WG: Bias and Bias Mitigation

Abstract

This discussion focused on definitions and categorization of bias (e.g., social bias, system bias, cognitive bias, and sample bias) and methods to identify and mitigate all in the context of text analysis and visualization. We (Figure 26) discussed the data processing pipelines from the NLP community and the data visualization community as a lens through which to discuss areas where bias can appear. A fundamental point when talking about bias is that biases can be found or introduced in every step of the pipeline. Locating bias in the pipeline can be a challenge. There can be bias in the data, this can be amplified or even introduced by the model, by choices on how to visualize the data, the transformations of the data as part of the visualization and in the eye of the beholder interpreting the results.