Latest release

MMT 2.5

@davidecaroselli davidecaroselli released this Jul 2, 2018

MMT release for Ubuntu 16.04

Changelog v2.5

  • Implemented private memories feature: you can specify a owner for each memory, that memory won't be accessible by any other user.
  • Implemented multilingual evaluate and translate commands. You can specify the source and target language with -s and -t command line options.
  • Fixed an important bug in sentence splitting impacting translation quality.
  • Improved server reliability with auto-recovery policy in case of problems with Neural Decoder python processes.
  • General improvement in pre-processing pipeline and URL tokenization.
  • Added support to .mqxliff files
  • Several bugfixes

Backwards-compatibility:

  • v2.4 engines or older are not compatible with version 2.5.

MMT 2.4

@davidecaroselli davidecaroselli released this Feb 16, 2018 · 120 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.4

  • Fixed Chinese pre/post-processing pipeline.
  • Enhanced Numeric Expression projection algorithm: reduced problems with "???" placeholders in the translation.
  • Several small bugfixes.

Backwards-compatibility:

  • v2.x engines with at least one CJKV language, are no more compatible with version 2.4
  • all other v2.x engines are directly compatible with version 2.4.

MMT 2.3

@AndRossi AndRossi released this Jan 15, 2018 · 209 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.3

  • Improved online adaptation quality employing a better selection algorithm.
  • Improved online adaptation speed reorganising the adaptation pipeline.
  • Fixed problem when importing tiny TMX files.
  • Several small bugfixes.

Backwards-compatibility:

  • v2.2 and v2.1 engines are directly compatible with version 2.3. However, Aligner is not optional anymore in 2.3 engines.
  • v2.0.x engines can be upgraded to version 2.3 using extras/porting/convert_from_v2.0.2_to_v2.1 script.
  • v1.x engines or older are not compatible with versions 2.x.

MMT 2.2

@AndRossi AndRossi released this Jan 3, 2018 · 222 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.2

  • Lighter XML encode: MMT now keeps quote and double quotes unchanged.
  • Improved XLIFF translation script with support for inner xml in source nodes.
  • Alignment now includes an alignment score.
  • A new filter now detects wrong languages during mmt create.
  • Important bugfix in mmt create when resuming a neural training.
  • Doubled speed in BPE creation during neural training .
  • Asymmetric encoder/decoder layers structure: it is now possible to specify the numbers of layers for encoder and for decoder separately.
  • General improvements.
  • Several small bugfixes.

Backwards-compatibility:

  • v2.1 engines are directly compatible with version 2.2.
  • v2.0.x engines can be upgraded to version 2.2 using extras/porting/convert_from_v2.0.2_to_v2.1 script. Additionally, you may need to update your CUDA driver to version 9.1 and your PyTorch version to 0.3. You will find more information on this here.
  • v1.x engines or older are not compatible with versions 2.x.

MMT 2.1

@AndRossi AndRossi released this Dec 15, 2017 · 260 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.1

  • Neural engines now use CUDA 9.1 drivers and pyTorch 0.3, providing a great boost in performances.

  • Improved neural decoder now makes use of an average of the last N checkpoints during the training.

  • New termination policy in neural training based on perplexity function trend, tested on a wide variety of language directions.

  • Translation requests now feature a priority value and are handled in a configurable queue.

  • MMT now supports sentence splitting for very long translation requests.

  • Enhanced support for XLIFF files through a new xliff-translate script.

  • General improvements.

  • Several small bugfixes.

Backwards-compatibility:

  • v2.0.x engines can be upgraded to version 2.1 using extras/porting/convert_from_v2.0.2_to_v2.1 script. Additionally, you may need to update your CUDA driver to version 9.1 and your PyTorch version to 0.3. You will find more information on this here.
  • v1.x engines or older are not compatible with versions 2.x.

MMT 2.0.2

@AndRossi AndRossi released this Nov 14, 2017 · 426 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.0.2:

  • Improved translation quality in Neural Adaptive Engine: alignments are derived directly from neural AttentionModel.
  • Aligner is now an optional component for Neural Adaptive Engine and can be disabled via config file.
  • Neural Decoder now supports N-Best feature.
  • APIs backward-compatibility restored.
  • Cleaning pipeline is now applied to imported TMX files too.
  • General improvements.
  • Several small bugfixes.

Backwards-compatibility:

  • v2.x engines are directly compatible with version 2.0.2.
  • v1.x engines or older are not compatible with versions 2.x

MMT 2.0.1

@AndRossi AndRossi released this Nov 2, 2017 · 471 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.0.1:

  • Fixed bug in Neural engine adaptivity; translation quality is strongly improved.

  • RAM consumption during Aligner training was drastically reduced.

  • Fixed severe bug in thread management.

  • Fixed severe bug in data stream sequence management: it prevented deletions to be completed and could make MMT get stuck on start.

  • Parameters of neural training are now settable from command line.

  • Bugfix in BPE model load during engine resume.

  • General improvements.

  • Several small bugfixes.

Backwards-compatibility:

  • v2.0 engines are directly compatible with version 2.0.1.
  • v1.x engines or older are not compatible with version 2.0.

MMT 2.0

@davidecaroselli davidecaroselli released this Oct 10, 2017 · 535 commits to master since this release

MMT release for Ubuntu 16.04

Changelog v2.0:

  • Brand new Neural-Adaptive MT engine with deep-neural networks for best translation quality. You can create a neural engine easily: just add --neural flag to mmt create command.
  • New Memory concept replaces old domain nomenclature.
  • Improved APIs for multi-language support.
  • Several bugfixes

Backwards-compatibility:

  • v1.x engines or older are not compatible with version 2.0.

MMT 1.0.3

@AndRossi AndRossi released this Sep 7, 2017 · 855 commits to master since this release

MMT release for Ubuntu 14.04 and Ubuntu 16.04

Changelog v1.0.3:

  • Bulk import is now allowed via shell command mmt import. Use --path argument to import all tmx and parallel files in a folder
  • Domain renaming is now allowed, both via shell command ./mmt rename and via API
  • Several bugfixes

Backwards-compatibility:

  • 1.0, 1.0.1 and 1.0.2 engines are directly compatible with version 1.0.3
  • 0.15.1 engines can be ported to version 1.0.3 just executing on them the extras/porting/model-port script

0.15 engines or older are not compatible with version 1.0.3.

MMT 1.0.2

@AndRossi AndRossi released this Jun 30, 2017 · 879 commits to master since this release

MMT release for Ubuntu 14.04 and Ubuntu 16.04

Changelog v1.0.2:

  • Improved CJK tokenization;
  • Implemented '--accuracy' option for 'mmt tune'
  • Enhanced SimpleRecaser
  • Bumped Cassandra to 3.11 and Kafka to 0.10.2.1.
  • Several bugfixes

Backwards-compatibility:

  • 1.0 and 1.0.1 engines are directly compatible with version 1.0.2
  • 0.15.1 engines can be ported to version 1.0.2 just executing on them the extras/porting/model-port script

0.15 engines or older are not compatible with version 1.0.2.