Semi-supervised emotion recognition in textual conversation via a context-augmented auxiliary training task

摘要

Recognizing emotions in textual conversations (ERC) is to identify the emotion of utterances by considering conversational context. Current supervise-based ERC methods require a large number of diverse conversations to train a model that leverages context effectively. However, the scarcity of annotated training data in most ERC corpora hinders their performance improvement. In this paper, we explore to use the easily accessible unlabeled data mixed with labeled data to help ERC models to improve performance. Considering collected unlabeled data may not share the same emotion class with the labeled set, we propose a semi-supervised ERC algorithm that leverages the unlabeled conversational data through a novel Context-augmented Auxiliary training Task (CAUXIT), which trains along with the original ERC task in a multitask fashion. Our idea is to utilize CAUXIT to learn a better utterance feature representation on top of existing ERC models. Especially, CAUXIT is designed to selectively mask the utterance based on a class-based sampling strategy and use the context, i.e., the rest utterances, to predict its emotion-related information rather than the lexical information of itself, which enhances the network's ability in making emotion inference through context and consequently improve the utterance feature representation. In addition to applying CAUXIT to unlabeled data, we also extend it to labeled data to further enrich the supervision signal. As shown in the experiments, applying CAUXIT on various ERC models achieves a significant improvement over the same network architectures trained on labeled data, which verifies our approach as an effective semi-supervised ERC framework.