A Cross-Language Attribute-Level Sentiment Analysis Approach Using TinyBERT and GCN

Abstract

A cross-language attribute-level sentiment analysis model (TinyBERT-GCN) based on TinyBERT and GCN is proposed to address the problems of existing cross-language attribute-level sentiment analysis methods, such as insufficient text feature extraction and easy to ignore cross-language semantic correlations at the word level. The model extracts contextual semantics through TinyBERT, fuses multilingual features using the Multi-Granular Interaction Module, and enhances the understanding of text syntax using GCN. The experimental results show that the proposed TinyBERT-GCN model can achieve ACC and F1 of 0.871 and 0.812 on SemEVAL-2016 dataset; and 0.848 and 0.821 on Taobao dataset, and 0.863 and 0.802 on SemEVAL-2014 dataset respectively. Compared with the other models, the proposed model not only improves the performance, but also reduces the computational cost, has better scalability, and is suitable for large-scale multilingual data processing. This model has important practical applications in market analysis, public opinion monitoring and decision making.