Learn to Speak Like A Native: AI-powered Chatbot Simulating Natural Conversation for Language Tutoring

Abstract In a globalizing world, the demand for acquiring a new language is expanding at a fast pace. However, language learners who have little opportunity to use the targeted language in their daily life are unable to increase their familiarity with the language given that contextual conversation is proven helpful in language acquisition. The current study dedicates to address this problem by proposing a language learning chatbot with which one can practice in a simulated setting. The chatbot is based on the deep neural network and is trained by MultiWOZ and personaChat collected in previous works. One chatbot structure is proposed to optimize the performance of the DNN chatbot. In comparison with the conventional chatbot, the new structure features a higher accuracy rate of generating appropriate utterance and suffers less from overfitting. Both machine and human evaluation show that the chatbot is able to produce transactional conversations that cover multiple domains of daily life conversations and incorporates a large vocabulary base as well as diverse sentence structures, which are beneficial for students.