Investigating an Effective Character-level Embedding in Korean Sentence Classification Won Ik Cho, Seok Min Kim, and Nam Soo Kim Human Interface Laboratory Department of Electrical and Computer Engineering and INMC Seoul National University 1 Gwanak-ro, Gwanak-gu, Seoul, Korea, 08826 {wicho,smkim}@hi.snu.ac.kr,
[email protected] Abstract used for many tasks such as machine translation (Ling et al., 2015), noisy document represen- Different from the writing systems of many tation (Dhingra et al., 2016), language correc- Romance and Germanic languages, some lan- guages or language families show complex tion (Xie et al., 2016), and word segmentation conjunct forms in character composition. For (Cho et al., 2018a). However, little consideration such cases where the conjuncts consist of was done for intra-language performance compari- the components representing consonant(s) and son regarding variant representation types. Unlike vowel, various character encoding schemes English, a Germanic language written with an can be adopted beyond merely making up a alphabet comprising 26 characters, many languages one-hot vector. However, there has been lit- used in East Asia are written with scripts whose tle work done on intra-language comparison characters can be further decomposed into sub-parts regarding performances using each represen- tation. In this study, utilizing the Korean lan- representing individual consonants or vowels. This guage which is character-rich and agglutina- conveys that (sub-)character-level representation tive, we investigate an encoding scheme that is for such languages has the potential to be managed the most effective among Jamo1-level one-hot, with more than just a simple one-hot encoding.