Extracting geographic locations from text

Develop and benchmark a modular NLP pipeline that uses smart gazetteer filtering and context extraction to reliably identify, resolve, and interpret study locations mentioned in scientific literature.

Posted: September 2026
Contact: Tim Repke

The Challenge

Extracting study locations from scientific articles requires knowing both where a place is and what it is. Standard geoparsing is a two-step process: identifying candidate text snippets and linking them to a knowledge base like GeoNames or Wikidata to extract structured geographic metadata (latitude/longitude, entity type like river or city, geopolitical structure, and history).

However, standard gazetteer queries suffer from candidate explosion (e.g., querying "Berlin" returns hundreds of hits), existing models are trained on news corpora rather than scientific literature, and standard tools fail to analyze the surrounding text to verify if statements are actually about the study site.

Project & Pipeline Goals

You will build and evaluate a modular pipeline using local LLMs, spaCy, or specialized NER models to identify, filter, match, and contextualize location mentions in academic text.

Your Tasks & Research Questions

  • Database Pre-Filtering: How can we best (pre)filter the GeoNames database to resolve the candidate explosion problem without losing valid locations?

  • High-Recall Mention Identification: Develop a high-recall method to recognize geographic entity candidates in text.

  • High-Precision Matching: Develop a high-precision disambiguation method to link text mentions to filtered GeoNames entries.

  • Structured Context Extraction: Extract surrounding sentence context and establish structured relations to confirm that statements are actually about the specific location.

  • Geoparser Horizon Scan: Perform a comprehensive overview and comparison of modern geoparsing tools evaluated to be better than Mordecai 3.

  • Evaluation Pipeline Replication: Build an evaluation pipeline using common benchmark data (specifically replicating the setup in arXiv:2303.13675).

  • Scientific Dataset Assembly: Identify and assemble a newer evaluation dataset tailored specifically to scientific literature (moving beyond traditional news corpora).

  • Comparative Evaluation: Benchmark your new approach against existing state-of-the-art systems.