Start Date
18-12-2020 9:40 AM
Description
GPT-3 is an incredibly powerful NLP model that has taken the NLP world by storm. However, it possesses the glaring flaw of bias: due to its training datasets being sourced from the internet, they contain biased and prejudiced text which shows through in some of GPT-3’s text output. In order to eliminate this bias in whatever way possible, it is necessary to first understand it and figure out what its biases are in the first place. Therefore, I propose an exhaustive analysis and evaluation of GPT-3’s biases towards a wide variety of people groups in the hopes of creating a reference for other research to use when investigating ways of combating bias in GPT-3.
Click below to download individual papers.
Identifying Bias in Language Modeling Algorithms
GPT-3 is an incredibly powerful NLP model that has taken the NLP world by storm. However, it possesses the glaring flaw of bias: due to its training datasets being sourced from the internet, they contain biased and prejudiced text which shows through in some of GPT-3’s text output. In order to eliminate this bias in whatever way possible, it is necessary to first understand it and figure out what its biases are in the first place. Therefore, I propose an exhaustive analysis and evaluation of GPT-3’s biases towards a wide variety of people groups in the hopes of creating a reference for other research to use when investigating ways of combating bias in GPT-3.
Comments
Mentor: Mandy Korpusik