Using Dependency Information to Enhance Attention Mechanism for Aspect-Based Sentiment Analysis

Attention mechanism has been justified beneficial to aspect-based sentiment analysis (ABSA). In recent years there arise some research interests to implement the attention mechanism based on dependency relations. However, the disadvantages lie in that the dependency trees must be obtained beforehand and are affected by error propagation problem. Inspired by the finding that the calculation of the attention mechanism is actually a part of the graph-based dependency parsing, we design a new approach to transfer dependency knowledge to ABSA in a multi-task learning manner. We simultaneously train an attentionbased LSTM model for ABSA and a graph-based model for dependency parsing. This transfer can alleviate the inadequacy of network training caused by the shortage of sufficient training data. A series of experiments on SemEval 2014 restaurant and laptop datasets indicate that our model can gain considerable benefits from dependency knowledge and obtain comparable performance with the state-of-the-art models which have complex network structures.