🤖 machine learning

Neural Recovery of Historical Lexical Structure in Bantu Languages from Modern Data

This paper demonstrates that neural transformer models trained on modern Bantu morphological data can successfully recover historical lexical structures and cognate patterns that align with established Proto-Bantu reconstructions.

Original authors: Hillary Mutisya, John Mugane

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hillary Mutisya, John Mugane

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Digital Time Machine: How AI is Uncovering the "Lost Language" of Africa

Imagine you are looking at a massive, ancient family tree. You see hundreds of cousins—some living in cities, some in villages, some speaking slightly different dialects—but you’ve never met the "Great-Great-Great-Grandparent" who started it all. You know they existed, but you’ve never heard their voice.

In linguistics, this "Great-Grandparent" is called Proto-Bantu. It is the ancient ancestor of over 500 languages spoken by 300 million people across Africa. For over a century, human scholars have been playing "detective," painstakingly comparing modern words to try and guess what that original ancestor sounded like.

A new research paper by Mutisya and Mugane asks a radical question: Can we teach an AI to do this detective work for us?

The Concept: The "DNA" of Words

Think of a language like a person’s appearance. You might have your father’s nose or your grandmother’s eyes. Even if you’ve never met your grandmother, you can look at your own face and see the "traces" of her features.

Words work the same way. Even though a word for "cow" might sound different in Swahili than it does in Zulu, they often carry a "genetic" fingerprint of the original Proto-Bantu word.

The researchers didn't teach the AI the ancient language (because nobody knows exactly how it sounded). Instead, they fed the AI modern data—the way people speak today. They used a specialized AI model called BantuMorph that is an expert at understanding the "grammar bones" (morphology) of these languages.

The Experiment: Sorting the "Family Reunion"

The researchers took 14 different languages from Eastern and Southern Africa and asked the AI to look at their "embeddings"—which you can think of as digital fingerprints for every word.

They then performed three main tasks:

The Cognate Hunt (Finding the Cousins): They asked the AI, "Which words across these different languages are actually just different versions of the same original word?" The AI found thousands of candidates. When they checked these against the "gold standard" of historical records, they were right 90.9% of the time for nouns! It successfully identified ancient words for "person," "cow," and "head."
The Number Test (The Unchanging Truth): Numbers are like the "heartbeat" of a language—they change very slowly over thousands of years. The AI successfully recovered ancient number patterns, like the word for "three" and "nine," which have remained remarkably stable across the continent.
The Family Tree (Mapping the Ancestry): They asked the AI to group the languages based on how similar they were. The AI didn't just guess; it perfectly recreated the "Guthrie Zones"—the established scientific map of how these language families are related. It was like giving the AI a pile of random photos and having it correctly sort them into "The Smith Family," "The Jones Family," and "The Miller Family."

The "Noise" in the System: Avoiding the Traps

The researchers had to be careful. Imagine if you tried to find family resemblances, but you accidentally included people who just happened to be wearing the same brand of t-shirt. You might think they are related, but they aren't!

In the digital world, "t-shirts" are loanwords. If many African languages all borrowed the word "hospital" or "school" from English or Arabic, the AI might think they are "related" because they use the same word. The researchers had to build filters to make sure the AI was looking at ancient heritage, not just modern trends.

Why This Matters: A High-Tech Magnifying Glass

Does this mean the AI is replacing human linguists? No.

Think of the AI as a high-powered metal detector. A metal detector can scan a massive beach and beep when it finds something shiny under the sand. It tells the treasure hunter, "Hey! Look over here!" But the human still has to go to that spot, dig carefully, and decide if they’ve found a gold coin or just a piece of a soda can.

This paper shows that AI can scan thousands of words in seconds, pointing linguists toward the most likely "ancient treasures." It’s a way to speed up the process of understanding the deep, shared history of the African continent.

Technical Summary: Neural Recovery of Historical Lexical Structure in Bantu Languages

1. Problem Statement

Historical linguistics traditionally relies on the comparative method—a labor-intensive, manual process of identifying regular sound correspondences across daughter languages to reconstruct ancestral "proto-languages" (e.g., Proto-Bantu). While computational linguistics has attempted to extract phylogenetic signals from multilingual models (like mBERT), there has been a lack of research demonstrating whether neural models trained specifically on morphological data can recover specific lexical cognates and proto-form correspondences. The core research question is: Can a neural model trained exclusively on modern morphological data recover cross-lingual lexical structures consistent with historical reconstruction?

2. Methodology

The researchers employed a multi-stage computational approach:

Model Architecture: They utilized BantuMorph v7, a character-level transformer (ByT5-small, 300M parameters) pre-trained on morphological paradigms across 16 Bantu languages.
Data Selection: The study focused on 14 Eastern and Southern Bantu languages (Guthrie zones C, E, G, H, J, N, S).
Embedding Extraction & Language Centering: Encoder embeddings were extracted from the final layer using mean pooling. To isolate morphological and semantic signals from language-specific biases (like writing systems or token distribution), the authors applied language centering—subtracting the per-language mean embedding.
Cognate Discovery: Using transfer learning, the authors identified "nearest neighbors" in the centered embedding space. Candidates were defined as lemmas shared across five or more languages.
Validation Framework:
- Historical Databases: Candidates were compared against the Bantu Lexical Reconstructions database (BLR3) and the ASJP basic vocabulary.
- Cross-Model Validation: To ensure results weren't artifacts of the training objective, they used NLLB-600M (a translation model with no morphological training) to see if it independently recovered the same cognate clusters and phylogenetic groupings.
Mitigation of Noise: The authors implemented strategies to filter out "modern artifacts," such as foreign proper nouns (e.g., "Argentina"), adapted loanwords (e.g., "hospital"), and domain bias (e.g., religious terminology).

3. Key Contributions

Automated Cognate Candidate Generation: Demonstrated that neural embeddings can generate large-scale lists of potential cognates (728 nouns, 1,525 verbs) that require significantly less manual labor than traditional methods.
Morphological Encoding of History: Proved that neural models learn to encode historical relationships (like noun class systems) as a byproduct of learning modern morphological rules.
Scalable Screening Tool: Positioned the neural approach as a "screening mechanism" that provides high-quality candidates for expert linguists to evaluate, rather than a replacement for the comparative method.

4. Results

High Accuracy in Noun Recovery: 10 of the top 11 noun candidates (90.9%) aligned with reconstructed Proto-Bantu forms. Notable successes included -ntU (‘person’), gombe (‘cow’), and mUn (‘person’).
Numeral Stability: The model successfully recovered highly stable Proto-Bantu numerals, such as tatu (‘three’) and kenda (‘nine’), across wide geographic ranges.
Verb Reconstruction: 12 verb cognates matched established Proto-Bantu roots, including -bon- (‘see’) and -jím- (‘stand’), with some appearing in all 14 studied languages.
Noun Class Preservation: The model captured the 13 productive noun classes, showing high cosine similarity (>0.83) within classes across languages ( $p < 10^{-9}$ ).
Phylogenetic Accuracy: Ward-linkage clustering and MDS projections successfully recovered known Guthrie-zone classifications and fine-grained sub-branching (e.g., the E-zone sub-tree), a result confirmed by both BantuMorph and NLLB ( $p < 0.01$ ).

5. Significance

This paper marks a significant step in Neural Historical Linguistics. It demonstrates that the "latent" knowledge within a transformer model, when trained on morphological patterns, can mirror the deep historical structures of a language family. While the authors caution that their results represent the recovery of shared lexical structure (which could include regional innovations or areal diffusion) rather than a definitive reconstruction of Proto-Bantu, the high alignment with BLR3 suggests that neural models can effectively bridge the gap between modern computational NLP and classical historical linguistics.