Abstract: In view of the fact that the current use of general sentiment dictionaries to enhance the sentiment tendency in the text cannot accurately identify the sentiment information between sentiment words and aspects in a specific domain, an aspect-based sentiment classification method with domain knowledge triples and graph convolutional networks (DKT-GCN) is proposed. First, the domain knowledge triples are constructed with reference to the span-based anti-bias aspect representation learning framework, and syntactic dependency analysis is conducted on the text sequence to extract the key structural information in the text. Next, the adjacency matrix is reconstructed using the constructed domain knowledge triples, thereby achieving domain knowledge enhancement. Then, the bidirectional encoder representation from transformers(BERT) model is used to capture the deep semantic and contextual information in the text, and generate hidden contextual representations of word vector containing context information for each word. Finally, the node information is aggregated through the graph convolutional networks model to better capture the semantic dependencies in the text, and the key aspect information is captured by adding mask operation, enabling the model to identify sentiment tendency more accurately. The experimental results show that the accuracy of proposed model on the four public benchmark datasets Lap14, Rest14, Rest15 and Rest16 reaches 79.15%, 86.34%, 84.50% and 91.56%, respectively, and the MF1 values are 75.72%, 79.83%, 71.84% and 80.78%, respectively, which are significantly higher than most comparison models. In addition, experimental results on laptop and restaurant datasets in SemEval 2014, SemEval 2015 and SemEval 2016 show that the proposed method can effectively improve the accuracy of sentiment classification.

Key words: aspect-based sentiment classification, domain knowledge triples, BERT model, graph convolutional networks

摘要： 针对目前使用通用情感词典来加强文本中的情感倾向，无法准确识别特定领域中情感词与方面之间情感信息的问题，提出一种基于领域知识三元组和图卷积网络的方面级情感分类方法（DKT-GCN）。参考基于跨度的反偏见方面表示学习框架构建领域知识三元组，对文本序列进行句法依存分析，提取出文本中的关键结构信息，并利用所构建的领域知识三元组重构邻接矩阵，从而实现领域知识增强；利用基于Transformers的双向编码器表示（BERT）模型捕捉文本中深层次的语义和上下文信息，为每个词生成包含上下文信息的词向量隐藏上下文表示；通过图卷积网络模型聚合节点信息以更好地捕捉文本中的语义依赖关系，并通过加入掩码操作捕获方面关键信息，使模型能够更精准地识别情感倾向。实验结果表明，所提模型在四个公共基准数据集Lap14、Rest14、Rest15和Rest16上的准确率分别达到79.15%、86.34%、84.50%和91.56%，MF1值分别为75.72%、79.83%、71.84%和80.78%，显著高于多数对比模型。此外，在SemEval 2014、SemEval 2015和SemEval 2016中笔记本电脑和餐厅数据集上的实验结果表明，该方法可以有效提升情感分类的准确性。

关键词: 方面级情感分类, 领域知识三元组, BERT模型, 图卷积网络