An entropy-based study of Simplification in ChatGPT translations compared to neural machine translation and human translation across genres

Abstract

This study investigates the phenomenon of simplification in Chinese-to-English translation across Human Translation (HT), neural machine translation (NMT), and large language model (LLM)-based translation, ChatGPT as an example. Employing entropy-based metrics (unigram entropy and Part-of-Speech (POS) entropy) to assess lexical and syntactic complexity, the research analyzes translations across three genres: political texts, fiction, and academic. Findings reveal that political and academic texts exhibit lexical simplification, and texts of all genres show a syntactic simplification trend, with the simplified degree varying across translation modes. While genre exerts minimal influence on lexical complexity, it significantly impacts syntactic complexity, with academic texts showing the lowest and fiction the highest complexity levels. Notably, ChatGPT’s translations consistently exhibit greater lexical complexity, as evidenced by higher unigram entropy scores compared to those of Neural Machine Translation. These results challenge the notion of simplification as a universal feature of translation, instead highlighting its probabilistic nature influenced by translation mode and genre. The study underscores the efficacy of entropy-based measures in capturing nuanced differences in translation complexity and advocates for a modal approach to translation studies that accounts for the unique characteristics of various translation methods.