A Bayesian account of pronoun and neopronoun acquisition

A major challenge to equity among members of queer communities is the use of one’s chosen forms of reference, such as personal names or pronouns. Speakers often dismiss their misuses of pronouns as “unintentional”, and claim that their errors reflect many decades of fossilized mainstream language use, as well as attitudes or expectations about the relationship between one’s appearance and acceptable forms of reference. We argue for explicitly modeling individual differences in pronoun selection and present a probabilistic graphical modeling approach based on the nested Chinese Restaurant Franchise Process (nCRFP) (Ahmed et al., 2013) to account for flexible pronominal reference such as chosen names and neopronouns while moving beyond form-to-meaning mappings and without lexical co-occurrence statistics to learn referring expressions, as in contemporary language models. We show that such a model can account for variability in how quickly pronouns or names are integrated into symbolic knowledge and can empower computational systems to be both flexible and respectful of queer people with diverse gender expression.

💡 Research Summary

The paper tackles a pressing equity issue for queer communities: the respectful use of chosen names and pronouns, including emerging neopronouns. Existing natural‑language‑processing (NLP) systems typically treat pronouns as fixed form‑to‑meaning pairs learned from large corpora, which leads to systematic misgendering because non‑binary forms are under‑represented or absent from training data. To move beyond this limitation, the authors propose a probabilistic graphical model that directly assigns a distribution over referring expressions (pronouns, names, titles) to each individual, without relying on external characteristics such as appearance or gendered stereotypes.

The core of the approach is the nested Chinese Restaurant Franchise Process (nCRFP), a Bayesian non‑parametric construction that simultaneously models an unbounded number of “topics” (here, social communities) and an unbounded vocabulary (including novel pronouns). Each person is treated as a “document” composed of a sequence of observed pronoun/name tokens. A speaker’s prior over pronoun usage (global P) and a speaker‑specific prior for a target individual (P_t) are combined with the community‑level topic distribution (T_d) to generate the pronoun used in a given interaction (pro_t,i). The model updates these priors in a Bayesian fashion whenever new evidence is observed, allowing a single exemplar of a new pronoun to instantly alter the speaker’s distribution.

Two hierarchical levels are described. At the single‑speaker level (Figure 1), the model captures how a speaker’s personal experience with a referent shapes pronoun choice, including the possibility of intentional alternation or stability. The authors introduce the notion of “prior rigidity” to quantify how strongly a speaker’s existing distribution resists change—a useful parameter for modeling binary‑oriented individuals versus those more open to gender‑diverse forms. At the community level (Figure 2), multiple speakers share and influence each other’s priors. Communities with inclusive norms (e.g., queer‑friendly groups) will have flatter, more uniform priors over a wide set of pronouns, whereas traditional cis‑binary communities will allocate most probability mass to “he/she” and assign near‑zero mass to neopronouns. The hierarchical Bayesian framework naturally captures these inter‑community differences through the hyper‑parameters of the nCRFP (α, γ, etc.).

The authors contrast their approach with standard Latent Dirichlet Allocation (LDA). LDA assumes a fixed vocabulary and a predetermined number of topics, making it ill‑suited for handling novel pronouns that appear after model training. In contrast, nCRFP’s infinite‑vocabulary property lets new tokens be added on the fly, and its nested structure permits the emergence of new community‑specific topics that reflect evolving linguistic practices.

Simulation experiments illustrate three key findings: (1) the model can incorporate a newly observed pronoun after a single exposure, updating the individual’s distribution rapidly; (2) speakers from low‑rigidity (inclusive) communities adapt faster than those from high‑rigidity (binary) communities; (3) the magnitude of misgendering errors correlates with the rigidity parameter, confirming the model’s ability to capture real‑world variability in pronoun usage. These results suggest that a system built on this framework could dynamically adjust to a user’s self‑identified pronouns in real time, without waiting for periodic software releases.

Potential applications are discussed. In conversational agents, chatbots, or automated text generation, the model could maintain a per‑user pronoun distribution that is updated whenever the user explicitly states a preference or when the system observes consistent usage. This would dramatically reduce inadvertent misgendering, improve user trust, and align AI behavior with emerging social norms. Moreover, because the model treats pronouns symbolically rather than as dense word embeddings, it avoids the gender‑subspace biases that plague current large language models (LLMs).

The conclusion emphasizes that representing individuals as probability distributions over referring expressions provides a principled, flexible, and equitable way to handle gender‑diverse language. The paper bridges cognitive‑linguistic insights on pronoun processing with computational Bayesian modeling, offering a concrete pathway toward more inclusive NLP systems. Future work includes incorporating relational social factors (e.g., speaker‑listener intimacy, power dynamics), extending the model to other referential forms such as titles and honorifics, and validating the approach on real conversational corpora with longitudinal pronoun change.