AugLy: Data Augmentations for Robustness

Zoe, Papakipos,; Bitton, Joanna

doi:10.48550/arxiv.2201.06494

Cited by 13 publications

(16 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Overall, we find that perturbation augmentation can mitigate demographic bias during classification without any serious degradation to task performance for most tasks on the GLUE benchmark (see Zhang et al, 2021;Sen et al, 2021;Papakipos and Bitton, 2022). Data augmentation has been shown to improve out of domain generalization (Ng et al, 2020;Wang et al, 2021a), and some work even use learned methods for doing data augmentation for syntactic alternatives (Ross et al, 2022).…”

Section: Measuring Fairness With the Fairscorementioning

confidence: 72%

“…Is it necesssary to train a perturber, or can we just use heuristics? Previous approaches to perturbing data relied on heuristic methods to generate counterfactual data, such as swapping in words from word lists or designing handcrafted grammars to generate perturbations (Zmigrod et al, 2019;Renduchintala and Williams, 2022;Papakipos and Bitton, 2022). However, heuristic approaches suffer from several weaknesses and training a controlled generation seq2seq model allows us to improve on many of them.…”

Section: Problems Arising From Heuristic Perturbationmentioning

confidence: 99%

See 1 more Smart Citation

Perturbation Augmentation for Fairer NLP

Qian¹,

Ross²,

Fernandes³

et al. 2022

Preprint

View full text Add to dashboard Cite

Unwanted and often harmful social biases are becoming ever more salient in NLP research, affecting both models and datasets. In this work, we ask: does training on demographically perturbed data lead to more fair language models? We collect a large dataset of human annotated text perturbations and train an automatic perturber on it, which we show to outperform heuristic alternatives. We find: (i) Language models (LMs) pre-trained on demographically perturbed corpora are more fair, at least, according to our current best metrics for measuring model fairness, and (ii) LMs finetuned on perturbed GLUE datasets exhibit less demographic bias on downstream tasks. We find that improved fairness does not come at the expense of accuracy. Although our findings appear promising, there are still some limitations, as well as outstanding questions about how best to evaluate the (un)fairness of large language models. We hope that this initial exploration of neural demographic perturbation will help drive more improvement towards fairer NLP.

show abstract

Section: Measuring Fairness With the Fairscorementioning

confidence: 72%

Section: Problems Arising From Heuristic Perturbationmentioning

confidence: 99%

Perturbation Augmentation for Fairer NLP

Qian¹,

Ross²,

Fernandes³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Many projects have proposed particular measurement templates, or prompts for the purpose of measuring bias, usually for large language models, (Rudinger et al, 2018;May et al, 2019;Sheng et al, 2019;Kurita et al, 2019;Webster et al, 2020;Gehman et al, 2020;Huang et al, 2020;Vig et al, 2020;Kirk et al, 2021a;Perez et al, 2022), and some even select existing sentences from text sources and swap demographic terms heuristically (Zhao et al, 2019;Wang et al, 2021;Papakipos and Bitton, 2022). Since one of our main contributions is the participatory assembly of a large set of demographic terms, our terms can be slotted into basically any templates to measure imbalances across demographic groups.…”

Section: Related Workmentioning

confidence: 99%

"I'm sorry to hear that": finding bias in language models with a holistic descriptor dataset

Smith¹,

Hall²,

Kambadur³

et al. 2022

Preprint

View full text Add to dashboard Cite

As language models grow in popularity, their biases across all possible markers of demographic identity should be measured and addressed in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes, and are commonly used with preset bias tests that presuppose which types of biases the models exhibit. In this work, we present a new, more inclusive dataset, HOLISTICBIAS, which consists of nearly 600 descriptor terms across 13 different demographic axes. HOLISTICBIAS was assembled in conversation with experts and community members with lived experience through a participatory process. We use these descriptors combinatorially in a set of bias measurement templates to produce over 450,000 unique sentence prompts, and we use these prompts to explore, identify, and reduce novel forms of bias in several generative models. We demonstrate that our dataset is highly efficacious for measuring previously unmeasurable biases in token likelihoods and generations from language models, as well as in an offensiveness classifier. We will invite additions and amendments to the dataset, and we hope it will help serve as a basis for easy-to-use and more standardized methods for evaluating bias in NLP models.

show abstract

“…It includes a reference set of 1 million images, a development set of 50,000 augmented query images (a subset of which are transformed copies of a reference image), and a training set of 1 million images. About 60% of the query images in DISC21 have been transformed using image augmentations from the AugLy library [Papakipos and Bitton, 2022]. The remaining 40% have been manually edited by humans.…”

Section: The Isc Challengementioning

confidence: 99%

Results and findings of the 2021 Image Similarity Challenge

Zoe¹,

Tolias²,

Jenícek³

et al. 2022

Preprint

View full text Add to dashboard Cite

The 2021 Image Similarity Challenge introduced a dataset to serve as a new benchmark to evaluate recent image copy detection methods. There were 200 participants to the competition. This paper presents a quantitative and qualitative analysis of the top submissions. It appears that the most difficult image transformations involve either severe image crops or hiding into unrelated images, combined with local pixel perturbations. The key algorithmic elements in the winning submissions are: training on strong augmentations, self-supervised learning, score normalization, explicit overlay detection, and global descriptor matching followed by pairwise image comparison 1

show abstract

AugLy: Data Augmentations for Robustness

Cited by 13 publications

References 11 publications

Perturbation Augmentation for Fairer NLP

Perturbation Augmentation for Fairer NLP

"I'm sorry to hear that": finding bias in language models with a holistic descriptor dataset

Results and findings of the 2021 Image Similarity Challenge

Contact Info

Product

Resources

About