A.N. Zhdanova · INFORMACIONNYE TEHNOLOGII 2026 · 2026
DOI: 10.17587/it.32.519-524
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The paper addresses the problem of short text classification under conditions of limited labeled data and class imbalance. A two-stage training strategy for a transformer-based model is proposed, combining pretraining on synthetic data with fine-tuning on real data. A comparison with simpler architectures (TF-IDF + logistic regression, FastText, CNN) justifies the choice of the transformer model. Experimental evaluation on a six-class classification task demonstrates a macro F1-score of 0.87 (accuracy 88.5 %), which exceeds the performance of training on real data only by 0.16 in macro F1. The largest improvement is observed for minority classes: the F1-score of the rare class increases from 0.43 to 0.82. An optimal volume of synthetic data is identified, and the contribution of both training stages is confirmed by an ablation study.
No comments yet — start the discussion below.