Front Matter

Jakub Piskorski, Pavel Přibáň, Preslav Nakov, Roman Yangarber, Michał Marcińczuk2025 · DOI

The proceedings of the 10th Workshop on Slavic Natural Language Processing (Slavic NLP 2025) represent a pivotal moment in the computational study of Slavic languages. As the field moves beyond traditional linguistic modeling, the integration of Large Language Models (LLMs) and multilingual transformers has catalyzed a new era of research. This year’s collection highlights a concerted effort to address the "resource gap" for languages such as Belarusian and Macedonian, while simultaneously tackling the societal challenges of propaganda and persuasion in digital discourse.

Addressing the Resource Gap

Slavic languages, spoken by over 400 million people, present a unique set of computational hurdles that often cause them to lag behind English in NLP development. These include rich inflectional morphology (where words change form significantly based on grammar), complex aspectual systems, and relatively free word order.

The workshop aims to bridge this gap, particularly for languages lacking official EU status or those spoken by smaller communities. The goal is to move beyond simple translation or transliteration, developing models that understand the nuanced linguistic and cultural fabric of Eurasia.

Methodological Innovations

The 2025 proceedings reflect a transition toward adapting state-of-the-art (SOTA) global methods to specific Slavic constraints. Key methodological contributions include:

  • LLM Adaptation: Utilizing few-shot prompting and fine-tuning for specialized tasks like humor detection and safety benchmarking.
  • Foundational Modeling: Developing language-specific foundational models (e.g., for Macedonian) to provide more robust alternatives to generic massive multilingual models.
  • Advanced Embedding Techniques: The use of "Hyperbolic Space" for hierarchical classification of propaganda and "Walking Embeddings" for detecting persuasion techniques.
  • Hybrid and Ensemble Systems: Combining rule-based logic with LLMs for Named Entity Recognition (NER) in domains like migration, and using weighted soft voting for complex tasks like Russian lemmatization.

New Benchmarks and Datasets

A significant portion of the workshop was dedicated to the release of critical new datasets and benchmarks designed to measure progress in the field:

  • RedSQL: A Russian Text-to-SQL benchmark for domain-specific applications.
  • DIACU: A dataset specifically designed for the diachronic (historical) analysis of Church Slavonic.
  • PL-Guard: A benchmark for testing the safety and alignment of language models specifically for the Polish language.
  • REPA: An annotation set for Russian error types to evaluate text generation capabilities.
  • Shared Task Dataset: A record-breaking dataset for the detection and classification of persuasion techniques in Slavic parliamentary debates and social media.

Key Insights: Safety, History, and Adversarial NLP

The research indicates a growing "Safety Shift," where researchers are increasingly focused on the ethics of AI within Slavic contexts. This includes benchmarking gender bias in Czech and Slovenian LLM outputs and recognizing that English-centric safety guardrails do not always translate effectively to Slavic cultural norms.

Furthermore, the application of NLP to digital humanities is evident in diachronic discovery, where researchers use word embeddings to track linguistic shifts over time in Croatian news and Church Slavonic. Perhaps most pressing is the "Persuasion Frontier"—the development of "adversarial" NLP tools to combat manipulation, propaganda, and disinformation in the Slavic digital sphere.

Results and Future Directions

The overarching result of the workshop is the successful convergence of LLMs with linguistically motivated problems. The Shared Task on Persuasion Techniques saw record participation, yielding novel methods like "explanation-based data augmentation" to tackle complex, multi-label classification. Additionally, several papers reported significant efficiency gains, achieving high accuracy with reduced parameter counts for tasks like Slovak summarization, making these tools more accessible for real-world deployment.

Despite these gains, the organizers acknowledge that challenges remain. Many Slavic dialects and smaller languages remain underrepresented in commercial NLP solutions. Future work must focus on "border effects" in dialect variation (such as the Rusyn case study) and the continued development of foundational resources for languages currently overlooked by major tech providers. As persuasion techniques become more sophisticated, the robustness of detection models against evolving disinformation tactics will remain a critical priority.

Summary updated

BibTeX
@misc{piskorski2025front,
  author = {Jakub Piskorski and Pavel Přibáň and Preslav Nakov and Roman Yangarber and Michał Marcińczuk},
  title = {{Front Matter}},
  year = {2025},
  doi = {10.18653/v1/2025.bsnlp-1.0},
}
Figure