychology 39: 106–124. Carnie, A. 2006. Syntax: A Generative Introduction. Blackwell. Chang, L. L., K.-J. Chen, and C.-R. Huang. 2000. Alternation across Semantic Field: A Study of Mandarin Verbs of Emotion. In Special Issue on Chinese Verbal Semantics. Computational Linguistics and Chinese Language Processing, vol. 5(1), ed. Yung-O Biq, 61–80. Chaumartin, F.-R. 2007. A Knowledge Based System for Headline Sentiment Tagging. In Proceedings of the 4th International Workshop on Semantic Evaluations. Chen, K.-J., and C.-R. Huang. 1990. Information-based Case Grammar. In Proceedings of the 13th International Conference on Computational Linguistics (COLING 1990), vol. 2, 54–59. Helsinki, Finland. August 20–25. Chen, Y., S.Y.M. Lee, and C.-R. Huang. 2009a. A Cognitive-based Annotation System for Emotion Computing. In Proceedings of the Third Linguistic Annotation Workshop (The LAW III). Chen, Y., S.Y.M. Lee, and C.-R. Huang. 2009b. Are Emotions Enumerable or Decomposable? And Its Implications for Emotion Processing. In Proceedings of The 23rd Pacific Asia Conference on Language, Information and Computation (PACLIC 23). Chen, Y., S.Y.M. Lee, and C.-R. Huang. 2010. Emotion Cause Detection with Linguistic Constructions. In Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010).
© Springer Nature Singapore Pte Ltd. 2019 S.Y.M. Lee, Emotion and Cause, Studies in East Asian Linguistics, https://doi.org/10.1007/978-981-10-6194-3
145
146
References
Chuang, Z. J., and C.H. Wu. 2002. Emotion Recognition from Textual Input Using an Emotional Semantic Network. In Proceedings of the International Symposium on Chinese Spoken Language Processing (Denver, CO). CKIP. 1995. The Content and Illustration of Sinica Corpus of Academia Sinica. Technical Report No. 95-102, Institute of Information Science, Academia Sinica. Comrie, B. 1978. Ergativity. In Syntactic Typology, ed. W.P. Lehmann, 329–394. Austin: University of Texas Press. Cruse, D.A. 1973. Some Thoughts on Agentivity. Journal of Linguistics 9: 11–23. Damasio, A. 2003. Looking for Spinoza: Joy, Sorrow, and the Feeling Brain. Orlando: Harcourt Harvest Books. Damasio, A.R. 1994. Descartes’ Error: Emotion, Reason, and the Human Brain. New York: G. P. Putnam’s Sons. Damasio, A.R. 1999. The Feeling of What Happens: Body and Emotion in the Making of Consciousness. London: Heinemann. Danisman, T., and A. Alpkocak. 2008. Feeler: Emotion Classification of Text Using Vector Space Model. In AISB 2008 Convention, Communication, Interaction and Social Intelligence, vol. 2, 53–59. Aberdeen, Scotland, April 1–4. Darwin, C. 1859. On the Origin of Species by Means of Natural Selection, or the Preservation of Favoured Races in the Struggle for Life. London: John Murray. Darwin, C. 1872. The Expression of Emotion in Man and Animals. London: Watts. Davidson, D. 1967. The Logical Form of Action Sentences. Essays on Actions and Events, 105– 148. Oxford: Clarendon Press. Descartes, R. 1649. The Passions of the Soul. In The Philosophical Writings of Descartes, vol. 1, ed. J. Cottingham et al., 325–404. Desmet, P.M.A. 2002. Designing Emotion. Doctoral Thesis. TU-Delft. Desmet, P.M.A., and P. Hekkert. 2007. Framework of Product Experience. International Journal of Design. 1 (1): 57–66. Doi, T. 2004. Understanding Amae: The Japanese Concet of Need-Love. Global Oriental. Dong, Z., and Q. Dong. 2000. Introduction to HowNet. http://www.keenage.com. Dowty, D. 1991. Thematic Proto-Roles and Argument Selection. Language 67 (3): 547–619. Ekman, P. 1984. Expression and the Nature of Emotion. In Approaches to Emotion, ed. K. Scherer, and P. Ekman, 319–343. Hillsdale, N.J.: Lawrence Erlbaum. Elliot, C.D. 1992. The Affective Reasoner: A Process Model of Emotions in a Multi-Agent System. Doctoral dissertation, Institute for the Learning Sciences, Northwestern University. Emde, R.N. 1980. Levels of Meaning for Infant Emotions: A Biosocial View. In Development of Cognition, Affect, and Social Relations: The Minnesota Symposium of Child Psychology, ed. W.A. Collins. Hillsdale, N.J.: Lawrence Erlbaum. Epstein, S. 1984. Controversial Issues in Emotion Theory. In Review of Personality and Social Psychology, ed. P. Shaver. Beverly Hills, Calif: Sage. Fehr, B., and J.A. Russell. 1984. Concept of Emotion Viewed from a Prototype Perspective. Journal of Experimental Psychology 113: 464–486. Fillmore, C. 1968. The Case for Case. In Universals in Linguistic Theory, ed. Bach and Harms. New York: Holt, Rinehart, and Winston, 1–88. Fillmore, C. 1976. Frame Semantics and the Nature of Language. In Annals of the New York Academy of Sciences: Conference on the Origin and Development of Language and Speech, vol. 280, 20–32. Fillmore, C., and B. Atkins. 1992. Toward a Frame-based Lexicon: The Semantic of Risk and Its Neighbors. In Frames, Fields, and Contrast, ed. A. Lehrer, and E. Kittay. Hillsdale: Lawrence Erlbaum Associates. Frijda, N.H. 1987. Emotion, Cognitive Structure, and Action Tendency. Cognition and Emotion 1: 115–143. Fromme, D.K., and C.S. O’Brien. 1982. A Dimensional Approach to the Circular Ordering of the Emotions. Motivation and Emotion 6 (4): 337–363.
References
147
Fujino, A., H. Isozaki, and J. Suzuki. 2008. Multi-label Text Categorization with Model Combination based on F1-score Maximization. In Proceedings of the International Joint Conference on Natural Language Processing. Hyderabad, India. January 7–12. Gao, H. 2001. The Physical Foundation of the Patterning of Physical Action Verbs: A Study of Chinese Verbs. Lund, Sweden: Lund University. Givón, T. 1993. English Grammar: A Function-Based Introduction, 2 Vols. Amsterdam: John Benjamins. Givón, T. 2009. The Ontogeny of Complex Verb Phrases: How Children Learn a Negotiate Fact and Desire. In Syntactic Complexity: Diachrony, Acquisition, Neuro-cognition, Evolution, ed. T. Givon, and M. Shibatani. Amsterdam: John Benjamins. Gray, J.A. 1982. The Neuropsychology of Anxiety: An Enquiry into the Function of the Septo-hippocampal System. New York: Oxford University Press. Harkins, J., and A. Wierzbicka (eds.). 2001. Emotions in Crosslinguistic Perspective. Berlin, New York: Mouton de Gruyter. Hopper, P.J., and S.A. Thompson. 1980. Transitivity in Grammar and Discourse. Language 56: 251–299. Huang, C. R. 2006. 大數與求真:如何以十億字語料庫進行語言分析與研究. Invited Speech, The 4th Annual Meeting of Society of Chinese Teachers in Taiwan. Kaohsiung, Taiwan, October 28. Izard, C.E. 1977. Human Emotions. New York: Plenum Press. Izard, C.E. 1993. Four Systems for Emotion Activation: Cognitive and Noncognitive Processes. Psychological Review. 100: 68. Jackendoff, R.S. 1972. Semantic Interpretation in Generative Grammar. Cambridge, MA: MIT Press. James, W. 1884. What is an Emotion? Mind 9 (34): 188–205. Ji, S., L. Tang, S. Yu, and J. Ye. 2008. Extracting Shared Subspace for Multi-label Classification. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Las Vegas, Nevada, USA. Johnson-Laird, P.N., and K. Oatley. 1989. The Meaning of Emotions: Analysis of a Semantic Field. Cognition and Emotion 3: 81–123. Johnson-Laird, P.N., and K. Oatley. 1992. Basic Emotions, Rationality, and Folk Theory. Cognition & Emotion 6 (3–4): 201–223. Keltner, D., K. Oatley, and J.M. Jenkins. 2014. Understanding Emotions. Hoboken: Wiley. Kemper, T.D. 1987. How Many Emotions are There? American Journal of Sociology 93: 263– 289. Kövecses, Z. 2000. Metaphor and Emotion: Language, Culture and Body in Human Feeling. Cambridge: Cambridge University Press. Kozareva, Z., B. Navarro, S. Vazquez, and A. Nibtoyo. 2007. UA-ZBSA: A Headline Emotion Classification through Web Information. In Proceedings of the 4th International Workshop on Semantic Evaluations. Kripke, S. 1972. Naming and Necessity. In Semantics of Natural Language, ed. D. Davidson and G. Harman, 253–355, 763–769. Dordrecht: Reidel. Lakoff, G. 1977. Linguistic Gestalts. In Papers from the Thirteenth Regional Meeting Chicago Linguistic Society, ed. Woodford A. Beach, S.E. Fox, and S. Philosoph, 236–286. Lee, S.Y.M., Y. Chen, and C.-R. Huang. 2009. Cause Event Representations for Happiness and Surprise. In Proceedings of the 23rd Pacific Asia Conference on Language, Information and Computation (PACLIC 23). Lee, S.Y.M., Y. Chen, and C.-R. Huang. 2010b. A Text-driven Rule-based System for Emotion Cause Detection. In Proceedings of NAACL-HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text. Los Angeles, CA. June 5–6. Lee, S.Y.M., Y. Chen, S.S. Li, and C.-R. Huang. 2010a. Emotion Cause Corpus: Corpus Construction and Analysis. In Proceedings of the 7th International Conference on Language Resources and Evaluation (LREC 2010). Malta, May 19–21.
148
References
Lee, Sophia Y.M., Y. Chen, C.-R. Huang, and S. Li. 2013. Detecting Emotion Causes with a Linguistic Rule-based Approach. Computational Intelligence, Special Issues on Computational Approaches to Analysis of Emotion in Text 29 (3): 390–416. Li, C.N., and S.A. Thompson. 1981. Mandarin Chinese: A Functional Reference Grammar. University of California Press. Li, W., and H. Xu. 2014. Text-Based Emotion Classification Using Emotion Cause Extraction. Expert Systems with Applications 41 (4): 1742–1749. Liu, H., H. Lieberman, and T. Selker. 2003. A Model of Textual Affect Sensing using Real-World Knowledge. In Proceedings of the 2003 International Conference on Intelligent User Interfaces, IUI 2003, 125–132. ACM Press. Liu, M.-C. 2002. Mandarin Verbal Semantics: A Corpus-based Approach, 2nd ed. Taipei: Crane. Liu, M.-C. 2009. Emotion Verbs in Mandarin: A Lexical-Constructional Approach. In Proceedings of Chinese Lexical Semantics Workshop 2009. Yantai. July 27–30. Liu, M.-C., and S.-M. Hong. 2008. Mandarin Emotion Verbs: A Frame-based Analysis. Journal of Chinese Language and Computing 18 (3): 107–119. Liu, M.-C., C.-Y. Hu, and P.-Y. Liao. 2009. Bridging Different Predicational Frames: The Semantic Shift of Mandarin Emotion Verbs “ke+V”. In Proceedings of Chinese Lexical Semantics Workshop 2009. Yantai, July 27–30. López, J.M., R. Gil, R. García, I. Cearreta, and N. Garay. 2008. Towards an Ontology for Describing Emotions. Lecture Notes in Computer Science. 5288: 96–104. Lutz, C.A. 1988. Unnatural Emotions: Everyday Sentiments on a Micronesian Atoll & their Challenge to Western Theory. Chicago and London: University of Chicago Press. Ma, W.Y., and C.-R. Huang. 2006. Uniform and Effective Tagging of a Heterogeneous Giga-word Corpus. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC2006). Genoa, Itlay. Malatesta, C.Z., and J.M. Haviland. 1982. Learning Display Rules: The Socialization of Emotion Expression in Infancy. Child Development 53: 991–1003. Masum, S.M., H. Prendinger, and M. Ishizuka. 2007. Emotion Sensitive News Agent: An Approach Towards User Centric Emotion Sensing from the News. In Proceedings of the IEEE/WIC/ACM International Conference on Web Intelligence. McCallum, A. 1999. Multi-label Text Classification with a Mixture Model Trained by EM. In Proceedings of AAAI’99 Workshop on Text Learning. Mihalcea, R., and H. Liu. 2006. A Corpus-based Approach to Finding Happiness. In Proceedings of the AAAI Spring Symposium on Computational Approaches to Weblogs. Mishne, Gilad. 2005. Experiments with Mood Classification in Blog Posts. In Proceedings of Style 2005—The 1st Workshop on Stylistic Analysis of Text for Information Access, at SIGIR 2005. Moens, M., and M. Steedman. 1988. Temporal Ontology and Temporal Reference. Computational Linguistics 14: 15–28. Mohammad, S., and P. Turney. 2013. Crowdsourcing a Word-emotion Association Lexicon. Computational Intelligence 29 (3): 435–465. Næss, Åshid. 2007. Prototypical Transitivity. Amsterdam: John Benjamins Publishing Company. Oatley, K., and J.M. Jenkins. 1996. Understanding Emotions. Cambridge: Blackwell Publishers Ltd. Oatley, K., and P.N. Johnson-Laird. 1996. The Communicative Theory of Emotions: Empirical Tests, Mental Models, and Implications for Social Interaction. In Striving and Feeling, ed. L.L. Martin, and A. Tesser, 363–393. Mahwah, NJ: Erlbaum. Obrenovic, Z., N. Garay, J. Miguel Lopez, I. da Fajardo, and I. Cearreta. 2005. An Ontology for Description of Emotional Cues. In Proceedings of Affective Computing and Intelligent Interaction (ACII 2005), 505–512, Beijing, China. Oltramari, 2006. “LexiPass” Methodology: A Conceptual Path from Frames to Senses and Back. In Proceedings of LREC 2006. Ortony, A., G.L. Clone, and A. Collins. 1988. The Cognitive Structure of Emotions. New York: Cambridge University Press.
References
149
Ortony, A., and T.J. Turner. 1990. What’s Basic about Basic Emotions? Psychological Review 97: 315–331. Osgood, C.E. 1966. Dimensionality of the Semantic Space for Communication via Facial Expressions. Scandinavian Journal of Psychology 7 (1): 1–30. Panksepp, J. 1982. Toward a General Psychobiological Theory of Emotions. Behavioral and Brain sciences 5 (3): 407–422. Picard, R.W. 2007. Toward Machines with Emotional Intelligence. In The Science of Emotional Intelligence: Knowns and Unknowns, ed. G. Matthews, M. Zeidner, and R.D. Roberts. Oxford, UK: Oxford University Press. Picard, R.W. 1995/2000. Affective Computing. Cambridge. MA: The MIT Press. Picard, R.W. 2010. Emotion Research by the People, for the People. Emotion Review 2 (3): 250– 254. Plutchik, R. 1962. The Emotions: Fact, Theories and a New Model. New York: Random House. Plutchik, R. 1980. Emotions: A Psychoevolutionary Synthesis. New York: Harper & Row. Plutchik, R. 1991. The Emotions. New York: University Press of America. Plutchik, R. 1994. The Psychology and Biology of Emotion. New York, NY: Harper-Collins. Polanyi, L., and A. Zaenen. 2004. Contextual Valence Shifters. In Computing Attitude and Affect in Text: Theory and Applications, ed. J.G. Shanahan, Y. Qu, and J. Wiebe, 1–10. Poria, S., A. Gelbukh, E. Cambria, A. Hussain, and G. Huang. 2014. EmoSenticSpace: A Novel Framework for Affective Common-sense Reasoning. Knowledge-based Systems 69: 108–123. Poria, S., A. Gelbukh, A. Hussain, N. Howard, D. Das, and S. Bandyopadhyay. 2013. Enhanced SenticNet with Affective Labels for Concept-based Opinion Mining. IEEE Intelligent Systems 28 (2): 31–38. Posner, J., J. Russell, and B. Peterson. 2005. The Circumplex Model of Affect: An Integrative Approach to Affective: Neuroscience, Cognitive Development, and Psychopathology. Development and Psychopathology 17: 715–734. Potts, C. 2007. Expressive Dimension. Theoretical Linguistics 33 (2): 165–198. Power, M.J. 1999. Sadness and Its Disorders. In Handbook of Cognition and Emotion, ed. T. Dalgleish, and M.J. Power, 497–519. New York: Wiley. Pustejovsky, J. 1991. The Syntax of Event Structure. Cognition 41: 47–81. Pustejovsky, J. 1995. The Generative Lexicon. Cambridge: MIT Press. Quan, C., and F. Ren. 2009. Construction of a Blog Emotion Corpus for Chinese Expression Analysis. In Proceedings of EMNLP. Richards, J., J. Platt, and H. Weber. 1985. Longman Dictionary of Applied Linguistics. Harlow: Longman. Robins, R.H. 1964. General Linguistics: An Introductory Survey. London: Longman. Russell, J.A. 1980. A Circumplex Model of Affect. Journal of Personality and Social Psychology 39: 1161–1178. Sabini, J., and M. Silver. 2005. Ekman’s Basic Emotions: Why Not Love and Jealousy? Cognition and Emotion 19: 693–712. Sajnani H., S. Javanmardi, D. McDonald, and C. Lopes. 2011. Multi-Label Classification of Short Text: A Study on Wikipedia Barnstars. In Proceedings of the AAAI-11 Workshop on Analyzing Microtext. San Francisco, CA. August 8. Saurí, R., J. Littman, R. Knippen, R. Gaizauskas, A. Setzer, and J. Pustejovsky. 2004. TimeML Annotation Guidelines. http://www.timeml.org. Scheff, T. 2015. Toward Defining Basic Emotions. Qualitative Inquiry 21: 111–121. Scott, J.P. 1980. The Function of Emotions in Behavioral Systems: A Systems Theory Analysis. In Emotion: Theory, Research, and Experience, vol. 1, ed. R. Plutchik, and H. Kellerman. New York: Academic Press. Shweder, R.A. 1991. Thinking through Cultures: Expeditions in Cultural Psychology. Cambridge, M.A: Harvard University Press. Solomon, R.C. 1976. The Passions: The Myth and Nature of Human Emotion. Garden City, NY: Doubleday.
150
References
Solomon, R.C. 2003. What is an Emotion? Classic and Contemporary Readings. New York: Oxford University Press. Spinoza, B. 1675. Ethics. In The Collected Works of Spinoza, ed. E. Curley. Princeton, N.J.: Princeton University Press. 1985. Sroufe, L.A. 1979. Socioemotional Development. In Handbook of Infant Development, ed. J.D. Osofsky. New York: Wiley. Stevenson, R., J. Mikels, and T. James. 2007. Characterization of the Affective Norms for English Words by Discrete Emotional Categories. Behavior Research Methods 39 (4): 1020–1024. Strapparava, C., and R. Mihalcea. 2008. Learning to Identify Emotions in Text. In Proceedings of the ACM Conference on Applied Computing ACM-SAC 2008: 1556–1560. Subasic, P., and A. Huettner. 2001. Affect Analysis of Text Using Fussy Semantic Typing. IEEE Transactions on Fuzzy Systems 9: 483–496. Talmy, L. 2000. Toward a Cognitive Semantics. Vol. 1 and 2. Cambridge: MIT Press. Tang, L., S. Rajan and V.K. Narayanan. 2009. Large Scale Multi-label Classification via Metalabeler. In Proceedings of the 18th International Conference on World Wide Web. Tokuhisa, M., J. Murakami, and S. Ikehara. 2007. Construction of Text-dialog Corpus with Emotion Tags Focusing on Facial Expression in Comics. Journal of Natural Language Processing 14 (3): 193–218. Tokuhisa, R., K. Inui, and Y. Matsumoto. 2008. Emotion Classification Using Massive Examples Extracted from the Web. In Proceedings of COLING. Trevarthen, C. 1984. Emotions in Infancy: Regulators of Contact and Relationships with Persons. In Approaches to Emotion, ed. K.R. Scherer, and P. Ekman. Hillsdale, N. J.: Lawrence Erlbaum. Trohidis, K., G. Tsoumakas, G. Kalliris, and I. Vlahavas. 2008. Multilabel Classification of Music into Emotions. In Proceedings of 9th International Conference on Music Information Retrieval (ISMIR 2008), 325–330, Philadelphia, PA, USA. Tsunoda, T. 1999. Transitivity. In Concise Encyclopedia of Grammatical Categories, ed. K. Brown, and J. Miller, 383–391. Amsterdam: Elsevier. Turner, J.H. 1996. The Evolution of Emotions in Humans: A Darwinian-Durkheimian Analysis. Journal for the Theory of Social Behaviour 26: 1–34. Turner, J.H. 2000. On the Origins of Human Emotions: A Sociological Inquiry into the Evolution of Human Affect. California: Stanford University Press. Turner, J.H. 2007. Human Emotions: A Sociological Theory. New York: Routledge. Van Valin, R. 1990. Semantic Parameters of Split Intransitivity. Language 66: 221–260. Wiebe, J., T. Wilson, and C. Cardie. 2005. Annotating Expressions of Opinions and Emotions in Language. Language Resources and Evaluation 39 (2–3): 165–210. Wierzbicka, A. 1972. Semantic Primitives. Frankfurt: Athenäum. Wierzbicka, A. 1992. Defining Emotion Concepts. Cognitive Science 16: 539–581. Wierzbicka, A. 1996. Semantics: Primes and Universals. Oxford: Oxford University Press. Wierzbicka, A. 1998. Anchoring Linguistic Typology in Universal Concepts. Linguistic Typology 2 (2): 141–194. Wierzbicka, A. 1999. Emotions Across Languages and Cultures: Diversity and Universals. Cambridge: Cambridge University Press. Wikipedia (2010). https://en.wikipedia.org/wiki/Dream_of_the_Red_Chamber#cite_note-1. Xu, L.H., H.F. Lin, and J. Zhao. 2008. Construction and Analysis of Emotional Corpus [情感语料 库的构建和分析]. Journal of Chinese Information Proceeding 22 (01): 116–122. Xu, X.Y., and J.H. Tao. 2003. The Study of Affective Categorization in Chinese [汉语情感系统 中情感划分的研究]. In Proceedings of the 1st Chinese Conference on Affective Computing and Intelligent Interaction. Beijing, China, December 8–9. Yang, J., D.B. Bracewell, F. Ren, and S. Kuroiwa. 2008. The Creation of a Chinese Emotion Ontology Based on HowNet. Engineering Letters 16 (1): 166–171. Ye, Z. 2001. An Inquiry into ‘Sadness’ in Chinese. In Emotions in Crosslinguistic Perspective, ed. J. Harkins, and A. Wierzbicka, 359–404. Berlin: Mouton de Gruyter.
References
151
Ye, Z. 2006. Why Are There Two ‘Joy-like’ ‘Basic’ Emotions in Chinese? Semantic Theory and Empirical Findings. In Love, Hatred and Other Passions: Questions and Themes on Emotions in Chinese Civilisation, ed. P. Santangelo, and D. Guida, 59–80. E.J. Brill: Leiden. Yip, P., and D. Rimmington. 2004. Chinese: A Comprehensive Grammar. London: Routledge. Zhao, C.L. 2007. Some Principles on the Co-occurrence of Affective Adjectives and Nouns. Studies of the Chinese Language 2: 125–132. Zhao, C.L. 2009. Some Principles on the Co-occurrence of Affective Adjectives and Verbal-constructions [狀位情感形容詞與述位動詞結構同現的原則]. In The 5th International Conference on Contemporary Chinese Grammar. Hong Kong, November 28–29. Zhu, S., X. Ji, W. Xu, and Y. Gong. 2005. Multi-labelled Classification Using Maximum Entropy Method. In Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Salvador, Brazil.