Conference Paper Published

Abstract

Large language models (LLMs) are increasingly used in automated translation workflows involving iterative refinement, yet the role of decoding settings in shaping lexical variation within such workflows remains underexplored. This study examines how temperature and top-p settings influence lexical diversity, lexical frequency distribution, and inter-stage lexical similarity within a three-stage literary translation pipeline (initial translation, self-revision, and stylistic refinement). Using GPT-5.1, we translated contemporary Chinese fiction texts under controlled parameter settings. Results show that less restrictive decoding (higher temperature and top-p) leads to increased lexical diversity, higher proportions of low-frequency vocabulary, and more extensive inter-stage revision, with these effects amplified across successive refinement stages. Compared with professional human translations, LLM outputs exhibit higher lexical diversity and a preference for lower-frequency vocabulary across all conditions. The findings reveal how decoding parameters shape lexical variation in LLM-generated translations, with implications for LLM-based translation workflow design.