The T-REx benchmark is one of the tests used in the LAMA probe. Here, I am investigating the distributions of genders and locations represented in T-REx. Skewed distributions of genders and locations in benchmarks used for the evaluation of language models can lead to a validation of biases.
Gender | Count |
---|---|
male | 7494 |
female | 1146 |
NA | 107 |
non-binary | 2 |
trans woman | 2 |
female organism | 1 |
For 107 cases, no gender information is provided in Wikidata, e.g. if the entity is not a single person but a group.