Hi all,
I have completed almost all the suggested and pending changes to the
evaluation module.Now module use Term Generator,Query Parser to index ,
form query object instead of modules from Trec harness code,
have flexibility for user to change Weighing scheme and enable bi-gram from
configuration itself.
After switching the indexing and query object to term generator and query
parser precision of the system have drastically increased .So problem of
precision was due to outdated modules of trec harness.Now low precision
problem is solved.
I have shared gimps of performance for all weighting scheme on FIRE Data
Set here: Xapian Evaluation
Results<https://docs.google.com/spreadsheet/pub?key=0AoCWuAKuwBGfdGljZGJ2MDJIY0dzdkVaTFFBLU1QQWc&single=true&gid=0&output=html>
.
Most of evaluation results seems fine.BM25(default Weighting scheme) have
outperform the current implementation of all the weighting scheme.But
according to surveys and initial paper we used as benchmark to implement
this Language Model Weighing scheme.Our Language Model Weighting scheme
should either perform equivalent or should outperform the BM25 Weighting
Scheme.
I am investigating on the reason for trashed performance of Language
Modelling but problem might be log trick, as some documents might be
getting negative score due to not getting on good transformation factor for
and still we need to investigate the results using different parameters and
Smoothing scheme.Results shown above are for Current Default Schemes.
Following the diversion from expected result, i should now see for problem
in implementation of Language Model and improve the scheme using the
evaluation result.
Please comments on the Evaluation Results,Language Model Weighting scheme
implementation, Future course of action.
Thanks,
--
with regards
Gaurav Arora
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.xapian.org/pipermail/xapian-devel/attachments/20120730/fc4cd0a5/attachment.html>
