lion-ai/MedKG-graph
收藏资源简介:
--- license: cc-by-4.0 language: - pl tags: - medical - knowledge-graph - polish - pharmacology task_categories: - question-answering - text-retrieval pretty_name: MedKG — Polish Medical Knowledge Graph size_categories: - 10K<n<100K --- # MedKG — Polish Medical Knowledge Graph A directed medical knowledge graph in **Polish**, extracted from Polish Wikipedia and drug leaflets (ChPL / charakterystyki produktów leczniczych), then **LLM-validated edge-by-edge**. - **36,544 directed edges**, **25,942 nodes**, **24 typed relation types** - Built for medical **question answering** and retrieval. ## How it was built 1. **Extraction** — Gemma-4-26b extracts triples from chunked Polish medical text with a schema-guided, qualifier-preserving prompt. 2. **Canonicalization + dedup** — BGE-M3 + Morfeusz lemmatization; `support` counts how many times an edge was independently extracted (corroboration). 3. **Validation & repair** — every corroborated edge (support≥2) is checked by **DeepSeek-V4** against a typed schema with domain/range + direction rules, and is either kept, **direction- fixed**, **predicate-retyped**, or dropped. Only kept/repaired edges are published. ## Quality (measured, not claimed) Evaluated with a **per-edge accuracy rubric** graded by an independent strong model (a physician- persona judge classifies each sampled edge CORRECT / QUESTIONABLE / WRONG). The rubric was validated against a negative control (a version with randomly permuted objects scored **2/10**). | Slice | Edge accuracy (physician-CORRECT) | Only ~WRONG | |---|---|---| | Full graph | **78%** | ~4% | | `support ≥ 3` | **82%** | — | | `support ≥ 5` | **88%** | — | Holistic OVERALL: **7/10** (7.3 on the highest-precision slice). Remaining imperfections are mostly *true-but-general* statements (e.g. `bakterie → powoduje → zapalenie płuc`) inherent to leaflet text, not factual errors. **Recommended use:** filter by `support` to trade recall for precision (`support>=3` is a good default; `support>=5` for the highest-precision core). ## Fields - `subject`, `predicate`, `object` — the directed relation (Polish). - `support` — corroboration count (confidence signal). - `rel_class` — `clinical` | `structural` | `taxonomic`. ## Relation schema (24 types) | predicate | edges | |---|---| | `ma działanie niepożądane` | 7920 | | `jest rodzajem` | 4413 | | `powoduje` | 3332 | | `objawia się` | 3053 | | `jest stosowany w leczeniu` | 2822 | | `jest czynnikiem ryzyka` | 1795 | | `wchodzi w interakcję z` | 1748 | | `jest przeciwwskazany przy` | 1726 | | `zawiera` | 1591 | | `zmniejsza` | 1063 | | `zwiększa stężenie` | 1004 | | `zwiększa` | 830 | | `występuje u` | 700 | | `ma powikłanie` | 673 | | `zmniejsza stężenie` | 638 | | `służy do rozpoznania` | 557 | | `jest częścią` | 542 | | `zapobiega` | 530 | | `działa na receptor` | 449 | | `zmniejsza ryzyko` | 388 | | `hamuje` | 323 | | `jest metabolizowany do` | 238 | | `jest metabolizowany przez` | 109 | | `jest preparatem zawierającym` | 100 | ## Related - QA dataset built on this graph: [`lion-ai/MedKG-QA`](https://huggingface.co/datasets/lion-ai/MedKG-QA) ## Limitations LLM-extracted from consumer/professional leaflets + Wikipedia; entities are **not** linked to a controlled vocabulary (UMLS/ATC/ICD), so near-duplicate concepts and general phrasings occur. Not a substitute for professional medical judgement.
MedKG is a directed medical knowledge graph in Polish, extracted from Polish Wikipedia and drug leaflets (ChPL), then LLM-validated edge-by-edge. It contains 36,544 directed edges, 25,942 nodes, and 24 typed relation types, built for medical question answering and retrieval.




