Discriminative Sample Selection for Statistical Machine Translation

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 626–635, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Discriminative Sample Selection for Statistical Machine Translation∗ Sankaranarayanan Ananthakrishnan, Rohit Prasad, David Stallard, and Prem Natarajan Raytheon BBN Technologies 10 Moulton Street Cambridge, MA, U.S.A. {sanantha,rprasad,stallard,prem}@bbn.com Abstract Production of parallel training corpora for the development of statistical machine translation (SMT) systems for resource-poor languages usually requires extensive manual effort. Ac- tive sample selection aims to reduce the la- bor, time, and expense incurred in produc- ing such resources, attaining a given perfor- mance benchmark with the smallest possible training corpus by choosing informative, non- redundant source sentences from an available candidate pool for manual translation. We present a novel, discriminative sample selec- tion strategy that preferentially selects batches of candidate sentences with constructs that lead to erroneous translations on a held-out de- velopment set. The proposed strategy supports a built-in diversity mechanism that reduces redundancy in the selected batches. Simu- lation experiments on English-to-Pashto and Spanish-to-English translation tasks demon- strate the superiority of the proposed approach to a number of competing techniques, such as random selection, dissimilarity-based se- lection, as well as a recently proposed semi- supervised active learning strategy. 1 Introduction Resource-poor language pairs present a significant challenge to the development of statistical machine translation (SMT) systems due to the latter’s depen- dence on large parallel texts for training. Bilingual human experts capable of producing the requisite ∗Distribution Statement “A” (Approved for Public Release, Distribution Unlimited) data resources are often in short supply, and the task of preparing high-quality parallel corpora is labori- ous and expensive. In light of these constraints, an attractive strategy is to construct the smallest pos- sible parallel training corpus with which a desired performance benchmark may be achieved. Such a corpus may be constructed by selecting the most informative instances from a large collection of source sentences for translation by a human ex- pert, a technique often referred to as active learn- ing. A SMT system trained with sentence pairs thus generated is expected to perform significantly better than if the source sentences were chosen using, say, a na¨ıve random sampling strategy. Previously, Eck et al. (2005) described a selec- tion strategy that attempts to maximize coverage by choosing sentences with the highest proportion of previously unseen n-grams. Depending on the com- position of the candidate pool with respect to the domain, this strategy may select irrelevant outliers. They also described a technique based on TF-IDF to de-emphasize sentences similar to those that have al- ready been selected, thereby encouraging diversity. However, this strategy is bootstrapped by random initial choices that do not necessarily favor sentences that are difficult to translate. Finally, they worked exclusively with the source language and did not use any SMT-derived features to guide selection. Haffari et al. (2009) proposed a number of fea- tures, such as similarity to the seed corpus, transla- tion probability, n-gram and phrase coverage, etc., that drive data selection. They …

Content truncated to 500 words