Interactive Graph Convolutional Network for Aspect-level Sentiment Analysis

A novel interactive graph convolutional network model is proposed to overcome the limitations of existing sentiment analysis models based on graph networks. This model effectively addresses the challenges of handling the internal semantic relations of aspect phrases and the sentiment interactions between different aspects within a sentence. Firstly, by leveraging the syntactic dependencies between context and aspect words, the aspect-inside module integrates the internal semantic correlations among aspect phrases and establishes sentiment interactions between different aspects, resulting in four types of adjacency matrix graphs. This allows for accurate modeling of aspect phrase semantics and the affective interactions between different aspects within a sentence. Secondly, an aspect interaction graph is constructed to capture cross-aspect relations, effectively resolving the issue of sentiment interactions between different aspects. Lastly, the inclusion of a global node in the graph convolutional network further addresses the problem of fluctuating accuracy when multiple aspect words are present in a sentence. The results of experimental on four publicly available datasets demonstrate the effectiveness of the proposed model, showing significant improvements in both accuracy and F₁ size.