Can We Translate Letters?

Proceedings of the Second Workshop on Statistical Machine Translation, pages 33–39, Prague, June 2007. c⃝2007 Association for Computational Linguistics Can We Translate Letters? David Vilar, Jan-T. Peter and Hermann Ney Lehrstuhl f¨ur Informatik 6 RWTH Aachen University D-52056 Aachen, Germany {vilar,peter,ney}@cs.rwth-aachen.de Abstract Current statistical machine translation sys- tems handle the translation process as the transformation of a string of symbols into another string of symbols. Normally the symbols dealt with are the words in differ- ent languages, sometimes with some addi- tional information included, like morpho- logical data. In this work we try to push the approach to the limit, working not on the level of words, but treating both the source and target sentences as a string of letters. We try to find out if a nearly unmodified state-of-the-art translation system is able to cope with the problem and whether it is ca- pable to further generalize translation rules, for example at the level of word suffixes and translation of unseen words. Experiments are carried out for the translation of Catalan to Spanish. 1 Introduction Most current statistical machine translation systems handle the translation process as a “blind” transfor- mation of a sequence of symbols, which represent the words in a source language, to another sequence of symbols, which represent words in a target lan- guage. This approach allows for a relative simplic- ity of the models, but also has drawbacks, as re- lated word forms, like different verb tenses or plural- singular word pairs, are treated as completely differ- ent entities. Some efforts have been made e.g. to integrate more information about the words in the form of Part Of Speech tags (Popovi´c and Ney, 2005), using addi- tional information about stems and suffixes (Popovi´c and Ney, 2004) or to reduce the morphological vari- ability of the words (de Gispert, 2006). State of the art decoders provide the ability of handling different word forms directly in what has been called factored translation models (Shen et al., 2006). In this work, we try to go a step further and treat the words (and thus whole sentences) as sequences of letters, which have to be translated into a new se- quence of letters. We try to find out if the trans- lation models can generalize and generate correct words out of the stream of letters. For this approach to work we need to translate between two related languages, in which a correspondence between the structure of the words can be found. For this experiment we chose a Catalan-Spanish corpus. Catalan is a romance language spoken in the north-east of Spain and Andorra and is considered by some authors as a transitional language between the Iberian Romance languages (e.g. Spanish) and Gallo-Romance languages (e.g. French). A common origin and geographic proximity result in a similar- ity between Spanish and Catalan, albeit with enough differences to be considered different languages. In particular, the sentence structure is quite similar in both languages and many times a nearly monotoni- cal word to word correspondence …

Content truncated to 500 words