Multiple graph convolutional networks for aspect-based sentiment analysis

Aspect-based sentiment analysis is a fine-grained sentiment analysis task that identifies the sentiment polarity of different aspects in a sentence. Recently, several studies have used graph convolution networks (GCN) to obtain the relationship between aspects and context words with the dependency tree of sentences. However, errors introduced by the dependency parser and the complexity and variety of sentence structures have led to incorrect predictions of sentiment polarity. Therefore, we propose a multiple GCN (MultiGCN) model to solve this problem. The proposed MultiGCN comprises a rational GCN (RGCN) to extract syntactic structure information of sentences, a contextual encoder to extract semantic content information of sentences, a common information extraction module to combine structure and content information, and a fusion mechanism that allows interaction among the aforementioned components. Further, we propose difference and similarity losses and combine them with traditional loss function to jointly minimize the difference between the values predicted by the model and those of the labels. The experimental results show that the prediction performance of our proposed method is more than that of the state-of-the-art models.