Hieu Hoang http://moses-smt.org/
On 7 January 2018 at 02:09, Jesvir Zuniega <[email protected]> wrote: > Hi, I am pretty new in Moses SMT and in the field of NLP. > I've successfully trained two translators for Bikol (ISO 639-3 bcl > <http://www-01.sil.org/iso639-3/documentation.asp?id=bcl>) to English and > an English to Bikol. Evaluation for both translators resulted to decent > BLEU scores which is around 33.0 with parameters (ie. -mbr -mp -s) but > without additional pre-processing techniques like (POS tagging, > morphological analysis, etc...). > And my questions are: > > 1. Does training an Unsupervised Transliteration Model could help to > improve the translation quality involving a language (Bikol) that consist > of alphanumeric characters only? > transliteration is usually used if you need to translate proper nouns which are not in your parallel corpus. So if you have lots of untranslated names etc in your output, then it would be worthwhile to look into it. 2. What are the other pre-processing techniques that does not require a > lexicon (Dictionary)? I am having difficulties to find a good source of > Bikol lexicons because of its unpopularity. > better tokenization, word error correction or normalization would be good start 3. Is it possible to train a translation model with a POS tagged corpus as > a source language to a target language which corpus is not POS tagged or > vice versa? > Not sure what you mean. Try it and see > > Thank you. :) > > _______________________________________________ > 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
