MSPF: A multi-semantic prompting fusion framework for emotion-cause pair extraction in conversations

Emotion-cause pair extraction in conversations (ECPEC) has garnered increasing attention but struggles with modeling multi-turn utterance dependencies. While semantic prompting improves language understanding, its high computational cost hinders widespread ECPEC adoption. To overcome these constraints, we innovatively develop a multi-semantic prompting fusion (MSPF) framework by introducing the pair-oriented sampling strategy, focusing on candidate utterance pairs and transforming the ECPEC task into a pair verification issue. This shift enables us to incorporate three specialized semantic prompts, including tagging, synonym, and causal claim prompts, designed to enrich the semantics of emotion sentiment and emotion-cause relationships. We further present a knowledge attention module for the integration of tagging and synonym prompts, and a two-layer attention pooling module for merging tagging and dual causal claim prompts. Experimental results demonstrate that our proposed MSPF models outperforms the best existing baselines by 4.91 %, 4.08 %, and 2.86 % in F1 score on the ConvECPE, ECPE-D-DD, and ECPE-D-IE (for the domain adaptation experiment) datasets, respectively, with further ablation analysis confirming the effectiveness of our framework.