摘要：

借助句法依赖信息和外部知识的图神经网络近年来成为方面级情感分析领域的一个研究热点，但是现有研究存在语法信息提取不充分和利用不合理等问题，同时未考虑增强文本方面词与意见词等关键节点的背景知识。此外，基于注意力机制的方法没有建立方面词与上下文词的语法信息交互，导致方面词错误地关注到与其语法无关的上下文词信息。提出一种基于虚拟依存关系与双知识增强的多交互图卷积网络模型。对方面词内每个单词构建依赖树，依据虚拟依存关系进行加权构造虚拟依存图，依据外部情感知识构造情感依存图，使用双通道图卷积神经网络处理虚拟依存图与情感依存图并进行融合，对融合后特定方面的特征表示进行语义和语法双交互。利用概念知识增强特定方面后的特征表示与上下文表示并进行知识注意力交互，对多种不同的增强表示进行融合从而实现不同表示间的共享与互补。实验结果表明，相较于经典的图卷积网络模型ASGCN，该模型在Rest15和Rest16数据集上的F1值分别提升4.71和8.57个百分点，具有较好的情感分类性能。

关键词: 方面级情感分析, 虚拟依存关系, 知识增强, 图卷积网络, 情感知识, 概念知识

Abstract:

Graph neural networks incorporating syntactic dependency information and external knowledge have recently gained significant attention in Aspect-Based Sentiment Analysis(ABSA). However, existing research exhibits limitations, such as inadequate extraction of grammatical information and suboptimal utilization. Moreover, the background knowledge pertaining to key nodes, such as context words and opinion words, has not been completely considered. Additionally, the attention mechanism-based method fails to establish a grammar information interaction between aspect words and context words, making aspect words mistakenly focus on irrelevant contextual information. A multi-interaction Graph Convolutional Network(GCN) model is proposed based on virtual dependencies and double knowledge enhancement. The model constructs a dependency tree for each word within an aspect word, generates a virtual dependency graph based on virtual relations, constructs an emotional dependency graph using external affective knowledge, and employs a dual channel graph convolutional neural network to process and fuse the virtual and affective dependency graphs. This approach enables semantic and grammatical double interaction on the feature representation of the fused specific aspect. Simultaneously, the feature and context representations of specific aspects, enhanced by conceptual knowledge, are used for knowledge-attention interaction. Finally, multiple enhanced representations are fused to facilitate sharing and complementation among different representations. Experimental results demonstrate that the proposed model achieves superior emotional classification performance compared to ASGCN. Specifically, the F1 values of the proposed model on the Rest15 and Rest16 datasets increase by 4.71 and 8.57 percentage points, respectively.

Key words: Aspect-Based Sentiment Analysis(ABSA), virtual dependency, knowledge enhancement, Graph Convolutional Network(GCN), emotional knowledge, conceptual knowledge