Hi Radu, Did you measure the performance before and after the changes?
Have you tried to create a custom feature generator according to the language constraints? Maybe it would be more flexible to let the framework decide the importance of the constraints according to the training corpus, taking other context information into account. I am interested on that because I'm working on pos-tagging improvements for Portuguese, and I was focusing on improving the context generators according to language particularities. Regards, William On Tue, Mar 15, 2011 at 2:40 PM, Radu Simionescu <[email protected]> wrote: > Yes, something like that > > > > Pe Mar 15, 2011, la 6:12 PM, Jörn Kottmann <[email protected]> a scris: > > > On 3/15/11 4:49 PM, Radu Simionescu wrote: > >> I can't wrap everything there. The validSequence method is called for > each word > >> in the sentence. But to generate the sequence of sets of possible > outputs I just > >> need to loop once through the input sentence, before anything else, and > store > >> these in an instance, and then use them. So, this is what I did right > now, I > >> implemented the SequenceValidator. > > > > That is actually a pattern you can observe in many places of our feature > generation. > > Is it this way possible for you to wrap everything? > > > >> I am reading the sequence of posibilities > >> from a static variable in the validSequence method. I am setting the > static > >> variable before every tagging call... but this can't really be > considered > >> wrapping. This is not healthy coding, in my opinion. > > This part I do not understand actually. You mean, you pass the possible > sequences > > first to the sequence validator and then you call the tag method? I am > little confused. > > > > Jörn >
