Combining Adversarial Training and Relational Graph Attention Network for Aspect-Based Sentiment Analysis with BERT

Abstract

Aspect-Based Sentiment Analysis (ABSA), also called Aspect Level Sentiment Classification (ALSC), is a common task in Natural Language Processing (NLP). Aspect-Based Sentiment Analysis mainly aims to extract and classify the sentiments objects in texts. In this paper, we propose a novel BERT-based ABSA model, which combines an adversarial training procedure with relational graph attention neural network (R-GAT). To our best knowledge, it is the first model that simultaneously using adversarial training, relational graph attention neural network and BERT for aspect-based sentiment analysis. In our proposed model, the BERT Encoder is used to extract the context feature vector, R-GAT is applied to integrate the typed syntactic dependency information. The proposed model also includes an adversarial training method to ensure the robustness of neural network, which is realized by artificially increasing data samples similar to the real-world examples using adversarial processes in the embedding space. Our experimental results on three benchmark datasets demonstrate that the proposed model is competitive compared to the other state-of-the-art methods.