Analysing reviews in other languages: Indian languages and Hinglish

What the data looks like

Reviews in other languages are a large part of review data for Indian brands. A single store's reviews can include English, Hindi in Devanagari, Hindi in Roman script, Marathi or Tamil, and sentences that switch between English and a local language mid-way. Any analysis that only reads English is reading part of its customers.

The challenges

• Code-mixing: "Staff bahut helpful tha but billing slow".
• Script variation: the same Hindi word spelled many ways in Roman script.
• Idioms and sarcasm that do not translate literally.
• Domain words: "making charges", "hallmark", "exchange" in local phrasing.

A practical approach

• Record the language and script of each review.
• Classify in the original language with a multilingual model where possible.
• Keep the original text and a translation for readers who need it.
• Check a sample in each language monthly with a native speaker.
• Report results by language as well as overall, at least at first.

Why not translate everything first?

Translating to English and then analysing is simple but lossy. Sarcasm, emphasis and mixed sentiment often change in translation, and machine translation of Roman-script Hindi is uneven. Classifying the original text avoids a whole layer of error.

An example

A retailer's English-only analysis showed staff complaints falling. Including Hindi and Hinglish reviews, which were about a third of the total at some stores, showed staff complaints flat at those stores. The difference was a group of customers whose complaints had simply not been counted.

Replying in their language

Common mistakes

• Dropping non-English reviews.
• Treating Roman-script Hindi as gibberish.
• Trusting translation for sentiment.
• Checking accuracy only in English.

Common questions

Checklist

• Language recorded per review.
• Originals classified.
• Translations kept for reading.
• Monthly native-speaker checks.
• Results reported by language.

In short

Indian review data is multilingual and code-mixed. Classify the original text with a multilingual model, keep translations for reading, check accuracy by language, and never drop non-English reviews.

Practical setup

Count last quarter's reviews by language and script. For languages above a few percent of volume, make sure your classification handles them and that someone can check a monthly sample. For rare languages, read them directly. Keep a short glossary of domain terms customers use in each language, such as local words for exchange, making charges or alteration, to help both people and models classify consistently.

Reporting by language

In the first months, report key topics by language as well as overall. If complaints in regional-language reviews differ from English ones, that is a finding in itself, often pointing to stores or customer groups that need different attention.

Using local teams

Store teams in each region read and write their local language natively. Involve them in checking samples and refining the glossary; they also know the local terms customers use for products and services, which improves classification for everyone.

Doing this with Falcon

Falcon builds the theme list from your own reviews, then labels every review against it with a multilingual model, scoring sentiment for each mention and keeping the exact quote behind each label. Texts too short to judge are marked unclear rather than guessed, and low-confidence labels are flagged for a person to check. Negative share by topic, location and month then shows what is rising where, with small samples flagged.