Optimizing context-based location extraction by tuning open-source LLMs with RAG
Text data such as news from media include different types of geographic information, represented by location, that indicates the whereabout of events or phenomena. Extracting the geographic locations from text within their contexts is challenging, even with Natural Language Processing (NLP) tools an...
Saved in:
| Main Authors: | , , , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-08-01
|
| Series: | International Journal of Digital Earth |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/17538947.2025.2521786 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Text data such as news from media include different types of geographic information, represented by location, that indicates the whereabout of events or phenomena. Extracting the geographic locations from text within their contexts is challenging, even with Natural Language Processing (NLP) tools and the latest Large Language Models (LLMs). We propose to optimize LLMs using Retrieval-Augmented Generation (RAG) and prompt-tuning methods, such as zero-shot and instruction-based prompting to improve the precision of extracting location information from news. Using Sudan conflict as an example, we extracted the corresponding locations and dates for conflict incidents. We compared runtime and accuracy of using various open-source LLMs, different hyperparameter settings, with and without RAG. Traditional Named Entity Recognition (NER), zero-shot prompting, instruction-based prompting, few-shot prompting, chain-of-thought (CoT) prompting, and RAG-based tuning were compared using an evaluation matrix. RAG-based tuning delivered the highest F1 score (>0.9) for extracting and associating location data with conflict incidents. This research highlights the benefits of using RAG for multi-incident context-based location extraction and provides insights into optimizing LLMs through prompt-tuning, hyperparameter adjustment, and model selection for location extraction tasks. The results can also be used to extract context-based locations or relevant information from text-based documents of other applications. |
|---|---|
| ISSN: | 1753-8947 1753-8955 |