An LSTM-CNN attention approach for aspect-level sentiment classification

Opinions in complex reviews often vary on different aspects of a thing. Coarse-grained sentiment analysis on a sentence can’t capture the sentiment polarity of it accurately. Therefore, aspect-level sentiment classification is a better choice because it is a fine-grained task in sentiment analysis. However, common methods used for sentiment classification are not suitable for fine-grained sentiment analysis, such as LSTM. LSTM works in a sequential way and manipulates each context word with the same operation, so that it cannot explicitly reveal the importance of each context word. To this end, we propose an Aspect-based LSTM-CNN Attention model for aspect-level sentiment classification. Our approach combines LSTM with CNN for simultaneously leveraging LSTM to handle long-range dependencies and CNN’s ability to identify local features. The features extracted by LSTM will be filtered again by convolution and pooling operations to find important local features. Then we introduce the attention mechanism to focus on the important information about the specific aspect. We experiment on the SemEval 2014 dataset and Chinese hotel reviews, and results show that our model can effectively improve the accuracy of aspect-level sentiment classification.