关键词: 方面情感三元组抽取, 细粒度词对标记, 结点自适应网络, 句法依赖树, 卷积注意力

Abstract: Sentiment analysis plays an important role in the field of natural language processing. As a sub-task of sentiment analysis, aspect sentiment triplet extraction can extract users’ opinions on products or services from reviews, thus providing key data support for merchants to make decisions, so it has great research value. However, there are some problems such as insufficient consideration of word pair boundary, redundant word pair labels and limited utilization of linguistic character. In order to solve these problems, an aspect sentiment triple extraction model based on fine-grained labeling and node adaptive network is proposed. Firstly, a fine-grained word pair labeling scheme is designed, where different labels are used at the head and tail boundaries of items and redundant inter-word labels in items are removed. Secondly, the node adaptive network is used to mine the syntactic dependency information and syntactic type information on the two channels respectively, and the convolutional attention is used to capture the global and local semantic dependencies of the comment text, so as to enhance the extraction of linguistic features of the model. Finally, an inference layer is used to obtain inter-word relations and a decoding layer is used to obtain aspect-opinion-sentiment triples. Experimental results show that the model not only solves the problems existing in the existing labeling schemes, but also efficiently explores linguistic features, and its comprehensive rating index value achieves better results on four public datasets.

Key words: aspect sentiment triple extraction, fine-grained word pair labeling, node adaptive network, syntactic dependency tree, convolutional attention