A Refined Multi-Grained Attention Graph for Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) is a finegrained task to determine the sentiment polarities towards given aspect terms, which plays an important role in understanding the opinions of the public in the era of the Internet. Recent studies tackled natural language as sequences and tend to apply larger total computation to improve model performance. However, with the instruction of prior domain knowledge, much more refined and efficient neural networks can be designed. In this paper, we embed grammar knowledge through dependency parsing to generate a graph from text, and propose a global-dependency multi-grained attention graph (G-D graph) that imitates the attention pattern of the human intensive reading process. Then we construct a specially designed refined intelligent graph neural network (GNN) with the guidance of refined intelligence theory to exploit the potential of the attention graph structure to the maximum extent. Extensive experiments are conducted to compare the performances of our refined intelligent model with the dynamic graph construction and representation learning model, which does not have any instruction from humans in the training progress. Experimental results show that our model is more effective and more efficient than baseline models in ABSA tasks.