Graph Attention Network with Dependency Parsing for Aspect-level Sentiment Classification
Ling Gan, Qiao Tang · 2021
The purpose of Aspect-level Sentiment Classification (ASC) is to judge the sentiment polarity of words or phrases with subjective emotions in a sentence. The existing mainstram method is to extract the context semantics of aspect targets words based on attention mechanism, but the attention mechanism is not sensitive to location information and not good at capture syntactic dependency information. To solve the existing problems, this paper proposes an Aspect-Specific Graph Attention Network (ASGAT) model to make up for the shortcoming of attention mechanism. Firstly, the dependency syntax tree of a sentence is obtained by dependency parsing, and then the adjacency matrix with local location information is generated by combining the local location relationship of aspect target words in the context. Then the context semantics of aspect target words is extracted by multi-layer Graph Attention Network (GAT), which makes the semantic features of extracted aspect targets more abundant. The model can capture the long-distance syntactic dependency and the local information of aspect words better. The proposed model has achieved better results on five benchmark datasets. The effectiveness of the model is verified by model comparision and experimental analysis.