Generative Artificial Intelligence in Education

Abstract Recent advances in generative artificial intelligence have opened new possibilities for the development of educational tools capable of supporting teachers in the creation of engaging learning activities. This thesis investigates how large language models (LLMs) can be integrated into educational workflows to automatically generate and adapt crossword puzzles for learning purposes. The work focuses on three main contributions: the construction of a large-scale linguistic dataset for Italian crossword clues, the design of an end-to-end architecture for automated crossword generation, and the introduction of a surprisal-based framework for estimating puzzle difficulty. The first contribution of the thesis is the creation of a large dataset of crossword clue–answer pairs in Italian. The dataset contains more than 125,000 pairs collected from publicly available crossword sources and subsequently processed through normalization, filtering, and annotation procedures. Building on this resource, the thesis proposes a modular architecture for automatic crossword generation. The pairs must then be revised by the user before being placed into a crossword grid generated by a dedicated grid-construction algorithm based on a backtracking search with scoring heuristics that prioritize grid density and word intersections. The thesis further investigates the problem of estimating crossword difficulty prior to user interaction. To address this challenge, it introduces a surprisal-based measure derived from language model probabilities. Surprisal quantifies how unexpected a specific answer is given its clue, and is interpreted as a proxy for the cognitive effort required to retrieve the correct solution. Empirical analyses show that surprisal correlates negatively with human accuracy and positively with solution time, indicating that higher surprisal values correspond to more difficult clues. These results suggest that language model probabilities can provide a principled signal for estimating puzzle difficulty. Together, the dataset, the generation architecture, and the surprisal-based difficulty model form a framework for scalable educational crossword creation. Beyond crossword generation itself, the thesis demonstrates how generative AI systems can be combined with validation mechanisms and cognitive metrics to support the development of adaptive and pedagogically grounded educational activities.