Emotion Cause Detection with Linguistic Constructions

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 179–187, Beijing, August 2010 Emotion Cause Detection with Linguistic Constructions Ying Chen*†, Sophia Yat Mei Lee†, Shoushan Li†, Chu-Ren Huang† * Dep. of Computer Engineering China Agricultural University †Dep. of Chinese and Bilingual Studies The Hong Kong Polytechnic University {chenying3176,sophiaym,shoushan.li,churenhuang}@gmail.com Abstract This paper proposes a multi-label ap- proach to detect emotion causes. The multi-label model not only detects mul- ti-clause causes, but also captures the long-distance information to facilitate emotion cause detection. In addition, based on the linguistic analysis, we cre- ate two sets of linguistic patterns during feature extraction. Both manually gener- alized patterns and automatically gener- alized patterns are designed to extract general cause expressions or specific constructions for emotion causes. Ex- periments show that our system achieves a performance much higher than a baseline model. 1 Introduction Text-based emotion processing has been a cen- ter of attention in the NLP field in the past few years. Most previous researches have focused on detecting the surface information of emo- tions, especially emotion classes, e.g., “happi- ness” and “anger” (Mihalcea and Liu 2006, Strapparava and Mihalcea 2008, Abbasi et al, 2008, Tokuhisa et al. 2008). Although most emotion theories recognize the important role of causes in emotion analysis (Descartes, 1649; James, 1884; Plutchik 1980, Wierzbicka 1999), very few studies explore the interactions be- tween emotion and causes. Emotion-cause in- teraction is the eventive relation which poten- tially yields the most crucial information in terms of information extraction. For instance, knowing the existence of an emotion is often insufficient to predict future events or decide on the best reaction. However, if the emotion cause is known in addition to the type of emotion, prediction of future events or assessment of po- tential implications can be done more reliably. In other words, when emotion is treated as an event, causal relation is the pivotal relation to discover. In this paper, we explore one of the crucial deep level types of information of emo- tion, i.e. cause events. Our study focuses on explicit emotions in which emotions are often presented by emotion keywords such as “shocked” in “He was shocked after hearing the news”. Emotion caus- es are the explicitly expressed propositions that evoke the presence of the corresponding emo- tions. They can be expressed by verbs, nomi- nalizations, and nominals. Lee et al. (2010a) explore the causes of explicit emotions by con- structing a Chinese emotion cause corpus. Based on this corpus, we formalize the emotion cause detection problem through extensive data analysis. We find that ~14% emotion causes are complicated events containing multi-clauses, to which previous cause detection systems can hardly be applied directly. Most previous cause detection systems focus on the causal relation between a pair of small-size text units, such as clauses or phrases. They are thus not able to detect emotion causes that are multi-clauses. In this paper, we formalize emotion cause detec- tion as a multi-label classification task (i.e. each instance may contain more …

Content truncated to 500 words