Adversarial Training: addressing biases in non-topical classification and regression tasks

ACL ARR 2025 July Submission94 Authors

Abstract: Many datasets for non-topical text classification contain topical shifts. Their presence in the data forces the classifiers to fit topics-related features instead of focusing on those relevant for the target class. The problem of topical shifts is also significant for the textual regression tasks. In our study, we estimate the effect of the topical shifts on performance of the classifiers and regressors in non-topical prediction tasks and try to reduce their impact by using adversarial methods. As two test tasks, we use sentiment analysis prediction on Amazon Reviews and identification of the education degree of the author on PASTEL. Each task is predicted as classification and regression. We show that Adversarial Domain Adaptation (ADA) helps to reduce the effect of topical shifts and to decrease the error in regression. Finally, we make a recommendation when to use ADA and how to select the hyperparameters for it.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Sentiment Analysis, adversarial training, domain adaptation, few-shot learning

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 94