Towards a more efficient bias detection in financial language models

TL;DR

This paper investigates bias detection in financial language models, demonstrating that leveraging cross-model insights and input reuse can significantly reduce detection costs while maintaining effectiveness.

Contribution

It introduces a cost-efficient bias detection approach using cross-model guidance and input reuse, validated on large-scale financial news data.

Findings

All models exhibit bias in both atomic and intersectional settings.

Consistent bias patterns enable reuse, reducing detection costs by up to 80%.

Up to 73% of biased behaviors can be identified with only 20% of input pairs.

Abstract

Bias in financial language models constitutes a major obstacle to their adoption in real-world applications. Detecting such bias is challenging, as it requires identifying inputs whose predictions change when varying properties unrelated to the decision, such as demographic attributes. Existing approaches typically rely on exhaustive mutation and pairwise prediction analysis over large corpora, which is effective but computationally expensive-particularly for large language models and can become impractical in continuous retraining and releasing processes. Aiming at reducing this cost, we conduct a large-scale study of bias in five financial language models, examining similarities in their bias tendencies across protected attributes and exploring cross-model-guided bias detection to identify bias-revealing inputs earlier. Our study uses approximately 17k real financial news sentences, mutated to construct over 125k original-mutant pairs. Results show that all models exhibit bias under both atomic (0.58\%-6.05\%) and intersectional (0.75\%-5.97\%) settings. Moreover, we observe consistent patterns in bias-revealing inputs across models, enabling substantial reuse and cost reduction in bias detection. For example, up to 73\% of FinMA's biased behaviours can be uncovered using only 20\% of the input pairs when guided by properties derived from DistilRoBERTa outputs.