Mit Mitig igating ing Gen Gender er Bia Bias in in NLP: Li - - PowerPoint PPT Presentation

Mit Mitig igating ing Gen Gender er Bia Bias in in NLP: Li Lite teratur ture Re Review

ACL 2019

Tony Sun · Andrew Gaut · Shirlyn Tang · Yuxin Huang Mai ElSherief · Jieyu Zhao · Diba Mirza · Elizabeth Belding · Kai-Wei Chang · William Yang Wang UC Santa Barbara, UCLA

man woman king queen computer programmer homemaker

Gender Bias Origins

Manifestation in Data and Representations

1

What is Gender Bias?

“Gender bias is the preference or prejudice toward

• ne gender over another”

2

Allocation vs Representation Bias

Text-based Dataset Pre-processed data “He is a nurse. She is a doctor.” User Inpu put Machine Translation

• n Mod
• del

Machine Translation

• n Outpu

put Ő ápolónő. Ő egy orvos “Sh She is is a nu nurse se. . He He is a doc

• ctor
• r.”

Gender Bias in Machine Translation

Propa

• pagation
• n of
• f gender bi

bias

3

Gen Gender er Bias in in NLP LP

Word Embeddings Language Modeling Coreference Resolution Machine Translation Speech Recognition Sentiment Analysis Caption Generation 4

Identifying Bias

Bias Evaluation Methods

Conclusion and Future Directions

Ideas Worth Exploring

The Pipeline

The Big Picture of Bias

Mitigating Bias

Gender Debiasing Models

5

Gender Bias Observation Task Specific Training Set Gender Bias Evaluation Test Set NLP Algorithm Debiasing Gender Observing Gender Bias in NLP Algorithm’s Predictions

Gen Gender er Debiasing Pipel eline

6

Gender Bias Evaluation

Method Categorizations

• 1. Vector Spaces
• 2. Gender Bias Evaluation Test Set (GBETs)

7

Gender Bias in Word Embeddings

Evaluating Bias: Vector Spaces

Bolukbasi et al., 2016 Caliskan et al, 2017

8

Captain Coach Receptionist Nurse Non-Bias Direction Female Bias Male Bias

Manizini et al., 2017 Gonen and Goldberg et al., 2019

Gender Bias Evaluation Test Sets

Evaluating Bias: GBETs

• New data sets for evaluating gender bias
• Traditional data sets lack gender-specific information
• Eliminate confounding variables

He called his mother

Sentences from

• WinoBias. (Zhao et

al., 2018)

9

She called her father

Rudinger et al., 2018 Zhao et al., 2018 Webster et al., 2018 Kiritchenko and Mohammad, 2018

We know there’s bias. Now what?

Mitigating the Bias

• 1. Inference vs Retraining
• 2. Training Data vs Algorithm
• 3. Key Debiasing Methods

10

Categorizing Debiasing Methods

Mitigating Bias: Retraining vs. Inference

Re Retra training In Infe ference

Adjust the model’s predictions at test time Fully retrain the model

11

Categorizing Debiasing Methods

Mitigating Bias: Data vs. Algorithm

Tr Training Data ta Al Algori rithm

Debias the algorithm. Retraining or Inference Debias the training data. Always Retraining

12

Debiasing via Data Augmentation

Mitigating Bias: Retraining + Data Method

Zhao et al., 2018

Original Corpus

• He is a doctor.
• The doctor called his mom.

Genderswapped Corpus

• She is a doctor.
• The doctor called her dad.

Coref efer eren ence e Res esolution Model el

13

Constraining Algorithm Predictions

Mitigating Bias: Inference + Algorithm Method

CRF Model Biased Prediction Constrained Predictions

14

Gender Imbalanced Training Data RBA

Five examples from the imSitu visual semantic role labeling dataset (Zhao et al., 2018)

Our Contributions

• We provide a comprehensive literature review
• f gender bias in NLP
• Critically discuss issues with the purpose of

identifying optimizations, knowledge gaps, and directions for future research

15

Future Directions

Ideas Worth Exploring

• 1. Non-English Languages
• 2. Non-binary Bias
• 3. Interdisciplinary Communication

16

#

Thanks for Listening

ACL 2019

Q & A

tonysun@ucsb.edu ajg@ucsb.edu

Mitigating Gender Bias in NLP: Literature Review