Sentiment analysis in Hindi and other Indian languages

Why it matters

For Indian brands, a large share of reviews include Hindi or another Indian language, often mixed with English. Sentiment analysis in Hindi and regional languages is therefore not an edge case; it is how many customers write.

What makes it hard

• Scripts: Devanagari, other Indian scripts, and Roman-script versions of the same languages.
• Spelling variation in Roman script: the same word spelled many ways.
• Code-mixing within a sentence.
• Negation and intensifiers placed differently from English: "accha nahi tha".
• Idioms and politeness forms that soften complaints.

Approaches

• Word lists per language: cheap but brittle, especially in Roman script.
• Translate then analyse in English: simple, but translation errors flip sentiment and lose nuance.
• Multilingual models that read the original text: usually the most robust for mixed language, when constrained to your categories.

Checking accuracy

• Sample reviews in each major language and script.
• Have native speakers label topic sentiment.
• Compare with the model's labels.
• Look especially at negation and mild complaints.

Examples

An example

A retailer translating Hinglish reviews to English before analysis found many polite complaints, such as "thoda slow tha" ("was a bit slow"), labelled neutral. Classifying the original text with a multilingual model captured them as negative waiting mentions, and waiting complaints at several stores turned out higher than reported.

Common mistakes

• Translating everything first.
• English-only accuracy checks.
• Ignoring Roman-script Hindi.
• Treating polite complaints as neutral.

Common questions

Checklist

• Languages and scripts counted.
• Originals classified.
• Native-speaker checks per language.
• Polite complaints captured.

In short

Indian-language and Hinglish reviews need analysis that reads the original text, handles scripts and code-mixing, and is checked by native speakers for each language.

Build a small glossary

Collect the words customers use in each language for your main topics: billing, exchange, making charges, alteration, delivery, staff. Include common Roman-script spellings. A glossary helps people tag consistently and helps check whether a model is picking up the right topics in each language.

Reporting by language

In the first months, show topic sentiment by language as well as overall. If customers writing in a regional language complain more about something, that may point to stores or customer groups that need different attention, such as signage, explanations or staff who speak the language.

Polite complaints

Many customers soften complaints: "thoda slow tha" ("it was a bit slow"), "theek tha" ("it was okay"). These are easy to label neutral, but in context they often mark a real issue. Include examples like these in your category definitions so they are counted as the mild complaints they are.

Where to start

Count last quarter's reviews by language, pick the two largest non-English groups, check a hundred reviews in each with a native speaker, and fix the biggest gaps first.

Doing this with Falcon

Falcon measures sentiment for every theme a review mentions rather than one score per review, keeps the exact words that support each label, and reports negative share by topic, store and month with small samples flagged and unclear texts kept visible. Every number opens the reviews behind it.