Aspect-based Sentiment Analysis with Graph Convolutional Networks and Path Distance Dependency Matrix

Aspect-based sentiment analysis aims to predict the polarity of sentiment concerning aspects. However, existing studies only consider the dependency semantics between the aspect word and emotional word, but ignore the contribution of each emotional word in expressing sentiment, the associated semantics between multi-aspect words, and the contribution of each aspect in the overall context. We present the Path Distance Dependency Graph Convolutional Network model(PDD-GCN) to address the aforementioned issue. Firstly, the path distance de pendency matrix S is proposed, which is calculated by Hadamard product of the syntactic dependency adjacency matrix M and the shortest distance weight matrix WP, and assigned higher weight for the word that is close to the aspect word. Secondly, the multi aspect global weight matrix Wgis addressed, which takes the proportion of a certain aspect's emotional words, modifiers and the total vocabularies in sentence as its weight parameter, the matrix S and Wgare further calculate to obtain the multi-aspect global weight matrix Sa, and solve the problems that the model loses the associated semantics between multi-aspect words and ignores their different contributions. Finally, the matrix S or Saand the sentence vector matrix X are used as the GCN input, it further aggregates the neighbor semantic of the aspect word and captures the correlation between features from different sources and realizes the fusion of syntactic semantics and sentence ontol ogy semantics, which further improves the model's aspect-level sentiment analysis capability. Extensive experiments on public datasets Rest14, Laptop and Twitter indicate that our model outperforms the state-of-the-art models on accuracy and F1 score. The code is available at https://github.com/MEywi/PDD-GCN.