Implicit Emotion Analysis Based on Improved Supervised Contrastive Learning

Implicit sentiment analysis (ISA) can be distinguished from traditional text by the fact that it does not rely on emotional words as clues, and expression is usually vaguer. Identifying implicit emotions more difficult, requires a deeper understanding of context, even when are absent. Researchers have focused context feature modeling developing sophisticated extraction mechanisms instead starting perspective. Enhancing difference in features samples an intuitive method to address this challenge. We proposed supervised contrastive learning (SCL) during training enables model conduct based emotion labels while weak features. SCL strengthen average embedding distance between texts with different enhance discrimination. Moreover, research indicates contextual information improve classification ability. Therefore, we applied straightforward fusion (bi-affine) over complicated approach. To evaluate effectiveness our method, conducted experiments SMP2019-ECISA (Chinese analysis) dataset. The results show 2.13% enhancement F1 value compared BERT baseline, proving methods.