Exploiting Narrative Context and A Priori Knowledge of Categories in Textual Emotion Classification

Proceedings of the 28th International Conference on Computational Linguistics, pages 5535–5540 Barcelona, Spain (Online), December 8-13, 2020 5535 Exploiting Narrative Context and A Priori Knowledge of Categories in Textual Emotion Classification Hikari Tanabe Tetsuji Ogawa Tetsunori Kobayashi Yoshihiko Hayashi Faculty of Science and Engineering, Waseda University Waseda-machi 27, Shinjuku, Tokyo 1620042, Japan tanabe@pcl.cs.waseda.ac.jp ogawa.tetsuji@waseda.jp koba@waseda.jp yshk.hayashi@aoni.waseda.jp Abstract Recognition of the mental state of a human character in text is a major challenge in natural language processing. In this study, we investigate the efficacy of the narrative context in recog- nizing the emotional states of human characters in text and discuss an approach to make use of a priori knowledge regarding the employed emotion category system. Specifically, we experi- mentally show that the accuracy of emotion classification is substantially increased by encoding the preceding context of the target sentence using a BERT-based text encoder. We also com- pare ways to incorporate a priori knowledge of emotion categories by altering the loss function used in training, in which our proposal of multi-task learning that jointly learns to classify posi- tive/negative polarity of emotions is included. The experimental results suggest that, when using Plutchik’s Wheel of Emotions, it is better to jointly classify the basic emotion categories with positive/negative polarity rather than directly exploiting its characteristic structure in which eight basic categories are arranged in a wheel. 1 Introduction Understanding of narrative text requires the ability to read between the lines and determine changes in the characters’ emotional states concerning the ongoing situation described in the text. Given this motivation, Rashkin et al. (2018) proposed a framework for annotating fully-specified chains of mental states with respect to characters’ motivations and emotional reactions. They publicized the resulting annotated corpus, called the Story Commonsense dataset. (Rashkin et al., 2018) further demonstrated that the mental state of a character in a narrative can be more accurately classified by using the preceding context of a target sentence. Although the results indicate the efficacy of the narrative context, the method might have certain limitations. The approach proposed by (Rashkin et al., 2018) focuses on the character-specific context that selectively concatenates the sub-set of sentences in which the target character appears. Although this account might be effective in excluding apparently irrelevant sentences, it may fail to extract implicit yet useful information from other parts of the context. Thus, we adopt a simple yet presumably effective method that encodes the entire preceding context along with the target sentence using BERT (Devlin et al., 2019). The task of emotion classification is generally formulated as a multi-label classification task (Zhou et al., 2016; He and Xia, 2018; Yu et al., 2018; Alswaidan and Menai, 2020) due to the innate com- plexity of human emotions: a sentence can express multiple emotions that an annotator may be unable to disassociate. Among the issues linked to multi-label classification, a crucial one is the handling of mutual dependencies among emotion categories. He and Xia (2018) proposed incorporating a term into the loss function so that it …

Content truncated to 500 words