FatmaElik/turkish-disaster-news-geonlp

Turkish Disaster News GeoNLP Dataset

Dataset Summary

This dataset was created and submitted as part of the [Uncharted Data Challenge by Adaption](https://www.kaggle.com/competitions/the-uncharted-data-challenge). The LLM-enhanced instruction pairs (turkish_earthquake_news.csv) were generated using [Adaptive Data by Adaption](https://www.adaption.ai) — an AI-powered data adaptation platform.

The first open-source Turkish-language disaster news dataset with district-level geocoding, humanitarian category labels, and multi-dimensional damage classification.

This dataset contains news articles from 9 Turkish national and local sources covering the 2023 Kahramanmaraş Earthquake (6 February 2023, Mw 7.8–7.6) and its multi-phase aftermath — from acute search & rescue through long-term reconstruction. Each record is enriched with verified district-level coordinates, CrisisNLP-aligned humanitarian categories, damage severity labels, urgency classifications, and transparent data quality flags.

Domain: Disaster Management · Turkish NLP · Crisis Informatics · GeoNLP Language: Turkish (tr) Coverage: 2023–2025 (acute response → shelter → recovery → reconstruction) Version: V12

Key Metrics (V12)

Note: The HF Dataset Viewer shows 591 rows — this is the core_benchmark split (train 472 + validation 59 + test 60). The full 838-record master dataset is available in turkish_disaster_news_geonlp_v9.csv.

Why This Dataset?

The Literature Gap

Existing Turkish NLP resources cover general domains:

• —BellaTurca / InstructTurca — large-scale general Turkish corpora, no disaster focus
• —turkish-wikiNER — Wikipedia-based NER, general domain
• —afetharita (2023) — tweet-level address extraction for acute response only

No open dataset exists combining:

• —Long-form Turkish disaster news (haber/article format)
• —District-level (ilçe) geocoding with confidence scoring
• —Humanitarian category labels aligned with CrisisNLP taxonomy
• —Multi-phase coverage: search & rescue → shelter → recovery → reconstruction
• —Transparent data quality flags per record

Disaster Context

The 2023 Kahramanmaraş earthquake sequence caused:

• —50,000+ deaths across 11 provinces
• —1.9 million housing units damaged or destroyed
• —3.3 million people displaced
• —USD 34.2 billion in direct damage (World Bank GRADE)

Turkish-language operational communications in disaster management are underserved by existing NLP tools. This dataset enables model training for real Turkish disaster discourse across all response phases.

Dataset Structure

Schema

dataqualityflags Reference

Damage Level Schema

Humanitarian Categories (CrisisNLP-aligned)

earthquake_damage · search_rescue · casualties · shelter · infrastructure · response · earthquake · recovery

Files

LLM Data Enhancement

The news corpus was enhanced using [Adaptive Data by Adaption](https://www.adaption.ai) as part of the Uncharted Data Challenge.

turkish_earthquake_news.csv contains 994 instruction pairs generated from the news corpus:

Quality improvement: Prompt word count increased by ~230% after Adaption enhancement (original: ~2 words → enhanced: ~6.6 words average).

# Load LLM-enhanced pairs
import pandas as pd
adp = pd.read_csv("turkish_earthquake_news.csv")
clean = adp[adp["prompt_safety_issues"].isna()]  # 983 safe pairs
print(f"Safe instruction pairs: {len(clean)}")

Sources

Geocoding Methodology

District-first (ilçe), province-fallback (il).

A curated reference table of 57 districts with OSM/Nominatim-verified coordinates was built from il_ilce_listesi.csv. Text and title are scanned for district and province names including Turkish suffix variants (da/de/ta/te/nda/ye/ya/nın/nın).

Fallback coordinate detection: Known generic Nominatim fallback coordinates (e.g. 37.132204, 31.092525) were detected and corrected or invalidated. A geocoder attractor bug — where 12 records across Hatay, Adana, Kahramanmaraş, Gaziantep, Osmaniye, and Adıyaman were incorrectly assigned Diyarbakır coordinates (37.91, 40.23) — was identified and corrected. All coordinates validated within Turkey's bounding box (lat: 35.8–42.1, lon: 26.0–44.8).

GeoJSON Output

disaster_heatmap.geojson — 647 geocoded points, compatible with:

• —Kepler.gl — heatmap and density visualization
• —QGIS — spatial analysis and overlay with satellite damage data
• —Leaflet / Mapbox — web mapping

Each feature includes damage_level, urgency, relevance_score, is_core, geocode_confidence, admin_level, and provenance.

Train / Validation / Test Split

Stratified by damage_level on the core_benchmark subset:

random_state=42

Pipeline Summary (V12)

Raw scrape (736)
  → Deduplication + encoding repair (ftfy + Turkish regex)
  → Humanitarian relevance scoring
  → District geocoding (57 ilçe reference)
  → Fallback coordinate detection + correction
  → Geocoder attractor bug fix (12 records corrected)
  → Label refinement (damage + urgency)
  → ZIP archive recovery (+238 rescued records)
  → Quality flag assignment per record
  → is_core / is_extended classification
  → LLM enhancement via Adaptive Data by Adaption (994 pairs)
  → Master V12 (838) → Core (591) → Extended (812)

Provenance tracking: Every record carries a provenance field:

• —original_pipeline — from main scraping pipeline
• —zip_recovered_good — rescued from archive (2+ disaster signals)
• —zip_recovered_weak — rescued with weak signal, flagged

Limitations

Quick Start

from datasets import load_dataset
import pandas as pd

# Geocoded news corpus
ds = load_dataset("FatmaElik/turkish-disaster-news-geonlp")
core = ds["train"].filter(lambda x: x["is_core"])
print(f"Core benchmark: {len(core)} records")

# LLM-enhanced instruction pairs
adp = pd.read_csv("turkish_earthquake_news.csv")
safe = adp[adp["prompt_safety_issues"].isna()]
print(f"Safe instruction pairs: {len(safe)}")

Intended Uses

Research: Turkish disaster NLP classifier training · GeoNLP geocoding benchmark · Multi-label humanitarian classification · Temporal disaster phase analysis · LLM fine-tuning on Turkish crisis text

Operational: Disaster situation awareness from news streams · Damage severity mapping · Integration with remote sensing assessments (Turkey-Earthquake-2023-GeoData)

Citation

@dataset{elik2026turkishdisasternlp,
  author    = {Elik, Fatma},
  title     = {Turkish Disaster News GeoNLP Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/FatmaElik/turkish-disaster-news-geonlp},
  note      = {2023 Kahramanmaraş Earthquake — V12, 838 records. LLM-enhanced with Adaptive Data by Adaption (adaption.ai).}
}

Author

Fatma Elik MSc İTÜ Satellite Communications & Remote Sensing İstanbul Metropolitan Municipality Research: Computer Vision · Earth Observation · Disaster Management · GeoNLP

Related Resources

• —Turkey-Earthquake-2023-GeoData — Remote sensing & geospatial
• —CrisisNLP — Crisis NLP benchmark datasets
• —afetharita — Tweet-level address NER (2023)
• —turkish-wikiNER — General Turkish NER
• —Adaptive Data by Adaption — LLM data enhancement platform

License

CC BY-NC 4.0 — Non-commercial research use only. Labels, geocoding, pipeline, and metadata are original contributions of the author. Article excerpts (500 chars) included for research purposes. Full article copyright belongs to respective publishers.