Source-linked AI summary
Automatically Neutralizing Subjective Bias in Text
Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, Diyi Yang
TL;DR
The paper addresses persistent inappropriate subjectivity in ostensibly objective text by introducing a neutralization task, a parallel corpus, and two encoder-decoder systems. Large-scale human evaluation suggests the algorithms reduce bias across four domains, while the authors identify single-word editing as a key scope limitation.
Problem
Inappropriate subjectivity remains widespread in texts intended to be neutral, motivating a task for generating more neutral text with otherwise similar meaning.
Method
The paper introduces the 180,000-pair Wiki Neutrality Corpus and two systems: interpretable MODULAR detection-plus-editing and more opaque CONCURRENT generation.
Results
Human evaluation suggests the algorithms identify and reduce bias across encyclopedias, news, books, and political speeches, outperforming style transfer and machine translation baselines.
Takeaways & Limitations
The work provides a first step toward automatically managing subjective bias in real-world expository writing.
Takeaways & Limitations
The experiments are limited to single-word edits, which constitute only a quarter of the edits and are probably among the simplest bias instances.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper frames inappropriate subjectivity as a persistent form of bias and introduces a corpus, task, and sequence-to-sequence algorithms for neutralizing it. Human evaluation suggests the proposed systems can reduce bias across several text domains.
- 62% of Americans believed their news was biased, while bias was reported as the media’s largest source of distrust.
- Subjective bias skews language that should be neutral through framing, presupposition, or opinion.Examples include factive “exposed,” subjective “pilfered,” and gender-presupposing “his.”
- The Wiki Neutrality Corpus contains 180,000 biased and neutralized sentence pairs harvested from Wikipedia edits.It is presented as the first parallel corpus of biased language.
- Neutralizing extends bias detection and classification into generating text with more neutral language and otherwise similar meaning.
- The MODULAR algorithm separates detection from editing, while CONCURRENT identifies subjectivity during generation.MODULAR uses a BERT-based detector and join embedding; CONCURRENT uses a BERT encoder and is more opaque.
- Large-scale human evaluation suggests the algorithms identify and reduce bias in encyclopedias, news, books, and political speeches.The systems also performed better than state-of-the-art style transfer and machine translation systems.
2 Wiki Neutrality Corpus (WNC)
The Wiki Neutrality Corpus (WNC) is built from Wikipedia edits that remove biased language, producing aligned biased–neutralized sentences and contextual data. It captures multiple subjectivity-bias types and supports analyses of where and by whom neutralizing edits occur.
- Corpus construction: 423,823 Wikipedia revisions from 2004–2019 were crawled for edits with NPOV-related justifications.The collection filters out multi-sentence, minimal, maximal, proper-noun-heavy, spelling, reference, hyperlink, and non-literary edits.
- Corpus construction: Sentence pairs were aligned with a size-5 sliding window using pairwise BLEU, then filtered by length-ratio outliers.Pairs beyond the 95th-percentile length-ratio threshold were discarded.
- Corpus composition: WNC contains 180k biased sentences with neutral counterparts plus 385k adjacent neutral sentences unchanged by editors.Neutralizing experiments focus on pairs where an editor modified or deleted a single source word.
- Bias categories: The corpus categorizes subjectivity bias as framing, epistemological, or demographic bias.Framing uses viewpoint-linked wording, epistemological bias alters proposition believability through features such as presupposition, and demographic bias presupposes characteristics of demographic groups.
- Bias categories: A hand-labeled sample of 500 examples estimates the distribution of all three bias types, with framing bias most common.The dataset itself does not include labels for these categories.
- Dataset properties: Neutralizing edits are most prevalent in history, politics, philosophy, sports, and language, and least prevalent in meteorology, science, landforms, broadcasting, and arts.These comparisons use a 44-category article-topic ontology.
- Dataset properties: Newcomers made neutralizing edits in 15% of cases versus 34% for other edits, a significant difference (χ̃2 p = 0.001).Editors were grouped as newcomers with less than one month of experience or experienced editors with more than one month.
3 Methods for Neutralizing Text
The paper presents MODULAR and CONCURRENT sequence-to-sequence methods for neutralizing subjective text while preserving its meaning. MODULAR separates bias detection from editing and uses detector-controlled hidden-state modification, whereas CONCURRENT directly generates neutralized text.
- The proposed task converts subjectively biased input into neutral output with otherwise similar meaning.
- MODULAR: MODULAR separates the process into BERT-based detection and LSTM-based editing, with pretrained modules later fine-tuned jointly.
- MODULAR: The detector estimates each input word’s probability of being subjectively biased using BERT embeddings and discrete expert features.
- MODULAR: The editing module uses a bi-LSTM encoder, attentional LSTM decoder, copy mechanism, and coverage mechanism to generate neutral text.
- MODULAR: The join embedding adds the same learned vector to encoder hidden states, gated by detector probabilities so subjective words receive stronger modification.
- MODULAR: A token-weighted loss emphasizes neutralized words, while decoder pretraining uses neutral text and a denoising autoencoder objective.
- CONCURRENT: CONCURRENT uses a BERT encoder with the same attentional LSTM decoder, copy and coverage mechanisms, and directly generates neutralized text.
4 Experiments
The experiments evaluate neutralization on Wikipedia and three out-of-domain datasets using automatic metrics and human judgments. MODULAR generally reduces bias more, while CONCURRENT better preserves fluency and meaning, with neither automatic metric replacing human evaluation.
- Wikipedia (WNC): All methods reduced bias in human evaluations, but many produced less fluent text.
- Metric Analysis: BLEU correlated more with fluency, while accuracy correlated more with subjective bias reduction.
- Metric Analysis: Neither BLEU nor accuracy was a true substitute for human judgment.
- Out-of-Domain Evaluation: On three out-of-domain datasets, 1,800 raters assessed 200 randomly sampled datapoints after inference.
- Wikipedia (WNC): MODULAR reduced bias better, whereas CONCURRENT preserved meaning and fluency better and achieved higher BLEU.
5 Error Analysis
Error analysis shows that neutralization errors mainly reflect difficult language understanding rather than fluent generation. Models also often delete words instead of finding suitable replacements, although some mismatches still successfully neutralize the source.
- Error Categories: The error analysis distinguishes no change, bad change, disfluency, and noisy target categories.
- Error Distribution: Most errors arose from the subtle language understanding required for bias neutralization, especially with factive and assertive verb replacements.
- Error Distribution: Many outputs that disagreed with Wikipedia’s target edit nevertheless successfully neutralized the source.
- Error Examples: Models tended to remove words rather than identify good replacements.
6 Algorithmic Analysis
The algorithmic analysis examines subjectivity detection, the join embedding, and user control. MODULAR’s detector slightly outperformed comparable BERT-based systems, the join embedding improved scores, and user-specified targets enabled controllable edits but could introduce downstream errors.
- Detecting Subjectivity: MODULAR explicitly detects problematic words before editing, whereas CONCURRENT lacks an interpretable detection window.
- Detecting Subjectivity: Human subjectivity detection was difficult, while MODULAR’s linguistic and category features gave it slight leverage over plain BERT-based models.
- Join Embedding: A gated join embedding combines separately pretrained models and serves as a learned conduit between detection and generation.
- Join Embedding: Removing the join embedding reduced performance from 93.52 BLEU and 46.8 Accuracy to 90.78 BLEU and 37.57 Accuracy.
- Join Embedding Control: User-specified probability distributions let MODULAR target particular words, but targeted control can introduce downstream generation errors.
7 Related Work
Related work spans subjective-bias detection, representation debiasing, and modular text generation. The paper distinguishes neutralizing biased text from removing bias dimensions in representations and connects MODULAR to soft word-level localization with a smooth conduit.
- Subjectivity Bias: Prior work detects subjective or persuasive frames and biased sentences, with this task extending detection into generation.
- Debiasing: Representation-debiasing studies remove prejudice directions or subspaces from embeddings or predictive-model hidden states.
- Neural Language Generation: MODULAR resembles prior methods that localize text style to selected words, but uses soft localization joined by a smooth conduit.
- Neural Language Generation: The paper’s examples illustrate context-sensitive word treatment and user-controlled selection of words to change.
8 Conclusion and Future Work
The proposed models offer useful suggestions for reducing subjective bias in real-world expository writing, but the study is limited to relatively simple single-word edits. Future work should address broader forms of bias and integrate fact-checking.
- The models provide useful suggestions for reducing subjective bias in news, books, and encyclopedias.
- The study’s scope is limited to single-word edits, which constitute a quarter of the edits and are probably among the simplest cases.
- Future work should tackle multi-word, multilingual, and cross-sentence bias.
- Integrating fact-checking could help determine when a presupposition is true and therefore not subjective.
- The join embedding mechanism could also be applied to other modular neural network architectures.