Automatic Construction of a Phonics Curriculum for Reading Education Using the Transformer Neural Network

Cassandra Potier Watkins (1) , Olivier Dehaene (2) , Stanislas Dehaene (1, 3)

1 UNICOG-U992 - Neuroimagerie cognitive - Psychologie cognitive expérimentale
2 Owkin France
3 Collège de France - Chaire Psychologie cognitive expérimentale

Key to effective phonics instruction is the teaching of graphemephoneme (GP) correspondences in a systematic progression that starts with the most frequent and consistent pronunciation rules. However, discovering the relevant rules is not a an easy task and usually requires subjective analysis by a native speaker and/or expert linguist. We describe GPA4.0, a submodule to the Transformer neural network model that automatizes the task of grapheme-tophoneme (g2p) transcription and alignment. The network is trained with four different languages of decreasing orthographic transparency (Spanish

Mots clés: phonics instruction; g2p; attention

Domaines: Environnements Informatiques pour l'Apprentissage Humain; Informatique et langage [cs.CL]; Réseau de neurones [cs.NE]; Sciences cognitives

Cassandra Potier Watkins, Olivier Dehaene, Stanislas Dehaene. Automatic Construction of a Phonics Curriculum for Reading Education Using the Transformer Neural Network. Seiji Isotani; Eva Millán; Amy Ogan; Peter Hastings; Bruce McLaren; Rose Luckin. Artificial Intelligence in Education. 20th International Conference, AIED 2019, Chicago, IL, USA, June 25-29, 2019, Proceedings, Part II, 11626, Springer International Publishing, pp.226-231, 2019, Lecture Notes in Computer Science, 978-3-030-23206-1. ⟨10.1007/978-3-030-23207-8_42⟩. ⟨hal-03702075⟩