Researchers Release Bilingual Tatarstan Geography Dataset with AI Question-Answering System

Researchers have developed a comprehensive approach to answering geography questions over multilingual place names, introducing both a novel dataset and a retrieval-based system designed to handle spatial data in Russian and Tatar.

The Dataset#

The team compiled a bilingual database containing nearly 10,000 location names drawn from the Tatarstan region. Each entry includes linguistic details, etymological background, and geographic coordinates—with over 93 percent of the toponyms precisely geolocated. From this foundation, they generated around 39,000 paired question-answer examples where the correct answer location can be verified automatically.

How the System Works#

The architecture pairs a hybrid retrieval component with a reading model. The retriever uses two complementary strategies: dense semantic search powered by multilingual-e5-large embeddings, combined with geometric filtering via KD-trees and haversine distance calculations to prioritize results by spatial proximity. The extraction phase relies on transformer-based readers fine-tuned for the task.

When evaluated on 500 test queries, the hybrid approach substantially outpaced simpler methods:

• Recall@1: 0.988 (finds the right answer in the top result 98.8% of the time)
• Recall@5: 1.000 (perfect recall within the top five results)
• Mean Reciprocal Rank: 0.994

Among the reader models tested—RuBERT, XLM-RoBERTa-large, and T5-RUS—XLM-RoBERTa-large delivered the strongest overall performance, achieving near-perfect extraction metrics (EM = 0.992, F1 = 0.994). An interesting finding emerged with RuBERT: it struggled on coordinate-based questions due to tokenization quirks, though a straightforward post-processing fix restored full accuracy.

Open Release and Applications#

All resources—including the dataset, question corpus, trained models, and a web demo—have been released publicly on Hugging Face. The authors note immediate utility for geocoding services, geospatial question-answering platforms, and digital humanities research.