加入梯度均衡机制的端到端方面级情感分析

Gradient harmonized for end-to-end aspect-based sentiment analysis

目前方面级情感分析任务存在类别标签不平衡的问题,会导致模型过多学习非方面词标签,影响模型的性能。基于BERT端到端方面级情感分类模型,针对其类别标签不平衡的问题进行深入研究,提出使用梯度均衡机制缓解类别标签不平衡的问题,采用指数滑动平均的方法缓解潜在离群点样本对实验结果造成的影响,提升模型的性能。在4个标准数据集中进行实验,实验结果表明,所提方法在实验结果中(F1值)相比许多强基线方法有较为明显提升。

Aspect-based sentiment analysis task has the problem of unbalanced label class,causing the model to learn too much non-aspect word labels,which affecting the performance of the model.Based on BERT end-to-end aspect-based sentiment analysis model,focusing on the problem of unbalanced label class,in-depth research was conducted,gradient harmonized mechanism was used to alleviate the problem of unbalanced labels.To improve the performance of the model,exponential moving average was used to alleviate the influence of potential outlier samples on the experimental results.Experiments were carried out on four benchmark datasets.Experimental results show that the proposed method is better than many strong baselines methods.