Linking Translation Memories with Example-Based Machine Translation

The paper reports on experiments which compare the translation outcome of three corpus-based MT systems, a string-based translation memory (STM), a lexeme-based translation memory (LTM) and the example- based machine translation (EBMT) sys- tem EDGAR. We use a fully automatic evaluation method to compare the outcome of each MT system and discuss the results. We investigate the benefits for the link- age of different MT strategies such as TM- systems and EBMT systems. is coded in the same way and the translation(s) of the most similar match string(s) are returned as best (available) translations. Each returned translation is associated a match score M between 0% and 100% which indicates the similarity of the search string and the match string. In the STM the surface forms of the reference trans- lation's source language sides are used as a match string, whereas the LTM match strings are based on the lexemes of the reference translation's source lan- guage content words. The representation in the LTM is thus an abstract of inflection and derivation, while the STM stores the surface forms of the match strings without performing any abstraction. Some TMs such as Translator's Workbench and TRANSIT follow the STM approach. The ZERES implementation is a mix- ture of the STM and the LTM approaches. EBMT system EDGAR relies on morphologic anal- ysis of both languages involved and on the induc- tion of translation templates from the analyzed ref- erence translations. EDGAR decomposes the transla- tion text at several levels of generalization by match- ing it against translation examples contained in a case base. The matched chunks are then specified and re- fined in the target language. Among the three MT sys- tems under consideration, EDGAR uses the most gen- eralized representations and is thus expected to have the broadest coverage, while the STM uses the least generalized representations and is therefore expected to yield most precise translations. In order to test this hypothesis empirically, we have used two bilingually aligned translation corpora: a reference corpus and a test corpus. Each system was trained with a reference corpus containing 303 German-English translation examples as produced by a car manufacturer. The test corpus contains 265 translation examples from the same company and the same sub-domain (repair instructions). The size of the sentences ranges from 1 up to 160 characters in length containing single numbers and word transla- tions, short imperative sentences, noun phrases and whole sentences with subordinate clauses. The test corpus and the reference corpus are from the same domain, with similar vocabulary and similar phrase structure.