Skip to main content

Alligator

Alligator is a Python library for entity linking over tabular data. Given a CSV table, it automatically:

  1. Identifies Named Entity (NE) columns (or accepts manual assignments)
  2. Fetches candidate Wikidata entities for each cell via the LAMAPI retrieval service
  3. Computes ~27 similarity and overlap features per candidate
  4. Runs a two-stage Keras ML pipeline (rank → rerank) to score and rank candidates
  5. Produces three types of SemTab-compatible annotations:
    • CEA — Cell Entity Annotation: links each NE cell to a Wikidata QID
    • CTA — Column Type Annotation: infers the Wikidata type for each NE column
    • CPA — Column Property Annotation: infers relationships between NE columns

Results are stored in MongoDB and can optionally be written back to CSV.

Key Features​

  • Automatic column type classification (NE / LIT / IGNORED) via column-classifier
  • SHA-256 keyed MongoDB TTL cache for API responses (avoids redundant lookups)
  • Fuzzy-retry candidate retrieval when exact match returns no results
  • 5-attempt exponential backoff on retrieval failures
  • Two-stage ML ranking: rank (local features) → rerank (global CTA/CPA frequency features)
  • Parallel multiprocessing workers for scalable throughput
  • FastAPI REST backend for integration into larger systems
  • Docker Compose setup for easy deployment

Architecture Overview​

CSV / DataFrame
      │
      ▼
DataManager       — column classification, MongoDB onboarding
      │
      ▼
WorkerManager     — N async workers: entity extraction → candidate fetch → feature computation
      │
      ▼
MLManager         — rank → compute global frequencies → rerank
      │
      ▼
OutputManager     — CSV output + MongoDB annotations