i'm not sure if anyone else replied to you. I assume you're using the phrase-based decoder
1. the decoder doesn't use <s> & </s>. It's added somewhere in the LM code, so just commnet it out. 2. You can get the longest phrase by just setting the order of the LM to be very high, eg. 100000. However, the LM order or state control the hypothesis recombination - when it considers 2 hypos to be identical and so throws away the worse hypo. If your LM needs the entire sentence constructed so far, then the decoder can't make that decision. So don't be surprised if your LM causes good hypotheses to be pruned. On 20 April 2011 15:11, Casey Kennington <[email protected]> wrote: > Moses support, > > For my masters thesis, I am making a custom language model that I want > to use with moses. I have been successful in being able to compile moses > and use my lm with it. However, to utilize my model the best, I want to > remove all padding (<s> and </s>) and I want moses to stop breaking up > the sentences into the order size; I always want moses to submit the > longest possible phrase that it has, especially during evaluation. I'm > looking at the code and see a few places where I could maybe do it, but > I wanted to check first to see if there wasn't some easier way to make > that happen. > > Thank you! > -Casey Kennington > _______________________________________________ > 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
