Researchers at Twitter announced the discovery of bias in named entities that form the linguistic foundation for a wide range of online properties, from search engines to knowledgebases. While most bias and fairness research focuses on a single variable like gender or race, this project explored the intersection of race and gender in the named entity recognition models. The bias derives, at least in part, from bias in the training data, an issue that compounds itself, as underrepresented classes of individuals are excluded now, which in turn excludes them from future training data sets.
Despite this insightful work, Twitter still struggles with the problem of bias itself, as revealed this week alongside claims of bias in Zoom's background isolation algorithm.