Automatic question generation with natural language processing

Abstract

In Natural Language Processing (NLP), Automatic Question Generation (AQG) is an important task that involves generating human-comprehensible questions from an input text. There are many useful applications in AQG, notably in educational settings, to create quizzes or reading comprehension papers. Many techniques have been studied for AQG, from rule-based algorithms to complex deep learning networks. As machine learning advances, transformer-based neural networks are more robust than rule-based logic in generating questions without prior knowledge of the grammar rules. However, current state-of-the-art models still perform worse in paragraph-level texts than sentence-level inputs. This FYP studies the pros and cons of different AQG methods in long paragraphs to design a multitasking model for AQG. By utilising transfer learning, this project’s best model is a robust AQG + summarisation T5 transformer, outperforming existing more complex Seq2seq and RNN models, achieving scores on BLEU-4 of 16.37, METEOR of 20.4, and ROUGE_L of 41.50. This project’s model does not require answer input, i.e., less information is given, but the performance is on par with many answer-aware models. The effectiveness of different AQG methods is analysed. Then, the T5-based AQG transfer learning model pipeline is designed, and different training parameters (batch size, epochs, learning rate) are tuned to optimise performance. The model’s predictions are also evaluated against human-generated questions and compared with existing models. Finally, a full-stack React Web Application using the model is implemented to demonstrate its application as a Quiz Generator.