Causal Models and Adversarial Training: Selecting the right properties for robust non-topical text classification

The vast majority of datasets for non-topical classification of texts contain distribution shifts.In most cases, those are 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.It causes a dramatic decrease in the accuracy of the trained models when the test data are taken from a different data source.To address this problem, we experiment with two techniques: causal models and adversarial domain adaptation.In our work, we apply CausalLM, Adversarial Domain Adaptation (ADA), and Energy-based ADA (EADA) for gender classification and compare the results.The results are novel for the non-topical classification task.We show that both causal and adversarial methods manage to make the model more resilient to the distribution shifts although it causes a decrease of accuracy when tested on the domain prevailing in the training dataset.Moreover, we describe the first attempt to reduce the impact of topical shifts in the task of non-topical classification with usage of causal methods.Besides, we provide a link to the GitHub repository with the code of our experiments to ensure their reproducibility: https://github.com/MikeLepekhin/CausalAndAdversarialMethods.