Nobody is perfect: ATR’s hybrid approach to spoken language translation

Nobody is Perfect: ATR’s Hybrid Approach to Spoken Language Translation Michael Paul, Takao Doi, Youngsook Hwang, Kenji Imamura, Hideo Okuma, Eiichiro Sumita ATR Spoken Language Communication Research Laboratories Hikaridai 2-2-2, Keihanna Science City, 619-0288 Kyoto {Michael.Paul,Takao.Doi,Youngsook.Hwang,Kenji.Imamura,Hideo.Okuma,Eiichiro.Sumita}@atr.jp Abstract This paper describes ATR’s hybrid approach to spoken lan- guage translation and it’s application to the IWSLT 2005 translation task. Multiple corpus-based translation engines are used to translate the same input, whereby the best trans- lation among the element MT outputs is selected according to statistical models. The evaluation results of the Japanese-to-English and Chinese-to-English translation tasks for different training data conditions showed the potential of the proposed hybrid approach and revealed new directions in how to improve the current system performance. 1. Introduction Corpus-based approaches to machine translation (MT) have achieved much progress over the last decades. There are two main strategies used in corpus-based translation: 1. Example-Based Machine Translation (EBMT) [1]: EBMT uses the corpus directly. EBMT retrieves the translation examples that are best matched to an input expression and then adjusts the examples to obtain the translation. 2. Statistical Machine Translation (SMT) [2]: SMT learns statistical models for translation from cor- pora and dictionaries and then searches for the best translation at run-time according to the statistical mod- els for language and translation. Despite a high performance on average, these approaches can often produce translations with severe errors. However, dif- ferent MT engines not always do the same error. Due to the particular output characteristics of each MT engine, quite dif- ferent translation hypotheses are produced. Thus, combining multiple MT systems can boost the system performance by exploiting the strengths of each MT engine. We propose a corpus-based approach that uses multiple translation engines in parallel. All engines translate the same input, whereby the best translation among the multiple MT outputs is selected according to multiple statistical language and translation models. The outline of our hybrid approach is given in Section 2. The proposed system was applied to two translation directions (Japanese-to-English, Chinese-to-English) and three data tracks (Supplied Data Track, Supplied+Tools Data Track, C-STAR Track). The evaluation of the obtained results is given in Section 3. 2. System Description We use an architecture in which multiple EBMT and SMT engines work in parallel and their outputs are passed to a post-process that selects the best candidate according to SMT models (cf. Figure 1). This section is structured as follows: (1) eight corpus- based MT engines are introduced in Section 2.1; (2) the SMT-based approach to select the best translation out of mul- tiple hypotheses is explained in Section 2.2; (3) the resources utilized for the IWSLT 2005 translation task are described in Section 2.3; and (4) a summary of the data tracks we partici- pated in and an overview on which MT engines were utilized for the respective tracks is given in Section 2.4. 2.1. MT Engines We employed the following four SMT and four EBMT sys- tems: An SMT engine that uses an example-based decod- ing method [SAT] (cf. …

Content truncated to 500 words