Abstract: "Artificial intelligence and machine learning are currently undergoing rapid growth. Concerns persist regarding their potential to perpetuate biases inherent in human language. This study demonstrates that standard machine learning applied to everyday language reproduces a range of human biases, from implicit associations to societal norms. Using the GloVe word embedding model trained on web text, the research reveals that language itself contains historical biases, whether neutral or problematic. New evaluation methods, WEAT and WEFAT, are introduced. These findings have broad implications for AI, psychology, sociology, and ethics, suggesting that biases may stem from everyday linguistic exposure."
Applicable Models: 
- GloVe (Opensource access)
Authors: Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan
Considerations: Although based in human associations, general societal attitudes do
  not always represent subgroups of people and cultures.
Datasets: .nan
Group: BiasEvals
Hashtags:
- Bias
- Word Association
- Embeddings
- NLP
Link: Semantics derived automatically from language corpora contain human-like biases
Modality: Text
Screenshots:
- Images/WEAT1.png
- Images/WEAT2.png
Suggested Evaluation: Word Embedding Association Test (WEAT)
Level: Model
URL: https://researchportal.bath.ac.uk/en/publications/semantics-derived-automatically-from-language-corpora-necessarily
What it is evaluating: Associations and word embeddings based on Implicit Associations
  Test (IAT)
Metrics: 
- Cosine Similarity
- Effect Size 
Affiliations: Princeton University, University of Bath
Methodology: Effect sizes between two sets of target words (e.g., programmer, engineer, scientist, ... and nurse, teacher, librarian, ...) and two sets of attribute words (e.g., man, male, ... and woman, female ...) are calculated using cosine similarity of the embeddings, with the null hypothesis that an unbaised model would have no difference betwewen the sets.