A Comparative Study of Attitude Resources Between LLM-Based Translation and Human Translation

Abstract

The unprecedented development of Large Language Models has spurred elevated expectations for human translation in specialized contexts. This study explores the differences between ChatGPT and human translation of President Xi’s New Year speeches in terms of reproducing the evaluative meanings from the perspective of Appraisal Theory. The objective is to examine the accuracy of both translations in conveying the attitude resources and attitude polarity of the source text, thereby highlighting the distinctions between human and LLM-based translations in terms of text comprehension and output quality. It is discovered that human translators perform better than LLM-based translation in reproducing and capturing the “affect”, “judgement” and “appreciation” features in the original texts. The advantages of human translation in handling appraisal resources lie in the emotional cognition, contextual understanding, and cultural perception abilities of human translators. While Large Language Models excel in language generation speed, they still have limitations in terms of emotional understanding, cultural adaptation, and information highlighting.