First Name Genderedness Table

• First name genderedness tables are structured mappings that assign gender probabilities to names using data-driven metrics and curated classifications.
• They are constructed from sources like SSA data, Wikidata, and expert curation, using statistical methods such as conditional probabilities and genderedness indexes.
• These tables support demographic inference, fairness audits, and NLP tasks, while addressing temporal, cultural, and methodological challenges.

A first name genderedness table is a structured representation mapping each first name to the probability or bias with which it is associated with a given gender category, typically male, female, or—more recently—neutral categories. Such tables can be constructed from large labeled datasets, expert curation, algorithmic inference, or multi-source consensus, and underpin empirical research on gender prediction, demographic analysis, and downstream applications that rely on automated or statistical assignment of gender based solely on name information.

1. Core Definitions and Formulations

First name genderedness is operationalized by various metrics, the most prominent being conditional probabilities and absolute or comparative “genderedness scores,” derived from labeled datasets or annotated corpora.

• Probability-based assignment: The estimated probability that a given name $n$ is associated with a particular gender $g$ is denoted $P(g|n)$. For binary settings, $g \in \{\text{male}, \text{female}\}$; many systems now acknowledge a third ("neutral" or "unisex") category (You et al., 7 Jul 2024).
• Genderedness index: For frequency data, the absolute imbalance is defined as

$G(n) = \frac{|f_n - m_n|}{f_n + m_n} \in [0,1]$

where $f_n$ and $m_n$ denote number of female and male bearers (Sullivan et al., 2020).

• Relative frequency: The masculinity score, or the frequency-based probability, is

$$g(n) = \frac{\text{male\_count}(n)}{\text{male\_count}(n)+\text{female\_count}(n)}$$

as implemented in large-scale Wikidata-based tables (Sainte-Marie et al., 9 Dec 2025).

• MLE and entropy approaches: For probabilistic name-gender assignments,

$$P_\text{m}(n) = \frac{\text{count}_\text{m}(n)}{\text{count}_\text{f}(n) + \text{count}_\text{m}(n)}, \quad P_\text{f}(n) = 1 - P_\text{m}(n)$$

and genderedness $g(n) = |P_\text{m}(n) - P_\text{f}(n)|$ (Krstovski et al., 2023).

Many frameworks now include context-conditional or meta-learned consensus probabilities, taxonomic labels based on entropy thresholds, and reliability annotations (Buskirk et al., 2022).

2. Data Sources and Construction Schemes

The construction of genderedness tables varies by data source, demographic, and intended application:

• Government and administrative datasets: E.g., U.S. Social Security Administration (SSA), IBGE (Brazil), INSEE (France), and others publish first name plus gender-by-year frequency tables, serving as canonical sources for frequency-based genderedness (Sullivan et al., 2020, Misa, 2022).
• Aggregated multi-source datasets: Some methods, such as the Cultural Consensus Theory (CCT) approach, harmonize reports from dozens of open and commercial sources (e.g., global registers, Facebook, Wikidata) to robustly estimate $P(\text{female}|n)$ for over 100,000+ unique names (Buskirk et al., 2022, Sainte-Marie et al., 9 Dec 2025).
• Expert-validated and curated lists: For controlled experiments or fairness studies, names may be manually labeled by consensus of native speakers and cultural experts, explicitly excluding ambiguous or unisex names (Sakunkoo et al., 15 Apr 2025).
• Probabilistic machine learning models: ML-based predictors employ n-gram features, orthographic patterns, or embeddings to infer $P(g|n)$, frequently with explicit "unisex"/"ambiguous"/"unknown" output classes in addition to hard male/female assignments (Zhao et al., 2019, Hu et al., 2021, Mueller et al., 2016).

Table Source	Key Columns	Coverage Scope
SSA, IBGE, INSEE	Name, fₙ, mₙ, G(n)	Country, 50–100 years
Wikidata	Name, male_count, female_count, genderedness	Global, all time
CCT-based (meta)	Name, $P_\text{m}$, $P_\text{f}$, entropy	Global, multi-century
ML-based	Name, $P_\text{m}$, $P_\text{f}$, confidence	Data-dependent

3. Algorithmic Methodologies and Statistical Frameworks

Several methodological paradigms produce genderedness tables:

• Direct frequency estimation: Maximum-likelihood frequencies from large annotated name-gender datasets. Common in demography and computational social science (Krstovski et al., 2023, Sainte-Marie et al., 9 Dec 2025).
• Meta-learning/Cultural Consensus: EM-based procedures estimate a consensus label $z_m$ for each name (interpreted as $P(\text{female}|n)$), and a competence $c_n$ per source, iteratively updating both until convergence. Taxonomic labels ("strong female", "weakly gendered") are derived from entropy $H(z_m)$ (Buskirk et al., 2022).
• Naive Bayes over n-grams: For multilingual or unknown names, a character n-gram Naive Bayes classifier outputs probability-based predictions, with Laplace smoothing:

$$P(g \mid n) = \frac{P(g)\prod_{t \in F(n)}P(t \mid g)}{\sum_{g'}P(g')\prod_{t \in F(n)}P(t \mid g')}$$

(Zhao et al., 2019).

• Logistic regression and ML: Features such as character n-grams, TF-IDF-weighted vectors, and handcrafted orthographic measures inform regularized regression or SVMs, yielding probabilistic $P(g|n)$ and genderedness scores (Mueller et al., 2016, Hu et al., 2021).
• LLM-based approaches: Recent studies probe foundational and fine-tuned LLMs' predictions for male, female, and neutral-gender names, typically via softmax over three logits and $\arg\max$ for prediction; these models systematically underperform on gender-neutral names compared to binary ones and show English/non-English performance gaps (You et al., 7 Jul 2024).
• Contextual embedding projection: For occupation–gender studies, models compute the projection of a name embedding onto a learned “gender direction” vector, correlating with real-world $P_\text{f}(n)$ and supporting context-sensitive analysis (An et al., 9 Mar 2025).

4. Cultural, Temporal, and Linguistic Variation

The gender association of first names is highly context-sensitive:

• Temporal drift: Several names change gender association over time, e.g., "Leslie", "Shelby", "Courtney" shifted from predominantly male to female in the mid-20th century U.S. This dynamic is quantitatively captured by $G(n,t) = F(n,t)/[F(n,t)+M(n,t)]$ and illustrated by evaluating $\Delta G(n)$ across decades (Misa, 2022).
• Country and language effects: The same name may be strongly gendered in one country but ambiguous or differently gendered elsewhere (e.g., "Andrea" is male in Italy, female in the US; "Dominique" is neutral in France) (Buskirk et al., 2022, Sullivan et al., 2020).
• Morphological cues: In Turkish, patterns such as -gül and -nar suffixes mark femininity, whereas -arslan or historical names are highly male, quantifiable via log-frequency gender bias $G(n) = \log(P_m(n)/P_f(n))$ (Herdağdelen, 2017).
• Orthographic and phonological features: Statistical classifiers exploit features such as the count of final vowels or the presence of "bouba"/"kiki" phonemes to boost prediction accuracy (Mueller et al., 2016).

5. Practical Applications and Limitations

First name genderedness tables are deployed for:

• Demographic inference: Large-scale gender assignment in big data pipelines for sociology, bibliometrics, epidemiology, and bias auditing (Sainte-Marie et al., 9 Dec 2025, Krstovski et al., 2023).
• Bias detection and fairness auditing: Quantifying model and system-level gender disparities in LLMs, user interfaces, and recommender systems, including assessing the impact of alphabetical ordering or status hierarchies in algorithmic outputs (Sullivan et al., 2020, Sakunkoo et al., 15 Apr 2025, An et al., 9 Mar 2025).
• Natural language processing: Enabling downstream tasks such as pronoun resolution, user personalization, and coreference in linguistically-diverse settings (e.g., for Persian, Turkish, multilingual datasets) (Bijary et al., 14 Sep 2025, Herdağdelen, 2017).
• Historical and bibliometric studies: Merging genderedness scores with author metadata to study citation impact, productivity, and historical gender shifts in scholarly authorship (Sainte-Marie et al., 9 Dec 2025, Misa, 2022).

However, the methodology faces several limitations:

• Ambiguity and exclusion of unisex names: Datasets built on expert curation may explicitly exclude ambiguous names, sacrificing recall of real-world non-binary labeling (Sakunkoo et al., 15 Apr 2025).
• Temporal and contextual misassignment: Use of fixed present-day genderedness tables for historical data can misclassify names that underwent temporal drift, introducing systematic bias ("female shift" phenomenon) (Misa, 2022).
• Coverage and sparsity: Some country-level datasets apply frequency cutoffs or exclude rare names, reducing coverage and possibly underestimating the incidence of unisex names (Sullivan et al., 2020).

6. Representative Genderedness Table Structures

Across methodologies, the standard schema for a first-name genderedness table is as follows:

Name	Source(s)	P_male	P_female	Genderedness/Label
John	SSA, Wikidata	0.988	0.012	Strong male
Mary	SSA, Wikidata	0.004	0.996	Strong female
Alex	SSA, Wikidata	0.500	0.500	Ambiguous/Unisex
Leslie	SSA, Wikidata	0.200	0.800	High female association

• Additional columns may include total counts, entropy-based labels, consensus reliability, or contextual probabilities by country or decade (Sainte-Marie et al., 9 Dec 2025, Buskirk et al., 2022, Misa, 2022).
• Binarized tables from curated sources use $P(g|n) \in \{0, 1\}$, while probabilistic tables support continuous $P(g|n) \in [0, 1]$ predictions and taxonomic stratification.

7. Contemporary Developments and Research Directions

Recent advances focus on expanding beyond binary categories, incorporating gender-neutral and ambiguous classes to align with evolving sociotechnical realities (You et al., 7 Jul 2024). Fine-tuned models (e.g., BERT/RoBERTa) improve accuracy for neutral names but still lag significantly compared to binary settings, particularly for non-English names.

There is increasing emphasis on open, interpretable, and consensus-driven methodologies, as well as the need for temporal and cultural calibration to support fairness and accuracy in both social science and computational systems (Buskirk et al., 2022, Misa, 2022). Ongoing challenges include the responsible treatment of unisex names, privacy concerns, and the ethical handling of non-binary and transgender identities, which are not adequately captured in most extant tables.

In summary, first name genderedness tables are indispensable infrastructure for gender inference tasks, but their design, interpretation, and application demand rigorous attention to statistical, cultural, and ethical complexities (Sainte-Marie et al., 9 Dec 2025, Krstovski et al., 2023, You et al., 7 Jul 2024).