The CMU-Avenue French-English Translation System

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 261–266, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics The CMU-Avenue French-English Translation System Michael Denkowski Greg Hanneman Language Technologies InstituteAlon Lavie Carnegie Mellon University Pittsburgh, PA, 15213, USA {mdenkows,ghannema,alavie}@cs.cmu.edu Abstract This paper describes the French-English trans- lation system developed by the Avenue re- search group at Carnegie Mellon University for the Seventh Workshop on Statistical Ma- chine Translation (NAACL WMT12). We present a method for training data selection, a description of our hierarchical phrase-based translation system, and a discussion of the im- pact of data size on best practice for system building. 1 Introduction We describe the French-English translation sys- tem constructed by the Avenue research group at Carnegie Mellon University for the shared trans- lation task in the Seventh Workshop on Statistical Machine Translation. The core translation system uses the hierarchical phrase-based model described by Chiang (2007) with sentence-level grammars ex- tracted and scored using the methods described by Lopez (2008). Improved techniques for data selec- tion and monolingual text processing significantly improve the performance of the baseline system. Over half of all parallel data for the French- English track is provided by the Giga-FrEn cor- pus (Callison-Burch et al., 2009). Assembled from crawls of bilingual websites, this corpus is known to be noisy, containing sentences that are either not par- allel or not natural language. Rather than simply in- cluding or excluding the resource in its entirety, we use a relatively simple technique inspired by work in machine translation quality estimation to select the best portions of the corpus for inclusion in our train- ing data. Including around 60% of the Giga-FrEn chosen by this technique yields an improvement of 0.7 BLEU. Prior to model estimation, we process all parallel and monolingual data using in-house tokenization and normalization scripts that detect word bound- aries better than the provided WMT12 scripts. After translation, we apply a monolingual rule-based post- processing step to correct obvious errors and make sentences more acceptable to human judges. The post-processing step alone yields an improvement of 0.3 BLEU to the final system. We conclude with a discussion of the impact of data size on important decisions for system building. Experimental results show that “best practice” deci- sions for smaller data sizes do not necessarily carry over to systems built with “WMT-scale” data, and provide some explanation for why this is the case. 2 Training Data Training data provided for the French-English trans- lation task includes parallel corpora taken from Eu- ropean Parliamentary proceedings (Koehn, 2005), news commentary, and United Nations documents. Together, these sets total approximately 13 million sentences. In addition, a large, web-crawled parallel corpus termed the “Giga-FrEn” (Callison-Burch et al., 2009) is made available. While this corpus con- tains over 22 million parallel sentences, it is inher- ently noisy. Many parallel sentences crawled from the web are neither parallel nor sentences. To make use of this large data source, we employ data se- lection techniques discussed in the next …

Content truncated to 500 words