Combining Unsupervised and Supervised Alignments for MT: An Empirical Study

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 667–673, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Combining Unsupervised and Supervised Alignments for MT: An Empirical Study Jinxi Xu and Antti-Veikko I. Rosti Raytheon BBN Technologies, 10 Moulton Street, Cambridge, MA 02138, USA {jxu,arosti}@bbn.com Abstract Word alignment plays a central role in statisti- cal MT (SMT) since almost all SMT systems extract translation rules from word aligned parallel training data. While most SMT systems use unsupervised algorithms (e.g. GIZA++) for training word alignment, super- vised methods, which exploit a small amount of human-aligned data, have become increas- ingly popular recently. This work empirically studies the performance of these two classes of alignment algorithms and explores strate- gies to combine them to improve overall sys- tem performance. We used two unsupervised aligners, GIZA++ and HMM, and one super- vised aligner, ITG, in this study. To avoid lan- guage and genre specific conclusions, we ran experiments on test sets consisting of two lan- guage pairs (Chinese-to-English and Arabic- to-English) and two genres (newswire and we- blog). Results show that the two classes of al- gorithms achieve the same level of MT perfor- mance. Modest improvements were achieved by taking the union of the translation gram- mars extracted from different alignments. Sig- nificant improvements (around 1.0 in BLEU) were achieved by combining outputs of differ- ent systems trained with different alignments. The improvements are consistent across lan- guages and genres. 1 Introduction Word alignment plays a central role in training sta- tistical machine translation (SMT) systems since al- most all SMT systems extract translation rules from word aligned parallel training data. Until recently, most SMT systems used GIZA++ (Och and Ney, 2003), an unsupervised algorithm, for aligning par- allel training data. In recent years, with the availabil- ity of human aligned training data, supervised meth- ods (e.g. the ITG aligner (Haghighi et al., 2009)) have become increasingly popular. The main objective of this work is to show the two classes (unsupervised and supervised) of al- gorithms are complementary and combining them will improve overall system performance. The use of human aligned training data allows supervised methods such as ITG to more accurately align fre- quent words, such as the alignments of Chinese par- ticles (e.g. “bei”, “de”, etc) to their English equiv- alents (e.g. “is/are/was/..”, “of”, etc). On the other hand, supervised methods can be affected by sub- optimal alignments in hand-aligned data. For exam- ple, the hand-aligned data used in our experiments contain some coarse-grained alignments (e.g. “lian- he guo” to “United Nations”) although fine-grained alignments (“lian-he” to “United” and “guo” to “Na- tions”) are usually more appropriate for SMT. Un- supervised methods are less likely to be affected by this problem. We used two well studied unsu- pervised aligners, GIZA++ (Och and Ney, 2003) and HMM (Liang et al., 2006) and one supervised aligner, ITG (Haghighi et al., 2009) as representa- tives in this work. We explored two techniques to combine different alignment …

Content truncated to 500 words