Inversion Transduction Grammar Constraints for Mining Parallel Sentences from Quasi-Comparable Corpora

Inversion Transduction Grammar Constraints for Mining Parallel Sentences from Quasi-Comparable Corpora⋆ Dekai Wu1 and Pascale Fung2 1 Human Language Technology Center, HKUST, Department of Computer Science 2 Department of Electrical and Electronic Engineering, University of Science and Technology, Clear Water Bay, Hong Kong dekai@cs.ust.hk, pascale@ee.ust.hk Abstract. We present a new implication of Wu’s (1997) Inversion Transduction Grammar (ITG) Hypothesis, on the problem of retriev- ing truly parallel sentence translations from large collections of highly non-parallel documents. Our approach leverages a strong language uni- versal constraint posited by the ITG Hypothesis, that can serve as a strong inductive bias for various language learning problems, resulting in both efficiency and accuracy gains. The task we attack is highly prac- tical since non-parallel multilingual data exists in far greater quantities than parallel corpora, but parallel sentences are a much more useful re- source. Our aim here is to mine truly parallel sentences, as opposed to comparable sentence pairs or loose translations as in most previous work. The method we introduce exploits Bracketing ITGs to produce the first known results for this problem. Experiments show that it obtains large accuracy gains on this task compared to the expected performance of state-of-the-art models that were developed for the less stringent task of mining comparable sentence pairs. 1 Introduction Parallel sentences are a relatively scarce but extremely useful resource for many applications including cross-lingual retrieval and statistical machine translation. Parallel sentences, or bi-sentences for short, can be exploited for a wealth of applications ranging from mining term translations for cross-lingual applications, to training paraphrase models and inducing structured terms for indexing, query processing, and retrieval. Unfortunately, far more is available in the way of monolingual data. High- quality parallel corpora are currently largely limited to specialized collections of government (especially UN) and certain newswire collections, and even then relatively few bi-sentences are available in tight sentence-by-sentence translation. ⋆ This work was supported in part by the Hong Kong Research Grants Council through grants RGC6083/99E, RGC6256/00E, DAG03/04.EG09, and RGC6206/03E. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 257–268, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 258 D. Wu and P. Fung Increasingly sophisticated methods for extracting loose translations from non- parallel monolingual corpora—and in particular, what have been called com- parable sentence pairs—have also recently become available. But while loose translations by themselves already have numerous applications, truly parallel sentence translations provide invaluable types of information for the aforemen- tioned types of mining and induction, which cannot easily be obtained from merely loose translations or comparable sentence pairs. In particular, truly par- allel bi-sentences are especially useful for extracting more precise syntactic and semantic relations within word sequences. We present a new method that exploits a novel application of Inversion Transduction Grammar or ITG expressiveness constraints (Wu 1995 [1], Wu 1997 [2]) for mining monolingual data to obtain tight sentence translation pairs, yielding accuracy significantly higher than previous known methods. We focus here on very non-parallel quasi-comparable monolingual corpora, which are avail- able in far …

Content truncated to 500 words