An Empirical Study of Korean Sentence Representation with Various Tokenizations

It is important how the token unit is defined in a sentence in natural language process tasks, such as text classification, machine translation, and generation. Many studies recently utilized the subword tokenization in language models such as BERT, KoBERT, and ALBERT. Although these language models achieved state-of-the-art results in various NLP tasks, it is not clear whether the subword tokenization is the best token unit for Korean sentence embedding. Thus, we carried out sentence embedding based on word, morpheme, subword, and submorpheme, respectively, on Korean sentiment analysis. We explored the two-sentence representation methods for sentence embedding: considering the order of tokens in a sentence and not considering the order. While inputting a sentence, which is decomposed by token unit, to the two-sentence representation methods, we construct the sentence embedding with various tokenizations to find the most effective token unit for Korean sentence embedding. In our work, we confirmed: the robustness of the subword unit for out-of-vocabulary (OOV) problems compared to other token units, the disadvantage of replacing whitespace with a particular symbol in the sentiment analysis task, and that the optimal vocabulary size is 16K in subword and submorpheme tokenization. We empirically noticed that the subword, which was tokenized by a vocabulary size of 16K without replacement of whitespace, was the most effective for sentence embedding on the Korean sentiment analysis task.