Automated Alignment in Multilingual Corpora

AUTOMATED ALIGNMENT IN MULTILINGUAL CORPORA J.A. Campbell Dept. of Computer Science, University College London, Gower Street, London WC1E 6BT, England (jac@cs.ucl.ac.u1c) and Alex Chengyu Fang Survey of English Usage, University College London, Gower Street, London WC1E 6BT, England (ucleacf@ucl.ac.uk) Abstract Experiences in computing alignments at the paragraph and sentence level within a project TRANSLEARN in the European Union's "LRE" programme of research and development in language engineering are reported. About 98% of the sentences in pairs of corpora in different languages have been aligned correctly by a method that uses dynamic programming on numbers of characters per sentence. This paral- lels the experience of previous researchers for English-French alignment. We have used Portuguese and Greek material in addition to these languages, from a set of 49 European Union official documents. It is argued that the key issue of automated alignment is now the automated improvement of the quality of alignment achieved by methods that rely only on character counts. Cues that are helpful to support such an improvement are identified special words, cognates, syntactic fragments, and a simple measure of seman- tic weight. A short account of their use in experiments is given. Introduction and Motivation There are many possible uses for multilingual corpora of parallel texts where alignments are indicated between corresponding parts of the textual material in different languages. In Europe, where the number of official languages of the European Union (EU) has recently risen to 11 and where documents recording formal acts of the EU have to be produced in all of them, there is no more pressing need than the need to give automated support to the human translators who are required to do this work. The project TRANSLEARN within the EU's LRE programme of research and develop- ment in language engineering addresses an important part of the needs of EU translators. Because of the repetition of various structures and formal details in EU documents, it is quite possible that a translator who is making a version of a document in language Y from a source in language X will be generating text that is very close to something in Y that is already part of the EU's files or archives. Differences may be just in inessential details like dates, serial numbers, and names of institutions or countries. In this situation, the translator will clearly be more cost-effective if it is possible to find and display candidate pieces of text from the files that are 185 very close to the language-Y version of the current piece of text in X that the translator is pro- cessing. (Here, "piece" refers typically to a sentence or a paragraph). Rather than having to spend significant amounts of time on tackling a complete translation from X to Y, the translator should then be able to produce the desired material much more quickly by localised editing of the retrieved text in Y. Alignment is always likely to be an important foundation-stone for any method that auto- mates this form of assistance …

Content truncated to 500 words