Research on Japanese Teaching System Based on Web Using Deep Learning Framework

Japanese as a language poses challenge to learners with complex grammar, multi-writings with the Hiragana, Katakana and Kanji systems, and words that are contextual in usage or meaning. Traditional computer-assisted language learning (CALL) systems based on static learning content or rules, are not easily adapted for individual learners' needs and do not keep pace with learners' progress. This research will propose a hybrid model consisting of a Deep Belief Network (DBN) and Convolutional Neural Network (CNN) capable of fully integrating with a Japanese language teaching platform based in J2EE. The DBN will be used for the extraction of features at a high-level, while semantics use of the word pictures is used to relate to grammar structures. The CNN model will be able to capture both the visual and spatial features of Japanese text with accuracy of printed or handwritten systems. The supporting datasets (Kuzushiji-MNIST - handwritten Japanese) and Joyo Kanji - a Japanese dataset, will consist of training data used for this hybrid DBN and CNN model. The experimental evaluations of the results indicate that the proposed hybrid model resulted in an accuracy of 97.6 % in character recognition and in sentence classification based on context, which surpassed accuracy of the DBN and CNN as stand-along models. The fully integrated application of the hybrid deep learning model into a J2EE framework provides scalable deployment, gives the ability for real-time feedback to students, and more personalized adaptive learner paths, albeit in an experimental setting, leads to greater levels of learner engagement and performance.