Presented at the 2023 International Conference on Machine Learning and Applications (ICMLA 2023), Jacksonville, FL, USA, pages 1228-1236, IEEE.
The paper combines GPT-3, GPT2-XL and BERT into a system that generates and verifies crossword clues for educational use, trained on over seven million clue-answer pairs from 1913 to mid-2021. Fine-tuned classifiers and zero-shot checks remove weak clue-answer pairs, and the evaluation results were positive.
Authors: Kamyar Zeinalipour, Tommaso Iaquinta, Giovanni Angelini, Leonardo Rigutini, Marco Maggini, Marco Gori.