Hieu Hoang wrote: > The 3rd & 4th columns of the phrase table contain alignment > information > about the words in the phrase. How it is created changed recently. > > However, this information isn't used by the decoder in the main > trunk so > shouldn't affect performance.
OK, so that's a red herring. > If the performance of the decoder is lower for the same weights, I > will be > very concerned. Can you tell me if this is the case? No, that's not what's happening. MERT arrives at different weights in the two versions - that's where they start to differ. With the exact same phrase table and reordering model, the two versions start diverging with the first iteration of MERT - about a point lower, which carries through to the last iteration, and then also a point lower evaluating on a held-out devtest. All of the data is exactly the same. Another difference I forgot about is that our new build is with IRST LM, the old one was with SRILM. Duh, that could very well send MERT in a different direction. I should have thought of that earlier. We may have to rebuild with SRI for me to get a better handle on the differences. > Which older version of Moses are you comparing it against? It's almost exactly a year old, sadly. What's the easiest way to tell what version it is? Miles asked about the size of the tuning set - it's 812 segments. That's not that small, is it? Thanks for your prompt replies and suggestions. - John D. Burger MITRE _______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support