One of the biggest stumbling blocks to bringing machine translation into an organization has been the complexity of the technology, requiring either dedicated linguists to configure the systems or handing off content, translation memories, and other linguistic assets to the MT developers — and ceding cash and control. We spoke to Asia Online CEO Dion Wiggins about the company’s Language Studio which he will demonstrate at this week’s Localization and Translation Thailand conference.

To date, statistical machine translation (SMT) developers have done all the configuration and training for these engines themselves, demanding a leap of faith from the buyers as to how they create these very opaque black boxes. Wiggins told us that Asia Online has automated its own internal processes — some 400 of them — into its Language Studio Pro tool which is available now in beta form and for commercial distribution in January 2010. It will initially be marketed to language service providers (LSPs) that want to offer MT to their clients such as Moravia Worldwide and to specialist integrators like Cross Language.

The first version of Language Studio Pro will have four major components, each meant to address a major aspect of training and configuring an MT solution, increasing the quality of the data, or managing related processes:

• Build an SMT engine. Statistical machine translation requires bilingual data from parallel corpora (the same documents in different languages). It can automatically align bilingual data, import existing translation memories, and validate alignment — thus providing vocabulary data for the SMT engine. One clever innovation lets Language Studio crawl a search engine’s output for candidate content sources behind the firewall or on the web, strip out any markup or formatting, and present it for consideration in training the engine. Besides data collection, this module provides tools to cleanse the data, develop a concordance, prepare translation memories and glossaries, create test sets, and optimize the engine. The focus is to create a clean data source, an issue that we have raised in our members-only report “TAUS and TDA Aim to Increase Translation Data Sharing.”

• Measure the results. This module gauges the output based on commonly used metrics such as BLEU scores, the translation error rate (TER), and human assessments. Research such as the “Study on the Impact of Data Consolidation and Sharing for Statistical Machine Translation” should inform this module.

• Refine the engine. A post-editing client module facilitates the process to improve the engine. It automatically analyzes any edits and uses the same crawling technology to compare and contrast human input with candidate data sources to create a sample concordance to reduce flagged errors in the future. We found post-processed MT to be a popular first step for users in our report on “The Business Case for Machine Translation.”
• Manage the system. Language Studio supports roles-based access to using the system — project manager, post-editors, human translator, and proof-reader. This module also manages system updates and coordinates corrections with project managers and participant input tracking.

Wiggins said that the goal of Language Studio Pro is to “put powerful tools into the hands of smart people to make them even smarter” when it comes to machine translation. This makes sense — to date, SMT training and refinement has been the exclusive province of the developers of MT software. Enabling translation automation druids at LSPs, specialty system integrators, and inside corporations to do the work themselves should increase the ability of those organizations to try out MT and tailor it to their specific needs. We think that this is a major step in making machine translation a full participant in global content or information management.