Pseudo dense counterfactual augmentation for aspect-based sentiment analysis

Abstract

Aspect-based sentiment analysis (ABSA) is a fine-grained text classification task, and the cutting-edge ABSA models have achieved outstanding performance. Unfortunately, the robustness of these ABSA models is neglected. ABSA models must face numerous challenges to be robust, and we concentrate on one of these challenges caused by negation words, such as “not”, “un-”. In the actual context, these negation words intuitively result in two problems: negative sensitivity and spurious correlation. First, a negation word tends to reverse the sentiment polarity of a sentence. Meanwhile, in the ABSA datasets, most sentences containing negation words express Negative polarities, which will lead the predictive model to learn the spurious correlation between negation words and polarities. To resolve these ambiguous issues, we are inspired by causal inference and propose a novel data augmentation framework, namely Pseudo Dense Counterfactual Augmentation (PDCaug) for ABSA. Specifically, we initialize a pseudo sequence and employ a multi-head multi-layer attention network to achieve counterfactual augmentation for a vanilla sentence in the hidden space. This pseudo sequence will be adversarially trained. PDCaug is a plug-and-play method for various ABSA models, so we evaluate it on discriminative models and generative prompt-based models. Our extensive experiments show that our PDCaug can significantly and consistently outperform several data augmentation methods and ABSA models.