Translation memory as a robust example-based translation system

82 Translation Memory as a Robust Example-based Translation System Hodász, Gábor**; Grőbler, Tamás*; Kis, Balázs* * MorphoLogic, Budapest {grobler,kis}@morphologic.hu ** Pázmány Péter Catholic University, Budapest Department of Information Technology hodasz@morphologic.hu Abstract This paper introduces a new approach to translation memories. The proposed translation technology uses linguistic analysis (morphology and parsing) to determine similarity between two source-language segments, and attempts to assemble a sensible transltion using translations of source-language chunks if the entire source segment was not found. This is achieved by integrating a rule-based machine translation (RBMT) engine. The drawback of this ap- proach is language-dependence; however, proper grammar acquisition methods are being developed to speed up grammar preparation for further language pairs. The paper discusses the basic processes of the proposed TM. Then the underlying machine translation engine is described. This is followed by an outline of the inegrated translation support tool where the proposed TM architec- ture is to become a core component. We argue that both the recall and the precision of a language-aware translation memory is higher than those of character-based systems. To this end, the paper introduces an evaluation scheme currently being prepared for use for testing the TM, followed by some estimated figures, the latter providing arguments in favour of the proposed TM scheme. 1. Introduction Commercially available translation memories do not involve linguistic knowledge. Instead, they deter- mine similarity between translation units on the ba- sis of pure mathematical distance calculations, with the majority of the algorithms using fuzzy indexes. Although most such systems recognize and handle non-translatable passages (such as dates, numbers, some abbreviations), and incorporate terminology management modules as well, existing translation memory systems do not systematically treat and recognize morpho-syntactic or even syntactic simi- larities, if the translation units themselves are too different in terms of characters. This paper presents a proposed system addressing these particular prob- lems, one that attempts to assemble translations from sub-sentence units stored in its database, both in the source and the target languages. The proposed language-aware translation mem- ory program is being implemented as a language- dependent translation support tool, first developed for the English-Hungarian language pair. The devel- opment process aims at surpassing the recall and precision of existing systems, resulting in a tool that offers translations more often, with the suggestions being closer to the desired translation. In the first version, the translation is assembled from stored translations of noun phrases and the morpho- syntactic skeleton of the source unit. The morpho- syntactic skeleton is a sequence of lemmas and morpho-syntactic parses of the words in the source unit, with a symbolic NP slot at the place of each noun phrase. This scheme may result ambiguities – an undesired phenomenon in CAT systems –, but the best translation will be a far closer approxima- tion of the desired one than in existing systems. Therefore, in the devised user interface, the user will be able to select the best translation from multiple suggestions. However, in the subsequent revision phase – when the human …

Content truncated to 500 words