Multimodal Political Bias Identification and Neutralization

Amun Kharel; Cedric Bernard; Chase Vickery; Xavier Pleimling

REVIEW 5 major objections 6 minor 30 references

Multimodal Political Bias Identification and Neutralization

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims to have built a multimodal pipeline that detects and neutralizes political bias in both news text and images, replacing slanted images with semantically similar, less biased ones.

desk verdict A candid course-project report whose architecture idea is fine but whose central 'succeeded' claim is unsupported by the evidence, mainly because the image-bias labels are outlet-level and circular. read the letter →

arxiv 2506.17372 v1 pith:SPD22QBM submitted 2025-06-20 cs.CY cs.AIcs.CV

Cedric Bernard , Xavier Pleimling , Amun Kharel , Chase Vickery This is my paper

classification cs.CYcs.AIcs.CV

keywords politicalbiasmultimodaldebiasingtextneutralizationimagescoringCLIPsemanticalignmentVisionTransformerBERTnewsretrieval

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that political bias in news is not only a text problem, and proposes a four-stage system that detects and neutralizes biased language in articles, scores images on a left-right bias scale, aligns images and text in a shared CLIP embedding space, and replaces each biased image with a semantically similar but less biased one. The load-bearing ambition is that both modalities can be de-biased together, so that an article can be rewritten and re-illustrated while keeping its meaning. The authors report that the text-bias detector often finds the most biased word, that neutralization produces replacements with an average cosine similarity of 0.3960, and that the image models train acceptably, although they also report that retrieval results were not yet satisfactory and that more training resources are needed. If the paper is right, a deployable pipeline of this kind could reduce the emotional slant of news before it reaches readers.

What carries the argument

The load-bearing object is the continuous bias score in [-1, 1] assigned to each news source and image. It feeds three places: the ViT regression target, the definition of the "bias neighborhood" (images within 10% of the anchor's score) used to form positive pairs in the bias-aware angular loss, and the evaluation metrics that measure retrieved-image bias. The second central object is the angular loss with a centroid term, which the paper adapts from cross-modal semantic alignment to simultaneously group images by topic and separate them by bias.

What would settle it

Give 100 randomly chosen article-image pairs to independent raters of both political leanings, asking each rater to pick the more neutral image from a retrieval pair; if agreement with the model's neutrality ranking is no better than chance, or if replacement images are consistently judged to shift bias in the opposite direction, the claim of neutralization fails. A quantitative version is to recompute Equations 3 and 4 under two independent bias-score annotations and check whether retrieved-image neutrality rankings are stable across the two label sets.

Watch

Extended reading notes

Core claim

The central claim is that political bias identification and neutralization can be done multimodally: a BERT-style detector flags biased tokens, a masked language model replaces them with neutral alternatives, CLIP is fine-tuned with an angular loss that pulls together images of the same topic and, via a new bias-aware version of the loss, keeps images with similar manual bias scores close while pushing apart images with different scores, and a ViT regressor assigns each image a bias score from -1 to 1. At inference, the de-biased text retrieves the nearest semantically similar image whose bias score is closer to zero, thereby replacing slanted images. The paper describes this as succeeding in devising an architecture that can de-bias emotionally charged language, while cautioning that results were limited by training time and resources.

Load-bearing premise

The architecture's idea of "neutral" is inherited from manual bias scores that the authors assigned to each news source, and every learned component—the ViT regressor, the bias-aware alignment pairs, and the retrieval evaluation—depends on those scores; if the scores are subjective or noisy, the retrieved "neutral" images merely mirror the annotators' political judgments.

Editorial extensions

If this is right

A newsroom or browser extension could feed a biased article through the pipeline and receive rewritten text plus a replacement image in one pass, preserving roughly the same subject matter.
Because image debiasing is retrieval-based rather than generative, the replacement is always a real image, so the system is fast and avoids fabricating new visual content.
The text module can be used alone on corpora like Wikipedia edits where paired neutral rewrites exist, making it testable without the image modules.
Error compounding is a direct concern: a wrong bias score on a source propagates through alignment, retrieval, and evaluation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper models bias on a single left-right spectrum, the system has no explicit handle on non-ideological bias or multi-issue ideology; a multi-dimensional score would likely change both training and retrieval behavior.
Since source-level scores are assigned by the authors for sites without public ratings, the entire neutrality target is partly subjective; an independent replication with multiple annotators or crowd-sourced scores would test whether "neutral" is stable across labelers.
The same architecture could transfer to other polarizing domains, such as product reviews or public-health messaging, whenever a bias score can be defined; the paper does not claim this, but its mechanism does not depend on politics specifically.
A testable extension would replace the fixed 10% bias neighborhood with a threshold scaled by annotator variance and measure whether retrieval neutrality improves.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a course project final report, not a research preprint, and the central claim—that the pipeline succeeds at neutralizing political bias in text and images—is not supported by the evidence in the paper. That said, the document is not a waste of time. The architecture is clearly described, the code is linked, and the ethics section is more thoughtful than many published papers.

What's actually new: the composition of four modules (BERT-based biased-word detection, text neutralization, bias-aware CLIP semantic alignment, ViT bias scoring, retrieval-based replacement) is novel in the sense that I haven't seen this exact multimodal pipeline. The bias-aware angular loss in Eq. 2 is a small modification of [27]'s angular loss with a 10% bias-neighborhood sampler. But no evaluation isolates that term, so we don't know if it helps. Text debiasing reuses [20] directly; image scoring uses [26]'s data and labels.

The soft spots are real and load-bearing. Text bias identification has only qualitative examples and no baseline (Section 4.1). Neutralization's only metric (cosine 0.396) is reported as a warning sign: the authors themselves note it's higher than a synonym pair, meaning the model is not producing synonyms. That's an honest report, but it also contradicts the conclusion's 'succeeded.' Image retrieval examples are off-topic; the ViT has no RMSE or R2; no human evaluation was run.

The deeper problem, which the stress-test note nails, is the image-bias label. The scalar bias b is an outlet-level score from -1 to 1, assigned by the authors, and every image from that outlet inherits it. That source-level score is used to train the ViT, to define the bias neighborhoods in the angular loss, and to evaluate retrieval neutrality in Eqs. 3-4. So the metrics measure consistency with the annotators' source scoring, not image-level visual bias. The model could be learning source identity. This circularity is only partially acknowledged in Section 6.

Who is this for? A reader interested in the design space of multimodal debiasing pipelines might skim it for the architecture, and the ethics discussion has useful questions. A serious referee should not spend time on this as a research claim. My recommendation: do not accept for peer review. If the authors want to make it a real paper, they need image-level bias annotations, baseline comparisons for each module, and a human evaluation. The current document is a candid course report.

Referee Report

5 major / 6 minor

Summary. This manuscript describes a four-module pipeline for reducing political bias in news text and images: (i) BERT-based biased-word detection and neutralization (Section 3.1), (ii) CLIP-based cross-modal embedding with a bias-aware angular loss (Section 3.2), (iii) ViT-based image bias scoring (Section 3.3), and (iv) retrieval-based image replacement (Section 3.2). The authors report qualitative token-level bias detection examples (Section 4.1), an average cosine similarity of 0.396 between original and replaced words (Section 4.2), a single qualitative retrieval example (Section 4.3), and unquantified training and validation loss trends for the ViT (Section 4.4). The paper concludes that the attempt succeeded, while also stating that more time and resources were needed and that some modules did not produce the desired results.

Significance. The task of detecting and mitigating political bias in both text and images is timely, and the proposed architecture is a reasonable conceptual framework if the individual modules can be made to work. The paper builds on existing datasets (the Politics Dataset of Thomas and Kovashka and the Wikipedia Neutrality Corpus), provides a GitHub link to the code, and includes a thoughtful ethical discussion in Section 6 that acknowledges subjectivity in source scoring, the single-spectrum assumption, censorship risks, and potential misuse. However, the paper does not provide sufficient evidence to establish any of its main claims. The single quantitative text metric is misinterpreted, the image evaluation is circular, key modules are unvalidated, and the conclusion contradicts the reported results. Consequently, the paper cannot currently be considered a valid scientific contribution.

major comments (5)

[Section 4.2, cosine similarity evaluation] The evaluation of bias word neutralization uses an average cosine similarity of 0.396 between the original and re-predicted words and compares this to 0.240 for the synonyms 'vacation'/'holiday'. The authors interpret the higher value as showing that the model 'is not producing synonymous words in the majority of cases.' This reading is backwards: a higher cosine with the original word means the predicted word is closer to the original than a typical synonym pair is, so the metric suggests the model is producing very similar (often identical) words, not unrelated ones. The metric is also uninformative without baselines such as random replacement, the identity mapping, or a contextual-appropriateness measure, and the qualitative example in the same section ('discharged' replaced by 'graduate') shows that cosine similarity does not capture whether the replacement fits the sentence context. This undermines the text-neutralization claim.
[Section 3 and Equations 3-4] The image-debiasing evaluation is circular. The 'ground truth bias' b(x) in Equations 3 and 4 is the manually assigned source-level score from -1 to 1 per website (Section 3), and the same b(x) is used to define the 10% bias neighborhoods in the angular loss (Equation 2) and to fine-tune the ViT in Section 3.3. Thus Equations 3-4 measure how close retrieved images are to the authors' source-level labels, not whether the images are visually neutral or less biased in any content-based sense. The paper's acknowledgment in Section 6 of possible human bias in scoring does not address this deeper issue: an outlet score is being treated as an image-level label without any image-level annotation. As a result, the reported image retrieval results cannot support the claim that neutral images are being retrieved.
[Section 4.1, bias word identification] Bias word identification is evaluated only with a few qualitative examples and no quantitative baseline. The authors state that 'there was no appropriate baseline to consider,' yet the WNC dataset provides ground-truth biased/neutral sentence pairs, and the system of Pryzant et al. [20] is a natural baseline since the detection module is taken from it. Without precision/recall or a comparison against [20] on a common test set, the claim in Section 5 that the model 'identified the biased text for most of the cases correctly' is unsupported. The statement in Section 4.1 that the model was trained for at most one epoch further weakens evidence of effectiveness.
[Section 4.4, ViT validation] The ViT image bias predictor is not validated. Section 4.4 states that the authors 'were unable to get the R2 and RMSE score' and only reports 'encouraging results for our training and validation loss,' without providing the loss values or curves. Without any quantitative performance measure, the claim in Section 5 that 'we received good loss scores' cannot be assessed, and the module's role in the retrieval pipeline is unconfirmed.
[Section 5, conclusion vs. reported results] The central claim in Section 5 that 'our attempt to de-bias both the text and images of our news succeeded' is contradicted by the paper's own findings: Section 4.3 reports that image retrieval 'did not produce the desired results,' Figure 8 shows retrieved images whose relevance to the query text is unclear, Section 4.2 describes a clear contextual failure ('discharged' → 'graduate'), and Section 4.4 provides no ViT performance numbers. The conclusion should be limited to the proposal of an architecture, not the demonstration of a working system, and the manuscript's current evidence cannot support the success claim.

minor comments (6)

[Throughout] There are numerous typographical and grammatical errors; for example, 'multitmodal' (Section 2.1), 'intial' (Section 4.5), 'we shared a score of 0' (Section 3), and 'B b(xa)' in Equation 2 should be 'B(xa)'.
[Equation 1] The angular loss is written without the hinge/truncation and margin term used in the original angular loss [29]; please clarify the exact formulation and specify the margin alpha value used.
[Sections 3.2 and 4.3] Section 4.3 says 'a pre-trained CLIP model is used and fine-tuned using a modified triplet loss' while Section 3.2 describes the angular loss and the new bias loss; the relationship between these losses and the term 'modified triplet loss' should be clarified.
[Section 4.5] The human evaluation form link in Section 4.5 is empty ('()'); include a working link or remove the claim that the form exists.
[Figures 3 and 4] Figures 3 and 4 are referenced but the manuscript as provided does not include the figure images; please ensure all figures are embedded so that the alignment and retrieval visualizations can be inspected.
[Section 3.1] The 'visual BERT' model fine-tuned with Wikipedia images and text is described without training details, hyperparameters, or results; specify which VisualBERT variant is used and how it is trained.

Circularity Check

2 steps flagged · score 6.0 of 10

Image 'neutrality' is defined, trained, and evaluated with the same manually assigned website-level bias scores; Eqs. 3-4 measure agreement with the model's own labels, not independent neutrality.

fitted input called prediction [Section 3; Section 3.2; Section 3.3; Section 4.3, Eqs. 3-4]
"Based on their political leaning, we labeled the website from a score of -1 to 1 on a continuous scale... For websites we did not have a score, we looked at their wording, sourcing, story choices, and political affiliation to give them a score from -1 to 1. ... We will fine-tune this model further with more images and its respective bias score. ... b is the bias function of an image that returns the ground truth bias for known images in the embedding space and estimated bias for newly scored input images."

Every image inherits the website-level manual bias score assigned by the authors (or taken from mediabiasfactcheck.com). That same score is used as the ViT training label, defines the 10% bias neighborhood in the embedding loss, and is the b(...) in Eqs. 3-4 that measures whether a retrieved replacement image is neutral or moved toward neutrality. The retrieval evaluation therefore is not an independent test of image neutrality; it measures agreement with the annotators' own source-level scale. A model could 'succeed' simply by retrieving images from websites scored near 0, even if the pictures themselves are equally charged.
self definitional [Section 3.2, Eq. 2; Section 4.3, Eq. 4]
"Aside from the angular loss that draws together distinct images of the same topic, the images should also be separated by bias. This process is done with another angular loss objective in which we create positive pairs using an image with similar bias of the anchor and negative pairs using an image with a different bias score. ... Equation 4 measures the average divergence towards neutrality of the retrieved images as compared to the original images in the test set."

The bias-aware angular loss is trained to place images with similar manual scores near each other and images with different manual scores apart. 'Divergence towards neutrality' in Eq. 4 is then computed with those same manual scores. Hence the evaluation criterion is literally the training objective: the metric cannot independently validate the debiasing claim because it is defined by the same labels that built the embedding.

full rationale

The text-debiasing chain is not circular: biased-word identification uses the external detector of Pryzant et al. and the Wikipedia Neutrality Corpus of biased/neutral pairs, and the neutralization evaluation (cosine similarity to the original word) is an external semantic metric, albeit a weak one. The circularity is confined to the image side. The paper assigns one bias score per website, either from mediabiasfactcheck.com or from the authors' own judgment, and every image from that website inherits the score. This score is then used in three roles: as the label for fine-tuning the ViT bias predictor, as the 10% 'bias neighborhood' that defines positive/negative pairs in the bias-aware angular loss, and as the ground-truth b(...) in Eqs. 3-4 that measures whether retrieved replacement images are neutral or moved toward neutrality. Because the evaluation target is identical to the training target, the image-neutralization claim is a self-consistency check rather than an independent demonstration: retrieval can lower |b| simply by returning images from sources the authors scored closer to 0. Section 6 acknowledges possible human bias in the manual scores but does not address the fact that the same scores define the objective, the training neighborhoods, and the metric. The paper's own qualitative retrieval result (Fig. 8) is inconclusive, and no human evaluation was completed. Overall this is partial circularity affecting the central image-side claim.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The system rests on manually assigned scalar bias labels, a hand-chosen neighborhood threshold, and an untested loss design. These are not derived from data or external benchmarks, and the paper provides no attempt to validate them independently of the authors' own annotations.

free parameters (2)

Manual news-source bias scores (-1 to 1) = -1 (far left) to +1 (right), 0 neutral; unlisted websites scored by authors' judgment
Used as ground truth for ViT training, bias-neighborhood sampling in Eq. 2, and retrieval evaluation in Eqs. 3-4. Not externally validated; authors acknowledge subjectivity in Section 6.
Bias-neighborhood threshold = 10% of bias score range
Section 3.2: 'Images within a bias score range of 10% of an anchor image are considered as part of the anchor's bias neighborhood. The range of 10% can be changed.' Hand-chosen because manual scores varied by 5-10%; no ablation.

assumptions (4)

ad hoc to paper Political bias is adequately represented by a single scalar left-right score for each article and image.
Section 3 assigns each source a score from -1 to 1; Section 6 admits this 'takes an America-centric view of political media' and reduces bias to 'a single spectrum.' The whole loss and retrieval design depends on this scalar.
domain assumption Dahlberg's public-sphere qualities are the correct normative standard for news neutrality, and reducing the scalar bias score moves articles toward those qualities.
Section 1 invokes Dahlberg's six qualities as the goal for debiasing, but the operationalized bias score is never validated against that standard.
ad hoc to paper A 10% bias-neighborhood angular loss preserves semantic alignment while separating images by bias in CLIP space.
Section 3.2 introduces Eq. 2. The paper provides no ablation; Figure 8 suggests the bias loss may interfere with semantic alignment ('it is also possible that some aspect of the bias loss interferes with the semantic alignment in a negative way').
domain assumption The pre-trained BERT, CLIP, Doc2Vec, and fastText models transfer to the political-news domain without further validation.
These are taken from [20], [21], [27], and [1] without domain adaptation checks; text results degrade on long sentences and punctuation, and image retrieval examples are off-topic.

invented entities (1)

Bias neighborhood in the shared CLIP embedding space
purpose: Defines positive pairs for the bias-aware angular loss (Eq. 2) so images with similar bias scores are pulled together and different-bias images are pushed apart.
Introduced ad hoc in Section 3.2 with a hand-set 10% radius; no external evidence that such neighborhoods correspond to meaningful political neutrality or that they are stable across topics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multimodal Political Bias Identification and Neutralization." pith.science (2026). https://pith.science/paper/SPD22QBM

@misc{pith2026250617372,
  author       = {Pith},
  title        = {Pith review of: Multimodal Political Bias Identification and Neutralization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPD22QBM}},
  note         = {Machine review of arXiv:2506.17372}
}

read the original abstract

Due to the presence of political echo chambers, it becomes imperative to detect and remove subjective bias and emotionally charged language from both the text and images of political articles. However, prior work has focused on solely the text portion of the bias rather than both the text and image portions. This is a problem because the images are just as powerful of a medium to communicate information as text is. To that end, we present a model that leverages both text and image bias which consists of four different steps. Image Text Alignment focuses on semantically aligning images based on their bias through CLIP models. Image Bias Scoring determines the appropriate bias score of images via a ViT classifier. Text De-Biasing focuses on detecting biased words and phrases and neutralizing them through BERT models. These three steps all culminate to the final step of debiasing, which replaces the text and the image with neutralized or reduced counterparts, which for images is done by comparing the bias scores. The results so far indicate that this approach is promising, with the text debiasing strategy being able to identify many potential biased words and phrases, and the ViT model showcasing effective training. The semantic alignment model also is efficient. However, more time, particularly in training, and resources are needed to obtain better results. A human evaluation portion was also proposed to ensure semantic consistency of the newly generated text and images.

Figures

Figures reproduced from arXiv: 2506.17372 by the authors.

**Figure 1.** illustrates our overall architecture. Images and Texts that are used to train this model will be retrieved from various left-leaning, and right-leaning websites from the Politics Dataset [26]. Based on their political leaning, we labeled the website from a score of -1 to 1 on a continuous scale. We scored -1 for far-left websites and 1 for rightleaning websites. We shared a score of 0 for neutral websites. For exa… view at source ↗

**Figure 2.** Text Bias Neutralization [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Semantic Alignment Model 3.2. Semantically Aligned Images The second primary step of the approach is performing semantic alignment based on the bias score introduced in the description of the overall architecture. We would do this by creating semantic neighborhoods following the process of [27]. Angular loss, an alternative to triplet loss, is used with a CLIP model to develop semantic alignment between images using… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Bias Loss. Text embeddings are represented as cir [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Image Bias Prediction from [20] were generated in conjunction with the neutralization portion of the algorithm, so there was no appropriate baseline to consider. Furthermore, it is difficult to perform quantitative results on our own dataset because we do not have a g…

**Figure 6.** Figure 6: Predicting the bias of BERT tokens with examples from [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Predicting the bias of BERT tokens with examples from [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: This is an example of retrieved images from the test [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages

[27]

Preserving Semantic Neighborhoods for Robust Cross-modal Retrieval

Christopher Thomas and Adriana Kovashka. Preserving Se- mantic Neighborhoods for Robust Cross-modal Retrieval, July 2020. arXiv:2007.08617 [cs]. 2, 4, 6

work page Pith review arXiv 2020
[20]

Automatically Neutralizing Subjective Bias in Text

Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. Automati- cally Neutralizing Subjective Bias in Text, Dec. 2019. arXiv:1911.09709 [cs]. 1, 2, 3, 5, 6

work page Pith review arXiv 2019
[26]

Predicting the Politics of an Image Using Webly Supervised Data

Christopher Thomas and Adriana Kovashka. Predicting the Politics of an Image Using Webly Supervised Data. 1, 2, 3

work page
[1]

English word vectors · fastText. 6

work page
[2]

Jost, Jonathan Nagler, Joshua A

Pablo Barber ´a, John T. Jost, Jonathan Nagler, Joshua A. Tucker, and Richard Bonneau. Tweeting From Left to Right: Is Online Political Communication More Than an Echo Chamber?Psychological Science, 26(10):1531–1542, Oct. 2015. Publisher: SAGE Publications Inc. 1

work page 2015
[3]

Slanted Images: Measuring Nonverbal Media Bias During the 2016 Election, Apr

Levi Boxell. Slanted Images: Measuring Nonverbal Media Bias During the 2016 Election, Apr. 2021. 2

work page 2016
[4]

Detecting Media Bias in News Articles using Gaussian Bias Distributions, Oct

Wei-Fan Chen, Khalid Al-Khatib, Benno Stein, and Henning Wachsmuth. Detecting Media Bias in News Articles using Gaussian Bias Distributions, Oct. 2020. arXiv:2010.10649 [cs]. 2, 3

arXiv 2020
[5]

Analyzing Political Bias and Unfairness in News Articles at Different Levels of Granularity

Wei-Fan Chen, Khalid Al-Khatib, Henning Wachsmuth, and Benno Stein. Analyzing Political Bias and Unfairness in News Articles at Different Levels of Granularity, Oct. 2020. arXiv:2010.10652 [cs]. 2

work page Pith review arXiv 2020

Show all 30 references

[6]

Learning to Model and Ignore Dataset Bias with Mixed Ca- pacity Ensembles, Nov

Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. Learning to Model and Ignore Dataset Bias with Mixed Ca- pacity Ensembles, Nov. 2020. arXiv:2011.03856 [cs]. 2

2020 arXiv
[7]

Echo Chamber or Public Sphere? Predicting Political Orien- tation and Measuring Political Homophily in Twitter Using Big Data.Journal of Communication, 64(2):317–332, Apr

Elanor Colleoni, Alessandro Rozza, and Adam Arvidsson. Echo Chamber or Public Sphere? Predicting Political Orien- tation and Measuring Political Homophily in Twitter Using Big Data.Journal of Communication, 64(2):317–332, Apr
[8]

The Habermasian public sphere: A spec- ification of the idealized conditions of democratic communi- cation.Studies in Social and Political Thought, Jan

Lincoln Dahlberg. The Habermasian public sphere: A spec- ification of the idealized conditions of democratic communi- cation.Studies in Social and Political Thought, Jan. 2004. 1

2004
[9]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, May 2019

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, May 2019. arXiv:1810.04805 [cs]. 3

2019 arXiv
[10]

Multi-Dimensional Gender Bias Classification, May 2020

Emily Dinan, Angela Fan, Ledell Wu, Jason Weston, Douwe Kiela, and Adina Williams. Multi-Dimensional Gender Bias Classification, May 2020. arXiv:2005.00614 [cs]. 2

2020 arXiv
[11]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition ...

2021 arXiv
[12]

A Survey on Bias in Visual Datasets, June 2022

Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A Survey on Bias in Visual Datasets, June 2022. arXiv:2107.07919 [cs]. 2

2022 arXiv
[13]

Detecting Political Bias in News Articles Using Headline Attention

Rama Rohit Reddy Gangula, Suma Reddy Duggenpudi, and Radhika Mamidi. Detecting Political Bias in News Articles Using Headline Attention. InProceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 77–84, Florence, Italy, Aug. 2...

2019
[14]

Quanti- fying Societal Bias Amplification in Image Captioning

Yusuke Hirota, Yuta Nakashima, and Noa Garcia. Quanti- fying Societal Bias Amplification in Image Captioning. In 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 13440–13449, New Or- leans, LA, USA, June 2022. IEEE. 2

2022
[15]

VisualBERT: A Simple and Per- formant Baseline for Vision and Language, Aug

Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A Simple and Per- formant Baseline for Vision and Language, Aug. 2019. arXiv:1908.03557 [cs]. 3

2019 arXiv
[16]

Towards Debiasing Sentence Representations, July 2020

Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, and Louis-Philippe Morency. Towards Debiasing Sentence Representations, July 2020. arXiv:2007.08100 [cs]. 1

2020 arXiv
[17]

Thomas Manzini, Yao Chong Lim, Yulia Tsvetkov, and Alan W. Black. Black is to Criminal as Caucasian is to Po- lice: Detecting and Removing Multiclass Bias in Word Em- beddings, July 2019. arXiv:1904.04047 [cs, stat]. 1

2019 arXiv
[18]

Examining and mitigating gender bias in text emotion detection task

Odbal, Guanhong Zhang, and Sophia Ananiadou. Examining and mitigating gender bias in text emotion detection task. Neurocomputing, 493:422–434, July 2022. 2

2022
[19]

Same Candidates, Different Faces: Uncov- ering Media Bias in Visual Portrayals of Presidential Can- didates with Computer Vision.Journal of Communication, 68(5):920–941, Oct

Yilang Peng. Same Candidates, Different Faces: Uncov- ering Media Bias in Visual Portrayals of Presidential Can- didates with Computer Vision.Journal of Communication, 68(5):920–941, Oct. 2018. 2

2018
[21]

Learning Transferable Visual Models From Natural Language Supervision, Feb

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, Feb. 2021. arXiv:210...

2021 arXiv
[22]

Linguistic Models for Analyzing and Detect- ing Biased Language

Marta Recasens, Cristian Danescu-Niculescu-Mizil, and Dan Jurafsky. Linguistic Models for Analyzing and Detect- ing Biased Language. InProceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650–1659, Sofia, Bulga...

2013
[23]

Robertson.Social Media and Civic Engagement: History, Theory, and Practice

Scott P. Robertson.Social Media and Civic Engagement: History, Theory, and Practice. Synthesis Lectures on Human-Centered Informatics. Springer International Pub- lishing, Cham, 2018. 1

2018
[24]

Determining Sub- jective Bias in Text through Linguistically Informed Trans- former based Multi-Task Network

Manjira Sinha and Tirthankar Dasgupta. Determining Sub- jective Bias in Text through Linguistically Informed Trans- former based Multi-Task Network. InProceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM ’21, pages 3418–3422, New Yor...

2021
[25]

Worst of Both Worlds: Biases Compound in Pre-trained Vision-and-Language Mod- els, May 2022

Tejas Srinivasan and Yonatan Bisk. Worst of Both Worlds: Biases Compound in Pre-trained Vision-and-Language Mod- els, May 2022. arXiv:2104.08666 [cs]. 2

2022 arXiv
[28]

A Multibias-mitigated and Sentiment Knowledge Enriched Transformer for Debiasing in Multimodal Conversational Emotion Recognition, July 2022

Jinglin Wang, Fang Ma, Yazhou Zhang, and Dawei Song. A Multibias-mitigated and Sentiment Knowledge Enriched Transformer for Debiasing in Multimodal Conversational Emotion Recognition, July 2022. arXiv:2207.08104 [cs]. 3

2022 arXiv
[29]

Deep Metric Learning with Angular Loss, Aug

Jian Wang, Feng Zhou, Shilei Wen, Xiao Liu, and Yuanqing Lin. Deep Metric Learning with Angular Loss, Aug. 2017. arXiv:1708.01682 [cs]. 4

2017 arXiv
[30]

De- moting Racial Bias in Hate Speech Detection, May 2020

Mengzhou Xia, Anjalie Field, and Yulia Tsvetkov. De- moting Racial Bias in Hate Speech Detection, May 2020. arXiv:2005.12246 [cs]. 2

2020 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [27]

Preserving Semantic Neighborhoods for Robust Cross-modal Retrieval

Christopher Thomas and Adriana Kovashka. Preserving Se- mantic Neighborhoods for Robust Cross-modal Retrieval, July 2020. arXiv:2007.08617 [cs]. 2, 4, 6

work page Pith review arXiv 2020

[2] [20]

Automatically Neutralizing Subjective Bias in Text

Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. Automati- cally Neutralizing Subjective Bias in Text, Dec. 2019. arXiv:1911.09709 [cs]. 1, 2, 3, 5, 6

work page Pith review arXiv 2019

[3] [26]

Predicting the Politics of an Image Using Webly Supervised Data

Christopher Thomas and Adriana Kovashka. Predicting the Politics of an Image Using Webly Supervised Data. 1, 2, 3

work page

[4] [1]

English word vectors · fastText. 6

work page

[5] [2]

Jost, Jonathan Nagler, Joshua A

Pablo Barber ´a, John T. Jost, Jonathan Nagler, Joshua A. Tucker, and Richard Bonneau. Tweeting From Left to Right: Is Online Political Communication More Than an Echo Chamber?Psychological Science, 26(10):1531–1542, Oct. 2015. Publisher: SAGE Publications Inc. 1

work page 2015

[6] [3]

Slanted Images: Measuring Nonverbal Media Bias During the 2016 Election, Apr

Levi Boxell. Slanted Images: Measuring Nonverbal Media Bias During the 2016 Election, Apr. 2021. 2

work page 2016

[7] [4]

Detecting Media Bias in News Articles using Gaussian Bias Distributions, Oct

Wei-Fan Chen, Khalid Al-Khatib, Benno Stein, and Henning Wachsmuth. Detecting Media Bias in News Articles using Gaussian Bias Distributions, Oct. 2020. arXiv:2010.10649 [cs]. 2, 3

arXiv 2020

[8] [5]

Analyzing Political Bias and Unfairness in News Articles at Different Levels of Granularity

Wei-Fan Chen, Khalid Al-Khatib, Henning Wachsmuth, and Benno Stein. Analyzing Political Bias and Unfairness in News Articles at Different Levels of Granularity, Oct. 2020. arXiv:2010.10652 [cs]. 2

work page Pith review arXiv 2020