基于局部全局上下文引导的方面级情感分析

Aspect Sentiment Analysis Based on Local-Global Context Guidance

情感分析是自然语言处理领域的重要方向之一,现有研究在探索方面词上下文对方面词情感极性的影响时仍存在句法信息捕捉困难、语义信息丢失、语义上下文缺失的问题。针对这些问题,提出一种新颖的结合局部全局上下文引导网络(LGCG)用于提升方面级情感分析的性能和表达能力。该方法首先通过构建文本的依赖句法分析树,为模型引入更多元化的信息特征;然后通过引入上下文聚焦机制将原始文本与依赖句法分析树的特征进行提炼,同时将提练后的局部特征向量与全局特征向量进行特征交互,有效保留方面词的上下文特征信息;最后使用特征聚合模块对局部全局特征进行聚合处理,提高了模型对方面词情感极性预测的准确度。在多个基准数据集上的实验结果表明,该模型相比于基线模型在准确率上分别提高了1.67%、1.67%、0.7%、0.16%,在F1值上分别提高了2.55%、2.03%、1.57%、2.08%。

Sentiment analysis is one of the important directions in the field of natural language processing.Existing researches on the influ-ence of context of exploration still has insufficient challenges such as difficulty in syntax information capture,loss of semantic information,and lack of semantic context.Aiming at these problems,propose a novel combination of local global context guidance network to improve the performance and expression ability in the aspect-based sentiment analysis.Specifically,in this method,a dependency syntax parsing tree is constructed firstly to introduce more diversified information features for the model;Then,by introducing the context focusing mechanism,the characteristics of the original text and dependency syntax parsing tree are refined.At the same time,the local feature vector of the refining is interacted with the global feature vector so as to retain the context feature information of the words effectively.Finally,the characteristic aggre-gation module is aggregated to the local global characteristics,which improves the accuracy of the model in emotional polarity prediction.The experimental results on the four benchmark datasets show that compared with the baseline models,the accuracy of the proposed model increas-es by 1.67%,1.67%,0.7%and 0.16%respectively,and the F1 value increases by 2.55%,2.03%,1.57%and 2.08%respectively.