Languages change over time, but they do not change at random. When a population splits, their shared language diverges into distinct dialects, and eventually, entirely different languages. Yet, the ghost of the ancestral tongue remains embedded in cognates—words across different languages that share a common genetic origin.
Historically, linguists used the comparative method to manually reverse-engineer these changes. This method relies on the regularity principle: sound changes operate like deterministic rules (e.g., Latin /f/ consistently becoming Spanish /h/ or remaining silent, as in filium $\rightarrow$ hijo). Computational historical linguistics aims to automate this process. This works because language evolution can be modeled as a mathematical transmission channel where discrete phonetic units undergo structured, tree-aligned mutations.
The First Hurdle: Identifying Cognates Amidst Divergence
Before a model can reconstruct a proto-language, it must first group related words together. This is exceptionally difficult because centuries of phonetic erosion can completely obscure surface-level similarities. As Kondrak notes, words derived from a common ancestor can undergo such extreme phonetic and semantic shifts that their shared origin becomes invisible to standard string-matching algorithms.
To capture these elusive relationships, researchers have moved beyond simple edit distances. For instance, Rama introduced gap-weighted string subsequences to handle sound losses, gains, and assimilations, allowing models to detect deep phonetic similarities even when words have diverged significantly. Furthermore, Nath et al. demonstrated that combining phonetic data with semantic and articulatory features (the physical mechanics of how sounds are produced) allows models to bridge the gap between closely related but low-resource languages.
Taking a completely different approach to avoid the messy problem of cross-lingual phonetic alignment, T. Mark Ellison and Simon Kirby developed a method called PHILOLOGICON. Instead of comparing words across languages directly, their method analyzes the internal similarity structure of a single language's vocabulary. Because a language's internal lexical matrix acts as a unique structural fingerprint, comparing these internal structures allows researchers to reconstruct language family trees without needing a universal phonetic space.
Modeling the Arrow of Time: Inducing Sound Laws
Once cognates are identified, the next step is to discover the exact rules—the sound laws—that governed their divergence.
Early computational approaches framed this as a generative process. Alexandre Bouchard et al. developed a probabilistic model that treats phonological shifts as stochastic edits moving down an evolutionary tree. By modeling sound changes as probabilistic transitions, the system can simultaneously infer the most likely tree topology and the ancestral words at the root.
More recently, researchers have leveraged deep learning to capture these rules implicitly or explicitly. Carlo Meloni et al. showed that neural sequence-to-sequence models can automatically reconstruct Latin proto-words from Romance descendants. Crucially, the model does not just memorize vocabulary; it internalizes complex phonological logic and groups sounds by physical properties (like voicing) without explicit human supervision.
To make these induced laws human-readable and highly precise, Atharva Naik et al. framed Sound Law Induction as a Programming by Example (PBE) task. Their system, PySLICoder, uses Large Language Models trained on synthetic data to generate explicit code representing historical sound shifts, combining the flexibility of neural networks with the rule-based rigor of traditional linguistics.
[ Proto-Language Word ]
│
┌──────────────┴──────────────┐
▼ ▼
Sound Law A Sound Law B
(e.g., p → f / #_) (e.g., p → b / V_V)
│ │
▼ ▼
[ Modern Reflex 1 ] [ Modern Reflex 2 ]
Predicting the Unobserved: Reflexes and Proto-Forms
If we can model the sound laws and the family tree, we can perform two predictive tasks: reconstruction (moving backward to find the proto-language) and reflex prediction (moving forward to predict how an ancestral word would manifest in a modern language).
This dual problem maps cleanly onto computational biology. V. S. D. S. Mahesh Akavarapu and Arnab Bhattacharya 0001 adapted the MSA Transformer—originally built for protein sequence alignment—to historical linguistics. By treating cognate sets like biological sequence alignments, the model leverages self-attention across both rows (languages) and columns (phonemes) to fill in missing words or predict ancestral forms.
Alternatively, Tiago Tresoldi framed reflex prediction as a classification problem. By enriching standard sequence alignments with multi-layered phonological features, machine learning models can classify exactly which sound a proto-phoneme will transform into within a specific target language. The utility of these approaches was validated in the SIGTYP 2022 Shared Task organized by Johann-Mattis List et al., which proved that convolutional neural networks and advanced sequence models could systematically learn these complex correspondences directly from data. For broader phylogenetic uncertainty, Gerhard Jäger demonstrated that Bayesian phylogenetic inference can systematically predict missing cognate words across entire language families by integrating sound correspondences with evolutionary trees.
The Data Engine: From Fragmentary Records to Global Databases
The transition from manual analysis to deep learning requires massive, structured datasets. Historically, computational historical linguistics was bottlenecked by fragmented data.
To solve this, Benoît Sagot developed methods to automatically parse crowdsourced platforms like Wiktionary into structured etymological databases containing nearly a million lexemes. Similarly, specialized regional benchmarks have emerged: Liviu P. Dinu et al. introduced RoBoCoP for Romance language cognates and borrowings, while Aryaman Arora et al. built JAMBU to unify the fragmented linguistic records of South Asian languages.
These large-scale data engines do more than just validate models; they allow us to tackle the most mysterious corners of linguistics. For instance, Frederic Blum et al. utilized supervised neural networks trained on global lexical and grammatical data to classify "linguistic isolates"—languages like Basque that have no known living relatives—and extract deep phonetic patterns from fragmentary historical records that elude traditional manual analysis.
Key Foundational Work
Measuring Language Divergence by Intra-Lexical Comparison (2006) — contributed 1 concept elements to this direction.
Identifying Cognates by Phonetic and Semantic Similarity (2001) — contributed 1 concept elements to this direction.