Vrednotenje metod za gradnjo grafov znanja na področju možganske kapi

The proliferation of biomedical literature on stroke creates a severe information-overload problem that hinders systematic research and discovery. Knowledge graphs (KGs) can address this by structuring unstructured text into machine-readable form, yet building high-quality, domain-specific KGs remains difficult. This thesis presents a systematic evaluation of four KG construction methods spanning three paradigms: rule-based (OpenIE), supervised (REBEL, ReLiK), and LLM-based (Gemma 2 9B), using a novel corpus of more than 433k PubMed abstracts. A key contribution is a new "LLM-as-a-judge'' evaluation framework that scores extracted facts on 10 clinically informed criteria, moving beyond traditional structural metrics to assess factual correctness, clinical relevance, and utility. The results benchmark these methods and provide a roadmap for building a comprehensive StrokeKG to accelerate future research and clinical decision-making.