Multilingual & Culturally-Adapted Materials

Moderate · Communication & Dissemination track · ~35 min hands-on + readings and quiz

← Back to the Communication & Dissemination track

Reach more communities by translating and adapting materials — with a quality check, not blind trust.

What you’ll be able to do

  • Translate outreach materials into another language with AI
  • Adapt for cultural context, not just words
  • Quality-check with back-translation and a human reviewer

Overview

AI translates quickly and can adapt tone for a community, but machine translation makes confident errors — especially with idioms, health terms, and culturally specific concepts. A human speaker should review anything that reaches the public.

Cultural adaptation is more than translation: examples, framing, and imagery may need to change. AI can propose options; community knowledge decides.

Practice activities

Activity 1 · Moderate — Translate with back-translation QA

Time ~15 min · Tools ChatGPT Edu (or an approved translation tool)

Goal. Translate an outreach message and catch errors before it goes public.

Setup. Use this pedestrian-safety outreach message and translate it into Spanish (the most common non-English language in Philadelphia):

Walk Safely in Your Neighborhood

Pedestrian crashes in Philadelphia peak between 4 and 7 PM, especially on wide, busy roads. Here’s how to stay safer:

  • Cross at marked crosswalks and intersections — drivers are more likely to see you there.
  • Make eye contact with drivers before stepping into the street.
  • At night, wear something light-colored or reflective. Carry a flashlight if you can.
  • If there’s no sidewalk, walk facing traffic so you can see oncoming cars.
  • Report broken streetlights, missing crosswalk markings, or damaged sidewalks to 311 (call 311 or use the Philly311 app).

Your safety matters. If a street in your neighborhood feels dangerous to walk on, you can request a traffic-safety review from the city at [link].

Steps.

  1. Translate:

    Translate this community health message into [language], keeping a warm, plain-language tone. [message]

  2. In a fresh chat, back-translate:

    Translate this [language] text back into English, literally. [translation]

  3. Compare the back-translation to your original and investigate any drift in meaning.

Expected result. A translation plus a back-translation you’ve compared.

Check your work. A bilingual colleague should review anything published — machine translation of health terms is error-prone.

Common pitfalls. Idioms and clinical terms mistranslate confidently. Never publish machine translation of health materials unreviewed.

Stretch (optional). Ask which sentences are most likely to be misunderstood in the target language.

Activity 2 · Moderate — Culturally adapt, not just translate

Time ~13 min · Tools ChatGPT Edu

Goal. Adapt framing and examples for a specific community, not just the words.

Setup. Your Spanish translation from Activity 1, plus this community context: the target audience is Spanish-speaking residents of North Philadelphia, many of whom are originally from Puerto Rico or the Dominican Republic. The neighborhood has high pedestrian crash rates, limited sidewalk infrastructure, and several schools within walking distance.

Steps.

  1. Surface adaptation needs:

    For [community], how might the examples, framing, or imagery in this message land differently? Suggest culturally appropriate adaptations. [message + community]

  2. Draft an adapted version; mark what a community reviewer must confirm.

  3. Check for harm:

    Flag anything in my adaptation that could unintentionally stereotype or exclude.

Expected result. An adapted draft plus a list to verify with community members.

Check your work. Cultural knowledge belongs to the community — AI suggestions are hypotheses to confirm, not facts.

Common pitfalls. AI can introduce stereotypes while “adapting.” Review with people from the community.

Stretch (optional). Make two variants and plan a quick community feedback check.

Activity 3 · Moderate — Scale translation QA across languages with Claude Cowork

Time ~20 min · Tools Claude Cowork (consumer/PI-purchased — not Drexel-supported; Low Risk Data only)

Goal. Run the same translate-then-back-translate QA check from Activity 1 across three languages at once, the way a real multilingual campaign actually goes out.

Setup. Real outreach rarely ships in one language. Spanish is Philadelphia’s most common non-English language, but DSPH projects often also need Vietnamese, Chinese, or others — and each version needs the same QA check. Instead of re-prompting a chat window three times, connect a folder and let Cowork work across all the files at once. Create a folder containing one file, original-message.txt, with the same pedestrian-safety message from Activity 1 (it’s already public and non-sensitive — keep it that way; nothing above Low Risk goes into Cowork). Connect that folder to a Cowork session.

Steps.

  1. Ask Cowork to translate into all three languages, one file per language:

    In the connected folder, read original-message.txt. Translate it into Spanish, Vietnamese, and Simplified Chinese, keeping a warm, plain-language tone. Save each as translation-[language].txt — three separate files.

  2. Then have it back-translate each one, as literally as possible:

    Now back-translate each translation-[language].txt into English, as literally as possible, without looking at the original. Save each as backtranslation-[language].txt.

  3. Finally, ask for a single consolidated QA table:

    Compare each back-translation against original-message.txt. Create qa-report.md with a table: one row per language, columns for sentences that drifted in meaning, the nature of the drift (idiom, health term, tone), and a severity flag. Flag any language where meaning shifted.

  4. Open the folder: you should have three translations, three back-translations, and one QA report. Read the flagged rows and spot-check them against the files.

Expected result. Seven files — one original, three translations, three back-translations — plus a consolidated QA table telling you where to look first in each language.

Check your work. The QA table is a triage step, not a sign-off. A bilingual human must review each language’s version before anything is published — same rule as Activity 1, now times three.

Common pitfalls. The same model translating and grading itself can miss its own confident errors — that’s exactly why the human review rule doesn’t scale away. And remember Cowork’s data status: Low Risk (public) content only, unless separately approved.

Stretch (optional). Ask Cowork to add a column to the QA table noting reading level per translation, and whether any language’s version drifted above plain-language level.

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources