Skip to content

Okapi at TREC-3: Probabilistic Retrieval Model

Authors: Stephen Robertson, Steve Walker, Susan Jones, et al. Year: 1994 Citations: 15,000+

Summary

BM25 probabilistic ranking function. Term frequency saturation and document length normalization. State-of-the-art retrieval baseline for decades.

Key Concepts

  • Probabilistic Ranking: Probability of relevance
  • TF Saturation: Diminishing returns on frequency
  • Document Length Normalization: Prevents length bias
  • BM25 Formula: k1, b parameters control behavior
  • Empirical Effectiveness: Consistently outperforms TF-IDF

Impact

  • 15,000+ citations
  • Standard baseline for retrieval evaluation
  • Used in Elasticsearch, Lucene, major search engines
  • Still competitive with neural methods
  • Foundation for understanding retrieval effectiveness
  • TF-IDF (Sparck Jones, 1972)
  • Dense Passage Retrieval (Karpukhin et al., 2020)