PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction

By Xinyao Ma, Maarten Sap, Hannah Rashkin, Yejin Choi. 2020

82 citations recorded by the source corpus.

Unconscious biases continue to be prevalent in modern text and media, calling for algorithms that can assist writers with bias correction. For example, a female character in a story is often portrayed as passive and powerless (“_She daydreams about being a doctor_”) while a man is portrayed as more proactive and powerful (“_He pursues his dream of being a doctor_”). We formulate **Controllable Debiasing**, a new revision task that aims to rewrite a given text to correct the implicit and potentially undesirable bias in character portrayals. We then introduce PowerTransformer as an approach that debiases text through the lens of connotation frames (Sap et al., 2017), which encode pragmatic knowledge of implied power dynamics with respect to verb predicates. One key challenge of our task is the lack of parallel corpora. To address this challenge, we adopt an unsupervised approach using auxiliary supervision with related tasks such as paraphrasing and self-supervision based on a reconstruction loss, building on pretrained language models. Through comprehensive experiments based on automatic and human evaluations, we demonstrate that our approach outperforms ablations and existing methods from related tasks. Furthermore, we demonstrate the use of PowerTransformer as a step toward mitigating the well-documented gender bias in character portrayal in movie scripts.

Subfields: Artificial Intelligence, Natural Language Processing, Prompting & Reasoning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Prompting & Reasoning

Citations in the corpus

RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models (2020, background): “Future work could also investigate infusing models with more sophisticated or nuanced representations of social biases (Ma et al., 2020).”
Societal Biases in Language Generation: Progress and Challenges (2021, methodology): “…by incorporating a bias control code through conditional training (Dinan et al., 2020a), by ap-pending a target value to inputs during training (Ma et al., 2020), by using a normative classiﬁer to produce reward values for backpropagation (Peng et al., 2020), or through adversarial training…”
Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey (2022, methodology): “rection (Ma et al., 2020) and semi-supervised methods to correct factual errors (Cao et al.”
Generating Training Data with Language Models: Towards Zero-Shot Language Understanding (2022, methodology): “Therefore, our method can be used in company with any bias reduction and correction techniques [15, 37] to mitigate the risks of PLMs.”
Gradient-based Constrained Sampling from Language Models (2022, background): “They also find useful applications in anonymizing protected attributes in written text (Reddy and Knight, 2016), as writing assistants to avoid users’ implicit biases (Ma et al., 2020; Field and Tsvetkov, 2020”
Explaining Toxic Text via Knowledge Enhanced Text Generation (2022, background): “For instance, debiasing methods that use information about toxic language may benefit from additional information given by detailed explanations of toxicity in text (Ma et al., 2020).”
A Survey on Gender Bias in Natural Language Processing (2021, methodology): “In the field of language generation, Ma et al. [2020] introduce controllable debiasing as an unsupervised text revision task that aims to correct the implicit bias against or towards a specific character portrayed in a language model generated text.”
Detect and Perturb: Neutral Rewriting of Biased and Sensitive Text via Gradient-based Decoding (2021, unknown): “Zmigrod et al. (2019) mitigate gender bias by converting between masculine-and feminine-inﬂected sentences with data augmentation; Ma et al. (2020) jointly train a reconstruction and an out-of-domain paraphrasing task to correct bias, which requires a parallel corpus with attribute-sensitive (e.g.…”
Controlled Text Generation as Continuous Optimization with Multiple Constraints (2021, background): “They also have many other positive use-cases, for example, anonymizing personal attributes in written text [54], and even aiding authors in avoiding implicit biases in their writing [44, 14].”
Bias and Fairness in Large Language Models: A Survey (2023, background): “Ma et al. (2020) focus specifically on editing the power dynamics and agency levels encoded in verbs, proposing a neural model that can reconstruct and paraphrase its input, while boosting the use of power- or agency-connoted words.”
Tuning Language Models as Training Data Generators for Augmentation-Enhanced Few-Shot Learning (2022, methodology): “In practice, our method can be combined with any bias reduction and correction strategies (Gehman et al., 2020; Ma et al., 2020) to reduce the adverse effects of PLMs. Limitations.”
A Review of Human Evaluation for Style Transfer (2021, background): “The most common QC strategies are to require workers to pass a qual-iﬁcation test (Jin et al., 2019; Li et al., 2016; Ma et al., 2020; Pryzant et al., 2020) to hire the top-ranked workers based on pre-computed scores that reﬂect the number of their past approved tasks (Kr-ishna et al., 2020; Li et…”
Deep Learning for Text Style Transfer: A Survey (2020, background): “2020), online text debiasing (Pryzant et al. 2020; Ma et al. 2020), simile generation (Chakrabarty, Muresan, and Peng 2020), and many others.”
Mix and Match: Learning-free Controllable Text Generationusing Energy Language Models (2022, methodology): “…by comparing the target agency level with that of the generated text, extracted using the connotation frames lexicon, and following the setup from Ma et al. (2020 For the results presented in Table 2, we ran the Gibbs chain for 8 epochs (8 iterations over all the tokens) for the conventional…”