Formalizing translation memories

MT Summit VII Sept. 1999 Formalizing Translation Memories Emmanuel Planas, Osamu Furuse NTT Cyber Solutions Laboratories 2-4 Hikaridai Seika-cho Soraku-gun, Kyoto 619-0237, Japan {planas, furuse} @soy.keel.ntt.co.jp Abstract The TELA structure, a set of layered and linked lattices, and the notion of Similarity between TELA structures, based on the Edit Distance, are introduced in order to formalize Translation Memories (TM). We show how this approach leads to a real gain in recall and precision, and allows extending TM towards rudimentary, yet useful Example-Based Machine Translation that we call Shallow Translation. 1 Introduction This paper describes a pioneer step towards a theory for Translation Memories (TM). We do not pretend to give "the" theoretical approach, but giving at least one arises interesting issues we propose to address now. The use of TM involves two phases. The first consists in accumulating translation units (TU). A TU is simply a source sentence (source TU) and its corresponding trans- lation, also called the target sentence (target TU), as the human translator has translated it. In the second phase, an input source sentence (INPUT) being given, the TM re- trieves the more similar source TU and proposes the cor- responding target TU as a close translation of the input sentence INPUT. The first issue we want to address is how to represent sentences in the TM. Does a linear string of characters suffice? By the way, are only characters involved in the representation of sentences extracted from current editing tools like Adobe Frame Maker, Interleaf of Microsoft Word? The first section, Structures, proposes answers to this issue. Another important question is: what does the more simi- lar sentence mean? Is it the number of different charac- ters between INPUT and source TU? In this case, in the example below, is sentence (0) closest to sentence (1) that has only four different letters, or to sentence (2) that has nine different letters? (0) "The wild child is destroying his new toy." (1) "The wild chief is destroying his new tool." (2) "The wild children are destroying their new toy." Furthermore, how to cope with non literal data like index marks and layout attributes that are inevitably found in common nowadays edited documents? The second sec- tion, Similarity, deals with these points. We have built a prototype implementing our proposals for the structures and the similarity issues. Section 4 shows that this leads to a real improvement of TM effi- ciency. This section also introduces the notion of "Shal- low Translation" as a natural extension of TM, and we give some examples of our prototype responses. 2 Structures 2.1 Real documents code Here is the display of a typical sentence, as it can be seen in classical word editors: Sentence (3a): as displayed word editors. Please suppose also that there is an index mark on "color". In the typical internal code used in word processors for- mats (Microsoft RTF™ in the example below), it would look like this: Sentence (3b): as coded in word editors. Current TM tools1 …

Content truncated to 500 words