Flexible text generation for counterfactual fairness probing

Abstract

A common approach for testing fairness issues in text-based classifiers is through the use of counterfactuals: does the classifier output change if a sensitive attribute in the input is changed? Existing counterfactual generation methods typically rely on wordlists or templates, producing simple counterfactuals that don't take into account grammar, context, or subtle sensitive attribute references, and could miss issues that the wordlist creators had not considered. In this paper, we introduce a task for generating counterfactuals that overcomes these shortcomings, and demonstrate how large language models (LLMs) can be leveraged to make progress on this task. We show that this LLM-based method can produce complex counterfactuals that existing methods cannot, comparing the performance of various counterfactual generation methods on the Civil Comments dataset and showing their value in evaluating a toxicity classifier. * Work done as a Google AI Resident.Original: True and the same goes with headscarves . Its not religious requirement but a cultural choice. Simple otherwise there would be no Muslim woman that don't wear them and clearly there are.Counterfactual: True and the same goes with yarmulkes . Its not a religious requirement but a cultural choice. Simple otherwise there would be no Jewish man that don't wear them and clearly there are.

Visualizing Linguistic Diversity of Text Datasets Synthesized by Large Language Models

“…Large language models (LLMs) are becoming ubiquitous for their ability to solve a wide range of linguistic tasks with prompting that does not require additional model training [1,6,22]. This ability also lets them generate smaller, more refined datasets for finetuning [13,25,27], benchmarking [29], low-resource tasks or languages [4,15], and counterfactual testing (e.g., examples that are identical except for having different religious or gender-based identities [12]).…”

AART: AI-Assisted Red-Teaming with Diverse Data Generation for New LLM-powered Applications

“…We address these limitations of human red teaming with a "plug-and-play" AI-assisted Red Teaming (AART) pipeline for generating adversarial testing datasets at scale by minimizing the human effort to only guide the adversarial generation recipe. Our work builds on recent automated red teaming (Perez et al, 2022), synthetic safety data generation (Fryer et al, 2022;Hartvigsen et al, 2022;Bai et al, 2022;Sun et al, 2023) and human-in-theloop methods . We adapt work on self-consistency (Wang et al, 2023a), chain-ofthought (Kojima et al, 2023Wei et al, 2022), and structured reasoning and data generation (Wang et al, 2023b;Xu et al, 2023;Creswell and Shanahan, 2022) and creatively apply them to the task of adversarial dataset creation.…”