一种基于多任务学习的方面级情感分析方法

AN ASPECT-LEVEL SENTIMENT ANALYSIS METHOD BASED ON MULTI-TASK LEARNING

针对方面类别或方面词项子任务的情感分析方法性能受单任务网络结构的限制,且相关任务的数据集不足。为此,提出一种基于多任务学习的方面级情感分析模型,共同学习两个子任务和方面词项重建任务。通过共享双向门控循环单元和Sentence-level的注意力对多个任务的句子向量编码,并利用Aspect-level的注意力针对不同方面捕获句子中的不同内容。使用双重注意力能更好地注意方面词和情感词,从而提高精确率。在SemEval三个数据集上的实验结果表明,该方法对方面级情感分析任务有明显的优化作用,与现有基线模型相比在精确率上有很好的提升。

The performance of those methods for either aspect category sentiment analysis(ACSA)or aspect term sentiment analysis(ATSA)is generally limited due to the single task network structure,and the datasets for each task are insufficient.To this end,an aspect-level sentiment analysis model based on multi-task learning is designed to jointly learn the two sub-tasks and aspect term reconstruction task.The multiple tasks sentence vectors were coded by sharing a bi-directional gated recurrent unit and a Sentence-level attention mechanism,and Aspect-level attention was used to capture different content in sentences for different aspects.The use of double attention can pay more attention to aspect words and sentiment words,thus improving the accuracy.The experimental results on the three SemEval datasets show that the method has a significant optimization effect on the aspect-level sentiment analysis task,and achieves better accuracy than the existing baseline model.