(035) previous ~ index ~ next

---

To: TDT distribution <tdt-distrib@ldc.upenn.edu>
From: George Doddington <doddington@nist.gov>
Subject: topic-weighted versus story-weighted error estimates
Date: Thu, 25 Feb 1999 21:44:52 -0500

---

Here are the results of a modeling exercise I did to determine
the optimum value for the variable weighting scheme proposed in
the TDT evaluation plan version 3.7. I'm including both an MS
word version and a postscript version. The results are that
the gains from the variable weighting scheme appear to be
small when compared to topic weighting. Topic weighting,
however, demonstrates a factor of 3 lower variance for the
error estimates than story weighting. The result of this
exercise is, therefore, that we will be using topic-weighting
exclusively in subsequent TDT evaluations.
--
George Doddington at NIST: doddington@nist.gov or 301/975-3261

Click below to retrieve the document:

Error_Variance.doc (MS Word file)
Error_Variance.ps (PostScript file)

---

(035) previous ~ index ~ next

Last updated Thu May 13 09:28:17 1999