Statistical Machine Translation with a Factorized Grammar

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 616–625, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Statistical Machine Translation with a Factorized Grammar Libin Shen and Bing Zhang and Spyros Matsoukas and Jinxi Xu and Ralph Weischedel Raytheon BBN Technologies Cambridge, MA 02138, USA {lshen,bzhang,smatsouk,jxu,weisched}@bbn.com Abstract In modern machine translation practice, a sta- tistical phrasal or hierarchical translation sys- tem usually relies on a huge set of trans- lation rules extracted from bi-lingual train- ing data. This approach not only results in space and efficiency issues, but also suffers from the sparse data problem. In this paper, we propose to use factorized grammars, an idea widely accepted in the field of linguis- tic grammar construction, to generalize trans- lation rules, so as to solve these two prob- lems. We designed a method to take advantage of the XTAG English Grammar to facilitate the extraction of factorized rules. We experi- mented on various setups of low-resource lan- guage translation, and showed consistent sig- nificant improvement in BLEU over state-of- the-art string-to-dependency baseline systems with 200K words of bi-lingual training data. 1 Introduction A statistical phrasal (Koehn et al., 2003; Och and Ney, 2004) or hierarchical (Chiang, 2005; Marcu et al., 2006) machine translation system usually re- lies on a very large set of translation rules extracted from bi-lingual training data with heuristic methods on word alignment results. According to our own experience, we obtain about 200GB of rules from training data of about 50M words on each side. This immediately becomes an engineering challenge on space and search efficiency. A common practice to circumvent this problem is to filter the rules based on development sets in the step of rule extraction or before the decoding phrase, instead of building a real distributed system. How- ever, this strategy only works for research systems, for which the segments for translation are always fixed. However, do we really need such a large rule set to represent information from the training data of much smaller size? Linguists in the grammar con- struction field already showed us a perfect solution to a similar problem. The answer is to use a fac- torized grammar. Linguists decompose lexicalized linguistic structures into two parts, (unlexicalized) templates and lexical items. Templates are further organized into families. Each family is associated with a set of lexical items which can be used to lex- icalize all the templates in this family. For example, the XTAG English Grammar (XTAG-Group, 2001), a hand-crafted grammar based on the Tree Adjoin- ing Grammar (TAG) (Joshi and Schabes, 1997) for- malism, is a grammar of this kind, which employs factorization with LTAG e-tree templates and lexical items. Factorized grammars not only relieve the burden on space and search, but also alleviate the sparse data problem, especially for low-resource language translation with few training data. With a factored model, we do not need to observe exact “template – lexical item” occurrences in training. New rules can be generated from …

Content truncated to 500 words