The CMU-UKA statistical machine translation systems for IWSLT 2007

The CMU-UKA Statistical Machine Translation Systems for IWSLT 2007 Ian Lane∗, Andreas Zollmann∗, Thuy Linh Nguyen∗, Nguyen Bach∗, Ashish Venugopal∗, Stephan Vogel∗, Kay Rottmann†, Ying Zhang∗ and Alex Waibel†∗ InterACT Research Laboratories: ∗Carnegie Mellon University, Pittsburgh, USA †University of Karlsruhe, Karlsruhe, Germany Abstract This paper describes the CMU-UKA statistical machine translation systems submitted to the IWSLT 2007 eval- uation campaign. Systems were submitted for three language-pairs: Japanese→English, Chinese→English and Arabic→English. All systems were based on a common phrase-based SMT (statistical machine translation) frame- work but for each language-pair a specific research prob- lem was tackled. For Japanese→English we focused on two problems: first, punctuation recovery, and second, how to incorporate topic-knowledge into the translation frame- work. Our Chinese→English submission focused on syntax- augmented SMT and for the Arabic→English task we fo- cused on incorporating morphological-decomposition into the SMT framework. This research strategy enabled us to evaluate a wide variety of approaches which proved effective for the language pairs they were evaluated on. 1. Introduction For the IWSLT 2007 evaluation campaign we focused on applying systems developed based on current re- search topics within our lab. This includes our work on syntax-augmented SMT (statistical machine transla- tion), morphological-decomposition for rich morphology languages, such as Arabic, and approaches to better cou- ple ASR (automatic speech recognition) and MT (machine translation) for spoken language translation. Three sys- tems were submitted for evaluation: Japanese→English, Chinese→English and Arabic→English, and all built upon our 2006 submissions for these tasks. This year we fo- cused on applying a number of distinct approaches for each language-pair. This enabled us to cover many research topics within spoken language translation. In Section 3, we describe our J→E submission sys- tem and compare various approaches to recover punctuation (Section 3.3). In Section 3.4 we investigate methods to incor- porate topic-knowledge into the translation framework via N-best list re-scoring. Our syntax-augmented SMT frame- work is detailed in Section 4. Sections 4.5 and 4.6 describe its application to the C→E task. Our A→E translation sys- tem which incorporates morphological decomposition is in- troduced in Section 5. 2. The CMU-UKA Phrase-based SMT System Our J→E and A→E systems built upon the STTK (Statisti- cal Translation Toolkit) framework used for our IWSLT 2006 submissions [1]. STTK implements phrase-based statistical machine translation using a log-linear model [2] in which a foreign language sentence f J 1 = f1, f2, . . . , fJ is translated into another language eI 1 = e1, e2, . . . , eI by searching for the hypothesis ˆeI 1 with maximum likelihood, given: ˆeI 1 = arg max eI 1 P(eI 1|f J 1 ) = arg max eI 1 exp ��M m=1 λmhm � eI 1, f J 1 �� � e′I′ 1 exp ��M m=1 λmhm � e′I′ 1 , f J 1 �� In this framework, the posterior probability P(eI 1|f J 1 ) is directly maximized using a log-linear combination of feature functions hm � eI 1, f J 1 � (during decoding the denominator …

Content truncated to 500 words