Pozivamo na Seminar Društva za jezičke resurse i tehnologije koji će se održati u četvrtak, 12. novembra 2026. godine od 18h, na Rudarsko-geološkom fakultetu (Đušina 7), u svečanoj sali.
Izlaganje će održati viši naučni saradnik dr Tatiana Batura (A.P. Ershov Institute of Informatics Systems) na temu: Information Extraction from News and Scientific Texts: Annotation Engineering, Benchmarks, and Practical Lessons.
Abstract:
Annotated data, rather than model architecture, remains the primary bottleneck in information extraction. Drawing on experience building specialized datasets across news, biomedical, and scientific domains in Russian, Kazakh, and English, this talk explores how to build high-quality NLP resources cost-effectively. We discuss the challenges of complex annotation schemes, such as nested named entities, semantic relations, and entity linking. To address this, we present practical strategies to speed up curation, including model-assisted pre-annotation, cost-effective LLM drafting, and human-in-the-loop validation. Beyond standard benchmark results, we briefly cover model performance in few-shot and cross-lingual transfer settings. The talk concludes with three empirical takeaways: nested entities behave fundamentally differently than multi-word terms, span boundary detection remains harder than label classification, and on entity linking a simple lexical baseline can match or even outperform a large language model.
'/%3E%3C/svg%3E){kind=small}
