A Virtual Reading Companion for Second Language Learners
BookBuddy is a human-centered AI design experiment that seeks to understand the possible benefits when children interact with an intelligent agent while developing language skills through reading. It consists of an array of chatbots. One chatbot, called BookBot, is an ensemble chatbot system that combines two model architectures based on recent advancements in deep learning neural networks. For general conversation, a sequence-to-sequence machine translation model was trained on movie scripts and other dialogue-based text to learn how to respond to popular English conversation expressions. This chatbot is designed to respond to open-ended questions like “how are you?” and “what books do you like to read?” For book-related conversation, this chatbot was combined with a generative question-answering model trained to answer context-specific questions related to the current book that the child is reading. This additional bot was trained on the Stanford Question Answering Dataset (SQuAD) and can answer questions about the characters and plots of the book. Additionally, the ensemble model can answer vocabulary questions, and basic arithmetic based questions, like, “what does recalcitrant mean?” or “what is 435 minus 381?” Finally, we added a few rule-based touches, like being able to ask the chatbot to tell a joke or tell time.