Supervised Bilingual Lexicon Induction with Multiple Monolingual Signals

Proceedings of NAACL-HLT 2013, pages 518–523, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Supervised Bilingual Lexicon Induction with Multiple Monolingual Signals Ann Irvine Center for Language and Speech Processing Johns Hopkins University Chris Callison-Burch⇤ Computer and Information Science Dept. University of Pennsylvania Abstract Prior research into learning translations from source and target language monolingual texts has treated the task as an unsupervised learn- ing problem. Although many techniques take advantage of a seed bilingual lexicon, this work is the first to use that data for super- vised learning to combine a diverse set of sig- nals derived from a pair of monolingual cor- pora into a single discriminative model. Even in a low resource machine translation setting, where induced translations have the potential to improve performance substantially, it is rea- sonable to assume access to some amount of data to perform this kind of optimization. Our work shows that only a few hundred transla- tion pairs are needed to achieve strong per- formance on the bilingual lexicon induction task, and our approach yields an average rel- ative gain in accuracy of nearly 50% over an unsupervised baseline. Large gains in accu- racy hold for all 22 languages (low and high resource) that we investigate. 1 Introduction Bilingual lexicon induction is the task of identifying word translation pairs using source and target mono- lingual corpora, which are often comparable. Most approaches to the task are based on the idea that words that are translations of one another have sim- ilar distributional properties across languages. Prior research has shown that contextual similarity (Rapp, 1995), temporal similarity (Schafer and Yarowsky, 2002), and topical information (Mimno et al., 2009) ⇤Performed while faculty at Johns Hopkins University are all good signals for learning translations from monolingual texts. Most prior work either makes use of only one or two monolingual signals or uses unsupervised meth- ods (like rank combination) to aggregate orthogonal signals (Schafer and Yarowsky, 2002; Klementiev and Roth, 2006). Surprisingly, no past research has employed supervised approaches to combine diverse monolingually-derived signals for bilingual lexicon induction. The field of machine learning has shown decisively that supervised models dramatically out- perform unsupervised models, including for closely related problems like statistical machine translation (Och and Ney, 2002). For the bilingual lexicon induction task, a super- vised approach is natural, particularly because com- puting contextual similarity typically requires a seed bilingual dictionary (Rapp, 1995), and that same dictionary may be used for estimating the param- eters of a model to combine monolingual signals. Alternatively, in a low resource machine transla- tion (MT) setting, it is reasonable to assume a small amount of parallel data from which a bilingual dic- tionary can be extracted for supervision. In this set- ting, bilingual lexicon induction is critical for trans- lating source words which do not appear in the par- allel data or dictionary. We frame bilingual lexicon induction as a binary classification problem; for a pair of source and tar- get language words, we predict whether the …

Content truncated to 500 words