contrastive-pairs / results.md
Martijn van Beers
Use subheadings
a80ec68
|
raw
history blame
729 Bytes
### Results for French and English language models
BERT is another popular type of language model, and CrowS-Pairs has originally been designed for BERT instead of GPT-2.
Here you can find the results for two languages, English and French, from the authors of [the extended version of CrowS-Pairs](https://aclanthology.org/2022.acl-long.583.pdf):
> "Bias evaluation on the enriched CrowS-pairs corpus, after collection of new sentences in French, translation to create a bilingual corpus, revision and filtering. A score of 50 indicates an absence of bias. **Higher scores indicate stronger preference for biased sentences.**"
![](https://github.com/clclab/contrastive-pairs-demo/raw/main/aggregated_results_crows-pairs.PNG)