Aspect-Level Sentiment Classification with Conv-Attention Mechanism

The aim of aspect-level sentiment classification is to identify the sentiment polarity of a sentence about a target aspect. Existing methods model the context sequence with recurrent network and employ attention mechanism to generate aspect-specific representations. In this paper, we introduce a novel mechanism called Conv-Attention, which can model the sequential information of context words and generate the aspect-specific attention at the same time via a convolution operation. Based on the new mechanism, we design a new framework for aspect-level sentiment classification called Conv-Attention Network (CAN). Compared to the previous attention-based recurrent models, the Conv-Attention Network can compute much faster. Extensive experimental results show that our model achieves the state-of-the-art performance while saving considerable time in model training and inferring.