A research team from Soochow University has developed a novel artificial intelligence (AI) method to improve emotion cause extraction in conversations, enabling machines to better understand the nuanced triggers behind human emotions. Published in Frontiers of Computer Science, this breakthrough addresses key challenges in identifying fine-grained emotional causes within complex dialogues, offering potential applications in mental health support, customer service chatbots, and human-computer interaction systems.
Emotion cause extraction in conversations is critical for developing empathetic AI but remains challenging due to the need to analyze contextual relationships and eliminate irrelevant information. Current methods often overlook semantic coherence between speakers or fail to align predictions across related tasks, leading to inaccuracies. The new approach, which integrates coreference resolution and advanced task alignment techniques, significantly improves detection accuracy while requiring fewer data samples compared to existing models.
By incorporating coreference resolution—identifying when words refer to the same entity—into the AI`s attention mechanism, the model better captures semantic connections between utterances. Additionally, the team introduced a non-linear position embedding method at both the utterance and token levels, helping the AI reduce interference from redundant information. A novel framework ensures consistency between coarse-grained (sentence-level) and fine-grained (span-level) emotion cause predictions, minimizing contradictory results through bidirectional optimization.
Experiments on benchmark datasets showed the model outperforms state-of-the-art methods, achieving higher accuracy with lower false positive and negative rates.
As AI becomes increasingly embedded in daily life, improving its ability to interpret human emotions accurately is essential. This research not only advances conversational AI but also provides a framework for understanding complex human interactions in areas like psychotherapy and social media analysis.
DOI:10.1007/s11704-025-40931-2