A pipelined multi-engine approach to Chinese-to-Korean machine translation: MATES/CK

MT Summit VII Sept. 1999 A Pipelined Multi-Engine Approach to Chinese-to-Korean Machine Translation: MATES/CK Min Zhang Key-Sun Choi KORTERM, CS Department, Korean KORTERM, CS Department, Korean Advanced Institute of Science and advanced Institute of Science and Technology, 373-1 Kusong-Dong Technology, 373-1 Kusong-Dong Yusong-ku Taejon 305-701 Korea Yusong-ku Taejon 305-701 Korea Abstract This paper presents MATES/CK, a Chinese- to-Korean machine translation system. We introduce the design philosophy, component modules, implementation and some other aspects of MATES/CK system in this paper. 1 Introduction A Chinese-to-Korean Machine Translation system MATES/CK has been developed as a research prototype and is still under upgrading now in KAIST (Korea Advanced Institute of Science and Technology). Up to now, though many different approaches (Choi et al. 1994; Su et al. 1995; Brown et al. 1996; Frederking et al. 1994) to MT have been advocated, it is generally agreed that no approach, whether rule-based, example-based, pattern-based or statistics-based, is completely adequate in all aspects to the machine translation task. In order to integrate the advantages of these approaches and get rid of their disadvantages in designing a hybrid MT system, we propose a new hybrid pipelined multi-engine approach to MT and apply this approach to our MATES/CK system. We introduce the design philosophy, component modules, implementation and some other aspects of MATES/CK system in this paper. 2 Design Philosophy The core idea of MATES/CK system is “pipelined multi-engine”. Each MT engine employs a different MT technology. When using the pipelined multi- engine MT approach, an MT task is divided into many sub-problems and we start up an engine to resolve the corresponding sub-problem that is most suitable for being resolved by the most appropriate engine. According to Frederking et al.'s definition (Frederking et al. 1994), multi-engine machine translation (MEMT) feeds an input text to several MT engines in parallel. But MATES/CK employs different engines serially, not in parallel. So we terms our proposed approach as a pipelined multi-engine approach to distinguish it from Frederking et al.'s definition (Frederking et al. 1994). The pipelined multi-engine MT model here also follows the typical three-phase scheme (analysis/transfer/synthesis) of a conventional transfer-based system. Rule-based Engine The rule-based engine is mainly used in the post- processing of Chinese morphological analysis and the pruning processing in the syntactical analysis stage (Zhang & Choi 1999). To improve the robustness of the rule-based engine, we propose a linguistic attribute knowledge classification method and a new attribute-pruning algorithm (Zhang 1997; Zhang & Choi 1999). Statistics-based Engine We use it in POS tagging, best syntactic tree selection, mapping pattern extraction, and lexical translation. A new probabilistic model was proposed and adopted to select the best syntactic tree from the syntactic tree candidate set (Zhang & Choi 1999). A new lexical selection algorithm was proposed by using Viterbi algorithm and some statistical knowledge (Zhang & Choi 1999). Pattern-based Engine Pattern-based engine is used in structural transfer. Our patterns are extracted from examples semi- automatically. We proposed a parameterized pattern- based transfer approach (Zhang & Choi 1999). 3 Translation Flow …

Content truncated to 500 words