计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (12): 122-131.DOI: 10.3778/j.issn.1002-8331.2203-0215

• 模式识别与人工智能 • 上一篇    下一篇

融合依存句法先验知识的方面级情感分析研究

方义秋,彭杨,葛君伟   

  1. 1.重庆邮电大学 计算机科学与技术学院,重庆 400065
    2.重庆邮电大学 软件工程学院,重庆 400065
  • 出版日期:2023-06-15 发布日期:2023-06-15

Aspect-Level Sentiment Analysis Research Integrating Dependent Syntactic Prior Knowledge

FANG Yiqiu, PENG Yang, GE Junwei   

  1. 1.School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
    2.School of Software Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Online:2023-06-15 Published:2023-06-15

摘要: 针对方面级情感分析中依存句法信息利用不足,以及上下文信息与依存句法信息学习分离的问题,提出融合先验知识与预训练模型的方面级情感分析研究方法。该方法将依存句法信息作为先验知识与句子一起输入到预训练模型,通过预训练模型在对句子进行上下文信息学习的同时,学习依存句法建模的方面词与情感词之间的远距离信息。进一步利用依存距离表示将预训练词向量进行动态权重匹配从而充分利用句法信息。通过自注意力机制学习词性信息以及依存关系标签信息,通过两种辅助信息增强句子表示。拼接三种信息作为情感分类器的输入。在四种基准数据集上与其他方面级情感分析方法相比,准确率和F1值都有所提高,验证了该方法提出的先验知识融合与动态权重匹配的有效性,在方面级情感分析任务上有较大的应用价值。

关键词: 依存句法, 先验知识, 预训练模型, 动态权重匹配, 自注意力机制

Abstract: In view of the insufficient utilization of dependent syntactic information in aspect-level sentiment analysis, as well as the separate learning of contextual information and dependent syntactic information, a research method of aspect-level sentiment analysis integrating prior knowledge and pre-training model is proposed. Firstly, this method inputs the dependent syntax information as prior knowledge and the sentence into the pre-training model. Through the pre-training model, the long-distance information between aspect words and emotional words of dependent syntactic modeling is learned while learning the context information of the sentence. Secondly, the dependency distance representation is further used to match the dynamic weight of the pre-training word vector, so as to make full use of the syntactic information. Thirdly, the part of speech information and the dependency label information are learned through self-attention mechanism, and using two auxiliary information to enhance sentence representation. Lastly, three kinds of information are spliced as the input of sentiment classifier. Compared with other aspect-level sentiment analysis methods, the accuracy and F1 value are increased on the four benchmark data sets, which verifies the effectiveness of prior knowledge fusion and dynamic weight matching proposed by this method, and has a great application value in aspect-level sentiment analysis tasks.

Key words: dependency syntax, prior knowledge, pre-training model, dynamic weight matching, self-attention mechanism