Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair Language Modeling and Translation

TL;DR

The paper addresses the dual nature of gender signals in language models by separating stereotypical bias from factual gender cues and proposes 2DAMA, a dual debiasing framework that combines DAMA (model editing), LEACE (concept erasure), and a covariance-based erasure step. It presents a theoretical basis and practical algorithm for selectively erasing bias while preserving useful factual gender information, with a bias-to-feature threshold guiding dimension selection. Empirical results across English and multilingual translation show substantial bias reduction with only modest degradation in non-gender tasks, and demonstrate cross-lingual applicability and bias sharing across languages. This work advances fair language technology by enabling more reliable gender representation and translation in morphologically rich languages while maintaining task performance.

Abstract

Mitigation of biases, such as language models' reliance on gender stereotypes, is a crucial endeavor required for the creation of reliable and useful language technology. The crucial aspect of debiasing is to ensure that the models preserve their versatile capabilities, including their ability to solve language tasks and equitably represent various genders. To address this issue, we introduce a streamlined Dual Dabiasing Algorithm through Model Adaptation (2DAMA). Novel Dual Debiasing enables robust reduction of stereotypical bias while preserving desired factual gender information encoded by language models. We show that 2DAMA effectively reduces gender bias in English and is one of the first approaches facilitating the mitigation of stereotypical tendencies in translation. The proposed method's key advantage is the preservation of factual gender cues, which are useful in a wide range of natural language processing tasks.

Figures (6)

• Figure 1: Dual character of gender signals encoded in language models: stereotypical cues are shown on the left, and factual gender cues are shown on the right-hand side. "Die Ärztin" and "der Arzt" are respectively female and male German translation for "the doctor".
• Figure 2: Schema (b) shows DAMA intervention in a language model layer. (a) We show the model's behavior when presented with a stereotypical prompt in three languages. Specifically, (c) shows the projections of the feed-forward latent vector ($\vec{u}$) onto the output space. With DAMA (lower arrow), we nullify the gender component of the representation. It results in balanced probabilities of gendered tokens in the model's output, as shown in (d). Adapted from limisiewicz2024debiasing.
• Figure 3: Visualization of dimensions and their variances related to stereotypical and factual gender signals identified by Dual Debiasing algorithm in 26th layer of Llama 2 13B. The red dots denote the bias-to-feature threshold $t=0.05$. In 2DAMA, the dimension is preserved if stereotypical covariance is below the threshold.
• Figure 4: The hyperparameter analysis for 2DAMA applied to Llama 2 13B model on performance and bias in language modeling. We measured bias on gendered prompts by linear coefficients: $a_s$ and $a_f$, the language modeling capabilities are measured by perplexity. Stars mark the performance of the best setting. The dashed line corresponds to the scores of the original model.
• Figure 5: The hyperparameter analysis for 2DAMA applied to ALMA-R 13B model on performance and bias in translation to German. We measured bias via WinoMT metrics $\Delta S$ and $\Delta G$. The translation quality to Germna is measured by chrf on WMT-22. Stars mark the performance of the best setting. The dashed line corresponds to the scores of the original model.

Theorems & Definitions (9)

• Theorem 1: LEACE
• Theorem 2: DAMA-LEACE
• Theorem 3: DUAL-DEBIASING
• Definition 1: Moore-Penrose Pseudoinvers
• Definition 2: Matrix Square Root
• Definition 3: Covariance Matrix
• Theorem 4: Gauss-Markov: Probabilistic Least Squares
• proof
• proof