BAB-QA: A New Neural Model for Emotion Detection in Multi-party Dialogue.

In this paper, we propose a new neural model BAB-QA to address the task of emotion detection in multi-party dialogues, which aims to detect emotion for each utterance in a dialogue among four label candidates: joy, sadness, anger, and neutral. A variety of models have been proposed to solve this task, but few of them manage to capture contextual information in a dialogue properly. Therefore, we adopt a Bi-directional Long Short-Term Memory network (BiLSTM) and an attention network to obtain representations of sentences and then apply a contextualization network to refine the sentence representations for classification. More importantly, we propose and incorporate a new module called QA network in our model, which is inspired by natural language inference tasks. This QA network enables our model to acquire better sentence encodings by modeling adjacent sentences in a dialogue as question-answer pairs. We evaluate our model in the benchmark EmotionX datasets provided by SociaINLP2018 and our model achieves the state-of-the-art performance.