Tuning was always gave better BLEU score by~8 points or more , suddenly now 
tune step (MERT) gives degradded BLEU score starting from iterration 1 till the 
last one !!!!
 
Marwa N. Refaie



To: [email protected]; [email protected]
From: [email protected]
Date: Wed, 6 Jan 2016 17:33:22 +0000
Subject: Re: [Moses-support] Baseline: Problem with tuned weights


  
    
  
  
    are you sure you got the source and target side of the tuning set
    the right way round?

    

    And the tuning set must be tokenized and lowercase/truecased EXACTLY
    the same way that you tokenized & lowercaseed/truecased the
    training data

    

    On 04/01/16 17:16, Raphael Höps wrote:

    
    Hello,
      

      

      I did the moses-baseline tutorial to train and tune and
      translation model for English to German. After finishing the
      system it seemed to work quite well at first but then I noticed
      that the tuning step seemed to actually having made my system
      worse! I really don't know what I did wrong. I sticked very close
      to the tutorial. Here is what I did in detail:
      

      

      1. Training the TM to working/train/model.
      

      2. Tuning with a corpus that is a cut-down version of
      news-test2008. The main result of this process are the weights of
      the new file mert-work/moses.ini, right?
      

      3. Filtering of mert-work/moses.ini to a testing corpus (cut-down
      version of newstest2011).
      

      4. Translating the testing corpus and calculating BLEU-score. I
      got a score of 7.42.
      

      5. In a second test I used the default moses.ini file instead of
      the tuned one (and the same filtered and binarized model) and got
      a score of 8.22 on the same testing corpus!
      

      

      Something is probably wrong with the tuned moses.ini file. To find
      out, I translated the corpus that was used for tuning with both
      ini-files and calculated the scores:
      

      Untuned: 7.01
      

      Tuned: 6.70 (!)
      

      

      Now this is really odd! Furthermore in the tuned moses.ini file
      there is the line:
      

      # BLEU 0.0755253 on dev
      /home/rh/Studium/aktuell/LSS/moses/mosesdecoder/corpus/dev-small.en
      

      Why do I get a score of 6.7 instead? The files dev-small.en and
      dev-small.de where my tuning corpora.
      

      

      Do you have any idea, what I might have done wrong?
      

      

      For the tuning step, I used:
      

      cd ~/working
      

      nohup nice ~/mosesdecoder/scripts/training/mert-moses.pl \
      

         ~/corpus/dev-small.en ~/corpus/dev-small.de \
      

         ~/mosesdecoder/bin/moses train/model/moses.ini --mertdir
      ~/mosesdecoder/bin/ \ &> mert.out &
      

      

      I appended mert.log and the tune moses.ini file. Did anyone ever
      build a system for English to German and can say something about
      the trained weights in moses.ini? Do they seem okay?
      

      

      Thank you very much for your help!
      

      Greetings,
      

      Raphi
      

      

      

      

      

      

      

      
      

      _______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

    
    

    -- 
Hieu Hoang
http://www.hoang.co.uk/hieu
  


_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support                           
          
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to