The TÜBÍTAK-UEKAE statistical machine translation system for IWSLT 2007

The TÜBİTAK-UEKAE Statistical Machine Translation System for IWSLT 2007 Coşkun Mermer, Hamza Kaya, Mehmet Uğur Doğan National Research Institute of Electronics and Cryptology (UEKAE) The Scientific and Technological Research Council of Turkey (TÜBİTAK) Gebze, Kocaeli 41470, Turkey {coskun,hamzaky,mugur}@uekae.tubitak.gov.tr Abstract We describe the TÜBITAK-UEKAE system that participated in the Arabic-to-English and Japanese-to- English translation tasks of the IWSLT 2007 evaluation campaign. Our system is built on the open-source phrase- based statistical machine translation software Moses. Among available corpora and linguistic resources, only the supplied training data and an Arabic morphological analyzer are used in the system. We present the run-time lexical approximation method to cope with out-of-vocabulary words during decoding. We tested our system under both automatic speech recognition (ASR) and clean transcript (clean) input conditions. Our system was ranked first in both Arabic-to- English and Japanese-to-English tasks under the “clean” condition. 1. Introduction Phrase-based statistical machine translation (SMT) has been the most actively researched machine translation approach in recent years. Since SMT is a corpus-based approach rather than relying on complicated linguistic rules, system development has been relatively easier and much less language-dependent. This has facilitated the emergence of publicly-available SMT development tools, which further lowered the entrance barrier to the research field. The International Workshop on Spoken Language Translation (IWSLT) series [1-3] have in addition provided common corpora and evaluation, creating an opportunity to comparatively evaluate machine translation technologies on the same test bed. This year’s workshop [4] further encouraged sharing of linguistic resources, emphasizing the cooperative nature of the workshop. In this paper, we report on our first-time participation in the IWSLT evaluation campaign. Among the four language tracks in IWSLT 2007, we participated in the Arabic-to- English and Japanese-to-English translation tasks. We built our baseline system using the open-source phrase-based statistical machine translation software Moses. Among shared corpora and tools, we used only the supplied training data and the Buckwalter Arabic Morphological Analyzer. In order to cope with previously unseen words during decoding, we present the run-time lexical approximation method, which replaces an out-of-vocabulary word with the closest known word having the same feature. We describe three different word-level features used in our system. For each language, both clean transcription and automatic speech recognition (ASR) outputs are translated, though only the 1- best hypotheses are used in the latter case. The rest of the paper is organized as follows. Section 2 describes the corpora and tools used in our system. Corpus preprocessing and system training are described in Sections 3 and 4, respectively. The run-time lexical approximation method is presented in Section 5. Results and discussion are provided in Section 6, followed by the conclusion. 2. Experimental setup and resources Our system is trained on the supplied BTEC corpus [5], whose characteristics are shown in Table 1. train and devsets1-3 are used for training, devset4 and devset5 are reserved for performance testing and tuning. Table 1: Supplied training corpus characteristics Corpus name Number of English references Number of source segments Avg. num. of tokens per segment (EN) …

Content truncated to 500 words