Online-adaptive Moses


Online-adaptive Moses (Bertoldi, 2014) is an enhanced version of the well-known state-of-the-art SMT Moses Toolkit, which permits the dynamic (on-the-fly) adaptation of its statistical models without the need of their reloading, according to any suggestions coming from users.

Online adaptation is achieved by means of cache-based language and translation models, which reward their content using either a parametrizable time-decaying scoring function or pre-defined constant values. The caches can be populated at any time from input using an xml-based annotation or pre-populated from file.

Online-adaptive Moses has been successfully integrated in adaptive MT server connected to the MateCat Tool; in this way, it adapts to user post-editing, and supports the usage of terminology.


The development of the online-adaptive cache-based version of Moses was supported by the EU-funded projects MateCat (ICT-2011.4.2-287688) and MosesCore (ICT-2011-7-288487).


The online-adaptive cache-based version of Moses is distributed under the GNU Lesser General Public License (LGPL).


Usage instructions are available in the official Moses documentation;

in particular at this URL for the cache-based Phrase Table

and at this URL for the cache-based Language Model.

Source code

Source code of the online-adaptive cache-based Moses is currently distributed in the branch "dynamic models" of the Moses official Github repository.

It can be downloaded directly using this link.


If you intend to use online-adaptive cache-based Moses please cite:

Nicola Bertoldi. 2014. Dynamic Models in Moses for Online Adaptation, The Prague Bulletin of Mathematical Linguistics, vol. 101, pp. 1-22.

Note that several improvements have been done since the publication of the paper. Please consider the description in the Moses documentation the only valid.


For more information about online-adaptive cache-based Moses and to report bugs please contact: bertoldi [at] fbk [dot] eu