Analysis and assessment of neural network-based text comprehension quiz generation algorithms

Last decades have been characterised by huge improvements in natural language processing tasks, achieving extraordinary results thanks to both extensive use of artificial intelligence and to availability of tons of data. Aim of this thesis is to overview the way neural networks have developed through the years, with particular focus on architecture that underlies latest state-of-the-art results in natural language processing applications, as well as to apply such technologies for complex tasks like generating multiple choice text comprehension questions. As part of a company project in Fluentify - online platform to learn languages -, this study takes advantage of the most advanced machine learning models, including transformers T5, GPT-2 and BERT - to test libraries that automatise the process of generating questions with multiple choice answers from a given text paragraph; moreover, in order to assess the accuracy of these techniques, a double evaluation system has been proposed for both questions and answers, looking at grammar, semantic and general sense for the former, and whether the answers contain the correct one as well as whether distractors make sense for the latter, to highlight strong points and weaknesses for these models and make suggestions for further improvements.