The CMU Statistical Machine Translation System for IWSLT2005

The CMU Statistical Machine Translation System for IWSLT 2005 Sanjika Hewavitharana, Bing Zhao, Almut Silja Hildebrand, Matthias Eck, Chiori Hori, Stephan Vogel and Alex Waibel Interactive Systems Laboratories, Language Technologies Institute Carnegie Mellon University, Pittsburgh, USA [sanjika, bzhao, silja, matteck, chiori, vogel+, ahw]@cs.cmu.edu Abstract In this paper we describe the CMU statistical machine translation system used in the IWSLT 2005 evaluation campaign. This system is based on phrase-to-phrase translations extracted from a bilingual corpus. We ex- perimented with two different phrase extraction methods; PESA on-the-fly phrase extraction and alignment free ex- traction method. The translation model, language model and other features were combined in a log-linear model during decoding. We present our experiments on model adaptation for new data in a different domain, as well as combining different translation hypotheses to obtain bet- ter translations. We participated in the supplied data track for man- ual transcriptions in the translation directions: Arabic- English, Chinese-English, Japanese-English and Korean- English. For Chinese-English direction we also worked on ASR output of the supplied data, and with additional data in unrestricted and C-STAR tracks. 1. Introduction Large vocabulary text translation has been the primary fo- cus in machine translation research during the past. Much improvements have been achieved with projects such as TIDES, which focused on large vocabulary text transla- tion. With the availability of reliable speech recognition systems and spoken language corpora, now the focus is shifting towards speech translation; and further towards speech-to-speech translation. With the IBM system [1] in early 90’s, statistical machine translation (SMT) has been the most promis- ing approach for machine translation. Many approaches for SMT have been proposed since then [2], [3], [4]. Whereas the original IBM system was based on purely word translation models, current SMT systems incorpo- rate more sophisticated models. The CMU statistical machine translation system uses phrase-to-phrase translations as the primary building blocks to capture local context information, leading to better lexical choice and more reliable local reorder- ing. In section 2, we describe the phrase alignment ap- proaches used by our system. The main obstacle in using additional data for a trans- lation task is that the new data may belong to a different domain. We explored methods of adapting both the trans- lation model and the language model to overcome this problem, which are described in section 3. Section 4 outlines the architecture of the decoder that combines the translation model, language model, and other models to generate the complete translation. When translating speech recognition output, we inte- grate multiple translation hypotheses into a single struc- ture and then derive the best hypothesis. This approach is described in section 5. Finally, in section 6 we give an overview of the data and tasks and present the results of the experiments we carried out for different data conditions. 2. Phrase Alignment In this evaluation, we applied a variation of the alignment-free approach, which is an extension to the previous work in [5] and [6] to extract bilingual phrase pairs for the supplied data …

Content truncated to 500 words