Source-linked AI summary
Detection and Resolution of Rumours in Social Media: A Survey
Arkaitz Zubiaga, Ahmet Aker, Kalina Bontcheva, Maria Liakata, Rob Procter
TL;DR
Social media rumours are difficult to resolve because research has often focused on veracity while assuming rumours are already identified. This survey synthesizes approaches across four rumour-classification components and provides a guide to the field’s state of the art, while noting that early detection and tracking remain limited.
Problem
Research has largely focused on rumour veracity while often assuming that rumours have already been identified, leaving earlier detection and tracking underdeveloped.
Method
The survey reviews scientific approaches for rumour detection, tracking, stance classification, and veracity classification in social media.
Results
The survey provides a guide to the state of the art in developing social-media rumour classification systems.
Takeaways & Limitations
Future research should prioritize rumour detection and tracking to reduce reliance on human-provided rumours and associated posts.
Takeaways & Limitations
Rumour tracking remains limited, and existing approaches often assume that keywords for collecting associated posts are known beforehand.
Abstract
from arXiv · showhide
Despite the increasing use of social media platforms for information and news gathering, its unmoderated nature often leads to the emergence and spread of rumours, i.e. pieces of information that are unverified at the time of posting. At the same time, the openness of social media platforms provides opportunities to study how users share and discuss rumours, and to explore how natural language processing and data mining techniques may be used to find ways of determining their veracity. In this survey we introduce and discuss two types of rumours that circulate on social media; long-standing rumours that circulate for long periods of time, and newly-emerging rumours spawned during fast-paced events such as breaking news, where reports are released piecemeal and often with an unverified status in their early stages. We provide an overview of research into social media rumours with the ultimate goal of developing a rumour classification system that consists of four components: rumour detection, rumour tracking, rumour stance classification and rumour veracity classification. We delve into the approaches presented in the scientific literature for the development of each of these four components. We summarise the efforts and achievements so far towards the development of rumour classification systems and conclude with suggestions for avenues for future research in social media mining for detection and resolution of rumours.
1 INTRODUCTION
Social media enables real-time information sharing and large-scale study of rumours, but its openness and lack of moderation also facilitate unverified information. This survey therefore examines rumour definitions, types, diffusion, and classification challenges.
- Motivation: Social media supports real-time information gathering during societal issues and breaking news because anyone with an internet-connected device can share updates.Its real-time, open participation also enables researchers to observe how information spreads and is discussed.
- Rumour Definition: A rumour is information whose truth value remains unresolved because it lacks supporting evidence or official confirmation from credible sources.The literature varies, with some studies incorrectly treating rumours as false information rather than unverified information.
- Rumour Types: The survey distinguishes newly emerging breaking-news rumours from long-standing rumours that circulate without their veracity being established.Breaking-news rumours require real-time processing and handling of unseen cases, whereas long-standing rumours may sustain ongoing interest despite unresolved truth.
- Rumours in Social Media: 44:1, 18:1, and 5:1 were the reported ratios of tweets supporting false rumours, indicating substantial support for inaccurate information in three examined cases.These ratios were reported as favouring tweets supporting false rumours.
- Rumours in Social Media: Users generally support unverified rumours early in breaking-news events, then increasingly support true rumours and debunk false rumours as time passes.This temporal shift was observed across rumours circulating during nine breaking-news events.
- Survey Scope: The survey is motivated by the need to improve information quality assurance because social media is open, lightly moderated, and easy to use for posting information.It proceeds from social-media opportunities and rumour datasets toward analysis of rumour classification systems.
2 SOCIAL MEDIA AS AN INFORMATION SOURCE: CHALLENGES POSED BY RUMOURS
Social media has become an important information source for news, emergencies, public opinion, and financial markets, but its potential for spreading misinformation and unsubstantiated claims creates challenges involving both long-standing and emerging rumours.
- Information-source uses: Social media supports information gathering across news, emergencies, public opinion, and financial markets.Applications include breaking-news diffusion, eyewitness reporting, crisis coordination, opinion measurement, and analysis of financial information.
- Challenges posed by rumours: Its increasing use as an information source has prompted studies of misinformation, unsubstantiated claims, credibility perceptions, and reliance on social media for news.These concerns arise from the coexistence of information-gathering potential with rumours and questionable claims.
- Dealing with Long-Standing Rumours: Long-standing rumours are known a priori and can be tracked while mining social media for opinions over extended periods.Examples include rumours about potential buyouts discussed for long periods in the financial domain.
- Dealing with Emerging Rumours: Emerging rumours arise suddenly during tracked events or topics, when information is released piecemeal and requires verification.This use case applies to news gathering, emergencies, and rumours anticipating political decisions affecting stock markets.
3 DATA COLLECTION AND ANNOTATION
This section reviews how social-media data and annotations are collected for rumour research, emphasizing platform APIs, collection constraints, sampling strategies, and task-dependent annotation. It distinguishes long-standing from emerging rumours and top-down from bottom-up sampling approaches.
- Data access through APIs: Social-media data collection generally uses platform APIs, whose documentation, access methods, and limitations must be understood before developing rumour classification systems.Twitter, Sina Weibo, and Facebook differ in API access, real-time capabilities, and content availability.
- Platform features and limitations: Twitter offers REST and streaming APIs, while Sina Weibo lacks an official streaming API and restricts some methods, including search access.Twitter’s most generous endpoint provides a randomly sampled 1% of the whole dataset; Sina Weibo methods are accessible only through its REST API.
- Platform features and limitations: Platform Terms of Service can restrict public dataset release, often permitting only unique content identifiers such as tweet IDs rather than raw data.Because these terms change frequently, they should be checked before releasing a dataset.
- Rumour collection strategies: Long-standing rumours can be collected using predefined keywords, whereas emerging rumours require broader real-time collection because their keywords are unknown beforehand.Keyword collection supports large-scale datasets and long-term opinion tracking, while emerging-rumour collection must capture relevant posts before or as the rumour occurs.
- Sampling strategies: Broad collections can use top-down sampling of known rumours or bottom-up sampling to discover rumour-related posts from events or unfiltered streams.Top-down sampling was prevalent in early rumour research, while bottom-up strategies emerged more recently for discovering newly emerging rumours.
- Annotation: Rumour annotation occurs at different levels depending on the task, beginning with identifying the rumourous subset within collected data.Top-down sampling may already isolate rumourous data, whereas bottom-up sampling requires further annotation to distinguish rumours from non-rumours.
4 CHARACTERISING RUMOURS: UNDERSTANDING RUMOUR DIFFUSION AND FEATURES
Research on rumour characterisation examines how rumours emerge, diffuse, evolve and elicit reactions, providing insights for rumour classification systems. Findings highlight strong-tie diffusion, temporal popularity changes and factors influencing tracking interest and propagation.
- Rumour discourse and reactions: Studies examine rumour discussions, their evolution over time and behavioural reactions expressed toward them.Research ranges from analyses of individual rumours to broader studies of larger rumour sets.
- Rumour discourse and reactions: Social media users commonly support and spread rumours irrespective of veracity, while earlier findings linked support to truth and denial to falsity.News organisations may support rumours early and issue corrections later where needed.
- Rumour diffusion: Rumours are more likely to spread across strong ties, and their popularity tends to fluctuate over time.These temporal patterns have been reported on social media and other internet platforms.
- Rumour emergence and propagation: Lack of an official source and personal involvement are the most important factors associated with interest in tracking a rumour.Anxiety was identified as less important, while poster credibility and rumour attractiveness may contribute to propagation.
- Scope: The survey notes that exhaustive analysis of rumour diffusion is outside its scope and directs readers to other surveys.Its focus is research on approaches for detecting and resolving rumours.
5 RUMOUR CLASSIFICATION: SYSTEM ARCHITECTURE
A typical rumour classification system comprises four components—detection, tracking, stance classification and veracity classification—though components may vary or be omitted depending on requirements. Together, they identify rumours, collect related posts, characterize user stances and assess truth using social-media and optional external data.
- System architecture: The architecture typically includes rumour detection, rumour tracking, stance classification and veracity classification, with components optionally omitted for specific requirements.The system begins by identifying unconfirmed information and then processes it through the remaining components.
- Rumour detection: Rumour detection classifies each post in a social-media stream as rumour or non-rumour, supporting identification of emerging rumours.Its output is a labelled stream of posts.
- Rumour tracking: Rumour tracking monitors social media for posts discussing an identified rumour while filtering out irrelevant posts.The input may be a post, sentence, or keyword set, and the output is a collection of related posts.
- Stance classification: Stance classification assigns predefined stance labels to posts associated with the same rumour, indicating how they orient toward its veracity.These labels can facilitate subsequent veracity classification.
- Veracity classification: Veracity classification predicts a rumour’s truth value using tracked posts, stance labels and optionally additional news, websites or databases.Its output may also provide URLs or other contextual data to help users assess reliability.
6 RUMOUR DETECTION
Rumour detection identifies social-media posts containing information that is unverified when posted, typically as a binary rumour-versus-non-rumour classification task. Research has progressed from detecting predefined rumours to identifying new rumours through sceptical-user queries or contextual analysis, with event context providing the current state-of-the-art approach.
- Task definition: Rumour detection classifies each social-media post as a rumour or non-rumour based on whether its information is unverified at posting time.Being a rumour does not imply that the information will later be judged true or false.
- Approaches: Early studies mainly detected rumours known a priori by classifying posts against predefined rumour examples.This work was described as limited to finding previously known rumours.
- Approaches: Zhao et al. detected new rumours by identifying sceptical or enquiring tweets, using manually curated regular expressions associated with questions about veracity.Their approach assumed that rumours provoke users to question or enquire about whether the information is true.
- Approaches: Zubiaga et al. learned context across a breaking-news story because an individual tweet may lack enough information to determine whether its underlying story is a rumour.The approach also avoided relying on enquiring tweets, which may not be triggered by every rumour.
- Generalisation: Feature patterns vary dramatically across events, making rumours difficult to distinguish at tweet level, whereas leveraging event context enables generalisability.Zubiaga et al. showed that contextual information can address the cross-event limitations identified by feature analysis.
- State of the Art: The state-of-the-art approach leverages earlier posts associated with an event to determine whether a tweet constitutes a rumour.Zubiaga et al. (2017) presented this context-based method as the state of the art for rumour detection.
7 RUMOUR TRACKING
Rumour tracking identifies subsequent posts associated with a detected rumour, typically as a binary classification task evaluated with precision, recall, and F1 score. Research remains scarce, with the tweet latent vector approach identified as the best approach while tracking newly emerging rumours remains unstudied.
- Task definition: Rumour tracking identifies subsequent posts associated with a detected rumour, usually by classifying each post as related or unrelated.Inputs may be rumour-specific streams filtered by keywords, broader event-related streams, or unrestricted post streams.
- Evaluation: Precision, recall, and F1 score are used to evaluate rumour tracking, with related posts forming the positive class.
- Datasets: Qazvinian et al.’s dataset contains over 10,000 tweets associated with 5 different rumours, annotated as related or unrelated.Unrelated tweets share characteristics such as overlapping keywords with related tweets, making classification more challenging.
- Research approaches: Research in rumour tracking is scarce, and the best approach is Hamidian and Diab’s tweet latent vector method.Qazvinian et al. used supervised learning with content, network, and Twitter-specific meme features, while Hamidian and Diab used latent tweet representations to address limited tweet length and context.
- Research gaps: Rumour tracking has not been studied for emerging rumours, so generalisable approaches enabling their tracking are still needed.Related event detection and tracking work uses keyword graphs to detect communities and track incoming event-related posts.
8 RUMOUR STANCE CLASSIFICATION
Rumour stance classification determines whether individual social-media posts support, deny, query, or comment on a rumour’s disputed veracity. Research spans annotated datasets, supervised and conversational approaches, and increasingly macro-averaged evaluation for skewed four-way labels, with RumourEval identifying leading systems.
- Task definition: Rumour stance classification assigns each post an orientation toward a rumour’s veracity: support, deny, query, or comment.The task applies to both breaking-news rumours and long-standing rumours discussed on social media.
- Evaluation and state of the art: Because four-way labels are generally skewed, evaluation increasingly emphasizes macro-averaged scores alongside micro-averaged precision, recall, F1, and accuracy.On the RumourEval stance shared task, Kochkina et al. (2017) was judged best performing, while Aker et al. (2017) later reported state-of-the-art results on the same dataset.
- Datasets: Publicly available resources include the PHEME stance dataset, with tweet-level support, deny, query, and comment annotations across nine events.Other datasets include 300 rumoured claims and a Fake News Challenge stance dataset composed of news articles rather than social-media posts.
- Research progression: Early studies progressed from Mendoza et al.’s manual analysis of 14 rumours, evenly split between true and false, to automatic supervised classification by Qazvinian et al.Qazvinian et al. classified tweets as supporting, denying, questioning, or neutral and also addressed rumour tracking.
- Methods: Later systems expanded supervised features and models, including time information, hashtags, LIWC features, logistic regression, naïve Bayes, random forests, and Gaussian processes.Lukasik et al. also introduced Brown clusters and evaluated cross-rumour training, reporting an accuracy of 67.7%.
- Conversational stance: Conversation-aware work modeled tree-structured Twitter discussions, while RumourEval evaluated four-way stance classification of individual tweets within conversational threads.Most RumourEval systems treated the task as single-tweet classification, except the best-performing approach.
9 RUMOUR VERACITY CLASSIFICATION
Rumour veracity classification determines whether a rumour is true, false, or unverified, and most research has focused on this task after rumours are identified. Studies use temporal, structural, linguistic, user, and verification features, with performance depending on propagation stage and evaluation setting.
- Task definition: Veracity classification assigns each rumour one of three labels: true, false, or unverified.Some studies instead formulate the task as binary true-versus-false classification.
- Research focus: Most rumour research focuses on veracity classification while assuming rumours have already been identified or supplied by humans.This emphasis means preceding components, especially rumour detection, are often skipped.
- Feature-based approaches: Temporal and structural features distinguish true from false rumours over long-term windows, whereas user and linguistic features support early veracity classification.Temporal and structural information is generally unavailable during the early stages of propagation.
- Feature-based approaches: Verification features incorporate source credibility, source identification, source diversity, locations, event propagation, and belief identification derived from stance classification.These features were informed by journalists and used to compare a proposed approach with earlier baseline systems.
- State of the Art: The RumourEval approach that aggregates individual-tweet stance to classify rumour veracity into three categories is identified as the current best-performing system.The survey cautions that comparing systems is difficult because studies use different datasets and evaluation methods.
10 APPLICATIONS
Applications for social-media rumour detection and verification range from research projects to operational tools for tracking, visualising, verifying, fact-checking, and assessing credibility. These systems support journalists and analysts, although some still require human input or do not determine veracity directly.
- Research projects: PHEME studied natural language processing for rumour detection and resolution, producing work on detection, stance, contradiction, modelling, visualisation, and journalistic verification.The European Commission-funded project ran from 2014–2017.
- Rumour tracking: Emergent automatically tracks social-media mentions of URLs associated with rumours, but humans must identify rumours and select the relevant URLs.Its work also examined media reporting and best practices for debunking misinformation.
- Rumour visualisation: TwitterTrails lets users investigate rumour origins, propagation, and refutations through burst, timeline, retweet, and co-retweet-network visualisations.It automatically answers questions about a rumour, including its originator and burst activity.
- Misinformation and fact-checking: Hoaxy collects, detects, and analyses online misinformation and fact-checking, while ClaimBuster identifies check-worthy claims and Una Hakika provides manually updated rumour search and data-collection tools.ClaimBuster’s demo covered claims from the 2016 U.S. election and user-submitted text.
- Verification platforms: Verification applications include InVID for authenticating newsworthy video, CrossCheck for collaborative election misinformation verification, Décodex for checking news-site reliability, and Check for newsroom verification.REVEAL similarly focused on journalistic and enterprise verification, especially image verification.
- Journalistic assessment: Seriously Rapid Source Review helps journalists assess Twitter sources around breaking news, while TweetCred assigns each tweet a credibility rating from 1 to 7 without determining story veracity.The source-review system can help find eyewitnesses who may confirm or debunk rumours.
11 DISCUSSION: SUMMARY AND FUTURE RESEARCH DIRECTIONS
The survey reviews progress toward four-component rumour classification systems—detection, tracking, stance classification, and veracity classification—while identifying gaps and priorities for future research. It highlights the need to strengthen early pipeline stages, improve evidence-based and explainable veracity decisions, expand datasets, and assess applicability beyond rumours.
- Summary: The survey synthesizes research on four rumour-classification components: detection, tracking, stance classification, and veracity classification.It presents a guide to the state of the art, focusing on rumours circulating in social media.
- Summary: Research has progressed substantially across all four components, but studies have adopted diverse approaches and have focused mainly on later-stage stance and veracity classification.Detection and tracking remain necessary prerequisites for those later stages, yet tracking has often been skipped or assumed.
- Future research directions: Future detection research should test event-detection techniques for rumours and use context and interactions because post content alone may be insufficient.A rumour detector must additionally determine whether a detected event constitutes a rumour.
- Future research directions: Publicly available datasets remain limited, so researchers are encouraged to release datasets that support research across datasets and comparison of approaches.The survey identifies dataset scarcity as an important limitation on rumour-classification development.
- Future research directions: Veracity systems should provide reasons alongside decisions and distinguish resolved rumours from unresolved cases where classification becomes an evidence-limited prediction task.The survey also calls for more research on exploiting conversational context for stance classification and on applying rumour approaches to hoaxes and fake news.