The GUS Framework offers a sophisticated approach to detecting social bias by examining three linguistic components: Generalizations, Unfairness, and Stereotypes, moving beyond simplistic binary classifications.

• Addresses limitations of traditional binary bias detection that often oversimplifies nuanced biases
• Implements a multi-dimensional approach targeting specific linguistic elements of bias
• Benchmarks performance across discriminative (encoder-only) and generative (decoder-only) language models
• Provides more actionable insights for content moderation and security systems

For security applications, this framework enables more precise identification of potentially harmful content while reducing false positives that can have significant emotional impact when content is incorrectly flagged.

The GUS Framework: Benchmarking Social Bias Classification with Discriminative (Encoder-Only) and Generative (Decoder-Only) Language Models