Returning to an old discussion we once had on this list.
I've tried to tune some systems with no reference in the configuration file to a language model so that I can do Moses justice when comparing it to a number of lines of work I have done that use only translation models and various schemes of phrase table filtering. It seems that the tuning script will not run when supplied with a config file minus a reference to a language model. I could run a baseline with the weights learned from tuning with a config file but again this would be almost as meaningless as comparing my work against a Moses system with default weights as the weights would not be optimised to a system with no language model. I'm trying to come up with some kind of work around so that I can make a TM only Moses based baseline with sensible weights in the config file that are optimised to a TM only situation (no LM). I'm running short of ideas about how best to approach this. I was wondering if the following approach might work. If I was to make some kind of minimal dummy LM file with only one entry for some work which doesn't exist would that be enough to fool the MERT script to run so I could get some sensible weights optimised to a no LM scenario? I'm open to any other suggestions of how I can obtain such weights if anybody knows a clever way of doing this. Just to be clear this support request is not an invitation to discuss whether or not one would want to tune a system with no language model or not.
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
