A Bayesian account of pronoun and neopronoun acquisition

Cassandra L Jacobs, Morgan GrobolNAACL 2025

TL;DR

This paper proposes a novel Bayesian probabilistic model to understand how individuals learn and adapt to new pronouns and names, particularly focusing on neopronouns and the experiences of queer individuals. By moving beyond simple form-to-meaning mappings, this approach aims to create more flexible and respectful AI systems that can accommodate diverse gender expressions.

What Problem Does This Paper Address?

The paper tackles a significant challenge in natural language processing (NLP) and computational linguistics: accurately and respectfully representing individuals' chosen forms of reference, especially personal names and pronouns. While conventional words have fixed meanings, names and pronouns are fluid and tied to personal identity, particularly for queer individuals who may use neopronouns or have evolving gender expressions. Current NLP systems often struggle with this variability, perpetuating biases and failing to accommodate non-binary identities. Dismissing misuses of pronouns as mere "unintentional" errors or blaming fossilized language use overlooks the dynamic nature of language and individual agency. The authors aim to bridge this gap by developing a computational model that respects individuals' right to self-determination in how they are referred to, moving beyond rigid form-to-meaning or form-to-feature mappings.

Methods

The core of the paper's contribution lies in its proposed probabilistic graphical modeling approach, specifically building upon the nested Chinese Restaurant Franchise Process (nCRFP). This framework allows for flexible pronominal reference learning without relying on lexical co-occurrence statistics or direct form-to-meaning mappings, which are common in contemporary language models.

The authors propose a single-speaker model that represents a speaker's evolving preferences for referring to a specific individual across interactions. This model incorporates variables such as:

  • prod: The pronoun produced for a referent.
  • Ud,i: The utterance containing the reference.
  • P: The speaker's general prior on pronoun production.
  • Pt: The speaker's prior on a specific individual's pronouns.
  • Td: The topic of the interaction.
  • pro: Pronoun usages witnessed by the speaker.

Crucially, this model allows for:

  1. Introduction of new forms: The vocabulary can expand to include novel names and neopronouns.
  2. Mixture of forms: Individuals can use a mix of pronouns, even alternating between them.
  3. Rapid adaptation: The model can quickly update its understanding based on new information, even a single exemplar.
  4. Individual variation: Adaptation can vary based on an individual's prior experience and comfort with such changes.

They extend this to a community model where individual preferences are influenced by community norms. This model treats pronouns as observable objects with probabilistic assignments to "topics" (communities of individuals), allowing for the learning of community-specific pronoun usage patterns.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and modeling approach rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Pronoun learning is dynamic and individual: The paper emphasizes that learning pronoun usage isn't a static process of mapping words to fixed meanings. Instead, it's an ongoing, adaptive process influenced by individual experience and community norms.
  • Respecting self-determination: The model prioritizes an individual's right to define how they are referred to, moving away from systems that infer identity from external characteristics.
  • Beyond form-to-meaning: By using a Bayesian approach and nCRFP, the authors propose a way to learn referring expressions that doesn't rely on traditional methods like word embeddings or co-occurrence statistics, which can encode biases.
  • Neopronouns as a natural extension: The framework is designed to accommodate the introduction and learning of novel pronouns, making it relevant for understanding the evolving landscape of gender expression.
  • Community influences individual priors: The model highlights how social interactions and community beliefs shape an individual's expectations and usage of pronouns.

Key Results

The paper presents a theoretical framework and a probabilistic model. It does not report quantitative results from experiments or evaluations on specific datasets. The "results" are the conceptual advancements and the proposed modeling approach itself.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Interaction dynamics: The current model simplifies certain interaction dynamics. For instance, it doesn't explicitly model how the "weight" or credibility of a pronoun usage might vary based on the speaker's relationship to the referent (e.g., closeness). The case where a speaker refers to themselves (s = t) is included but could be given more specific treatment.
  • Scope of referring expressions: The model primarily focuses on pronouns. The authors note that pronoun usage is often informed by other referring expressions like names, titles, and terms of address. Future work could integrate these into a richer model.
  • Non-linguistic social dynamics: The community model does not explicitly incorporate non-linguistic social factors that might influence pronoun usage.
  • Complexity of priors: While the model folds other referring expressions into priors (P), a more detailed examination of their internal structure could lead to a more nuanced model.

Future work could explore more sophisticated models that balance increased complexity with improved insights, particularly in capturing the nuances of social influence and the interplay of various referring expressions.

Summary updated

BibTeX
@inproceedings{jacobs2025bayesian,
  author = {Cassandra L Jacobs and Morgan Grobol},
  title = {{A Bayesian account of pronoun and neopronoun acquisition}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure