Abstract
Emotion Recognition in Conversations (ERC) is the task of identifying the emotions of utterances from speakers in a conversation. The major challenge of ERC is how to aggregate useful contextual information from multiple utterances. Existing research in this area has mainly focused on (1) the context modeling of utterances and (2) the interaction between speakers in conversation. Another concern with ERC is that emphatically enhancing context modeling might over contextualize surrounding information and neglect the valuable influence of context-free semantic information from the utterance. In this work, we present the cross-level attention to preserve context-free semantics of utterances and prevent over-contextualizing. Additionally, multiparty attention masks are introduced to better model complex speaker interactions by separating the conversation into speakers and others. The proposed methods are integrated with the transformer-based architecture resulting in the Cross-Level Attention with Multi-Party mask model (CLAMP). The experimental results indicate that CLAMP empirically achieves competitive performance on the IEMOCAP dataset. The proposed method presents and improvement on context utilization when adding conversation utterances. Additionally the contribution of the proposed components is demonstrated in an ablation study.