MBMT: Memory-based machine translation

MBMT: Memory-based machine translation

MBMT is a software package for training and running a machine translation system. It is based on the k-nearest neighbor classifier as implemented in TiMBL, and also features a memory-based target language model based on WOPR, a memory-based language model. It assumes a word-aligned bilingual parallel training corpus, such as produced by GIZA++.

Features

Generates a machine translation model from a word-aligned parallel corpus

Based on k-nearest neighbor classification, as implemented in TiMBL

Fast training and translation

MBMT is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation.

Credits

MBMT is written by Antal van den Bosch and Peter Berck, with contributions from Ko van der Sloot.

See also

PBMBMT, phrase-based memory-based machine translation, is an extension to MBMT based on variable-width n-grams, or phrases, instead of the fixed-width trigrams of MBMT. Coded by Maarten van Gompel.

Archived versions

version 0.1

Download and installation

mbmt-0.1.tar.gz (139 Kb)

To install, please follow these basic instructions:

MBMT relies on an installation and availability in $PATH of the following two packages:

Timbl version 6.1.4 (or higher);
WOPR version 1.4.7 (or higher)

The tarball will unpack ('tar zxvf mbmt-0.1.tar.gz') in a directory called 'mbmt-0.1'.

In the 'mbmt-0.1' directory, issue a './configure' command, followed by 'make'.

If you want to install the software elsewhere, issue a './configure --prefix <install-dir>', followed by 'make' and 'make install'.

For step-by-step instructions on running MBMT, see this howto. Alternatively, the software comes with a shell script, mbmt.sh:

The mbmt.sh script included in the package runs a full training and translation process, based on a GIZA++-aligned 'A3.final' file, and a target-language text (one sentence per line);
For example, run the script by
1. changing to the 'etc/' directory,
2. moving the following two files there: JRC-Acquis.sample.A3.final and JRC-Acquis.source.test.txt, and
3. issuing 'mbmt.sh JRC-Acquis.sample.A3.final JRC-Acquis.source.test.txt'.
(These Dutch-English files are small extracts from the JRC-Acquis multilingual parallel corpus).

MBMT has been compiled successfully with gcc (4.0 - 4.2), on Intel/AMD platforms running several versions of Linux and the Mac OS X platform.

References

For more information and background on MBMT, see

Van den Bosch, A. and Berck, P. (2009). Memory-based machine translation and language modeling. The Prague Bulletin of Mathematical Linguistics No. 91, pp. 17-26.
Van den Bosch, A., Stroppa, N., and Way, A. (2007). A memory-based classification approach to marker-based EBMT. In F. Van Eynde, V. Vandeghinste, and I. Schuurman (Eds.), Proceedings of the METIS-II Workshop on New Approaches to Machine Translation, pp. 63-72. Leuven, Belgium.
Stroppa, N., Van den Bosch, A., and Way, A. (2007). Exploiting source similarity for SMT using context-informed features. In A. Way and B. Gawronska (Eds.), Proceedings of the 11th International Conference on Theoretical Issues in Machine Translation (TMI 2007), Skïvde University Studies in Informatics 2007:1, pp. 231-240.

Sponsor

MBMT is developed as part of the Implicit Linguistics project, funded by NWO, the Netherlands Organisation for Scientific Research.

Antal.vdnBosch@uvt.nl | Last update: Wed Nov 4 2009