Multi-Label Emotion Recognition Model Integrating Multi-Perspective Common Sense Knowledge

Emotion recognition models aim at deciphering the nuanced emotions within sentences through the analysis of textual content. Studies have revealed that integrating commonsense knowledge into these models can markedly improve their accuracy. Commonsense knowledge represents the underlying context and cultural information exchanged in communication, spanning various aspects such as intentions, reactions, and impacts. Nonetheless, current emotion recognition models typically merge commonsense knowledge with text vectors in a simplistic manner, neglecting the multifaceted nature of this knowledge. This study introduces the MER-MPCK (Multi-label Emotion Recognition with Multiple Perspectives Commonsense Knowledge) model, which adeptly fuses extrinsic knowledge with textual data via multi-head attention mechanisms. The MER-MPCK model employs the COMET commonsense generation tool to distill nuanced commonsense insights from five distinct perspectives: intent of speaker, reaction of speaker, effect on speaker, effect on others and reaction of others. The model is structured into three components: the Text Semantic Module employs the RoBERTa architecture to capture semantic details from the text; the Commonsense Knowledge Extraction Module utilizes the COMET model to derive detailed commonsense knowledge based on the target text and reasoning relations; and the Commonsense Fusion Prediction Module employs multi-head attention, using commonsense knowledge as the Query and text semantics as the Key and Value, to generate sentence embeddings for emotion prediction. Experimental results on the GoEmotions dataset confirm the MER-MPCK model's superior performance in emotion recognition over existing baseline models.