The University of Edinburgh system description for IWSLT 2007

The University of Edinburgh System Description for IWSLT 2007 Josh Schroeder, Philipp Koehn School of Informatics University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW Scotland, United Kingdom j.schroeder@ed.ac.uk, pkoehn@inf.ed.ac.uk Abstract We present the University of Edinburgh’s submission for the IWSLT 2007 shared task. Our efforts focused on adapting our statistical machine translation system to the open data conditions for the Italian-English task of the evaluation cam- paign. We examine the challenges of building a system with a limited set of in-domain development data (SITAL), a small training corpus in a related but distinct domain (BTEC), and a large out of domain corpus (Europarl). We concentrated on the corrected text track, and present additional results of our experiments using the open-source Moses MT system with speech input. 1. Introduction The IWSLT 2007 shared task offered the University of Ed- inburgh a chance to expand our experience with spoken lan- guage translation and with translation using data from multi- ple corpora. We focused on the Italian-English challenge task because it offered a chance to explore spontaneous speech as well as an opportunity to use a corpus we are familiar with, Europarl[1], as an additional data resource. In this paper we first present a summary of the phrase- based statistical machine translation system used for this shared task. We go on to discuss the data sources we used to train the system, and show the results of our analysis of the domains of the development test data sets compared to the corpora used for training. Having explained the framework and data used, in Sec- tion 5 we present the results of our experiments in cross- domain adaptation. In Section 6 we describe the experiments we conducted with ASR inputs to our system. 2. Framework 2.1. The Moses MT system The open source Moses[2] MT system was originally devel- oped at the University of Edinburgh. It received a major boost through a 2006 Johns Hopkins workshop, and is now used at several academic institutions as the basic infrastruc- ture for statistical machine translation research. The Moses system is an implementation of the phrase- based machine translation approach[3]. In this approach, an input sentence is first split into text chunks (so-called phrases), which are then mapped one-to-one to target phrases using a large phrase translation table. Phrases may be re- ordered, but typically a maximum movement reordering limit is used. Phrase translation probabilities, reordering probabilities and language model probabilities are combined to give each possible sentence translation a score. The best-scoring trans- lation is searched for by the decoding algorithm and out- putted by the system as the best translation. The different system components hi (phrase translation probabilities, lan- guage model, etc.) are combined in a log-linear model to obtain the score for the translation e of an input sentence f: score(e, f) = exp � i λi hi(e, f) (1) The weights of the components λi are set by minimum error rate training on held-out development data[4]. The ba- sic components used in …

Content truncated to 500 words