Customizing an MT system for unsupervised automatic email translation

BUILDING AN ENVIRONMENT FOR UNSUPERVISED AUTOMATIC EMAIL TRANSLATION Salvador Climent Joaquim Moré Antoni Oliver Interdisciplinary Internet Institute (IN3) / Universitat Oberta de Catalunya (UOC) Av. Tibidabo 39 08035 Barcelona scliment@uoc.edu jmore@uoc.edu aoliverg@uoc.edu Abstract By this paper, we present the INTERLINGUA Project1: its design and current work. The goal of the project is achieving fully-automatic (no pre-edition, no post-edition) translation of emails in the virtual campus of the Open University of Catalonia (UOC). The problem of un- supervised machine translation of emails is discussed. Then we describe the strat- egy designed to build the system, includ- ing a multiple-level evaluation process and the building of several automatic pre- edition, post-edition and unknown-word extraction modules. Last, the work carried on so far on building such decision-taking modules is presented. 1 Introduction and Rationale The UOC (www.uoc.edu) is a virtual University currently offering 17 official university degrees, 1 Ph.D. program, and several dozens of other courses. Communication between students, profes- sors, and supervisors is completely performed via email or email-like means –e.g. kinds of news- groups– within a virtual campus and a system of virtual classrooms. Courses are taught in Catalan and/or Spanish. 1 Funded by the Interdisciplinary Internet Institute; IN3- IR 226 Many of the students are Catalan speakers, which means that, due to the linguistic situation in Cata- lonia, they are fluent in both Catalan and Spanish. Another part of them are Spanish speakers living in Catalonia, which most of the times can read Cata- lan but can’t write it properly. Last, due to the re- cent expansion of the UOC to the rest of Spain and South-America, there is a new sector of students who are strict monolingual speakers of Spanish. Such a situation might lead to gradual substitution of Catalan by Spanish in the classrooms. A statisti- cal study on the language used to write and to reply messages at the UOC, covering 1 year of 4 news- groups, shows that, although 68.9% of the users can be considered spontaneous users of Catalan, 42.9% of these Catalan-speakers code-switch to Spanish when replying to messages in that lan- guage. The INTERLINGUA Project aims to over- come such effect by allowing effective cross- linguistic communication using machine transla- tion (MT). We foresee that the goal (MT for communication, not for gisting or budget cutting in document trans- lation) is reachable since Catalan and Spanish are two Romance languages structurally quite similar at all levels. This makes that MT systems perform at high levels of quality between the pair, as pre- liminary tests of translation of pre-edited texts have shown. It seems clear that MT between Cata- lan and Spanish (and vice versa), when using a knowledge-rich system, just needs of good lexi- cons and a tuning effort on solving some reluctant ambiguities to produce fully comprehensible and faithful texts. Notwithstanding, it is clear that the special task of machine-translating emails in the environment we have described above shows a number of addi- tional important problems, which can be classified in three …

Content truncated to 500 words