CMU Syntax-Based Machine Translation at WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 365–371, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics CMU Syntax-Based Machine Translation at WMT 2011 Greg Hanneman and Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {ghannema, alavie}@cs.cmu.edu Abstract We present the Carnegie Mellon University Stat-XFER group submission to the WMT 2011 shared translation task. We built a hy- brid syntactic MT system for French–English using the Joshua decoder and an automati- cally acquired SCFG. New work for this year includes training data selection and grammar filtering. Expanded training data selection significantly increased translation scores and lowered OOV rates, while results on grammar filtering were mixed. 1 Introduction During the past year, the statistical transfer ma- chine translation group at Carnegie Mellon Univer- sity has continued its work on large-scale syntactic MT systems based on automatically acquired syn- chronous context-free grammars (SCFGs). For the 2011 Workshop on Machine Translation, we built a hybrid MT system, including both syntactic and non-syntactic rules, and submitted it as a constrained entry to the French–English translation task. This is our fourth yearly submission to the WMT shared translation task. In design and construction, the system is sim- ilar to our submission from last year’s workshop (Hanneman et al., 2010), with changes in the meth- ods we employed for training data selection and SCFG filtering. Continuing WMT’s general trend, we worked with more data than in previous years, basing our 2011 system on 13.9 million sentences of parallel French–English training data and an En- glish language model of 1.8 billion words. Decod- ing was carried out in Joshua (Li et al., 2009), an open-source framework for parsing-based MT. We managed our experiments with LoonyBin (Clark and Lavie, 2010), an open-source tool for defining, mod- ifying, and running complex experimental pipelines. We describe our system-building process in more detail in Section 2. In Section 3, we evaluate the sys- tem’s performance on WMT development sets and examine the aftermath of training data selection and grammar filtering. Section 4 concludes with possi- ble directions for future work. 2 System Construction 2.1 Training Data Selection WMT 2011’s provided French–English training data consisted of 36.8 million sentence pairs from the Eu- roparl, news commentary, UN documents, and Giga- FrEn corpora (Table 1). The first three of these are, for the most part, clean data resources that have been successfully employed as MT corpora for a number of years. The Giga-FrEn corpus, though the largest, is also the least precise, as its Web-crawled data sources are less homogeneous and less structured than the other corpora. Nevertheless, Pino et al. (2010) found significant improvements in French– English MT output quality by including it. Our goal for this year was to strike a middle ground: to avoid computational difficulties in using the entire 36.8 million sentence pairs of training data, but to mine the Giga-FrEn corpus for sentences to increase our system’s vocabulary coverage. Our method of training data selection proceeded as follows. We …

Content truncated to 500 words