A Semi-Supervised Batch-Mode Active Learning Strategy for Improved Statistical Machine Translation

Proceedings of the Fourteenth Conference on Computational Natural Language Learning, pages 126–134, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics A Semi-Supervised Batch-Mode Active Learning Strategy for Improved Statistical Machine Translation Sankaranarayanan Ananthakrishnan, Rohit Prasad, David Stallard and Prem Natarajan BBN Technologies 10 Moulton Street Cambridge, MA, U.S.A. {sanantha,rprasad,stallard,prem}@bbn.com Abstract The availability of substantial, in-domain parallel corpora is critical for the develop- ment of high-performance statistical ma- chine translation (SMT) systems. Such corpora, however, are expensive to pro- duce due to the labor intensive nature of manual translation. We propose to al- leviate this problem with a novel, semi- supervised, batch-mode active learning strategy that attempts to maximize in- domain coverage by selecting sentences, which represent a balance between domain match, translation difficulty, and batch di- versity. Simulation experiments on an English-to-Pashto translation task show that the proposed strategy not only outper- forms the random selection baseline, but also traditional active learning techniques based on dissimilarity to existing training data. Our approach achieves a relative im- provement of 45.9% in BLEU over the seed baseline, while the closest competitor gained only 24.8% with the same number of selected sentences. 1 Introduction Rapid development of statistical machine transla- tion (SMT) systems for resource-poor language pairs is a problem of significant interest to the research community in academia, industry, and government. Tight turn-around schedules, bud- get restrictions, and scarcity of human translators preclude the production of large parallel corpora, which form the backbone of SMT systems. Given these constraints, the focus is on making the best possible use of available resources. This usually involves some form of prioritized data col- lection. In other words, one would like to con- struct the smallest possible parallel training corpus that achieves a desired level of performance on un- seen test data. Within an active learning framework, this can be cast as a data selection problem. The goal is to choose, for manual translation, the most infor- mative instances from a large pool of source lan- guage sentences. The resulting sentence pairs, in combination with any existing in-domain seed par- allel corpus, are expected to provide a significantly higher performance gain than a na¨ıve random se- lection strategy. This process is repeated until a certain level of performance is attained. Previous work on active learning for SMT has focused on unsupervised dissimilarity measures for sentence selection. Eck et al. (2005) describe a selection strategy that attempts to maximize cov- erage by choosing sentences with the highest pro- portion of previously unseen n-grams. However, if the pool is not completely in-domain, this strat- egy may select irrelevant sentences, whose trans- lations are unlikely to improve performance on an in-domain test set. They also propose a technique, based on TF-IDF, to de-emphasize sentences sim- ilar to those that have already been selected. How- ever, this strategy is bootstrapped by random ini- tial choices that do not necessarily favor sentences that are difficult to translate. Finally, they work exclusively with the source language and do not use …

Content truncated to 500 words