Using Collocations and K-means Clustering to Improve the N-pos Model for Japanese IME

Kana-Kanji conversion is known as one of the representative applications of Natural Language Processing (NLP) for the Japanese language. The N-pos model, presenting the probability of a Kanji candidate sequence by the product of bi-gram Part-of-Speech (POS) probabilities and POS-to-word emission probabilities, has been successfully applied in a number of well-known Japanese Input Method Editor (IME) systems. However, since N-pos model is an approximation of n-gram word-based language model, important word-to-word collocation information are lost during this compression and lead to a drop of the conversion accuracies. In order to overcome this problem, we propose ways to improve current N-pos model. One way is to append the highfrequency collocations and the other way is to sub-categorize the huge POS sets to make them more representative. Experiments on large-scale data verified our proposals.