Microsoft Cambridge at TREC 14: Enterprise Track.
Nick Craswell, Hugo Zaragoza, Stephen E. Robertson · 2005
mbination and 3) The tuning and ranking framework we used this year. To calculate BM25F [4] we first calculate a normalised term frequency for each field: x d,f,t := x d,f,t (1 +B f ( l d,f l f - 1)) # {SUBJECT, BODY, QUOTED} indicates the field type, x d,f,t is the term frequency of term t in the field type f of document d, l d,f is the length of that field, and l f is the average field length for that field type. B f is a field-dependant parameter similar to the B parameter in BM25. In particular, if B f = 0 there is no normalisation and if B f = 1 the frequency is completely normalised w.r.t. the average field length. These term frequencies can then be combined in a linearly weighted sum to obtain the final term pseudo-frequency: x d,t = W f x d,f,t (2) with weight parameters W f . This is then used in the usual BM25 saturating function. This leads the following ranking function, which we refer to as BM25F: BM25F (d) := t#q#d x d,t K 1 + x d,t t (3)