Towards Equitable, Diverse and Inclusive Language Models in AI-enhanced Education Technology for Language Learners

Andrew Caines, Paula Buttery, Graham Seed

Abstract: Artificial intelligence (AI) plays an important and growing part in education technology for language learning and assessment. The role of AI ranges from automated assessment to task selection, and automated feedback to content creation. Recent trends in language-related AI – ‘natural language processing', as it is known – involve large language models which are capable of generating fairly coherent and grammatical texts, holding chat conversations with end-users, and performing various tasks such as text classification or labelling. In this chapter we examine these trends more closely, discuss the datasets, methods and ideology underpinning them, and explore how this branch of AI presents benefits and opportunities for creators of education technology. We show that, as with the field of natural language processing in general, there is a bias towards the English language, with resources for other languages being more limited. In addition, certain varieties of English dominate the training data, as it tends to come from the Web, and the training methods promote a frequency-driven notion of grammaticality. We explore why these factors may be problematic with regard to equality, diversity and inclusion, and consider some of the steps which might mitigate them in the training of large language models and the development of education technology for language learners.

External IDs:doi:10.4324/9781003398172-29