AI summaryⓘ
The authors studied how automatic language speaking tests might be unfairly influenced by things like a speaker's native language or age instead of their actual speaking skill. They used a method called Concept Activation Vectors (CAVs) to check if certain characteristics are encoded in the models and if they affect scoring. They tested this on two modern speaking assessment systems based on BERT and Whisper models and also tried improving CAVs with sparse autoencoders (SAEs). Their results show that whether a concept can be detected and whether it influences the score both depend on the specific model design, and that SAEs help find concepts but can reduce how sensitive scores are to them. This work helps understand and audit potential biases in automated speaking tests better.
Automatic speaking assessmentSecond language (L2) proficiencyConcept Activation Vectors (CAVs)Transformer modelsBERTWhisper modelSparse autoencoders (SAEs)Bias detectionModel interpretabilityGradient-based sensitivity
Abstract
Automatic speaking assessment systems are increasingly deployed in high-stakes settings to mark second language (L2) learners' speaking tests, making it critical to show that their scores depend on speaking proficiency rather than irrelevant speaker attributes such as first language (L1) or age. Transformer-based foundation models have improved the accuracy of these L2 speaking graders, but their black-box representations make fairness and interpretability analysis more difficult. Building on prior work that used Concept Activation Vectors (CAVs) to detect bias towards unwanted attributes (`concepts') in feature-based graders, we extend CAV-based analysis to two neural speaking assessment systems: a text-based BERT grader and a speech-and-text multimodal grader based on Whisper. CAVs represent human-interpretable concepts as directions in a model's activation space, allowing us to distinguish between whether a concept is encoded in a model's internal representations and whether it influences the predicted score, the latter quantified using a gradient-based sensitivity metric. Since CAVs rely on linear separability, which is less likely in complex neural embedding spaces, we also investigate whether sparse autoencoders (SAEs) provide cleaner concept directions by learning CAVs in a sparse latent space and mapping them back to activation space. Our analysis shows that concept recoverability depends strongly on the representation and architecture being probed, rather than on the concept alone. Sensitivity to concepts is also architecture-dependent. SAEs make concepts more linearly recoverable, but attenuate the original activation-space sensitivity, especially in low-dimensional layers. These findings highlight the need to distinguish concept recoverability from concept influence when auditing bias in speaking assessment systems.