Emotion Detection in Unfix-length-Context Conversation

TL;DR

This paper introduces a novel emotion detection method in conversations that dynamically models variable-length context using speaker-aware modules and a top-k normalization layer, improving accuracy over existing baselines.

Contribution

The paper proposes new modules for modeling inner- and inter-speaker dependencies and a top-k normalization layer to select optimal context windows for emotion prediction.

Findings

Outperforms strong baselines on three datasets

Effective modeling of speaker dependencies improves accuracy

Dynamic context window selection enhances emotion detection

Abstract

We leverage different context windows when predicting the emotion of different utterances. New modules are included to realize variable-length context: 1) two speaker-aware units, which explicitly model inner- and inter-speaker dependencies to form distilled conversational context, and 2) a top-k normalization layer, which determines the most proper context windows from the conversational context to predict emotion. Experiments and ablation studies show that our approach outperforms several strong baselines on three public datasets.