Borrowing from cousins: morphological discovery for Giriama

Giriama is a Bantu language spoken by about 600,000 people on the Kenyan coast, in and around Malindi and Kilifi. It’s part of the Mijikenda group — the family of related Bantu languages that share the coastal strip with Swahili and have been there for a long time. It has a writing tradition, a translated Bible, oral literature, and ongoing community use. What it doesn’t have is much of an NLP footprint.

When we went to look for computational morphological resources for Giriama, we found 91 labeled paradigms. Not 91 verbs in 91 forms — 91 lemma-feature-form triples total, each one a word, its morphological features, and its surface form. The kind of thing you’d plausibly hand-collect in an afternoon as a starting point for a language documentation project. Compare that to the 11,749 paradigms for Swahili in the same source, and you have a sense of the scale problem.

This post is the practitioner’s account of our paper on zero-shot morphological discovery for Giriama. We were trying to answer a specific question: can you do useful morphological work on a language with this little data, by borrowing from its better-resourced cousins?

The short answer is yes. The longer answer involves cross-lingual transfer, unsupervised clustering, and the careful business of telling the difference between “the model discovered something real” and “the model overfit to its own training set.”

What “low-resource” actually looks like

Most papers on low-resource NLP work with thousands of examples. That’s already a small dataset by deep learning standards, but it’s not really low-resource in the field linguistics sense. Many living languages have on the order of a few dozen to a few hundred fully annotated paradigms in any computational form, if any at all. The methods that work at the “thousand examples” scale don’t always survive at the “fifty examples” scale.

Giriama at 91 paradigms is in this harder regime. Standard supervised learning on 91 examples will overfit to the specific examples and generalize poorly. The vocabulary is too small to discover noun class patterns. The morphological complexity is too rich to be captured by 91 word forms.

But Giriama isn’t isolated. It’s a close cousin of Kikuyu, Kamba, and Kimeru — languages with substantially more documentation. Roughly 60% of the Giriama vocabulary overlaps with Swahili. The grammar shares the standard Bantu skeleton: noun classes, agglutinative verbs, agreement chains. If a model knows Bantu morphology generally, learning Giriama specifically should be a small step.

The question is how to take that small step without enough data to do it the normal way.

Two methods that complement each other

We ended up using two methods in parallel and combining their outputs:

Cross-lingual transfer. Run a Bantu-family pre-trained model (BantuMorph, our model trained jointly on 16 Bantu languages) on Giriama text. For each Giriama word, look at where it lands in the model’s embedding space and find the nearest Swahili neighbors. The Swahili neighbor’s noun class is a starting hypothesis for the Giriama word’s noun class. This works well for cognates (words shared between Giriama and Swahili) but it doesn’t work for Giriama-specific innovations, words that don’t have a Swahili counterpart.

Unsupervised clustering. Take all the Giriama words, project them into a low-dimensional space using UMAP (a dimensionality-reduction technique that preserves local structure), and run k-means clustering. Each cluster gets a candidate noun class assignment, derived from whatever class is most consistent within the cluster’s known examples. This works on words that don’t have Swahili cognates, including innovations specific to Giriama, but it requires enough Giriama vocabulary to form meaningful clusters.

Cross-lingual transfer is good at finding cognates and bad at innovations. Unsupervised clustering is good at innovations and bad at one-off cognates. Combining them via weighted voting captures both: where they agree, you get high-confidence assignments; where they disagree, you get useful flags for human review.

The mechanic is straightforward but the practical value is that you can run both methods on every word and use the agreement structure to know where to trust each.

What we found

After running the pipeline on Giriama text and validating against the 91 labeled paradigms (held out for evaluation), we ended up with:

• 2,455 noun class assignments for previously unannotated words — a 27× expansion over what existed before
• 78.2% lemmatization accuracy when externally validated against 444 known Giriama verb paradigms collected from a separate source
• 97.3% segmentation accuracy and 86.7% lemmatization accuracy on a larger 19,624-word v3 corpus expansion across all word classes

The first number is the headline, but the second one, the external validation, is the one that mattered for trust. Anything you can produce from a 91-example training set will look good when you evaluate against the same 91 examples. The 444 paradigms were collected separately, by different people, and the 78.2% lemmatization on that set is a real signal that the pipeline learned generalizable patterns and not just memorized its training set.

Two morphological patterns we didn’t expect

The most linguistically interesting outcome wasn’t the noun class assignments. It was two morphological patterns that we hadn’t seen documented in published Giriama grammars:

An a- prefix variant for Class 2. Standard Bantu Class 2 is wa- (and Giriama is no exception — wantu “people” follows this pattern). But the model found a substantial set of Giriama words where a Class 2 plural is marked with bare a- instead of wa-, apparently the result of vowel coalescence (the merger of two adjacent vowels) when wa- combines with vowel-initial roots. The a- variant appears with 95.1% consistency in the corpus. It’s the kind of pattern that a linguist working on the language would eventually notice; the model surfaced it from 91 paradigms and an unlabeled corpus.

A contracted k’- prefix. Some Giriama nouns appear with a k’- (glottalized k) prefix that doesn’t quite match any of the standard Bantu noun-class markers. The model found this with 98.5% consistency across the relevant words. Whether this is a contracted form of ki- (Class 7) or a distinct innovation is something a phonetician should weigh in on; we can’t tell from text alone.

One important caveat. We didn’t discover these patterns in the linguistic sense. A native speaker noticing the same patterns wouldn’t be able to publish a “I found a new prefix” paper, because the patterns are presumably part of the productive grammar that speakers know intuitively. What we did is surface them from text, in a form that lets a linguist verify or reject them, in much less time than it would take to discover them through traditional fieldwork. The model doesn’t replace the linguistic work; it accelerates it.

When this approach works and when it doesn’t

The Giriama case is favorable in a specific way: it has close, well-resourced relatives. Kikuyu, Kamba, and Kimeru aren’t English-rich-resource by global standards, but they have hundreds of paradigms each, and the morphology is structurally similar to Giriama’s. The transfer step has somewhere productive to transfer from.

Take that away and the approach struggles. We’ve tried similar pipelines on languages that don’t have close, well-resourced relatives — Maasai (Nilotic, no documented Bantu relative), Somali (Cushitic, in a different family entirely). The transfer signal is much weaker because the model can’t bootstrap from a related language’s morphology. You’re back to needing more direct supervision or more sophisticated unsupervised methods.

This means the practical envelope for the approach is something like: small target language + close, well-resourced relative + a few hundred examples in the target. Outside that envelope, it gets harder. Inside, the recipe works pretty consistently.

Where this fits

If you’re working on a low-resource language with a close, well-resourced relative, you don’t need 10,000 paradigms to do useful morphological work. You can do useful work with a few dozen, drawing on the relative’s resources via cross-lingual transfer plus targeted clustering. The output isn’t a finished morphological analyzer; it’s a candidate set for expert review, with confidence scores attached.

For language documentation projects, this matters because it changes the time budget. The bottleneck for documenting an undocumented language has historically been fieldwork — collecting examples, sitting with native speakers, building the lexicon and grammar from scratch. Computational methods can’t replace that work, but they can let a single linguist cover much more ground per hour. Surfacing 2,455 noun class candidates that a linguist can verify or reject in a focused review session is a different workflow than building those 2,455 entries by hand.

For NLP research, the more interesting claim is that “low-resource” methods that work at the thousand-example scale aren’t necessarily the ones that work at the dozens-of-examples scale. Giriama with 91 paradigms is harder than Swahili with 11,749, and it requires different methods — cross-lingual transfer becomes essential, not optional, and unsupervised methods carry weight they don’t carry in higher-resource regimes.

For Giriama specifically, what came out the other end was 2,455 candidate noun class assignments and two prefix patterns — a- and k’- — sitting in a form a linguist can verify or reject. That’s not a solved language, but it’s a different starting point than 91 paradigms.

The paper is at arxiv.org/abs/2604.22723. Code and discovered lexicons are linked from the paper.