Nominal Class Assignment in Swahili: A Computational Account

Giada Palmieri, Konstantinos Kogkalidis

Introduction

In Bantu languages like Swahili, nouns are categorized into a complex system of "nominal classes"—often referred to as genders, though they far exceed the binary or ternary systems found in Indo-European languages. These classes are not merely decorative; they govern the grammatical agreement (concord) of verbs and adjectives within a sentence.

Mapping of Swahili nominal classes to their respective noun affixes and subject concords.

A long-standing debate in linguistics concerns whether these assignments are arbitrary historical artifacts or if they are rooted in semantic principles. While some classes clearly target specific categories—such as humans or plants—others appear to be "catchalls" for disparate concepts. This study employs modern natural language processing (NLP) techniques to quantitatively map the semantic cohesion of the Swahili nominal system.

Methodology

To investigate the semantic basis of noun classes without the confounding influence of Swahili morphology (like prefixes), the researchers adopted a two-pronged computational approach:

  1. Supervised Classification: A MiniLMv2 language model was fine-tuned to predict a noun's class based solely on its English definition. By using translations, the model was forced to rely on semantic content rather than morphological cues.
  2. Unsupervised Topic Modeling: Swahili nouns were mapped to English WordNet synsets. The researchers then analyzed "hypernym paths" to identify broad taxonomic generalizations and used weighted Pointwise Mutual Information (wPMI) to determine the statistical relevance of high-level concepts (e.g., "artifact," "abstraction") to specific classes.

The study utilized a cleaned dataset of 6,341 records derived from the TUKI Swahili-English Dictionary, focusing on the nine most populous classes that represent approximately 98% of the data.

Results and Semantic Insights

The supervised model significantly outperformed a random baseline, achieving a Macro F1 of 34.5 compared to 14.3. This performance confirms that noun class assignment contains a strong semantic signal.

Confusion matrix showing the performance of the MiniLMv2 model in predicting noun classes from English definitions.

Key Findings:

  • Semantic Anchors: Class 1/2 ($a-/wa-$), which contains humans and causal agents, showed the highest cohesion with an 89.4 F1 score. Class 11/14 ($u-$) emerged as a primary home for abstract concepts (60.0 F1).
  • The "Catchall" Class: Class 9/10 ($i-/zi-$) proved to be semantically disparate, supporting the theory that it serves as a residual category for loanwords and miscellaneous items.
  • Overlapping Domains: The model struggled to distinguish between Classes 5/6, 7/8, and 9/10, as all three heavily feature "human-made objects." This overlap suggests a shared semantic space for artifacts across different grammatical categories.
  • Understudied Classes: Analysis of the $ya-$ and $i-$ classes revealed they are predominantly populated by specific types of abstractions, potentially related to uncountability.

Conclusion

This research provides the first quantitative taxonomic map of the Swahili nominal system, demonstrating that while some classes act as "catchalls," many exhibit predictable semantic patterns. Future work may extend this methodology to other Bantu languages to determine if these semantic topographies are cross-linguistically consistent or if they reflect unique cultural worldviews.

Summary updated

BibTeX
@inproceedings{palmieri2024nominal,
  author = {Giada Palmieri and Konstantinos Kogkalidis},
  title = {{Nominal Class Assignment in Swahili: A Computational Account}},
  booktitle = {CLiC-it},
  year = {2024},
}
Figure