Abstract
This paper explores the long-standing linguistic debate regarding whether nominal class assignment in Swahili is semantically motivated or arbitrary. Using transformer-based language models and WordNet taxonomies, the authors quantify the semantic logic behind Swahili’s 18 nominal classes. By utilizing English-language proxies to avoid morphological bias, the study demonstrates that semantic cohesion varies significantly across classes, with some acting as highly predictable categories and others as residual "catch-all" buckets.
Introduction
In Swahili, every noun belongs to a "nominal class," a system akin to grammatical gender but significantly more complex. These classes determine the agreement patterns for verbs and adjectives. A central question in Bantu linguistics is whether these assignments follow a semantic logic or are merely historical accidents.
A major hurdle in computational analysis is morphological confounding. Swahili noun prefixes often signal their class (e.g., m-toto for Class 1). To isolate the role of meaning, this study uses English translations of definitions as input, effectively "blinding" the models to Swahili word structure and testing if semantic content alone can predict class assignment.
Methodology
The researchers employed two primary computational strategies to analyze the TUKI Swahili-English Dictionary (6,341 records):
- Supervised Learning: The authors fine-tuned MiniLMv2, a compact transformer model, to classify nouns based on their English definitions (e.g., "a young human being" → Class 1/2).
- Unsupervised Topic Modeling: Swahili nouns were linked to WordNet synsets. By calculating weighted Pointwise Mutual Information (wPMI) across the WordNet hierarchy, the authors identified taxonomic descriptors (e.g., "organism," "abstraction") most strongly associated with each class.
Results and Discussion
The study reveals that semantic cohesion exists on a spectrum. While some classes are "semantic fortresses," others are far more heterogeneous.
- Class 1/2 (a-/wa-): Primarily contains humans and causal agents, showing the highest predictability.
- Class 3/4 (u-/i-): Functions as a "botanical" class, dominated by plants and trees.
- Class 9/10 (i-/zi-): Acts as a "residual catch-all," absorbing loanwords and artifacts, which results in lower predictive accuracy.
As shown in Table 2, the supervised model achieved a Macro F1 score of 34.5, significantly outperforming the random baseline of 14.3. Class-specific performance highlights the internal logic of the system: Class 1/2 reached an F1 of 89.4, whereas the "catch-all" Class 9/10 scored 35.3.
The WordNet analysis further validated these findings by "rediscovering" classic linguistic hypotheses. For instance, Class 11 (u-) was strongly linked to "abstraction," confirming its role in representing qualities and states.
Conclusion
This computational approach provides robust evidence for semantic structure within the Swahili nominal class system. While the reliance on English translations may introduce some bias, the methodology is typologically flexible and could be extended to other Bantu languages. Future work may look beyond "is-a" taxonomic relations to part-whole (meronym) structures to uncover even deeper semantic patterns.
Summary updated