TS3: an Improved Version of the Bilingual Concordancer TransSearch

Proceedings of the 13th Annual Conference of the EAMT, pages 20–27, Barcelona, May 2009 TS3: an Improved Version of the Bilingual Concordancer TransSearch St´ephane Huet, Julien Bourdaillet and Philippe Langlais DIRO - Universit´e de Montr´eal C.P. 6128, succursale Centre-ville H3C 3J7, Montr´eal, Qu´ebec, Canada {huetstep,bourdaij,felipe}@iro.umontreal.ca Abstract Computer Assisted Translation tools re- main the preferred solution of human translators when publication quality is of concern. In this paper, we present our ongoing efforts conducted within TS3, a project which aims at improv- ing the commercial bilingual concordancer TransSearch. The core technology of this Web-based service mainly relies on sentence-level alignment. In this study, we discuss and evaluate the embedding of sta- tistical word-level alignment. 1 Introduction Although the last decade witnessed an impressive amount of effort devoted to improving the current state of Machine Translation (MT), professional translators still prefer Computer Assisted Transla- tion (CAT) tools, among which translation mem- ory (TM) systems and bilingual concordancers. Both tools exploit a TM composed of a bitext: a set of pairs of units (typically sentences) that are in translation relation. Whereas a TM system is a translation device, a bilingual concordancer is conceptually simpler, since its main purpose is to retrieve from a bitext, the pairs of units that con- tain a query (typically a phrase) that a user man- ually submits. It is then left to the user to locate the relevant material in the retrieved target units. As simple as it may appear, a bilingual concor- dancer is nevertheless a very popular CAT tool. In (Macklovitch et al., 2008), the authors report that TransSearch,1 the commercial web-based con- cordancer we focus on in this study, received an av- c⃝ 2009 European Association for Machine Translation. 1www.tsrali.com erage of 177 000 queries a month over a one-year period (2006–2007). Figure 1 provides a screenshot of a session with the current concordancer TransSearch. A user submitted the multi-word query in keeping with to which the system responded with a web- page showing the first 25 pairs of sentences in the TM that contain an occurrence of the query. As can be observed, nothing in the target material re- trieved is emphasized, which forces the user to read the examples retrieved until an appropriate translation was found. Figure 1: Screenshot of TransSearch. Two of the first 25 matches returned to the user for the query in keeping with. The main objective of the TS3 project is to auto- matically identify (highlight) in the retrieved mate- rial the different translations of a user query. Iden- tifying translations offers interesting prospects for user-efficient interactions. Although the definitive look-and-feel of the new prototype is not settled yet, Figure 2 shows an interface where the user can consult the most likely translations automatically 20 identified. Of course, she can still consult the pairs of sentences containing the query, but can as well click a given translation to see related matches. Figure 2: A hypothetical interface which exploits translation spotting. The reminder of this paper is organized as fol- …

Content truncated to 500 words