An Improved Statistical Transfer System for French-English Machine Translation

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 140–144, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics An Improved Statistical Transfer System for French–English Machine Translation Greg Hanneman, Vamshi Ambati, Jonathan H. Clark, Alok Parlikar, Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {ghannema, vamshi, jhclark, aup, alavie}@cs.cmu.edu Abstract This paper presents the Carnegie Mellon University statistical transfer MT system submitted to the 2009 WMT shared task in French-to-English translation. We de- scribe a syntax-based approach that incor- porates both syntactic and non-syntactic phrase pairs in addition to a syntactic grammar. After reporting development test results, we conduct a preliminary anal- ysis of the coverage and effectiveness of the system’s components. 1 Introduction The statistical transfer machine translation group at Carnegie Mellon University has been devel- oping a hybrid approach combining a traditional rule-based MT system and its linguistically ex- pressive formalism with more modern techniques of statistical data processing and search-based de- coding. The Stat-XFER framework (Lavie, 2008) provides a general environment for building new MT systems of this kind. For a given language pair or data condition, the framework depends on two main resources extracted from parallel data: a probabilistic bilingual lexicon, and a grammar of probabilistic synchronous context-free grammar rules. Additional monolingual data, in the form of an n-gram language model in the target language, is also used. The statistical transfer framework op- erates in two stages. First, the lexicon and gram- mar are applied to synchronously parse and trans- late an input sentence; all reordering is applied during this stage, driven by the syntactic grammar. Second, a monotonic decoder runs over the lat- tice of scored translation pieces produced during parsing and assembles the highest-scoring overall translation according to a log-linear feature model. Since our submission to last year’s Workshop on Machine Translation shared translation task (Hanneman et al., 2008), we have made numerous improvements and extensions to our resource ex- traction and processing methods, resulting in sig- nificantly improved translation scores. In Section 2 of this paper, we trace our current methods for data resource management for the Stat-XFER sub- mission to the 2009 WMT shared French–English translation task. Section 3 explains our tuning pro- cedure, and Section 4 gives our experimental re- sults on various development sets and offers some preliminary analysis. 2 System Construction Because of the additional data resources provided for the 2009 French–English task, our system this year is trained on nearly eight times as much data as last year’s. We used three officially pro- vided data sets to make up a parallel corpus for system training: version 4 of the Europarl cor- pus (1.43 million sentence pairs), the News Com- mentary corpus (0.06 million sentence pairs), and the pre-release version of the new Giga-FrEn cor- pus (8.60 million sentence pairs)1. The combined corpus of 10.09 million sentence pairs was pre- processed to remove blank lines, sentences of 80 words or more, and sentence pairs where the ra- tio …

Content truncated to 500 words