Affiliations: [a] Instituto Caro y Cuervo, Calle 10 #4-69, Bogotá, D.C., Colombia
| [b]
Microsoft Research, One Microsoft Way, Redmond, WA, US
| [c] Universidad Nacional de Colombia, Ciudad Universitaria, Bogotá, D.C., Colombia
| [d] CIC, Instituto Politécnico Nacional, Av. Juan de Dios Bátiz, 07738, Mexico City, Mexico
Correspondence:
[*]
Corresponding author. Sergio Jimenez, Instituto Caro y Cuervo, Calle 10 #4-69, Bogotá, D.C., Colombia. E-mail: [email protected].
Abstract: In this paper, the use of collection term frequencies (i.e. the total number of occurrences of a term in a document collection) in the BM25 retrieval model is investigated by modifying its term frequency (TF) and inverse document frequency (IDF) components. Using selected examples extracted from TREC collections, it was observed that the informative nature, for retrieval purposes, of terms, either with the same TF (in a document) or IDF (in a collection) may be better revealed with the use of collection term frequencies (CTF). From three new heuristics based on those observations and deviations from a random Poisson model, collection term frequencies were integrated to TF and IDF factors. The novel formulations were tested by employing the TREC-1 to TREC-8 collections in the ad hoc task, for which BM25 was first developed and tested. Consistent and significant improvements were observed in mean average precision (MAP) reaching up to 17.67% for the TREC-8 dataset, and 7.16% averaged over all tested collections. These results were considerably better in comparison to other approaches surveyed aiming to improve BM25, proving in this way the effectiveness of the proposed heuristics and formulae. The proposed approach requires only additional offline pre-computations and does not entail extra computational complexity for retrieval while keeping the original spirit and parameter robustness of BM25.
Keywords: BM25, tf-idf, collection term frequency, information retrieval
heuristics, TREC collections, deviation from randomness
