Hiya john 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. 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? If so, would it be possible to send me a copy of the model for me to debug? Which older version of Moses are you comparing it against? -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of John D. Burger Sent: 02 August 2008 06:04 To: [email protected] Subject: [Moses-support] Trying to debug reduced performance with new Moses Hi - I'm trying to replicate an experiment done with an older version of Moses, against the latest version. Everything goes identically up to tuning, where the newer version starts with a lower BLEU and runs only 8 as opposed to 19 iterations of MERT, resulting in a lower evaluation score. Looking more closely at the data files, the phrase tables seem to be slightly different. All of the difference have to do with the third and fourth columns of the phrase tables, as in this example: < " ||| " . ||| (0,1) ||| (0) (0) ||| 0.000201491 0.0130605 0.00038117 0.000751688 2.718 --- > " ||| " . ||| (0) ||| (0) () ||| 0.000201491 0.0130605 0.00038117 0.000751688 2.718 For most of the differences, the older phrase table has more indices, but there are some where the reverse is true. The weights are not different. statmt.org seems to be down at the moment, so I'm not sure what those columns represent. These seem to be the =only= differences between the two versions, in the main (huge) phrase table as well as all the filtered versions. Of course, the MERT code, etc., is presumably different, but I thought I'd ask about this anyway. Does it mean anything? Thanks! - John D. Burger MITRE _______________________________________________ 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
