pith. sign in

arxiv: 1906.10861 · v2 · pith:5G6QNVHMnew · submitted 2019-06-26 · 💻 cs.SI · cs.CR· cs.LG

Assessing Post Deletion in Sina Weibo: Multi-modal Classification of Hot Topics

Meisam Navaki Arefi , Rajkumar Pandi , Michael Carl Tschantz , Jedidiah R. Crandall , King-wa Fu , Dahlia Qiu Shi , Miao Sha This is my paper

Pith reviewed 2026-05-25 15:21 UTC · model grok-4.3

classification 💻 cs.SI cs.CRcs.LG
keywords Weibo censorshippost deletionsentiment analysismulti-modal classificationsocial media monitoringChinese internettopic categorizationcontent moderation
0
0 comments X

The pith

Sentiment is the only consistent predictor of which Weibo posts are deleted across 14 topic categories.

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

The paper collects censored and uncensored Weibo posts and classifies them into 14 categories using both text and image content. It then measures how topic, sentiment, and other factors relate to deletion decisions. The central result is that sentiment alone tracks censorship rates reliably no matter which category a post falls into. Other signals, such as whether a post concerns protests or politicians, vary in strength by topic. The work also reports that deleted posts are removed after roughly three hours on average.

Core claim

A dataset of 994 text posts and 18,966 images spanning 14 categories with censorship potential is assembled and processed with deep learning, CNN localization, and NLP methods. The resulting analysis identifies sentiment as the sole factor whose association with deletion holds steady across every category examined. Categories tied to anti-government actions or named politicians show high deletion rates, while crisis-related categories show lower rates. Across all categories, censored posts are removed after an average of three hours, a pattern consistent with leaked Sina Weibo logs.

What carries the argument

Multi-modal classification of posts into 14 topic categories, followed by comparison of deletion rates against sentiment and other attributes to isolate consistent predictors.

If this is right

  • Censorship on Weibo tracks negative sentiment uniformly even when the underlying topics differ.
  • Posts about protests or politicians are deleted at higher rates than posts about crises such as rainstorms.
  • Deleted posts across categories are removed after an average of three hours.
  • Leaked internal logs align with the observed deletion patterns derived from external data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Platforms could build early-warning systems that flag high-sentiment posts before deletion occurs.
  • Extending the same method to other languages or platforms would test whether sentiment remains the dominant signal outside Weibo.
  • If labeling bias exists in the 14 categories, new topics that emerge could change which factors appear consistent.

Load-bearing premise

The 14 categories and the labeling of the 994 posts plus 18,966 images capture the full range of potentially censored content without selection or labeling errors that would alter which factors appear consistent.

What would settle it

A replication that applies the same multi-modal labeling and analysis to a fresh collection of Weibo posts and finds that another attribute, such as image content type or specific keyword presence, predicts deletion more consistently than sentiment across the same categories.

Figures

Figures reproduced from arXiv: 1906.10861 by Dahlia Qiu Shi, Jedidiah R. Crandall, King-wa Fu, Meisam Navaki Arefi, Miao Sha, Michael Carl Tschantz, Rajkumar Pandi.

Figure 1
Figure 1. Figure 1: An example of image augmentation. such comprehensive list of topics that China censors. However, we have tried to pick general categories so that they can be applied for analyzing any other Chinese platforms that practice censorship. Training Dataset: To assemble a training dataset, we utilized Google Image Search to find images of 200 × 200 pixels or bigger per category. As has been done by other studies … view at source ↗
Figure 2
Figure 2. Figure 2: Confusion matrix of image classifier. To reduce the incidence of classifying images that belong to none of our categories as belonging to the most similar category, we used two approaches at the same time: i) Using an “Other” class: as described in the previous section, ii) Using a confidence level threshold: a confidence level threshold of 80% is used to decide whether to accept the classifier’s decision … view at source ↗
Figure 3
Figure 3. Figure 3: Examples of highlighted images. 2. We added more diverse images to that category in the CCTI14 dataset to address the problem identified in (1). 3. If the false positive rate decreased, then we kept the diverse images. 4. Else, go to (1). Following this type of methodology generally for all categories helped us increase the robustness of the trained classifier [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Confusion matrix of the text classifier. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of false positives. Category #Censored posts #Uncensored posts Censorship Rate Bo Xilai 665 336 64% Deng Xiaoping 281 125 70% Fire 431 530 45% Injury/Dead Body 1799 1029 51% Liu Xiaobo 184 123 60% Mao Zedong 1093 486 70% People’s Congress 145 113 56% Policeman 1311 927 59% Protest 536 220 71% Prurient/Nudity 2664 2551 51% Rainstorm 153 207 43% Winnie the Pooh 160 177 48% Xi Jinping 1745 1029 63% Z… view at source ↗
Figure 6
Figure 6. Figure 6: Categories vs. life time 9 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Widespread Chinese social media applications such as Weibo are widely known for monitoring and deleting posts to conform to Chinese government requirements. In this paper, we focus on analyzing a dataset of censored and uncensored posts in Weibo. Despite previous work that only considers text content of posts, we take a multi-modal approach that takes into account both text and image content. We categorize this dataset into 14 categories that have the potential to be censored on Weibo, and seek to quantify censorship by topic. Specifically, we investigate how different factors interact to affect censorship. We also investigate how consistently and how quickly different topics are censored. To this end, we have assembled an image dataset with 18,966 images, as well as a text dataset with 994 posts from 14 categories. We then utilized deep learning, CNN localization, and NLP techniques to analyze the target dataset and extract categories, for further analysis to better understand censorship mechanisms in Weibo. We found that sentiment is the only indicator of censorship that is consistent across the variety of topics we identified. Our finding matches with recently leaked logs from Sina Weibo. We also discovered that most categories like those related to anti-government actions (e.g. protest) or categories related to politicians (e.g. Xi Jinping) are often censored, whereas some categories such as crisis-related categories (e.g. rainstorm) are less frequently censored. We also found that censored posts across all categories are deleted in three hours on average.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper assembles a dataset of 994 Weibo posts (and 18,966 associated images) spanning 14 hand-labeled categories, applies multi-modal CNN/NLP analysis, and reports that (i) sentiment is the sole consistent predictor of censorship across categories, (ii) anti-government/protest and politician-related categories are censored more frequently than crisis-related ones, and (iii) censored posts are deleted after an average of three hours. These empirical patterns are said to align with leaked Sina Weibo logs.

Significance. If the category labels and consistency claims survive validation, the work supplies a rare multi-modal empirical window into real-world censorship dynamics on a major platform and identifies a single cross-topic signal (sentiment) that could be tested in follow-up studies.

major comments (3)
  1. [Abstract] Abstract and methods description: the manual assignment of 994 posts to 14 categories is presented without inter-annotator agreement, label-validation metrics, or selection protocol; because all downstream claims about consistent indicators and differential censorship rates rest on these labels, the absence of these statistics is load-bearing.
  2. [Abstract] Abstract: the claim that 'sentiment is the only indicator of censorship that is consistent across the variety of topics' is stated as an empirical finding, yet no table, figure, or statistical test is referenced that demonstrates consistency after controlling for category or that rules out image-only or text-only baselines.
  3. [Abstract] Abstract: the reported three-hour average deletion time and the differential censorship frequencies by category are given without error bars, sample sizes per category, or any ablation showing that the multi-modal (CNN) features contribute explanatory power beyond text features alone.
minor comments (2)
  1. [Abstract] The abstract supplies dataset sizes but does not indicate whether the 18,966 images are unique or how they were paired with the 994 posts.
  2. [Abstract] Notation for the 14 categories is introduced without an explicit list or example posts in the provided abstract.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We address each major comment point by point below, with a focus on clarifying the labeling process, strengthening statistical reporting, and adding requested validations and ablations where feasible. Revisions will be made to the abstract and methods/results sections accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract and methods description: the manual assignment of 994 posts to 14 categories is presented without inter-annotator agreement, label-validation metrics, or selection protocol; because all downstream claims about consistent indicators and differential censorship rates rest on these labels, the absence of these statistics is load-bearing.

    Authors: We agree that documentation of the labeling process is essential. The 994 posts were drawn from a larger crawl of Weibo content and assigned to the 14 categories by the authors using domain knowledge of sensitive topics in the Chinese context; category definitions were based on prior literature on Weibo censorship. We will add a full description of the selection protocol and category definitions to the methods section, along with the distribution of posts per category as a form of label validation. However, labeling was performed by a single annotator, so inter-annotator agreement cannot be computed retroactively. This constitutes a partial revision. revision: partial

  2. Referee: [Abstract] Abstract: the claim that 'sentiment is the only indicator of censorship that is consistent across the variety of topics' is stated as an empirical finding, yet no table, figure, or statistical test is referenced that demonstrates consistency after controlling for category or that rules out image-only or text-only baselines.

    Authors: The results section of the manuscript reports the multi-modal (CNN + NLP) analysis and identifies sentiment as the sole consistent predictor via per-category comparisons. We will revise the abstract to explicitly reference the relevant tables/figures and statistical tests that demonstrate this consistency after controlling for category. We will also add an explicit comparison to text-only and image-only baselines to rule out that the signal is driven by a single modality. revision: yes

  3. Referee: [Abstract] Abstract: the reported three-hour average deletion time and the differential censorship frequencies by category are given without error bars, sample sizes per category, or any ablation showing that the multi-modal (CNN) features contribute explanatory power beyond text features alone.

    Authors: We will update both the abstract and results sections to report per-category sample sizes (summing to the total of 994 posts), standard errors or confidence intervals around the three-hour average deletion time, and a new ablation study comparing the full multi-modal model against a text-only baseline to quantify the added value of the CNN image features. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical observations on external dataset

full rationale

The paper collects a dataset of 994 posts and 18,966 images, manually assigns them to 14 categories, then applies CNN/NLP analysis to observe patterns such as sentiment consistency and differential censorship rates. These are direct empirical measurements on the assembled data rather than any derivation, fitted parameter renamed as prediction, or self-citation chain that reduces the result to its inputs by construction. The match to external leaked Sina Weibo logs supplies independent corroboration. No equations, ansatzes, or uniqueness theorems appear; the central claims remain falsifiable observations outside any definitional loop.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper relies on standard supervised-learning assumptions (i.i.d. samples, label accuracy) and on the authors' choice of 14 categories; no free parameters, invented entities, or non-standard axioms are visible in the abstract.

axioms (2)
  • domain assumption The 14 topic categories chosen by the authors form a representative partition of content that may be censored.
    Abstract states the dataset was categorized into these 14 categories without justifying completeness or lack of overlap.
  • domain assumption Human or model labels for the 994 posts and 18,966 images are sufficiently accurate to support claims about consistency across categories.
    No inter-annotator agreement or validation procedure is mentioned.

pith-pipeline@v0.9.0 · 5832 in / 1482 out tokens · 20022 ms · 2026-05-25T15:21:26.082135+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 4 internal anchors

  1. [1]

    https : / / chinadigitaltimes

    Five years of sensitive words on june fourth. https : / / chinadigitaltimes . net / 2016 / 06 / five-years-sensitive-words-june-fourth/ . Accessed: 2019-02-20

    work page 2016

  2. [2]

    https://en.wikipedia.org/wiki/July_2009_Urumqi\_riots

    July 2009 urumqi riots. https://en.wikipedia.org/wiki/July_2009_Urumqi\_riots. Accessed: 2018- 08-30

    work page 2009

  3. [3]

    https://drive.google.com/file/d/ 1xjHWI0Ih3abONxL0WoXj2swlcBuiRCXO/view

    Who not what: The logic of china’s information control strategy. https://drive.google.com/file/d/ 1xjHWI0Ih3abONxL0WoXj2swlcBuiRCXO/view. Accessed: 2019-03-14. 11 A PREPRINT - J ULY 4, 2019

    work page 2019

  4. [4]

    Online:https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-of-users/ (2017)

    Most popular social networks worldwide as of july 2018, ranked by number of active users. Online:https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-of-users/ (2017)

    work page 2018

  5. [5]

    Establishing a database for studying human face photograph memory

    BAINBRIDGE , W., I SOLA , P., B LANK , I., AND OLIVA, A. Establishing a database for studying human face photograph memory. In Proceedings of the Annual Meeting of the Cognitive Science Society (2012), vol. 34

    work page 2012

  6. [6]

    A., I SOLA , P., AND OLIVA, A

    BAINBRIDGE , W. A., I SOLA , P., AND OLIVA, A. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General 142, 4 (2013), 1323

    work page 2013

  7. [7]

    Censorship and deletion practices in chinese social media

    BAMMAN , D., O’C ONNOR , B., AND SMITH , N. Censorship and deletion practices in chinese social media. First Monday 17, 3 (2012)

    work page 2012

  8. [8]

    A coefficient of agreement for nominal scales

    COHEN , J. A coefficient of agreement for nominal scales. Educational and psychological measurement 20, 1 (1960), 37–46

    work page 1960

  9. [9]

    N G, J., R UAN, L., T SUI , L., AND XIONG , R

    CRETE -N ISHIHATA , M., K NOCKEL , J., M ILLER , B., Q. N G, J., R UAN, L., T SUI , L., AND XIONG , R. Analyzing censorship of the death of liu xiaobo on wechat and weibo

    work page

  10. [10]

    Assessing censorship on microblogs in china: Discriminatory keyword analysis and the real-name registration policy

    FU, K.- W., C HAN , C.- H., AND CHAU, M. Assessing censorship on microblogs in china: Discriminatory keyword analysis and the real-name registration policy. IEEE Internet Computing 17, 3 (2013), 42–50

    work page 2013

  11. [11]

    KING , G., P AN, J., AND ROBERTS , M. E. How censorship in china allows government criticism but silences collective expression. American Political Science Review 107, 2 (2013), 326–343

    work page 2013

  12. [12]

    Q., S ENFT , A., AND CRANDALL , J

    KNOCKEL , J., C RETE -NISHIHATA , M., N G, J. Q., S ENFT , A., AND CRANDALL , J. R. Every rose has its thorn: Censorship and surveillance on social video platforms in china. In 5th{USENIX} Workshop on Free and Open Communications on the Internet ({FOCI} 15) (2015)

    work page 2015

  13. [13]

    An analysis of automatic image filtering on wechat moments

    KNOCKEL , J., R UAN, L., AND CRETE -N ISHIHATA , M. An analysis of automatic image filtering on wechat moments. In 8th USENIX Workshop on Free and Open Communications on the Internet (FOCI 18) (2018), USENIX Association

    work page 2018

  14. [14]

    The stanford corenlp natural language processing toolkit

    MANNING , C., S URDEANU , M., B AUER , J., F INKEL , J., B ETHARD , S., AND MCCLOSKY , D. The stanford corenlp natural language processing toolkit. In Proceedings of 52nd annual meeting of the association for computational linguistics: system demonstrations (2014), pp. 55–60

    work page 2014

  15. [15]

    The limits of commercialized censorship in china

    M ILLER , B. The limits of commercialized censorship in china

    work page

  16. [16]

    Y., F ELDMAN , A., AND LEBERKNIGHT , C

    NG, K. Y., F ELDMAN , A., AND LEBERKNIGHT , C. Detecting censorable content on sina weibo: A pilot study. In Proceedings of the 10th Hellenic Conference on Artificial Intelligence (2018), ACM, p. 41

    work page 2018

  17. [17]

    Linguistic Characteristics of Censorable Language on SinaWeibo

    NG, K. Y., F ELDMAN , A., P ENG , J., AND LEBERKNIGHT , C. Linguistic characteristics of censorable language on sinaweibo. arXiv preprint arXiv:1807.03654 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  18. [18]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    SIMONYAN , K., AND ZISSERMAN , A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  19. [19]

    Understanding data augmentation for classification: when to warp?

    WONG , S. C., G ATT, A., S TAMATESCU , V., AND MCDONNELL , M. D. Understanding data augmentation for classification: when to warp? arXiv preprint arXiv:1609.08764 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  20. [20]

    A., O LIVA, A., AND TORRALBA , A

    XIAO, J., H AYS, J., E HINGER , K. A., O LIVA, A., AND TORRALBA , A. Sun database: Large-scale scene recognition from abbey to zoo. In Computer vision and pattern recognition (CVPR), 2010 IEEE conference on (2010), IEEE, pp. 3485–3492

    work page 2010

  21. [21]

    Improved Relation Classification by Deep Recurrent Neural Networks with Data Augmentation

    XU, Y., J IA, R., M OU, L., L I, G., C HEN , Y., L U, Y., AND JIN, Z. Improved relation classification by deep recurrent neural networks with data augmentation. arXiv preprint arXiv:1601.03651 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  22. [22]

    Learning deep features for discriminative localization

    ZHOU , B., K HOSLA , A., L APEDRIZA , A., O LIVA, A., AND TORRALBA , A. Learning deep features for discriminative localization. In Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on (2016), IEEE, pp. 2921–2929

    work page 2016

  23. [23]

    R., AND WALLACH , D

    ZHU, T., P HIPPS , D., P RIDGEN , A., C RANDALL , J. R., AND WALLACH , D. S. The velocity of censorship: High-fidelity detection of microblog post deletions. In USENIX Security Symposium (2013), pp. 227–240. 12

    work page 2013