Report: Reform Exams for Language Disorders
A new report is calling for reforms to assessments to better support students with Developmental Language Disorder (DLD). The report highlights the need for evaluation methods that are fair and accessible. This has implications for the design of adaptive assessment modules in edtech tools to ensure they can evaluate children with language-based learning differences equitably.
- Developmental Language Disorder (DLD) is a hidden disability affecting approximately 7.6% of children, which equates to roughly two students in every classroom of 30. These students are six times more likely to have reading difficulties and four times more likely to struggle with math compared to their peers. - The report calls for reform because current UK exam accommodations often rely on reading and vocabulary scores to determine eligibility for support, failing to account for broader language comprehension deficits central to DLD. As a result, many students are denied adjustments like extra time or human readers. - Machine learning models are being used to create more equitable and precise language assessments, such as the Duolingo English Test. These systems use natural language processing to estimate the difficulty of test items directly, enabling computer-adaptive tests that adjust to the user's ability level without the need for expensive human pilot testing. - Research into adaptive testing demonstrates that deep reinforcement learning (DRL) can optimize question selection based on a learner's real-time performance. This approach can improve the accuracy of proficiency predictions to over 95% while simultaneously reducing the time required for assessment. - A significant challenge for AI tutors is that speech recognition models, including state-of-the-art systems, perform poorly with children's voices. Children's smaller, developing vocal tracts create different acoustic properties, and their speech has greater variability in pitch, rhythm, and disfluencies like false starts. - The performance gap in speech recognition for children is largely a data problem; most ASR systems are trained on vast datasets of adult speech. Even models trained on hundreds of thousands of hours of audio struggle without specific fine-tuning on representative datasets of child speech. - For children with DLD, reading comprehension is often impacted by underlying difficulties in listening comprehension. Evidence-based instructional methods like systematic phonics, which explicitly teaches the relationship between letters and sounds in a structured sequence, are particularly effective for these learners. - Ofqual, the UK's exams regulator, has issued guidance that assessments should use accessible language and that a test of numeracy, for example, should not be obstructed by overly complex text. However, the new report argues these principles are not being applied effectively for students with DLD.
