Applying Statistical Post-Editing to English-to-Korean Rule-based Machine Translation System

Copyright 2012 by Ki-Young Lee and Young-Gil Kim 26th Pacific Asia Conference on Language,Information and Computation pages 318–324 Applying Statistical Post-Editing to English-to-Korean Rule-based Machine Translation System Ki-Young Lee and Young-Gil Kim Natural Language Processing Team, Electronics and Telecommunications Research Institute, 138 Gajeongno, Yuseong-gu, Daejeon, Korea {leeky, kimyk}@etri.re.kr Abstract Conventional rule-based machine translation system suffers from its weakness of fluency in the view of target language generation. In particular, when translating English spoken language to Korean, the fluency of translation result is as important as adequacy in the aspect of readability and understanding. This problem is more severe in language pairs such as English-Korean. It’s because English and Korean belong to different language family. So they have distinct characteristics. And this issue is very important factor which effects translation quality. This paper describes a statistical post-editing for improving the fluency of rule-based machine translation system. Through various experiments, we examined the effect of statistical post- editing for FromTo-EK1 system which is a kind of rule-based machine translation system, for spoken language. The experiments showed promising results for translating diverse English spoken language sentences. 1 FromTo-EK is an English-to-Korean rule-based machine translation system for some various domains (patent, paper, email and messenger). 1 Introduction There have been many improvements in machine translation from rule-based machine translation (RBMT) to the latest statistical machine translation (SMT). Approaches for machine translation can be typically classified into conventional rule-based approach and statistical approach (Jin et al., 2008). RBMT translates a source sentence to a target sentence through analysis process, transfer process and generation process using analysis rules, dictionaries and transfer rules as its main translation knowledge. On the other hand, SMT system accomplishes translation using translation model and language model obtained from training large parallel corpus composed of source sentences and the corresponding target sentences (Koehn et al. 2003). Comparing two approaches, they have opposite features. That is, rule-based approach is better than statistical approach in the aspect of translation accuracy. However, fluency is contrary to each other. The language pairs that linguistic differences are huge such as English-Korean show these kinds of features apparently. We aim to improve the translation fluency of RBMT system by introducing SPE. The proposed method is similar to SMT. Difference is the composition of parallel corpus used to build statistical models. To build model for post-editing, parallel corpus should be ready, which is composed of the pairs of the sentence translated by RBMT and the corresponding correct sentence 318 Copyright 2012 by Ki-Young Lee and Young-Gil Kim 26 th Pacific Asia Conference on Language, Information and Computation pages 318-324 translated by human translators. Using this parallel corpus, we can build statistical model to post-edit RBMT results. Also, we explain some points to consider when applying SPE to English-to-Korean translation by various experiments.  Our method consists of the following steps: Constructing parallel corpus composed of translation results by English-to-Korean RBMT system and translation results by human translator of English source sentences.  Building translation model and language model for …

Content truncated to 500 words