Lance,
There was no change to Maxent. Maybe someone else may know. I only
know of one change to the name finder that allows it to report
back-to-back spans that report differing names. It shouldn't have
affected any of the models but how the models reported the spans was
wrong before... or should I say was correct; but, the wrong output was
pushed to the output instead of the correct span detected.
The name finder suffered a performance improvement without changing any
of the models that were trained properly.
James
On 6/16/2013 6:30 PM, Lance Norskog wrote:
My Lucene OpenNLP patch (LUCENE-2899) contains small test data sets
just to create unit tests. The patch runs MaxEnt on this test data and
then uses the .bin files to run simple unit tests. These datasets are
completely bogus, they only exist to demonstrate a complete round trip.
The chunker output changed slightly from 1.5.2-incubation to 1.5.3.
Was this expected? Was there some change in MaxEnt that caused
generated models to change? If there was, that's fine, as long as
someone expected this. But it does mean that the old models on
Sourceforge may be slightly wrong.
Lance
