Attention-based ResNet for Chinese Text Sentiment Classification

摘要

Identifying sentiment polarity of a document is a building block of sentiment classification and natural language processing tasks, it aims to automate the prediction of sentiment orientation in a document. In general, recently fast-growing Deep Neural Networks (DNN) method has been extensively used as a sentiment learning approach. But the dominant approach for sentiment classification tasks are recurrent neural networks, in particular LSTM, and convolutional neural networks. However, these architectures are rather shallow in comparison to the Residual Neural Networks (ResNet) which have pushed in computer vision. We present a model using ResNet for high-level document representation, and attention mechanism to capture the crucial components for document. The experimental results show that using up to 2 ResNet block and attention mechanism achieve state-of-the-art performance on three public sentiment classification datasets.