Perceptron model normalization not working?

[Jörn Kottmann <ko...@gmail.com>]

Hi all,

after experimenting with the perceptron sequence training for
the name finder I found an issue with the normalization of
the perceptron model.

The perceptron models eval methods outputs scores which
indicate how likely an even is, when they are normalized
the scores should be between zero and one.

I observed that the score also are Infinity, which does
not work that well for beam search, depending on the scores outputted
it is not able to find a sequence at all.

Why is a score Infinity? They are normalized with the exponential
function which returns Infinity if the value for example is 850.

Any suggestions how we should fix the normalization?

Thanks,
Jörn

Re: Perceptron model normalization not working?

[Jörn Kottmann <ko...@gmail.com>]

On 5/24/11 10:57 PM, Jason Baldridge wrote:
> I think there are actually some problems to fix with the Perceptron that I
> will get to on Thursday or so after I get a few major things off my plate.
> Will explain then.

Thanks, we should also perform the fixes/refactoring to eliminate the 
duplicate
code in the perceptron sequence trainer like you did in the non-sequence 
perceptron
trainer.

It seems also that the stop criteria in the non-sequence perceptron trainer
does not really work, I did a few test with a very high iteration count, 
and the accuracy
just started jumping back and forth between two values.

Beside that I already fixed a performance bug in the sequence perceptron
trainer, see OPENNLP-185.

Jörn

Re: Perceptron model normalization not working?

[Jason Baldridge <ja...@gmail.com>]

I think there are actually some problems to fix with the Perceptron that I
will get to on Thursday or so after I get a few major things off my plate.
Will explain then.

On Tue, May 24, 2011 at 3:44 PM, Jörn Kottmann <ko...@gmail.com> wrote:

> Hi all,
>
> after experimenting with the perceptron sequence training for
> the name finder I found an issue with the normalization of
> the perceptron model.
>
> The perceptron models eval methods outputs scores which
> indicate how likely an even is, when they are normalized
> the scores should be between zero and one.
>
> I observed that the score also are Infinity, which does
> not work that well for beam search, depending on the scores outputted
> it is not able to find a sequence at all.
>
> Why is a score Infinity? They are normalized with the exponential
> function which returns Infinity if the value for example is 850.
>
> Any suggestions how we should fix the normalization?
>
> Thanks,
> Jörn
>



-- 
Jason Baldridge
Assistant Professor, Department of Linguistics
The University of Texas at Austin
http://www.jasonbaldridge.com
http://twitter.com/jasonbaldridge