关键词: 基于方面的情感分析, 不确定性感知, 非似然学习, 最小熵, 对比学习

Abstract: Existing models still face multiple challenges in the Aspect Sentiment Quad Prediction (ASQP) task. First, existing models have difficulty in dealing with implicit sentiment expressions (such as implicit aspects or opinions). Implicit sentiment expressions lack clear lexical clues, which makes it difficult for the model to accurately capture sentiment tendencies. Second, a quadruple prediction is considered accurate only if all predicted elements exactly match the correct elements, while the model can generate confusing synonyms or synonyms that make the quadruple prediction completely wrong. Existing models are committed to improving the probability of predicting correct words, but ignore the probability of suppressing confusing words. At the same time, the cross-entropy loss used by existing models makes the model overconfident about wrong predictions, lacks modeling of uncertainty, and is difficult to actively suppress high-risk errors. These problems together limit the performance of the model in the aspect-based sentiment analysis task. To address these problems, a supervised contrastive generative sentiment analysis method based on uncertainty-aware unlikelihood learning (SCUAUL) is proposed. Firstly, supervised contrastive learning is used to shorten the semantic space distance of similar samples (such as the same sentiment polarity) through contrastive loss, thereby enhancing the model's ability to distinguish key features of input data (such as sentiment polarity, implicit aspects, etc.). Secondly, Monte Carlo Dropout (MC Dropout) is used to capture the inherent uncertainty of the model and discover confusing words, and the generation probability of easily confused words is dynamically suppressed through marginalized non-likelihood learning, the generation probability of correct words is maintained, and the minimum entropy constraint is combined to balance the generation diversity and accuracy. The average results of five experiments on the Rest15 and Rest16 datasets show that compared with baseline models such as AugABSA and PARAPHRASE, the precision of SCUAUL is improved by 0.4~3.98 and 0.38~3.83 percentage points, the recall is improved by 0.3~2.87 and 0.43~2.88 percentage points, and the F1 score is improved by 0.35~3.43 and 0.83~3.37 percentage points, respectively, which verifies the effectiveness of SCUAUL in aspect-based sentiment analysis tasks.

Key words: aspect-based sentiment analysis, uncertainty perception, non-likelihood learning, minimum entropy, contrastive learning