FBK@IWSLT 2007

FBK @ IWSLT 2007 N. Bertoldi, M. Cettolo, R. Cattoni, M. Federico FBK - Fondazione B. Kessler, Trento, Italy surname@itc.it Abstract This paper reports on the participation of FBK (for- merly ITC-irst) at the IWSLT 2007 Evaluation. FBK partic- ipated in three tasks, namely Chinese-to-English, Japanese- to-English, and Italian-to-English. With respect to last year, translation systems were developed with the Moses Toolkit and the IRSTLM library, both available as open source soft- ware. Moreover, several novel ideas were investigated: the use of confusion networks in input to manage ambiguity in punctuation, the estimation of an additional language model by means of the Google’s Web 1T 5-gram collection, the combination of true case and lower case language models, and finally the use of multiple phrase-tables. By working on top of a state-of-the art baseline, experiments showed that the above methods accounted for significant BLEU score im- provements. 1. Introduction This paper presents work carried out at FBK (formerly ITC- irst) to develop speech translation systems for three transla- tion tasks of the IWSLT 2007 Evaluation, namely Chinese- to-English (CE), Japanese-to-English (JE), and Italian-to- English (IE). All three systems are based on different set-ups of the same translation engine, namely the Moses statistical ma- chine translation (SMT) Toolkit [1].1 The FBK team joined the Moses open source project in 2006 and indeed discon- tinued the development of its own decoding software [2]. This paper is organized as follows. Section 2 gives a short introduction of the general architecture of the systems, which basically takes advantage of experience gained in the past IWSLT evaluations. Section 3 focuses on the novel aspects that were investigated specifically for IWSLT 2007, namely the management of punctuation and the use of additional lan- guage resources. Sections 4 to 6 discuss details related to the development and experimentation of each single translation system. Section 7 summarizes set-ups and official results of the primary and contrastive submissions. Section 8 provides some general discussion and conclusions about our partici- pation to IWSLT 2007. 1Available from http://www.statmt.org/moses 2. System description In its simplest configuration, the speech translation system provides the translation of the best transcription hypothesis generated by the ASR module. This version is used in the manual transcription input condition. Nevertheless, a tighter ASR-SMT integration [3] is considered as well. Given the word lattice produced by any ASR engine: (i) the word lat- tice is transformed into a compact structure, called confusion network (CN); (ii) punctuation is inserted in the CN; (iii) the optimal translation is computed from the CN; (iv) finally, if needed, case information is added to the translation. The insertion of punctuation (step ii), described in Sec- tion 3.1, is required whenever the final translation should be enriched with this information but it is not included in the input lattices. 2.1. Extraction of Confusion Network A word lattice contains several transcription alternatives ex- plored during the ASR process, but its topology is very com- plex and redundant. A simpler and more compact way of representing these alternatives …

Content truncated to 500 words