READ-IT: Tools for understanding and supporting children's reading development

Literacy skills are vital to a modern economy, yet many children fail to learn to read effectively by the time they leave school, with an increasing gap in attainment between the lowest and highest reading skills since COVID-19, at an estimated annual cost to the UK economy of £20 billion. Learning to read is slow, varies substantially in pace between children, and requires substantial classroom time to support. It is thus a key area where unlocking the potential of AI to assist in optimising the process can benefit educational outcomes. However, there are four challenges facing us before this can be achieved, which we will meet in our READ-IT project. First, individualised data on children learning to read are required in order to develop AI-inspired models, and these data do not exist. Second, we need to develop models that are able to process relatively sparse data produced during individual children’s early reading. Third, we need to understand how these models simulate and predict human behaviour, such that their use in educational settings is responsible and ethical in guiding children’s reading. Fourth, we need to incorporate these model predictions into tools that are useful, accessible, and useable by teachers in the classroom supporting their children. There are four stages to this project, each relating to these challenges. First, we will understand individual children’s reading development by tracking, for the first time, 80 primary school children’s word-by-word, week-by-week reading over two years of early literacy training in the classroom. Second, we will model these data in order to predict children’s individualised reading development. We will apply cognitive-inspired neural-network (CI-NN) models of reading, as well as long-short term (LSTM) and large language models (LLM) to determine whether predictive models with or without knowledge of the language are able to better simulate individual children’s reading development. Third, we will analyse the models to ensure responsible use of predictive models, by determining the causal properties of predictions from the LSTM and LLM approaches, coupled with explainable-deep-neural-networks and interpretable-by-design deep learning algorithms. Fourth, we will design tools that will support children’s reading development, by individualised tuning of reading materials that scaffold the child’s reading skills at that point in their learning, developing interfaces that are useable by teachers in classroom settings. By the end of this project, we will have individualised data and predictive models of children’s reading development. We will have a set of analytical methods such that we can uncover the causal properties of deep-learning and LLM models used to predict individual behaviour. We will have educational-technology tools that can apply in the classroom, co-created with teachers, charitable organisations, and children’s book publishers. The applications of bespoke reading support for individual children has the potential to transform children’s literacy in early schooling, empowering teachers with the knowledge they need to address areas of weakness in individual children’s reading, potentially closing the attainment gap in children's skills in school, and ultimately benefiting the UK economy.