Can ChatGPT Break Through the ‘Fortress’ of Literary Translation? A Computational Stylistic Analysis of Human vs. Machine Translation

I went to this conference under the scholarship from Tongji University. Thanks!

Presented by:
KONG Delu
PhD Candidate, Department of English Language and Literature
School of Foreign Studies, Tongji University, Shanghai
Contact: Kongdelu2009@hotmail.com

Research Overview

Corpus:

• Chinese prose classics: “Moonlight on the Lotus Pond” and “The View of Father’s Back” by Zhiqing ZHU
• Translation variants:
  • 7 Human Translations (HT)
  • 14 Neural Machine Translations (NMT)
  • ChatGPT-generated translations

Methodology:

• Text similarity algorithms (Word2Vec embeddings)
• Cluster analysis for reliability verification
• Corpus-based stylistic comparison

Key Findings

1. Stylistic Variation Spectrum:
   • HT: Highest variation (Avg similarity = 0.42)
   • ChatGPT: Intermediate variation (Avg similarity = 0.67)
   • NMT: Most consistent (Avg similarity = 0.81)
2. Human-Like Tendencies:
   • ChatGPT achieved 28% greater stylistic diversity than NMT
   • Captured 61% of human translator variation range
3. Critical Limitations:
   • Failed to replicate human “stylistic leaps” in literary adaptation
   • Showed systematic bias in metaphor handling