2.3 KiB
promix - for training translation model interpolation weights using PRO
Author: Barry Haddow <bhaddow [AT] inf.ed.ac.uk>
ABOUT
The code here provides the "inner loop" for a batch tuning algorithm (like MERT) which optimises phrase table interpolation weights at the same time as the standard linear model weights. Interpolation of the phrase tables uses the "naive" method of tmcombine.
Currently the interpolation only works for two phrase tables, but will soon be extended to work for more than two.
REQUIREMENTS
The scripts require the Moses Python interface (in contrib/python). It should be built first, following the instructions in that directory.
The scripts also require scipy and numpy. They have been tested with the following versions:
- Python 2.7
- Scipy 0.11.0
- Numpy 1.6.2
Run the test.py script to check that everything is functioning correctly.
USAGE
Since the code in this directory provides the inner loop for a batch tuning algorithm,
it is run from the increasingly inaccurately named mert-moses.pl. If you want to run
the optimiser directly, run main.py -h for usage.
A sample command for mert-moses.pl is as follows:
MOSES/scripts/training/mert-moses.pl \
input-file ref-file \
decoder \
ini-file \
--promix-training MOSES/contrib/promix/main.py \
--maximum-iterations 15 \
--promix-table phrase-table-1 \
--promix-table phrase-table-2 \
--filtercmd "MOSES/scripts/training/filter-model-given-input.pl --Binarizer MOSES/bin/processPhraseTable" \
--nbest 100 --working-dir ./tmp --decoder-flags "-threads 4 -v 0 " \
--rootdir MOSES/scripts -mertdir MOSES/bin \
--return-best-dev
Note that promix training requires a filter and binarise script, and that the phrase table
referenced in the ini file is not used. The argument --return-best-dev is not essential,
but recommended.
REFERENCES
The code here was created for:
Haddow, Barry (2013) Applying Pairwise Ranked Optimisation to Improve the Interpolation of Translation Models. In: Proceedings of NAACL 2013
See also:
Sennrich, Rico (2012). Perplexity Minimization for Translation Model Domain Adaptation in Statistical Machine Translation. In: Proceedings of EACL 2012.