An Emotion Cause Detection Method Based on XLNet and Contrastive Learning
Hai Feng Zhang, Cheng Zeng, Peng He · Proceedings/Proceedings of the ... International Conference on Software Engineering and Knowledge Engineering · 2022
Emotion cause detection is a new direction in the field of emotion research and is a fine-grained analysis of emotion. However, research on emotion cause detection is still challenging due to the extreme complexity of human emotions, the difficulty of tracing emotions back to their origins, and the fact that emotion cause detection corpus annotation requires a lot of manual involvement. An emotion cause detection method incorporating contrastive learning is proposed to address this problem, which combines the autoregressive language model XLNet and a contrastive learning approach to introduce a difficult sample generation strategy and a word repetition strategy in the positive/negative example comparison pattern in the training data,and design a loss function that incorporates the classification task and the comparison learning task. Experiments on the Weibo sports game commentary dataset show better performance in terms of accuracy, macro-average F1 values, and a 2.73 percentage point improvement in accuracy compared to the baseline XLNet, demonstrating the effectiveness of the proposed method.