A Modular Architecture for Typologically Controlled Lexicon Generation

Sankalp Tattwadarshi Swain; Dhruv Kumar

cs.CLApr 7, 2026

A Modular Architecture for Typologically Controlled Lexicon Generation

Authors: Sankalp Tattwadarshi Swain, Dhruv Kumar

Organizations: Birla Institute of Technology and Science, Pilani Pilani, Rajasthan 333031, India

Abstract

Constructing artificial lexicons that are pronounceable, typologically plausible, and semantically structured remains an open challenge in computational linguistics. Existing conlang generators either lack formal phonotactic guarantees or delegate generation to opaque, non-reproducible LLM-based pipelines. We propose a modular framework that samples phoneme inventories from PHOIBLE, generates word forms under interchangeable phonological grammars (deterministic, OT, and MaxEnt), and assigns meanings via a Swadesh--Leipzig--Jakarta ontology with explicit form--meaning alignment. Evaluation on character $n$ -gram perplexity, log-likelihood, and KL divergence against PHOIBLE across lexicon sizes of 100-5,000 forms shows that probabilistic grammars consistently outperform deterministic and random baselines on both phonotactic coherence and typological realism.

Figures & tables

System	Inventory Source	Word Form Generation	Phonotactics	Semantics	Typological Grounding
Heyer (2021)	User-defined templates	Template sampling	Hard rule filters (user-specified)	None	None
Cai & Martens (2023)	Designer-specified parameters	Syllable template expansion	Configurable rules; no formal grammar	None	None
Goldwater & Johnson (2003)	Real language phoneme sets	MaxEnt-ranked OT candidates	MaxEnt over ranked OT constraints	None	None
Hayes & Wilson (2008)	Real phoneme inventories	MaxEnt-weighted candidate selection	Maximum Entropy with learned constraint weights	None	None
Futrell et al. (2017)	Real lexicons (14 languages)	Bayesian generative model (stochastic memoization)	Probabilistic; phonological feature hierarchy	None	Multi-lingual; no typological sampling
Pimentel et al. (2020)	Real lexicons (106 languages)	Not generative; complexity measurement only	LSTM / trigram LM; bits-per-phoneme measure	None	Cross-linguistic comparison; no inventory sampling

Table 1: Comparison of lexicon generation systems across five dimensions: inventory source, word form generation, phonotactics, semantics, and typological grounding. Our system is the only one that unifies all five within a single pipeline.

Relationship	Statistic	Value	p -value
Consonant count vs. ejective presence	$r$	$0.324$	$<0.001$
Consonant count vs. click presence	$r$	$0.188$	$<0.001$
Inventory size vs. click presence	$r$	$0.188$	$<0.001$
Vowel count vs. vowel length contrast	$r$	$0.412$	$<0.001$
Ejective $\leftrightarrow$ uvular	$\chi^{2}$	$0.000$	$1.000$
Pharyngeal $\leftrightarrow$ uvular	$\chi^{2}$	$0.000$	$1.000$

Table 2: Pearson correlation coefficients (

r

) and chi-square statistics (

\chi^{2}

) computed over PHOIBLE inventories. Correlation values measure symmetric associations between inventory size and marked segment presence; chi-square values reflect feature co-occurrence tests. All correlation results are significant at

p<0.001

Implication	Total $X$	Violations	$P(Y\mid X)$	Failure Rate
Pharyngeal $\Rightarrow$ uvular	$15$	$0$	$1.000$	$0.00000$
Nasal vowel $\Rightarrow$ oral vowel	$688$	$0$	$1.000$	$0.00000$
Voiced obs. $\Rightarrow$ voiceless obs.	$2355$	$6$	$0.997$	$0.00255$
Fricative $\Rightarrow$ stop	$3020$	$0$	$1.000$	$0.00000$
Front rounded $\Rightarrow$ front unrounded	$273$	$0$	$1.000$	$0.00000$

Table 3: Conditional probability results for the five implicational universals incorporated into the sampler. Total

X

denotes the number of inventories in which the antecedent feature is present; Violations counts cases where the consequent feature is absent;

P(Y\mid X)

is the empirical conditional probability; and Failure Rate is the complementary proportion. Four of the five implications are fully categorical (

P=1.0

); the voiced–voiceless obstruent implication is a near-universal (

P=0.997

Figure 1: Overview of the lexicon generation pipeline. PHOIBLE-sampled inventories feed four interchangeable phonological grammars; the generated word space is then mapped to meanings via the semantic assignment module.

Panel of Figure 2 — Figure 2: Phonotactic well-formedness as a function of lexicon size. OT and MaxEnt consistently outperform the deterministic and random baselines across all lexicon sizes on both metrics.

Figure 3: KL divergence $D_{\mathrm{KL}}(P\,\|\,Q)$ between the generated phoneme distribution and the PHOIBLE global reference, as a function of lexicon size. Lower values indicate greater typological realism.

Panel of Figure 4 — Figure 4: Cross-grammar evaluation matrices. Rows denote the training grammar; columns denote the test grammar. Diagonal entries reflect within-grammar fit; off-diagonal entries reveal distributional transferability across formalisms.

Linguistic Typology

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

See all 26 papers

Self-Supervised Lexical Representation Learning for Fast, Large-Scale Phylogenetic Inference
Sep 4, 2026Tim WientzekPhylogenetic Inference Self-Supervised Learning
YallaMorph: A Benchmark for Evaluating Arabic Morphological Generation in Large Language Models
Sep 9, 2026Mahmoud Reda, Salam Khalifa, Reham Marzouk +1Multilingual Language Model Evaluation Arabic NLP
New York University Abu Dhabi · Stony Brook University · Mohamed bin Zayed University of Artificial Intelligence
Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach
Sep 23, 2026MinJu Jeon, Younghan Park, Han Sung Park +3G2P Conversion
NAVER Cloud · Hanyang University · Carnegie Mellon University +1

Sep 4, 2026J/K move · Enter open · S save

Self-Supervised Lexical Representation Learning for Fast, Large-Scale Phylogenetic Inference

Tim Wientzek

Phylogenetic Inference Self-Supervised Learning Contrastive Learning

Open paper AskPDF arXiv

Computational phylogenetics has become an essential tool in historical linguistics, yet its application at a global scale remains constrained by two factors: the labor-intensive manual annotation of cognacy judgments required for character-based methods and the substantial computational cost of inference on large datasets. This paper introduces a fully self-supervised contrastive learning framework that learns lexical representations directly from raw IPA-transcribed wordlists, without requiring cognacy annotations, alignments, or additional expert input. The model employs a dual contrastive objective: a word-level loss that organizes phonetically similar forms into a coherent space, and an auxiliary language-level loss that encourages the lexical space to reflect broader phonological properties of languages. From the resulting word representations, pairwise language distances are derived and used to infer a global phylogenetic tree of 3,399 language varieties. The inferred tree achieves a generalized quartet distance (GQD) to the Glottolog reference tree competitive with multiple baselines, while requiring only minutes of computation on a standard notebook GPU. Furthermore, the same representations capture diachronic concept stability: variance in pairwise distances across languages yields stability rankings that correlate significantly with established rankings. Ablation studies confirm that both the language-level objective and the use of phonetic feature vectors improved the inferred trees topology with regards to GQD. The framework thus provides a computationally efficient and fully automatic alternative for large-scale phylogenetic inference and offers a unified representation supporting downstream analyses at both the language and concept level.

Abstract

Linguistic Typology

Explore similar work

Self-Supervised Lexical Representation Learning for Fast, Large-Scale Phylogenetic Inference

YallaMorph: A Benchmark for Evaluating Arabic Morphological Generation in Large Language Models

Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach