Source Error-Projection for Sample Selection in Phrase-Based SMT for Resource-Poor Languages

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 819–827, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Source Error-Projection for Sample Selection in Phrase-Based SMT for Resource-Poor Languages Sankaranarayanan Ananthakrishnan, Shiv Vitaladevuni, Rohit Prasad, and Prem Natarajan Raytheon BBN Technologies 10 Moulton Street Cambridge, MA 02138, U.S.A. {sanantha,svitalad,rprasad,pnataraj}@bbn.com Abstract The unavailability of parallel training cor- pora in resource-poor languages is a ma- jor bottleneck in cost-effective and rapid deployment of statistical machine transla- tion (SMT) technology. This has spurred significant interest in active learning for SMT to select the most informative sam- ples from a large candidate pool. This is especially challenging when irrelevant outliers dominate the pool. We propose two supervised sample selection methods, viz. greedy selection and integer lin- ear programming (ILP), based on a novel measure of benefit derived from error anal- ysis. These methods support the selec- tion of diverse and high-impact, yet rel- evant batches of source sentences. Com- parative experiments on multiple test sets across two resource-poor language pairs (English-Pashto and English-Dari) reveal that the proposed approaches achieve BLEU scores comparable to the full sys- tem using a very small fraction of all avail- able training data (ca. 6% for E-P and 13% for E-D). We further demonstrate that the ILP method supports global constraints of significant practical value. 1 Introduction The laborious and time-consuming nature of pro- ducing parallel training corpora for the develop- ment of high-quality SMT systems cannot be over- stated. Barring a few mainstream languages, the vast majority of language pairs can be classified This paper is based upon work supported by the DARPA TRANSTAC Program. The views expressed are those of the author and do not reflect the official policy or position of the Department of Defense or the U.S. Government. as “resource-poor” as far as availability of a usable SMT system is concerned. Active learning can re- duce human labor, turn-around time and monetary cost of developing SMT systems with little or no loss in translation accuracy. In its simplest form, active learning for building parallel corpora involves selecting “high-value” samples from a large monolingual corpus of source sentences (the candidate pool) for transla- tion by a bilingual human expert. The notion of “value” depends on the selection method, and can be derived using unsupervised, semi-supervised, or supervised techniques. For instance, Eck et al. (2005) define high-value source sentences as those that contain a large number of previously unseen n-grams. While it aims to increase coverage of the training set, the main deficiency of this approach is its tendency to pick irrelevant outliers if the can- didate pool contains data from unrelated regimes. Haffari et al. (2009) propose a number of fea- tures, such as similarity to the seed corpus, trans- lation probability, n-gram and phrase coverage as unsupervised measures of the value of candidate samples. Additionally, a linear combination of these features is proposed as a supervised measure of value for ranking candidate sentences. The pa- rameters of this model are …

Content truncated to 500 words