Resource alignment for machine translation or implicit transfer

MT Summit VIII Santiago de Compostela (Spain), 18 - 22 September 2001 Resource Alignment and Implicit Transfer Jean Senellart, Mirko Plitt, Christophe Bailly, Francoise Cardoso Systran, Autodesk 1, rue du Cimetière - BP.7 95237, Soisy-sous-Montmorency Cedex France senellart@systran.fr, mirko.plitt@eur.autodesk.com, bailly@systran.fr, cardoso@systran.fr Abstract In this article we present the concept of “implicit transfer” rules. We will show that they represent a valid compromise between huge direct transfer terminology lists and large sets of transfer rules, which are very complex to maintain. We present a concrete, real-life application of this concept in a customization project (TOLEDO project) concerning the automatic translation of Autodesk (ADSK) support pages. In this application, the alignment is moreover combined with a graph representation substituting linear dictionaries. We show how the concept could be extended to increase coverage of traditional translation dictionaries as well as to extract terminology from large existing multilingual corpora. We also introduce the concept of "alignment dictionary" which seems promising in its ability to extend the pragmatic limits of multilingual dictionary management. Keywords syntactic alignment, customization, dictionary management, automatic translation, implicit transfer Introduction Everyday, feedback on Systran’s free translation services on the Web shows that despite the fact that Systran dictionaries contain an impressive number of entries, the size of the linguistic resources is still a bottleneck before high quality of general translation can be achieved. Continuous work on enrichment is thus a requirement in order to satisfy the more and more demanding translation user and to increase the translation quality. In a “classic” incremental model this task is complex; a great number of rules is accumulated and must be maintained and coordinated. It is not easy to always understand the final behaviour of the whole set when new rules are added. If we consider, in addition, parallel maintenance of several target languages, maintenance and quality control becomes even more complex. At the same time, experience with customization service shows that a single customization in a very narrow domain may be asking for very large glossaries, leading quickly to the same type of complexity in the handling of the resources mentioned above. If we analyse the distribution of entry complexity in an average Systran main dictionary of 200,000 entries, we see that about 80% of the entries are simple lexicalized entries. For part of these entries, we could have considered more complex coding, extending the capacity of variation of the entry, or restricting its application to a given context. Practically, on a large scale, this fine-tuning is not feasible considering human limits of organization, let alone any cost considerations. For such dictionary size, high-level organization of the information is thus a requirement. In this article we present the basic structuring work in progress on our multilingual resources. We will show, with examples, the following sources of complexity in dictionary • management and our solutions: Great redundancy in the linguistic information. This can be reduced by giving the system a real capacity to handle implicit information. via our alignment module. • Lack of factorisation of …

Content truncated to 500 words