Dialogue relation extraction based on Multi-Perspective Contextual Interaction Attention Network

Dialogue relation extraction aims to identify semantic relations between entity pairs from dialogue text. Multi-party dialogues present unique challenges due to dynamic speaker interactions and cross-turn dependencies, which existing methods fail to adequately model. We propose the Multi-Perspective Contextual Interaction Attention Network (MPCIA-Net) to address these limitations through two key components: (1) The Dual-stream Context Disentanglement Module ( DCDM ), which decouples context representations from the base encoder through parallel feature streams and integrates fine-grained contextual features; (2) The Dual-Attention Fusion Module ( DAF ), which processes features from dual perspectives—speaker perception for pronoun resolution through dialogue participant identity modelling, and utterance perception for capturing long-range cross-turn dependencies. The representations from both channels are merged through an adaptive attention mechanism for final relation prediction. Experimental results demonstrate that MPCIA-Net achieves an F 1 score of 68.70% on the DialogRE dataset, surpassing current state-of-the-art methods. • A novel dialogue relation extraction model MPCIA-Net is proposed. • DCDM module resolves speaker-utterance entanglement and reference chains. • Asymmetric attention mechanism enhances long-distance dependency modeling.