Beyond Static Word Lists for Bias Mitigation

52 concepts50 papers1 proposal

When early natural language processing models exhibited social biases, the computer science community responded with an intuitive fix: if a model treats certain demographic groups unfairly, we should explicitly define those groups and balance their representations. This gave rise to dictionary-based debiasing. By compiling static lists of identity terms (e.g., "man", "woman", "gay", "Muslim") or stereotypical attributes (e.g., "doctor", "nurse"), researchers could programmatically swap tokens, project embeddings onto orthogonal "neutral" subspaces, or penalize specific vocabulary associations.

However, treating fairness as a lookup table introduces a fundamental flaw. Language is fluid, context-dependent, and infinitely creative. A static dictionary acts as a rigid, brittle filter trying to catch a shifting smoke screen.

A growing body of research demonstrates that static word lists create severe coverage gaps and fail to capture how bias actually manifests. To build truly fair models, the field is shifting toward dynamic, context-aware, and automated discovery mechanisms that look at how language behaves, rather than what words are used.

The Failure Modes of Static Dictionaries

To understand why the field is moving away from predefined lexicons, we must first examine the three core ways they fail in real-world language applications.

1. Context Blindness and Framing Effects

Bias is rarely a property of an isolated word; it emerges from how words interact. A toxicity filter that relies on a static list of identity terms will often suffer from "keyword bias," incorrectly flagging neutral or positive sentences simply because they contain a word like "gay" or "black."

As Marta Recasens et al. demonstrate in Linguistic Models for Analyzing and Detecting Biased Language, bias in collaborative environments like Wikipedia is a dynamic property driven by framing and epistemological presuppositions. A static list of prohibited words cannot predict actual bias because it ignores the surrounding syntactic structure and historical edit patterns.

2. The Coverage Gap and Intersectional Blindness

If a demographic group, intersectional identity, or derogatory slang term is omitted from a human-curated list, the model remains completely blind to it. In Debiasing Pre-trained Contextualised Embeddings, Masahiro Kaneko et al. point out that target-word lists fail to generalize to emergent or nuanced social prejudices that fall outside the initial lexicon. This problem compounds when dealing with intersectional biases or subtle, non-explicit forms of prejudice that lack simple linguistic markers, a limitation highlighted by Haoyu Zhang et al. in Joint Vision-Language Social Bias Removal for CLIP.

3. Temporal and Domain Drift

Language evolves rapidly. Coded language, political jargon, and online slang shift faster than researchers can update datasets. Rowan Hall Maudslay argues in It's All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitution that dictionary-based methods cannot account for linguistic drift, failing to recognize when previously gendered terms become neutral or when new associations emerge.

Furthermore, a dictionary built on generic web text fails when applied to specialized domains. For instance, Thiemo Wambsganss et al. show in Bias at a Second Glance: A Deep Dive into Bias for German Educational Peer-Review Data Modeling that fixed word lists face severe out-of-vocabulary issues when applied to niche datasets like educational peer reviews, leading to incomplete fairness assessments.

[Static Lexicon Approach]
"The doctor walked in." ──> Look up "doctor" in list ──> Swap with "nurse" (Brittle, misses context)

[Dynamic Contextual Approach]
"The doctor walked in." ──> Analyze model attention/SHAP ──> Identify latent stereotypical framing

Shifting from Tokens to Contextual Behavior

To overcome these limitations, modern mitigation strategies move the intervention from the lexicon level to the representation level. Instead of telling the model which words to watch out for, these methods analyze how the model processes sentences to identify and penalize biased reasoning.

One prominent approach uses post-hoc explanations to regularize models during training. In Contextualizing Hate Speech Classifiers with Post-hoc Explanation, Brendan Kennedy 0001 et al. introduced a regularization technique that penalizes a classifier when its internal explanations reveal an over-reliance on identity terms. By forcing the model to focus on the surrounding context rather than the mere presence of a keyword, they successfully mitigated keyword bias without needing an exhaustive list of every protected group.

Similarly, we can use causal inference to dynamically isolate biased signals. Shaofei Shen et al. developed a framework in Words Can Be Confusing: Stereotype Bias Removal in Text Classification at the Word Level that uses SHAP (Shapley Additive exPlanations) values to dynamically detect which words are driving stereotypical predictions in real time. Once these words are identified contextually, a counterfactual inference step "subtracts" the biased signal from the prediction, bypassing the need for any static, pre-defined dictionary.

Automating the Discovery of Bias

If human curation is the bottleneck, why not let the models find their own biases? A parallel track of research focuses on automating the generation of prompts and lexicons, allowing the system to adapt to the specific corpus it is processing.

In Auto-Debias: Debiasing Masked Language Models with Automated Biased Prompts, Yue Guo et al. designed a system that automatically searches for "biased prompts" that trigger stereotypical responses in masked language models. The system then fine-tunes the model to equalize its outputs across these discovered contexts. While the framework still anchors on a baseline set of demographic terms, the behavioral triggers are discovered dynamically rather than hardcoded by hand.

To eliminate human curation entirely, Yi Cai 0005 et al. proposed an end-to-end framework in Power of Explanations: Towards automatic debiasing in hate speech detection. By leveraging a proxy metric called False Positive Proportion (FPP), the system automatically detects when a model is making "wrong decisions for wrong reasons"—such as associating benign cultural terms with toxicity—and corrects its behavior through a staged training process without any human-curated sensitive word lists.

[Automated Discovery Loop]
Freeze Model ──> Generate Prompts ──> Measure Output Disparity ──> Identify Biased Latent Concepts ──> Update Model Weights

When explicit lexical information is required, researchers are finding ways to make the lexicons dynamic. For example, Marion Bartl et al. developed a scalable methodology in Inferring Gender: A Scalable Methodology for Gender Detection with Online Lexical Databases that automatically infers lexical gender by parsing online dictionary definitions. This allows the bias detection framework to expand its vocabulary dynamically, matching the scale and evolution of natural language.

Why This Matters

Moving away from static word lists is not just about improving performance metrics; it fundamentally changes what it means to "debias" a model.

Static lists treat bias as a superficial vocabulary problem—as if scrubbing a few dozen words could erase systemic disparities. Dynamic and contextual approaches recognize that bias is embedded in the structural relationships, framing choices, and latent representations of language. By shifting the focus from individual tokens to holistic model behavior, these modern techniques offer a more robust, generalizable, and theoretically sound path toward algorithmic fairness.

Key Foundational Work

Linguistic Models for Analyzing and Detecting Biased Language (2013) — contributed 2 concept elements to this direction.
Active learning for detection of stance components (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_lexicon_dependent_bias_mitigation_7a32b1d4,
  title = {lexicon-dependent bias mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexicon-dependent-bias-mitigation/rdr_7a32b1d4b89248568fb5d02c5be4efa5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_lexicon_dependent_bias_mitigation_7a32b1d4_v1,
  title = {lexicon-dependent bias mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexicon-dependent-bias-mitigation/rdr_7a32b1d4b89248568fb5d02c5be4efa5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake