Caliskan et al: Semantics derived automatically from language corpora contain human-like biases

Tags: papers, nlp, bias nlp project

Summary

• The researchers showed that using biases are inherent to the statistical frequency of specific words - distributional hypothesis in linguistics
• They developed a method similar to the Implict Association Test called Word Embedding Association Test, which measures the cosine similarity of specific words.
• Largely a study that shows that all the stereotypes and biases show from the IAT are apart of word embeddings, meaning that bias is not necessarily escapable here by tuning the dataset
• Note that this only captured word embeddings and only captured Glove.
  • Likely, if not certain, that the same biases are in fasttext or w2v
• Paper was from 2017, so no mention of sentence embeddings
• Relates specifically to the work done by Bolukbasi et al who proposed a method to debias word embeddings (T. Bolukbasi, K.-W. Chang, J. Y. Zou, V. Saligrama, A. T. Kalai, Adv. Neural Inf. Process. Syst. 2016, 4349–4357 (2016).)