Emotion Profiling in LLM-Based Literary Translation: Systematic Shifts Across MT and Post-Editing

Researchers at arXiv analyzed how large language models introduce distinctive emotional signatures when translating literary works, finding that LLM translations preserve author's voice less effectively than human translations. Post-editing partially corrects these emotional distortions, but MT systems consistently exhibit model-specific emotional fingerprints that deviate from human translation norms.

This study addresses a critical limitation in machine translation technology that extends beyond surface-level accuracy metrics. While translation quality is typically measured through BLEU scores and semantic fidelity, the research reveals that LLMs inject measurable emotional biases into translated text—a dimension that substantially impacts literary and creative works where tone and voice are paramount. The researchers employed multilingual emotion modeling to detect these patterns across an LLM translation of Margaret Atwood's Oryx and Crake, comparing outputs against post-edited versions and professional human translations.

The findings suggest each LLM architecture introduces consistent emotional distortions, meaning GPT-based systems may systematically over-emphasize certain emotional registers while dampening others. This occurs because language models learn statistical patterns from training data that don't perfectly align with the nuanced emotional palette of literary authors. Post-editing by human translators partially corrects these artifacts but cannot fully restore an author's original emotional intent, indicating a fundamental gap in how current systems process affective dimensions of language.

For the translation industry, this research signals that automated MT systems require supplementary emotional analysis layers to serve literary markets effectively. Publishing houses and translation platforms relying on LLM-assisted workflows may inadvertently homogenize translated literature, reducing stylistic diversity in international markets. The implications extend to localization services, accessibility tools, and any domain where emotional authenticity matters. Organizations adopting these systems should implement emotion-aware post-editing protocols and acknowledge that creative translation remains poorly suited to current automation approaches.

• →LLM translation systems introduce statistically significant and model-specific emotional fingerprints that deviate from human translation norms.
• →Post-editing partially mitigates emotional distortions but cannot fully restore an author's original emotional intent in translated works.
• →Each LLM architecture exhibits consistent emotional biases, suggesting systematic patterns in how these models process affective language.
• →Literary translation quality metrics must expand beyond semantic accuracy to include emotional authenticity and author voice preservation.
• →Organizations using LLM-assisted translation for creative content should implement emotion-aware quality assurance protocols before publication.