cnupo23/korean-medicine-embedding-dataset

Korean Medicine Embedding Dataset (한의학 임베딩 데이터셋)

A contrastive-learning dataset for training and evaluating domain-adapted text embedding models for Korean Medicine (KM, 한의학). Each example is a query–positive–negatives triplet built from Korean Medicine terminology dictionaries and a curated KM ontology, capturing clinically meaningful semantic relations such as disease–symptom, disease–prescription, prescription–indication, prescription–herb, and herb–indication.

This dataset accompanies the study "A Domain-Adapted Embedding Model for Semantic Retrieval of Korean Medicine Knowledge", where it was used to fine-tune BGE-M3 with LoRA and ontology-guided hard-negative sampling. To our knowledge, it is the first publicly released embedding dataset specifically targeting Korean traditional medicine.

Why this dataset

General-purpose and even biomedical embedding models do not capture the conceptual and linguistic structure of Korean Medicine. KM interprets disease states in functional/relational terms and relies on symbolic concepts (e.g., 음양 yin–yang, 한열 cold–heat). It also shares historical roots with Traditional Chinese Medicine (TCM) but developed an independent theoretical system (e.g., Donguibogam, Sasang constitutional medicine), and KM texts mix Hangeul (한글), Hanja (한자, traditional characters), and Hangeul–Hanja parallel notation, where homophonous Sino-Korean terms can differ in meaning depending on the underlying characters. This dataset was constructed to inject that domain knowledge and these orthographic characteristics into embedding training.

Dataset structure

Data fields

Each line is a JSON object with the following fields:

Field	Type	Description
query	string	A KM term or concept (term, disease, pattern, prescription, herb).
pos	list[string]	Semantically related positive text(s).
neg	list[string]	10 negatives per query: 3 similarity-based hard negatives + 7 random negatives.

Example (Hangeul–Hanja format):

{
  "query": "사군산(使君散)",
  "pos": ["사군자(使君子), 빈랑자(檳榔子), 웅황(雄黃)"],
  "neg": ["목통(木通), 청피(靑皮), 내복자(萊菔子), 회향(茴香), ...", "..."]
}

Text representation formats

To help models generalize across the notations used in KM texts, every instance is provided in three text formats, applied identically to the query, pos, and neg fields:

• Hangeul (한글 전용) — Korean phonetic spelling only
• Hanja (한자 전용) — traditional Chinese characters only
• Hangeul–Hanja (병기) — Korean with Hanja in parentheses, e.g. 풍한해수(風寒咳嗽)

The same query rendered in each format:

Format	Example
Hangeul	풍한해수
Hanja	風寒咳嗽
Hangeul–Hanja	풍한해수(風寒咳嗽)

Splits

The base dataset contains 113,299 query–positive pairs. Each of the three formats yields 113,299 instances, split 9:1 into train/test. The three formats are merged for training and kept separate for evaluation.

File	Split	Format	Examples
split/unified_train.jsonl	train	all three formats combined	305,907
split/test_both.jsonl	test	Hangeul–Hanja	11,330
split/test_chinese.jsonl	test	Hanja	11,330
split/test_korean.jsonl	test	Hangeul	11,330

Per-format, single-notation versions of the full data (before splitting) are also provided under transformed/:

File	Format	Examples
transformed/train_data_both.jsonl	Hangeul–Hanja	113,299
transformed/train_data_chinese.jsonl	Hanja	113,299
transformed/train_data_korean.jsonl	Hangeul	113,299

Naming note: both = Hangeul–Hanja, korean = Hangeul, chinese = Hanja.

Dataset composition

Positive pairs were generated from two source types — KM terminology and the KM ontology — spanning 16 relation types:

Source	Query–Positive relation	Pairs
KM Terminology	Term–Definition	29,278 (25.8%)
KM Terminology	Term–Synonym	33,243 (29.3%)
KM Ontology	Disease–Disease	3,266 (2.9%)
KM Ontology	Disease–Etiology	617 (0.5%)
KM Ontology	Disease–Symptom	6,374 (5.6%)
KM Ontology	Disease–Treatment Principle	1,137 (1.0%)
KM Ontology	Disease–Prescription	2,548 (2.2%)
KM Ontology	Disease–Pattern Concept	665 (0.6%)
KM Ontology	Disease–Prognosis	126 (0.1%)
KM Ontology	Prescription–Indication	15,770 (13.9%)
KM Ontology	Prescription–Herb	16,271 (14.4%)
KM Ontology	Prescription–Efficacy	940 (0.8%)
KM Ontology	Herb–Alias	1,358 (1.2%)
KM Ontology	Herb–Property	565 (0.5%)
KM Ontology	Herb–Efficacy	569 (0.5%)
KM Ontology	Herb–Indication	572 (0.5%)
Total		113,299 (100%)

Construction pipeline

1. Source curation. Positive pairs were derived from KM terminology resources (standard KM terminology, lay-friendly term dictionary, synonyms) and a pre-corrected/verified KM ontology covering diseases, patterns (변증), prescriptions, and herbs.
2. Positive pair construction. Term–definition and term–synonym links were taken from terminology data; ontology node-to-node relations were extracted and converted into query–positive pairs.
3. Hard-negative mining (ontology-guided). All pairs were placed on an ontology graph with bidirectional edges. For each query:
   • Nodes within 2 hops of the query were treated as semantically related and excluded from negative candidates (to avoid false negatives).
   • 10 negatives were assembled: 3 similarity-based hard negatives (cosine similarity in the 0.4–0.6 range under BGE-M3 — hard but not near-duplicates) and 7 random negatives (excluding 2-hop neighbors and already-selected negatives).
4. Multi-format generation & splitting. Each instance was rendered in Hangeul, Hanja, and Hangeul–Hanja; each format was split 9:1; the train portions were unified.

How the dataset was used (reference results)

The dataset was used to fine-tune BGE-M3 (568M params) with LoRA (rank=16, α=32, dropout=0.1) and MultipleNegativesRankingLoss (best configuration: batch size 256, learning rate 1e-3, 20 epochs). Evaluated with a candidate pool of 1 positive + 10 negatives per query, the fine-tuned model (568M params) outperformed the base model and all compared large-scale general-purpose embedding models spanning 8B–27B parameters (e.g., Octen-Embedding-8B, F2LLM-v2-14B, Harrier-OSS-v1-27B) on all retrieval metrics. The best configuration used the Hangeul–Hanja format:

Metric	BGE-M3 (base)	BGE-M3 + LoRA
Recall@1	0.5583	0.8422
MRR@10	0.6607	0.8971
NDCG@10	0.7315	0.9218

LoRA fine-tuning improved retrieval consistently across all three formats, with the Hangeul–Hanja parallel notation giving the strongest results. The strongest baseline (Octen-Embedding-8B) reached only MRR@10 0.8052, well below the proposed 568M model's 0.8971. A domain-specific semantic-separability metric, SimGap (mean cosine similarity of positive pairs minus that of negative pairs), also rose sharply after fine-tuning — from 0.1245 to 0.5117 on the Hangeul–Hanja format — indicating a more discriminative embedding space for KM concepts.

Intended uses

• Training / fine-tuning Korean Medicine–specific embedding models for semantic retrieval.
• Building RAG systems, clinical decision support, and knowledge-retrieval pipelines over KM literature, guidelines, and ontology-based knowledge bases.
• Studying the effect of Hangeul / Hanja / Hangeul–Hanja representations on retrieval in Sino-Korean technical domains.

Limitations

• Train and test were derived from the same terminology/ontology resources via random 9:1 splitting, so the same entity may appear in both splits; cold-start / fully unseen-entity generalization is not guaranteed.
• The data reflects terminology and ontology relations, not real clinical notes or prescription records, which use non-standardized, abbreviated, and varied expressions.
• Negatives were mined under BGE-M3 similarity; the difficulty distribution is tied to that base model.

Citation

If you use this dataset, please cite the accompanying paper:

@article{kim2026km_embedding,
  title   = {A Domain-Adapted Embedding Model for Semantic Retrieval of Korean Medicine Knowledge},
  author  = {Kim, Young-Eun and Keum, Kimun and Kim, Sang-Kyun and Kim, Anna and Son, Mi Ju and Seo, SuMin and Jang, Yunji and Oh, Yongtaek and Yi, Mun Yong},
  year    = {2026},
  note    = {Dataset: https://huggingface.co/datasets/cnupo23/korean-medicine-embedding-dataset}
}

License

Released under the Apache License 2.0. The underlying terminology and ontology resources are subject to their original terms; please verify source-specific conditions before use.