This study explores the automated prediction of text complexity levels for short Russian texts on the Common European Framework of Reference for Languages (CEFR) scale. The dataset consists of 7,322 nonfictional fragments (15–30 words) extracted from textbooks for learners of Russian as a second language and filtered according to linguistic feature distributions typical of each CEFR level, with additional validation conducted by 4 human experts. Each text fragment was annotated with 127 linguistic features, including lexical, morphological, syntactic, and length-based characteristics. We evaluate several approaches to text complexity assessment: traditional machine learning classifiers, fine-tuned transformer models, and instruction-based large language models (LLMs). Among all models, RuBERT achieved the best strict F1-score (47.8%) and the lowest mean absolute error (0.56), while instruction-based LLMs such as YandexGPT captured overall complexity trends but underperformed in exact classification. Feature ablation experiments demonstrated that lexical features are the most informative for CEFR prediction. Our findings confirm that fine-tuned language models currently offer the most reliable results for short-text CEFR assessment in Russian, whereas instruction-based LLMs show potential for qualitative analysis of text difficulty patterns.

In this work, we compared various architectures of surface morpheme segmentation algorithms using data from 20 languages belonging to different language families and with varying data availability. We compare conventional subword and statistical baselines, CNN and LSTM models trained from scratch, fine-tuned multilingual encoders, prompted frontier LLMs, and LoRA-tuned decoder-only LLMs. To the best of our knowledge, this is the first systematic study of decoder-only LLMs for morpheme segmentation. Our results show that XLM-Roberta-large fine-tuned for character-level sequence labeling with whole-word context remains the strongest and most stable language-agnostic approach, achieving the highest average normalized score and improving previous SOTA results for five languages. At the same time, LoRA fine-tuning of Qwen3-32B yields competitive results and substantially outperforms XLM-Roberta-large for Dravidian and Indo-Aryan languages, demonstrating that decoder-only LLMs can be effectively adapted to the task. Prompting frontier LLMs, however, remains inconsistent. Finally, root-disjoint evaluation and error analysis show that generalization to unseen roots and dataset-specific segmentation conventions remain central challenges.

Tokenization is a central component of pretrained language models, especially for morphologically rich low-resource languages. We present a controlled study of morphology-aware tokenization for Belarusian masked language models. To isolate the effect of tokenization, we pretrain three RoBERTa-family models on the same Belarusian HPLT corpus, varying only the tokenizer: standard WordPiece, WordPiece trained on morpheme-segmented text, and a dictionary-based hybrid morpheme tokenizer with WordPiece fallback. We evaluate the models on masked language modeling, Universal Dependencies morphology-sensitive tasks, and BelarusianGLUE. Morphology-aware tokenization substantially improves morpheme-boundary preservation and yields clear gains on explicitly morphological tasks, especially fine-grained feature prediction and lemmatization, where models must recover information encoded in inflectional structure. However, these gains do not uniformly transfer to broad sentence-level NLU: on BelarusianGLUE, morphology-aware tokenizers improve some tasks but remain comparable to or weaker than standard WordPiece on others. These findings suggest that morpheme-informed tokenization is a targeted inductive bias rather than a universal replacement for statistical subword segmentation.

Pre-trained language models have significantly advanced natural language processing (NLP), particularly in analyzing languages with complex morphological structures. This study addresses lemmatization for the Russian language, the errors in which can critically affect the performance of information retrieval, question answering, and other tasks. We present the results of experiments on generative lemmatization using pre-trained language models. Our findings demonstrate that combining generative models with the existing solutions allows achieving performance that surpasses current results for the lemmatization of Russian. This paper also introduces Rubic2, a new ensemble approach that combines the generative BART-base model, fine-tuned on a manually annotated data set of 2.1 million tokens, with the neural model called Rubic which is currently used for morphological annotation and lemmatization in the Russian National Corpus. Extensive experiments show that Rubic2 outperforms current solutions for the lemmatization of Russian, offering superior results across various text domains and contributing to advancements in NLP applications.

Automatic morpheme segmentation algorithms are applicable in various tasks, such as building tokenizers and language education. For Slavic languages, the development of such algorithms is complicated by the rich derivational capabilities of these languages. Previous research has shown that, on average, these algorithms have already reached expert-level quality. However, a key unresolved issue is the significant decline in performance when segmenting words containing roots not present in the training data. This problem can be partially addressed by using pre-trained language models to better account for word semantics. In this work, we explored the possibility of fine-tuning BERT-like models for morpheme segmentation using data from Belarusian, Czech, and Russian. We found that for Czech and Russian, our models outperform all previously proposed approaches, achieving word-level accuracy of 92.5-95.1%. For Belarusian, this task was addressed for the first time. The best-performing approach for Belarusian was an ensemble of convolutional neural networks with word-level accuracy of 90.45%.

This paper describes our system for SemEval-2021 Task 5 on Toxic Spans Detection. We developed ensemble models using BERT-based neural architectures and post-processing to combine tokens into spans. We evaluated several pre-trained language models using various ensemble techniques for toxic span identification and achieved sizable improvements over our baseline fine-tuned BERT models. Finally, our system obtained a F1-score of 67.55% on test data.