Character Decomposition for Japanese-Chinese Character-Level Neural Machine Translation

After years of development, Neural Machine Translation (NMT) has produced richer translation results than ever over various language pairs, becoming a new machine translation model with great potential. For the NMT model, it can only translate words/characters contained in the training data. One problem on NMT is handling of the low-frequency words/characters in the training data. In this paper, we propose a method for removing characters whose frequencies of appearance are less than a given minimum threshold by decomposing such characters into their components and/or pseudo-characters, using the Chinese character decomposition table we made. Experiments of Japanese-to-Chinese and Chinese-to-Japanese NMT with ASPEC-JC (Asian Scientific Paper Excerpt Corpus, Japanese-Chinese) corpus show that the BLEU scores, the training time and the number of parameters are varied with the number of the given minimum thresholds of decomposed characters.