Emotion Classification by Jointly Learning to Lexiconize and Classify

Proceedings of the 28th International Conference on Computational Linguistics, pages 3235–3245 Barcelona, Spain (Online), December 8-13, 2020 3235 Emotion Classification by Jointly Learning to Lexiconize and Classify Deyu Zhou ∗ South China University of Technology chawdoe@163.com Shuangzhi Wu Tencent / Beijing, China frostwu@tencent.com Qing Wang ∗ University of Illinois at Urbana Champaign qwang55@illinois.edu Jun Xie Tencent / Beijing, China stiffxie@tencent.com Zhaopeng Tu Tencent AI Lab tuzhaopeng@gmail.com Mu Li Tencent / Beijing, China limugx@qq.com Abstract Emotion lexicons have been shown effective for emotion classification (Baziotis et al., 2018). Previous studies handle emotion lexicon construction and emotion classification separately. In this paper, we propose an emotional network (EmNet) to jointly learn sentence emotions and con- struct emotion lexicons which are dynamically adapted to a given context. The dynamic emotion lexicons are useful for handling words with multiple emotions based on different context, which can effectively improve the classification accuracy. We validate the approach on two represen- tative architectures – LSTM and BERT, demonstrating its superiority on identifying emotions in English tweets. Our model outperforms several approaches proposed in previous studies and achieves new state-of-the-art on the benchmark Twitter dataset. 1 Introduction Tweet Emotion This is a joke really how long will he keep diving and ducking. disgust That’s the joke. I know it’s incense. joy Table 1: Example sentences and their emotions. The last several years have seen a land rush in research on identification of emotions in short text such as Twitter or product reviews due to its greatly commercial value. For example, the emotions (e.g., anger or joy) expressed in prod- uct reviews can be a major factor in deciding the marketing strategy for a company (Meisheri and Dey, 2018). The SOTA approaches to this task (Baziotis et al., 2018; Meisheri and Dey, 2018) generally employ pre-defined emotion lexicons, which have two major limitations: 1. Most established emotion lexicons were created for a general domain, and suffer from limited cov- erage and inaccuracies when applied to the highly informal short text. 2. The pre-defined lexicons suffer from the ambiguity problem: the emotion of a word is highly in- fluenced by the context. Table 1 shows an example. The word “joke” carries different emotions according to different context. In this work, we tackle these challenges by jointly learning to construct the emotion lexicons and classify the emotions of short texts. Specifically, we propose a novel emotional network (EmNet), which consists of three main components: 1. Sentence encoder encodes the input sentence into semantic hidden states, which can be implemented as either LSTM or BERT. ∗ Work done while the author was an intern at Tencent. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons. org/licenses/by/4.0/. 3236 2. Emotion generator leverages both the semantic hidden states and word embeddings to construct word emotions, which dynamically adapt to the sentence context. 3. Emotion classifier classifies the sentence emotions based on both the encoded sentence representa- tions and generated word emotions. With the newly introduced emotion generator, our …

Content truncated to 500 words