Statistical Transfer Systems for French-English and German-English Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 163–166, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Statistical Transfer Systems for French–English and German–English Machine Translation Greg Hanneman and Edmund Huber and Abhaya Agarwal and Vamshi Ambati and Alok Parlikar and Erik Peterson and Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {ghannema, ehuber, abhayaa, vamshi, aup, eepeter, alavie}@cs.cmu.edu Abstract We apply the Stat-XFER statistical transfer machine translation framework to the task of translating from French and German into En- glish. We introduce statistical methods within our framework that allow for the principled extraction of syntax-based transfer rules from parallel corpora given word alignments and constituency parses. Performance is evaluated on test sets from the 2007 WMT shared task. 1 Introduction The Carnegie Mellon University statistical trans- fer (Stat-XFER) framework is a general search- based and syntax-driven framework for develop- ing MT systems under a variety of data condi- tions (Lavie, 2008). At its core is a transfer en- gine using two language-pair-dependent resources: a grammar of weighted synchronous context-free rules (possibly augmented with unification-style fea- ture constraints), and a probabilistic bilingual lexi- con of syntax-based word- and phrase-level transla- tions. The Stat-XFER framework has been used to develop research MT systems for a number of lan- guage pairs, including Chinese–English, Hebrew– English, Urdu–English, and Hindi–English. In this paper, we describe our use of the frame- work to create new French–English and German– English MT systems for the 2008 Workshop on Sta- tistical Machine Translation shared translation task. We first describe the acquisition and processing of resources for each language pair and the roles of those resources within the Stat-XFER system (Sec- tion 2); we then report results on common test sets (Section 3) and share some early analysis and future directions (Section 4). 2 System Description Building a new machine translation system under the Stat-XFER framework involves constructing a bilingual translation lexicon and a transfer gram- mar. Over the past six months, we have developed new methods for extracting syntax-based translation lexicons and transfer rules fully automatically from parsed and word-aligned parallel corpora. These new methods are described in detail by Lavie et al. (2008). Below, we detail the statistical meth- ods by which these resources were extracted for our French–English and German–English systems. 2.1 Lexicon The bilingual lexicon is automatically extracted from automatically parsed and word-aligned paral- lel corpora. To obtain high-quality statistical word alignments, we run GIZA++ (Och and Ney, 2003) in both the source-to-target and target-to-source di- rections, then combine the resulting alignments with the Sym2 symmetric alignment heuristic of Ortiz- Mart´ınez et al. (2005)1. From this data, we extract a lexicon of both word-to-word and syntactic phrase- to-phrase translation equivalents. The word-level correspondences are extracted di- rectly from the word alignments: parts of speech for these lexical entries are obtained from the preter- 1We use Sym2 over more well-known heuristics such as “grow-diag-final” because Sym2 has been shown to give the best results for the node-alignment subtask …

Content truncated to 500 words