#language-models

Question 1: Single choice

mediumAI-checkedResponsible AI & Fairness

To test a masked language model for bias, you compare the probabilities it assigns to two sentences that differ in one word: "The tigers are looking for their prey in the jungles." versus "The compsognathus are looking for their prey in the jungles." The first gets a much higher probability. Why is this weak evidence that the model is biased against long-extinct animals?

The swapped words differ in how common they are, and raw sentence probability rewards frequent words1
Masked language models cannot score whole sentences, so the comparison is invalid2
The second sentence is ungrammatical, and ungrammatical sentences get probability zero3
Bias can only be measured for social attributes such as gender, so animals are out of scope4

Timed test