Hi all, I have just read all your mails about the tagger and I would make to things clear:
- There is no way of changing the def-mult without retraining because every def-mult ends up being a HMM state, so adding one of them or removing one of them changes the topology of the HMM and needs retraining. When doing this retraining, the most clean way would be to regenerate the .dic file. - When you make use of apertium-tagger-apply-new-rules you are introducing some new zeroes in the transition probabilities. This can be safely done without retraining, but the changes introduced cannot be undo. Important: be sure that the states (labels and adef-mult) are exactly the same used when obtaining the probabilities you are changing. To know why the tagger was crashing with segmentation fault, I would need additional information. Fran, do you know when this happened? Was it reading the dic file, initialising the probabilities or doing a Baum-Welch iteration? Cheers -- Felipe El 03/08/11 20:05, Francis Tyers escribió: > El dc 03 de 08 de 2011 a les 15:59 +0100, en/na Jimmy O'Regan va > escriure: >> On 3 August 2011 14:36, Francis Tyers<[email protected]> wrote: >>> El dc 03 de 08 de 2011 a les 14:38 +0100, en/na Jimmy O'Regan va >>> escriure: >>>> On 3 August 2011 14:26, Francis Tyers<[email protected]> wrote: >>>>> Anyone got any idea how to fix it ? Also, is there any way to update the >>>>> def-mults without retraining the tagger ? I just want to fix the problem >>>> >>>> apertium-tagger-apply-new-rules >>> >>> Doesn't work. >>> >> >> Are you actually /updating/ an existing .prob file, or just trying to >> replace the coarse categories in an unrelated file with another set? >> The former should work, the latter won't. New coarse tags require >> (re)training, there's no way around that. > > Actually, I just managed to train it now by commenting out all the > dictionary entries. And just leaving two coarse tags. Now we get the > desired results: > > $ echo "govoriću" | lt-proc sh-mk.automorf.bin | apertium-tagger -g > sh-mk.prob > ^govoriti<vblex><imperf><iv><inf>+htjeti<vbmod><clt><futI><p1><sg>$ > > :) > > Fran > > > > ------------------------------------------------------------------------------ > BlackBerry® DevCon Americas, Oct. 18-20, San Francisco, CA > The must-attend event for mobile developers. Connect with experts. > Get tools for creating Super Apps. See the latest technologies. > Sessions, hands-on labs, demos& much more. Register early& save! > http://p.sf.net/sfu/rim-blackberry-1 > _______________________________________________ > Apertium-stuff mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/apertium-stuff -- Felipe Sánchez Martínez Dep. de Llenguatges i Sistemes Informàtics Universitat d'Alacant, E-03071 Alacant (Spain) Tel.: +34 965 903 400, ext: 2966 Fax: +34 965 909 326 http://www.dlsi.ua.es/~fsanchez ------------------------------------------------------------------------------ BlackBerry® DevCon Americas, Oct. 18-20, San Francisco, CA The must-attend event for mobile developers. Connect with experts. Get tools for creating Super Apps. See the latest technologies. Sessions, hands-on labs, demos & much more. Register early & save! http://p.sf.net/sfu/rim-blackberry-1 _______________________________________________ Apertium-stuff mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/apertium-stuff
