Phonological Fossils: Machine Learning Detection of Non-Mainstream Vocabulary in Sulawesi Basic Lexicon

Mukhlis Amien; Go Frendi Gunawan

cs.CLMar 11, 2026

Phonological Fossils: Machine Learning Detection of Non-Mainstream Vocabulary in Sulawesi Basic Lexicon

Authors: Mukhlis Amien, Go Frendi Gunawan

Abstract

Basic vocabulary in many Sulawesi Austronesian languages includes forms resisting reconstruction to any proto-form with phonological patterns inconsistent with inherited roots, but whether this non-conforming vocabulary represents pre-Austronesian substrate or independent innovation has not been tested computationally. We combine rule-based cognate subtraction with a machine learning classifier trained on phonological features. Using 1,357 forms from six Sulawesi languages in the Austronesian Basic Vocabulary Database, we identify 438 candidate substrate forms (26.5%) through cognate subtraction and Proto-Austronesian cross-checking. An XGBoost classifier trained on 26 phonological features distinguishes inherited from non-mainstream forms with AUC=0.763, revealing a phonological fingerprint: longer forms, more consonant clusters, higher glottal stop rates, and fewer Austronesian prefixes. Cross-method consensus (Cohen's kappa=0.61) identifies 266 high-confidence non-mainstream candidates. However, clustering yields no coherent word families (silhouette=0.114; cross-linguistic cognate test p=0.569), providing no evidence for a single pre-Austronesian language layer. Application to 16 additional languages confirms geographic patterning: Sulawesi languages show higher predicted non-mainstream rates (mean P_sub=0.606) than Western Indonesian languages (0.393). This study demonstrates that phonological machine learning can complement traditional comparative methods in detecting non-mainstream lexical layers, while cautioning against interpreting phonological non-conformity as evidence for a shared substrate language.

Figures & tables

Language	Total	Has Cognacy	Residual	% Residual	Coverage
Muna	219	185 (84%)	26	11.9%	84%
Bugis	242	180 (74%)	49	20.2%	74%
Toraja-Sa’dan	216	171 (79%)	32	14.8%	79%
Wolio	254	171 (67%)	68	26.8%	67%
Makassar	217	137 (63%)	67	30.9%	63%
Tolaki	209	75 (36%)	114	54.5%	36%

Table 1: Per-language residual rates after rule-based cognate subtraction (E022).

Model	Classifier	AUC	F1	Accuracy
A (Full)	XGBoost	1.000 $\pm$ 0.001	1.000 $\pm$ 0.001	1.000 $\pm$ 0.001
	Random Forest	1.000 $\pm$ 0.000	1.000 $\pm$ 0.000	1.000 $\pm$ 0.001
	Logistic Reg.	1.000 $\pm$ 0.000	1.000 $\pm$ 0.000	1.000 $\pm$ 0.000
B (Phon.)	XGBoost	0.760 $\pm$ 0.007	0.822 $\pm$ 0.005	0.741 $\pm$ 0.008
	Random Forest	0.762 $\pm$ 0.006	0.788 $\pm$ 0.006	0.717 $\pm$ 0.008
	Logistic Reg.	0.747 $\pm$ 0.003	0.748 $\pm$ 0.005	0.683 $\pm$ 0.006

Table 2: Stratified 5-fold CV results (

\times

10 seeds). Model A includes cognacy features (circular baseline); Model B uses only phonological features (scientific claim).

Held-out language	AUC	F1	Acc.	$N_{\mathrm{test}}$	$N_{\mathrm{substr}}$
Tolaki	0.806	0.530	0.364	209	134
Makassar	0.747	0.781	0.659	217	80
Bugis	0.727	0.799	0.707	242	62
Wolio	0.697	0.798	0.669	254	83
Toraja-Sa’dan	0.696	0.787	0.690	216	45
Muna	0.618	0.465	0.402	219	34

Table 3: Leave-One-Language-Out results for Model B (XGBoost). Each language was held out as the test set in turn.

Variant	CV AUC	LOLO mean AUC	LOLO $\geq$ 0.65
Full Model B (27 features)	0.760 $\pm$ 0.007	0.715	5/6
Ablated ( $-$ coverage, 26 feat.)	0.763 $\pm$ 0.007	0.722	6/6
Pure phonological (25 feat.)	0.727 $\pm$ 0.007	0.701	5/6

Table 4: Feature ablation results (XGBoost). Removing the cognacy coverage confound improves performance.

Figure 1: SHAP beeswarm plot for Model B (XGBoost). Each dot represents one form; color indicates feature value (red = high, blue = low). Positive SHAP values push toward substrate classification.

Figure 2: Four-quadrant comparison of rule-based (E022) and ML (E027 Model B) substrate predictions. CS = Consensus Substrate, CA = Consensus Austronesian, RO = Rule-Only, MO = ML-Only.

Figure 3: Distribution of cross-linguistic pairwise Levenshtein distances for consensus substrate forms (solid line) vs. null distribution from 1,000 random concept draws (histogram). Substrate forms are not more similar across languages than expected by chance ( $p$ = 0.569).

Group	$N_{\mathrm{langs}}$	Mean Rule%	Mean ML%	Mean AUC	Mean $P_{\mathrm{sub}}$	Notable outliers
Original 6	6	28.0	23.1	0.890	0.326	—
Sulawesi exp.	8	49.9	62.4	0.685	0.606	Bol.Mong. (9.2%), Gorontalo (84.2%)
W. Indonesian	6	28.4	35.3	0.634	0.393	Acehnese (62.9%), Sundanese (1.8%)
E. Indonesian	2	34.9	51.9	0.661	0.520	—

Table 5: Expansion validation results by geographic group. “Original” = the 6 training languages. AUC is computed against each language’s own rule-based residual labels.

Figure 4: Predicted substrate rates (ML) vs. rule-based residual rates for 22 languages grouped by geographic region. Languages are ordered by ML substrate rate within each group.

Self-Supervised Lexical Representation Learning for Fast, Large-Scale Phylogenetic Inference
Sep 4, 2026Tim WientzekPhylogenetic Inference Self-Supervised Learning
Scaling Self-Supervised Speech Models Uncovers Deep Linguistic Relationships: Evidence from the Pacific Cluster
Mar 7, 2026Minu Kim, Hoirin Kim, David R. MortensenSelf-Supervised Speech Representation Learning Language Identification
Thomas Lord Department of Computer Science, University of Southern California, USA · School of Electrical Engineering, KAIST, Republic of Korea · Language Technologies Institute, Carnegie Mellon University, USA
Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages
Apr 20, 2026V. S. D. S. Mahesh Akavarapu, Michael Daniel, Gerhard JägerASR Evaluation Automatic Speech Recognition
University of Tübingen · University of Jena

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

Explore similar work

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

Scaling Self-Supervised Speech Models Uncovers Deep Linguistic Relationships: Evidence from the Pacific Cluster

Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages