A Self-attention Hybrid Network for Aspect-level Sentiment Classification

Abstract

Aspect-level sentiment classification is a task to determine the sentiment polarity of an aspect target in a sentence. In this paper, we propose a hybrid network based on self-attention mechanism, including self-attention convolutional neural network module and target toward sentence attention module to achieve sentiment classification. Specifically, the self-attention convolutional neural module can explore the remote context semantic information, overcoming the shortcomings of the traditional convolution neural network that can only obtain local features because of the limitation of the convolution kernel and pooling. In the target toward sentence attention module, the attention network could compute the hidden state of the sentence obtained by bidirectional gated recurrent unit (Bi-GRU) and the target embedding matrix, and further obtain the important part of the sentence for the target. In addition, in order to express the relative relationship, we add location embedding to the input layer. Experiments on restaurant and laptop datasets show self-attention hybrid network (SAHN) achieves state-of-the-art accuracy relative to other models in aspect-level sentiment classification.