A new context‐dependent term weight computed by boost and discount using relevance information

Abstract

We studied the effectiveness of a new class of contextdependent term weights for information retrieval. Unlike the traditional term frequency-inverse document frequency (TF-IDF ), the new weighting of a term t in a document d depends not only on the occurrence statistics of t alone but also on the terms found within a text window (or "document-context") centered on t. We introduce a Boost and Discount (B&D) procedure which utilizes partial relevance information to compute the contextdependent term weights of query terms according to a logistic regression model. We investigate the effectiveness of the new term weights compared with the contextindependent BM25 weights in the setting of relevance feedback. We performed experiments with title queries of the TREC-6, -7, -8, and 2005 collections, comparing the residual Mean Average Precision (MAP) measures obtained using B&D term weights and those obtained by a baseline using BM25 weights. Given either 10 or 20 relevance judgments of the top retrieved documents, using the new term weights yields improvement over the baseline for all collections tested. The MAP obtained with the new weights has relative improvement over the baseline by 3.3 to 15.2%, with statistical significance at the 95% confidence level across all four collections.

Fast Forward Index Methods for Pseudo-Relevance Feedback Retrieval

“…Figure 3 shows the plots of the mean average precision (MAP) and P@20 varying with the number of query expansion terms N QE , for three TREC collections. In our comparison of runtimes with various indexes as reported later (Section 4.5.4), we follow the calibration of N QE = 80 in Dang et al [2010]. This choice of N QE value is supported by the saturation of MAP for all the collections as shown in Figure 3(a).…”

“…In the current study, we follow the ranking score of Dang et al [2010] to select N QE terms from the top N PRF passages to yield the vector of expansion terms, q QE PRF . Since our current focus is on indexing methods rather than on retrieval effectiveness, we do not present the details of query expansion, which may be found in Dang et al [2010]. An expanded query vector q PRF is obtained by mixing the initial query q and the vector q QE PRF :…”

“…, q k ) are relevant. Scores are assigned to the words in the feedback passages, with the highest-ranking words being selected to expand the original query [Harman 1992;Dang et al 2010]. In the current study, we follow the ranking score of Dang et al [2010] to select N QE terms from the top N PRF passages to yield the vector of expansion terms, q QE PRF .…”

“…Scores are assigned to the words in the feedback passages, with the highest-ranking words being selected to expand the original query [Harman 1992;Dang et al 2010]. In the current study, we follow the ranking score of Dang et al [2010] to select N QE terms from the top N PRF passages to yield the vector of expansion terms, q QE PRF . Since our current focus is on indexing methods rather than on retrieval effectiveness, we do not present the details of query expansion, which may be found in Dang et al [2010].…”

“…With an inverted index, the retrieval time varies linearly with the number of query expansion terms (N QE ) [Carpineto and Romano 2012]. This can be a concern as some past research works found that it was necessary to use quite a large number of expansion terms in PRF, with N QE as high as 80 [Dang et al 2010] or even 200 [Abdul-Jaleel et al 2004]. Lavrenko and Allan [2006] had noted a large increase in retrieval time in RM compared with retrieval without query expansion.…”