Using Context to Help Predict Speaker's Emotions in Social Dialogue

Mei Si

Abstract: Emotion plays a vital role in social interaction. Often, the speaker’s attitude is as essential as, if not more important than, his/her words for communication purposes. In this paper, we present experiments for using conversational context to help text-based emotion detection. We used data from the Dialogue Emotion Recognition Challenge – EmotionX. BERT is used for encoding the input sentences. We explore four ways for encoding the input by varying whether to concatenate a dialogue history with the current sentence and whether to add the speaker’s name as part of the input. Our results indicate that adding context can improve the results of emotion detection when the emotion categories do not overlap with each other.