Geocoding & NER Service

This repo provides a unified service that processes text to extract and resolve geolocations, as well as general named entities. It combines natural language processing (NLP) for entity extraction with geocoding to map locations to geographic coordinates.

Location extraction is based on the RobBERT NER model and currently only suitable for Ghent. General NER supports multiple languages (Dutch, German, English) and extraction methods (regex, spaCy, Flair, composite). The service uses spaCy for NLP processing.

Features

• Extracts location entities from text using an NLP model
• Extracts general entities (PERSON, ORG, DATE, etc.) with multi-language support
• Resolves extracted locations to latitude/longitude using Nominatim
• Load data from local triple store
• Stores all annotations in SPARQL triplestore with full provenance

Requirements

• Modify the docker-compose.yaml file and fill in the MU_SPARQL_ENDPOINT of the local triple store (hosted in the app-decide Docker container). This may require modifying the network name as well to match the docker network app-decide is running in.

Running Nominatim with Docker

Running Nominatim has been included in the included docker compose file. Note that on the first start, Nominatim might be busy for an hour to build its database. As a volume is mounted for the database, this setup happens once.

Usage

Run

docker compose up

and send a request to the /notify-change endpoint on the port as configured in the docker-compose.yaml file.

Original demo code

The original demo code can be found in the demo folder.

Segmentation

The service supports two different segmentation models, configurable via config.json.

1. LLMSegmentor (Generic LLM)

Default usage for general LLMs (OpenAI, Azure OpenAI, Mistral, Ollama, etc.). It instructs the model to return JSON structure.

Configuration in config.json:

"segmentation": {
  "model_name": "gpt-4.1",       // Your model deployment name
  "api_key": "YOUR_KEY",
  "endpoint": "YOUR_ENDPOINT",
  "temperature": 0.1,
  "max_new_tokens": 14000
}

2. GemmaSegmentor (Specialized Model)

Uses the specialized wdmuer/decide-marked-segmentation model which is trained to output XML tags directly. Use this if you want to reproduce the original/legacy behavior.

To enable this mode, you must set the model_name specifically:

"segmentation": {
  "model_name": "wdmuer/decide-marked-segmentation",
  "max_new_tokens": 4000
  // ... other fields as needed
}