基于局部特征聚焦的方面级情感分析

Aspect-based sentiment analysis based on local feature focusing

现有方面级情感分析模型忽略了各词间句法关系且未能针对性地提取语义信息。为此,提出一种可聚焦局部上下文特征的方面级情感分析模型,其核心思想在于构建局部上下文加权邻接图和动态赋权方法,通过图卷积神经网络生成聚焦于局部上下文信息的方面词特征。具体地,首先采用局部上下文动态赋权方式增加局部上下文的关注度;其次,在提取句法依存关系的基础上为上下文各节点赋权,构建针对局部上下文赋权的邻接图;最后,由图卷积神经网络提取聚焦于局部上下文信息的方面词特征。在公开数据集上的实验结果表明,与ASGCN相比,提出模型在restaurant和laptop数据集中的宏F1值分别提高了1.76%和1.12%,经过局部上下文加权,聚焦局部特征所得信息有助于提高分类效果。

The existing aspect-based sentiment analysis models ignore the syntactic relationship between words and fail to extract targeted semantic information. To alleviate the problem, this paper proposed an aspect-based sentiment analysis model to focus on local contextual features. The core idea was to construct local context weighted adjacency graph and dynamic weighting method, and generated aspect word features focusing on local context information through graph convolutional neural network. Specifically, the model adopted the local context dynamic weighting method to increase the attention to the local context during the feature extraction process. Secondly, it assigned weights to the context nodes based on the syntactic dependency relationship, and constructed an adjacency graph for the local context weighting. Finally, under the influence of the multi-layer graph convolutional neural network, it continuously extracted the aspect word features focusing on the local context information. The experimental results shows that, compared with ASGCN, the macro-F1value on the restaurant and laptop datasets increase by 1.76% and 1.12%, respectively. With the local context weighting, local feature focusing can help to improve the classification effect.