BLINK is a Python library for entity linking, mapping mentions in text to entities in Wikipedia. It targets researchers and developers needing state-of-the-art Wikification capabilities, offering a two-stage BERT-based approach for high accuracy.

How It Works

BLINK employs a two-stage architecture. First, a bi-encoder independently embeds mention context and entity descriptions for efficient retrieval in a dense space. Second, a cross-encoder refines candidate selection by processing concatenated mention and entity text. This approach achieves state-of-the-art results by balancing retrieval speed with deep contextual understanding. FAISS integration enables further optimization for large-scale retrieval.

Quick Start & Requirements

• Install via pip install -e git+git@github.com:facebookresearch/BLINK#egg=BLINK.
• Requires Python 3.7+.
• Download pre-trained models using ./download_blink_models.sh.
• Optional: FAISS index can be built using python blink/build_faiss_index.py.
• Interactive demo: python blink/main_dense.py.
• Official Docs: https://github.com/facebookresearch/BLINK

Highlighted Details

• State-of-the-art performance on multiple entity linking benchmarks (AIDA-YAGO2, ACE 2004, etc.).
• Supports fast mode using only the bi-encoder for speed gains.
• Integrates FAISS for efficient exact/approximate retrieval.
• Includes ELQ for end-to-end entity linking on questions.

Maintenance & Community

• Developed by Facebook AI Research.
• The project appears actively maintained, with recent updates noted in the README.

Licensing & Compatibility

• MIT License.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The project's knowledge base is based on a 2019 Wikipedia dump, which may not reflect the most current information. The older Solr-based IR system is deprecated.