Re: Corpora: generalisation in text

Ted E. Dunning (ted@aptex.com)
Thu, 10 Sep 1998 17:29:42 -0700

It is actually quite instructive to check out the docs for Tom's and
Peter's program. It actually is able to grade essays on more than
just the surface characteristics. As you might guess, Tom and crowd
have extensively compared the accuracy to human performance in grading
essays. The results are quite impressive. My guess is that the good
essays fall into relative narrow bins with the bad ones falling into
fairly easy to classify sorts of excreta. Thus the grading ability of
LSI is not so much a measure of its ability to recognize good essays
as a measure of its ability to measure secondary characteristics which
correlate highly with good essays.