On Defining Erasure Harms for NLP

The paper introduces a structured definition of "erasure" in NLP, addressing the lack of a cohesive framework for identifying and measuring this harm. It critiques existing conceptualizations for being either overly broad or context-specific, and provides a clear articulation of the necessary components for practitioners to operationalize and measure erasure effectively. This work is significant for AI engineers as it lays groundwork for assessing representational harms in NLP systems, enabling more responsible deployment of these technologies.