How to Overcome the Domain Barriers in Pattern-Based Machine Translation System

How to Overcome the Domain Barriers in Pattern-Based Machine Translation System* Sung-Kwon Choi a, Ki-Young Lee a, Yoon-Hyung Roh a, Oh-Woog Kwon a, and Young-Gil Kim a aNatural Language Processing Team, Electronics and Telecommunications Research Institute, 161 Gajeong-dong, Youseong-gu, Daejon, Korea {choisk, leeky, yhroh, ohwoog, kimyk}@etri.re.kr Abstract. One of difficult issues in pattern-based machine translation system is maybe to find how to overcome the domain difference in adapting a system from one domain to other domain. This paper describes how we have resolved such barriers among domains as default target word of any domain, domain-specific patterns, and domain adaptation of engine modules in pattern-based machine translation system, especially English-Korean pattern-based machine translation system. For this, we will discuss two types of customization methods which mean a method adapting an existing system to new domain. One is the pure customization method introduced for patent machine translation system in 2006 and another is the upgraded customization method applied to scientific paper machine translation system in 2007. By introducing an upgraded customization method, we could implement a practical machine translation system for scientific paper translation within 8 months, in comparison with the patent machine translation system that was completed even in 24 months by the pure customization method. The translation accuracy of scientific paper machine translation system also rose 77.25% to 81.10% in spite of short term of 8 months. Keywords: Customization, Machine Translation, Pattern-based MT, Scientific Paper MT. 1. Introduction The use of on-line systems is the biggest growth area in the use of machine translation. People are translating web pages or very large documents by using machine translation system as the solution, as human translation of pages which need to be continually updated or are very large scale is not feasible (Mellebeek et.al., 2005). Electronics and Telecommunications Research Institute (ETRI, henceforth) in Korea has developed the web-based English-Korean machine translation system till 2004, under assumption that as the size of patterns grows, the performance of the system can be incrementally improved (Hong et. al., 2003). During 2 years (2005- 2006) it implemented an English-Korean patent machine translation system on the basis of the web-based English- Korean machine translation system. The English-Korean patent machine translation system was installed in International Patent Assistance Center (IPAC, henceforth) under Ministry of * This work was supported by the IT R&D program of MKE/IITA, Domain Customization Machine Translation Technology Development for Korean, Chinese, and English. Copyright 2008 by Sung-Kwon Choi, Ki-Young Lee, Yoon-Hyung Roh, Oh-Woog Kwon, and Young- Gil Kim 161 22nd Pacific Asia Conference on Language, Information and Computation, pages 161–168 Commerce, Industry and Energy in Korea and provides the patent attorneys and the patent examiners with the on-line English-Korean machine translation service for electro-electric patent documents (http://www.ipac.or.kr) because it helped them understand the existing English patent documents easier and more rapidly (Choi et. al., 2007). It was due to the customization method (Kwon et. al., 2007) that we could change successfully an existing machine translation system from general domain to patent domain. Figure 1 …

Content truncated to 500 words