MILAB at SemEval-2019 Task 3: Multi-View Turn-by-Turn Model for Context-Aware Sentiment Analysis

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 256–260 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 256 MILAB at SemEval-2019 Task 3: Multi-View Turn-by-Turn Model for Context-Aware Sentiment Analysis Yoonhyung Lee, Yanghoon Kim, and Kyomin Jung Seoul National University, Seoul, Korea {cpi1234, ad26kr, kjung}@snu.ac.kr Abstract This paper describes our system for SemEval- 2019 Task 3: EmoContext, which aims to pre- dict the emotion of the third utterance consid- ering two preceding utterances in a dialogue. To address this challenge of predicting the emotion considering its context, we propose a Multi-View Turn-by-Turn (MVTT) model. Firstly, MVTT model generates vectors from each utterance using two encoders: word-level Bi-GRU encoder (WLE) and character-level CNN encoder (CLE). Then, MVTT grasps contextual information by combining the vec- tors and predict the emotion with the contex- tual information. We conduct experiments on the effect of vector encoding and vector com- bination. Our final MVTT model achieved 0.7634 microaveraged F1 score. 1 Introduction Sentiment analysis is a task of identifying emo- tional information from text materials and has been studied by various research fields since it can be applied to many applications such as public sur- vey and market analysis. However, most studies in sentiment analysis have only focused on a single sentence (Socher et al., 2011) or a single document (Pang and Lee, 2005). It is still hard to predict the emotion of a sentence with extra contextual infor- mation because the emotion can be understood dif- ferently depending on its context. SemEval-2019 Task 3: EmoContext (Chatterjee et al., 2019) pro- vides a dataset of dialogues which consist of three utterances between two users (Figure 1). Partic- ipants are required to predict the emotion of the third utterance among ‘Happy’, ‘Sad’, ‘Angry’, and ‘Others’, considering its context of two pre- ceding utterances. In this paper, we propose a Multi-View Turn- by-Turn (MVTT) model which encodes each ut- terance separately and combines the encoded vec- Figure 1: An example of a dialogue between two users. tors to get the contextual information. MVTT model first generates vectors from each utterance with two encoders: word-level Bi-GRU encoder (WLE) and character-level CNN encoder (CLE). The two-encoder strategy makes MVTT model more robust to the noisy texts which have a lot of typos and abbreviations. Then, MVTT extracts contextual information by combining the vectors and makes a prediction. We compare the MVTT model with some variants focusing on utterance vector encoding and utterance vector combination methods with microaveraged F1 score (F1) which is the main evaluation metric. This paper is organized as follows: Section 2 describes MVTT model architecture in detail. Section 3 describes dataset and various methods that we use to reflect the dataset’s characteristics to our training. Section 4 compares our results of MVTT model and other variants, and Section 5 outlines our conclusions. 2 System Description This section describes our Multi-View Turn-by- Turn (MVTT) model which consists of two en- coders: word-level Bi-GRU encoder (WLE) and character-level CNN encoder (CLE), which …

Content truncated to 500 words