Data Caricatures: On the Representation of African American Language in Pretraining Corpora
This highlights a bias in language model training data that could perpetuate stereotypes and exclude AAL speakers, addressing a fairness issue in NLP.
The study evaluated the representation of African American Language (AAL) in 12 English pretraining corpora, finding it underrepresented at 0.007% to 0.18% of documents and that over 25% of AAL texts in C4 may reinforce harmful stereotypes.
With a combination of quantitative experiments, human judgments, and qualitative analyses, we evaluate the quantity and quality of African American Language (AAL) representation in 12 predominantly English, open-source pretraining corpora. We specifically focus on the sources, variation, and naturalness of included AAL texts representing the AAL-speaking community. We find that AAL is underrepresented in all evaluated pretraining corpora compared to US demographics, constituting as few as 0.007% and at most 0.18% of documents. We also find that more than 25% of AAL texts in C4 may be perceived as inappropriate for LLMs to generate and to reinforce harmful stereotypes. Finally, we find that most automated filters are more likely to conserve White Mainstream English (WME) texts over AAL in pretraining corpora.