An Axiomatic Study of Learned Term-Weighting Schemes

Created by W.Langdon from gp-bibliography.bib Revision:1.3872

@InProceedings{Cummins:2007:SIGIR,
  author =       "Ronan Cummins and Colm O'Riordan",
  title =        "An Axiomatic Study of Learned Term-Weighting Schemes",
  booktitle =    "SIGIR 2007 workshop: Learning to Rank for Information
                 Retrieval",
  year =         "2007",
  editor =       "Thorsten Joachims and Hang Li and Tie-Yan Liu and 
                 ChengXiang Zhai",
  month =        "27 " # jul,
  organisation = "Microsoft",
  keywords =     "genetic algorithms, genetic programming",
  URL =          "http://ww2.it.nuigalway.ie/cirg/localpubs/axioms.pdf",
  abstract =     "At present, there exists many term-weighting schemes
                 each based on different underlying models of retrieval.
                 Learn- ing approaches are increasingly being applied to
                 the term- weighting problem, further increasing the
                 number of useful term-weighting approaches available.
                 Many of these term- weighting schemes have certain
                 features and properties in common. As such, it is
                 beneficial to formally model these common features and
                 properties.

                 In this paper, we introduce a term-weighting scheme
                 that has been developed incrementally using an
                 evolutionary learn- ing approach. We analyse one such
                 term-weighting function produced from the evolutionary
                 approach by decomposing it into inductive query and
                 document growth functions. Con- sequently, we show that
                 it is consistent with a number of axioms previously
                 postulated for term-weighting schemes. Interestingly,
                 we show that a further constraint can be de- rived from
                 the resultant scheme.

                 Finally, we empirically validate our analysis, and the
                 newly developed constraint, by showing that the newly
                 developed nonparametric term-weighting scheme can
                 outperform BM25 and the pivoted document length
                 normalisation scheme over many different query types
                 and collections. We conclude that the scheme produced
                 from the learning approach adds further evidence to the
                 validity of the axioms.",
  notes =        "https://research.microsoft.com/en-us/um/beijing/events/LR4IR-2007/",
}

Genetic Programming entries for Ronan Cummins Colm O'Riordan

Citations