Textual Biases in NLP

• Textual biases are systematic distortions in language data and model outputs, shaped by lexical choices, co-occurrence patterns, and annotation artifacts.
• They span overt lexical biases to subtle multimodal and implicit style cues, impacting fairness and interpretability in NLP applications.
• Quantification methods like PMI, WEAT, and adversarial techniques provide actionable insights to mitigate these biases in language models.

Textual biases are systematic patterns in language data or model behavior where certain associations, outcomes, or representations are unevenly distributed across social, demographic, semantic, or modal dimensions. Such biases manifest both in human-generated texts and in the predictions or latent representations produced by trained models. Textual biases are not monolithic: they include direct lexical and co-occurrence biases, as well as subtle forms that arise from annotation artifacts, implicit style markers, multimodal integration failures, and the framing of factual content. Quantification and mitigation of textual bias are critical for fair, robust, and interpretable NLP systems.

1. Categories and Formalizations of Textual Bias

Lexical and First-order Co-Occurrence Biases

Lexical bias refers to the use of words or constructions whose sentiment, subjectivity, or connotation may slant the presentation of facts ("draconian" vs. "tough") (Fan et al., 2019). First-order co-occurrence bias arises when specific terms (e.g., gendered words) are statistically over- or under-represented in proximity to concepts of interest, as formalized by pointwise mutual information (PMI)-based metrics (Valentini et al., 2021, Borenstein et al., 2023).

Second-order and Embedding-based Biases

Second-order bias captures indirect associations, such as when embeddings or similarity structures absorb higher-order statistical relations (e.g., through skip-gram negative sampling or GloVe) (Valentini et al., 2021). Embedding-based bias metrics include the Word Embedding Association Test (WEAT), cosine-difference benchmarks, and effect sizes based on permutation or resampling (Borenstein et al., 2023).

Informational and Coverage Bias

Informational bias, distinct from lexical bias, results from the factual selection and framing of content (e.g., which events or entities are mentioned) (Fan et al., 2019). This requires modeling bias at the sentence, span, or document level, often by comparing across sources or by extracting spans associated with framing effects.

Implicit Bias and Style

Implicit bias refers to differences arising from latent style markers or authorial cues, even in the absence of explicit group identifiers. Such biases manifest where models systematically favor or penalize outputs associated with certain social groups, not due to content but due to correlated linguistic patterns (Liu et al., 2021).

Annotation Artifacts and Dataset-Induced Biases

Crowdsourced or protocol-driven data collection can inject artifacts such as superficial lexical cues strongly predictive of labels (e.g., “nobody” → contradiction in NLI), leading to inflated model performance and spurious correlations (Tan et al., 2019, Clark et al., 2019).

Modality and Multimodal Integration Bias

In vision-language systems, "textual bias" refers to the over-reliance on language modalities at the expense of visual input, as in failures of VLMs to ignore images and base decisions on text alone (Restrepo et al., 31 Jul 2025, Wang et al., 2 Apr 2025). Such biases can be exposed by selective modality shifting or adversarial prompts.

2. Quantifying and Modeling Textual Bias

First-order PMI-based Metrics

Bias across groupings (e.g., gender, race) is quantified by

$$\operatorname{Bias}_{\mathrm{PMI}}(C;A,B) = \operatorname{PMI}(A,C) - \operatorname{PMI}(B,C)$$

with

$$\operatorname{PMI}(X,Y) = \log \frac{P(X,Y)}{P(X)P(Y)}$$

BiasPMI can be re-expressed as a log conditional-probability ratio, and under data sparsity, as a log-odds ratio, enabling standard error and confidence interval estimation: $$SE = \sqrt{\frac{1}{f_{A,C}}+\frac{1}{f_{B,C}}+\frac{1}{f_{A,nC}}+\frac{1}{f_{B,nC}}}$$ and

$$\mathrm{CI}_{95\%} = [ \operatorname{Bias}_{\mathrm{PMI}} \pm 1.96 \cdot SE ]$$

(Valentini et al., 2021).

Embedding-based and WEAT Metrics

Association scores for target w and attribute sets A, B: $$\text{bias}(w,A,B) = \cos(v_w, v_A) - \cos(v_w, v_B)$$ The WEAT effect size is

$$d = \frac{ \operatorname{mean}_{x\in X} s(x,A,B) - \operatorname{mean}_{y\in Y} s(y,A,B)}{\operatorname{std}_{w\in X\cup Y} s(w,A,B)}$$

Statistical significance assessed via permutation tests (Borenstein et al., 2023).

Relative/Predictive Scoring

Pairwise ranking models (e.g., Bradley-Terry frameworks) learn interpretable bias scores via supervised comparison of document pairs labeled as more or less biased (Suresh et al., 2023).

Annotation Artifact Quantification

Bias in NLI is revealed by "hypothesis-only" classification: in SNLI, accuracy rises to 64% (chance: 33.3%) using only the hypothesis with no premise, exposing strong lexical-label associations (Tan et al., 2019).

3. Empirical Analysis and Case Studies

Table: Empirical Categories of Textual Bias in NLP Research

Bias Type	Core Mechanism	Exemplary Method or Metric
PMI/Co-occurrence	Asymmetric context-target associations	log odds ratio, PMI
Embedding/Second-Order	Similarity in vector space	WEAT, cosine-difference
Informational	Framing via content selection	Span annotation, sequence tagging
Implicit (Style/Author)	Latent demographic style patterns	Saliency overlap, adversarial training
Annotation Artifact	Protocol-induced label predictors	Hypothesis-only accuracy
Modality/Multimodal	Overweighting of text in VLMs	Selective Modality Shifting

Interpretable Example (Gender Bias, Wikipedia Nursing Context)

BiasPMI(nurse; female,male) = 1.3172 implies "nurse" is 3.73x more likely in female context than male (+273%) (Valentini et al., 2021).

Historical and Intersectional Biases

Temporal analysis of 18th–19th-century Caribbean newspapers via embeddings and PMI demonstrates distinct, non-additive compound biases at the intersection of gender and race; e.g., "elderly" strongly biased toward non-white females, not predictable from additive gender and race effects alone (Borenstein et al., 2023).

Healthcare and Social Stereotype Biases

Analysis of RedPajama-Data-1T and related corpora reveals a 3.79-fold overrepresentation of Black race mentions with disease concepts compared to US population share. Gender markers are much more frequently associated with diseases than race markers across web-scale text corpora (Hansen et al., 2024).

Implicit Bias in Author Style

Text classifiers trained on group-imbalanced data produce divergent false-positive/negative rates for demographic subgroups even absent explicit identity terms. Mitigation via adversarial feature-corrector layers that penalize reliance on demographic-style tokens sharply reduces demographic parity disparities while maintaining accuracy (Liu et al., 2021).

Informational vs. Lexical Media Bias

Span-level annotations in political news indicate that informational bias (selection/framing) is almost three times more frequent than lexical bias (word choice), and largely eludes standard span-level classification (BERT token-level F1=18.7% vs. lexical F1=26%) (Fan et al., 2019).

Modality Bias in Vision-Language Systems

Systematic text dominance in clinical VLMs (accuracy drop >20 pp under text-swap, NFR > 0.60) demonstrates strong textual shortcutting, with attention focused on text-tokens even when images are diagnostic. Architectural mitigation strategies (cross-modal penalties, co-training objectives) are needed to enforce true visual integration (Restrepo et al., 31 Jul 2025).

4. Mitigation Strategies and Algorithmic Advances

Adversarial, Ensemble, and Debiasing Frameworks

• Ensemble-based learning: Training a "bias-only" model to capture dataset-dependent cues, and a robust ensemble partner forced to focus on non-bias features (Bias-Product, Learned-Mixin, entropy penalty), delivering up to +9 pp OOD gain in textual entailment (Clark et al., 2019).
• Adversarial correction: Gradient reversal between task and group-adversary objectives to upweight task-relevant and downweight demographic-style tokens (Debiased-TC), yielding DPD drops from 32.52%→0.73% in sentiment tasks (Liu et al., 2021).
• Mask-shift de-biasing: Tagging and masking bias-laden tokens then iterative MLM filling, guided by bias-probability minimization (Dbias pipeline), achieving Disparate Impact (DI) ≈ 1.0 and G-AUC=0.78 (Raza et al., 2022).

Corpus-level and Experimental Interventions

• Dataset pruning: Algorithmic removal of low-entropy, label-predictive examples reduces hypothesis-only accuracy (64%→56%) and forces models to use truly discriminative features (Tan et al., 2019).
• Representation balancing: Quantifying and rebalancing disease–demographic associations to match real-world prevalence, with context-aware filtering and counterfactual fine-tuning (Hansen et al., 2024).
• Temporal and intersectional diagnostics: Embedding-based WEAT and groupwise PMI time-series illuminate the evolution and compounding of bias across underexplored axes (Borenstein et al., 2023).

Architectural Remedies for Multimodal Bias

• Fusion modifications addressing late-stage text-attention dominance in VLMs, with early fusion and contrastive pretraining to integrate visual-semantic features (Wang et al., 2 Apr 2025).
• Decoding-by-perturbation (DeP): Applying controlled textual perturbations, measuring attention variance and logit drift to counteract co-occurrence and language priors at generation time (Jia et al., 14 Apr 2026).

5. Societal and Practical Implications

Downstream Harms and Fairness Concerns

Textual bias in large-scale training data is directly inherited by LLMs and downstream systems. In healthcare, disease–demographic co-occurrence skews (e.g., “Black” overrepresentation) risk reinforcing stereotypes, with clinical recommendations or risk predictions reflecting biased text statistics rather than epidemiological realities (Hansen et al., 2024). Similar compounding is evident for gender, race, and intersectionality in both historical and contemporary corpora (Borenstein et al., 2023).

Interpretability and Transparency

Metrics such as PMI-based bias measures, relative pairwise scoring, and word-level attributions provide interpretable, first-order diagnostics for practitioners and editors. The explicit mapping of bias to example spans or tokens is foundational for actionable remediation—whether through editorial review, dataset curation, or automated de-biasing flows (Suresh et al., 2023, Fan et al., 2019).

Content Moderation and Model Robustness

Architectural fixation on textual tokens in VLMs enables adversarial attacks—e.g., masking hate speech in ASCII art—calling for advances in multimodal embedding fusion and bias-aware content filtering (Wang et al., 2 Apr 2025).

6. Open Challenges and Research Directions

Contextual and Higher-order Bias Modeling

Detection of informational and coverage bias—"what is left unsaid"—remains challenging due to the lack of modeling context across documents and sources. Multi-task models are needed to jointly surface lexical, informational, and intersectional bias at various text granularities (Fan et al., 2019).

Extension Beyond Binary and Sparse Axes

Current bias benchmarks often focus on binary gender or race, neglecting continuous, intersectional, and non-binary identities. Richer annotation and corpus construction, guided by intersectionality theory, are required for nuanced benchmarks (Borenstein et al., 2023).

Inference-time Correction and Fair Selection under Bias

Efficient best-alternative selection when automated judgment is cheap but biased, and human audit is costly, is an emerging problem. PP-LUCB algorithms and anytime-valid confidence sequences allow cost-effective identification of optimal configurations under instance-dependent, arm-dependent textual bias (Ao et al., 11 Mar 2026).

Evaluative and Societal Feedback Loops

Evaluating bias correction in practice requires not only technical metrics (e.g., Disparate Impact, Equalized Odds, G-AUC) but also consideration of long-term societal impacts and user-centered perceptions—such as how readers interpret biased annotations in data visualizations (Stokes et al., 2024).

Tooling and Standardization

Publicly available packages (e.g., Dbias), code, and detailed annotation protocols are essential for reproducible experimentation and deployment in both media and clinical/NLP applications (Raza et al., 2022, Hansen et al., 2024).

---

Textual biases encompass a spectrum of phenomena at the interface of corpus composition, model architecture, annotation protocol, and real-world deployment. Their systematic identification, quantification, and mitigation are foundational challenges in the development of fair, interpretable, and robust language and multimodal systems.