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

Key to effective phonics instruction is the teaching of grapheme-phoneme (GP) correspondences in a systematic progression that starts with the most frequent and consistent pronunciation rules. However, discovering the relevant rules is not 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-to-phoneme (g2p) transcription and alignment. The network is trained with four different languages of decreasing orthographic transparency (Spanish < Portuguese < French < English). Our results show that the Transformer model improves on the current state-of-the-art in g2p transcription and that the attention mechanism allows for the alignment of graphemes to their corresponding phonemes. From the g2p aligned words, our software provides an optimally ordered phonics progression based on frequency and consistency in the target language, as well as an ordered list of words that teachers can use. This work exemplifies a practical way that neural networks can be used to develop educational materials for research and teachers. Submodules and phonics output are available at, https://github.com/OlivierDehaene/GPA4.0.