A generative AI-driven interactive listening assessment task.

Introduction: Assessments of interactional competence have traditionally been limited in large-scale language assessments. The listening portion suffers from construct underrepresentation, whereas the speaking portion suffers from limited task formats such as in-person interviews or role plays. Human-delivered tasks are challenging to administer at large scales, while automated assessments are typically very narrow in their assessment of the construct because they have carried over the limitations of traditional paper-based tasks to digital formats. However, computer-based assessments do allow for more interactive, automatically administered tasks, but come with increased complexity in task creation. Large language models present new opportunities for enhanced automated item generation (AIG) processes that can create complex content types and tasks at scale that support richer assessments.

Methods: This paper describes the use of such methods to generate content at scale for an interactive listening measure of interactional competence for the Duolingo English Test (DET), a large-scale, high-stakes test of English proficiency. The Interactive Listening task assesses test takers' ability to participate in a full conversation, resulting in a more authentic assessment of interactive listening ability than prior automated assessments by positing comprehension and interaction as purposes of listening.

Results and discussion: The results of a pilot of 713 tasks with hundreds of responses per task, along with the results of human review, demonstrate the feasibility of a human-in-the-loop, generative AI-driven approach for automatic creation of complex educational assessments at scale.