An Automatic Reviser: The TransCheck System

An Automatic Reviser: The TransCheck System Jean-Marc Jutras RALI, Universit6 de Montr6al C.P. 6128, succ. Centre-ville, Montr6al (QC), Canada jutras@iro.umontreal.ca Abstract Over the past decade or so, a lot of work in computational linguistics has been directed at finding ways to exploit the ever increasing volume of electronic bilingual corpora. These efforts have allowed for substantial expansion of the computational toolbox. We describe a system, TransCheck, which makes intensive use of these new tools in order to detect potential translation errors in preliminary or non-revised translations. Introduction For the sake of argument, let's consider a translator to be a black box with source text in and target text out. We feed that box with texts and, to be really tricky, we input the same text a couple of times. Looking at the results, the first thing we notice is that though the different translations are quite similar, they're not exactly the same. Nothing to worry about, this may simply exemplify the potential for synonymy and paraphrase. But let's further suppose the text to translate is too big for one individual to translate in the given time frame. In realistic conditions, such a text would be split among perhaps half a dozen translators, each with his own vocabulary, experience and stylistic preferences, which would normally lead to the well known problem of non-uniformity of the translation. It is therefore part of the normal translation process to have a reviser look at a translator's output. His job will be to spot any typos (taken in a very broad sense to include missing chapters!). Usually, at this point the translator probably has submitted the preliminary version to a spell checker, so what could be done automatically at that level has already been done. No automatic detection of typical translation mistakes has been attempted though. That's the gap TransCheck is designed to fill. The concept of a "translation checker" was initially proposed in Isabelle and al. [8] and eventually led to a demonstration prototype concerned with the detection of a very restricted type of mistake: deceptive cognates. In comparison, the system described in this paper goes much further toward a "real " usable translation checker by allowing for the detection of errors of omission, the comparison of diverse numerical expressions and the flagging of inconsistent terminology. On the interface side, it allows for the automatic alignment of the source and target texts, the flagging of potential mistakes and the possibility of saving any modifications made to the target text. I Error detection Complete automatic modelling of the translation process is still far beyond our technical ability. The same is true of our ability to detect all types of translation mistakes. We can however, for certain well- defined sub-types of mistake, devise specific mechanisms. And if a program capable of detecting all mistakes of translation would undoubtedly be extremely useful, so would 127 one capable of detecting frequent mistakes, especially when time is short and a thorough revision isn't possible. Errors are then bound to …

Content truncated to 500 words