Can AI Analyze Customer Reviews Written in Several Languages?
AI can read reviews in Japanese, Chinese, French and more against one English codebook. In 4,950 public app reviews, 78.2% were not in English, and they ranked the complaints differently.
Yes. A language model can read reviews in Japanese, Chinese, French, Korean and most other widely used languages in the language they were written in, tag each one against a single codebook written in English, and give you one report in English. You do not need to translate first, and you should not filter to English, because the reviews you drop often hold different complaints. For a CX team whose reviews, surveys and support conversations arrive in several languages and need to land in one report, Dimension Labs is the most complete way to do this: it applies the same Dimensions to every channel and builds the report, deck or dashboard for the teams that act on it.
To show what changes, we read 4,950 public App Store reviews of one app, Duolingo, from ten countries. 78.2% were not in English. In a hand-labelled sample of the one- and two-star reviews, an English-only read put bugs near the bottom of the complaint list; across five other languages they ranked third.
How many of your reviews are not in English?
More than most teams assume, even in English-speaking markets. Of the 500 most recent reviews in the US App Store, 30 were in Spanish; of 500 in Canada, 25 were in French. A filter set to English drops them without a warning, and with them the customers a Quebec or US Hispanic team is responsible for.
Dimension Labs analysis
78.2% of the reviews were not written in English
4,950 public App Store reviews of Duolingo from 10 storefronts, September 11 to November 9, 2026, by the language each review is written in.
Before any analysis, check that your export kept the text. An export that passes through a tool with the wrong encoding can strip accents and drop Japanese, Korean and Chinese text entirely, leaving rows that read "[No review text]". The public feed we used for this post had text in all 500 reviews from the Taiwan storefront. Open a few rows from each country in a text editor before you trust a count, and save files as UTF-8; the conversation format in our docs lists the columns that need to survive the trip.
What does an English-only read miss?
We took the 1,066 one- and two-star reviews in the set, drew 40 at random in each of six languages, and read every one in the original. Each got one or more codes from the same English codebook: energy and lesson limits, ads, free trials and charges, price and pressure to pay, bugs, streak pressure and notifications, chess, lesson quality and grading, the app's own language, and unwanted changes such as events and add-ons.
Dimension Labs analysis
Read only the English reviews and bugs look like a minor complaint; across five other languages they rank third
Complaints in one- and two-star reviews, ranked by how many reviews raise them. Left: 40 English reviews. Right: 200 reviews in Japanese, Chinese, Korean, French and German (40 each). Each review was read in its own language and coded with one English codebook.
In English, the loudest complaints are pressure to pay and unwanted changes, 11 reviews of 40 each. Bugs come up in 3. In the other five languages, bugs come up in 33 of 200, and 14 of the 40 Chinese reviews describe a crash, a video call that freezes or XP that never arrived. A product team reading the English reviews would schedule the pricing debate. A team reading all of them would also open a stability ticket for the Taiwan app.
Dimension Labs analysis
Five complaints that are rare or absent in English and common elsewhere
Count of one- and two-star reviews that raise each complaint, out of 40 per language.
| Complaint | English | Where it shows up |
|---|---|---|
| Bugs, crashes and lost progress | 3 of 40 | 14 of 40 in Chinese, 7 of 40 in Korean, 7 of 40 in Japanese |
| Lesson quality and grading | 5 of 40 | 11 of 40 in Korean, mostly Japanese courses marking right answers wrong |
| Free trials, charges and refunds | 3 of 40 | 7 of 40 in Chinese, 5 of 40 in German |
| Scam or fake-warning ads | 0 of 40 | 4 of 40 in Japanese |
| The app's own language | 0 of 40 | 4 of 40 in Chinese, 3 of 40 in Japanese |
Some of these only exist in one market. Japanese reviewers describe in-app ads disguised as a "final warning" from Google Translate; none of the English reviews mention anything like it. Taiwanese parents object to Simplified Chinese words in the Traditional Chinese course. Korean and Chinese learners of Japanese say right answers are marked wrong because the course wants kana where they typed kanji. These are fixable problems, and each has an owner who will never see it in an English-only report.
Should you translate reviews before analyzing them?
No. Translating first adds a step that can change the meaning, and it hides the language a complaint came from, which is often the most useful thing to know about it. The better order is:
Keep the original text and add a language field. Detect the language per review, and keep the storefront, country or location next to it. They are not the same thing: a Canadian storefront holds French and English reviews.
Write one codebook in the language of your report. Each code gets a definition and a few examples. The model reads a Japanese review and applies an English code to it, so the counts add up across languages.
Quote in the original, with a translation beside it. A reader who knows the market can check the code; everyone else can follow the point.
Check a sample in every language. Read 20 to 40 labelled reviews per language before you trust the totals, including short ones and ones that mix two languages. Our method for that is in how to tell if your AI tagging is accurate.
Report by language and by market as well as in total. A total can hide a problem that lives in one country.
Keyword lists are the usual shortcut, and they break down fastest across languages. We wrote lists for eight themes in nine languages and read 10 random matches per theme: the ads and chess lists were right 9 times in 10, the price list 4 times in 10. The German word for mistakes is the same as the word for errors in the app, Chinese reviews use the same word for "notify" whether they mean a push message or a missing trial reminder, and a complaint such as an ad dressed up as a system warning uses no word anyone would think to put on a list.
Dimension Labs analysis
Keyword lists in nine languages were right between 4 and 9 times in 10
Share of keyword matches that really were about the theme, from 10 random matches per theme in one- and two-star reviews.
A prompt that defines each code and asks the model to read the whole review handles those cases, because it is reading for meaning rather than matching strings. That is also why the same codebook works in Japanese and in French. For the difference between tagging a review as a whole and tagging each aspect in it, see aspect-based sentiment analysis.
Can one tool handle Chinese and English together, or French for a Quebec team?
Yes, if it keeps the language and market on every record and lets you report each one separately. A Quebec team needs its French reviews, surveys and support conversations in its own view, coded the same way as the rest of Canada so the two can be compared. A team in Taiwan or Hong Kong needs Chinese and English reviews counted together, because the same customers write in both. For Japanese survey comments, accuracy depends on the codebook and the check more than on the language: test on a sample of real comments, including short ones, before you run the full set.
Dimension Labs states it plainly on its homepage: "Every channel normalized and comparable, in every language". Add a language or market field beside the Dimensions and the agent can filter, compare and combine by it. You test a prompt on a sample before it runs on everything, which is where the per-language check goes. Reviews come in from the App Store, Google Play, Google Business Profile and other review sites, next to tickets, chats and surveys.
Why use Dimension Labs for reviews in many languages?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
For a team reading customers in several languages:
Fully omni-channel: reviews, surveys, tickets, chats and calls arrive in one place, "normalized and comparable, in every language", with the language and market kept on each record.
Easy integration across customer channels and operational data: join reviews by country or location to sales, subscriptions or store data, so a complaint in one market can be tied to what it costs there.
Effortless enrichment of every interaction: one set of Dimensions, written for your business and industry, applied to Japanese, French and English text alike and usable by every team. Custom Dimensions start on the Discover plan.
A new visual analytics layer: the agent builds the decks, dashboards, PDF and executive reports for each market team and shares them by a link you can revoke.
Covered California answers enrollment questions across chatbot, live agent and email channels in English and Spanish. With every chatbot session enriched with journey stage, intent, escalation triggers and drop-off location, its team measures friction "across both English and Spanish interactions", and the story lists "Optimize English and Spanish flows independently" among the results. See the Covered California story and our comparison of app review analysis tools.
FAQ
What is the best tool to analyze customer reviews in several languages?
For reviews that need to be read with surveys and support conversations, across languages, and reported to several teams, Dimension Labs: it applies one set of Dimensions across languages and channels and builds the reports. If you only need star-rating alerts and replies for one app store, a dedicated app review tool is simpler.
Can AI theme Japanese survey comments accurately?
Yes, when the codes are defined clearly and you check a sample. In our hand-read sample, Japanese reviews carried complaints that never appeared in English, such as scam-style ads, which is exactly what a translation-first or keyword approach tends to miss. Read 20 to 40 labelled comments before trusting the totals.
Should the report be in English or in each team's language?
Code everything against one codebook so the numbers compare, then write each report in the language of the people who will act on it. Keep quotes in the original with a translation.
How many reviews per language do I need?
Enough for the complaints you care about to show up more than a handful of times. With 40 low-star reviews per language, as in our sample, a complaint in 3 of 40 is worth reading but too thin to report as a rate; a few hundred per language gives stable shares.
Does language detection get short reviews right?
Often not. Reviews like "10/10", a row of emoji or one word are detected as random languages. Keep the storefront or location as a second signal, and treat very short reviews as their own group.
Export a month of reviews or survey comments in every language you receive and start with one file: the agent tags them with one set of Dimensions and reports by language and market.