ABSTRACT

Aspect-based sentiment analysis (ABSA) is a key problem in text analysis. However, previous work ignores the fact that the joint effects of local and global features affect the classification accuracy. Therefore, an ABSA model based on local position-part of speech (POS) awareness and global dense connection (LPP-GDC) is proposed to fully grasp the information from both local and global features concurrently. First, the BERT pre-trained model is used to obtain the word vectors. Second, position-POS awareness mechanism is designed to focus on words of important POS in the local context. Then, the multi-head self-attention mechanism and a densely connected graph convolutional network are constructed to capture the global information. Finally, the results are obtained by the dynamic feature fusion method. Experiments on three public datasets show that LPP-GDC obtains state-of-the-art performance.

摘要

基于方面的情感分析（ABSA）是文本分析中的一个关键问题。然而，先前的工作忽略了局部和全局特征的联合效应影响分类精度的事实。因此，提出了一种基于局部位置词性（POS）感知和全局密集连接（LPP-GDC）的ABSA模型，以同时充分掌握局部和全局特征的信息。首先，使用BERT预训练模型获得词向量。其次，位置 POS 感知机制旨在关注本地上下文中重要 POS 的单词。然后，构建多头自注意力机制和密集连接的图卷积网络来捕获全局信息。最后通过动态特征融合方法得到结果。