Cue word guided question generation with BERT model fine-tuned on natural question dataset

Abstract

This thesis aims to develop an efficient question generator for an automated tutoring system. Given a context passage with an answer, the question generator asks questions to help the reader learn new material. By utilizing the BERT model, this thesis experiments on generating type-specific questions with a cue word. This thesis also uses an RNN encoder-decoder architecture for question generation on SQuAD as a comparative baseline and fine-tune the BERT question generation model on Google's Natural Question dataset. Ultimately, I deliver a RESTful API by the end of this year-long master program.