Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation

Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation Philipp Koehn, Amittai Axelrod, Alexandra Birch Mayne, Chris Callison-Burch, Miles Osborne, David Talbot pkoehn@inf.ed.ac.uk, amittai@mit.edu a.c.birch-mayne@sms.ed.ac.uk, chris@linearb.co.uk miles@inf.ed.ac.uk, d.r.talbot@sms.ed.ac.uk School of Informatics University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, Scotland, UK Abstract Our participation in the IWSLT 2005 speech translation task is our first effort to work on limited domain speech data. We adapted our statistical machine trans- lation system that performed successfully in previous DARPA competitions on open domain text translations. We participated in the supplied corpora transcription track. We achieved the highest BLEU score in 2 out of 5 language pairs and had competi- tive results for the other language pairs. 1 Introduction The statistical machine translation group at the Uni- versity of Edinburgh has been focused on open do- main text translation, so we welcomed the challenge to work on the IWSLT 2005 limited domain speech translation task. We participated in the transcription translation tasks for all five language pairs, using only the supplied corpora. Our MT system was originally developed for translation of European parliament texts from Ger- man to English (Koehn et al., 2003). We extended the system while working on the DARPA challenges to translate Chinese and Arabic news texts into En- glish (Koehn, 2004a; Koehn et al., 2005). Now, we were faced with the challenge of speech data in mostly Asian languages. The translation of transcribed speech differs in many ways from our traditional translation scenario: Much less training data is available, the domain is more limited, and the text style is very different — short questions and statements. In some respect, the task is easier, since smaller training corpora result in faster training times for the system. But it also meant that we had to re-examine various components of our system. In this paper we present an overview of our cur- rent out-of-the-box system in the next section. It in- cludes a more detailed treatment of models added over the last year, especially a novel lexicalised re- ordering model. Experimental work went into the adaptation of our system to the IWSLT’05 translation tasks. This is described in Section 3. We used a Linux cluster of about 50 machines, which allowed extensive optimi- sation of key components of our system, especially word alignment, lexicalised reordering, and reorder- ing limits. Finally, we report on our results in the competi- tion and some post-evaluation analysis. 2 System Description The system employs a phrase-based statistical ma- chine translation model (Koehn et al., 2003) that uses the Pharaoh decoder (Koehn, 2004b). In this section, we will give an overview of the system. 2.1 Phrase-Based Statistical MT In phrase-based SMT models, the input (foreign) sentence is segmented into so-called phrases, which may be any sequences of adjacent words that do not have to be linguistically motivated. Each phrase is mapped into the target language (English). Phrases may be reordered. See Figure 1 for an illustration. Figure 1: Phrase-based SMT: Input is segmented into phrases, each is …

Content truncated to 500 words