What is Needed the Most in MT-Supported Paper Writing

What is Needed the Most in MT-Supported Paper Writing Chang Hyun Kim, Oh-Woog Kwon, Young Kil Kim ETRI, NLP Team 161 Gajeong-dong, Yuseong-gu, 305-350 Daejeon, Korea {chkim,ohwoog,kimyk}@etri.re.kr Abstract. This paper addresses our system which provides an effective method to write an English paper suitable for international conferences and analyze the system’s pros and cons through the user’s data collected from operating the system for 6 months. The system consists of Korean-English paper MT module supported by user interaction environment. Our original Korean-English paper MT system was quite useful for understanding, but not satisfactory for writing. So, we analyzed our system to trace what caused such dissatisfaction. We classified the analysis results into three main categories, that is, the errors in the source sentence itself, the errors of our MT system, and the absence of the appropriate domain-specific expression information. For each category we provide an alternative method and show the effectiveness through analyzing the user’s data. We can confirm that our system can be used quite usefully for paper writing. Keywords: Machine Translation, User Interaction, Korean-English Translation, Paper writing 1. Introduction Many Koreans who are not fluent in English have difficulty in writing a scientific paper or technical documents in English. Although the state-of-the-art Korean-English MT system is quite useful for understanding, many people hesitate to use the MT system for paper writing. Understanding does not necessarily require perfect sentences, but writing papers does require impeccable grammars and correct, native expressions. The main purpose of the original Korean-English paper MT system (Kim, 2007) was to help researchers or students to submit their papers to a conference or an academic journal. This system had been developed through customization of the patent MT system (Hong, 2005), which is currently serviced by KIPO (Korean Intellectual Property Office) and is being used by more than 20 countries with positive feedbacks from foreign users. The customization process includes construction of translation resources specialized in scientific papers, modification of the engine to reflect the linguistic characteristics of academic papers. Moreover, a Controlled- Language (CL) guided Korean rewriting checker is provided to correct the errors in Korean spelling and sentence structure. Language model component reports the unlikely or unnatural English expression. Several beta testers of the original MT system reported that it was very helpful in writing a paper, but that was not enough. They said that the user interface was inconvenient, and they wanted to understand what caused the mistranslations, and how to correct them. Besides, the MT output still contained erroneous expressions even the users rewrite sentences according to the guidelines of the CL-checker. We analyzed those reports and found 3 main reasons: the errors in the source sentence itself, the errors of our MT system, and the absence of the appropriate domain-specific expression information. In this paper, we provide alternative methods to cope with those problems within our user interaction environment. As a result, authors can interact with the system through modification of source sentences, correction of engine errors, and correction of target sentence …

Content truncated to 500 words