Interpretable Relevant Emotion Ranking with Event-Driven Attention

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 177–187, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 177 Interpretable Relevant Emotion Ranking with Event-Driven Attention Yang Yang† Deyu Zhou∗† Yulan He§ Meng Zhang† †School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China §Department of Computer Science, University of Warwick, UK {yyang, d.zhou}@seu.edu.cn, y.he@cantab, m.zhang@seu.edu.cn. Abstract Multiple emotions with different intensities are often evoked by events described in doc- uments. Oftentimes, such event information is hidden and needs to be discovered from texts. Unveiling the hidden event information can help to understand how the emotions are e- voked and provide explainable results. How- ever, existing studies often ignore the latent event information. In this paper, we proposed a novel interpretable relevant emotion rank- ing model with the event information incor- porated into a deep learning architecture using the event-driven attentions. Moreover, corpus- level event embeddings and document-level event distributions are introduced respectively to consider the global events in corpus and the document-specific events simultaneously. Ex- perimental results on three real-world corpora show that the proposed approach performs re- markably better than the state-of-the-art emo- tion detection approaches and multi-label ap- proaches. Moreover, interpretable results can be obtained to shed light on the events which trigger certain emotions. 1 Introduction The advent and prosperity of social media enable users to share their opinions, feelings and attitudes online. Apart from directly expressing their opin- ions on social media posts, users can also vote for their emotional states after reading an article on- line. An example of a news article crawled from Sina News Society Channel together with its as- sociated emotion votes received from readers is il- lustrated in Figure 1. Treating these emotion votes as labels for the news article, we can define the e- motion detection problem as an emotion ranking problem that ranks emotions based on their in- tensities. Moreover, some of the emotion labels ∗Corresponding author 女子虐打小孩被拍 一段近5分钟的女子虐打儿童的视频引发网友关注。一名儿童被女子用 小棍连续用力抽打。整个过程中，儿童哭泣声不断。打人家长是否违法 等在进一步调查中。 A women beating her child abusively were photographed A nearly five-minute video of women beating a child aroused the attention of the public. A child was beaten severely and continuously by a woman with a stick. The child kept crying throughout the whole process. Further investigations will be conducted to determine whether the parent beating the child is illegal or not……. Moved Shocked Funny Sad Strange Angry 2 30 1 86 1 159 Figure 1: An example of an online news article with the readers’ votes on various emotion categories. Words highlighted in red are event-indicative words. could be considered as irrelevant emotions. For example, the emotion categories ‘Moved’, ‘Fun- ny’ and ‘Strange’ in Figure 1 only received one or two votes. These emotion votes could be nois- es (e.g., readers accidentally clicked on a wrong emotion button) and hence can be considered as irrelevant emotions. We need to separate …

Content truncated to 500 words