Low-Resource Languages and the Representational Gap in Large Language Models: A Theoretical Analysis of Data Scarcity, Transfer and Linguistic Equity

Abstract

Large language models perform far worse in the languages spoken by most Africans than in the handful of languages that dominate their training data, and the difference is usually described as a data problem that more data will solve. This paper argues that the description is incomplete in a way that matters for what should be done about it. Drawing on the recent survey literature on language models for low-resource languages, on empirical work on tokenization, transfer and adaptation for African languages, and on philosophical analyses of language-technology bias as epistemic injustice, it develops a theoretical account of the representational gap as the compound product of four distinct mechanisms: scarcity of pre-training data, tokenization inefficiency that taxes every input and output in the language, capacity dilution across languages in a shared model, and the absence of evaluation resources that would make the other three visible. Because the mechanisms multiply rather than add, interventions on one leave the gap largely intact, which explains why a decade of data-collection effort has narrowed it less than expected. The paper organises the mitigation strategies that the literature documents into a design space, shows which mechanism each addresses, and argues that linguistic equity should be adopted as an explicit evaluation criterion for language models, operationalised as capability parity, cost parity and evaluation coverage across languages. It derives six propositions, considers implications for Nigerian and other African languages, and proposes a research agenda in which universities act as stewards of the linguistic data on which equity depends.