基于语法类型依赖的图注意力网络细粒度情感分析方法

Aspect-based sentiment analysis based on dependency syntax type graph attention network

现实生活中人们在网络上会产生大量的评价性文本,对这些文本进行细粒度情感分析,可以帮助决策者快速了解人们对某类产品或事件的看法和意见,提供决策支持。细粒度情感分析旨在利用神经网络来对输入句子进行编码,提取目标与其上下文之间的关系,从而进行情感倾向的判断。最近,人们把图神经网络应用于该任务,取得了很好的结果。然而这些方法没有利用依赖类型标签信息,也没有考虑到噪声的影响,为解决这2个问题,本文研究了一种新型图神经网络拓展来利用依赖类型信息,并应用本文提出的语法语义交互融合的方法来降低噪声影响,在4个公开数据集上取得了最优的效果。

In current life,a large number of evaluative texts could be generated on the Internet.Sentiment analysis of these texts can help decision makers quickly understand people′s views and opinions on a certain type of product or event,and provide decision support.Aspect-level sentiment analysis aims to use neural network to encode the input sentences,extract the relationship between the aspect and context,and judge the emotional tendency.Recently,graph neural network has been applied to this task and achieved fine results.However,these methods do not use the dependency type label information,and do not take into account the impact of noise.To solve these two problems,the paper studies a new graph neural network extension to utilize the dependency type information,and applies the syntax-semantic interaction proposed in this paper to reduce the influence of noise.The fusion method has achieved the best results on 4 public datasets.