The problem is linguistic before it is algorithmic

Automated moderation systems do not read every comment in the same way. A comment can mix English with a Romanized Indian language, use slang or abbreviations, substitute emoji for words, and rely on conversational context. Those features make a surface-level decision unreliable.

This is a technical explanation, not a claim that any one platform misses every comment or that SABHYA has a measured accuracy advantage. Public evidence about a platform, an academic benchmark and a product-specific production result are different things.

Code-mixing changes the signal

Hinglish often combines English vocabulary with Hindi grammar, idiom and slang. A classifier trained on monolingual English may recognize individual words while missing the meaning of the sentence as a whole. The same problem appears in other code-mixed conversations: language boundaries are fluid, and the spelling is not standardized.

Romanization removes useful script cues

When Hindi, Bengali or Punjabi is written with Latin characters, a script-based detector may treat it as English or as an unknown variant. Spelling, transliteration and abbreviation choices vary between people. A router can identify a likely processing path, but routing alone does not decide whether a comment is harmful.

Slang is context-dependent

Some words are ordinary in one context and abusive in another. A blocklist can catch known terms, but it cannot reliably distinguish quotation, banter, reclamation and targeted harassment without context. Emoji and character substitutions make brittle keyword systems easier to evade.

What the research supports

Shared-task datasets and multilingual NLP research have treated code-mixed and low-resource language classification as distinct evaluation problems. Results from one dataset or model should not be presented as a universal production guarantee. A responsible system needs route-specific validation, calibrated decision regions, abstention and human review for unresolved cases.

How SABHYA approaches uncertainty

SABHYA uses an explicit language route such as eng, hin, ben, pan or UNCERTAIN. The route is evidence for downstream processing, not a toxicity verdict. The product exposes three moderation decisions: KEEP, HIDE and REVIEW. A review decision means a person needs to decide; it is separate from whether an Instagram action was attempted or succeeded.

Automatic hiding is disabled until a route-specific high-precision acceptance gate and operational controls are in place. Provider failure or unresolved uncertainty must not silently become KEEP.

Limitations

SABHYA does not claim equal performance across native and Romanized scripts, does not claim to outperform Meta, and does not publish production benchmark numbers without a documented methodology. Real-world performance depends on the route, policy version, data quality, context and the account’s configured rollout state.

References

Last reviewed: 18 September 2026.