Advanced NLP Interview Questions #4 - The WEAT Bias Detection Trap

Why your embeddings look ‘mathematically perfect’ while secretly encoding 100 years of human bias.

You’re in a final round ML Interview at Google DeepMind. The interviewer sets a trap:

“How do you prove your word embeddings aren’t biased before we ship?”

95% of candidates fail immediately by citing the classic textbook example.

They say “I’d check the vector arithmetic. If 𝐊𝐢𝐧𝐠 - 𝐌𝐚𝐧 + 𝐖𝐨𝐦𝐚𝐧 = 𝐐𝐮𝐞𝐞𝐧, the geometry is sound.”

The interviewer sighs. They just proved your model knows dictionary definitions. They failed to prove it’s safe.

Checking definitions ignores 𝐒𝐭𝐞𝐫𝐞𝐨𝐭𝐲𝐩𝐢𝐜𝐚𝐥 𝐏𝐫𝐨𝐣𝐞𝐜𝐭𝐢𝐨𝐧𝐬.

Because your model was trained on 𝘊𝘰𝘮𝘮𝘰𝘯 𝘊𝘳𝘢𝘸𝘭 or 𝘛𝘩𝘦 𝘗𝘪𝘭𝘦, it encodes 100 years of historical bias. While it gets King/Queen right, it silently learns:

𝘋𝘰𝘤𝘵𝘰𝘳 - 𝘔𝘢𝘯 + 𝘞𝘰𝘮𝘢𝘯 = 𝘕𝘶𝘳𝘴𝘦

𝘗𝘳𝘰𝘨𝘳𝘢𝘮𝘮𝘦𝘳 - 𝘔𝘢𝘯 + 𝘞𝘰𝘮𝘢𝘯 = 𝘏𝘰𝘮𝘦𝘮𝘢𝘬𝘦𝘳

You aren’t detecting bias, you are just confirming the model understands gender.

-----

𝐓𝐡𝐞 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧: To pass, you need to implement 𝐓𝐡𝐞 𝐖𝐄𝐀𝐓 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (Word Embedding Association Test).