Improving multimodal fusion with Main Modal Transformer for emotion recognition in conversation

作者：

Highlights：

• Modal with different representational abilities should be learned differently.

• Modal with stronger representation ability after feature extraction as the main modal.

• Preserve the integrity of the main modal features, enhancing the weak modal feature.

• Design an emotional cue extractor to enrich the conversation information.

摘要

•Modal with different representational abilities should be learned differently.•Modal with stronger representation ability after feature extraction as the main modal.•Preserve the integrity of the main modal features, enhancing the weak modal feature.•Design an emotional cue extractor to enrich the conversation information.

论文关键词：Emotion recognition in conversation,Main modal,Transformer,Multihead attention,Emotional cues

论文评审过程：Received 16 July 2022, Revised 28 September 2022, Accepted 30 September 2022, Available online 8 October 2022, Version of Record 21 October 2022.

论文官网地址：https://doi.org/10.1016/j.knosys.2022.109978