Multi-Task Learning and Adapted Knowledge Models for Emotion-Cause Extraction

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3975–3989 August 1–6, 2021. ©2021 Association for Computational Linguistics 3975 Multi-Task Learning and Adapted Knowledge Models for Emotion-Cause Extraction Elsbeth Turcan1∗ Kasturi Bhattacharjee3 Shuai Wang3 Yaser Al-Onaizan3 1Department of Computer Science, Columbia University 2Data Science Institute, Columbia University 3Amazon AI {eturcan, smara}@cs.columbia.edu {wshui, ranubhai, kastb, onaizan}@amazon.com Rishita Anubhai3 Smaranda Muresan2,3 Abstract Detecting what emotions are expressed in text is a well-studied problem in natural lan- guage processing. However, research on finer- grained emotion analysis such as what causes an emotion is still in its infancy. We present solutions that tackle both emotion recognition and emotion cause detection in a joint fash- ion. Considering that common-sense knowl- edge plays an important role in understanding implicitly expressed emotions and the reasons for those emotions, we propose novel meth- ods that combine common-sense knowledge via adapted knowledge models with multi-task learning to perform joint emotion classifica- tion and emotion cause tagging. We show per- formance improvement on both tasks when in- cluding common-sense reasoning and a multi- task framework. We provide a thorough analy- sis to gain insights into model performance. 1 Introduction Utterance and document level emotion recognition has received significant attention from the research community (Mohammad et al., 2018; Poria et al., 2020a). Given the utterance Sudan protests: Out- rage as troops open fire on protestors an emotion recognition system will be able to detect that anger is the main expressed emotion, signaled by the word "outrage". However, the semantic informa- tion associated with expressions of emotion, such as the cause (the thing that triggers the emotion) or the target (the thing toward which the emotion is directed), is important to provide a finer-grained understanding of the text that might be needed in real-world applications. In the above utterance, the cause of the anger emotion is the event “troops open fire on protestors”, while the target is the en- tity "troops" (see Figure 1) . ∗Work done during an internship with Amazon AI. Research on finer-grained emotion analysis such as detecting the cause for an emotion expressed in text is in its infancy. Most work on emotion-cause detection has utilized a Chinese dataset where the cause is always syntactically realized as a clause and thus was modeled as a classification task (Gui et al., 2016). However, recently Bostan et al. (2020) and Oberländer and Klinger (2020) argued that in English, an emotion cause can be expressed syntactically as a clause (as troops open fire on protestors), noun phrase (1,000 non-perishable food donations) or verb phrase (jumped into an ice-cold river), and thus we follow their approach of framing emotion cause detection as a sequence tagging task. We propose several ways in which to approach the tasks of emotion recognition and emotion cause tagging. First, these two tasks should not be in- dependent; because the cause is the trigger for the emotion, knowledge about what the cause is should narrow down what emotion may be ex- pressed, and vice versa. Therefore, we …

Content truncated to 500 words