GL-GCN: Global and Local Dependency Guided Graph Convolutional Networks for aspect-based sentiment classification

Aspect-based sentiment classification, which aims at identifying the sentiment polarity of a sentence towards the specified aspect, has become a crucial task for sentiment analysis. Existing methods have proposed effective models and achieved satisfactory results, but they mainly focus on exploiting local structure information of a given sentence, such as locality, sequentiality or syntactical dependency constraints within the sentence. Recently, some research works, which utilizes global dependency information, has attracted increasing interest and significantly boosts the performance of text classification. In this paper, we simultaneously introduce both global structure information and local structure information into the task of aspect-based sentiment classification, and propose a novel aspect-based sentiment classification approach, i.e., Global and Local Dependency Guided Graph Convolutional Networks (GL-GCN). In particular, we exploit the syntactic dependency structure as well as sentence sequential information (e.g., the output of BiLSTM) to mine the local structure information of a sentence. On the other hand, we construct a word-document graph using the entire corpus to reveal the global dependency information between words. In addition, an attention mechanism is leveraged to effectively fuse both global and local dependency structure signals. Extensive experiments are conducted on five benchmark datasets in terms of both Accuracy and F1-Score, and the results illustrate that our proposed framework outperforms state-of-the-art methods for aspect-based sentiment classification. The model is implemented using PyTorch and is trained on GPU GeForce GTX 2080 Ti.

Automated summary

The study introduces a Global and Local Dependency Guided Graph Convolutional Networks (GL-GCN) approach, which integrates global and local structure information into aspect-based sentiment classification. Experimental results demonstrate that this method outperforms existing approaches in terms of accuracy and F1-Score.