Computational Typology

Gerhard Jäger

cs.CLApr 22, 2025

Computational Typology

Authors: Gerhard Jäger

Organizations: University of Tübingen, Seminar für Sprachwissenschaft, Keplerstr. 2, 72074 Tübingen, Germany

Abstract

Typology is a subfield of linguistics that focuses on the study and classification of languages based on their structural features. Unlike genealogical classification, which examines the historical relationships between languages, typology seeks to understand the diversity of human languages by identifying common properties and patterns, known as universals. In recent years, computational methods have played an increasingly important role in typological research, enabling the analysis of large-scale linguistic data and the testing of hypotheses about language structure and evolution. This article provides an illustration of the benefits of computational statistical modeling in typology.

Figures & tables

	Absolute (exceptionless)	Statistical (tendencies)
Unconditional (unrestricted)	Type 1. “Unrestricted absolute universals” All languages have property X	Type 2. “Unrestricted tendencies” Most languages have property X
Conditional (restricted)	Type 3. “Exceptionless implicational universals” If a language has property X, it also has property Y	Type 4. “Statistical implicational universals” If a language has property X, it will tend to have property Y

Table 1: Typology of linguistic universals (from Evans & Levinson , (2009) ), Table 1

Panel of Figure 1 — Figure 1: Correlations between linguistic and non-linguistic traits from Roberts & Winters ((2013)) . Left: Correlation between inflectional synthesis of verbs and siesta habits. Right: Correlation between presence of tone and presence of acacia trees. Sources: doi.org/10.1371/journal.pone.0070902.g005 and doi.org/10.1371/journal.pone.0070902.g008

Affix Type	Postposition	Preposition	% Preposition
Strongly suffixing	169	45	21
Weakly suffixing	30	26	46
Equal prefixing and suffixing	37	36	49
Weakly prefixing	17	26	60
Strongly prefixing	3	33	92

Table 2: Distribution of affix types across postpositions and prepositions, with preposition percentages.

Figure 2: Distribution of affix types across postpositions and prepositions.

Figure 3: Scatterplot of population size and phoneme inventory size.

Figure 4: Map of population size and phoneme inventory size.

$\displaystyle\rho$	$\displaystyle\sim\text{Uniform}(-1,1)$	(1)
$\displaystyle\sigma_{1},\sigma_{2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(2)
$\displaystyle\Sigma$	$\displaystyle=\begin{pmatrix}\sigma_{1}^{2}&\rho\sigma_{1}\sigma_{2}\\ \rho\sigma_{1}\sigma_{2}&\sigma_{2}^{2}\end{pmatrix}$	(3)
$\displaystyle z_{i}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma)$	(4)
$\displaystyle c_{1}$	$\displaystyle\sim\mathcal{N}(0,2)$	(5)
$\displaystyle\zeta_{k}$	$\displaystyle\sim\text{LogNormal}(0,2),\quad c_{k+1}=c_{k}+\zeta_{k}+10^{-2}$	(6)
$\displaystyle\alpha$	$\displaystyle\sim\mathcal{N}(0,2)$	(7)

Model 1: Bivariate normal model for latent variables in the affix–adposition association.

Parameter	Mean	SE_Mean	SD	2.5%	97.5%	n_eff	Rhat
cutpoints 1	0.08	0.00	0.33	-0.54	0.81	21692	1.00
cutpoints 2	1.72	0.02	1.05	0.67	4.57	2120	1.00
cutpoints 3	4.28	0.06	2.50	1.91	11.09	1958	1.00
cutpoints 4	6.69	0.09	3.88	3.06	17.19	1936	1.00
$\alpha$	-1.08	0.01	0.61	-2.73	-0.41	7092	1.00
$\sigma_{1}$	4.46	0.07	2.96	1.46	12.41	1869	1.00

Table 3: Posterior summary statistics for the bivariate model associating affix type and adposition type.

$\displaystyle\rho$	$\displaystyle\sim\text{Uniform}(-1,1)$	(11)
$\displaystyle\sigma_{1},\sigma_{2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(12)
$\displaystyle\Sigma$	$\displaystyle=\begin{pmatrix}\sigma_{1}^{2}&\rho\sigma_{1}\sigma_{2}\\ \rho\sigma_{1}\sigma_{2}&\sigma_{2}^{2}\end{pmatrix}$	(13)
$\displaystyle\mu$	$\displaystyle\sim\mathcal{N}(\mathbf{0},4\mathbf{I})$	(14)
$\displaystyle x_{i}$	$\displaystyle\sim\mathcal{N}(\mu,\Sigma)$	(15)

Model 2: Bivariate normal model for segment inventory size/population size association.

Parameter	Mean	SD	2.5%	97.5%	n_eff	Rhat
$\mu_{1}$	0.00	0.029	-0.057	0.057	20455	1.00
$\mu_{2}$	0.00	0.029	-0.057	0.057	20573	1.00
$\sigma_{1}$	1.00	0.021	0.962	1.043	21378	1.00
$\sigma_{2}$	1.00	0.021	0.962	1.042	21594	1.00
$\rho$	0.39	0.025	0.341	0.437	19679	1.00

Table 4: Posterior summaries for the Bayesian linear bivariate model modelling log-transformed phoneme inventory size and log-transformed population size.

$\displaystyle\rho_{f},\rho_{l}$	$\displaystyle\sim\text{Uniform}(-1,1)$	(17)
$\displaystyle\sigma_{f,1},\sigma_{f,2},\sigma_{l,1},\sigma_{l,2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(18)
$\displaystyle\Sigma_{f}$	$\displaystyle=\begin{pmatrix}\sigma_{f,1}^{2}&\rho_{f}\sigma_{f,1}\sigma_{f,2}\\ \rho_{f}\sigma_{f,1}\sigma_{f,2}&\sigma_{f,2}^{2}\end{pmatrix}$	(19)
$\displaystyle\Sigma_{l}$	$\displaystyle=\begin{pmatrix}\sigma_{l,1}^{2}&\rho_{l}\sigma_{l,1}\sigma_{l,2}\\ \rho_{l}\sigma_{l,1}\sigma_{l,2}&\sigma_{l,2}^{2}\end{pmatrix}$	(20)
$\displaystyle z_{f,j}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{f})\quad\text{for each family }j$	(21)
$\displaystyle z_{l,i}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{l})$	(22)
$\displaystyle z_{i}$	$\displaystyle:=z_{f,f_{i}}+z_{l,i}$	(23)

Model 3: Hierarchical bivariate model with family-level random effects for affix–adposition association.

f_{i}

denotes the family of language

i

Parameter	Mean	SE_Mean	SD	2.5%	97.5%	n_eff	Rhat
cutpoints 1	0.251	0.002	0.367	-0.421	1.027	24925	1.00
cutpoints 2	1.579	0.009	0.588	0.798	3.035	4244	1.00
cutpoints 3	3.522	0.020	1.098	2.393	6.450	3012	1.00
cutpoints 4	5.186	0.030	1.574	3.713	9.454	2832	1.00
$\alpha$	-2.937	0.004	0.728	-4.588	-1.739	30423	1.00
$\sigma_{f,1}$	2.929	0.016	0.920	1.938	5.346	3183	1.00

Table 5: Posterior summaries for the ordinal–binary bivariate model with family-level and language-level covariance. Cutpoints are used to map a latent continuous variable to ordinal affix positions; correlations are estimated separately at the family (

f

) and language (

l

) level.

$\displaystyle\rho_{f},\rho_{l}$	$\displaystyle\sim\text{Uniform}(-1,1)$	(30)
$\displaystyle\sigma_{f,1},\sigma_{f,2},\sigma_{l,1},\sigma_{l,2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(31)
$\displaystyle\Sigma_{f}$	$\displaystyle=\begin{pmatrix}\sigma_{f,1}^{2}&\rho_{f}\sigma_{f,1}\sigma_{f,2}\\ \rho_{f}\sigma_{f,1}\sigma_{f,2}&\sigma_{f,2}^{2}\end{pmatrix}$	(32)
$\displaystyle\Sigma_{l}$	$\displaystyle=\begin{pmatrix}\sigma_{l,1}^{2}&\rho_{l}\sigma_{l,1}\sigma_{l,2}\\ \rho_{l}\sigma_{l,1}\sigma_{l,2}&\sigma_{l,2}^{2}\end{pmatrix}$	(33)
$\displaystyle\mu$	$\displaystyle\sim\mathcal{N}(\mathbf{0},4\mathbf{I})$	(34)
$\displaystyle z_{f,j}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{f})\quad\text{for each family }j$	(35)
$\displaystyle z_{l,i}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{l})$	(36)

Model 4: Bivariate model with family-level random intercepts for segment inventory/population size association.

Parameter	Mean	SE_Mean	SD	2.5%	97.5%	n_eff	Rhat
$\mu_{1}$	-0.494	0.001	0.063	-0.619	-0.370	2797	1.00
$\mu_{2}$	-0.325	0.001	0.072	-0.469	-0.182	2506	1.00
$\sigma_{l,1}$	0.668	0.000	0.015	0.640	0.698	25769	1.00
$\sigma_{l,2}$	0.658	0.000	0.015	0.630	0.687	26397	1.00
$\sigma_{f,1}$	0.693	0.000	0.052	0.597	0.800	10667	1.00
$\sigma_{f,2}$	0.860	0.001	0.062	0.746	0.986	13196	1.00

Table 6: Posterior summaries for the extended bivariate model with both residual- and family-level covariance components.

Figure 5: Top: Phylogenetic tree of the world’s languages. The tree was pruned to the 422 languages of the affix–adposition study. Bottom: Subtree for the Indo-European languages.

$\displaystyle\rho_{f},\rho_{l}$	$\displaystyle\sim\text{Uniform}(-1,1)$	(41)
$\displaystyle\sigma_{f,1},\sigma_{f,2},\sigma_{l,1},\sigma_{l,2},\lambda_{1},\lambda_{2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(42)
$\displaystyle\Sigma_{f}$	$\displaystyle=\begin{pmatrix}\sigma_{f,1}^{2}&\rho_{f}\sigma_{f,1}\sigma_{f,2}\\ \rho_{f}\sigma_{f,1}\sigma_{f,2}&\sigma_{f,2}^{2}\end{pmatrix}$	(43)
$\displaystyle\mu_{\text{adp}}$	$\displaystyle\sim\mathcal{N}(0,2),\quad\mu=(0,\mu_{\text{adp}})$	(44)
$\displaystyle z_{f,j}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{f})\quad\text{for each family }j$	(45)
$\displaystyle w_{\text{root}}$	$\displaystyle\sim\mathcal{N}(\mu,S(\infty))$	(46)
$\displaystyle w_{c}\mid w_{p}$	$\displaystyle\sim\mathcal{N}\left(\mu+e^{-\Lambda t_{c}}(w_{p}-\mu),\ S(t_{c})\right)\quad\text{for each branch }p\to c$	(47)

Model 5: Phylogenetic correlation model for affix–adposition association.

w

is the pair of latent traits at the nodes of the tree,

t_{c}

the length of the branch leading to node

c

\Lambda=\mathrm{diag}(\lambda_{1},\lambda_{2})

, and

w_{i}

the value at the tip for language

i

Parameter	Mean	SE_Mean	SD	2.5%	97.5%	n_eff	Rhat
cutpoints 1	-0.192	0.006	1.216	-2.722	2.226	37571	1.00
cutpoints 2	2.570	0.026	1.646	-0.148	6.330	3864	1.00
cutpoints 3	6.672	0.064	2.960	2.802	13.797	2150	1.00
cutpoints 4	10.290	0.097	4.269	5.033	20.706	1947	1.00
$\mu_{\text{adp}}$	-1.076	0.010	1.909	-4.699	2.736	35442	1.00
$\sigma_{l,1}$	2.552	0.028	1.239	1.035	5.579	1936	1.00

Table 7: Posterior summaries for the bivariate Ornstein–Uhlenbeck model with family-level correlation

\rho_{f}

, correlation

\rho_{l}

between the random changes of the two latent traits, and mean-reversion rates

\lambda_{1},\lambda_{2}

. Cutpoints map the latent variable for the ordinal response to observed categories.

Figure 6: Ancestral state reconstructions for the Indo-European sub-tree. The left panel shows the reconstructed values for the affix latent variable, while the right panel shows the reconstructed values for the adposition latent variable. Markers at the leaves indicate observed features.

$\displaystyle\rho_{f},\rho_{l}$	$\displaystyle\sim\text{Uniform}(-1,1)$	(55)
$\displaystyle\sigma_{f,1},\sigma_{f,2},\sigma_{l,1},\sigma_{l,2},\lambda_{1},\lambda_{2}$	$\displaystyle\sim\text{LogNormal}(0,1)$	(56)
$\displaystyle\Sigma_{f}$	$\displaystyle=\begin{pmatrix}\sigma_{f,1}^{2}&\rho_{f}\sigma_{f,1}\sigma_{f,2}\\ \rho_{f}\sigma_{f,1}\sigma_{f,2}&\sigma_{f,2}^{2}\end{pmatrix}$	(57)
$\displaystyle\mu$	$\displaystyle\sim\mathcal{N}(\mathbf{0},4\mathbf{I})$	(58)
$\displaystyle z_{f,j}$	$\displaystyle\sim\mathcal{N}(\mathbf{0},\Sigma_{f})\quad\text{for each family }j$	(59)
$\displaystyle w_{\text{root}}$	$\displaystyle\sim\mathcal{N}(\mu,S(\infty))$	(60)
$\displaystyle w_{c}\mid w_{p}$	$\displaystyle\sim\mathcal{N}\left(\mu+e^{-\Lambda t_{c}}(w_{p}-\mu),\ S(t_{c})\right)\quad\text{for each branch }p\to c$	(61)

Model 6: Phylogenetic correlation model for segment inventory/population size association.

x_{i}

is the pair of standardized log population size and log inventory size of language

i

;

w

t_{c}

and

\Lambda

are as in Model 5 .

Parameter	Mean	SE_Mean	SD	2.5%	97.5%	n_eff	Rhat
$\sigma_{l,1}$	0.483	0.000	0.020	0.446	0.524	2605	1.00
$\sigma_{l,2}$	0.494	0.001	0.022	0.453	0.540	1929	1.00
$\lambda_{1}$	0.215	0.001	0.029	0.160	0.275	3084	1.00
$\lambda_{2}$	0.226	0.001	0.033	0.165	0.297	2398	1.00
$\mu_{1}$	-0.474	0.001	0.118	-0.708	-0.238	13542	1.00
$\mu_{2}$	-0.327	0.001	0.118	-0.563	-0.095	14491	1.00

Table 8: Posterior summaries for parameters in the bivariate phylogenetic OU model with family-level correlation. The 95% credible interval of the correlation

\rho_{l}

between the random changes of the two variables includes 0.

Model Type	Affix–Adposition		Segment Inventory–Population Size
	Bayes factor (log)	$\Delta$ elpd	Bayes factor (log)	$\Delta$ elpd
Vanilla Model	$-196$	$-273$	$-752$	$-933$
Hierarchical Model	$-27$	$-59$	$-133$	$-207$
Phylogenetic Model	$0$	$0$	$0$	$0$

Table 9: Model comparison for both case studies. Columns 2–3 correspond to the affix–adposition case study, and columns 4–5 to the segment inventory–population size case study. Higher elpd and log-Bayes values indicate better model fit. All values are differences to the phylogenetic model. The standard errors of the elpd differences are 16.4 (vanilla) and 7.9 (hierarchical) for the affix–adposition models, and 38.6 and 24.7 for the segment inventory–population size models; the Monte Carlo standard errors of the log marginal likelihoods are below 0.3.

Linguistic Typology

Linguistic typology compares languages to identify patterns in their sounds, grammar, meaning, and history.

See all 26 papers

Sep 7, 2026

LLM Agents as Computational Typologists

Linguistic typology relies on expert analysis of reference grammars across languages, making large-scale crosslinguistic comparison labor-intensive and unscalable. We introduce AUTOTYPOLOGIST, an LLM agent for evidence-grounded typological analysis over reference grammars. The agent is capable of retrieving relevant grammar sections, analyzing interlinear glossed text (IGT), and iteratively reasoning over typological hypotheses using a ReAct-style workflow. We evaluate the system on TYPOLOGICAL FEATURE CODING against expert annotations and TYPOLOGICAL HYPOTHESIS TESTING with typological universals using 25 open-source reference grammars. Operating under different information constraints in TYPOLOGICAL FEATURE CODING, the agent can synthesize information from reference grammar prose but still faces challenges with only IGTs in the target language. In TYPOLOGICAL HYPOTHESIS TESTING, the agent can synthesize crosslinguistic evidence and identify both supporting cases and counterexamples. These findings suggest that LLM agents can support scalable and inspectable typological analysis, while still requiring expert validation.

Changbing Yang, Christopher Hammerly, Freda Shi +1

University of British Columbia · University of Waterloo · Vector Institute

LLM Agents Linguistic Typology

Ask PDF

Sep 30, 2026

Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies

Cross-lingual transfer describes how knowledge in a source language benefits a target language. Measuring it quantitatively requires broad multilingual pre-training, as prior work has done with cross-lingual transfer matrices. We ask whether transfer is predictable from freely available typological features, and whether the prominence of high-resource source languages reflects typology or data quality and quantity. We show that typological databases contain cheap and dense signals about cross-lingual transfer. Our typology-only random forest on a 24-language prior-work transfer matrix scores leave-one-language-out

ρ{=}0.705

and

R^2{=}0.49

, beating a non-typological control at

ρ{=}0.62

, which verifies the ability of typology-only predictions to reconstruct costly measured cross-lingual transfer. The signal survives leave-one-script-out and leave-one-family-out protocols, so script and family confounding do not explain the effect. By decomposing the transfer into a typology term and a resource-and-script bias term, we find the best-source ranking sensitive to this bias. In contrast, typology is not affected by this bias, which makes it a zero-compute screening tool that replaces hundreds of training runs with a model fit. Our code is available here.

Dalton Raphael Harmsen, Swier Garst, Thomas van Osch +2

AMOR/e Lab, Eindhoven University of Technology, Eindhoven, The Netherlands · SURF, Amsterdam, The Netherlands

Task Transferability Estimation Linguistic Typology Cross-Lingual Transfer

Ask PDF

Sep 3, 2026

Typological Feature Prediction with Large Language Models: An In-Context Learning Approach

Typological features are widely used in multilingual NLP, and the prediction of such features holds downstream utility. However, existing methods to predict missing values lack interpretable justifications for predictions, while their performance across resource levels and feature types remains underexplored. Given LLMs' abilities in meta-linguistic reasoning and in providing rationales, we investigate LLMs' performance in typological feature prediction via an in-context learning approach with linguistic data from URIEL+ and Glottolog. We find that zero-shot prompting is insufficient, but when given phylogenetic and geographic neighbour evidence, LLMs substantially outperform all baselines without disadvantaging low-resource languages. We further find that most LLM rationales are consistent with the provided evidence, offering a step toward explainable typological feature prediction.

Qianwen Wang, York Hay Ng, Aditya Khan +1

University of Toronto, Canada · Columbia · Ontario Tech University, Canada

In-Context Learning Linguistic Typology

Ask PDF

Abstract

Linguistic Typology

Explore similar work

LLM Agents as Computational Typologists

Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies

Typological Feature Prediction with Large Language Models: An In-Context Learning Approach