Using Bi-textual Alignment for Translation Validation: the TransCheck system

Using Bi-textual Alignment for Translation Validation: the TransCheck System Elliott Macklovitch1 Centre for Information Technology Innovation ABSTRACT We describe the first prototype version of TransCheck, a system for automatically detecting certain types of translation errors that is based on the notion of bi-text, or aligned corpora of translated texts. We analyse the preliminary results obtained from applying TransCheck to five lengthy samples of published translations, discuss some of the problems that currently lie beyond the system's scope, and conclude by briefly sketching our plans for extending the first TransCheck prototype. 1. Introduction Translators, like all people who draft a lot of texts, will commonly run a spelling or style checker on their translations before delivering them, in the hope of detecting simple kinds of errors they may have overlooked. Because these writing aids are monolingual, however, they are entirely useless in detecting even the most flagrant kinds of translation errors, since these are bilingual in nature and depend on relations between two texts in different languages. This paper will describe a novel kind of writing aid that is specifically designed to detect translation errors, errors like those reproduced in Table 1 at the end of the paper. All the examples listed in Table 1 are authentic: they are taken from the Canadian Hansards, and so the errors they contain were actually committed by translators who are among the most qualified in Canada.2 The errors in Table 1 were detected automatically by a translation checking system called TransCheck. TransCheck is one of a family of new translation support tools being developed at the CITI (see reference [1]), all of which are based on bi-textual alignment algorithms like those originally proposed by Gale and Church [2] and Brown et al. [3]. What these alignment algorithms do, in a word, is to calculate, with a high degree of accuracy, the links between corresponding segments in a source text and its translation, regardless of the size of the two texts.3 Following Harris [4], a corpus of translated texts aligned in this way is often referred to as a bi-text. In the references just cited, the aligned segments of a bi-text are generally taken to be sentences, but in principle linkages between larger or smaller units are also possible; and, of course, one sentence in the source can be explicitly linked with two or more sentences in the target, and vice versa. 1. Please see the acknowledgments at the end of this paper. 2. In their defence, it must be said that these translators work under enormous pressure to produce the previous day's parliamentary debates in time for publication the next morning. Under these conditions, and given the volume of texts we scanned, what is really surprising is the relative paucity of the errors we detected in the Hansards. 3. For details on the CITI's current alignment algorithm, which has a success rate of over 98%, see Simard et al. [7]. 157 In a number of recent CITI publications (including [1], [5] & [6]), it …

Content truncated to 500 words