Pith. sign in

REVIEW 3 major objections 4 minor 203 references

A single text-based pipeline can measure political partisanship on different social media platforms using only post text and external media-bias ratings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:31 UTC pith:ZO7DSQB6

load-bearing objection A clean, worthwhile extension of Waller & Anderson to cross-platform text, but the validation is circular and the seed pairs mix topics, so the portability claim doesn't hold as written. the 3 major comments →

arxiv 2607.21842 v1 pith:ZO7DSQB6 submitted 2026-07-23 cs.SI cs.LG

Quantifying Political Partisanship for Cross-Platform Analyses

Fathima Ameen , Christopher G. Healey This is my paper
classification cs.SI cs.LG
keywords political partisanshipsocial mediacross-platform analysisembedding spacemedia bias ratingsBlueskyTruth Socialsocial dimension
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that political partisanship in social media posts can be measured with one portable, text-only pipeline, so scores mean the same thing on different platforms. The method embeds posts into a semantic space, groups them into topics, labels clusters using media-bias ratings of cited news sources, and builds a partisanship axis from opposing clusters. Posts are then scored by projecting onto that axis. The paper shows the scores correlate with held-out media-bias ratings on the platforms used to build the axis and on an unseen Twitter corpus, and that the scores track event-driven shifts in sentiment, such as Democrats' fading support for Biden on Bluesky. If true, this gives researchers a general tool for comparing polarization, echo chambers, and radicalization across the fragmented social media ecosystem.

Core claim

The central claim is that a partisanship score can be defined purely by text, anchored by external media-bias ratings, and transferred across platforms. After embedding posts and clustering them into topics, the method labels each cluster by the aggregated bias scores of news outlets cited within it. It then selects pairs of left- and right-labeled clusters whose centroids are most similar—reasoning that such pairs differ mainly in political framing, not topic—and forms the partisanship axis as the mean difference between their centroids. Every post's score is the projection of its embedding onto that axis. Validation shows the axis correlates with held-out media-bias scores both within the

What carries the argument

The load-bearing object is the partisanship axis: a direction in the embedding space defined as the average difference between centroid pairs of semantically similar but ideologically opposite topic clusters. It is built from clusters labeled left or right via Empirical-Bayes-shrunk media-bias scores of cited news sources, and each post receives a partisanship score by projecting its normalized embedding onto the axis. The axis is what makes scores comparable across platforms, because it is fit once on text alone and then applied to any post.

Load-bearing premise

The axis assumes that pairs of topic clusters whose embeddings are similar differ mainly in political ideology rather than in subject matter, style, or platform dialect; if that assumption fails, the axis measures something other than partisanship.

What would settle it

Take the fixed partisanship axis and apply it to posts from a platform or community whose dominant topics differ from those in the seed clusters (for example, local news or non-political content that still cites rated outlets); if the correlation between projected scores and media-bias ratings disappears or reverses, the axis is measuring topic separation rather than general partisanship.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can compare partisanship distributions across platforms sharing the same axis, making cross-platform polarization studies feasible without re-training classifiers.
  • The method works on any text-based platform where users share rated news outlets at non-trivial rates, lowering barriers for studying fringe alt-tech platforms.
  • Because scores are continuous and platform-agnostic, they enable temporal tracking of partisan sentiment toward candidates and issues within a single platform (for example, the observed leftward shift for Harris and rightward drift for Biden on Bluesky).
  • The validation approach—checking projected scores against independent media-bias ratings—offers a general template for trusting embedding-based political dimensions.

Where Pith is reading between the lines

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

  • The method's correlation on X (r≈0.19) is weaker than in-distribution; a fair reading suggests the axis transfers direction but the magnitude of scores may be platform-specific, so cross-platform score distributions may need calibration even if rank order is meaningful.
  • A natural testable extension is to apply the same pipeline to multi-party or non-U.S. contexts using local bias-rating tools; the paper's own limitation discussion suggests the axis assumption is U.S.-centric, and the cluster-pair reasoning may need topic-matching refinements to avoid conflating partisanship with topic differences.
  • If the method is applied to platforms with different language conventions (for example, non-English), both the transformer encoder and the bias-anchoring signal would need to be language-matched; the paper's claims are scoped to English-language political discourse.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a text-based, platform-portable method for measuring political partisanship in social media posts. Posts are embedded with a transformer encoder, clustered into topic groups, labeled using aggregated AllSides media-bias scores of cited news outlets, and a partisanship axis is constructed as the difference between centroids of left- and right-labeled cluster pairs. Individual posts are scored by projection onto this axis. The method is applied to ~1.3M posts from Bluesky and Truth Social from May–October 2024. The authors report significant correlations between projected scores and AllSides bias scores on held-out posts from the same platforms (r=0.365) and on an independent X/Twitter corpus (r=0.193), and they present within-platform temporal patterns as evidence that the axis captures partisanship beyond platform identity.

Significance. If the method and validation were sound, this would be a valuable contribution: a text-only pipeline that could measure partisanship comparably across platforms without platform-specific features would enable cross-platform polarization research in the fragmented alt-tech ecosystem. The application to Bluesky and Truth Social is timely, and the detailed pipeline is clearly described. However, the current evidence for the central claim is substantially weakened by circularity in the validation design and by a topic confound in the seed-pair construction. The contribution therefore remains promising but not yet established by the manuscript as written.

major comments (3)
  1. [3 (Cluster Labeling; Social Dimension Construction) and 4 (Results)] The validation is circular in two ways. First, the same AllSides ratings are used to label clusters, to define the left/right pools for seed selection, and as the outcome variable in both validation correlations. A post's AllSides score influences the axis indirectly through the cluster labels, so the reported 'held-out' correlation is not an independent check. Second, the number of seed pairs (4) was chosen after testing 1–9 and selecting the 'most robust output,' which is model selection on the validation metric. Consequently the reported r=0.365 and the 100th-percentile result in Figure 3 are optimistically biased. The authors should prespecify the number of pairs or use a nested validation procedure, and validate against an external partisan signal that does not feed into clustering/labeling.
  2. [3 (Social Dimension Construction) and Table 3] The central assumption behind the axis is that high centroid similarity pairs differ primarily in partisanship rather than content. Table 3 contradicts this: pair 2 is 'Pro-Palestine, Biden-Israel criticism' versus 'Immigration & border security outrage'; pair 4 is 'Trump controversies and criticism' versus 'Conspiracy theories & Extremism.' The resulting axis is a linear combination of topic differences, not a pure partisan dimension. Because left- and right-leaning outlets systematically cover different topics, the positive validation correlations can arise from topic separation alone. The X-transfer correlation (r=0.193) is weak and does not rule out this alternative. The authors should require seed pairs to be topically matched (e.g., same issue discussed from opposing stances) or otherwise control for topic in validation.
  3. [4 (Sensitivity to Seed Pair Selection)] The exhaustive-pair baseline does not address the topic confound. Many arbitrary cluster pairs may yield positive correlations simply because they separate news sources by topic. Showing that the chosen axis outperforms the majority of arbitrary pairs is consistent with the axis capturing a topic dimension that happens to align with AllSides bias. A more informative baseline would match seed pairs on topic or hold topic fixed. Additionally, because the number of pairs was selected using the same validation data, the percentile comparison is not an honest null distribution.
minor comments (4)
  1. [2 (Literature Review)] Typo: 'mainstream platformsm' should be 'mainstream platforms.'
  2. [4 (Figure 3)] The caption says 'Each bar represents the number of cluster pairs' but does not specify the binning for this histogram. If the selected axis is marked, clarify how its location is indicated.
  3. [3 (Validation)] The term 'held-out' is misleading for the in-distribution validation: posts in the seed clusters are excluded, but their AllSides scores were used to label clusters. Consider using 'excluded from axis construction' or similar.
  4. [3 (Cluster Labeling)] Equation for Empirical Bayes shrinkage: define σ̂²_c precisely; it appears to be the variance of bias scores across posts cited within cluster c, but this is not stated in the text.

Circularity Check

3 steps flagged

The AllSides signal used to build the axis is also used as the validation ground truth, and the seed-pair count is selected on the same metric; the in-distribution validation is therefore partly circular, though the X transfer test provides some independent support.

specific steps
  1. self definitional [Section 3, Cluster Labeling using News Domains; Section 3, Validation]
    "We validate the partisanship axis in two ways, both of which exploit the fact that posts citing news sources carry an independent ground-truth signal—the AllSides bias score of the cited outlet—against which our projection-based partisanship scores can be compared."

    The 'independent ground-truth signal' is the same AllSides score used to label clusters ('Clusters are then labeled right-leaning if θc ≥ 0.5, left-leaning if θc ≤ −0.5') and to select the seed pairs ('we select the top 4 seed pairs based on cosine similarity' among left/right clusters). The axis d is the mean of differences between centroids of AllSides-labeled clusters, so projected post scores are a direct function of AllSides labels. The in-distribution validation therefore compares the method's output to its own labeling input; a positive correlation is an internal consistency check, not an independent test that the axis measures partisanship.

  2. fitted input called prediction [Section 3, Social Dimension Construction]
    "We chose four pairs after testing between one and nine seed pairs. Our analysis showed four pairs produced the most robust output."

    The number of seed pairs is a hyperparameter selected by testing candidate settings and picking the one with the 'most robust output'—in context, the validation correlation with AllSides scores. The reported in-distribution correlation (r = 0.365) is then presented as validation evidence, but the chosen configuration was selected on that same metric. The 'prediction' is partly a fitted value; the paper does not report whether the selected k=4 is held out from a separate model-selection procedure.

  3. other [Section 3, Social Dimension Construction; Table 3]
    "Pairs with high centroid similarity are presumed to differ primarily along the partisan dimension rather than in content, making them ideal seed pairs. ... Table 3: 'Pro-Palestine, Biden-Israel criticism' vs 'Immigration & border security outrage'"

    The axis is defined as the difference between centroids of such seed pairs, so the partisanship score is, by construction, a linear contrast between those cluster topics. The paper's own Table 3 shows that the second pair contrasts Pro-Palestine discourse with immigration discourse—different topics, not the same topic framed from opposing partisan perspectives. Since left- and right-leaning news outlets systematically emphasize different issues, the AllSides validation correlation can arise from topic separation alone, without the axis measuring general partisanship. The validation cannot distinguish topic-contrast from partisan valence, so the central portability claim is not independently established by the in-distribution correlation.

full rationale

The derivation chain is not fully circular: the axis is built from cluster centroids, the in-distribution evaluation uses held-out posts not in the seed clusters, and the out-of-distribution X test is genuinely unseen during axis construction. The exhaustive sensitivity analysis also shows that most arbitrary cluster-pair axes do not correlate, so the selected axis carries real information. However, the AllSides signal enters both as the labeling/seed-selection input and as the validation outcome. The in-distribution correlation therefore measures how well the axis reproduces the same external label source it was built from; it is a consistency check rather than an independent confirmation. The seed-pair count (k=4) was selected by testing 1-9 and choosing the 'most robust output,' which is effectively tuning on the validation metric. Finally, the assumption that high-similarity left/right cluster pairs differ 'primarily along the partisan dimension rather than in content' is contradicted by Table 3's own topic summaries (e.g., Pro-Palestine vs. immigration), so the axis may encode topic differences that correlate with AllSides because left- and right-leaning outlets cover different topics. The X transfer result (r=0.193) provides some independent support, but it is weak and could also reflect cross-platform topic differences. Overall, the central in-distribution validation is partially circular and the method's portability claim is not as strongly established as presented. Score 6 reflects partial circularity, not full equivalence.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central method rests on AllSides as ground truth, on embedding-geometry assumptions, and on a linear axis fitted through the same external label source. No independent human-annotated labels or external benchmarks are used. The free parameters (seed-pair count, k values, thresholds, PCA cutoff) are methodologically consequential and several are chosen by inspecting the output.

free parameters (5)
  • Number of seed pairs = 4
    Authors tested 1-9 seed pairs and chose 4 based on 'most robust output,' which is a validation-based selection that inflates the reported correlation.
  • MiniBatch k-means k = 15 (dedup); 36 (topic clustering)
    k=15 is used for per-platform-month semantic deduplication; k=36 is chosen by elbow/kneed. Both determine which clusters become seed candidates and shape the axis.
  • Near-duplicate cosine threshold = 0.90
    Arbitrary threshold that removes posts closer to the cluster centroid when pairwise similarity exceeds 0.90; directly affects which posts remain.
  • PCA variance retention = 95% (104 components)
    Component count changes the geometry of the embedding space and therefore the clusters and centroids used to build the axis.
  • Cluster label thresholds = ±0.5
    Clusters with shrunken mean bias ≥0.5 are labeled right, ≤-0.5 left, else neutral/no signal; no external justification for these cutoffs.
axioms (5)
  • domain assumption AllSides media-bias ratings are a valid external ground truth for partisanship of news domains.
    Cluster labels, seed-pair selection, and all validation correlations depend on AllSides scores; the paper does not independently validate this rating source.
  • domain assumption Qwen3 embedding cosine similarity aligns with political ideology and topic structure across platforms.
    The whole pipeline assumes that semantically similar posts (and centroids) in this embedding space are informative about partisanship, not just about platform or style.
  • ad hoc to paper High centroid-similarity cluster pairs differ primarily in partisanship, not in topic.
    Explicitly stated in Section 3 as the seed-pair selection premise; if false, the axis captures topic or style rather than partisanship.
  • domain assumption Posts without news citations can be scored using an axis built from clusters labeled by news-citing posts.
    The method applies the axis to all posts, but the labels and centroid differences come only from posts that cite AllSides-rated domains; generalizing to non-citing posts is assumed.
  • domain assumption Keyword/hashtag-filtered posts are representative of each platform's political discourse.
    Truth Social was collected via trending hashtags and political keywords, while Bluesky used a keyword search; different sampling strategies may confound cross-platform comparisons.
invented entities (1)
  • Partisanship axis d no independent evidence
    purpose: A latent linear dimension in embedding space used to score all posts by projection; defined as the mean of centroid differences of left/right seed clusters.
    The axis is constructed using AllSides-labeled clusters and validated against the same AllSides signal, so it does not have a falsifiable handle outside the paper's correlation framework.

pith-pipeline@v1.3.0-alltime-deepseek · 13975 in / 13046 out tokens · 133085 ms · 2026-08-01T06:31:48.803832+00:00 · methodology

0 comments
read the original abstract

Research on political polarization on social media depends on the ability to reliably measure partisanship in user-generated content. However, existing approaches are typically tailored to platform-specific properties, such as structural affordances or linguistic conventions, which hurts generalizability across platforms. This limitation is increasingly consequential as the social media ecosystem fragments and fringe, alt-tech platforms emerge alongside mainstream ones. We propose a text-based, platform-portable methodology for measuring political partisanship in social media posts, anchored by an external news-credibility signal. Posts are embedded using a transformer-based sentence encoder and clustered into topic groups, which are labeled using the aggregated AllSides media bias scores of cited news outlets. A partisanship axis is then constructed in the embedding space as the difference between centroids of oppositely labeled clusters, and individual posts are scored by projection onto this axis. We apply the method to a corpus of approximately 1.3 million posts collected from Bluesky and Truth Social during the six months preceding the 2024 U.S. presidential election, providing the first cross-platform comparison of partisanship distributions on these two ideologically asymmetric platforms. The resulting partisanship scores correlate significantly with held-out AllSides media bias scores both in-distribution and out-of-distribution on an independent Twitter corpus, and recover within-platform partisan dynamics that platform identity alone cannot explain.

Figures

Figures reproduced from arXiv: 2607.21842 by Christopher G. Healey, Fathima Ameen.

Figure 1
Figure 1. Figure 1: Overview of the social dimension construction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Monthly share of total posts and average engage [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity of the partisan axis to seed pair selec [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Mean partisanship score (zp) of posts containing each keyword, stratified by platform and month. Negative scores (blue) indicate left-leaning partisan alignment; positive scores (red) indicate right-leaning alignment. Keywords are grouped into candidate names (top) and party-affiliation terms (bottom). Highlighted with a purple border are the scores for the Demo￾cratic candidates in the month of Septem… view at source ↗

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

203 extracted references · 74 canonical work pages · 9 internal anchors

  1. [1]

    Proceedings of the first instructional conference on machine learning , volume=

    Using tf-idf to determine word relevance in document queries , author=. Proceedings of the first instructional conference on machine learning , volume=. 2003 , organization=

    2003

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Matryoshka representation learning , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  4. [4]

    Proceedings of the Second Workshop on NLP and Computational Social Science , pages=

    community2vec: Vector representations of online communities encode semantic relationships , author=. Proceedings of the Second Workshop on NLP and Computational Social Science , pages=

  5. [5]

    Vaccines , volume=

    \# Scamdemic,\# Plandemic, or\# Scaredemic: what Parler social media platform tells us about COVID-19 vaccine , author=. Vaccines , volume=. 2021 , publisher=

    2021

  6. [6]

    Proceedings of the ACM on Human-Computer Interaction , volume=

    Content Moderation and Hate Speech on Alternative Platforms: A Case Study of BitChute , author=. Proceedings of the ACM on Human-Computer Interaction , volume=. 2025 , publisher=

    2025

  7. [7]

    2018 IEEE/ACM international conference on advances in social networks analysis and mining (ASONAM) , pages=

    Inside the right-leaning echo chambers: Characterizing gab, an unmoderated social system , author=. 2018 IEEE/ACM international conference on advances in social networks analysis and mining (ASONAM) , pages=. 2018 , organization=

    2018

  8. [8]

    Addawood, Aseel and Bashir, Masooda , year =. ". Proceedings of the. doi:10.18653/v1/W16-2801 , abstract =

    work page doi:10.18653/v1/w16-2801

  9. [9]

    Annotation and

    Schaefer, Robin and Stede, Manfred , file =. Annotation and

  10. [10]

    it - Information Technology , author =

    Argument. it - Information Technology , author =. 2021 , pages =. doi:10.1515/itit-2020-0053 , abstract =

    work page doi:10.1515/itit-2020-0053 2021

  11. [11]

    Proceedings of the International AAAI Conference on Web and Social Media , author =

    Online. Proceedings of the International AAAI Conference on Web and Social Media , author =. 2024 , pages =. doi:10.1609/icwsm.v18i1.31383 , abstract =

    work page doi:10.1609/icwsm.v18i1.31383 2024

  12. [12]

    Investigating Chain-of-thought with ChatGPT for Stance Detection on Social Media

    Zhang, Bowen and Fu, Xianghua and Ding, Daijun and Huang, Hu and Dai, Genan and Yin, Nan and Li, Yangyang and Jing, Liwen , month = oct, year =. Investigating. doi:10.48550/arXiv.2304.03087 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.03087

  13. [13]

    Proceedings of the International AAAI Conference on Web and Social Media , author =

    Stance. Proceedings of the International AAAI Conference on Web and Social Media , author =. 2024 , pages =. doi:10.1609/icwsm.v18i1.31360 , abstract =

    doi:10.1609/icwsm.v18i1.31360 2024

  14. [14]

    IEEE Transactions on Computational Social Systems , author =

    Improved. IEEE Transactions on Computational Social Systems , author =. 2023 , pages =. doi:10.1109/TCSS.2023.3320723 , language =

    arXiv 2023

  15. [15]

    Yuan, Jiaqing and Xi, Ruijie and Singh, Munindar P , file =. A

  16. [16]

    Information Processing & Management , author =

    Is my stance the same as your stance?. Information Processing & Management , author =. 2022 , pages =. doi:10.1016/j.ipm.2022.103070 , abstract =

    arXiv 2022

  17. [17]

    2024 , pages =

    Proceedings of the International AAAI Conference on Web and Social Media , author =. 2024 , pages =. doi:10.1609/icwsm.v18i1.31424 , abstract =

    work page doi:10.1609/icwsm.v18i1.31424 2024

  18. [18]

    Li, Hao and Schlegel, Viktor and Sun, Yizheng and Batista-Navarro, Riza and Nenadic, Goran , month = nov, year =. Large. doi:10.48550/arXiv.2506.16383 , abstract =

    doi:10.48550/arxiv.2506.16383

  19. [19]

    Exploring the

    Chen, Guizhen and Cheng, Liying and Luu, Anh Tuan and Bing, Lidong , year =. Exploring the. Proceedings of the 62nd. doi:10.18653/v1/2024.acl-long.126 , language =

    doi:10.18653/v1/2024.acl-long.126 2024

  20. [20]

    ACM Transactions on Intelligent Systems and Technology , author =

    Prompting and. ACM Transactions on Intelligent Systems and Technology , author =. 2025 , pages =. doi:10.1145/3725816 , abstract =

    work page doi:10.1145/3725816 2025

  21. [21]

    Reasoner

    Yuan, Jiaqing and Xi, Ruijie and Singh, Munindar P , month = sep, year =. Reasoner. Proceedings of the 36th. doi:10.1145/3720553.3746668 , abstract =

    arXiv

  22. [22]

    How would Stance Detection Techniques Evolve after the Launch of ChatGPT?

    Zhang, Bowen and Ding, Daijun and Jing, Liwen and Dai, Genan and Yin, Nan , month = aug, year =. How would. doi:10.48550/arXiv.2212.14548 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.14548

  23. [23]

    Körner, Erik and Wiedemann, Gregor and Hakimi, Ahmad Dawar and Heyer, Gerhard and Potthast, Martin , year =. On. Proceedings of the 2021. doi:10.18653/v1/2021.emnlp-main.795 , abstract =

    work page doi:10.18653/v1/2021.emnlp-main.795 2021

  24. [24]

    Neural Computing and Applications , author =

    A systematic review of machine learning techniques for stance detection and its applications , volume =. Neural Computing and Applications , author =. 2023 , pages =. doi:10.1007/s00521-023-08285-7 , abstract =

    work page doi:10.1007/s00521-023-08285-7 2023

  25. [25]

    Zhang, Bowen and Dai, Genan and Niu, Fuqiang and Yin, Nan and Fan, Xiaomao and Wang, Senzhang and Cao, Xiaochun and Huang, Hu , month = nov, year =. A. doi:10.48550/arXiv.2409.15690 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2409.15690

  26. [26]

    Aker, Ahmet and Derczynski, Leon and Bontcheva, Kalina , month = sep, year =. Simple. doi:10.48550/arXiv.1708.05286 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1708.05286

  27. [27]

    Predicting

    Poddar, Lahari and Hsu, Wynne and Lee, Mong Li and Subramaniyam, Shruti , month = nov, year =. Predicting. 2018. doi:10.1109/ICTAI.2018.00021 , abstract =

    arXiv 2018

  28. [28]

    Liang, Bin and Chen, Zixiao and Gui, Lin and He, Yulan and Yang, Min and Xu, Ruifeng , month = apr, year =. Zero-. Proceedings of the. doi:10.1145/3485447.3511994 , abstract =

    arXiv

  29. [29]

    2024 , pages =

    Proceedings of the International AAAI Conference on Web and Social Media , author =. 2024 , pages =. doi:10.1609/icwsm.v18i1.31342 , abstract =

    work page doi:10.1609/icwsm.v18i1.31342 2024

  30. [30]

    Examining the

    Ng, Lynnette Hui Xian and Cruickshank, Iain J and Lee, Roy , file =. Examining the

  31. [31]

    Integrated or

    Xia, Yan and Monti, Corrado and Keller, Barbara and Kivelä, Mikko , file =. Integrated or

  32. [32]

    Large Language Models Meet NLP: A Survey

    Qin, Libo and Chen, Qiguang and Feng, Xiachong and Wu, Yang and Zhang, Yongheng and Li, Yinghui and Li, Min and Che, Wanxiang and Yu, Philip S. , month = aug, year =. Large. doi:10.48550/arXiv.2405.12819 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.12819

  33. [33]

    Minaee, Shervin and Mikolov, Tomas and Nikzad, Narjes and Chenaghlu, Meysam and Socher, Richard and Amatriain, Xavier and Gao, Jianfeng , month = mar, year =. Large. doi:10.48550/arXiv.2402.06196 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.06196

  34. [34]

    Jamuie, Adelani and Blessing, Funke , file =. Natural

  35. [35]

    and Sap, Maarten , year =

    Antoniak, Maria and Field, Anjalie and Mun, Jimin and Walsh, Melanie and Klein, Lauren F. and Sap, Maarten , year =. Riveter:. Proceedings of the 61st. doi:10.18653/v1/2023.acl-demo.36 , abstract =

    doi:10.18653/v1/2023.acl-demo.36 2023

  36. [36]

    Sentence-

    Reimers, Nils and Gurevych, Iryna , month = aug, year =. Sentence-. doi:10.48550/arXiv.1908.10084 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1908.10084 1908

  37. [37]

    Proceedings of the 16th

    Alhuzali, Hassan and Ananiadou, Sophia , year =. Proceedings of the 16th. doi:10.18653/v1/2021.eacl-main.135 , abstract =

    work page doi:10.18653/v1/2021.eacl-main.135 2021

  38. [38]

    Unfiltered

    Shah, Kashish and Gerard, Patrick and Luceri, Luca and Ferrara, Emilio , year =. Unfiltered. doi:10.2139/ssrn.5018881 , abstract =

    work page doi:10.2139/ssrn.5018881

  39. [39]

    Balduf, Leonhard and Sokoto, Saidu and Ascigil, Onur and Tyson, Gareth and Scheuermann, Björn and Korczyński, Maciej and Castro, Ignacio and Król, Michaŀ , month = nov, year =. Looking. Proceedings of the 2024. doi:10.1145/3646547.3688407 , abstract =

    arXiv 2024

  40. [40]

    IEEE Data Descriptions , author =

    Descriptor:. IEEE Data Descriptions , author =. 2024 , pages =. doi:10.1109/IEEEDATA.2024.3474640 , abstract =

    arXiv 2024

  41. [41]

    Emergent: a novel data-set for stance classification , shorttitle =

    Ferreira, William and Vlachos, Andreas , year =. Emergent: a novel data-set for stance classification , shorttitle =. Proceedings of the 2016. doi:10.18653/v1/N16-1138 , abstract =

    doi:10.18653/v1/n16-1138 2016

  42. [42]

    PolitiSky24: U.S. Political Bluesky Dataset with User Stance Labels

    Rostami, Peyman and Rahimzadeh, Vahid and Adibi, Ali and Shakery, Azadeh , month = jun, year =. doi:10.48550/arXiv.2506.07606 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.07606

  43. [43]

    Politics and polarization on Bluesky

    Salloum, Ali and Quelle, Dorian and Iannucci, Letizia and Bovet, Alexandre and Kivelä, Mikko , month = jun, year =. Politics and polarization on. doi:10.48550/arXiv.2506.03443 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.03443

  44. [44]

    Proceedings of the International AAAI Conference on Web and Social Media , author =

    Truth. Proceedings of the International AAAI Conference on Web and Social Media , author =. 2023 , pages =. doi:10.1609/icwsm.v17i1.22211 , abstract =

    work page doi:10.1609/icwsm.v17i1.22211 2023

  45. [45]

    Balasubramanian, Ashiwin and Zou, Vito and Narayana, Hitesh and You, Christina and Ferrara, Emilio , year =. A. doi:10.2139/ssrn.5018883 , language =

    work page doi:10.2139/ssrn.5018883

  46. [46]

    Allaway, Emily and McKeown, Kathleen , month = oct, year =. Zero-. doi:10.48550/arXiv.2010.03640 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.03640 2010

  47. [47]

    PLOS ONE , author =

    ". PLOS ONE , author =. 2024 , note =. doi:10.1371/journal.pone.0310330 , abstract =

    doi:10.1371/journal.pone.0310330 2024

  48. [48]

    doi:10.5281/ZENODO.14653265 , abstract =

    Gildersleve, Patrick and Beers, Anna and Ito, Viviane and Orozco, Agustin and Tripodi, Francesca , month = jan, year =. doi:10.5281/ZENODO.14653265 , abstract =

    work page doi:10.5281/zenodo.14653265

  49. [49]

    Shah, Chaitya and Konka, Ritesh and Malpani, Gautam and Mehta, Swapneel and Ng, Lynnette Hui Xian , month = jun, year =. Can. doi:10.48550/arXiv.2406.03354 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.03354

  50. [50]

    IEEE Transactions on Neural Networks and Learning Systems , author =

    A. IEEE Transactions on Neural Networks and Learning Systems , author =. 2021 , pages =. doi:10.1109/TNNLS.2020.2978386 , abstract =

    arXiv 2021

  51. [51]

    2017 , pages =

    Communications of the ACM , author =. 2017 , pages =. doi:10.1145/3065386 , abstract =

    doi:10.1145/3065386 2017

  52. [52]

    Teng, Fengwei and Shi, Quan and Yu, Zhaoyang and Zhang, Jiayi and Luo, Yuyu and Wu, Chenglin and Guo, Zhijiang , month = dec, year =. Atom of. doi:10.48550/arXiv.2502.12018 , abstract =

    doi:10.48550/arxiv.2502.12018

  53. [53]

    Symmetry , author =

    Convolutional. Symmetry , author =. 2024 , pages =. doi:10.3390/sym16111516 , abstract =

    work page doi:10.3390/sym16111516 2024

  54. [54]

    Behaviormetrika , author =

    Can’t see the forest for the trees:. Behaviormetrika , author =. 2024 , pages =. doi:10.1007/s41237-023-00205-2 , abstract =

    work page doi:10.1007/s41237-023-00205-2 2024

  55. [55]

    Chain-of-

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H and Le, Quoc V and Zhou, Denny , file =. Chain-of-

  56. [56]

    Wang, Xuezhi and Wei, Jason and Schuurmans, Dale and Le, Quoc and Chi, Ed and Narang, Sharan and Chowdhery, Aakanksha and Zhou, Denny , month = mar, year =. Self-. doi:10.48550/arXiv.2203.11171 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.11171

  57. [57]

    Baldock, Robert J N and Maennel, Hartmut , file =. Deep

  58. [58]

    Albalak, Alon and Elazar, Yanai and Xie, Sang Michael and Longpre, Shayne and Lambert, Nathan and Wang, Xinyi and Muennighoff, Niklas and Hou, Bairu and Pan, Liangming and Jeong, Haewon and Raffel, Colin and Chang, Shiyu and Hashimoto, Tatsunori and Wang, William Yang , month = aug, year =. A. doi:10.48550/arXiv.2402.16827 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.16827

  59. [59]

    , month = feb, year =

    Han, Song and Mao, Huizi and Dally, William J. , month = feb, year =. Deep. doi:10.48550/arXiv.1510.00149 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1510.00149

  60. [60]

    and Wilson, Andrew Gordon , month = dec, year =

    Stanton, Samuel and Izmailov, Pavel and Kirichenko, Polina and Alemi, Alexander A. and Wilson, Andrew Gordon , month = dec, year =. Does. doi:10.48550/arXiv.2106.05945 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.05945

  61. [61]

    Neural Networks , author =

    Towards dropout training for convolutional neural networks , volume =. Neural Networks , author =. 2015 , pages =. doi:10.1016/j.neunet.2015.07.007 , abstract =

    work page doi:10.1016/j.neunet.2015.07.007 2015

  62. [62]

    Going deeper with convolutions , isbn =

    Szegedy, Christian and. Going deeper with convolutions , isbn =. 2015. 2015 , pages =. doi:10.1109/CVPR.2015.7298594 , abstract =

    arXiv 2015

  63. [63]

    Tsitsulin, Anton and Palowitch, John and Perozzi, Bryan and Müller, Emmanuel , file =. Graph

  64. [64]

    Compressing

    Chen, Wenlin and Wilson, James T and Tyree, Stephen and Weinberger, Kilian Q and Chen, Yixin , file =. Compressing

  65. [65]

    Importance

    Molchanov, Pavlo and Mallya, Arun and Tyree, Stephen and Frosio, Iuri and Kautz, Jan , month = jun, year =. Importance. doi:10.48550/arXiv.1906.10771 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1906.10771 1906

  66. [66]

    Distilling the

    Hinton, Geoffrey and Vinyals, Oriol and Dean, Jeff , month = mar, year =. Distilling the. doi:10.48550/arXiv.1503.02531 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.02531

  67. [67]

    Liebenwein, Lucas and Baykal, Cenk and Carter, Brandon and Gifford, David and Rus, Daniela , month = mar, year =. Lost in. doi:10.48550/arXiv.2103.03014 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2103.03014

  68. [68]

    Language

    Brown, Tom B and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M and Wu, Jeffrey and Win...

  69. [69]

    Frankle, Jonathan and Carbin, Michael , month = mar, year =. The. doi:10.48550/arXiv.1803.03635 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1803.03635

  70. [70]

    Wu, Haibing and Gu, Xiaodong , editor =. Max-. Neural. 2015 , note =. doi:10.1007/978-3-319-26532-2_6 , abstract =

    work page doi:10.1007/978-3-319-26532-2_6 2015

  71. [71]

    Dropout:

    Srivastava, Nitish and Hinton, Geoffrey and Krizhevsky, Alex and Sutskever, Ilya and Salakhutdinov, Ruslan , file =. Dropout:

  72. [72]

    Lin, Zhenghao and Gou, Zhibin and Gong, Yeyun and Liu, Xiao and Shen, Yelong and Xu, Ruochen and Lin, Chen and Yang, Yujiu and Jiao, Jian and Duan, Nan and Chen, Weizhu , file =. Not

  73. [73]

    Fantastically

    Lu, Yao and Bartolo, Max and Moore, Alastair and Riedel, Sebastian and Stenetorp, Pontus , year =. Fantastically. Proceedings of the 60th. doi:10.18653/v1/2022.acl-long.556 , abstract =

    doi:10.18653/v1/2022.acl-long.556 2022

  74. [74]

    Sahoo, Pranab and Singh, Ayush Kumar and Saha, Sriparna and Jain, Vinija and Mondal, Samrat and Chadha, Aman , month = mar, year =. A. doi:10.48550/arXiv.2402.07927 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.07927

  75. [75]

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , month = jun, year =. Deep. 2016. doi:10.1109/CVPR.2016.90 , abstract =

    doi:10.1109/cvpr.2016.90 2016

  76. [76]

    Mulchandani, Varun and Kim, Jung-Eun , year =

  77. [77]

    Lee, Namhoon and Ajanthan, Thalaiyasingam and Torr, Philip H. S. , month = feb, year =. doi:10.48550/arXiv.1810.02340 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1810.02340

  78. [78]

    Li, Hao and Kadav, Asim and Durdanovic, Igor and Samet, Hanan and Graf, Hans Peter , month = mar, year =. Pruning. doi:10.48550/arXiv.1608.08710 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1608.08710

  79. [79]

    Dong, Qingxiu and Li, Lei and Dai, Damai and Zheng, Ce and Ma, Jingyuan and Li, Rui and Xia, Heming and Xu, Jingjing and Wu, Zhiyong and Liu, Tianyu and Chang, Baobao and Sun, Xu and Li, Lei and Sui, Zhifang , month = oct, year =. A. doi:10.48550/arXiv.2301.00234 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.00234

  80. [80]

    Carlini, Nicholas and Jagielski, Matthew and Papernot, Nicolas and Terzis, Andreas and Tramer, Florian and Zhang, Chiyuan , file =. The

Showing first 80 references.