uznlp-uz/uz_edbench

---
license: cc-by-4.0
language:
- uz
pretty_name: UZ-EDBench v1.0
size_categories:
- 1K<n<10K
task_categories:
- text-classification
configs:
- config_name: default
default: true
data_files:
- split: train
path: UZ-EDBench.tsv
sep: "\t"
- config_name: doctors
data_files:
- split: train
path: UZ-EDBench.Doctors.tsv
sep: "\t"
---

# Uzbek Medical Entity Benchmark (UZ-EDBench)

## 📌 Description

**UZ-EDBench** is a structured Uzbek-language dataset designed for **medical entity recognition and classification** in a low-resource setting. The dataset is distributed in **TSV format** and consists of annotated tokens and domain-specific entity labels.

It includes:

* a main annotated corpus (`UZ-EDBench.tsv`)
* a structured list of medical specialists (`UZ-EDBench.Doctors.tsv`)

This dataset addresses the lack of:

* Uzbek medical NLP benchmarks
* annotated corpora for clinical entity extraction
* structured taxonomies of medical specialists

---

## 🧠 Task Definition

The dataset supports:

### 1. Named Entity Recognition (NER)

* **Input:** tokenized Uzbek text
* **Output:** entity labels (BIO format)

### 2. Entity Classification

* **Input:** token / span
* **Output:** entity type (medical category)

---

## 📊 Dataset Structure

### 🔹 Main File: `UZ-EDBench.tsv`

* Format: **TSV (tab-separated)**
* Each row = one token
* Sentence boundaries may be separated by empty lines

Typical format:

```tsv id="a1xk9d"
token label
Bemor O
kardiolog B-DOCTOR_TYPE
qabuliga O
keldi O
```

---

### 🔹 Auxiliary File: `UZ-EDBench.Doctors.tsv`

This file contains structured information about **medical specialists (doctor types)** used in annotation.

Typical structure:

```tsv id="b8j2md"
doctor_type description
kardiolog Yurak kasalliklari bo‘yicha mutaxassis
nevrolog Asab tizimi mutaxassisi
```

---

## 🏷 Tagset (Entity Labels)

The dataset uses a **domain-specific BIO tagging scheme**.

### 🔹 Core Medical Entities

| Tag | Description |
| ------------------------- | --------------------- |
| B-DISEASE / I-DISEASE | Kasallik nomlari |
| B-SYMPTOM / I-SYMPTOM | Belgilar (simptomlar) |
| B-DRUG / I-DRUG | Dori vositalari |
| B-TREATMENT / I-TREATMENT | Davolash usullari |
| B-TEST / I-TEST | Tibbiy tekshiruvlar |
| B-ANATOMY / I-ANATOMY | Tana qismlari |

---

### 🔹 Doctor Types (Shifokor turlari)

| Tag | Description |
| ----------------------------- | -------------------- |
| B-DOCTOR_TYPE / I-DOCTOR_TYPE | Tibbiy mutaxassislik |

Examples include:

* kardiolog
* terapevt
* nevrolog
* pediatr
* jarroh
* dermatolog

The full list is provided in:
👉 `UZ-EDBench.Doctors.tsv`

---

### 🔹 BIO Tagging Scheme

| Tag | Meaning |
| ----- | ------------------- |
| B-XXX | Beginning of entity |
| I-XXX | Inside entity |
| O | Outside entity |

Example:

```text id="v3n9sj"
kardiolog B-DOCTOR_TYPE
shifokor I-DOCTOR_TYPE
```

---

## 🧾 Example

```text id="g4u7sl"
Token Label
Bemor O
nevrolog B-DOCTOR_TYPE
qabuliga O
bosh B-ANATOMY
og‘rig‘i B-SYMPTOM
bilan O
keldi O
```

---

## 📏 Evaluation Protocol

Recommended metrics:

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

Evaluation should follow the **CoNLL NER standard**.

---

## 📊 Data Splits

*Predefined splits are not included.*

Recommended split:

* Train: 80%
* Validation: 10%
* Test: 10%

---

## 🎯 Use Cases

* 🏥 Uzbek medical NER systems
* 🤖 Fine-tuning transformer models (BERT, RoBERTa, Qwen, etc.)
* 📊 Clinical text mining
* 🧠 Healthcare AI assistants
* 🔍 Information extraction from Uzbek medical text

---

# UZ-EDBench

This repository contains two tab-separated subsets:

- `default`: the main triage benchmark in `UZ-EDBench.tsv`
- `doctors`: the doctor label reference table in `UZ-EDBench.Doctors.tsv`

## ⚙️ Loading the Dataset

```python id="y1p3z7"
from datasets import load_dataset

main = load_dataset("ruhilloalaev/uz_edbench", name="default")
doctors = load_dataset("ruhilloalaev/uz_edbench", name="doctors")
```

---

## ⚠️ Notes

* Data is in **Uzbek (Latin script)**
* Format: **TSV (tab-separated)**
* Domain: **medical / healthcare**

Text may include:

* morphological variation
* domain-specific terminology
* spelling inconsistencies

---

## 📜 License

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