Rule-based Reordering and Post-Processing for Indonesian-Korean Statistical Machine Translation

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 287–295 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Candy Olivia Mawalim, Dessi Puji Lestari and Ayu Purwarianti Rule-based Reordering and Post-Processing for Indonesian-Korean Statistical Machine Translation Candy Olivia Mawalim, Dessi Puji Lestari, Ayu Purwarianti School of Electrical Engineering and Informatics Institut Teknologi Bandung candyoliviamawalim@gmail.com,dessipuji@gmail.com,ayu@informatika.org Abstract This paper presents several experiments on constructing Indonesian–Korean Statistical Machine Translation (SMT) system. A parallel corpus containing around 40,000 segments on each side has been developed for training the baseline SMT system that is built based on n-gram language model and the phrase-based translation table model. This system still has several problems, including non-translated phrases, mistranslation, incorrect phrase orders, and remaining Korean particles in the target language. To overcome these problems, some techniques are employed i.e. POS (part-of-speech) tag model, POS-based reordering rules, multiple steps translation, additional post-process, and their combinations. We then test the SMT system by randomly extracting segments from the parallel corpus. In general, the additional techniques lead to better performance in terms of BLEU score compared to the baseline system 1 Introduction Statistical Machine Translation (SMT) is a corpus- based MT for automatic translation. It has been growing rapidly since this approach gives some advantages, including language-independent and low-cost construction (Koehn, 2010). In the case of Indonesian–Korean translation, there has not been much research done in this field. It is probably because of the difficulty in constructing parallel corpus since both Indonesian and Korean are low- resource languages. As a valuable resource in developing SMT, we first construct a parallel corpus obtained from Korean learning books, drama and movie subtitles, and Bible text. By using this corpus, we construct the baseline SMT system. Phrase-based translation model is used since the previous studies have shown that phrase-based variant of SMT gives better performance than word-based variant of SMT (Koehn, 2010). After the baseline system has been built, we analyze the problems found on the translation results. Based on these problems, we investigate several additional techniques which can be used to overcome them. These additional techniques are tested with random segments from the parallel corpus. The quality of each system is determined by using smoothed BLEU metric, known as BLEU+1 (Lin and Och, 2004). BLEU score is calculated by multiplying the geometric mean of the test corpus’ modified precision scores with the exponential brevity penalty factor (Papineni, et al. 2002). 2 Related Work Parallel corpus is a valuable component needed in SMT to train models, optimize the model parameters, and test the translation quality. However, a good parallel corpus of low-resource languages such as Indonesian and Korean is hard to obtain. Therefore, we do not only use books as the source for constructing, but also subtitles and Bible. Automatic parallel corpus extraction from movie subtitles has been introduced in (Caroline et al., 2007). From this study, it was reported that 37,625 287 aligned pairs with a precision of 92.3% was obtained from 40 movies. Using Bible as the parallel corpus …

Content truncated to 500 words