Affective Event Classification with Discourse-enhanced Self-training

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5608–5617, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5608 Affective Event Classification with Discourse-enhanced Self-training Yuan Zhuang, Tianyu Jiang, Ellen Riloff School of Computing University of Utah {yyzhuang, tianyu, riloff}@cs.utah.edu Abstract Prior research has recognized the need to asso- ciate affective polarities with events and has produced several techniques and lexical re- sources for identifying affective events. Our re- search introduces new classification models to assign affective polarity to event phrases. First, we present a BERT-based model for affective event classification and show that the classi- fier achieves substantially better performance than a large affective event knowledge base. Second, we present a discourse-enhanced self- training method that iteratively improves the classifier with unlabeled data. The key idea is to exploit event phrases that occur with a coref- erent sentiment expression. The discourse- enhanced self-training algorithm iteratively la- bels new event phrases based on both the classifier’s predictions and the polarities of the event’s coreferent sentiment expressions. Our results show that discourse-enhanced self- training further improves both recall and preci- sion for affective event classification. 1 Introduction In recent years, researchers have been tackling the problem of identifying affective events, which are events that have a positive or negative effect on peo- ple who experience the event. For example, events that are typically positive include being hired for a new job, breaking a sports record, or buying a home. Conversely, events that are typically neg- ative include being fired from a job, breaking an arm, or having your house burn down. People’s world knowledge about events and how they im- pact people is sufficient for humans to infer the affective state of someone who experiences such an event, even if that person does not explicitly express an emotion. Consequently, we will refer to these events as having positive or negative polarity with respect to an implicit affective state. Research has shown that recognizing affective events is im- portant for a variety of natural language processing tasks, including narrative text comprehension and summarization (Lehnert, 1981; Goyal et al., 2013), dialogue systems (Andr´e et al., 2004), response generation (Ritter et al., 2011), and sarcasm detec- tion (Riloff et al., 2013). Much of the prior work on recognizing affective events has focused on producing lexical resources of verbs or event phrases with corresponding af- fective polarity values (Goyal et al., 2010, 2013; Rashkin et al., 2016; Ding and Riloff, 2016, 2018). These resources reflect substantial progress toward recognizing affective events in text, but their cover- age is limited by their fixed content. We hypothe- sized that deep learning architectures that encode rich meaning representations could lead to a more effective approach for identifying affective events. Specifically, neural classification models have the capacity to generalize across lexically and syntacti- cally different phrases that are semantically similar, and similar events are usually associated with the same affective polarity. To explore this approach, we created a BERT-based model for affective event classification and …

Content truncated to 500 words