Detecting stereotypes in the GPT-2 language model using CrowS-Pairs

GPT-2 is a language model which can score how likely it is that some text is a valid English sentence: not only grammaticality, but also the 'meaning' of the sentence is part of this score. CrowS-Pairs is a dataset with pairs of more and less stereotypical examples for different social groups (e.g., gender and nationality stereotypes). We sample 10 random pairs from CrowS-Pairs and show whether the stereotypical example gets a higher score ('is more likely'). If GPT-2 systematically prefers the stereotypical examples, it has probably learnt these stereotypes from the training data.

The colors indicate whether the stereotypical or the less stereotypical examples gets the higher score, the intensity of the color how strong the preference is.