Undesirable biases in NLP: Averting a crisis of measurement (2022, background): “In the related field of algorithmic fairness, some works have made explicit the distinction between construct and operationalizations (e.g., Friedler et al., 2021; Jacobs & Wallach, 2021).”
Five sources of bias in natural language processing (2021, background): “For example, Friedler et al. (2021) defined bias as fairness in algorithms by capturing all the latent features (i.e., demographics) in the data.”
Bias and Fairness in Large Language Models: A Survey (2023, background): “Attention Gaci et al. (2022) L = ∑ S∈S ∑L l=1 ∑H h=1 ∥∥Al,h,S,G :σ,:σ − O :σ,:σ ∥∥22 +λ ∑ S∈S ∑L l=1 ∑H h=1 ∑|G| i=2 ∥∥Al,h,S,G :σ,σ+1 − A :σ,σ+i ∥∥22 Attanasio et al.”
The Fallacy of AI Functionality (2022, background): “However, this intuition is only valid when certain conditions and assumptions are met [67, 124, 198], such as the measurement validity of the data and labels.”
Measuring Bias (2023, background): “Contrary to bias, fairness is a subjective concept that depends on the goal to achieve in a given context, and many different measures have been proposed [10], [24] and perhaps is impossible to achieve all the time given the variety of theories of justice and cultural differences [13].”