What It Takes to Control Societal Bias in Natural Language Processing
summary
Examines sources of societal bias in natural language processing models, focusing on how data, model architectures, and learning objectives contribute to biased behavior. Describes approaches for measuring and mitigating bias in NLP systems, including dataset construction, representation learning, and algorithmic techniques to control or reduce biased outcomes.
AI-written summary · about
chapters (25)
- 0:00Intro
- 1:08What It Takes to Control Societal Bias in Natural Language Processing
- 2:54NLP Models are Brittle
- 3:26NLP Models is lack of commonsense
- 3:42NLP Models Exhibit Societal Bias
- 4:06A carton of ML (NLP) pipeline
- 5:38Representational Harm in NLP: Word Embeddings can be Sexist
- 8:32Bias in Multi-lingual Embedding
- 9:32Contextualized Representation
- 10:34Unequal Treatment of Gender
- 15:51Performance Gap
- 18:45Bias in Visual-and-Language Models
- 20:54im Situ Visual Semantic Role Labeling (VSRL)
- 22:12Algorithmic Bias in Applications
- 22:46Defining Dataset Bias (events) Training Gender Radio
- 24:47Leakage of Gender
- 25:50What It Takes to Control Societal Bias in NLP?
- 26:29Outline
- 26:46Approach 1: Surgery representation space
- 28:22Approach 1: Make Gender Information Transparent in Word Embedding
- 31:11How to deal with bias in data
- 32:03Gender bias in Coref System
- 32:54Reducing Bias Amplification (RBA)
- 36:08Lagrangian Relaxation
- 37:47Conclusions
related talks
31mApproaches to Fairness and Bias Mitigation in Natural Language Processing
42mTowards Inclusive AI: Understanding bias in ML models
38mSocial Implications of Bias in Machine Learning
42mUnderstanding the nature of Bias in Natural & AI Systems
37mDigital Discrimination: Cognitive Bias in Machine Learning
23mDeveloping an intersectional framework to analyze biases in artificial intelligence and deep neural
47mFight AI Bias With… Bias
38mSocial Implications of Bias in Machine Learning