The CUED HiFST System for the WMT10 Translation Shared Task

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 155–160, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The CUED HiFST System for the WMT10 Translation Shared Task Juan Pino Gonzalo Iglesias‡1 Adri`a de Gispert Graeme Blackwood Jamie Brunning William Byrne Department of Engineering, University of Cambridge, Cambridge, CB2 1PZ, U.K. {jmp84,gi212,ad465,gwb24,jjjb2,wjb31}@eng.cam.ac.uk ‡ Department of Signal Processing and Communications, University of Vigo, Vigo, Spain Abstract This paper describes the Cambridge Uni- versity Engineering Department submis- sion to the Fifth Workshop on Statistical Machine Translation. We report results for the French-English and Spanish-English shared translation tasks in both directions. The CUED system is based on HiFST, a hierarchical phrase-based decoder imple- mented using weighted finite-state trans- ducers. In the French-English task, we investigate the use of context-dependent alignment models. We also show that lattice minimum Bayes-risk decoding is an effective framework for multi-source translation, leading to large gains in BLEU score. 1 Introduction This paper describes the Cambridge University Engineering Department (CUED) system submis- sion to the ACL 2010 Fifth Workshop on Statis- tical Machine Translation (WMT10). Our trans- lation system is HiFST (Iglesias et al., 2009a), a hierarchical phrase-based decoder that generates translation lattices directly. Decoding is guided by a CYK parser based on a synchronous context- free grammar induced from automatic word align- ments (Chiang, 2007). The decoder is imple- mented with Weighted Finite State Transducers (WFSTs) using standard operations available in the OpenFst libraries (Allauzen et al., 2007). The use of WFSTs allows fast and efficient exploration of a vast translation search space, avoiding search errors in decoding. It also allows better integration with other steps in our translation pipeline such as 5-gram language model (LM) rescoring and lattice minimum Bayes-risk (LMBR) decoding. 1Now a member of the Department of Engineering, Uni- versity of Cambridge, Cambridge, CB2 1PZ, U.K. # Sentences # Tokens # Types (A)Europarl+News-Commentary FR 1.7 M 52.4M 139.7k EN 47.6M 121.6k (B)Europarl+News-Commentary+UN FR 8.7 M 277.9M 421.0k EN 241.4M 482.1k (C)Europarl+News-Commentary+UN+Giga FR 30.2 M 962.4M 2.4M EN 815.3M 2.7M Table 1: Parallel data sets used for French-to- English experiments. We participated in the French-English and Spanish-English translation shared tasks in each translation direction. This paper describes the de- velopment of these systems. Additionally, we re- port multi-source translation experiments that lead to very large gains in BLEU score. The paper is organised as follows. Section 2 describes each step in the development of our sys- tem for submission, from pre-processing to post- processing. Section 3 presents and discusses re- sults and Section 4 describes an additional experi- ment on multi-source translation. 2 System Development We built three French-English and two Spanish- English systems, trained on different portions of the parallel data sets available for this shared task. Statistics for the different parallel sets are sum- marised in Tables 1 and 2. No additional parallel data was used. As will be shown, the largest paral- lel corpus gave the best results in French, but this was not the case in …

Content truncated to 500 words