Evaluation of African American Language Bias in Natural Language Generation

Abstract

Warning: This paper contains content and language that may be considered offensive to some readers.While biases disadvantaging African American Language (AAL) have been uncovered in models for tasks such as speech recognition and toxicity detection, there has been little investigation of these biases for language generation models like ChatGPT.We evaluate how well LLMs understand AAL in comparison to White Mainstream English (WME), the encouraged "standard" form of English taught in American classrooms.We measure large language model performance on two tasks: a counterpart generation task, where a model generates AAL given WME and vice versa, as well as a masked span prediction (MSP) task, where models predict a phrase hidden from their input.Using a novel dataset of AAL texts from a variety of regions and contexts, we present evidence of dialectal bias for six pre-trained LLMs through performance gaps on these tasks.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 23 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.