Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study analyzes the emotional profiles of translations produced by large language models (LLMs) and how post-editing alters these profiles to align more closely with human translation norms. By comparing LLM translations of Margaret Atwood's "Oryx and Crake" with post-edited versions and human translations, the authors reveal that machine translation (MT) systems leave distinct emotional fingerprints that significantly affect the preservation of the author's voice. The findings highlight that while post-editing can enhance emotional fidelity, it may still fall short of achieving the nuanced emotional expression found in human translations.
LLM translations leave unique emotional fingerprints that can distort an author's voice, even after post-editing.
This paper investigates whether LLM translations exhibit identifiable emotional profiles and how post-editing reshapes them toward human-like norms. We compare LLM translations of Margaret Atwood's Oryx and Crake with their post-edited versions and a human translation, using a large-scale corpus of contemporary Italian science-fiction as a baseline. We examine emotion through lexicon-based and multilingual modeling, conducting a fine-grained analysis of emotional variation across systems. We find that MT systems introduce model-specific and statistically significant emotional fingerprints across translations, leading to a limited preservation of an author's voice.