Using the Apertium Spanish-Brazilian Portuguese machine translation system for localization

Using the Apertium Spanish-Brazilian Portuguese machine translation system for localization Franc¸ois Masselot & Petra Ribiczey Autodesk Development S`arl Rue Puits-Godet 6 CH-2002. Neuchatel. Switzerland {Francois.Masselot Petra.Ribiczey}@eur.autodesk.com Gema Ram´ırez-S´anchez Prompsit Language Engineering, S.L. Av. Sant Francesc, 74, 1L 03195. L’Altet. Spain gramirez@prompsit.com Abstract We present a user case of the free/open- source Spanish ↔ Brazilian Portuguese Apertium machine translation system in- side the localization workflow of Au- todesk. This system, initially developed to perform general-domain translations, has been customized by Prompsit to fit with Autodesk needs and by respecting the lo- calization workflow as much as possible. This original scenario shows that post- edited machine translation can generate immediately significant productivity gains with publication-ready linguistic quality. 1 Introduction Autodesk is a leader company in design software and localizes its products and their user documen- tation from English into 20 languages. Offering products in the user’s language gives an obvious competitive advantage and helps gaining market share. However, when fully-fledged localization involves costs out of proportion with the sales fore- casts in emerging markets, low-cost localization solutions must be explored, otherwise the localiza- tion might be not happen at all. One of the 3D- mechanical design flagship products of Autodesk was in this situation: the sales team requested in vain a Portuguese version for the Brazilian market for several years. At Autodesk, low cost localization consists of localizing only the software user interface (UI), and not the user documentation, or the software UI and the Getting Started manuals, while leaving the reference, developers and other advanced manuals in English. This approach was chosen, resulting in c⃝ 2010 European Association for Machine Translation. 318,000 words of software UI and 111,000 words of the manuals (10% of the total documentation set) to be translated into Portuguese. In addition, machine translation (MT) combined with a full re- vision of the machine-translated material by pro- fessional translators was considered to help save on the translation costs and time. Autodesk had been preparing the introduction of MT in its localization workflow for several months. The long-term MT solution was designed around statistical MT (Plitt and Masselot, 2010), but a combination of circumstances opened the door to an unusual, later proving to be successful, localization scenario for this project. Instead of translating from English into Brazilian Portuguese, we decided to use the Spanish version as the ba- sis for the translation, resulting in chain translation between English and Spanish and then Spanish and Brazilian Portuguese. Once the Spanish version was completed, the plan was to machine-translate Spanish into Portuguese with the rule-based Aper- tium MT system, and perform the final revision between English and Portuguese, to cover for the mistakes resulting from chain translation. The use of standard CAT tools working at sentence level supported such a 3-way localization. The reasons for our choice were the following: • Autodesk’s in domain English → Portuguese data (600,000 words) appeared to be insuffi- cient to obtain satisfactory results with statis- tical MT. • An evaluation of several rule-based MT engines …

Content truncated to 500 words