Gender Neutrality
Runs on the assistant's replies
Stops the assistant from guessing someone's gender, and changes those words to ones that work for any gender.
🔗 View the full list of replacement words
What happens​
↻ Fixes & passes on — The reply is sent as normal, with the words changed. The user never sees the original wording.
Example​
| Assistant would say | Aapka beta bada hoke doctor ya engineer ban sakta hai. |
| Assistant actually sends | Aapka bachcha bada hoke doctor ya engineer ban sakta hai. |
Only the gender word changes. The advice stays the same.
Works only when the reply is in English, Hindi or Hinglish. Replies in other languages are sent unchecked.
Without this guardrail​
The assistant guesses a gender — like calling a child "beta" (son) — even when the user never said one. Risk: some users feel left out, and your programme feels less welcoming.
How well it works​
| Accuracy | Response time |
|---|---|
| 81% | 0.28 ms |
Tested on 600 real-style messages across English, Hindi and Hinglish.
Known limitations​
- Swapping a word can make the sentence sound a little clunky.
e.g. changing "beta" to "bachcha" mid-sentence can read slightly awkwardly. - In Hindi, gender hides inside action words too, not just in words like "he" or "she."
e.g. "woh gaya" (he went) vs "woh gayi" (she went) — swapping one word doesn't fix the verb ending. - Sometimes there's no neutral word to use, so it falls back to the boy/man version.
e.g. there's no common neutral form for "beta," so it becomes "bachcha" (child) instead. - It only knows the words on its list — anything not on that list slips through.
e.g. a new spelling or slang term for a gendered word won't be caught if it isn't already in the list. - It can't always tell a real question about gender apart from the assistant assuming one.
e.g. "scholarships for girls" is a genuine, valid question — the check must be careful not to "fix" something that wasn't wrong.
Technical details
| Method | Accuracy | Precision | Recall | F1 | Sample size |
|---|---|---|---|---|---|
| Rule-based | 80% | 76% | 87% | 81% | 600 |
Dataset: Evaluation dataset (600 samples) ↗
Evaluated on 600 real-style messages across English, Hindi and Hinglish. Precision (76%) means about 1 in 4 flags is a false alarm; recall (87%) means most real gender assumptions are caught, though some still slip through. F1 (81%) balances the two.