THU_NGN at SemEval-2019 Task 3: Dialog Emotion Classification using Attentional LSTM-CNN

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 340–344 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 340 THU NGN at SemEval-2019 Task 3: Dialog Emotion Classification using Attentional LSTM-CNN Suyu Ge , Tao Qi , Chuhan Wu , Yongfeng Huang Department of Electronic Engineering, Tsinghua University Beijing 100084, China {gesy17,qit16,wuch15,yfhuang}@mails.tsinghua.edu.cn Abstract With the development of the Internet, dialog systems are widely used in online platforms to provide personalized services for their users. It is important to understand the emotions through conversations to improve the qual- ity of dialog systems. To facilitate the re- searches on dialog emotion recognition, the SemEval-2019 Task 3 named EmoContext is proposed. This task aims to classify the emo- tions of user utterance along with two short turns of dialogues into four categories. In this paper, we propose an attentional LSTM- CNN model to participate in this shared task. We use a combination of convolutional neural networks and long-short term neural networks to capture both local and long-distance con- textual information in conversations. In ad- dition, we apply attention mechanism to rec- ognize and attend to important words within conversations. Besides, we propose to use ensemble strategies by combing the variants of our model with different pre-trained word embeddings via weighted voting. Our model achieved 0.7542 micro-F1 score in the final test data, ranking 15th out of 165 teams. 1 Introduction The analysis of emotions in dialog systems where limited number of words appear with strong se- mantic relations between them deserves special attention in domain of natural language process- ing (NLP) due to both interesting language nov- elties and wide future prospects (Gupta et al., 2017). By analyzing the emotions through conver- sations, service providers can design better chat- ting strategies according to users’ emotion pat- terns, which can improve user experience. There- fore, SemEval-2019 task 3 (Chatterjee et al., 2019) aims to call for research in this field. Given a tex- tual dialogue, i.e., a user utterance along with two turns of context, systems need to classify the emo- tion of user utterance into four emotion classes: happy, sad, angry or others. The field of sentiment analysis has been ex- tensively studied. For example, SemEval-2018 Task 2 (Barbieri et al., 2018) once called for the study on relevance between tweet texts and emo- jis. However, understanding textual conversations is challenging in absence of voice modulations and facial expressions, which participants at this task are asked to deal with. Apart from diminishing the negative impact caused by class size imbalance, ambiguity, misspellings and slang, their systems should mainly focus on capturing the intricate in- terplay between two turns of conversations. Traditional sentiment analysis requires a lot of feature engineering, such as n-grams and features extracted from sentiment lexicons (Mohammad and Turney, 2013; Kiritchenko et al., 2014a), and then feed them into a classifier such as Support Vector Machines (SVM) (Bollen et al., 2011; Kir- itchenko et al., 2014b). However, manual feature engineering usually needs a large amount of …

Content truncated to 500 words