Dependency-Based Decipherment for Resource-Limited Machine Translation

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1668–1676, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Dependency-Based Decipherment for Resource-Limited Machine Translation Qing Dou and Kevin Knight Information Sciences Institute Department of Computer Science University of Southern California {qdou,knight}@isi.edu Abstract We introduce dependency relations into deci- phering foreign languages and show that de- pendency relations help improve the state-of- the-art deciphering accuracy by over 500%. We learn a translation lexicon from large amounts of genuinely non parallel data with decipherment to improve a phrase-based ma- chine translation system trained with limited parallel data. In experiments, we observe BLEU gains of 1.2 to 1.8 across three different test sets. 1 Introduction State-of-the-art machine translation (MT) systems apply statistical techniques to learn translation rules from large amounts of parallel data. However, par- allel data is limited for many language pairs and do- mains. In general, it is easier to obtain non parallel data. The ability to build a machine translation system using monolingual data could alleviate problems caused by insufficient parallel data. Towards build- ing a machine translation system without a paral- lel corpus, Klementiev et al. (2012) use non paral- lel data to estimate parameters for a large scale MT system. Other work tries to learn full MT systems using only non parallel data through decipherment (Ravi and Knight, 2011; Ravi, 2013). However, the performance of such systems is poor compared with those trained with parallel data. Given that we often have some parallel data, it is more practical to improve a translation sys- tem trained on parallel corpora with non parallel Figure 1: Improving machine translation with deci- pherment (Grey boxes represent new data and process). Mono: monolingual; LM: language model; LEX: trans- lation lexicon; TM: translation model. data. Dou and Knight (2012) successfully apply decipherment to learn a domain specific translation lexicon from monolingual data to improve out-of- domain machine translation. Although their ap- proach works well for Spanish/French, they do not show whether their approach works for other lan- guage pairs. Moreover, the non parallel data used in their experiments is created from a parallel corpus. Such highly comparable data is difficult to obtain in reality. In this work, we improve previous work by Dou and Knight (2012) using genuinely non parallel data, 1668 and propose a framework to improve a machine translation system trained with a small amount of parallel data. As shown in Figure 1, we use a lexi- con learned from decipherment to improve transla- tions of both observed and out-of-vocabulary (OOV) words. The main contributions of this work are: • We extract bigrams based on dependency re- lations for decipherment, which improves the state-of-the-art deciphering accuracy by over 500%. • We demonstrate how to improve translations of words observed in parallel data by us- ing a translation lexicon obtained from large amounts of non parallel data. • We show that decipherment is able to find cor- rect translations for OOV words. • We use a …

Content truncated to 500 words