基于语法和语义分割的跨领域方面级情感分类

Separated syntax and semantics modeling for cross-domain aspect-level sentiment classification

神经网络在方面级情感分类任务上已经取得了良好的性能.然而,由于复杂且耗时的数据标注流程,方面级情感分类在很多领域上是低资源甚至是零资源的,这限制了该任务在实际场景中的应用.为了解决这个挑战性的问题,本文关注跨领域的方面级情感分类,并提出一种基于语法和语义分割的跨领域方面情感分类方法.具体而言,针对不同领域用词差异造成的领域漂移和注意力泛化问题,本文首次提出利用单纯的语法信息来获取可在领域之间迁移的语法注意力,并引入与目标领域相近的文档情感分类任务来增强神经网络模型对目标领域的情感识别能力,最终从语法和语义两个层面分别提升模型的注意力机制和文本上下文表示.实验在6个跨领域方面级情感分类任务上进行,结果表明,与其他9种基线方法相比,本文的方法在6个任务上都取得了最先进的性能,在平均准确率和平均macro-F1两个指标上比之前最好的模型DIFD分别提升7.14%和7.6%.此外,即使以大规模预训练模型BERT,BERT-ADA,RoBERTa等作为骨干网络,本文的方法仍能实现3.5%以上的平均准确率提升和平均macro-F1提升.

Recently,neural network-based methods have achieved significant success in aspect-level sentiment classification(ALSC).However,the ALSC task is low-resource or even zero-resource in many domains because of complex and time-consuming data annotations.Data scarcity limits the applications of the ALSC to most domains in real-world scenarios.We thus focus on the cross-domain ALSC task.In previous attention-based methods,the distribution difference of words among domains leads to domain shift and poor generalization of the attention mechanism.To address these issues,we propose a separate syntax and semantics modeling(S3M)method to improve the attention of the neural model from the perspective of syntax and context representations from the perspective of semantics.Specifically,we find that the syntax pattern can be transferred across domains and thus propose syntax-based attention to capture the opinion expression of the given aspect.Besides,we introduce document-level sentiment classification data similar to target domain to help the model capture the sentiment semantics of target domain.We conduct experiments across six cross-domain ALSC tasks.Experimental results show that our S3M method achieves state-of-the-art performance above the benchmark.This method outperforms the previous best method DIFD by 7.14%in average accuracy and 7.6%in average macro-F1.When using BERT,BERT-ADA,or RoBERTa as a backbone,S3M still achieves over 3.5%improvement in average accuracy and in average macro-F1.