Bilingual Restaurant Menu Generator With AI: A Translation QA Workflow

The ALIGN framework for bilingual menu control

Use **ALIGN** as a repeatable hand-off between menu operations, translation and publishing:

1. **Approve the master.** Establish one source language and freeze the item list for the translation round. Include names, descriptions, ingredients, modifiers, prices, dietary labels and approved allergen or cross-contact wording. A half-finished source produces two uncertain versions. 2. **Lock terminology.** Create a glossary for recurring ingredients, techniques, place names, brands and dietary terms. Record the exact target-language form and identify terms that must not be translated. 3. **Interpret dish names.** Assign one treatment to each name: retain, transliterate, translate or explain. This is a policy decision, not an invitation for the model to improvise item by item. 4. **Gate critical facts.** Compare factual fields with recipe, supplier and kitchen records. AI may draft language; it cannot inspect those records or determine whether conditions have changed. 5. **Normalise every format.** Reconcile language order, item order, prices, labels and notices in the hosted, QR and print presentations. Approve the content and the rendered layout separately.

ALIGN separates three questions that are often collapsed into one: Is the wording fluent? Is the underlying information true? Is the approved information visible in the format the guest receives? Different reviewers may own those answers.

Build one master record before translating

Choose the language that the kitchen and front-of-house team can verify most confidently. Give every item a stable internal identifier; the identifier should remain unchanged even if the displayed name changes. This makes it possible to match corresponding rows without relying on similar wording.

For each item, prepare these fields:

Update this master first whenever a recipe, supplier, price or availability status changes. Then reopen the target-language row. Do not silently repair only the translated version, because that creates two competing sources.

A glossary should be equally controlled. Keep separate entries for a word’s culinary meaning and its everyday meaning when confusion is possible. Include capitalisation, accents and plural forms. For each entry, record a preferred target term, a prohibited alternative if one causes ambiguity, and a short note explaining the choice. The note helps the next reviewer apply the decision consistently rather than starting over.

• source name and description;
• confirmed ingredients and preparation details that belong on the menu;
• modifiers, portions and prices;
• approved dietary labels;
• approved allergen and cross-contact wording, where used;
• availability status and menu section;
• revision owner and review date.

Choose a dish-name treatment deliberately

The right treatment depends on what the name communicates to the intended guest. Apply the following decision table before asking AI for target copy.

| Treatment | Choose it when | Target-menu pattern | Human check | |---|---|---|---| | Retain | The original name is a recognised identity or proper name | Original name + concise target-language explanation | Spelling, accents, cultural context | | Transliterate | Readers need the sound represented in another writing system | Transliteration + short explanation | Pronunciation convention and ambiguity | | Translate | The name is descriptive and a natural equivalent is clear | Natural target-language name | Meaning, tone and ingredient accuracy | | Explain | A literal version would not tell the guest what the dish is | Original name + factual explanation | Cuisine knowledge and source-record match |

Do not make “retain everything” or “translate everything” the default. A short explanation can preserve the original name while helping a guest understand the principal ingredient or preparation. Avoid adding an ingredient, origin story or quality claim merely to make the wording more appealing.

Bilingual Menu Translation QA Matrix

Create one row per item. Copy this matrix into a spreadsheet or content workflow and expand the evidence references as needed.

| Field | Source-language entry | Target-language entry | Verification or decision | Owner | Status | |---|---|---|---|---|---| | Item ID | Stable internal code | Same code | Exact match | Menu owner | Open / pass | | Dish name | Approved display name | Drafted name | Retain / transliterate / translate / explain | Language reviewer | Open / pass / escalate | | Description | Approved factual copy | Drafted copy | Meaning and tone reconciled | Language reviewer | Open / pass / revise | | Ingredients | Current menu-facing list | Translated list | Checked against current recipe record | Kitchen owner | Open / pass / stop | | Allergens and cross-contact wording | Approved source wording | Translated wording | Checked against restaurant records and applicable review process | Safety owner | Open / pass / stop | | Dietary labels | Approved labels | Equivalent labels | Criteria and wording confirmed | Safety owner | Open / pass / stop | | Modifiers and portions | Approved choices | Translated choices | Options, quantities and defaults match | Service owner | Open / pass / revise | | Price | Approved price | Same price and currency display | Character-for-character comparison | Menu owner | Open / pass / stop | | Glossary term | Term ID or “none” | Approved target term | Glossary match | Language reviewer | Open / pass / revise | | QR layout | Source rendering | Target rendering | No clipping, omission or wrong order | Publisher | Open / pass | | Print layout | Source proof | Target proof | Legibility, line breaks and notices checked | Publisher | Open / pass | | Review record | Version and date | Reviewer and date | Accountability complete | Menu owner | Open / pass | | Escalation | Question or “none” | Decision or pending | Named decision-maker | Menu owner | Open / resolved / stop |

A row is ready only when every required field is passed or deliberately marked not applicable. “Looks fine” is not a status. If a field has no current evidence, mark it **stop** rather than asking AI to fill the gap.

Example: a hypothetical Spanish-to-English item review

This example is invented solely to demonstrate the matrix; it is not a customer menu, a test result or an authoritative translation for another restaurant.

Assume the approved Spanish master row is:

The AI draft returns “Mushroom tacos — corn tortillas, roasted mushrooms, pickled onion and seed sauce”, with “Add avocado” as the modifier. Instead of approving the smooth wording immediately, the team processes it field by field:

1. **Name treatment:** Translate. “Mushroom tacos” is descriptive, but a fluent reviewer checks whether the chosen mushroom wording fits the actual varieties and local usage. 2. **Ingredient reconciliation:** Each listed component maps to the approved source. “Seed sauce” is understandable but too unspecific for the restaurant’s own approved record, so the language reviewer proposes “sesame seed sauce”. The kitchen owner, not the AI, confirms that this remains factually accurate. 3. **Safety gate:** The source description does not display the restaurant record note. Publication stops until the designated safety owner decides the approved sesame and shared-surface wording for both languages. The model is not asked to invent it. 4. **Modifier and price:** “Add avocado” and 12 are compared with the master fields, including any currency display used by the full menu. 5. **Format check:** Reviewers inspect both language layouts. If the English safety wording wraps below another item or disappears from the print proof, the content may be correct while the format still fails. 6. **Approval record:** The row receives the target-language reviewer, kitchen or safety owner, date and version. Any later recipe change reopens T-17 in both languages.

The useful lesson is the stop condition: a fluent draft does not overrule an incomplete source menu. The workflow exposes the missing decision before publication.

• **Item ID:** T-17
• **Name:** Tacos de hongos
• **Description:** Tortillas de maíz, setas asadas, cebolla encurtida y salsa de semillas
• **Modifier:** Añadir aguacate
• **Price:** 12
• **Restaurant record note:** the seed sauce contains sesame; preparation uses a shared surface

Implementation steps from draft to publication

1. Define ownership.

Name the person responsible for the master menu, the target-language reviewer, the kitchen or safety reviewer, and the publisher. One person may hold several roles, but each approval must still be explicit.

2. Freeze a translation batch.

Assign a version and date to the source menu. Remove unavailable items and resolve unfinished source copy before generation. Keep late changes in a change log instead of editing an untracked copy.

3. Prepare the prompt inputs.

Give the AI only approved source fields, the target language, desired reading level or tone, glossary entries and dish-name policy. Instruct it not to add ingredients, origin claims, dietary labels or preparation details. Ask for field-preserving output so omissions are easier to detect.

4. Run language and fact reviews separately.

The language reviewer checks naturalness, meaning, register and cultural clarity. Operational owners check recipes, modifiers, prices, availability, dietary labels and safety wording. Fluency is not evidence that those facts are current.

5. Resolve exceptions.

Escalate unfamiliar regional terms, disputed names and any wording with legal or safety implications. Record the decision in the glossary or item row so future batches use the approved treatment.

6. Proof the actual guest formats.

Check mobile and print versions rather than reviewing only a text document. Compare item count, sequence, prices, modifiers, language labels and notices. Look for clipping, displaced accents, poor line breaks and text that becomes too small after the second language is added.

7. Publish and maintain.

Archive the approved version and keep a simple change log. A recipe, supplier, allergen process, price or item-status change should reopen all affected language rows. If you need an editable starting point, [MenuCrafters](https://menucrafters.com/) publicly supports editable menu drafts, translation of individual items, a hosted menu and print-ready PDF output; the same human verification gates still apply.

Pre-publication stop/go checklist

**Go only when:**

**Stop and escalate when:**

• every visible item has the same stable ID, status, modifiers and price in both languages;
• dish-name treatment follows the recorded policy;
• glossary terms are applied consistently;
• a fluent reviewer has approved the target-language meaning and tone;
• restaurant owners have verified recipe-, supplier-, dietary- and allergen-dependent fields from current records;
• QR and print proofs preserve item order, notices and readable formatting;
• reviewer names, dates and version are recorded.
• a source field is missing, contradictory or out of date;
• a recipe, supplier product or preparation condition cannot be checked;
• an allergen, cross-contact or dietary claim depends on an assumption;
• cultural nuance or transliteration is disputed;
• jurisdiction-specific wording has not been reviewed through the restaurant’s appropriate process;
• AI adds a fact that is absent from the approved master;
• the second language causes critical wording to be clipped, separated or unreadable.

Limitations and when this approach is not appropriate

AI cannot inspect a kitchen, recipe file, supplier specification, package label or preparation surface. It cannot establish whether an allergen or dietary statement is true, whether a product changed, or whether wording meets requirements in a particular jurisdiction. It also cannot guarantee culturally appropriate phrasing simply because a sentence is grammatical.

The matrix adds control, but it adds work. It may be disproportionate for a temporary internal draft that guests will never see; it is appropriate when menu content is moving towards publication. It also does not replace professional language, safety or legal review where the restaurant’s risk, audience or local obligations call for it.

A bilingual layout can become crowded. If both languages cannot remain readable in print, use a carefully signposted alternative format rather than reducing type until it is difficult to read. A hosted menu can make revisions easier, but it does not fix weak source records or remove the need to check the live rendering.

Finally, a two-language workflow does not automatically serve every guest. Language choice, reading level, accessibility and staff support remain separate operational decisions. Add another language only when the restaurant can maintain it through the same update and approval process.