How to Create More Ethical Language Models

Total time needed: ~15 minutes
Learning Objectives
Learn about bias and toxicity in language models, create a more ethical language model, or adjust a language model to be more ethical.
Potential Use Cases
You need to create an unbiased, non-toxic language model, or you need to adjust an existing language model to be unbiased and non-toxic.
Target Audience
BEGINNERNLP practitioners creating or using a language model, or people studying AI ethics.
Go through the following annotated items in order:
ARTICLE 1. Bias in Natural Language Processing (NLP): A Dangerous But Fixable Problem
  • What kinds of biases are present in NL and how can they be fixed?
10 minutes
ARTICLE 2. GeDi: A Powerful New Method for Controlling Language Models
  • How can I use the GeDi package to control bias in a language model?
10 minutes
ARTICLE 3. Controlling Text Generation for Language Models
  • How do I use the PPLM approach to control text generation from a language model?
10 minutes
PAPER 4. Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science
  • How do I make people aware of the potential biases in my data?
15 minutes
PAPER 5. Mitigating Gender Bias in Natural Language Processing: Literature Review
  • What are the areas of research on ways to mitigate bias in NLP?
10 minutes
PAPER 6. Towards Controllable Biases in Language Generation
  • What is another area of research for tools to decrease biases in text generation?
10 minutes

Concepts Covered