摘要： 方面级情感分析是一种细粒度文本情感分析技术，可以判断文本目标方面的情感倾向，被广泛应用于商品评价、教育评价等领域，可以辅助用户更全面地了解实体属性并做出精准决策。但是现有方面级情感分析技术大多存在文本句法依存关系特征以及外部知识特征提取不充分的问题，为此，利用图卷积神经网络可以处理异构数据的特点，构建一种语义增强的方面级文本情感分析模型。将文本的词嵌入向量输入双向门控循环神经网络以提取文本和目标方面词的上下文语义信息，依据句法依存关系类型构建加权句法依存图，根据文本单词和外部知识库构建知识子图，使用图卷积神经网络处理加权句法依存图和知识子图，从而获取融合文本句法结构信息的文本特征和体现外部知识信息的目标方面特征，在此基础上，拼接两组特征向量完成情感极性分类。实验结果表明，在Laptop14、Restaurat14和Restaurat15数据集上，该模型的F1值分别达到77.34%、76.58%和68.57%，相比ATAE-LSTM、TD-LSTM、ASGCN等基线模型，其F1值分别平均提高7.28%、5.71%和6.28%，所提模型通过提取文本句法依存关系特征以及外部知识特征获得了更好的情感分析性能。

关键词: 方面级情感分析, 图卷积神经网络, 句法依存图, 知识图谱, 注意力机制

Abstract: Aspect-level sentiment analysis is a fine-grained text sentiment analysis technology that determines the sentiment tendency of text targets.It is widely used in fields such as product and education evaluation，and assists users in establishing a more comprehensive understanding of entity attributes and making accurate decisions.However，a key challenge is that the majority of existing aspect-level sentiment analysis techniques do not sufficiently extract text syntax dependency and external knowledge features.Therefore，a semantic enhanced aspect-level text sentiment analysis model is proposed herein using graph convolution neural networks to process heterogeneous data.A word embedding vector of the text is input into a bidirectional gated circular neural network to extract contextual semantic information of text and target aspect words，a weighted syntactic dependency graph is constructed based on the type of syntactic dependency，and a knowledge subgraph is constructed based on the text words and external knowledge base，and a graph convolutional neural network is used to process the weighted syntactic dependency graph and knowledge subgraph，to obtain text features that integrate text syntactic structure information and target aspect features that reflects external knowledge information.On this basis，two sets of feature vectors are spliced to complete sentimental polarity classification.Experimental results demonstrate that on the Laptop14，Restaurant14，and Restaurant15 datasets，the model F1 values reach 77.34%，76.58%，and 68.57%，respectively.Compared with baseline models such as ATAE-LSTM，TD-LSTM，and ASGCN，the model F1 values increase by an average of 7.28%，5.71%，and 6.28%，respectively.The proposed model achieves an improved sentimental analysis performance by extracting textual syntactic dependency and external knowledge features.

Key words: aspect-level sentiment analysis, graph convolution neural network, syntactic dependency graph, knowledge graph, attention mechanism

中图分类号: 

• TP391