classify() method does not return correct distribution

[timjzee]

I'm using the LaMachine virtual environment on Ponyland

classifier = timbl.TimblClassifier("pl_type.master", "-mO:I1 -k 5 -G 0")

Should return probability distribution that adds up to 1, but ...

classifier.classify(("administrateur", "n", "i", "=", "str", "a", "=", "t", "|", "r", "-", "-", "+", "r"))

returns:

{'EN': 1, 'S': 1}

But the same classifier returns the correct distribution if the test() method is used instead:

{ EN 0.0526316, S 0.947368 }

[proycon]

The classify() only returns a class distribution only when you instantiate the TimblClassifier with the keyword argument dist=True, by default it doesn't do that indeed (because it takes some extra computation), similarly you can also decide whether to normalize the distribution (normalize=True (default)) or not.

[timjzee]

Thanks for responding! That still doesn't work for me though. Let's take the Readme example:

classifier = timbl.TimblClassifier("test-norm", "-a 0 -k 5 -G 0", dist=True, normalize=True)
classifier.append( (1,0,0), 'financial')
classifier.append( (0,1,0), 'furniture')
classifier.append( (0,0,1), 'geographic')
classifier.train()
classifier.classify( (1,0,0) )

Still returns:

('financial', {'financial': 1, 'furniture': 1, 'geographic': 1}, 0.0)

[proycon]

Hmm... that might be a bug after all then.. Could you try setting the timbl options +v+db +v+di explicitly and see if that makes a difference?

[timjzee]

That isn't it either. The training procedure actually shows that those options are added twice now (because it automatically adds them):

Calling Timbl API for training: -F Tabbed -a 0 -k 5 -G 0 +v+db +v+di +v+db +v+di

I think it's a bug in the C code, but I don't know C, so I got stuck there.

[proycon]

I'll do some extra debugging to see what's going on, I'm not sure either at this stage, it's been a while since I've looked at it. Let's hope we can solve it without dragging main developer @kosloot from his retirement ;)

[timjzee]

Haha, fingers crossed. As the test() method works perfectly fine, I don't actually need classify(). It would just be more convenient. Thanks for taking a look!

[kosloot]

awaking from retirement coma...
checking normalization code in https://github.com/proycon/python-timbl/blob/master/src/timblapi.cc
line 248.
this is dead wrong imho.
so retiring again

[proycon]

Indeed, that code looks very weird... I hope I didn't write it but I fear I might have... :$ Will investigate further...

[proycon]

I fixed the problem and released v2020.05.09. It's deployed it in LaMachine on ponyland as well. Can you give it a try to see if everything now works as expected? Neither the dist argument I mentioned previously nor the explicit +v+db +v+di options should be necessary anymore.

[timjzee]

I tested it with the Readme example and my own larger dataset. It works! Thanks for fixing!