International Workshop on Spoken Language Translation (IWSLT) 2010

Paris, France
December 2-3, 2010

The QMUL System Description for IWSLT 2010

Sirvan Yahyaei (1), Christof Monz (2)

(1) School of Electronic Engineering and Computer Science, Queen Mary University of London, UK
(2) ISLA, Informatics Institute, University of Amsterdam, The Netherlands

The QMUL submission to IWSLT 2010 is a phrase-based statistical MT system. A multi-stack, multi-beam decoder with several features, with weights tuned on the provided development data through Minimum Error Rate Training (MERT) algorithm. This year QMUL participated in Arabic- English, French-English and Turkish-English language pairs of the BTEC task.
   A discriminative reordering model is added as a feature to improve the reordering capabilities of the decoder. In addition, an algorithm is devised to determine the best distortion limit for each hypothesis expansion. Improvements in quality were also gained by different means in different stages of the training and decoding.

Full Paper

Bibliographic reference.  Yahyaei, Sirvan / Monz, Christof (2010): "The QMUL system description for IWSLT 2010", In IWSLT-2010, 157-162.