Machine Translation Continues Its Journey to the Promised Land

Article Details

Global Watchtower

Common Sense Advisory Blogs

Posted by Donald A. DePalma on November 15, 2012 in the following blogs: Technology, Translation and Localization
Pages \| \|

Until a few years ago, anyone wanting to use machine translation (MT) had to buy it from a commercial supplier or build it themselves (see “Trends in Machine Translation,” Oct11). Then along came the Moses decoder, an open-source statistical MT (SMT) engine that is available at Github. Two commercialized variants -- a new do-it-yourself (DIY) product and introduction of a do-it-all solution -- continue the Moses momentum (see “Getting to the Promised Land of Machine Translation,” Jan12).

On the DIY side, MT software developer Precision Translation Tools announced DoMT, a solution that buyers can install on their Windows and Linux computers − Mac and server editions are in the works. Building on its community edition of Moses (DoMYCE), Bangkok-based PT Tools has added a commercial version with the following features:

A graphical user interface. While commonplace in most software products, open-source Moses has remained old-school in its use of command lines. DoMT’s tools lower the usage barriers for anyone wanting to bring machine translation behind their firewall. The GUI supports product installation, preparation of the corpus used to train the engine, training and evaluation of the resulting MT models, and workflow integration.
Distribution flexibility. A change in the end-user license agreement (EULA) increases the value of the MT engines generated by DoMT because licensees now have the royalty-free right to distribute the models they generate. Developers can customize MT engines for specific industries and sublicense them for use on their customers’ servers, thus expanding the machine translation ecosystem to include third-party applications.
Improved output quality. Managing Director Tom Hoar cited several user cases that demonstrate high quality scores based on small, seemingly inadequate training sets. He ascribed the improvements to work on the kernel and the corpus preparation kernel, Corpus Filtergraph, a toolbox that PT Tools developed and open-sourced for extracting, filtering, aligning, and selecting datasets for training MT engines. It also allows users to transform translation memories and other linguistic assets into training data.

On the “do-it-all” side of the Moses market, Xcelerator Machine Translation Solutions in Ireland has been beta-testing cloud-based KantanMT since the middle of this year. Founder Tony O’Dowd told us that their goal was to “demystify and simplify” machine translation, two critical requirements for market growth that we flagged in our 2011 report on machine translation.

Xcelerator is targeting the growing numbers of small- to medium-sized language service providers (LSPs) that want to – or must – offer machine translation to their clients. Our recent research on pricing shows that while a very small percentage of translation agencies use MT on every job, nearly a quarter of small LSPs, and more than half of medium-sized LSPs, will apply it if requested to do so by their clients (see “Trends in Translation Pricing,” Sep12). However, most providers don’t have the technical expertise to install an MT solution, train engines for their clients, and integrate it with their workflow. KantanMT aims to fill that gap.

LSPs can create custom engines by uploading translation memories to KantanMT.com where they are validated, cleansed, and incorporated into training for their own engines. They can then upload their clients’ files for translation using those engines. The site includes MT analytics for scoring output according to BLEU and F-Measure (see “Automated Translation Technology,” Nov06). According to O’Dowd, nearly 260 companies have signed up for a free test-drive. These beta users have created more than 10,000 engines to date. They own any data they create, along with related statistics. All user information is encrypted on the KantanMT multi-tenant servers.

Both PT Tools and KantanMT emphasize that users own all data and related statistics, and control their use. This may not be the case with all MT engines and portals, so buyers should carefully read terms and conditions from their prospective suppliers.

The bottom line: These new products from Precision Translation Tools and Xcelerator Machine Translation Solutions advance machine translation on three critical axes: engine training and refinement, improved linguistic quality, and increasing accessibility for translators and content specialists (see “Developers Work to Broaden Use of Machine Translation,” Mar10).

Categories