Stereotype and Categorical Bias Evaluation via Differential Cosine Bias Measure

A vast range of Natural Language Processing (NLP) systems that are in use today have direct impact on humans. While machine learning models are expected to automatically infer world knowledge from historical texts, we should also be cognizant not to let NLP applications consume undesired societal stereotype bias back. Many bias evaluation measures have been designed and experimented to check whether unwanted stereo-type bias is present in the model or not. Upon performing various experiments, we found out that the most popular bias measures do not always indicate bias accurately. In addition to these experimental findings, we also propose our novel Differential Cosine Bias measure with examples of unwanted stereotype biases as well as necessary categorical bias that is based on knowledge. Our experiments show that our bias measure is a potential indicator of bias in NLP models compared to the popular bias evaluation measures.