The NICT/ATR speech translation system for IWSLT 2007

The NICT/ATR Speech Translation System for IWSLT 2007 Andrew Finch, Etienne Denoual, Hideo Okuma, Michael Paul, Hirofumi Yamamoto, Keiji Yasuda, Ruiqiang Zhang, Eiichiro Sumita ATR Spoken Language Communication Research Labs Hikaridai 2-2-2, Keihanna Science City, 619-0288 Kyoto Andrew.Finch@atr.jp Abstract This paper describes the NiCT-ATR statistical machine trans- lation (SMT) system used for the IWSLT 2007 evaluation campaign. We participated in three of the four language pair translation tasks (CE, JE, and IE). We used a phrase-based SMT system using log-linear feature models for all tracks. This year we decoded from the ASR n-best lists in the JE track and found a gain in performance. We also applied some new techniques to facilitate the use of out-of-domain external resources by model combination and also by uti- lizing a huge corpus of n-grams provided by Google Inc.. Using these resources gave mixed results that depended on the technique also the language pair however, in some cases we achieved consistently positive results. The results from model-interpolation in particular were very promising. 1. Introduction Phrase-based statistical machine translation approaches con- tinue to dominate the field of machine translation. All of the systems for each of the languages pairs we submitted results for differ in important respects from other systems, however they are all based around a fairly typical phrased- based machine translation system built within the framework of a feature-based exponential model containing the follow- ing features: • Phrase translation probability from source to target • Inverse phrase translation probability • Lexical weighting probability from source to target • Inverse lexical weighting probability • Phrase penalty • Language model probability • Simple distance-based distortion model • Word penalty The basic framework within which all the systems were constructed is shown in Figure 1, and the corresponding overview of the translation process is shown in Figure 1. With the exception of the experiments where factored models were required, the decoder used for the training and decod- ing of the test data was the in-house multi-stack phrase-based decoder CleopATRa. This paper is constructed as follows, firstly for each lan- guage pair we provide a description of the components of the system that are specific for that language pair. This descrip- tion includes details of any segmentation, external resources and any specific modelling techniques that were employed. Next, we detail those parts of the process that are common to all systems. In particular we focus on the techniques we used to utilize a large corpus of word n-grams provided by Google Inc. The next sections present and discuss our exper- imental results, and finally we conclude and propose avenues for future research. 2. Japanese-English 2.1. Corpora In addition to the supplied corpus, we also drew on resources from the following corpora: • The Tanaka corpus (203K sentence pairs) • The Yomiuri News corpus (202K sentence pairs) • The SLDB corpus (72K sentence pairs) • The Chinese Olympic corpus included in the Chinese- LDC (Code: 2004-863-009) (104K sentence pairs) 2.2. Pre-processing The data was segmented using the publicly …

Content truncated to 500 words