From Confucius to computational linguistics: quantifying cross-linguistic semantic similarity and semantic fidelity using large language models

Abstract Translating The Analects from classical Chinese into English frequently results in semantic losses that obscure understanding and introduce ambiguities. This study aims to quantify the semantic similarities and differences between various English translations of The Analects and the original text, uncover the root causes of these discrepancies, and quantitatively assess the translations’ semantic fidelity. The study constructed a parallel corpus, incorporating translations by D.C. Lau, James Legge, William Jennings, Edward Slingerland, and Burton Watson, along with the original text. By leveraging large language models from Baidu AI, Zhipu AI, and OpenAI, the study established a framework for cross-linguistic semantic similarity. The findings reveal that the semantic alignment between the translations and the original text predominantly ranges from 60 per cent to 80 per cent. Instances of high similarity demonstrate that, despite linguistic variations, the core messages and logical structures of the original can be effectively conveyed in English. This ensures an accurate representation of meanings and maintains structural integrity. Conversely, instances of low similarity highlight the translators’ challenges in maintaining semantic fidelity to the source text while ensuring readability. Finally, this study assigns semantic fidelity scores to the five translations of The Analects and offers a reference for the quantification of semantic fidelity in translation evaluation.