Abstract
Aspect-based sentiment analysis aims to determine the sentiment polarity of specific aspects of a given sentence. In the current research, the pre-trained model BERT combined with graph convolutional neural networks is usually used. However, in the word vectorization process of BERT, words not in the pre-trained vocabulary list are sliced, and the sliced words cannot correspond to the syntactic adjacency matrix, leading to the problem of biased syntactic information in aspect word convolution learning. Meanwhile, most models rely too much on syntactic information when constructing the adjacency matrix required for graph convolutional networks and lack understanding of the context and semantic information of aspect words. In view of this, the aspect word information augmentation network model is proposed by constructing a word-level BERT and performing vector summation fusion on the BERT-sliced words to ensure the alignment of sentence vectors with the adjacency matrix information during the convolution process. In addition, the double adjacency matrix strategy is used in the convolution stage to take into account the syntactic and semantic information of the sentence. Experiments demonstrate that the algorithm proposed in this paper outperforms state-of-the-art baseline methods in fine-grained sentiment analysis tasks.