计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (9): 212-218.DOI: 10.3778/j.issn.1002-8331.2212-0375

• 模式识别与人工智能 • 上一篇    下一篇

基于位置增强词向量和GRU-CNN的方面级情感分析模型研究

陶林娟,华庚兴,李波   

  1. 华中师范大学 计算机学院,武汉 430079
  • 出版日期:2024-05-01 发布日期:2024-04-29

Aspect-Level Sentiment Analysis Based on Location-Enhanced Word Embeddings and GRU-CNN Model

TAO Linjuan, HUA Gengxing, LI Bo   

  1. School of Computer Science, Central China Normal University, Wuhan 430079, China
  • Online:2024-05-01 Published:2024-04-29

摘要: 方面级情感分析旨在判断一段文本中特定方面词的情感倾向,其核心问题是方面词的上下文如何准确表征。与现有研究主要关注注意力机制的改进不同,该文从词语表征和上下文编码模型两个方面进行改进。在词语表征方面,通过BERT模型和位置度量公式获得增强的词向量表示;在上下文编码模型方面,使用GRU-CNN网络提取文本语义特征。在SemEval2014 Task4数据集上的实验表明,提出的模型在Restaurant和Laptop领域中的准确率分别达到了85.54%和80.35%,证实了所提出模型的有效性。

关键词: 方面级情感分析, 卷积神经网络, 预训练词向量, 位置函数, 注意力机制

Abstract: Aspect-level sentiment analysis aims to judge the emotional attitude of a specific aspect-level word according to the given context. The core problem is how to accurately represent the context of the aspect word. Different from the existing research which mainly focuses on the improvement of attention mechanism, this paper focuses on two aspects: word representation and context encoding. In terms of word representation, the location-enhanced word representation is obtained through the BERT model and position measurement formula. In terms of context encoding, GRU-CNN network is used to extract semantic features of text. The experiments on the SemEval2014 Task4 dataset show that the accuracy of the proposed model on Restaurant and Laptop datasets respectively reaches to 85.54% and 80.35%, which proves the efficiency of the proposed model.

Key words: aspect-level sentiment analysis, convolutional neural network (CNN), pre-trained word vectors, position function, attention mechanism