From: Roger P. Menezes <rogerm@ya...> - 2006-04-25 06:16:49
Checked out the scaling technique to curb overflow problems and I'm fine
with things now. I had another question:
double crfScore1 = crf.score(testRecord);
double crfScore2 = crf.score(testRecord);
Now crfScore1 and crfScore2 are marginally different. I looked up the
viterbiSearch() function and it had the following statement about which
I'm still confused.
corrScore += (Ri.get(dataSeq.y(i)) + ((i >
This is even before the dataSeq has been assigned labels. Could you
please clarify how dataSeq.y(i) is used here?
Roger P. Menezes wrote:
> Dear M'am,
> I am trying few techniques from McCallum's paper "Confidence
> Estimation for IE". Are you doing any work on confidence estimation
> that can be made public now? Even if you have ideas and are willing to
> share that would be fine. We wanted to have confidence estimates for
> every field extracted before showing it to the user?
> I tried calculating Z(x) and was running into overflow problems. Was
> adding up the (prevAlpha[yp] + Mi[yp][yi] + Ri[yi]) for all previous
> nodes in the trellis structure to get the alpha[yi] (in a separate
> vitterbisearch() function). I saw the current cvs code which has
> functionality to calculate Z(x) in getLogZx(DataSequence
> dataSequence). Does this calculate logZx? And further, where am I
> going wrong in my approach?