Transformer and Multi-scale Convolution for Target-Oriented Sentiment Analysis

Yinxu Pan, Binheng Song, Ningqi Luo, Xiaojun Chen, Hengbin Cui

Abstract: Target-oriented sentiment analysis aims to extract the sentiment polarity of a specific target in a sentence. In this paper, we propose a model based on transformers and multi-scale convolutions. The transformer which is based solely on attention mechanisms generalizes well in many natural language processing tasks. Convolution layers with multiple filters can efficiently extract n-gram features at many granularities on each receptive field. We conduct extensive experiments on three datasets: SemEval ABSA challenge Restaurant and Laptop dataset, Twitter dataset. Our framework achieves state-of-the-art results, including improving the accuracy of Restaurant dataset to 84.20% (5.81% absolute improvement), improving the accuracy of the Laptop dataset to 78.21% (4.23% absolute improvement), and improving the accuracy of the Twitter dataset to 72.98% (0.87% absolute improvement).