Multi-hypergraph Neural Networks for Emotion Recognition in Multi-party Conversations

Abstract

Emotion recognition in multi-party conversations (ERMC) is becoming increasingly popular as an emerging research topic in natural language processing. Although previous work exploited inter-dependency and self-dependency among participants, they paid more attention to the use of specific-speaker contexts. Specific-speaker context modeling can well consider the speaker’s self-dependency, but inter-dependency has not been fully utilized. In this paper, two hypergraphs are designed to model specific-speaker context and non-specific-speaker context respectively, so as to deal with self-dependency and inter-dependency among participants. To this end, we design a multi-hypergraph neural network for ERMC, namely ERMC-MHGNN. In particular, we combine average aggregation and attention aggregation to generate hyperedge features, which can make better use of utterance information. Extensive experiments are conducted on two ERC benchmarks with state-of-the-art models employed as baselines for comparison. The empirical results demonstrate the superiority of this new model and confirm that further exploiting inter-dependency is of great value for ERMC. In addition, we also achieved good results on the emotional shift issue.