Improving machine-translated novels via style transfer — looking for advice on the faithfulness/fluency tradeoff [P]

Summary

Hey all. I recently started working on a project to improve machine-translated webnovels via style transfer. The basic idea is to take the clunky translated prose and rewrite it to something that reads like it was written by a professional author, while remaining as faithful as possible to the original text. The source material is mostly amateur/MTL output full of direct sentence structure translations carried over from Chinese, awkward honorifics, over-translated idioms, that kind of thing. The

SpinGraph analysis pending — check back after processing.

• What does "Safe AI" look like? [D]
• Small Language Model SLM [D]
• Tom Yeh's AI by hand? is it worth it? [D]
• I built my 'first' flow matching image generator, here's what I learned [P]
• H64LM: A 249M-parameter Mixture-of-Experts Transformer built from scratch in PyTorch [P]
• Contrastive Decoding Diffing (CDD): recovering verbatim finetuning data from logits alone, no weight access needed[R]