Automatic Quiz Generation System

Quizzes provide a quick way of assessing students’ knowledge and understanding on specific topics. The manual creation of quizzes however is a very demanding task. In this project, we present an Automatic Quiz Generation System (AQGS) that generates quizzes from a given knowledge source (text) without human intervention. We use natural language processing tech- niques in combination with Transformers to generate high quality quiz questions in an efficient manner. Our system generates quizzes consisting of multiple choice questions (MCQs), True/False questions, and open-ended questions. We fine-tuned a pre-trained Text-to-Text Transfer Transformer (T5) model on a question generation task with the SQuAD dataset (reading com- prehension dataset consisting of 100,000+ questions). The fine-tuned T5 model generates stem questions for the MCQs. We select alternatives answers (distractors) for the MCQs from pub- licly available lexical databases such as WordNet and Sense2Vec. We transform quantitative and qualitative terms in adjectival phrases to generate True/False questions, and we generate open-ended questions by first generating a "yes" or "no" question and adding "open question extension terms" such as briefly explain your answer, discuss, argue, and why. We evaluate the questions generated by our system using both standard automatic evaluation metrics and hu- man judgments. The results of the evaluation shows that the system generates questions that are syntactically correct, semantically correct, contextually relevant and complete.