The limitations of large language models in short drama subtitle translation from the perspective of Skopos theory

Abstract

Driven by China's cultural export strategy, vertical short dramas have surged in popularity; however, subtitle translation speed and quality have emerged as critical bottlenecks hindering their global reach. Although large language models are widely used for preliminary subtitle translation, they remain inadequate in handling culture-loaded expressions, conveying emotional nuances, and restructuring syntactic patterns. Drawing on the three rules of Skopos theory, this study conducts a comparative analysis of Chinese-to-English subtitle translations generated by DeepSeek-V4-Pro, Doubao 2.1 Turbo, and DeepSeek-R1, using two widely viewed short dramas as case studies. The findings reveal that the core deficiency is not grammatical inaccuracy but a lack of Skopos-oriented awareness: while the models can transfer semantic content, they fail to calibrate their outputs according to dramatic context and character traits, resulting in a persistent gap between semantic fidelity and pragmatic effectiveness. Each model exhibits distinct shortcomings, and none adequately reconciles fidelity with colloquial naturalness.