Convolutional attention neural network over graph structures for improving the performance of aspect-level sentiment analysis

Abstract

Recently, aspect-level sentiment analysis methods using graph convolutional network (GCN)-based structures with fairly good performance have been introduced. However, previous GCN-based methods often experience one of the following limitations. First, GCNs usually use edges with binary weights. However, binary weights are not helpful in many tasks. Second, these GCNs only focus on extracting node features from some single words or phrases and ignore their context in the entire sentence or paragraph or only consider the information of independent phrases when determining the relation between two graph edges overlooking the semantic relation among these phrases. Finally, no studies simultaneously use the information on the context, the semantic relation, and the sentiment knowledge among words or phrases to build GCNs for aspect-level sentiment analysis. Therefore, to resolve these limitations, in this study, we propose a new method, the CANN-SSCG model, as follows. First, we built three separate heterogeneous graphs, namely, syntax-based, semantic-based, and context-based graphs. Second, we constructed a general heterogeneous graph (SSC graph) by combining the three constructed graphs. We then converted the nodes of the SSC graph into sentence vectors using a GCN with two layers (creating an SSC-GCN). Finally, we used a convolutional neural network algorithm with attention to position embeddings (CANN) on the output of the SSC-GCN model for aspect-level sentiment analysis. The experiments, which used three different datasets, including reviews and tweets, showed that the proposed method yields promising results based on the F1score.