Abstract: Existing models use graph neural network based on dependency trees for aspect-based sentiment analysis,which improves the classification performanceof the model to a certain extent.However,due to technical limitations of dependency parsing,the inaccuracy of the dependency parsing results leads to a large amount of noise in the dependency tree,which makes the performance improvement of the model is limited.In addition,some sentences themselves do not conform to the standard syntactic structure.Previous studies utilized syntactic and semantic information with the same confidence level without fully considering the difference in their contributions to determining the polarity of aspect words,resulting in poor model performance on the corres-ponding datasets.To overcome these challenges,a dual feature adaptive fusion network based on dependency type pruning is proposed in this paper.Specifically,the model uses a novel hybrid approach,named dependency type pruning and adjacency matrix smoothing,to mitigate the noise generated by dependency parsing.In addition,the model fully considers the availability of syntactic information of sentences through a dual feature adaptive fusion module to combine syntactic features and semantic features for aspect-level sentiment analysis in a more flexible way.Extensive experiments on five publicly available datasets demonstrate that the proposed method significantly outperforms baseline models.

Key words: Aspect-based sentiment analysis, Graph neural networks, Dependency type pruning, Dual feature adaptive fusion, Deep learning, Natural language processing