BUILDING AN AUTOMATIC ENGLISH QUESTION GENERATION SYSTEM FROM PDF DOCUMENTS USING PRE-TRAINED LANGUAGE MODELS COMBINED WITH DISTRACTOR GENERATION.

Abstract

Generating high-quality assessment questions is a time-consuming and laborintensive process in educational settings. To address this challenge, this study proposes a multi-tasking automated question generation system (QAG) capable of generating short-answer questions, fill-in-the-blank questions, and multiple-choice questions (MCQs) from input text content. The proposed system employs a sequential learning framework based on a pre-trained T5 (Text-to-Text Transfer Transformer) model, refined to unify different question generation tasks within a single text-to-text model. To support multi-tasking learning, the model is trained on a combined dataset constructed from SQuAD and RACE. Furthermore, to enhance the quality of the generated multiple-choice questions, particularly the reasonableness and variety of incorrect answer options, a hybrid incorrect answer option generation module is integrated. This module combines named entity recognition (NER), external knowledge bases, and semantic similarity filtering based on Sentence-BERT, enabling the system to generate contextually relevant distractors that are distinct from the correct answer. The project aims to develop a system capable of efficiently generating diverse and coherent assessment questions across multiple formats, minimizing the manual effort required in creating educational content and supporting the creation of scalable assessments.