STC: Stacked Two-stage Convolution for Aspect Term Extraction

Aspect term extraction (ATE) aims to extract aspect terms from reviews as opinion targets for sentiment analysis. Although some of the previous works prove that dependency relationship between aspect terms and context is useful for ATE, they have barely tried to use graph neural networks to capture valuable information in dependency patterns automatically. In this paper, we propose a novel sequence labeling method for ATE, which exploits convolutional neural network (CNN) to capture local information of a sentence, and further aggregate k-order neighbor nodes’ information via graph convolutional network (GCN) over dependency tree. Differently from approaches based on sequential networks like recurrent neural network (RNN), our convolution model can be calculated in parallel, which improves the training and inference speed. Experimental results show that our approach outperforms other baseline methods, which don't rely on pre-trained transformer model.