Aspect-Level Sentiment Analysis of Graph Convolutional Networks Based on Position and Sentiment Knowledge

To address the problems that the graph convolutional network model using syntactic dependencies is susceptible to performance degradation due to the influence of syntactic noise when dealing with spoken and multivariate online comment data and the sensitivity of the model to the extraction of sentiment features of aspect words due to the existence of multiple pairs of dependencies between aspect words and contexts, this paper proposes a graph convolutional network model based on position and sentiment knowledge. Based on the syntactic dependencies combined with the positional relationships between aspect words and contexts, a positional weight adjacency matrix is constructed to reduce the influence of sentence noise. At the same time, the sentiment enhancement graph network is constructed using external sentiment knowledge to enhance the sentiment representation of contextual words that depend on aspect words to solve the problem of decreasing model sensitivity. According to the results of comparison experiments and ablation experiments on openly available datasets, the model proposed in this study executes better than the benchmark model in terms of reducing the impact of sentence noise and improving the sensitivity of sentiment feature capture.