ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models

Anna Wr\'oblewska; Mateusz Kubita; Soveatin Kuntur; Tymoteusz Urban; Wojciech Michaluk

REVIEW 4 major objections 5 minor 19 references

ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A hybrid model that pairs 1,000-dimension text embeddings with 115 handcrafted style measures claims 0.909 F1 for clickbait detection.

desk verdict A practical clickbait-detection extension with a plausible but unverifiable 0.909 F1; the undocumented data split and undefined features make it impossible to audit as written. read the letter →

arxiv 2607.20463 v1 pith:MZPLAICQ submitted 2026-05-18 cs.AI

Wojciech Michaluk , Tymoteusz Urban , Mateusz Kubita , Soveatin Kuntur , Anna Wróblewska This is my paper

classification cs.AI

keywords clickbaitdetectionbrowserextensionhybridfeaturesinformativenessmeasuresbaitnessscoretextembeddingscontentspoilinginterpretability

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

This paper tries to establish that clickbait detection can be both accurate and practical enough for a real-time browser extension. It claims that fusing compact semantic embeddings with cheap linguistic informativeness measures—readability, sentiment, punctuation patterns, stop-word ratios, and a custom 'baitness' score—lets a gradient-boosted tree classifier reach an F1 of 0.909 on a balanced combined dataset of 40,000 headlines. This outperforms fine-tuned language models (0.842) and pure embeddings (0.864), while remaining fast enough for pre-click warnings and post-click explanations. A sympathetic reader would care because the system goes beyond black-box classification: it gives users an interpretable probability score and a generated spoiler that satisfies curiosity without opening the article.

What carries the argument

The load-bearing mechanism is the hybrid feature vector: text embeddings (projected from 3,072 to 1,000 dimensions to cut cost without losing predictive power) concatenated with 15 linguistically motivated informativeness measures. Central among these is the custom 'baitness' score, a composite that aggregates eye-catching cues (capitalization, numerals, second-person pronouns, exaggerated punctuation) and curiosity-inducing patterns, and which the authors say separates classes effectively. The combined vector is fed to a gradient-boosted decision tree, which learns to weigh deep semantic similarity against surface-level rhetorical tricks.

What would settle it

Deduplicate the combined dataset by fuzzy string similarity, then re-run the same 1,000-dimensional embedding plus 115-feature gradient-boosted pipeline on a fresh sample of headlines from the same news domains; if the F1 drops substantially below 0.909, train/test leakage is inflating the reported performance.

Watch

Extended reading notes

Core claim

The paper presents ClickGuard, a client–server system whose detection core is a tree-based classifier trained on the concatenation of 1,000-dimensional dense text embeddings and 115 handcrafted informativeness measures. The authors claim the classifier reaches a test F1 of 0.909 on a balanced subset of 20,000 clickbait and 20,000 non-clickbait headlines drawn from open datasets, with a confusion matrix showing near-symmetric precision and recall. They further claim the hybrid representation beats both a fine-tuned transformer baseline and an embeddings-only model by roughly 4.5 percentage points, and that the resulting speed/accuracy tradeoff enables a browser extension to score headlines on

Load-bearing premise

The reported F1 rests on the assumption that the balanced dataset’s train/test split contains no near-duplicate headlines across its three pooled sources, so the test set measures generalization to truly new headlines rather than recall of memorized variants.

Editorial extensions

If this is right

If the reported F1 holds, a lightweight classifier can score headlines before a click, making real-time browser protection feasible without a large language model call per page.
Post-click spoilers generated by a compact generative model can close the curiosity gap, reducing the reward for clicking manipulative content.
The style-based feature set enables interpretable warnings: users see which cues (superlatives, punctuation, readability) drove the score, unlike black-box filters.
The measured 4.5-point gain from adding handcrafted features to embeddings demonstrates that lexical and rhetorical signals carry information that pure semantics misses.

Reading between the lines

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

The 'baitness' composite score could be reused as a standalone, auditable risk indicator by other content-moderation or news-reader tools, independent of the full classifier.
The feature set is plausibly portable to other languages with modest changes to stop-word lists and punctuation conventions, though the paper does not test this explicitly.
A likely real-world failure mode is distribution shift: the model was tuned on news headlines, so performance on social media posts, video titles, or non-English text may be lower; this is testable with fresh samples.

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 an applied systems paper, not a methods advance. What's genuinely new is the integration of a pre-click warning badge and a post-click spoiler into one browser extension, plus an empirical F1 of 0.909 for a hybrid XGBoost model on a combined clickbait dataset. The confusion matrix arithmetic checks out, and the baseline comparison is useful. That said, the paper cannot be accepted as-is. The biggest issue is that the central result rests on an undocumented data split: no deduplication, no stratification by source, no stated binarization of Clickbait Challenge 2017's continuous labels, no seed. With ~40k headlines from three different sources, near-duplicate surface forms could easily leak across train/test and inflate the F1. The stress-test note is right that this is load-bearing.

Beyond that, the 15-vs-115 feature count appears in consecutive paragraphs; the 'baitness' score is named as a key contribution but never defined; and no code, data, or split is released. These are fixable reportability problems, but they are exactly the kind of thing that makes the headline claim impossible to audit. The spoiling module is presented as a contribution but has no evaluation at all — it's a demo with GPT-4o-mini, not a system with measurements.

What's good: the system architecture is sensible, the authors compare against reasonable baselines (TF-IDF, Word2Vec, RoBERTa, pure OpenAI embeddings), and the 4.5-point gain from adding linguistic features to embeddings is a concrete, somewhat interesting empirical observation — though the recipe itself is standard and cited as such. The demo video and the interpretability angle are nice touches for a practitioner audience.

Who is this for? Someone building a deployable clickbait tool might find the integration ideas worth borrowing. A researcher looking for methodological novelty or rigorous evaluation will be disappointed. I agree with the reader's conditional verdict: every flaw is concrete and addressable, but they need to be addressed before the paper can be trusted.

Recommendation: send to peer review, but only with the expectation of a major revision. Require the authors to document the data split and label binarization, define the informativeness measures and baitness score, and release code or at least provide error bars. Without those, it's a demo, not a paper.

Referee Report

4 major / 5 minor

Summary. The paper presents ClickGuard, a browser extension for clickbait detection and spoiling. Detection is performed by an XGBoost classifier trained on a concatenation of OpenAI text-embedding-3-large vectors projected to 1,000 dimensions and handcrafted "informativeness" features, including a custom "baitness" score. On a balanced combined dataset of roughly 40,000 headlines from Kaggle and the Clickbait Challenge 2017, the system reports F1 = 0.909, outperforming TF-IDF, Word2Vec, fine-tuned RoBERTa, and embeddings-only XGBoost baselines. The extension provides pre-click risk badges, post-click probability scores with feature-based explanations, and a GPT-4o-mini-generated spoiler. The central contribution is a practical, hybrid, low-latency detection pipeline combined with a user-facing spoiling mechanism.

Significance. If the reported result is reproducible, the paper demonstrates that a hybrid XGBoost model combining projected LLM embeddings with cheap linguistic features can exceed pure transformer/embedding baselines by roughly 4.5 percentage points while remaining light enough for a browser extension. This is a useful practical result for real-time clickbait filtering, and the confusion matrix in Table 1 is arithmetically consistent with the claimed F1. The evaluation is a held-out classifier measurement rather than a circular derivation, which is a genuine strength. The significance, however, hinges on experimental transparency: the data-split protocol, feature definitions, baitness formula, and embedding-projection procedure are underspecified, so the headline number cannot currently be independently verified.

major comments (4)

[§4 'Training Dataset' and Table 1] The reported F1=0.909 cannot be verified or interpreted without a precise evaluation protocol. The paper states a balanced subset of 20,000 clickbait and 20,000 non-clickbait headlines, but does not specify (i) how Clickbait Challenge continuous probability labels were binarized, (ii) whether near-duplicates across Kaggle and Clickbait Challenge were removed, (iii) whether the split was stratified by source, or (iv) the random seed. Table 1 sums to 4,000, implying a 10% test split, yet the test size is never stated. If duplicates appear in both train and test, XGBoost can memorize surface forms and F1 is an optimistic upper bound; without source stratification, the baselines in Table 2 are not guaranteed to be compared on the same distribution. Please specify these details and add robustness analyses.
[§1/§4.1 vs §4.1.1/Table 2] The number of informativeness features is internally contradictory: the abstract, introduction, and Table 2 say '15', while §4.1.1 says the XGBoost model was trained on '115 informativeness measures'. This is not a harmless typo, because the input representation is defined as these features plus 1000-d embeddings; without knowing whether the feature vector has 15 or 115 dimensions, the model cannot be reproduced and the claimed 4.5-point gain over embeddings-only models cannot be re-run. Please reconcile the count and provide the complete feature list.
[§4.1 'Baitness' score] The custom 'baitness' measure is presented as a key component of the hybrid model and of the user-facing explanation, but no formula, weights, or algorithm are given, and the cited reference [4] appears unrelated. The statement that it 'proved effective at separating classes' is therefore unsupported. A precise definition (or a pointer to a release with code) is required for reproducibility and for assessing the interpretability claims.
[§4.1 'Deep Semantic Embeddings'] The projection from 3,072 to 1,000 dimensions is mentioned only as an ablation outcome; no method (e.g., PCA vs linear layer), no fitting procedure, and no statement that the projection was fit exclusively on the training split are provided. If the projection or XGBoost hyperparameters were chosen with access to the test split, the reported F1 is not an unbiased estimate. Please document the projection and confirm that no test-set information was used in any model selection step.

minor comments (5)

[Abstract] The demo video link contains a duplicated LaTeX artifact ('https://...}{https://...') and should be cleaned.
[Throughout] 'One implemented/used/developed' is nonstandard; use 'we' for clarity.
[Figure 2 caption] Missing space in 'architecture ofClickGuard'.
[References] Reference [17] is cited as a ResearchGate link with an unusual title; please use a citable formal version if one exists.
[§4.2] The spoiling module is described but not evaluated; a short human or automatic evaluation of spoiler usefulness would support the system-level claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the F1=0.909 claim is an empirical held-out classifier measurement, not a derivation that reduces to its inputs.

full rationale

The paper's central result is an experimental evaluation: an XGBoost model is trained on concatenated embeddings and handcrafted linguistic features and then evaluated on a test split of public datasets. This is a standard supervised-learning benchmark, not a derivation chain in which a predicted quantity is defined in terms of the fitted parameters or in which a fitted parameter is renamed as a prediction. The custom 'baitness' measure is an input feature computed from headline text, not a quantity that is itself derived from the model's output or from the labels, so there is no self-definitional loop. The paper cites external prior work for features and baselines, and no load-bearing claim depends on a self-citation. Possible experimental weaknesses—unstated binarization of Clickbait Challenge labels, lack of deduplication, single unseeded split, potential train/test leakage—would affect the validity or generalizability of the reported F1, but they are not circularity in the sense of a claimed derivation being equivalent to its inputs by construction. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on standard supervised-learning machinery plus several hand-tuned choices whose values are unreported (baitness weights, XGBoost parameters, split details) and four domain assumptions about label quality, embedding stability, feature informativeness, and split integrity.

free parameters (5)

embedding_dimension_after_projection = 1000 (from 3072)
Selected via an ablation on the same combined dataset; the paper states 1000 dims 'maintained predictive performance' (§4.1). This selection used the target dataset, so it is a fitted hyperparameter.
baitness_score_weights = not specified
The custom Baitness composite aggregates capitalization, numerals, and curiosity patterns with unspecified weights/formula; 'proved effective at separating classes' (§4.1). Its calibration is a free parameter that cannot be checked.
class_balance_target = 20,000 clickbait / 20,000 non-clickbait
Balanced subset size chosen by hand 'to prevent classifier bias' (§4, Training Dataset); affects precision-recall trade-off and the reported F1.
xgboost_hyperparameters = not reported
Learning rate, max depth, tree count, etc. are not listed; these are fitted on the training split and materially affect F1.
train_test_split = single 90/10 split; seed not reported
The confusion matrix sums to 4,000 for a 40,000-sample balanced subset, implying a 10% test split; no stratification, deduplication, or seed is described.

assumptions (4)

domain assumption Labels of the three combined public datasets are correct, mutually consistent, and directly comparable (binary clickbait labels from two Kaggle sets and Clickbait Challenge 2017).
Invoked in §4 (Training Dataset); the paper does not audit label distributions, annotator agreement, or how the Challenge probability scores were binarized. A mismatch would bias F1.
domain assumption OpenAI text-embedding-3-large is a fixed, reproducible feature extractor across experiments and at deployment time.
Invoked in §4.1; proprietary API embeddings can change on the provider's schedule, and the claim relies on the 1000-dim projection being stable.
domain assumption The 15 handcrafted informativeness features (FRES, polarity, subjectivity, stop-word ratio, bait punctuation, baitness) capture stylistically meaningful clickbait signal not present in the embeddings.
Asserted in §4.1 and the Abstract; the paper gives examples but no full enumeration or theoretical/empirical grounding.
domain assumption A single random train/test split faithfully represents out-of-distribution real-world headline distributions.
Invoked by the F1 claim in §4.1.1; no cross-validation, no per-source holdout, and no deduplication of near-identical headlines across the combined datasets.

invented entities (1)

Baitness score
purpose: A composite informativeness measure for the hybrid classifier's feature vector and for post-hoc explanations; aggregates eye-catchingness (capitalization, numerals) and curiosity-inducing patterns.
Introduced in §4.1 with no formula, weights, or external validation; its claimed class-separation power is asserted but not shown. As a new construct it carries no falsifiable handle outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models." pith.science (2026). https://pith.science/paper/MZPLAICQ

@misc{pith2026260720463,
  author       = {Pith},
  title        = {Pith review of: ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MZPLAICQ}},
  note         = {Machine review of arXiv:2607.20463}
}

read the original abstract

This paper presents an AI-driven browser extension that identifies clickbait to help users avoid misleading Internet articles. Moving beyond traditional detection, the application employs a hybrid machine learning architecture that combines transformer-based embeddings with linguistically motivated features and a custom "baitness" score. After evaluating various natural language processing techniques -- from classic vectorizers to large language model (LLM) embeddings -- an XGBoost-based model was developed that achieves an F1-score of 91% on the open combined dataset. Most importantly, the tool can warn users before and after they access a clickbait article. After opening an article, the user receives a percentage score indicating the likelihood that it is clickbait. The prediction is explained based on the analyzed metrics, including those specifically developed within the proposed system. The browser extension also provides a clickbait spoiler -- a one- to two-sentence summary of the entire article. Demo video:https://www.youtube.com/watch?v=IJ1gkQV82C4}{https://www.youtube.com/watch?v=IJ1gkQV82C4

Figures

Figures reproduced from arXiv: 2607.20463 by the authors.

**Figure 2.** Client–server architecture of ClickGuard, consisting of the browser extension and a Python REST API backend. Our goal was to address this challenge by filling the gap between theoretical NLP advancements and practical, usercentric tools. We developed ClickGuard, a browser extension that not only detects clickbait but also informs Internet users about the potential threat of clicking a clickbait article. It is worth… view at source ↗

**Figure 3.** ClickGuard operating in pre-click mode. Icons next to headlines represent the clickbait probability calculated by a hybrid XGBoost model [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: ClickGuard in post-click mode with clickbait spoiling (from 3,072) maintained predictive performance while significantly reducing computational overhead [18]. Informativeness Measures: To capture stylistic nuances not always preserved in semantic embeddings, one comput…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 4 canonical work pages

[4]

Multimodal clickbait detection by de-confounding biases using causal representation inference

Jianxing Yu, Shiqi Wang, Han Yin, Zhenlong Sun, Ruobing Xie, Bo Zhang, and Yanghui Rao. Multimodal clickbait detection by de-confounding biases using causal representation inference. https://arxiv.org/abs/ 2410.07673, 2024. arXiv preprint, accessed 29 October 2024

arXiv 2024
[1]

You won’t believe what’s in this paper! clickbait, relevance and the curiosity gap.Journal of Pragmatics, 175:53–66, 2021

Kate Scott. You won’t believe what’s in this paper! clickbait, relevance and the curiosity gap.Journal of Pragmatics, 175:53–66, 2021

2021
[2]

Multi-modal soft prompt-tuning for chinese clickbait detection.Neurocomputing, 614:128829, 2025

Ye Wang, Yi Zhu, Yun Li, Liting Wei, Yunhao Yuan, and Jipeng Qiang. Multi-modal soft prompt-tuning for chinese clickbait detection.Neurocomputing, 614:128829, 2025

2025
[3]

A multimodal ensemble-based framework for detecting fake news using visual and textual features.Mathematics, 14(2):360, 2026

Muhammad Abdullah, Hongying Zan, Arifa Javed, Muhammad Sohail, Orken Mamyrbayev, Zhanibek Turysbek, Hassan Eshkiki, and Fabio Caraffini. A multimodal ensemble-based framework for detecting fake news using visual and textual features.Mathematics, 14(2):360, 2026

2026
[5]

RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, July 2019

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, July 2019

arXiv 1907
[6]

Clickbait spoiling via question answering and passage retrieval.https://arxiv.org/abs/2203.10282, 2022

Matthias Hagen, Maik Fröbe, Artur Jurk, and Martin Potthast. Clickbait spoiling via question answering and passage retrieval.https://arxiv.org/abs/2203.10282, 2022. arXiv preprint, accessed 29 October 2024

arXiv 2022
[7]

Summary of ChatGPT-related research and perspective towards the future of large language models (LLMs).Meta-Radiology, 1(2):100017, September 2023

Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, Zihao Wu, Lin Zhao, Dajiang Zhu, Xiang Li, Ning Qiang, Dingang Shen, Tianming Liu, and Bao Ge. Summary of ChatGPT-related research and perspective towards the future of large language models (LLMs).Meta-Radiology, 1(2):100017, Sept...

2023
[8]

Automatic detection of clickbait headlines using semantic analysis and machine learning techniques.Applied Sciences, 13(4), January 2023

Mark Bronakowski, Mahmood Al-khassaweneh, and Ali Al Bataineh. Automatic detection of clickbait headlines using semantic analysis and machine learning techniques.Applied Sciences, 13(4), January 2023. Accessed: 30 October 2024. DOI:https://doi.org/10.3390/app13042456

work page doi:10.3390/app13042456 2023

Show all 19 references

[9]

Clickbait detection using deep recurrent neural network.Applied Sciences, 12(1), 2022

Abdul Razaque, Bandar Alotaibi, Munif Alotaibi, Shujaat Hussain, Aziz Alotaibi, and Vladimir Jotsov. Clickbait detection using deep recurrent neural network.Applied Sciences, 12(1), 2022. Accessed: 29 October 2024. DOI: https://doi.org/10.3390/app12010504

2022 doi
[10]

Identifying clickbait in online news using deep learning

Andry Chowanda, Nadia Nadia, and Lie Kolbe. Identifying clickbait in online news using deep learning. Bulletin of Electrical Engineering and Informatics, 12:1755–1761, 06 2023. Accessed: 29 October 2024. DOI: 10.11591/eei.v12i3.4444

2023 doi
[11]

CliNe AI chrome extension

Janakar Patel and Abhi Prajapati. CliNe AI chrome extension. https://github.com/Janakarpatel/CliNe. AI_Chrome-Extension, 2023

2023
[12]

Clickbait spoiling via question answering and passage retrieval.arXiv, 03 2022

Matthias Hagen, Maik Fröbe, Artur Jurk, and Martin Potthast. Clickbait spoiling via question answering and passage retrieval.arXiv, 03 2022. Accessed: 29 October 2024. DOI: https://doi.org/10.48550/arXiv. 2203.10282

2022 doi arXiv
[13]

Clickbait dataset

Aman Anand. Clickbait dataset. https://www.kaggle.com/datasets/amananandrai/ clickbait-dataset/, 2019. Accessed: 28 October 2024

2019
[14]

News clickbait dataset

Vikas Singh. News clickbait dataset. https://www.kaggle.com/datasets/vikassingh1996/ news-clickbait-dataset, 2020. Accessed: 28 October 2024

2020
[15]

Clickbait detection challenge 2017

Webis. Clickbait detection challenge 2017. https://webis.de/events/clickbait-challenge/ shared-task.html, 2017. Accessed: 28 October 2024

2017
[16]

Clickbait challenge at SemEval 2023 — clickbait spoiling

PAN Webis. Clickbait challenge at SemEval 2023 — clickbait spoiling. https://pan.webis.de/semeval23/ pan23-web/clickbait-challenge.html, 2023

2023
[17]

Using random forest to learn imbalanced data

Chao Chen and Leo Breiman. Using random forest to learn imbalanced data. https://www.researchgate. net/publication/254196943_Using_Random_Forest_to_Learn_Imbalanced_Data, 01 2004. Accessed: 30 October 2024

2004
[18]

Sentence-BERT: Sentence embeddings using siamese BERT-networks

Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019. Accessed: 30 October 2024.https://arx...

2019 arXiv
[19]

Schweidel, and Daniel Dan

Martin Reisenbichler, Thomas Reutterer, David A. Schweidel, and Daniel Dan. Frontiers: Supporting content marketing with natural language generation.Marketing Science, 41(3):441–452, 2022. 6

2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [4]

Multimodal clickbait detection by de-confounding biases using causal representation inference

Jianxing Yu, Shiqi Wang, Han Yin, Zhenlong Sun, Ruobing Xie, Bo Zhang, and Yanghui Rao. Multimodal clickbait detection by de-confounding biases using causal representation inference. https://arxiv.org/abs/ 2410.07673, 2024. arXiv preprint, accessed 29 October 2024

arXiv 2024

[2] [1]

You won’t believe what’s in this paper! clickbait, relevance and the curiosity gap.Journal of Pragmatics, 175:53–66, 2021

Kate Scott. You won’t believe what’s in this paper! clickbait, relevance and the curiosity gap.Journal of Pragmatics, 175:53–66, 2021

2021

[3] [2]

Multi-modal soft prompt-tuning for chinese clickbait detection.Neurocomputing, 614:128829, 2025

Ye Wang, Yi Zhu, Yun Li, Liting Wei, Yunhao Yuan, and Jipeng Qiang. Multi-modal soft prompt-tuning for chinese clickbait detection.Neurocomputing, 614:128829, 2025

2025

[4] [3]

A multimodal ensemble-based framework for detecting fake news using visual and textual features.Mathematics, 14(2):360, 2026

Muhammad Abdullah, Hongying Zan, Arifa Javed, Muhammad Sohail, Orken Mamyrbayev, Zhanibek Turysbek, Hassan Eshkiki, and Fabio Caraffini. A multimodal ensemble-based framework for detecting fake news using visual and textual features.Mathematics, 14(2):360, 2026

2026

[5] [5]

RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, July 2019

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, July 2019

arXiv 1907

[6] [6]

Clickbait spoiling via question answering and passage retrieval.https://arxiv.org/abs/2203.10282, 2022

Matthias Hagen, Maik Fröbe, Artur Jurk, and Martin Potthast. Clickbait spoiling via question answering and passage retrieval.https://arxiv.org/abs/2203.10282, 2022. arXiv preprint, accessed 29 October 2024

arXiv 2022

[7] [7]

Summary of ChatGPT-related research and perspective towards the future of large language models (LLMs).Meta-Radiology, 1(2):100017, September 2023

Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, Zihao Wu, Lin Zhao, Dajiang Zhu, Xiang Li, Ning Qiang, Dingang Shen, Tianming Liu, and Bao Ge. Summary of ChatGPT-related research and perspective towards the future of large language models (LLMs).Meta-Radiology, 1(2):100017, Sept...

2023

[8] [8]

Automatic detection of clickbait headlines using semantic analysis and machine learning techniques.Applied Sciences, 13(4), January 2023

Mark Bronakowski, Mahmood Al-khassaweneh, and Ali Al Bataineh. Automatic detection of clickbait headlines using semantic analysis and machine learning techniques.Applied Sciences, 13(4), January 2023. Accessed: 30 October 2024. DOI:https://doi.org/10.3390/app13042456

work page doi:10.3390/app13042456 2023