... dif-ferences. For the TREC dataset, BM25 performedbetter than the other models, and for the NTCIRdataset, VS performed better.Our features of the topic association modelshow mild improvement over the ... mea-sures. Among the topic association models, PMIperforms the best in MAP and R-prec, while WPachieved the biggest improvement in P@10.Since BM25 performs the best among the wordgeneration models, ... is the frequency of w in d, | d | is the number of unique terms in d, avgdl is the average| d | of all documents, N is the number of doc- uments in the collection, df(w) is the number ofdocuments...