Improved Review Sentiment Analysis with a Syntax-Aware Encoder

Review sentiment analysis has drawn a lot of active research interest because of the explosive growth in the amount of available reviews in our day-to-day activities. The current review sentiment classification work often models each sentence as a sequence of words, thus simply training sequence-structured recurrent neural networks (RNNs) end-to-end and optimizing via stochastic gradient descent (SGD). However, such sequence-structured architectures overlook the syntactic hierarchy among the words in a sentence. As a result, they fail to capture the syntactic properties that would naturally combine words to phrases. In this paper, we propose to model each sentence of a review with an attention-based dependency tree-LSTM, where a sentence embedding is obtained relying on the dependency tree of the sentence as well as the attention mechanism in the tree structure. Then, we feed all the sentence representations into a sequence-structured long short-term memory network (LSTM) and exploit attention mechanism to generate the review embedding for final sentiment classification. We evaluate our attention-based tree-LSTM model on three public datasets, and experimental results turn out that it outperforms the state-of-the-art baselines.