Feedback Engineering: Correct AI Translation to Human Accuracy

Workshop: Prompt Engineering for the Creative Translator · Article 3 of 4

By this point in the series, you have two powerful tools. Article 1 gave you context and persona — the framing that tells the model what kind of text it is handling and what kind of translator to be. Article 2 gave you Chain-of-Thought prompting — the technique that forces the model to reason through translation risks before committing to output.

Between them, these two approaches handle most of the work. But not all of it. There will always be a gap between the first output and the output you actually want to deliver to a client. Closing that gap is what this article is about.

Most translators close it the same way: they read the output, feel vaguely dissatisfied, and type something like “try again” or “make it more natural.” The model produces a second version — sometimes better, sometimes worse, always unpredictable. Several rounds later, they either give up or rewrite the whole thing manually.

Feedback engineering is the alternative. It is the discipline of giving the model correction instructions that are specific enough to work, structured enough to be repeatable, and targeted enough to fix one problem at a time without breaking what was already right.

Why Vague Feedback Fails

Before building the framework, it is worth understanding why generic feedback produces generic results.

When you tell a model “make this more natural,” you are giving it a direction without a destination. Natural according to whom? For what register? In which dialect? The model has no way to anchor the instruction, so it makes an arbitrary adjustment — usually toward the most generic Arabic prose in its training data, which is precisely the wrong direction for literary or specialized work.

The same problem applies to instructions like “this doesn’t sound right,” “the tone is off,” “make it flow better,” or “try a different approach.” Each of these tells the model that something is wrong without telling it where the problem is or what a correct version would look like.

Vague feedback produces vague corrections. The model cannot fix what it cannot locate. Your job in feedback engineering is to locate the problem precisely before you ask for the fix.

The Three-Layer Diagnostic

Every translation problem lives at one of three levels. Identifying the right level is the first — and most important — step in feedback engineering.

Layer 1 — Lexical: The wrong word was chosen. The problem is at the level of a specific term, name, or phrase. The surrounding structure is fine; only this element needs to change.

Layer 2 — Structural: The right words are present but in the wrong arrangement. The problem is at the level of sentence construction, clause order, or paragraph sequencing. The vocabulary is acceptable; the architecture needs rebuilding.

Layer 3 — Tonal / Register: Both words and structure are technically acceptable, but the overall feel is wrong. The register is too formal or too casual. The voice is off. The rhythm does not match the source. This is the hardest layer to diagnose because it requires comparing a feeling against a standard — but it can be done systematically.

Before writing any correction prompt, ask yourself: which layer is the problem actually at? Most translators misdiagnose this. They feel a tonal problem and try to fix it at the lexical level — swapping words when the real issue is sentence length or clause rhythm. Or they feel a structural problem and try to fix it by adding register instructions when the issue is actually word order.

Layer 1 Corrections: Targeted Lexical Feedback

Lexical corrections are the most straightforward. The template is:

In the translation you just produced, the word/phrase [X] is not working.

Problem: [explain specifically why — wrong connotation / wrong register /
culturally inappropriate / too archaic / too colloquial / imprecise / etc.]

Preferred option: Use [Y] instead.
Reason: [Y] better captures [specific quality — the legal register / the ironic
tone / the technical precision / the emotional weight / etc.]

Keep everything else in the translation unchanged.
Regenerate only the affected sentence(s).

The last two lines are critical. “Keep everything else unchanged” prevents the model from deciding to improve parts of the translation you were already satisfied with. “Regenerate only the affected sentence(s)” saves time and keeps corrections granular.

Here is a concrete example. The model translated a passage about digital identity and produced:

“يشعر المستخدم بالضياع داخل الفضاء الرقمي.”

The word ضياع is accurate but too melodramatic for a technical-analytical passage. Your correction prompt:

In the sentence "يشعر المستخدم بالضياع داخل الفضاء الرقمي":

Problem: "الضياع" carries emotional and existential weight that is too strong
for this analytical passage. The source text is clinical and observational,
not lyrical.

Preferred option: Replace with "الارتباك" or "فقدان الاتجاه."
Reason: These terms match the technical-analytical register of the passage
without the melodrama.

Keep everything else unchanged.

Layer 2 Corrections: Structural Feedback

Structural corrections are more complex because you are asking the model to rebuild, not replace. The template:

The following sentence/paragraph in the translation has a structural problem:

[quote the problematic passage]

Problem: [describe the structural issue — the main clause is buried / the
subordinate clauses are in the wrong order / the parallel structure is broken /
the sentence is too long / the Arabic connector obscures the logical
relationship / etc.]

What I need: [describe the target structure — shorter sentences / the main
claim front-loaded / the logical dependency made explicit / etc.]

Source sentence for reference: [paste the original English]

Rebuild only this sentence/paragraph. Keep the rest unchanged.

A typical structural problem in English-to-Arabic translation is what we might call connector inflation. English uses punctuation and short sentences to signal logical relationships. Arabic grammar pulls toward connectors — إذ، حيث، كما أن، والذي — that can bury the main idea under layers of subordination. When a translated paragraph feels dense and hard to follow, the problem is usually structural rather than lexical.

In that case, your correction might say: “The paragraph has become one long sentence connected by three instances of حيث. Break it into three separate sentences. Keep the vocabulary. Change only the architecture.”

Layer 3 Corrections: Register and Tone

Register corrections are the most nuanced because they require you to articulate a feeling. The key is anchoring the instruction to something concrete rather than leaving it as a vague aesthetic preference. Three anchors that work well:

Anchor 1 — A reference publication or author: “The register should read like Al-Hayat newspaper, not like a university textbook. Read like a Naguib Mahfouz interior monologue, not like a government report.”

Anchor 2 — A spectrum position: “On a scale from 1 (highly formal classical Arabic) to 10 (Egyptian street conversation), this translation should sit at approximately 3 — Modern Standard Arabic that reads easily without being journalistic.”

Anchor 3 — A list of specific contrasts: “The current version uses passive constructions where the source is active; uses abstract nouns where the source uses verbs; and uses multi-syllabic formal terms where the source uses short, punchy vocabulary.” This is the most powerful anchor for literary translation because it gives concrete structural targets rather than a general emotional impression.

The translation is technically accurate but the register is wrong.

Current problem: The translation reads as formal written Arabic appropriate
for academic publication. The source text is a first-person literary essay
with an intimate, slightly conversational tone.

Target register: Modern Standard Arabic that reads like a literary magazine.
Accessible but not colloquial. Personal but not chatty. Precise but not clinical.

Specific changes needed:
- Replace passive constructions with active voice where the source uses active
- Replace abstract nouns (التحقق، الاستيعاب) with verb forms (يتحقق، يستوعب)
- Shorten sentences that run over three clauses

Regenerate the full passage with only these adjustments. Do not change the meaning.

The Surgical Correction Principle

The single most important rule in feedback engineering: fix one problem per correction prompt.

The temptation is to bundle everything into a single correction: “The register is too formal, the sentence structure is too complex, the word choice in the third paragraph is wrong, and the ending doesn’t land.” This produces output that changes in unpredictable ways — the model tries to address everything simultaneously and overshoots on some dimensions while undershooting on others.

Instead, work in sequence:

1. Fix the most fundamental problem first — usually structural. If the architecture is wrong, lexical fixes will not hold.
2. Once the structure is right, address register.
3. Once register is right, address specific lexical choices.
4. Finally, review rhythm and flow as a whole.

This sequence matters because lower-level corrections are fragile. Lexical choices get disturbed when structural or tonal corrections are applied afterward. Work top-down, not bottom-up.

Think the way a film editor works: fix the cut before you color-grade. If the structure changes, you have to color-grade again. Fix structure first, then surface.

Preserving What Is Already Working

One of the most common feedback engineering mistakes is failing to tell the model what to keep alongside what to change. When you ask for a correction, the model will sometimes decide — helpfully, from its perspective — to improve parts of the translation you were perfectly satisfied with. You get back a version that fixed the problem you identified but broke three things you had not mentioned.

The solution is to make preservation explicit:

In the following correction, keep these elements exactly as they are:
- The translation of paragraph 1 (the opening is working well)
- The term "الفاعلية المعرفية" (this is the correct technical term for this field)
- The sentence structure of the final paragraph

Change only:
- [specify the exact change]

Produce the full translation with only this modification applied.

This pattern — explicit preservation alongside explicit change — dramatically reduces variance in correction outputs. The model has a clear boundary around what is fixed and what is adjustable.

A Complete Correction Workflow: From Draft to Delivery

Here is a full feedback engineering session in practice, applied to a short literary passage. We start with a hypothetical first output and work through two correction rounds.

Source text (English):
“The city had not changed, but he had. He stood at the corner where the old café used to be and felt the particular loneliness of returning.”

First output (from a well-contextualized prompt):
“لم تتغيّر المدينة، لكنه هو الذي تغيَّر. وقف عند الزاوية التي كان المقهى القديم يقع فيها، وأحسّ بالوحدة الخاصة التي يشعر بها من يعود إلى مكانٍ ما.”

Assessment: The translation is accurate — but “الوحدة الخاصة التي يشعر بها من يعود إلى مكانٍ ما” unpacks “the particular loneliness of returning” into an explanatory relative clause that drains the compression the source depends on. The original’s power is in its brevity. Three words carry a whole recognizable emotional category. The translation describes the feeling rather than naming it.

Correction Round 1 — Structural:

In the second sentence, the phrase "الوحدة الخاصة التي يشعر بها من يعود إلى مكانٍ ما"
has been unpacked where the source compresses.

Problem: "the particular loneliness of returning" names a recognizable emotional
experience in three words. The translation turns it into a relative clause that
explains rather than evokes.

Target: A compressed Arabic phrase of similar economy — something that names
the emotion rather than describes it. Options: وحشة العودة / غربة العائد / وحدة من يعود

Keep the first sentence exactly as is. Apply the correction to the second sentence only.

Second output (after Round 1):
“لم تتغيّر المدينة، لكنه هو الذي تغيَّر. وقف عند الزاوية التي كان المقهى القديم يقع فيها، وأحسّ بوحشة العودة.”

Assessment: The compression is restored. But “الزاوية التي كان المقهى القديم يقع فيها” reads slightly administrative for a literary passage. A small lexical fix remains.

Correction Round 2 — Lexical:

Minor lexical fix: "الزاوية التي كان المقهى القديم يقع فيها" reads as slightly
administrative. Replace with "زاوية المقهى القديم" — shorter, more idiomatic,
appropriate for the literary register. Keep everything else exactly as is.

Final output (after Round 2):
“لم تتغيّر المدينة، لكنه هو الذي تغيَّر. وقف عند زاوية المقهى القديم، وأحسّ بوحشة العودة.”

Two targeted corrections. No full regenerations. The final output is publishable.

When to Stop Correcting and Rewrite Manually

Feedback engineering has limits. There are cases where iterative correction becomes less efficient than simply rewriting a sentence or paragraph yourself. Knowing when to stop is a professional skill in its own right.

The rule of thumb: if you have made three correction attempts at the same problem and the model keeps missing it, the problem is at a level of nuance your prompt cannot fully capture in natural language. At that point, write the sentence yourself, paste it back into the conversation, and tell the model: “Use this sentence exactly as written. Continue applying the same approach to the rest of the passage.”

Manual rewriting and AI correction are not competing strategies — they are complementary tools. Use AI for the 90% that responds to structured prompting. Use your own professional judgment for the 10% that requires a translator’s ear.

For a broader understanding of how iterative self-review works at the model architecture level — including why some corrections are more stable than others — see our article on self-reflection prompting: (See our article: Self-Reflection and Recursive Self-Improvement Prompting)

What’s Next

Article 4 closes this series with the question of permanence. The prompts you develop — your persona instructions, your Chain-of-Thought templates, your correction patterns for specific text types and dialects — represent real intellectual capital. Article 4 shows you how to organize them into a searchable, reusable prompt library so that every project makes the next one faster. (See our article: Building Your Translator’s Prompt Library for Dialects and Specialized Terminology)

If you want to explore how correction and verification prompting can be applied beyond translation — to fact-checking, logical consistency, and output validation — see our piece on anti-hallucination prompting: (See our article: Anti-Hallucination Prompting: Self-Consistency, Chain-of-Verification and Reliable AI Outputs)

References

1. Madaan, A. et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023. arxiv.org/abs/2303.17651
2. Saunders, W. et al. (2022). Self-critiquing models for assisting human evaluators. arxiv.org/abs/2206.05802
3. Nida, E. & Taber, C. (1969). The Theory and Practice of Translation. Brill.
4. Nord, C. (1997). Translating as a Purposeful Activity: Functionalist Approaches Explained. St. Jerome Publishing.