Existing IR-based expert finding generally follows two methods, i.e. the proflfe-based method and the voting-based one. However, neither the expert-relevant data collected in the proflfe- based method nor the query-refevant data used for the voting- based method is compietefy accurate within the confines of current refevance ranking approaches. This problem has been rarely discussed, but impedes expert finding. On this issue, we provide a feasibie sofution, that is, the coifection can be flftered to generate a subset of high-precision reievant data for further processing. In this paper, we propose two perspectives of filtering approaches, i.e. the query-centered perspective and the expert-centered one. For both perspectives, some specific strategies are also discussed and experimented under the CERC collection using the TMJAC model, a voting-based method. On such basis, the different preferences of two perspectives are revealed. Further, to examine the stability of filtering, we examine the filtering strategies using a profile-based method and also testify the effects under the W3C collection. In conclusion, the filtering we proposed is a universal approach of improving expert finding performance.
[1]
Peter Bailey,et al.
The CSIRO enterprise search test collection
,
2007,
SIGF.
[2]
Marc Moens,et al.
Named Entity Recognition without Gazetteers
,
1999,
EACL.
[3]
M. de Rijke,et al.
Formal models for expert finding in enterprise corpora
,
2006,
SIGIR.
[4]
Yiqun Liu,et al.
THUIR at TREC 2005: Enterprise Track
,
2005,
TREC.
[5]
Yiqun Liu,et al.
A CDD-based formal model for expert finding
,
2007,
CIKM '07.
[6]
Yiqun Liu,et al.
THUIR at TREC 2007: Enterprise Track
,
2007,
TREC.
[7]
Stephen E. Robertson,et al.
Simple BM25 extension to multiple weighted fields
,
2004,
CIKM '04.
[8]
Alfred Kobsa,et al.
Expert-Finding Systems for Organizations: Problem and Domain Analysis and the DEMOIR Approach
,
2003,
J. Organ. Comput. Electron. Commer..