Hi Liang, Here's a hack to avoid changing Moses code: First, append a unique identifier to every word in your source file. E.g., a source file with two lines like the following:
this is sentence 1 . sentence 2 would become: this~~1 is~~2 sentence~~3 1~~4 .~~5 sentence~~6 2~~7 Then when generating your sentence-specific phrase tables for these sentences, use the same IDs in the source words in those phrase table entries. Then concatenate all the sentence-specific phrase tables together and train as usual. The problem with this is that the concatenated phrase tables become really large, but you can filter the phrase tables a bit. This is what we did in here <http://ttic.uchicago.edu/~kgimpel/papers/gimpel+smith.smt08.pdf> (Sec. 5), which made things possible to run and didn't seem to affect the results. Kevin On Sun, Dec 13, 2015 at 2:41 AM, 姚亮 <[email protected]> wrote: > Dear support-Team, > I wanted to add a new feature into moses decoder which relies on > source contexts of the sentence to be translated. > my idea are as follows： > 1, for each test/dev sentence, a phrase table of all potential > phrases > that can be used during decoding is extracted from the aligned training > set. > 2, the translation score of phrase pairs in the source contexts > are computed and then added into the phrase table as a new feature. > therefore , in my setting, for each sentence , i got an phrase > table. but i don't know how to tuning on this situation. > i know maybe i can add a new Feature Function as the moses > tutorial says, but it diffcult for me to implement the code. > could you give my some advices? > > Thanks > Liang > > > > > _______________________________________________ > 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
