In this paper, we propose a model-agnostic cost-effective approach to developing bilingual base large language models (LLMs) to support English and any target language. The method includes vocabulary expansion, initialization of new embeddings, model training and evaluation. We performed our experiments with three languages, each using a non-Latin script—Ukrainian, Arabic, and Georgian.Our approach demonstrates improved language performance while reducing computational costs. It mitigates the disproportionate penalization of underrepresented languages, promoting fairness and minimizing adverse phenomena such as code-switching and broken grammar. Additionally, we introduce new metrics to evaluate language quality, revealing that vocabulary size significantly impacts the quality of generated text.

While the evaluation of multimodal English-centric models is an active area of research with numerous benchmarks, there is a profound lack of benchmarks or evaluation suites for low- and mid-resource languages. We introduce ZNO-Vision, a comprehensive multimodal Ukrainian-centric benchmark derived from the standardized university entrance examination (ZNO). The benchmark consists of over 4300 expert-crafted questions spanning 12 academic disciplines, including mathematics, physics, chemistry, and humanities. We evaluated the performance of both open-source models and API providers, finding that only a handful of models performed above baseline. Alongside the new benchmark, we performed the first evaluation study of multimodal text generation for the Ukrainian language: we measured caption generation quality on the Multi30K-UK dataset. Lastly, we tested a few models from a cultural perspective on knowledge of national cuisine. We believe our work will advance multimodal generation capabilities for the Ukrainian language and our approach could be useful for other low-resource languages.

Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), yet achieving high performance for low-resource languages remains challenging due to limited annotated data and linguistic complexity. Ukrainian exemplifies these issues with its rich morphology and scarce NLP resources. Recent advances in Large Language Models (LLMs) demonstrate their ability to generalize across diverse languages and domains, offering promising solutions without extensive annotations. This research explores adapting state-of-the-art LLMs to Ukrainian through prompt engineering, including chain-of-thought (CoT) strategies, and model refinement via Supervised Fine-Tuning (SFT). Our best model achieves 0.89 F1 on the NER-UK 2.0 benchmark, matching the performance of advanced encoder-only baselines. These findings highlight practical pathways for improving NER in low-resource contexts, promoting more accessible and scalable language technologies.

This paper introduces UAlign, the comprehensive benchmark for evaluating the alignment of Large Language Models (LLMs) in the Ukrainian language. The benchmark consists of two complementary components: a moral judgment dataset with 3,682 scenarios of varying ethical complexities and a dataset with 1,700 ethical situations presenting clear normative distinctions. Each element provides parallel English-Ukrainian text pairs, enabling cross-lingual comparison. Unlike existing resources predominantly developed for high-resource languages, our benchmark addresses the critical need for evaluation resources in Ukrainian. The development process involved machine translation and linguistic validation using Ukrainian language models for grammatical error correction. Our cross-lingual evaluation of six LLMs confirmed the existence of a performance gap between alignment in Ukrainian and English while simultaneously providing valuable insights regarding the overall alignment capabilities of these models. The benchmark has been made publicly available to facilitate further research initiatives and enhance commercial applications.Warning: The datasets introduced in this paper contain sensitive materials related to ethical and moral scenarios that may include offensive, harmful, illegal, or controversial content.

Detecting manipulation techniques in online text is vital for combating misinformation, a task complicated by generative AI. This paper compares machine learning approaches for multi-label classification of 10 techniques in Ukrainian Telegram content (UNLP 2025 Shared Task 1). Our evaluation included TF-IDF, fine-tuned XLM-RoBERTa-Large, PEFT-LLM (Gemma, Mistral) and a RAG approach (E5 + Mistral Nemo). The fine-tuned XLM-RoBERTa-Large model, which incorporates weighted loss to address class imbalance, yielded the highest Macro F1 score (0.4346). This result surpassed the performance of TF-IDF (Macro F1 0.32-0.36), the PEFT-LLM (0.28-0.33) and RAG (0.309). Synthetic data slightly helped TF-IDF but reduced transformer model performance. The results demonstrate the strong performance of standard transformers like XLM-R when appropriately configured for this classification task.

In this paper, we present our solutions for the two UNLP 2025 shared tasks: manipulation span detection and manipulation technique classification in Ukraine-related media content sourced from Telegram channels. We experimented with fine-tuning large language models (LLMs) with up to 12 billion parameters, including both encoder- and decoder-based architectures. Our experiments identified Gemma 3 12b with a custom classification head as the best-performing model for both tasks. To address the limited size of the original training dataset, we generated 50k synthetic samples and marked up an additional 400k media entries containing manipulative content.

This paper presents a new Universal Dependencies (UD) treebank based on Ukrainian parliamentary transcripts, complementing the existing UD resources for Ukrainian. The corpus includes manually annotated texts from key historical sessions of the Verkhovna Rada, capturing not only official rhetoric but also features of colloquial spoken language. The annotation combines UDPipe2 and TagText parsers, with subsequent manual correction to ensure syntactic and morphological accuracy. A detailed comparison of tagsets and the disambiguation strategy employed by TagText is provided. To demonstrate the applicability of the resource, the study examines vocative and nominative case variation in direct address using a large-scale UD-annotated corpus of parliamentary texts.

Large Language Models (LLMs) have demonstrated remarkable performance across various domains, but they often inherit biases present in the data they are trained on, leading to unfair or unreliable outcomes—particularly in sensitive areas such as hiring, medical decision-making, and education. This paper evaluates gender bias in LLMs within the Ukrainian language context, where the gendered nature of the language and the use of feminitives introduce additional complexity to bias analysis. We propose a benchmark for measuring bias in Ukrainian and assess several debiasing methods, including prompt debiasing, embedding debiasing, and fine-tuning, to evaluate their effectiveness. Our results suggest that embedding debiasing alone is insufficient for a morphologically rich language like Ukrainian, whereas fine-tuning proves more effective in mitigating bias for domain-specific tasks.

We developed a methodology and a framework for automatically evaluating and filtering large-scale parallel corpora for neural machine translation (NMT). We applied six modern Quality Estimation (QE) models to score 55 million English-Ukrainian sentence pairs and conducted human evaluation on a stratified sample of 9,755 pairs. Using the obtained data, we ran a thorough statistical analysis to assess the performance of selected QE models and build linear, quadratic and beta regression models on the ensemble to estimate human quality judgments from automatic metrics. Our best ensemble model explained approximately 60% of the variance in expert ratings. We also found a non-linear relationship between automatic metrics and human quality perception, indicating that automatic metrics can be used to predict the human score. Our findings will facilitate further research in parallel corpus filtering and quality estimation and ultimately contribute to higher-quality NMT systems. We are releasing our framework, the evaluated corpus with quality scores, and the human evaluation dataset to support further research in this area.

In this paper we introduce the first effort to adapt large language models (LLMs) to the Ukrainian dialect (in our case Hutsul), a low-resource and morphologically complex dialect spoken in the Carpathian Highlands. We created a parallel corpus of 9852 dialect-to-standard Ukrainian sentence pairs and a dictionary of 7320 dialectal word mappings. We also addressed data shortage by proposing an advanced Retrieval-Augmented Generation (RAG) pipeline to generate synthetic parallel translation pairs, expanding the corpus with 52142 examples. We have fine-tuned multiple open-source LLMs using LoRA and evaluated them on a standard-to-dialect translation task, also comparing with few-shot GPT-4o translation. In the absence of human annotators, we adopt a multi-metric evaluation strategy combining BLEU, chrF++, TER, and LLM-based judgment (GPT-4o). The results show that even small(7B) finetuned models outperform zero-shot baselines such as GPT-4o across both automatic and LLM-evaluated metrics. All data, models, and code are publicly released at: https://github.com/woters/vuyko-hutsul.

Text preprocessing is a fundamental component of high-quality speech synthesis. This work presents a novel rule-based phonemizer combined with a sentence-level lexical stress prediction model to improve phonetic accuracy and prosody prediction in the text-to-speech pipelines. We also introduce a new benchmark dataset with annotated stress patterns designed for evaluating lexical stress prediction systems at the sentence level.Experimental results demonstrate that the proposed phonemizer achieves a 1.23% word error rate on a manually constructed pronunciation dataset, while the lexical stress prediction pipeline shows results close to dictionary-based methods, outperforming existing neural network solutions.

This paper presents the results of the UNLP 2025 Shared Task on Detecting Social Media Manipulation. The task included two tracks: Technique Classification and Span Identification. The benchmark dataset contains 9,557 posts from Ukrainian Telegram channels manually annotated by media experts. A total of 51 teams registered, 22 teams submitted systems, and 595 runs were evaluated on a hidden test set via Kaggle. Performance was measured with macro F1 for classification and token‐level F1 for identification. The shared task provides the first publicly available benchmark for manipulation detection in Ukrainian social media and highlights promising directions for low‐resource propaganda research. The Kaggle leaderboard is left open for further submissions.

We participated in the Fourth UNLP shared task on detecting social media manipulation in Ukrainian Telegram posts, addressing both multilabel technique classification and token-level span identification. We propose two complementary solutions: for classification, we fine-tune the decoder-only model with class-balanced grid-search thresholding and ensembling. For span detection, we convert causal LLM into a bidirectional encoder via masked language modeling pretraining on large Ukrainian and Russian news corpora before fine-tuning. Our solutions achieve SOTA metric results on both shared task track. Our work demonstrates the efficacy of bidirectional pretraining for decoder-only LLMs and robust threshold optimization, contributing new methods for disinformation detection in low-resource languages.

Recent advances in multilingual language modeling have highlighted the importance of high-quality, large-scale datasets in enabling robust performance across languages. However, many low- and mid-resource languages, including Ukrainian, remain significantly underrepresented in existing pretraining corpora. We present Kobza, a large-scale Ukrainian text corpus containing nearly 60 billion tokens, aimed at improving the quality and scale of Ukrainian data available for training multilingual language models. We constructed Kobza from diverse, high-quality sources and applied rigorous deduplication to maximize data utility. Using this dataset, we pre-trained Modern-LiBERTa, the first Ukrainian transformer encoder capable of handling long contexts (up to 8192 tokens). Modern-LiBERTa achieves competitive results on various standard Ukrainian NLP benchmarks, particularly benefiting tasks that require broader contextual understanding or background knowledge. Our goal is to support future efforts to develop robust Ukrainian language models and to encourage greater inclusion of Ukrainian data in multilingual NLP research.

Synonym-based adversarial tests reveal fragile word patterns that accuracy metrics overlook, while virtually no such diagnostics exist for Ukrainian, a morphologically rich and low‐resource language. We present the first systematic robustness evaluation under synonym substitution in Ukrainian. Adapting TextFooler and BERT‐Attack to Ukrainian, we (i) adjust a 15000‐entry synonym dictionary to match proper word forms; (ii) integrate similarity filters; (iii) adapt masked‐LM search so it generates only valid inflected words. Across three text classification datasets (reviews, news headlines, social‐media manipulation) and three transformer models (Ukr‐RoBERTa, XLM‐RoBERTa, SBERT), single‐word swaps reduce accuracy by up to 12.6, while multi‐step attacks degrade performance by as much as 40.27 with around 112 model queries. A few‐shot transfer test shows GPT‐4o, a state‐of‐the‐art multilingual LLM, still suffers 6.9–15.0 drops on the same adversarial samples. Our results underscore the need for sense‐aware, morphology‐constrained synonym resources and provide a reproducible benchmark for future robustness research in Ukrainian NLP.

This paper presents a pipeline for generating gender-balanced datasets through sentence-level gender swapping, addressing the gender-imbalance issue in Ukrainian texts. We select sentences with gender-marked entities, focusing on job titles, generate their inverted alternatives using LLMs and human-in-the-loop, and fine-tune Aya-101 on the resulting dataset for the task of gender swapping. Additionally, we train a Named Entity Recognition (NER) model on gender-balanced data, demonstrating its ability to better recognize gendered entities. The findings unveil the potential of gender-balanced datasets to enhance model robustness and support more fair language processing. Finally, we make a gender-swapped version of NER-UK~2.0 and the fine-tuned Aya-101 model available for download and further research.

In this paper, we introduce OmniGEC, a collection of multilingual silver-standard datasets for the task of Grammatical Error Correction (GEC), covering eleven languages: Czech, English, Estonian, German, Greek, Icelandic, Italian, Latvian, Slovene, Swedish, and Ukrainian. These datasets facilitate the development of multilingual GEC solutions and help bridge the data gap in adapting English GEC solutions to multilingual GEC. The texts in the datasets originate from three sources: Wikipedia edits for the eleven target languages, subreddits from Reddit in the eleven target languages, and the Ukrainian-only UberText 2.0 social media corpus. While Wikipedia edits were derived from human-made corrections, the Reddit and UberText 2.0 data were automatically corrected with the GPT-4o-mini model. The quality of the corrections in the datasets was evaluated both automatically and manually. Finally, we fine-tune two open-source large language models — Aya-Expanse (8B) and Gemma-3 (12B) — on the multilingual OmniGEC corpora and achieve state-of-the-art (SOTA) results for paragraph-level multilingual GEC. The dataset collection and the best-performing models are available on Hugging Face.

This paper addresses the challenges of sentiment analysis in Ukrainian social media, where users frequently engage in code-switching with Russian and other languages. We introduce COSMUS (COde-Switched MUltilingual Sentiment for Ukrainian Social media), a 12,224-texts corpus collected from Telegram channels, product‐review sites and open datasets, annotated into positive, negative, neutral and mixed sentiment classes as well as language labels (Ukrainian, Russian, code-switched). We benchmark three modeling paradigms: (i) few‐shot prompting of GPT‐4o and DeepSeek V2-chat, (ii) multilingual mBERT, and (iii) the Ukrainian‐centric UkrRoberta. We also analyze calibration and LIME scores of the latter two solutions to verify its performance on various language labels. To mitigate data sparsity we test two augmentation strategies: back‐translation consistently hurts performance, whereas a Large Language Model (LLM) word‐substitution scheme yields up to +2.2% accuracy. Our work delivers the first publicly available dataset and comprehensive benchmark for sentiment classification in Ukrainian code‐switching media. Results demonstrate that language‐specific pre‐training combined with targeted augmentation yields the most accurate and trustworthy predictions in this challenging low‐resource setting.

This paper presents one of the top-performing solutions to the UNLP 2025 Shared Task on Detecting Manipulation in Social Media. The task focuses on detecting and classifying rhetorical and stylistic manipulation techniques used to influence Ukrainian Telegram users. For the classification subtask, we fine-tuned the Gemma 2 language model with LoRA adapters and applied a second-level classifier leveraging meta-features and threshold optimization. For span detection, we employed an XLM-RoBERTa model trained for multi-target, including token binary classification. Our approach achieved 2nd place in classification and 3rd place in span detection.

The Russia-Ukraine war has transformed social media into a critical battleground for information warfare, making the detection of manipulation techniques in online content an urgent security concern. This work presents our system developed for the UNLP 2025 Shared Tasks, which addresses both manipulation technique classification and span identification in Ukrainian Telegram posts. In this paper, we have explored several machine learning approaches (LR, SVC, GB, NB) , deep learning architectures (CNN, LSTM, BiLSTM, GRU hybrid) and state-of-the-art multilingual transformers (mDeBERTa, InfoXLM, mBERT, XLM-RoBERTa). Our experiments showed that fine-tuning transformer models for the specific tasks significantly improved their performance, with XLM-RoBERTa large delivering the best results by securing 3rd place in technique classification task with a Macro F1 score of 0.4551 and 2nd place in span identification task with a span F1 score of 0.6045. These results demonstrate that large pre-trained multilingual models effectively detect subtle manipulation tactics in Slavic languages, advancing the development of tools to combat online manipulation in political contexts.