uznlp-uz/uz_medner

Uzbek Medical NER Dataset (UzMedNER)

📌 Description

This dataset introduces UzMedNER, a structured Named Entity Recognition (NER) resource for the Uzbek language in the medical domain. It is designed to support token-level sequence labeling tasks and facilitate research in low-resource biomedical NLP.

The dataset consists of manually annotated Uzbek text where each token is labeled using a predefined tagset representing medical and related entity types.

UzMedNER addresses the lack of:

• domain-specific annotated corpora in Uzbek
• standardized NER benchmarks for medical text
• resources for training sequence labeling models in low-resource settings

---

🧠 Task Definition

This dataset is designed for:

Named Entity Recognition (NER)

• Input: tokenized Uzbek sentence
• Output: sequence of entity labels (BIO tagging scheme)

Example:

Bemor B-DISEASE diabet I-DISEASE bilan O kasallangan O .

---

📊 Dataset Structure

The dataset is stored in TSV format with token-level annotations.

Typical format:

token	label
Bemor	O
diabet	B-DISEASE
bilan	O
kasallangan	O

• Each row = one token
• Labels follow BIO tagging scheme
• Sentences are separated by empty lines

---

🏷 Tagset (Entity Types)

The dataset uses a BIO-based tagging scheme with the following entity categories:

Tag	Description
B-DISEASE / I-DISEASE	Disease names
B-SYMPTOM / I-SYMPTOM	Symptoms
B-DRUG / I-DRUG	Medications
B-TREATMENT / I-TREATMENT	Medical treatments
B-ANATOMY / I-ANATOMY	Body parts
B-TEST / I-TEST	Medical tests
O	Outside (non-entity token)

Note: Exact tag inventory is defined in the accompanying tagset.tsv file.

---

🧾 Example

Token        Label
Bemor        O
yurak        B-ANATOMY
og‘rig‘i     B-SYMPTOM
bilan        O
shifoxonaga  O
murojaat     O
qildi        O

---

📏 Evaluation Protocol

Recommended evaluation metrics:

• Precision
• Recall
• F1-score (entity-level)
• Token-level accuracy

Evaluation should follow standard CoNLL NER evaluation.

---

📊 Data Splits

Note: predefined splits may be added in future versions.

Recommended split:

• Train: 80%
• Validation: 10%
• Test: 10%

---

🎯 Use Cases

This dataset can be used for:

• 🏥 Medical NER in Uzbek
• 🤖 Fine-tuning transformer models (BERT, RoBERTa, Qwen, etc.)
• 📊 Sequence labeling research
• 🔍 Clinical text mining
• 🧠 Biomedical NLP for low-resource languages

---

⚙️ Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ruhilloalaev/UzMedNER", "default")

---

⚠️ Notes

• Data is in Uzbek (Latin script)

• Annotation follows BIO scheme

• Domain: medical / clinical language

• Some entities may exhibit:

  • morphological variation
  • spelling inconsistencies
  • domain-specific abbreviations

---

📜 License

This dataset is released under the CC-BY-4.0 License.