Joint LSTM with multi-CNN network by hierarchical attention for aspect-based sentiment classification

Aspect-Based Sentiment Analysis (ABSA) is a challenging task in recent natural language processing research. It aims at extracting people's sentiment polarity on a specific category. In this paper, we explore that the key phrases in sentence are highly relevant to certain aspect words which influence the final result. We propose a joint LSTM with Multi-CNN network by hierArchical aTtention (MAT) model to achieve this goal. Specifically, we design a double embedding with fully connected layer module to improve the final performance. MAT shows the effectiveness on the experiments of datasets from SemEval 2014.