Learning Robust Pair Confidence for Multimodal Emotion-Cause Pair Extraction

The paper introduces Robust Pair Confidence Learning (RPCL), a novel framework for enhancing pair confidence in multimodal emotion-cause pair extraction (MECPE). RPCL employs a confidence-difference margin constraint to improve the separation between gold pairs and hard negatives, resulting in a 2.58 to 2.83 percentage point increase in mean Pair F1 across the ECF, MECAD, and MEC4 datasets, while also enhancing mean Pair AUPRC. This approach highlights the importance of structured pair confidence in training models, offering practitioners a method to achieve more reliable extraction of emotion-cause pairs in multimodal contexts.