An attention-based ensemble model for emotion recognition in conversation

Emotion Recognition in Conversation (ERC) has become a major area of interest in the field of Natural Language Processing (NLP). The aim of this task is to achieve accurate classification of the emotions conveyed by speakers during conversation. For this purpose, it is essential to capture not only the literal meaning of words but also the relationships and interactions between different modalities. Most state-of-the-art approaches primarily focus on textual data to analyze speaker dependencies and contextual information within conversation.

This limitation restricts their ability to fully exploit information from multiple modalities, such as text and audio. Moreover, current ERC models struggle with the class imbalance issue, resulting in an inadequate performance for recognizing minority class emotions. To address these challenges, we propose an ensemble of attention-based cross-modal framework for ERC.

This framework addresses the challenges of insufficient utilization of multi-modal information and class imbalance in emotion classification. Experiments are conducted using the benchmark MELD dataset to evaluate the performance of the proposed model. Our model achieves a weighted average accuracy (WAA) of 64.33% and a weighted F1 (WF1) of 62.09%, for ERC outperforming state-of-the-art approaches.