#20daysofresponsibleAI - Twitter Hashtag

over 2 years ago

How Gender Debiasing Affects Internal Model Representations, and Why It Matters (NAACL 2022) Paper: https://t.co/8aBS6IJTXf TLDR: Intrinsic bias metrics can correlate well with extrinsic metrics, with some changes, and can be more reliable Day 20 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. How Gender Debiasing Affects Internal Model Representations, and Why It Matters (NAACL 2022)

Paper: https://t.co/8aBS6IJTXf

TLDR: Intrinsic bias metrics can correlate well with extrinsic metrics, with some changes, and can be more reliable

Day 20 of #20daysofresponsibleAI #NLP https://t.co/bOS5ABW29v

1

8

0

1

659

Vipul Gupta

@vipul_1011

over 2 years ago

Debiasing should be Good and Bad (ACL 2023) Paper: https://t.co/erPvi2lqR6 TLDR: Proposes that if bias specifications are reversed, output of debiasing method should also get reversed & find current methods to be inconsistent with this Day 19 of #20daysofresponsibleAI #NLProc

vipul_1011's tweet photo. Debiasing should be Good and Bad (ACL 2023)

Paper: https://t.co/erPvi2lqR6

TLDR: Proposes that if bias specifications are reversed, output of debiasing method should also get reversed & find current methods to be inconsistent with this

Day 19 of #20daysofresponsibleAI #NLProc https://t.co/DTATrfjSLd

1

14

3

8

3K

Vipul Gupta

@vipul_1011

over 2 years ago

When Do Pre-Training Biases Propagate to Downstream Tasks? (EACL 2023) Paper: https://t.co/q3rHBmeo3E TLDR: Presence of high biases in pre-trained LMs can lead to high hallucination rates in downstream tasks! Day 18 of #20daysofresponsibleAI #NLProc

vipul_1011's tweet photo. When Do Pre-Training Biases Propagate to Downstream Tasks? (EACL 2023)

Paper: https://t.co/q3rHBmeo3E

TLDR: Presence of high biases in pre-trained LMs can lead to high hallucination rates in downstream tasks!

Day 18 of #20daysofresponsibleAI #NLProc https://t.co/2bA2RWqSVx

1

19

3

12

3K

Vipul Gupta

@vipul_1011

over 2 years ago

Debiasing Masks: A New Framework for Shortcut Mitigation in NLU Paper: https://t.co/084PECg9Ik TLDR: Finds a subset of weights in the model which causes bias and prunes it to zero during inference. Requires no retraining! Day 15 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. Debiasing Masks: A New Framework for Shortcut Mitigation in NLU

Paper: https://t.co/084PECg9Ik

TLDR: Finds a subset of weights in the model which causes bias and prunes it to zero during inference. Requires no retraining!

Day 15 of #20daysofresponsibleAI #NLP https://t.co/EPvOTypieJ

1

2

0

358

Vipul Gupta

@vipul_1011

over 2 years ago

The Tail Wagging the Dog (ACL 2023 Outstanding Paper) Paper: https://t.co/TvF55jnqgC TLDR: Social bias benchmarks are very sensitive to small modifications e.g. synonymizing words in templates, resulting in very different bias values Day 14 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. The Tail Wagging the Dog (ACL 2023 Outstanding Paper)

Paper: https://t.co/TvF55jnqgC

TLDR: Social bias benchmarks are very sensitive to small modifications e.g. synonymizing words in templates, resulting in very different bias values

Day 14 of #20daysofresponsibleAI #NLP https://t.co/ApBrAMFGpr

1

2

1

429

Vipul Gupta

@vipul_1011

over 2 years ago

“How Biased are Your Features?” (FAccT 2023) Paper: https://t.co/NEsKe1n6uf TLDR: They measure the contribution of individual features to the overall bias of a classifier by fitting a set-additive model to the classifier Day 13 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. “How Biased are Your Features?” (FAccT 2023)

Paper: https://t.co/NEsKe1n6uf

TLDR: They measure the contribution of individual features to the overall bias of a classifier by fitting a set-additive model to the classifier

Day 13 of #20daysofresponsibleAI #NLP https://t.co/KbnjY1pqRQ

1

10

2

1

2K

Vipul Gupta

@vipul_1011

over 2 years ago

Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model Paper: https://t.co/pQm6io2W0F TLDR: Biased machine translation models create gender-biased text from unbiased text, later used to train unbiased seq-to-seq model Day 12 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model

Paper: https://t.co/pQm6io2W0F

TLDR: Biased machine translation models create gender-biased text from unbiased text, later used to train unbiased seq-to-seq model

Day 12 of #20daysofresponsibleAI #NLP https://t.co/UiucDut4XE

1

5

1

0

1K

Vipul Gupta

@vipul_1011

over 2 years ago

FairPrism: Evaluating Fairness-Related Harms in Text Generation (ACL 2023) Paper: https://t.co/G5NF6GuXbr TLDR: A new dataset for different types of harm. Addresses issues such as clearer coverage specification, annotators disagreement Day 11 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. FairPrism: Evaluating Fairness-Related Harms in Text Generation (ACL 2023)

Paper: https://t.co/G5NF6GuXbr

TLDR: A new dataset for different types of harm. Addresses issues such as clearer coverage specification, annotators disagreement

Day 11 of #20daysofresponsibleAI #NLP https://t.co/lsmUjB3yft

1

5

1

1K

Vipul Gupta

@vipul_1011

over 2 years ago

BERTScore is Unfair: On Social Bias in Language Model-Based Metrics for Text Generation (EMNLP 2022) Paper: https://t.co/DKRcyOzCRX TLDR: Pre-trained LM based metrics have higher bias on 6 sensitive attributes including race and gender Day 10 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. BERTScore is Unfair: On Social Bias in Language Model-Based Metrics for Text Generation (EMNLP 2022)

Paper: https://t.co/DKRcyOzCRX

TLDR: Pre-trained LM based metrics have higher bias on 6 sensitive attributes including race and gender

Day 10 of #20daysofresponsibleAI #NLP https://t.co/Qd1KMyir2F

1

42

8

20

6K

Vipul Gupta

@vipul_1011

over 2 years ago

NLPositionality: Characterizing Design Biases of Datasets and Models (ACL 2023 Outstanding Paper) Paper: https://t.co/aMZud0oq74 TLDR: Authors impart their own identities, backgrounds into datasets and models which leads to design biases Day 9 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. NLPositionality: Characterizing Design Biases of Datasets and Models (ACL 2023 Outstanding Paper)

Paper: https://t.co/aMZud0oq74

TLDR: Authors impart their own identities, backgrounds into datasets and models which leads to design biases

Day 9 of #20daysofresponsibleAI #NLP https://t.co/Sxu4lPPT5b

1

19

3

6

3K

Vipul Gupta

@vipul_1011

over 2 years ago

Tracking the Trails of Political Biases Leading to Unfair NLP Models (ACL 2023 Best Paper) Paper: https://t.co/VjrMg7fcy3 TLDR: Using political spectrum theories they show political bias propagates from train data to LMs to downstream tasks Day 8 of #20daysofresponsibleAI #NLP

vipul_1011's tweet photo. Tracking the Trails of Political Biases Leading to Unfair NLP Models (ACL 2023 Best Paper)

Paper: https://t.co/VjrMg7fcy3

TLDR: Using political spectrum theories they show political bias propagates from train data to LMs to downstream tasks

Day 8 of #20daysofresponsibleAI #NLP https://t.co/siaiVmZJYF

1

8

0

6

2K

Vipul Gupta

@vipul_1011

over 2 years ago

Can Voice Assistants Be Microaggressors? Paper: https://t.co/nxcXcV4TNK TLDR: They found that high errors/biases in speech recognition harm marginalized groups more psychologically Thanks @kimiwenzel for the suggestion and great work! Day 7 of #20daysofresponsibleAI #NLProc

vipul_1011's tweet photo. Can Voice Assistants Be Microaggressors?

Paper: https://t.co/nxcXcV4TNK

TLDR: They found that high errors/biases in speech recognition harm marginalized groups more psychologically

Thanks @kimiwenzel for the suggestion and great work!

Day 7 of #20daysofresponsibleAI #NLProc https://t.co/MOAOIXpezI

1

4

1

0

748

Vipul Gupta

@vipul_1011

over 2 years ago

Stereotyping Norwegian Salmon Paper: https://t.co/6dXlGvh54D TLDR: Identified pitfalls in bias benchmark datasets about what they aim to measure vs how well they measure it 🧵⬇️ Thanks @shaily99 for the suggestion! Day 6 of #20daysofresponsibleAI #NLProc #bias

vipul_1011's tweet photo. Stereotyping Norwegian Salmon

Paper: https://t.co/6dXlGvh54D

TLDR: Identified pitfalls in bias benchmark datasets about what they aim to measure vs how well they measure it 🧵⬇️

Thanks @shaily99 for the suggestion!

Day 6 of #20daysofresponsibleAI #NLProc #bias https://t.co/5mJ3YDDqbU

1

7

1

0

2K

Vipul Gupta

@vipul_1011

over 2 years ago

Using Natural Sentences for Understanding Biases in Language Models Paper: https://t.co/7mx76QU0GQ TLDR: Quantifying bias is very sensitive to the design choice by authors. Real-world (Natural) sentences should be used instead 🧵⬇️ Day 4 of #20daysofresponsibleAI #NLP #bias

2

0

1

672

Top Tweets for #20daysofresponsibleAI

Last Seen Hashtags on Sotwe

Trends for you

Most Popular Users