Bias Analysis of L2 Speaking Assessment Systems Using Concept Activation Vectors

2026-08-06

AI Summary

This study addresses bias in automated scoring systems for high-stakes language assessments, where speaker attributes such as native language and age—irrelevant to proficiency—can unduly influence outcomes. It extends Concept Activation Vectors (CAVs) to multimodal speech scoring models based on BERT and Whisper, integrating sparse autoencoders (SAEs) to enhance the linear recoverability and interpretability of such extraneous concepts in embedding spaces. The findings reveal that both the recoverability of these concepts and their impact on scores are highly dependent on model architecture and representation layer. While SAEs improve linear separability of concepts, they simultaneously reduce sensitivity to these concepts in lower-dimensional layers of the original space, uncovering a decoupling between the mere presence of a concept and its actual influence on model predictions.

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.

Problem

• bias
• L2 speaking assessment
• fairness
• concept activation vectors
• interpretability

Innovation

• Concept Activation Vectors
• Bias Analysis
• Sparse Autoencoders
• Multimodal Assessment
• Interpretability