Dear Moses expert,

I currently experiment translating at *morpheme level* (each morpheme
treated as a token). Instead of using a morpheme-based Language Model (LM),
I want to modify the Moses code to use a *word-based* LM so as to better
capture long-range word sequence constraint. I'd like to seek your advices
on whether my modification is sufficient, and if there's any issue that I
need to consider.

I intervened the Language Model scoring mechanism by modifying the two
methods LanguageModel::CalcScore and LanguageModel::Evaluate.
The modification idea is that in the two methods *CalcScore* and *Evaluate*,
the phrase input is a morpheme sequence, so I'll first put them back into
word form. The *contextFactor* that is used to query value from the LM will
be supplied with word tokens (instead of morpheme tokens).

As far as I know, the CalcScore method is only used for each target phrase
separately, so modifying it is relatively straight forward. Modifying
Evaluate method, on the other hand, is more complicated as it is for a
hypothesis, and involves phrases from previous hypothesis. I ensure that
every n-gram scored is in words and has part of it from previous
hypothesis's phrase.

* Here's an example about the method Evaluate:
** assume we have: prev phrase = A B1 ,  current phrase = B2 C1 C2 D1 D2 E
(where B1 B2 are two morphemes of a word B, similarly for C and D. A and E
are stand-alone morphemes.
** After concatenating into word form, they are A B C D E
So the 3-gram LM score will be computed for "A B C" and  "B C D" (since they
all contain part of the previous phrase)
The last n-gram C D E will be used to get the LMState

What I do not really understand is the use of the LMState: why do we need to
use the last n-gram to get the state? and how is it reused in later score
calculation?

Please correct me if my understanding is not correct. Any comments and
advices are very much appreciated.

Thank you in advance,

Regards,
Thang


-- 
Luong Minh Thang
WING group, School of Computing, National University of Singapore
http://wing.comp.nus.edu.sg/~lmthang
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to