nel: The Entity Linking framework
nel is an fast, accurate and highly modular framework for linking entities in documents.
Out of the box, nel provides:
- named entity recognition
- coreference clustering and candidate generation
- multipple entity disambiguation feature models
- a supervised learning-to-rank framework for entity disambiguation
- a supervised nil detection system with configurable confidence thresholds
- basic nil clustering for out-of-KB entities
- support for evaluating linker performance and running error analysis
nel is modular, it can:
- link entity mentions to any knowledge base you like (not just Wikipedia and Freebase!)
- update, rebuild and redeploy models as a knowledge base changes over time
- retrain recognition and disambiguation classifiers on your own corpus of documents
- adapt linking pipelines to meet performance, precision and recall tradeoffs
nel is flexible, you can run it:
- ad-hoc, from python or as a web service
- offline, in parallel over a corpus of pre-processed documents
- with markdown, html or custom document formats (e.g. CoNLL, TAC)
nel is open-source software released under an MIT license.
You're free to copy, modify and deploy the code in any setting you like - no strings attached.
Checkout the setup guide for details.
pip install git+http://email@example.com/wikilinks/nel.git
To link entities, nel first needs some model of who or what an entity is.
nel uses models from the sift framework for entity linking.
To build models from scratch, you need a corpus of documents that link to entities in the KB.
Wikipedia is a good staring point for notable named entities.
See the model build guide to get started.