Study on Mistype Classification in Japanese Input Using Machine Learning

When considering usability of electronic devices such as computers and smartphones, how to input characters without mistakes is a very important point in the interface. Therefore in this paper, we examine classification of mistypes for Japanese input using machine learning. First, we classified the category of mistypes in Japanese input into six types (replacement, removal, exchange, insertion, involvement, repetition). We asked 12 subjects to input the words displayed and saw a trend. Experimental results showed that there was a common trend such as many replacement errors in all subjects, and characteristic mistypes were found for each subject. Next, we made the neural network (RNN · LSTM) to learn the mistype data of all subjects and classified. As a result, the correct answer rate of RNN was 10%, and the correct answer rate of LSTM was 88%. We are planning to use LSTM for correcting mistype in the future.