EmoSense at SemEval-2019 Task 3: Bidirectional LSTM Network for Contextual Emotion Detection in Textual Conversations

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 210–214 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 210 EmoSense at SemEval-2019 Task 3: Bidirectional LSTM Network for Contextual Emotion Detection in Textual Conversations Sergey Smetanin National Research University Higher School of Economics Moscow, Russia sismetanin@gmail.com Abstract In this paper, we describe a deep-learning sys- tem for emotion detection in textual conversa- tions that participated in SemEval-2019 Task 3 “EmoContext”. We designed a specific archi- tecture of bidirectional LSTM which allows not only to learn semantic and sentiment fea- ture representation, but also to capture user- specific conversation features. To fine-tune word embeddings using distant supervision we additionally collected a significant amount of emotional texts. The system achieved 72.59% micro-average F1 score for emotion classes on the test dataset, thereby significantly outper- forming the officially-released baseline. Word embeddings and the source code were released for the research community. 1 Introduction Emotion detection has emerged as a challeng- ing research problem that can make some valu- able contribution not only in basic spheres like medicine, sociology and phycology but also in more innovative areas such as human-computer interaction. Nowadays, people increasingly com- municate using text messages with dialogue sys- tems, for which it is crucial to provide emotion- ally aware responses to users. The SemEval-2019 Task 3 “EmoContext” is focused on the contex- tual emotion detection in textual conversation. In EmoContext, given a textual user utterance along with 2 turns of context in a conversation, we must classify whether the emotion of the next user ut- terance is “happy”, “sad”, “angry” or “others” (4- point scale). For a detailed description see (Chat- terjee et al., 2019). In this paper, we present bidirectional LSTM for contextual emotion detection in textual con- versations that participated in SemEval-2019 Task 3 “EmoContext”. The proposed architecture aims to capture not only semantic and sentiment fea- ture representation from the conversation turns, but also to capture user-specific conversation fea- tures. We avoided using traditional NLP features like sentiment lexicons and hand-crafted linguistic features by substituting them with word embed- dings which were calculated automatically from the text corpora. Based on this paper, we make the following contributions1 freely available for the research community: • The source code of the deep-learning system for emotion detection. • Word embeddings fine-tuned for emotional detection in short texts. The rest of the article is organized as follows. Section 2 gives a brief overview of the related work. In section 3 we describe the proposed ar- chitecture of LSTM used in our system. Section 4 is focused on the texts pre-processing and train- ing process. Section 5 lays emphasis on the differ- ent system architectures and approaches we have tried. In conclusion, the performance of our sys- tem and further ways of research are presented. 2 Related Work In recent years deep learning techniques have cap- tured the attention of researchers due to their abil- ity to significantly outperform traditional methods in sentiment analysis task (Tang et …

Content truncated to 500 words