Overview
Explore the critical topic of bias and fairness in Natural Language Processing (NLP) through this comprehensive lecture from CMU's Advanced NLP course. Delve into various types of bias present in NLP models and learn effective strategies for bias prevention. Examine allocational harm, stereotyping, and biases in human annotation. Discover bias detection techniques, including word embedding association tests and null hypothesis testing. Analyze bias in word and sentence embeddings, error rates, and language disparities across different cities. Investigate counterfactual evaluation methods and explore mitigation strategies such as feature representation, bias sentence embeddings, and data augmentation techniques. Gain valuable insights into the current landscape of bias research in NLP and its implications for developing fair and equitable language models.
Syllabus
Introduction
NLP Systems
Allocational Harm
Stereotyping
Bias in human annotation
Bias detection techniques
Word embedding association test
Null hypothesis
Word embeddings
Sentence embeddings
Error rates
Difference by city
Language disparities
Counterfactual evaluation
Mitigating biases
Feature and variant rep representations
Bias sentence embeddings
Soft devices
Data augmentation
Augmentation with humans
Bias research
Taught by
Graham Neubig