Aligning Clattses in Parallel Texts

Aligning Clauses in Parallel Texts Sotii·is Boutsis and Stelios Piperidis Institute for Language and Speech Processing - ILSP Artemidos & Epidavrou 151-25, Athens, Greece tel:+301 6800959, fax:+301 6854270 email: {sboutsis, spip}@ilsp.gr National Technical University of Athens - NTUA Abstract This paper describes a method for the automatic align- ment of parallel texts at clause level. The method fea- tures statistical techniques coupled with shallow linguis- tic processing. It presupposes a parallel bilingual corpus and identifies alignments between the clauses of the source and target language sides of the corpus. Parallel texts are first statistically aligned at sentence level and then tagged with their part-of-speech categories. Regular grammars functioning on tags, recognize clauses on both sides of the parallel text. A probabilistic model is ap- plied next, operating on the basis of word occurrence and co-occurrence probabilities and character lengths. De- pending on sentence size, possible alignments arc fed into a dynamic progranuning framework or a simulated annealing system in order to find or approxim~te the best alignment. 1he method has been tested on a Small Eng~ lish-Greek corpus consisting of texts relevant to software systems and has produced promising results in terms of correctly identified clause alignments. Introduction The availability of large collections of texts in electronic fom1, has given rise to a wide range of applications aim~ ing at the elicitation of linguistic resources such as tTanslation dictionaries, transfer grammars and retTieval of translation examples (Dagan et al., 1991; Matsumoto et al., 1993), or even the building of fully-blown machine translation systems (Brown et al., 1990). The pmpose of this paper is to describe a technique for extracting trans- lation correspondences at bellow sentence level by em- ploying statistical techniques coupled with shallow lin- guistic processing catering for the segmentation of sen- tences into clauses. Statistical processing has proved powerful for the exh·action of translation equivalences at sentence and intra-sentence level. Brown et al. (1991) described a method based on the number of words contained in sen- tences. The general idea is that the closer in length two 17 sentences are, the most likely they are to align. Moreo- ver, certain anchor points and paragraph markers are considered. Dynamic progra111111ing and HMMs are pipe- lined to produce alignments at sentence level. The method has been applied to the Hansard-Corpus, achiev- ing an accuracy of 96%-97%. Gale and Church ( 1991) proposed a method that relies on a simple statistical model of character lengths. The model is based on the observation that longer sentences in one language tend to be translated into longer sentences in the other language while shorter ones tend to be translated into shorter ones. A probabilistic score is assigned to each pair of proposed sentence pairs, and a dynamic programming framework calculates the most probable alignment. Although the apparent efficacy of the Gale-Church algorithm is unde- niable and validated on different pairs of languages, rt faces problems when handling complex alignments(l-0, 1-2, 2-2). 'Simard et al. (1992) argue that a small amount of lin- guistic information …

Content truncated to 500 words