A novel weight-oriented graph convolutional network for aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) determines the sentiment polarity of specific aspects mentioned in the review. However, some existing ABSA studies have limitations, such as the model only detecting aspect-relevant semantics when using the attention mechanism and ignoring the aspect's long-distance dependence when introducing aspect position information. This study proposes a multiweight graph convolutional network (MWGCN) to address the above-mentioned limitations. MWGCN aims to design two weighting methods, multigrain dot-product weighting (MGDW) and the way (LCG), to create a local context weighted adjacency graph. The MGDW method retains the overall context semantics while emphasizing aspect-related features. Furthermore, the adjacency graph constructed by LCG emphasizes the importance of local context words and helps to avoid the aspect's long-distance dependence. A multilayer graph convolutional network (GCN) is also used to extract contextual features that integrate syntactic information and capture aspect features that focus on local context words. We performed several experiments on five datasets; the experimental results verify the MWGCN generalization and further prove that with MGDW and LCG, the features extracted using GCN help improve the MWGCN effect.