Semantic Errors in Machine Translation: A Study of Colloquial Arabic-English Captions on TikTok

This study investigates the semantic accuracy of TikTok's machine translation (MT) system, focusing on auto-generated English captions translated from colloquial Arabic content posted by the verified MBC1 and Shahid TikTok accounts. The primary objective is to identify and categorize semantic errors using Sayogie's (2014) framework, which classifies meaning into three interrelated dimensions: grammatical, contextual, and referential. Employing a descriptive qualitative approach, the research analyses a sample of colloquial Arabic captions to evaluate the extent to which TikTok's translation system preserves the intended meaning. Findings reveal a consistent presence of all three error types, underscoring the system's limitations in processing idiomatic, context-dependent expressions typical of collo-quial Arabic. These inaccuracies primarily result from the system's reliance on literal translation strategies, which fail to account for figurative language, cultural references, and emotional nuance. While TikTok's MT feature enhances accessibility for multilingual users, it remains inadequate in preserving semantic integrity, particularly in informal, culturally embedded content. To address these challenges, the study recommends developing more context-aware, dialect-sensitive models capable of handling the pragmatic and cultural complexity of colloquial speech. The findings contribute to current debates in MT evaluation by highlighting the need to prioritize semantic adequacy, especially for low-resource and dialect-rich languages such as Arabic.