Comparative analysis of the adversarial methods for non-topical classification of texts

Non-topical text classification is widely used in modern applications. One of the issues related to this problem is the presence of biases and shifts in the distribution in the training text datasets. The most significant type of shift is the topical shift. To handle this issue we apply competitive methods such as Adversarial Domain Adaptation, Energy-based ADA, BERT with contrast loss function, ADA with contrast loss function.
In this paper, we first modify the contrast loss function to reduce the influence of thematic shifts and show that the use of adversarial methods improves the accuracy and reliability of classifiers for the task of determining the gender of the author of a text. We also apply LLaMA-3B and show that the large language models attain lower accuracy in the few-shot mode and require more time for prediction than the pre-trained models based on smaller architectures.