Why Your Literary Translation Needs a Human Eye

Why Is My Literary Translation So Stiff?

You've just run your latest short story through DeepL, and the result reads like a robot wrote it. The plot is there, but the soul is missing. You're not alone. Every literary translator has faced this moment of doubt. The question is: what do you do about it?

Here's the blunt truth: machine translation (MT) is not ready to replace you. It's a tool, not a translator. And for literary work, it's a tool that needs a skilled human hand to guide it. The machines are getting better, sure, but they still stumble on the very things that make literature sing: idiom, nuance, and cultural context.

The Limits of Machine Translation

Machine translation has evolved dramatically since the rule-based systems of the 1950s, when the Georgetown-IBM experiment used a 250-word vocabulary and six grammar rules to translate Russian sentences into English (Georgetown-IBM experiment). Today, neural machine translation (NMT) systems like Google Translate and DeepL analyze whole sentences, not just words, and they've gotten impressively fluent (Machine translation). But fluency isn't the same as fidelity.

Consider a line like "She had a heart of gold." A literal MT output might render it as "Her heart was made of gold," which loses the idiomatic weight. And when you're translating poetry, where every word carries multiple layers of meaning, a machine's tendency to pick the most literal option can flatten the text into prose.

Why Human Post-Editing Is Non-Negotiable

This is where human post-editing comes in. The ISO 18587:2017 standard defines full post-editing as a process that requires human intervention to ensure the translation meets a quality level comparable to human translation (ISO). For literary work, you need full post-editing, not just a quick once-over. You need to rework sentences, choose the right metaphor, and capture the author's voice.

Machine translation output often requires human post-editing, especially for idiomatic expressions, cultural nuance, and domain-specific terminology (Machine translation). Literary translation is the ultimate test of this. A machine might translate "raining cats and dogs" literally, but a human post-editor knows to render it as "raining buckets" or find an equivalent idiom in the target language.

What About the New Evaluation Metrics?

You might have heard about BLEU, the metric that IBM researchers introduced in 2002 to score machine translation by comparing n-grams to human references (IBM Research). But BLEU is increasingly seen as unreliable because it correlates poorly with human judgment (IBM Research). That's why newer metrics like COMET and BLEURT are becoming standard (IBM Research).

But here's the thing: these metrics measure how close MT output is to a reference, not how it reads as literature. For literary translation, you need a human reader who can judge whether the text flows, whether it evokes the same feelings, and whether it honors the original's style.

Using MT as a Starting Point

That doesn't mean you should throw MT out. Used wisely, it can be a time-saver. You can run a draft through DeepL to get a rough version, then spend your energy on polishing and re-creating. This is what computer-assisted translation (CAT) tools have done for years: they split documents into segments, pre-translate them from translation memory, and fall back on machine translation when no match is found (Google Translator Toolkit).

But remember: MT is fully automated and replaces human translation during the translation phase, while CAT is machine-assisted human translation where the translator keeps control (Machine translation). For literary work, you want to keep control. Use MT as a drafting tool, not as the final word.

How to Work with MT Effectively

So, how do you actually do this? First, accept that MT output will be uneven. Some sentences will be perfect; others will be unusable. Your job is to identify which is which.

• Read the MT output aloud. If it sounds like a robot, rewrite it.
• Check for idioms and cultural references that MT might have mangled.
• Pay close attention to dialogue, where tone and rhythm matter most.

For example, let's say you're translating a French novel into English. A sentence like "Il a pris son courage à deux mains" literally means "He took his courage in both hands," but the English idiom is "He plucked up his courage." A machine might produce the literal version, but a human post-editor would know to use the idiomatic English.

What I'd Actually Do

Here's my recommendation: if you're a literary translator, don't rely on MT as your primary tool. Use it as a source of inspiration or a quick draft, but always do full post-editing yourself. And if you're a publisher or an author hiring a translator, spec that the final work must be human-translated, with MT used only as a reference if at all.

The market is shifting—language services revenue dipped to $49.68 billion in 2023 (CSA Research)—but literary translation is a craft that resists automation. The best translators I know use MT sparingly and rely on their own ear for language. That's what I'd do: let the machine do the grunt work, but never let it have the last word.

Sources

• Machine translation - https://en.wikipedia.org/wiki/Machine_translation
• IBM Research BLEU - https://research.ibm.com/blog/bleu-nlp-benchmark-anniversary
• Google Translator Toolkit - https://en.wikipedia.org/wiki/Google_Translator_Toolkit
• CSA Research - https://csa-research.com/l/media/Language-Services-and-Technology-Industry-Faces-Revenue-Decline-but-Remains-Poised-for-Transformation