TokyoTech_NLP at SemEval-2019 Task 3: Emotion-related Symbols in Emotion Detection

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 350–354 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 350 TokyoTech NLP at SemEval-2019 Task 3: Emotion-related Symbols in Emotion Detection Zhishen Yang Tokyo Institute of Technology zhishen.yang @nlp.c.titech.ac.jp Sam Vijlbrief Delft University of Technology s.vijlbrief @student.tudelft.nl Naoaki Okazaki Tokyo Institute of Technology okazaki @c.titech.ac.jp Abstract This paper presents our contextual emotion de- tection system in approaching the SemEval- 2019 shared task 3: EmoContext: Contextual Emotion Detection in Text. This system co- operates with an emotion detection neural net- work method (Poria et al., 2017), emoji2vec (Eisner et al., 2016) embedding, word2vec embedding (Mikolov et al., 2013), and our proposed emoticon and emoji preprocessing method. The experimental results demon- strate the usefulness of our emoticon and emoji prepossessing method, and representations of emoticons and emoji contribute model’s emo- tion detection. 1 Introduction Social media and online text applications have been gaining popularity in recent years. Users post videos, pictures, and text to share their daily life as well as to communicate with others. This vast amount of multimodal data greatly facilitates vari- ous user analysis tasks such as sentiment analysis. Emotion detection as part of sentiment analysis can be conducted with user’s multimodal data such as facial expression and voice data in addition to text data. Therefore, emotion detection becomes a challenging problem when only textual data is available for extracting the contextual and senti- ment features (Chatterjee et al., 2019). For exam- ple, without proper visual and voice data, ”Why did you not call me last night” may be classified as sad or angry without an appropriate understanding of context. The SemEval-2019 shared task 3: EmoContext: Contextual Emotion Detection in Text is the task to detect an emotion of a three-turn conversation (Chatterjee et al., 2019). We can consider this task as contextual sentiment analysis, as it requires de- tecting the emotion of the third-turn conversation by comprehensively understanding the contextual Emotion class Distribution others 50% happy 14% sad 18% angry 18% Table 1: Emotion class distribution in the training data set. relationship and sentiment features from both lan- guage and emotion-related symbols. The task organizers provided a training set that consisted of 30,160 three-turn conversations. Meanwhile, as Table 1 shows, the emotion class distribution in the training data set was unbal- anced: 50% of samples were from “others” class. The task organizers also mentioned about the real- life distribution of emotion class: 88% of data is classified as “others” (Chatterjee et al., 2019). We also observed that over 28% of conversations in the training data set contained emoticons or emoji; users use them along with text to express emotion in a conversation. These two observations chal- lenge us to have a method that can learn emotion and contextual features from unbalanced training data and the emotion-related symbols (emoticons and emoji). The rest of this paper is organized as follows: Section 2 introduces the related work to the task; Section 3 explains our method in …

Content truncated to 500 words