Data Scarcity vs. Algorithmic Bias in Persian LLMs: A Critical Analysis

The rapid advancement of Large Language Models (LLMs) has revolutionized natural language processing, yet their transformative potential is unevenly distributed. For low-resource languages such as Persian, significant challenges persist, hindering the development of robust and equitable models. Among these challenges, data scarcity and algorithmic bias are particularly critical obstacles; however, these issues are often analyzed in isolation. This paper posits that data scarcity and algorithmic bias are not independent problems but are fundamentally intertwined in a self-perpetuating “vicious cycle.” We argue that the chronic lack of high-quality, representative data is a primary driver that actively creates and systematically amplifies social and representational biases within Persian LLMs. Through a critical analysis, we deconstruct the stages of this cycle, supported by a structured review of the limitations inherent in existing Persian corpora and connecting these data-level deficiencies to documented forms of algorithmic bias. This analysis reveals that simplistic solutions focusing solely on increasing data quantity are insufficient. Understanding this cyclical dynamic is crucial for guiding future research towards a more holistic approach, focusing on community-driven data curation and the development of fairness-aware methodologies to build genuinely equitable and reliable NLP technologies for the Persian language.

کلیدواژه ها:

Large Language Models (LLMs) ، Persian NLP ، Algorithmic Bias ، Data Scarcity ، Low-Resource Languages ، Vicious Cycle ، AI Ethics ، Corpus Linguistics