Towards the Integration of MT into a LSP Translation Workflow

Towards the Integration of MT into a LSP Translation Workflow David Vilar1, Michael Schneider2, Aljoscha Burchardt1, and Thomas Wedde2 1DFKI , Language Technology Lab , Berlin, Germany {name.surname}@dfki.de 2beo GmbH, Stuttgart, Germany {name.surname}@beo-doc.de Abstract This user study reports on an ongoing pilot that aims at using machine translation on a large scale, for the translation of technical documentation for a globally acting auto- motive supplier. The pilot is conducted by a language service provider and a research institution. First results go beyond expec- tations. 1 Introduction In real-world translation environments efficiency, both in terms of cost and time, is of critical im- portance. Even more when the volume of texts to translate is large. Machine translation (MT) seems to be a good candidate for achieving these goals, but somehow surprisingly the economic fea- sibility of MT and the fitness for real-world needs of professional translators and Language Service Providers (LSPs) have been hardly analysed so far. The MT community tries to broaden the do- mains the translation systems are applied to. In the early years, research on statistical machine transla- tion concentrated on restricted domains, the touris- tic domain being a typical example. As the qual- ity of the translations got better, the difficulty of the task was increased by moving to richer do- mains. The WMT evaluations are another exam- ple of this trend. In the first editions (Koehn and Monz, 2005) the data the systems were trained and evaluated on consisted only of the proceedings of the European Parliament. In more recent editions (Callison-Burch et al., 2011) the (parallel) training data still mostly consists of europarl data, but the evaluation has moved to the news domain, with a much wider variety of topics. c⃝ 2012 European Association for Machine Translation. The goal of this research direction is clear: to produce an “universal translator” that is able to translate any type of text. This is however a very optimistic goal and current systems are still very far from it. And it also may not be the optimal goal for professional translators. When a LSP has a translation request, it is usually accompanied by guidelines of style, vocabulary, etc. Also, the do- main is usually quite restricted. Not much topic variation can be expected from, say, user manuals of heavy machinery. As such, the research community has perhaps overlooked a potential niche where machine trans- lation, in its current development state, can prove to be beneficial. At the same time, potential cus- tomers are reluctant when it comes to financing the development of specialised MT engines as their idea is that MT comes for free. In this paper we present a pilot study where we analyze how a state- of-the-art machine translation system performs in a real-life environment. The work is a collabora- tion between a LSP (beo) providing the experience on real-life translation tasks for Bosch, and a re- search institution (DFKI) providing the know-how about statistical machine translation. The paper is structured as follows: Section 2 presents the …

Content truncated to 500 words