BM25L

BM25L is a variant of the Okapi BM25 ranking function used in information retrieval. It was introduced to address limitations of BM25 in handling documents of varying lengths, particularly very long documents, by modifying the document-length normalization in the term-frequency component. BM25L retains the core structure of BM25, including the inverse document frequency factor and the general tf-k1-b normalization, but introduces an additional length-based adjustment controlled by a parameter, commonly denoted l. This adjustment changes how term frequency scales with document length, aiming to reduce the bias toward short documents while avoiding excessive penalization of long ones.

In practice, BM25L computes a score for a query by summing contributions from query terms, with the term-frequency contribution being adjusted by the length-aware term to reflect document length more smoothly. Compared with BM25, BM25L often yields improvements on corpora with wide variation in document length and when long documents contain many query terms. The method is simple to implement and can be incorporated into existing BM25-based ranking pipelines with minimal changes, using the same tokenization and indexing infrastructure.

BM25L is part of the family of BM25 variants that includes BM25+, BM25F and others. It is evaluated in information retrieval research and used as a strong baseline for experiments involving document-length effects.