Aspect-level sentiment classification via location enhanced aspect-merged graph convolutional networks

Aspect-level sentiment classification (ALSC) is a fine-grained sentiment analysis task that needs to predict the sentiment polarities of the given aspect terms in the sentence. Recently, emerging research has taken syntactic dependency tree as input and used graph convolutional neural network (GCN) to process ALSC tasks. However, existing GCN-based researches only consider the syntactic connections between words, ignoring the semantic relevance within aspectual entities. To address this deficiency, we propose a graph convolutional network based on Merger aspect entities and Location-aware transformation (MLGCN). Specifically, we use a specific token to replace the aspect entity, whether single-word or multi-word. The merged syntactic dependency graph is obtained through parsing for the sentence after merging aspect words. Then, we feed the sentence into an encoder and apply a novel location-aware function designed in this paper to the encoding result to enhance the model's attention to the opinion entities. Finally, the dependency graph and the processed sentence encoding are fed to the graph convolutional network for training. Experimental results on five benchmark datasets show that the model proposed in this paper has good performance and achieves satisfactory results, exceeding the vast majority of previous work.