Beyond Text: Multimodal Credibility Assessment Approaches for Online User-Generated Content
Abstract
1 Introduction
![](/cms/10.1145/3673236/asset/6e2c0199-108b-41ac-bc2e-3408f1a14aff/assets/images/medium/tist-2023-10-0694-f01.jpg)
2 Background
![](/cms/10.1145/3673236/asset/89fa3219-1994-4af8-9eb5-c626d0e467a2/assets/images/medium/tist-2023-10-0694-f02.jpg)
![](/cms/10.1145/3673236/asset/bf1a0253-6011-4911-b49c-69fbe69bc97d/assets/images/medium/tist-2023-10-0694-f03.jpg)
![](/cms/10.1145/3673236/asset/8dfed525-5914-4624-8568-8e078a6b8825/assets/images/medium/tist-2023-10-0694-f04.jpg)
2.1 Motivation
2.2 Key Challenges
![](/cms/10.1145/3673236/asset/2cc2803f-266b-4c70-bb76-7a80692cb023/assets/images/medium/tist-2023-10-0694-f05.jpg)
3 Credibility Assessment of UGC
![](/cms/10.1145/3673236/asset/c9d80273-74ab-4708-84c0-a05f3d5d46c7/assets/images/medium/tist-2023-10-0694-f06.jpg)
3.1 Review Methodology
![](/cms/10.1145/3673236/asset/4c828c07-e760-479c-ae4e-cbf81e27a574/assets/images/medium/tist-2023-10-0694-f07.jpg)
![](/cms/10.1145/3673236/asset/8b00df16-56d5-4c76-a0b3-769d392bfb38/assets/images/medium/tist-2023-10-0694-f08.jpg)
![](/cms/10.1145/3673236/asset/f1a7bc62-bbec-4637-be26-ac256c2ac675/assets/images/medium/tist-2023-10-0694-f09.jpg)
4 Content-Level Credibility Assessment
4.1 Content Credibility Assessment Based on Unimodal UGC
4.2 Content Credibility Assessment Based on Multimodal UGC
Author | UM | MM | Modality | Approach | Model | Platform | Observation |
---|---|---|---|---|---|---|---|
Castillo et al. 2011 [28] | ✓ | \(\times\) | Metadata | ML | J48 Decision Tree | Credible news is propagated by users with large message history; message with single/few origins and re-posts. | |
Qazvinian et al. 2011 [90] | ✓ | \(\times\) | Metadata | ML | MLE | Content-, network-, and microblog-specific features are effective. | |
Feng et al. 2012 [89] | ✓ | \(\times\) | Text | ML | Parse trees with SVM | Multiple | Suspicious content identified by deep syntactic patterns. |
Zhang et al. 2012 [84] | ✓ | \(\times\) | Text | ML | SVM | Not specified | Feature selection methods are improved by CHI statistics and hypothesis testing. |
Yang et al. 2012 [82] | ✓ | \(\times\) | Metadata | ML | SVM | User device and location can classify rumors. | |
Briscoe et al. 2013 [26] | ✓ | \(\times\) | Network | ML | Social network graphs | Not specified | Corroboration and degree centrality are indicators of credibility. |
Briscoe et al. 2014 [88] | ✓ | \(\times\) | Text | ML | SVM, gradient boosting | Self-curated | Linguistic cues are present in social media deception. |
Gupta et al. 2014 [85] | ✓ | \(\times\) | Metadata | ML | SVM-Rank | Credibility evaluation models and features change with time. | |
Liu et al. 2015 [35] | ✓ | \(\times\) | Metadata | ML | SVM | ML model with selected features works much faster than manual verification. | |
Ma et al. 2015 [86] | ✓ | \(\times\) | Metadata | ML | SVM | Twitter, Weibo | Time series analysis of features improves rumor detection. |
Zhao et al. 2015 [123] | ✓ | \(\times\) | Text | ML | Decision tree | Tweets asking verification questions and corrections to controversial statements are signals of early rumors detection. | |
Wu et al. 2015 [124] | ✓ | \(\times\) | Metadata | ML, graph | Graph kernel-based hybrid classifier | False rumors have a different re-post pattern. | |
Mukharjee et al. 2015 [29] | ✓ | \(\times\) | Metadata | graph | Continuous metadata random field | Websites | Credibility of articles is language, topic, perspective, and time-line dependent. |
Kumar et al. 2016 [125] | ✓ | \(\times\) | Metadata | ML | Random forest | Wikipedia | Automated classifiers outperform humans by a big margin. |
Rubin et al. 2016 [87] | ✓ | \(\times\) | Text | ML | SVM | Websites | Inter-link between deception and satire, irony and humor. |
Zeng et al. 2016 [36] | ✓ | \(\times\) | Metadata | ML | Regression | Crowd correction is an effective means to prevent misinformation. | |
Wang et al. 2016 [126] | ✓ | \(\times\) | Metadata | ML | Expectation maximization | Jointly estimates the theme awareness and reliability of sources as well as the theme relevance. | |
Jin et al. 2016 [91] | ✓ | \(\times\) | Network | Graph | Iterative deduction | Twitter, Weibo | Conflicting social viewpoints are effective in credibility propagation network for microblogs. |
Rosas et al. 2017 [127] | ✓ | \(\times\) | Text | ML | SVM | News Websites | Linguistic features can give insights into fake and real news. |
Wang et al. 2017 [128] | ✓ | \(\times\) | Network | ML | MLE | Facebook, Twitter Weibo | Different communities have different user interests and different rumor propagation characteristics. |
Ma et al. 2017 [129] | ✓ | \(\times\) | Metadata | ML | Tree kernel | Kernel-based approaches have more higher dimension information carrying capabilities over feature-based methods. | |
Tacchini et al. 2017 [130] | ✓ | \(\times\) | Metadata | Graph | Harmonic Boolean label crowdsourcing | Mapping the diffusion pattern of information can be a useful component of automatic hoax detection systems. | |
Ruchansky et al. 2017 [94] | ✓ | \(\times\) | Metadata | DL, graph | CSI | Twitter, Weibo | Text, article responses, and users are three aspects to identify fake news. |
Yu et al. 2017 [93] | ✓ | \(\times\) | Text | DL | CNN | Twitter, Weibo | CNN can extract scattered key features and shape high-level interactions. |
Jin et al. 2017 [10] | \(\times\) | ✓ | Metadata, text, image | DL | LSTM, VGG 19 | Weibo, Twitter | Attention-based RNN mechanism can detect rumors. |
Wang et al. 2018 [131] | ✓ | \(\times\) | Text | ML | Decision tree | Websites | Claim-relevance discovery can help identify online misinformation. |
Potthast et al. 2018 [132] | ✓ | \(\times\) | Text | ML | Random forest | Websites | Style-based fake news detection may not work effectively. |
Tschiatscheket et al. 2018 [133] | ✓ | \(\times\) | Network | Graph | Bayesian inference | The approach to fake news detection need to learn about user's flagging behavior. | |
Kim et al. 2018 [134] | ✓ | \(\times\) | Metadata | Others | Bayesian inference | Twitter, Weibo | Crowd-powered procedure can reduce the spread of fake news using scalable online algorithms. |
Huh et al.2018 [96] | ✓ | \(\times\) | Metadata | Others | ResNet | Columbia, Carvalho, RT | EXIF metadata can be used as a supervisory signal for training a model to determine whether an image is self-consistent. |
Wang et al. 2018 [135] | ✓ | \(\times\) | Text | DL | CNN | Weibo, Twitter | CNN-based multiscale feature attention can select effective features from text. |
Yang et al. 2019 [136] | ✓ | \(\times\) | Metadata | Graph | Bayesian network with Gibbs sampling | Websites | Users’ engagements on social media can be used to identify their opinions toward the authenticity of the news. |
Khan et al. 2019 [137] | ✓ | \(\times\) | Text | DL | LSTM, VGG 19 | Websites | Accuracy of the model is proportional to article length. |
Reis et al. 2019 [38] | ✓ | \(\times\) | Metadata | ML | Random forest, extreme gradient boosting | Buzzfeed | Language, source, temporal, and engagement features can be combined for better analysis of fake news. |
Shu et al. 2019 [98] | ✓ | \(\times\) | Text | DL | Sentence-comment co-attention network | Gossipcop, Politifact | Credibility of users and user engagement can be explored to enhance model performance. |
Shu et al. 2019 [138] | ✓ | \(\times\) | Metadata | ML, graph | Tri-relationship optimization | Websites | There is a relationship among publisher, news, and social media engagement. |
Khattar et al. 2019 [106] | \(\times\) | ✓ | Text, image | DL | Word2vec, VGG 19 | Twitter, Weibo | Attention mechanism helps to improve the model performance by considering similar parts of image and text. |
Singhal et al. 2019 [109] | \(\times\) | ✓ | Text, image | DL | BERT, VGG 19 | Twitter, Weibo | Multimodal methods can be employed for fake news detection. |
Yang et al. 2019 [139] | \(\times\) | ✓ | Text, image | DL | TI-CNN | Websites | CNN models can be trained much faster than LSTM and many other RNN models. |
Qi et al. 2019 [95] | ✓ | \(\times\) | Image | DL | MVNN | MediaEval, Weibo | The pixel domain and frequency domain information is important for detecting fake news. |
Nguygen et al. 2019 [140] | \(\times\) | ✓ | Text, network | DL, graph | Markov random field | Weibo, Twitter, PHEME | The correlations among news articles are effective cues for online news analysis. |
Palod et al. 2019 [116] | ✓ | \(\times\) | Text | DL | Word2vec, LSTM | FVC, VAVD | Simple features extracted from metadata are not helpful in identifying fake videos. |
Tanwar et al. 2020 [112] | \(\times\) | ✓ | Metadata, text | DL | Word2vec, CNN | MediaEval | More explicit feature based on textual information or user profile data can be explored to improve our accuracy. |
Giachanou et al. 2020 [108] | \(\times\) | ✓ | Text, image | DL | Word2vec, CNN | Politifact, Gossipcop, MediaEval | Combining textual, visual, and text-image similarity information is beneficial for the task of fake news detection. |
Shah et al. 2020 [105] | \(\times\) | ✓ | Text, image | ML | Cultural algorithm with radial kernel function | Twitter, Weibo | Cultural algorithm can extract optimum features from text and images. |
Zhou et al. 2020 [107] | \(\times\) | ✓ | Text, image | DL | SAFE | Politifact, Gossipcop | Multimodal features and the cross-modal relationship (similarity) are essential. |
Xia et a. 2020 [141] | \(\times\) | ✓ | Metadata, network | DL | Encoders | Twitter, Weibo | State independent and time-evolving networks can assist in rumor detection. |
Silva et al. 2021 [110] | \(\times\) | ✓ | Metadata, network | DL, graph | Domain-agnostic news classification | Politifact, Gossipcop, CoAID | Modality of news records (propagation network and text) provides unique knowledge. |
Song et al. 2021 [8] | \(\times\) | ✓ | Text, image | DL | Cross-modal attention residual and multichannel CNN | Twitter, Weibo | Keep the unique properties for each modality while fusing the relevant information between different modalities. |
Wang et al. 2022 [142] | ✓ | \(\times\) | Text | DL, graph | Elementary discourse unit | Websites | Granularity between word and sentence with improved text representation can improve fake news detection. |
Li et al. 2022 [117] | \(\times\) | ✓ | Metadata, text | DL | CNN | Bilibili | DL can be more appropriate over ML for misleading video detection. |
Choi et al. 2022 [118] | \(\times\) | ✓ | Text, video | DL | Fake news video detection model | FVC, VAVD | Domain knowledge is effective in assessing fake news videos. |
Wang et al. 2023 [99] | ✓ | \(\times\) | Text | DL | First-order-logic-guided knowledge | Wikipedia | LLMs can be used to better understand context. |
Chen et al. 2023 [102] | ✓ | \(\times\) | Text | DL | LLMs, humans | LLM-generated data | Detecting LLM-generated misinformation is more challenging than human-written misinformation with similar semantics. |
4.3 Discussion
5 User-Level Credibility Assessment
Social Media Platform | Definition of Fake/Suspicious Profiles |
---|---|
Twittera | Fake accounts are operated to engage in spam, interfere in civic processes, carry out financial scams, artificially inflate engagement, or abuse and harass others. |
Facebookb | Fake profile is a profile where someone is pretending to be something or someone that doesn’t exist. These profiles can include profiles for fake or made-up people, celebrities, or organizations. |
LinkedInc | A profile may be fake if it appears empty or contains profanity, fake names, or impersonates public figures. |
5.1 User Credibility Assessment Based on Unimodal UGC
5.2 User Credibility Assessment Based on Multimodal UGC
Author | UM | MM | Modality | Approach | Model | Platform | Observation |
---|---|---|---|---|---|---|---|
Lee et al. 2011 [159] | ✓ | \(\times\) | Metadata | ML | Random forest | Mis-classified users have a low standard deviation of numerical IDs of following and followers. | |
Zhu et al. 2012 [148] | ✓ | \(\times\) | Metadata | ML | Collective matrix factorization, SVM | Renren | Based on users’ social actions and social relations, spammers can be detected. |
Ahmed et al. 2013 [65] | ✓ | \(\times\) | Metadata | ML | J48 | Facebook, Twitter | Different profile-based features have varied impact on detection capabilities. |
Alowibdi et al. 2014 [149] | ✓ | \(\times\) | Metadata | ML | Bayesian classifier | Combination of multiple profile's characteristics from each Twitter user improves deception detection. | |
Cao et al. 2014 [160] | ✓ | \(\times\) | Metadata | Graph | User similarity graph | Accounts that act similarly at around the same time for a sustained period can be grouped. | |
Ruan et al. 2015 [55] | ✓ | \(\times\) | Metadata | Statistical | Behavioral patterns | Impostors’ social behaviors can hardly conform to the authentic user's behavioral profile. | |
Xiao et al. 2015 [161] | ✓ | \(\times\) | Metadata | ML | Random forest | Basic distribution, pattern, and frequency features can be used to identify a cluster of fake users. | |
Gurajala et al. 2015 [80] | ✓ | \(\times\) | Metadata | ML | Pattern matching | Activity-based profile-pattern detection scheme provides a quick way to identify potential spammers. | |
Tsikerdekis et al. 2016 [146] | ✓ | \(\times\) | Network | Others | Common contribution network | The proposed model can have high computational overhead with more users. | |
Cresci et al. 2016 [71] | ✓ | \(\times\) | Metadata | Pattern Matching | DNA-inspired techniques | DNA-inspired techniques can be used to model user behavior. | |
Singh et al. 2016 [162] | ✓ | \(\times\) | Metadata | ML | Random forest | Behavioral characteristics can be used to differentiate between spammers and genuine users. | |
Zoubi et al. 2017 [163] | ✓ | \(\times\) | Metadata | ML | Naive Bayes and decision tree | Suspicious words and the repeated words greatly influence the detection process's accuracy. | |
Kaur et al. 2018 [143] | ✓ | \(\times\) | Text | ML | K-NN | AHP-TOPSIS method to rank and give appropriate weights to different features for each user improves results. | |
Walt et al. 2018 [150] | ✓ | \(\times\) | Metadata | ML | Random forest | Engineered features are not successful in detecting fake accounts generated by humans. | |
Caruccio et al. 2018 [164] | ✓ | \(\times\) | Metadata | Statistical | Relaxed functional dependencies | Automatic procedures can hardly simulate all human behaviors. | |
Agarwal et al. 2019 [78] | ✓ | \(\times\) | Text | ML | Random forest | Three emotion categories, fear, surprise, and trust, are found least in the posts of fake users. | |
Khaled et al. 2018 [145] | ✓ | \(\times\) | Metadata | ML | SVM-NN | Spearmans Rank correlation technique selects the best features and removes redundancy. | |
Rathore et al. 2018 [75] | ✓ | \(\times\) | Metadata | ML | Bayesian network | Profile- and content-based features enhance spammer detection. | |
Singh et al. 2019 [165] | ✓ | \(\times\) | Metadata | ML | SVM | Malicious users also tend to use spam bots to increase their followers count. | |
Akyon et al. 2019 [73] | ✓ | \(\times\) | Metadata | ML | SVM | Normalization and feature selection algorithms can be used to mitigate bias in the dataset. | |
Zarei et al. 2019 [166] | ✓ | \(\times\) | Metadata | ML | k-Means | Text can be considered to understand what users publish and find a pattern in it. | |
Yuan et al. 2019 [147] | ✓ | \(\times\) | Network | Graph | Graph inference | Account registration information can be used for early sybil detection. | |
Wanda et al. 2020 [152] | \(\times\) | ✓ | Metadata, Network, Text | DL | CNN | Fake accounts should be detected quickly before interacting with real users by capturing informative features from content posted and metadata of profile. | |
Adikari et al. 2020 [53] | ✓ | \(\times\) | Metadata | ML | SVM with polynomial kernel | PCA-based feature selection followed by SVM modeling with the polynomial kernel can be used for identifying fake profiles when a limited number of profile features are publicly available. | |
Breuer et al. 2020 [167] | ✓ | \(\times\) | Network | Graph | SybilEdge | Focusing on interaction of new fake users with other users is insightful. | |
Fazil et al. 2022 [153] | \(\times\) | ✓ | Metadata, Text | DL | LSTM | Using the description text improved the cross-domain performance of the model. | |
Wanda et al. 2022 [168] | ✓ | \(\times\) | Metadata | DL | CNN | Adding the Gaussian function to the non-linear classifier helps achieve better performance. | |
Verma et al. 2022 [169] | \(\times\) | ✓ | Metadata, Text | DL | RoBERTa, BiLSTM, random forest | Using ML, DL, and pretrained models with voting classifier on text and numeric metadata improves model performance. | |
Goyal et al. 2023 [156] | \(\times\) | ✓ | Network, Text, Image | DL, Graph | CNN, LSTM, GCN | Incorporating multimodal data allows for improved detection of bogus accounts. | |
Breuer et al. 2023 [157] | ✓ | \(\times\) | Network | Graph | Preferential attachment k-class classifier | Analyzing the friend request behavior of new users can give insightful information in early detection. | |
Khan et al. 2024 [158] | \(\times\) | ✓ | Metadata, Network, Text | DL, Graph | Graph neural networks | User's profile, content shared, and user–user interaction network provides important cues in identifying malicious users. |
5.3 Discussion
6 Research Gaps
7 Multimodal Datasets
Dataset | Key Modalities | Size | Link for the Dataset |
---|---|---|---|
FakeNews-Kaggle | Text, image, video | 11,941 fake and 8,074 real news | https://www.kaggle.com/competitions/fake-news/data |
FakeNewsNet [172] | Text, image, spatial, temporal, user, network | 17,441 real and 5,755 fake news articles with user activity data | https://github.com/KaiDMML/FakeNewsNet |
Fakeddit [173] | Text, image, and other metadata of post | 628,501 fake and 527,049 real posts | https://github.com/entitize/fakeddit |
Breaking News dataset [174] | Text, image | 100,000 news articles | https://www.iri.upc.edu/people/aramisa/BreakingNews/ |
Tampered News dataset | Text, image | 72K news articles | https://data.uni-hannover.de/dataset/tamperednews |
News400 dataset [175] | Text, image | 400 news articles | https://github.com/TIBHannover/cross-modal_entity_consistency |
Twitter dataset [176] (MediaEval) | Text, image | Around 6,000 real and 9,000 fake posts | https://github.com/MKLab-ITI/image-verification-corpus |
Weibo dataset [10] | Text, image, social context | Around 50,000 tweets | https://github.com/yaqingwang/EANN-KDD18/tree/master/data/weibo |
NewsBag, NewsBag++ [177] | Text, image | 200K real and 15K fake news articles (NewsBag) 200K real and 389K fake news articles (NewsBag++) | Available on request |
Multimodal Entity Image Repurposing [178] | Text, image, spatial, temporal | 57,940 packages | https://github.com/Ekraam/MEIR/tree/master/dataset |
Youtube videos dataset [179] | Video and audio | 121 video clips (61 deceptive, 60 truthful) | Available on request |
WIT: Wikipedia- based image text dataset [180] | Text, image | 37.6 million image-text sets and 11.5 million unique images | https://github.com/google-research-datasets/wit |
VisualNews dataset [181] | Text, image, and other metadata | Over 1 million images and more than 600,000 articles | https://github.com/FuxiaoLiu/VisualNews-Repository Available on request |
NewsCLIPpings dataset [182] | Text, image, metadata | Over 986K unique images with captions | https://github.com/g-luo/news_clippingsdata-format Available on request |
COSMOS dataset [183] | Text, image | Over 200K images with 450K corresponding text captions | https://shivangi-aneja.github.io/projects/cosmos/ Available on request |
COVID-VTS [184] | Text, video, speech | 10K video-claim pairs | https://drive.google.com/drive/folders/1xT4QaOZQlZtW9Ul36VCJ4arZQ94-Ok3V |
FakeSV [185] | Text, video, audio, metadata | Over 3.6K videos in Chinese with metadata | https://github.com/ICTMCG/FakeSV |
FVC [186] | Text, video, audio, metadata | 200 fake and 180 real videos | https://github.com/MKLab-ITI/fake-video-corpus |
YouTubeAudit-data [187] | Video, title (text), metadata | 56K videos, covers 5 popular misinformative topics | https://social-comp.github.io/YouTubeAudit-data/ |
ReCOVery dataset [188] | Text, image, metadata | Over 1.7K multimodal news articles; 93K users sharing 140K tweets | https://github.com/apurvamulay/ReCOVery |
8 Scope for Future Work
9 Conclusion
Footnote
References
Index Terms
- Beyond Text: Multimodal Credibility Assessment Approaches for Online User-Generated Content
Recommendations
Exploring the user-generated content (UGC) uploading behavior on youtube
WWW '14 Companion: Proceedings of the 23rd International Conference on World Wide WebYouTube is the world's largest video sharing platform where both professional and non-professional users participate in creating, uploading, and viewing content. In this work, we analyze content in the music category created by the non-professionals, ...
User generated content and credibility evaluation of online health information
This meta-analysis addresses credibility concerns for online health information.A collection of empirical studies addressing user-generated content was analyzed.We synthesized 22 effect sizes drawn from empirical studies of 1346 participants.Source ...
Comments
Information & Contributors
Information
Published In
![cover image ACM Transactions on Intelligent Systems and Technology](/cms/asset/fe2abc04-2236-46c1-a48a-b3e3e8853efc/3613688.cover.jpg)
Publisher
Association for Computing Machinery
New York, NY, United States
Publication History
Check for updates
Author Tags
Qualifiers
- Research-article
Contributors
Other Metrics
Bibliometrics & Citations
Bibliometrics
Article Metrics
- 0Total Citations
- 559Total Downloads
- Downloads (Last 12 months)559
- Downloads (Last 6 weeks)76
Other Metrics
Citations
View Options
Login options
Check if you have access through your login credentials or your institution to get full access on this article.
Sign in