European Language Translation with Weighted Finite State Transducers: The CUED MT System for the 2008 ACL Workshop on SMT

Proceedings of the Third Workshop on Statistical Machine Translation, pages 131–134, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics European Language Translation with Weighted Finite State Transducers: The CUED MT System for the 2008 ACL Workshop on SMT Graeme Blackwood, Adri`a de Gispert, Jamie Brunning, William Byrne Machine Intelligence Laboratory Department of Engineering, Cambridge University Trumpington Street, Cambridge, CB2 1PZ, U.K. {gwb24|ad465|jjjb2|wjb31}@cam.ac.uk Abstract We describe the Cambridge University En- gineering Department phrase-based statisti- cal machine translation system for Spanish- English and French-English translation in the ACL 2008 Third Workshop on Statistical Ma- chine Translation Shared Task. The CUED system follows a generative model of trans- lation and is implemented by composition of component models realised as Weighted Fi- nite State Transducers, without the use of a special-purpose decoder. Details of system tuning for both Europarl and News translation tasks are provided. 1 Introduction The Cambridge University Engineering Department statistical machine translation system follows the Transducer Translation Model (Kumar and Byrne, 2005; Kumar et al., 2006), a phrase-based generative model of translation that applies a series of transfor- mations specified by conditional probability distri- butions and encoded as Weighted Finite State Trans- ducers (Mohri et al., 2002). The main advantages of this approach are its mod- ularity, which facilitates the development and eval- uation of each component individually, and its im- plementation simplicity which allows us to focus on modeling issues rather than complex decoding and search algorithms. In addition, no special-purpose decoder is required since standard WFST operations can be used to obtain the 1-best translation or a lat- tice of alternative hypotheses. Finally, the system architecture readily extends to speech translation, in which input ASR lattices can be translated in the same way as for text (Mathias and Byrne, 2006). This paper reviews the first participation of CUED in the ACL Workshop on Statistical Machine Trans- lation in 2008. It is organised as follows. Firstly, section 2 describes the system architecture and its main components. Section 3 gives details of the de- velopment work conducted for this shared task and results are reported and discussed in section 4. Fi- nally, in section 5 we summarise our participation in the task and outline directions for future work. 2 The Transducer Translation Model Under the Transducer Translation Model, the gen- eration of a target language sentence tJ 1 starts with the generation of a source language sentence sI 1 by the source language model PG(sI 1). Next, the source language sentence is segmented into phrases accord- ing to the unweighted uniform phrasal segmenta- tion model PW (uK 1 , K|sI 1). This source phrase se- quence generates a reordered target language phrase sequence according to the phrase translation and re- ordering model PR(xK 1 |uK 1 ). Next, target language phrases are inserted into this sequence according to the insertion model PΦ(vR 1 |xK 1 , uK 1 ). Finally, the sequence of reordered and inserted target language phrases are transformed to word sequences tJ 1 under the target …

Content truncated to 500 words