Concord begets concord: A Bayesian model of nominal concord typology

Kyle Mahowald, Dan Jurafsky, Mark NorrisProceedings of the Linguistic Society of America 2021 · DOI

Introduction

Linguistic complexity is often viewed through the lens of trade-offs, but new research suggests that in the realm of grammar, features may be mutually reinforcing. This study investigates nominal concord—the requirement for modifiers like adjectives, numerals, or demonstratives to agree with the nouns they modify in categories such as gender, number, or case. Using a Bayesian mixed-effect model, the authors test whether the presence of one type of concord makes others more or less likely.

The Competing Hypotheses

The research evaluates two primary theories regarding the statistical relationship between different types of concord within a language:

  1. Concord begets concord: Agreement is "sticky." If a language possesses the morphological machinery for one type of concord, it is statistically more likely to extend that machinery to other categories.
  2. A little concord goes a long way: Agreement is "repellent." To maintain system efficiency and minimize "costly" complexity, a language with one type of agreement might avoid others.

Methodology and Data

The study utilizes the Norris (2020) database, comprising 174 languages from 105 distinct families. This dataset was specifically curated to ensure geographic and genetic balance, avoiding the Eurocentric biases common in older typological samples.

Table 1: Distribution of nominal concord across major language families in the sample.

To analyze this data, the researchers employed a Bayesian hierarchical (mixed-effect) model implemented in Stan. By treating language family and geographic area as random effects, the model distinguishes between universal linguistic tendencies and similarities inherited from common ancestors. A key innovation involved using the model’s posterior draws to simulate thousands of hypothetical languages to calculate conditional probabilities.

Key Insights

The results provide overwhelming support for the "concord begets concord" hypothesis. The model demonstrates that linguistic agreement systems are self-reinforcing rather than compensatory.

  • The Power of Gender: Grammatical gender systems act as a powerhouse; their presence significantly increases the probability of every other type of concord.
  • Adjective-Demonstrative Symmetry: The study found a near-perfect correlation ($r=0.92$) between adjective and demonstrative concord, suggesting these loci are deeply linked in the mental grammar.
  • Word Order Neutrality: Contrary to some functionalist theories, the relative order of nouns and modifiers (e.g., whether an adjective precedes or follows a noun) does not significantly impact the likelihood of concord.

The model achieved a high degree of predictive power, with 92% accuracy for known languages and 87% for unseen families. It also highlighted significant "areal hotspots" where concord is most prevalent.

Table 2: Regional distribution of nominal concord in the study sample.

As shown in the data, the probability of a language having any form of concord is highest in Australia and Africa, while it is notably lower in South America and Papunesia. For an "average" language, the model-predicted probability of having any concord is 0.71 in Australia and 0.66 in Africa.

Conclusion

By applying modern statistical rigor to typological data, this research confirms that nominal concord is a cohesive grammatical phenomenon. Rather than being a zero-sum game of complexity, grammatical agreement tends to cluster, suggesting that once a language adopts the "logic" of concord, it is likely to apply it across its nominal system. Future research will look to integrate these findings with clausal syntax and evolutionary models of language change.

Summary updated

Abstract

Nominal concord is a phenomenon whereby nominal modifiers (e.g., adjectives, demonstratives, numerals) agree with their nominals along various dimensions (e.g., gender, number, case, definiteness). Here, drawing on a rich and typologically diverse database of nominal concord (Norris 2020), we build a Bayesian mixed effect model of nominal concord. Specifically, we consider two competing hypotheses regarding the statistical relationship between different types of concord within a language: (1) concord begets concord: the presence of some type of concord in a language makes it more likely that it has other types of concord vs. (2) a little concord goes a long way: if a language has some kind of concord, it is less likely to have other types of concord. We present evidence strongly in favor of the first hypothesis, that concord begets concord. Languages with nominal concord tend to have concord in more than one place and of more than one type. Using posterior draws from our model, we also provide quantitative evidence for a number of the tendencies described by Norris (2019a). Future work will build on this model to understand the functional role of nominal concord in language systems, how it evolves, and how it co-evolves with other typological features.
BibTeX
@inproceedings{mahowald2021concord,
  author = {Kyle Mahowald and Dan Jurafsky and Mark Norris},
  title = {{Concord begets concord: A Bayesian model of nominal concord typology}},
  booktitle = {Proceedings of the Linguistic Society of America},
  year = {2021},
  doi = {10.3765/plsa.v6i1.4988},
}
Figure