Automatically Neutralizing Subjective Bias in Text

By Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, Diyi Yang. 2019

207 citations recorded by the source corpus.

Texts like news, encyclopedias, and some social media strive for objectivity. Yet bias in the form of inappropriate subjectivity — introducing attitudes via framing, presupposing truth, and casting doubt — remains ubiquitous. This kind of bias erodes our collective trust and fuels social conflict. To address this issue, we introduce a novel testbed for natural language generation: automatically bringing inappropriately subjective text into a neutral point of view (“neutralizing” biased text). We also offer the first parallel corpus of biased language. The corpus contains 180,000 sentence pairs and originates from Wikipedia edits that removed various framings, presuppositions, and attitudes from biased sentences. Last, we propose two strong encoder-decoder baselines for the task. A straightforward yet opaque concurrent system uses a BERT encoder to identify subjective words as part of the generation process. An interpretable and controllable modular algorithm separates these steps, using (1) a BERT-based classifier to identify problematic words and (2) a novel join embedding through which the classifier can edit the hidden states of the encoder. Large-scale human evaluation across four domains (encyclopedias, news headlines, books, and political speeches) suggests that these algorithms are a first step towards the automatic identification and reduction of bias.

Subfields: Artificial Intelligence, Natural Language Processing, Safety, Robustness & Interpretability

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 Safety, Robustness & Interpretability

Citations in the corpus

Language (Technology) is Power: A Critical Survey of “Bias” in NLP (2020, neutral): “…(2019); Prates et al. (2019); Prost et al. (2019); Qian et al. (2019); Swinger et al. (2019); Zhao et al. (2019); Zhou et al. (2019); Ethayarajh (2020); Huang et al. (2020); Jia et al. (2020); Popović et al. (2020); Pryzant et al. (2020); Vig et al. (2020); Wang et al. (2020); Zhao et al. (2020).”
Societal Biases in Language Generation: Progress and Challenges (2021, methodology): “Other works present parallel corpora of biased versus unbiased framings and presuppositions (Pryzant et al., 2020) and AAE versus WAE equivalents (Groen-wold et al., 2020).”
Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey (2022, background): “The first set of works treats the task of rewriting as a sequence labeling task, where each token in the output sequence is either substituted, deleted, or kept the same (Pryzant et al., 2020; He et al., 2021b).”
PEER: A Collaborative Language Model (2022, background): “• WNC (Pryzant et al., 2020) is a dataset where the task is to remove or mitigate biased words to make sentences more neutral;”
Towards Controllable Biases in Language Generation (2020, methodology): “Furthermore, Bordia and Bowman (2019) introduce a regularization loss term when training a language model, and Pryzant et al. (2020) propose encoder-decoder systems for neutralizing subjective biases in text.”
PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction (2020, methodology): “Controllable Debiasing is a new formalization of the unsupervised stylistic rewriting task, contrasting with supervised approaches which beneﬁt from parallel corpora (e.g., Xu et al., 2012, 2015; Rao and Tetreault, 2018; Pryzant et al., 2020).”
Investigating Societal Biases in a Poetry Composition System (2020, background): “, 2018; Kiritchenko and Mohammad, 2018), and language modeling (Sheng et al., 2019; Bordia and Bowman, 2019; Pryzant et al., 2020; Huang et al., 2019).”
Civil Rephrases Of Toxic Texts With Self-Supervised Transformers (2021, background): “When para-phrase generation aims at being explicitly attribute-invariant, it is referred as obfuscation or neutralization (Emmery et al., 2018; Xu et al., 2019b; Pryzant et al., 2020).”
Deep Learning for Text Style Transfer: A Survey (2020, background): “2020), biasedness (Pryzant et al. 2020), medical text simplification (Cao et al.”
Explaining Toxic Text via Knowledge Enhanced Text Generation (2022, methodology): “We incorporate expert knowledge into the generation process using the join embedding technique (Pryzant et al., 2020) along with toxicity classification models.”
Cross-Loss Influence Functions to Explain Deep Network Representations (2020, methodology): “We leverage a larger dataset of Wikipedia edits from Pryzant et al. (Pryzant et al., 2019) containing 180,000 sentences pairs, each consisting of a point-of-view biased sentence and its neutralized counterpart.”
A Framework for the Computational Linguistic Analysis of Dehumanization (2020, background): “, 2018), and adapting natural language generation methods to “neutralize” text (Pryzant et al., 2019).”
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…”
Wordcraft: Story Writing With Large Language Models (2022, background): “Such applications are often powered by models trained for a particular purpose, such as autocompletion [40], or neutralizing bias [35].”
Towards Facilitating Empathic Conversations in Online Mental Health Support: A Reinforcement Learning Approach (2021, methodology): “Finally, supervised methods commonly used for similar tasks such as style transfer [29, 55] and content debiasing [37, 46] usually require a large parallel dataset.”
Detect and Perturb: Neutral Rewriting of Biased and Sensitive Text via Gradient-based Decoding (2021, unknown): “, 2018); training-time methods (Huang et al., 2020; Liu et al., 2020a,b; Kaneko and Bollegala, 2021; Pryzant et al., 2020); and inference-time methods.”
A Transformer-based Framework for Neutralizing and Reversing the Political Polarity of News Articles (2021, methodology): “From the results, shown in Table 3, it can seen that for headlines, our method outperforms the prior work for the headlines [20, 81] by at least 10% for all flipping directions.”