Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (15): 169-176.DOI: 10.3778/j.issn.1002-8331.2204-0439

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Multilayer Gating and Relational Graph Attention Fusion Network for Aspect-Based Sentiment Analysis

LUO Rongrong, GONG Hongfang, XU Dan   

  1. School of Mathematics and Statistics, Changsha University of Science and Technology, Changsha 410114, China
  • Online:2023-08-01 Published:2023-08-01

融合多层门控与关系图注意力的方面情感分析

罗容容,龚红仿,徐丹   

  1. 长沙理工大学 数学与统计学院,长沙 410114

Abstract: As the mainstream techniques in current aspect-based sentiment analysis field, serialization neural network and graph neural networks focus on semantic and syntactic relation modeling, respectively. Aiming at the problems that serialization neural network cannot accurately capture the long-distance dependency of complex sentences, and graph neural network lacked local sequence semantics as well as fine-grained final sentiment representation, multilayer gating and relational graph attention fusion network is designed. The pre-trained model ERNIE 2.0 generated contextualized representations. Aspect-based gated recurrent unit captured sequence semantic information about aspect items. Relational graph attention network learned higher-order syntactic features in local sequence semantics. Finally, the feature refinement dual channels consisting of dual refinement gate network strengthen the interaction between specific aspect and context, filter irrelevant contextual information in semantics and syntax, and obtain aspect sentiment-enhanced representation with semantic and syntactic relations. Experiments on Twitter and SemEval2014 datasets show that the proposed fusion network has superior classification performance compared to eight advanced baseline methods.

Key words: aspect-based sentiment analysis, multilayer gating network, relational graph attention network, feature interactive refinement, local feature extraction

摘要: 作为目前方面情感分析领域的主流技术,序列化神经网络和图神经网络分别聚焦于语义和句法关系建模。针对序列化神经网络无法准确捕获复杂句的远距离依赖关系,图神经网络缺乏局部序列语义以及精细的最终情感表达等问题,设计了一种多层门控与关系图注意力混合网络。采用预训练模型ERNIE 2.0生成语境化表示,构建方面门控循环单元捕获有关方面的序列语义信息,使用关系图注意力网络学习局部序列语义中的高阶句法特征。最后由双重蒸馏门控网络构成的特征蒸馏双通道,强化特定方面与上下文之间的交互,过滤语义和句法中的冗余信息,获取兼具语义和句法关系的方面情感增强表示。在Twitter和SemEval2014数据集上进行的实验表明,相较于八种先进基线方法,所提出的混合网络具有更优的分类性能。

关键词: 方面情感分析, 多层门控网络, 关系图注意力网络, 特征交互蒸馏, 局部特征提取