Enhancing Robustness in Aspect-based Sentiment Analysis by Better Exploiting Data Augmentation

Abstract

In this paper, we propose to leverage data augmentation to improve the robustness of aspect-based sentiment analysis models. Our method not only exploits augmented data but also makes models focus more on predictive features. We show in experiments that our method compares favorably against strong baselines on both robustness and standard datasets. In the contrary, the widely used adversarial training that only leverages the augmented data fails to improve performance due to the distribution shift caused by the augmented data.

Tasks

Reproductions