Zero-Shot Morphological Discovery in Low-Resource Bantu Languages via Transfer Learning and Unsupervised Clustering

Forum topic · 小凯 · 2026-04-28

Summary

This arXiv paper (2504.19767, published April 2025) by Hillary Mutisya and John Mugane presents a method for discovering morphological features in low-resource Bantu languages by combining cross-lingual transfer learning with unsupervised clustering. Applied to Giriama (nyf), a language with only 91 labeled paradigms, the pipeline discovers noun class assignments for 2,455 words and identifies two previously undocumented morphological patterns: an a- prefix variant for Class 2 (vowel coalescence of wa-, with 95.1% consistency) and a contracted k'- prefix (98.5% consistency). External validation on 444 known Giriama verb paradigms confirms 78.2% lemmatization accuracy. After expanding the corpus to 19,624 words, the method achieves 97.3% segmentation and 86.7% lemmatization rates across all major word classes. The work demonstrates that computational morphology can be applied to extremely low-resource languages with minimal labeled data, offering a practical pathway for linguistic documentation and NLP resource creation in underrepresented languages.

Paper Overview

Field: NLP Authors: Hillary Mutisya, John Mugane Published: 2025-04-28 arXiv: 2504.19767

Abstract

We present a method for discovering morphological features in low-resource Bantu languages by combining cross-lingual transfer learning with unsupervised clustering. Applied to Giriama (nyf), a language with only 91 labeled paradigms, our pipeline discovers noun class assignments for 2,455 words and identifies two previously undocumented morphological patterns: an a- prefix variant for Class 2 (vowel coalescence of wa-, 95.1% consistency) and a contracted k'- prefix (98.5% consistency). External validation on 444 known Giriama verb paradigms confirms 78.2% lemmatization accuracy, while a corpus expansion to 19,624 words achieves 97.3% segmentation and 86.7% lemmatization rates across all major word classes.

Key Contributions

Combines cross-lingual transfer learning with unsupervised clustering to work in a zero-shot setting with extremely limited labeled data (only 91 labeled paradigms for Giriama).
Discovers noun class assignments for 2,455 words, including two previously undocumented patterns:
An a- prefix variant for Class 2, formed by vowel coalescence of the standard wa- prefix (95.1% consistency).
A contracted k'- prefix (98.5% consistency).
External validation on 444 known Giriama verb paradigms yields 78.2% lemmatization accuracy.
Scaling the corpus from 91 labeled paradigms to 19,624 words boosts performance to 97.3% segmentation and 86.7% lemmatization across all major word classes.

Significance

The results show that morphology in severely under-resourced Bantu languages can be uncovered computationally without substantial annotation, opening a scalable path for linguistic documentation and downstream NLP tools for low-resource languages.