Translating Chinese Fiction: The Four-Layer Surgery Model
Machine translation gives you a passing grade on Chinese fiction; the gap to polished prose is four layers of structural loss — density, context decoding, voiceprint, and active editing. A workable workflow uses LLMs as decoders rather than generators.
Machine-translating Chinese web fiction into English produces a passing-grade draft. The gap between passing and polished is four layers of structural loss, each requiring different intervention. Density surgery addresses the fact that Chinese is a high-density language and idioms are mood-compression packages. 草长莺飞 (four characters, 'grass grows, orioles fly') automatically unpacks for a Chinese reader as 'late spring, warmth, everything growing, optimism' — translating it as 'the grass was growing and the orioles were flying' is a corpse translation. The right English is 'a fine spring day.' Chinese-to-English: subtract. English-to-Chinese: add. When translated novels feel overexplained, the translator was filling in compression blanks for Western readers and killed the original compression in the process. Context decoding is where machine translation fails hardest. Idioms shift meaning by scene — 落花有意，流水无情 means heartbreak in a romance scene but impermanence in an ascension scene. Subtext in dialogue is more dangerous than insults: Yinyang guaiqi (阴阳怪气), weaponized politeness via smiling compliment, is recognized on sight by Chinese readers but invisible to English readers unless body language carries the tell. The rule is to translate what was not said, not what was — translate the knife, not the smile. Voiceprint override is the translator's own language habits smoothing distinct authors into the same voice. Running a symphony and a rock album through the same MP3 encoder produces the same compression artifacts. This is why translated Chinese novels often 'feel like they were all written by the same person' — the translation process ate the differences. For informational content it is harmless; for fiction it replaces the author's ghost with the translator's. Active editing is where human judgment lives: restructuring stacked Chinese adjectives ('moonlight as if washed clean, gentle breeze, lake shimmering with ripples' becomes 'Moonlight on still water'), cutting storyteller-style narration that reads as author intrusion in English, and pushing physical details into political metaphor ('a pot of cold tea' becomes 'a pot of tea no one would finish' when the alliance just cracked). The workable workflow: run source through Google Translate and DeepL in parallel; mark the places they diverge meaningfully as the spots context decoding broke; feed marked passages to an LLM as a decoder (ask what the subtext is, not for a translation); then do the manual restructuring. Read the final version standalone without the Chinese present — if it stands up as English, it is done. The LLM as decoder move — asking what a passage means rather than asking for translation — is the single most useful trick. A related concept is Liubai (留白, 'leaving things unsaid'), the Chinese writing principle where empty space between sentences carries as much weight as the words. The reader is expected to fill in what was not written. Western prose works the opposite way: show it, say it, make sure nothing is missed. Neither is better; they are built on different assumptions about the reader's job. The translation damage is that the liubai was there in the original and the translator filled in the blanks 'to make sure nothing got missed' — what gets read is not what the author wrote.