Directed Acyclic Graphs With Prototypical Networks for Few-Shot Emotion Recognition in Conversation

Emotion recognition in conversation is the task of recognizing the emotion in each utterance of a conversation, and it is an active field of study with various applications. Many studies have been conducted in well-designed settings in which all of the emotion labels in the training set are available. However, a rich dataset with emotion labels for all utterances is rare. In this study, we address this problem by resorting to few-shot learning. Specifically, we propose Directed Acyclic Graph (DAG)-based prototypical networks, namely ProtoDAG, to better capture contextual information in conversations, which in turn leads to accurate emotion predictions. In our model, the DAG layers are tailored into prototypical networks, which is learned end-to-end. Our experiments with popular benchmark datasets demonstrate that our model achieves state-of-the-art results outperforming the existing few-shot learning model by a significant margin and is even competitive with fully supervised baseline models for emotion recognition in conversation.