I2R Chinese-English translation system for IWSLT 2007

I2R Chinese-English Translation System for IWSLT 2007 Boxing Chen, Jun Sun, Hongfei Jiang, Min Zhang, Ai Ti Aw Department of Human Language Technology Institute for Infocomm Research, Singapore {bxchen, visjs, vishfj, mzhang, aaiti}@i2r.a-star.edu.sg Abstract In this paper, we describe the system and approach used by Institute for Infocomm Research (I2R) for the IWSLT 2007 spoken language evaluation campaign. A multi-pass approach is exploited to generate and select best translation. First, we use two decoders namely the open source Moses and an in-home syntax-based decoder to generate N-best lists. Next we spawn new translation entries through a word-based n-gram language model estimated on the former N-best entries. Finally, we join the N-best lists from the previous two passes, and select the best translation by rescoring them with additional feature functions. In particular, this paper reports our effort on new translation entry generation and system combination. The performance on development and test sets are reported. The system was ranked first with respect to the BLEU measure in Chinese-to-English open data track. 1. Introduction This paper describes the statistical machine translation (SMT) system and approach undertaken by the Institute for Infocomm Research (I2R) for the International Workshop on Spoken Language Translation (IWSLT) 2007. We submitted runs under the open data conditions for Chinese-to-English task. A typical state-of-the-art SMT system applies two-pass search strategy [1, 2]. In the first pass, N-best translations are generated by a decoding algorithm; while in the second pass, the best translation is computed by rescoring and reranking the N-best translations with additional feature functions. In our MT system, we introduce an intermediate (re-generation) pass before rescoring. The N-best lists generated separately by different decoders are consolidated first before further expanded to create a new list by applying a generative n-gram language model, estimated on the joined N-best lists [3]. In our experiment, three N-best lists are generated in the first pass. One is provided by an in-home syntax-based decoder. The other two are generated by Moses decoder [4] with different data preprocessing. This paper is organized as follows. Section 2 presents the structure of our system with emphasis on the syntax-based decoder. We also describe the n-gram expansion, system combination and rescoring modules. Section 3 reports the experimental setups and results with discusses on the results obtained while Section 4 concludes the paper. 2. System Description Figure 1 depicts our system architecture. A source input sentence is preprocessed with two different settings. It is then passed to the phrase-based decoder (Moses) and the syntax-based decoder to produce three N-best translation lists. All the three N-best lists go through the n-gram expansion to create new translation hypotheses, and formed the fourth N-best list. All four N-best lists are then combined, and the best translation is selected by rescoring approach. We will detail the following components namely phrase-based system, syntax-based system, n-gram expansion, system combination and rescoring. 2.1. Phrase-based system Phrase-based statistical machine translation systems are usually modeled through a log-linear framework [5, 6] by introducing the hidden word alignment variable …

Content truncated to 500 words