Source-linked AI summary
A Survey on Automated Fact-Checking
Zhijiang Guo, Michael Schlichtkrull, Andreas Vlachos
TL;DR
Rapidly spreading misinformation and the time demands of manual checking motivate automated fact-checking. The paper surveys NLP-based research by unifying its stages, datasets, and models into a common framework, then identifies challenges for future research. Its central outcome is a synthesis spanning claim detection, evidence retrieval, verdict prediction, and justification production.
Problem
Manual fact-checking can take hours or days, and rapidly spreading information makes manual validation insufficient.
Method
The paper surveys automated fact-checking and unifies prior definitions, datasets, and modelling strategies in a framework of claim detection, evidence retrieval, verdict prediction, and justification production.
Results
The survey provides an extensive overview of existing datasets and modelling strategies and identifies vital challenges for future research.
Takeaways & Limitations
Automated fact-checking research can be compared across a shared framework connecting its task formulations, datasets, modelling strategies, and research challenges.
Takeaways & Limitations
Current fact-checking research often assumes a single trustworthy authoritative source, although available sources may disagree or be untrustworthy.
Abstract
from arXiv · showhide
Fact-checking has become increasingly important due to the speed with which both information and misinformation can spread in the modern media ecosystem. Therefore, researchers have been exploring how fact-checking can be automated, using techniques based on natural language processing, machine learning, knowledge representation, and databases to automatically predict the veracity of claims. In this paper, we survey automated fact-checking stemming from natural language processing, and discuss its connections to related tasks and disciplines. In this process, we present an overview of existing datasets and models, aiming to unify the various definitions given and identify common concepts. Finally, we highlight challenges for future research.
1 Introduction
The survey frames automated fact-checking as an NLP-based response to the scale and speed of misinformation. It unifies prior work into a framework covering claim detection, evidence retrieval, verdict prediction, and justification production, while connecting the field to related research.
- Manual fact-checking can take hours or days, while tight deadlines and rapidly spreading information make manual validation insufficient.
- Earlier surveys omitted claim detection or recent work on explainable verdicts, which this survey explicitly includes.
- Unlike fake-news and social-media-focused surveys, this work concentrates on general-domain claim veracity across domains and sources.
- The survey organizes automated fact-checking into claim detection, evidence retrieval, verdict prediction, and justification production.
- It surveys datasets and modelling strategies across these stages and identifies research challenges for future work.
2 Task Definition
Automated fact-checking is defined as a pipeline that selects claims, retrieves evidence, predicts veracity, and produces justifications. The survey also describes alternative formulations, dataset structures, and assumptions underlying these stages.
- Task Definition: The framework comprises claim detection, evidence retrieval, and claim verification, with verification divided into verdict prediction and justification production.
- Claim Detection: Claim detection selects claims for verification using check-worthiness, rumour detection, or whether assertions are verifiable with readily available evidence.
- Evidence Retrieval: Evidence retrieval finds text, tables, knowledge bases, images, or metadata that indicate a claim’s veracity, rather than relying only on claim-surface patterns.
- Verdict Prediction: Verdict prediction assigns truthfulness labels from retrieved evidence, ranging from binary true/false or supported/refuted labels to finer-grained categories.
- Datasets: Claim-detection datasets vary in input format, evidence type, and output representation, including social-media posts, statements, text, metadata, labels, and rankings.
- Justification Production: Justifications should explain how evidence, assumptions, and reasoning produce a verdict; merely presenting retrieved evidence is a weak baseline.
3 Datasets
The survey organizes fact-checking datasets around inputs, evidence, and outputs, covering natural and artificial claims across diverse domains. Dataset designs vary in claim formats, evidence availability, verdict labels, and justification support.
- Dataset axes: Datasets are analyzed along input, evidence, and output axes aligned with the three-stage fact-checking framework.The output axis includes verdicts and justifications, while the framework brings together efforts using different terminology.
- 3.1 Input: Social-media datasets use source tweets, individual posts, or collections of related posts as claim-detection inputs.Examples include PHEME, RumourEval, CheckThat, and thread-based representations designed to compensate for limited context in individual posts.
- 3.1 Input: Other datasets use debate transcripts, fact-checking websites, domain-specific claims, question-answering answers, passages, or artificially constructed claims.Artificial construction includes mutating Wikipedia sentences and representing claims as subject-predicate-object triples.
- 3.2 Evidence: Evidence ranges from metadata and textual sources to Wikipedia sentences, knowledge graphs, and tables, but each form has grounding or coverage limitations.Metadata does not ground claims; graph topology indicates plausibility without conclusive evidence; retrieved evidence may contain irrelevant information or omit necessary information.
- 3.3 Verdict & Justification: Verdict labels range from binary true/false decisions to multi-class schemes, while newer datasets combine labels with extracted evidence.Label sets can range from 2 to 27, and some studies normalize heterogeneous labels through post-processing rules.
- 3.3 Verdict & Justification: Most datasets lack journalist-written textual justifications, and available gold explanations may rely on fact-checking articles unavailable during inference.This creates a mismatch when systems are expected to generate justifications from retrieved evidence rather than from supplied articles.
4 Modelling Strategies
Modelling strategies commonly use separate pipeline components for claim detection, evidence retrieval, verification, and justification production, although joint approaches also exist. The survey highlights challenges involving check-worthiness, multi-evidence reasoning, evidence-free predictions, and explanation faithfulness.
- Overall strategies: The most common strategy builds separate models for each component and applies them in a pipeline, while joint approaches model multiple components together.Joint approaches use end-to-end learning or joint output distributions.
- Claim Detection: Claim detection is typically framed as classification of whether claims are checkable or check-worthy, but real-world claim volume makes evidence use difficult.Neural sequence and graph models use surrounding social-media context, while joint rumour systems can make preliminary predictions without evidence.
- Evidence Retrieval and Verification: Evidence retrieval and verification are usually separate pipeline modules, though some systems train them jointly.Verification is related to textual entailment, and retrieval uses search APIs, indices, entity linking, ranking functions, or dense representations.
- Evidence Retrieval and Verification: Multiple evidence pieces require aggregation, either by concatenating them before entailment modelling or by using specialized aggregation components.Both real-world and research-created claims often require reasoning over combined evidence.
- Claim Verification: Some verification systems omit retrieval and rely on surface forms, metadata, or pretrained-model knowledge as evidence.These approaches do not explicitly retrieve supporting evidence for the claim.
- Justification Production: Justification methods include attention-based selections, interpretable rules, extractive summaries, and abstractive text generation, each with distinct limitations.Attention may be unfaithful or unreadable, rules are restricted to knowledge-base representations, and generated explanations may be plausible but inconsistent with predictions.
5 Related Tasks
Automated fact-checking overlaps with misinformation, disinformation, propaganda, fake-news, and previously fact-checked claim detection, but these tasks differ in scope and objectives. The survey focuses on verifying general-domain claims through evidence across domains and sources.
- Misinformation and Disinformation: Misinformation contradicts or distorts verifiable facts, whereas disinformation is misinformation deliberately propagated to deceive.The distinction is based on intent, not merely on the truth value of the claim.
- Misinformation and Disinformation: Propaganda detection additionally targets techniques such as appeals to emotion, logical fallacies, whataboutery, and cherry-picking.This makes propaganda detection related to, but not identical with, disinformation detection.
- Detecting Previously Fact-checked Claims: Previously fact-checked claim detection matches new claims against verified claims, often reducing the task to sentence-level similarity.The survey distinguishes this from verification based on finding evidence, while noting that repeated misleading claims make the task useful.
6 Research Challenges
Automated fact-checking faces challenges involving labels, source reliability, subjectivity, dataset bias, multimodality, multilinguality, faithfulness, and the timing of intervention. Future systems must address these constraints while supporting earlier responses to misinformation.
- Choice of Labels: Fine-grained labels can encode composite claims, while binary truth labels may miss misleading presentations such as cherry-picking and technical truth.The paper identifies claim decomposition and new labels for misleading forms as possible responses.
- Sources & Subjectivity: Trustworthy sources may contradict one another, challenging systems that rely on a single authoritative source.Proposed directions include assessing source credibility, agreement, and coherence with known facts.
- Sources & Subjectivity: Practical deadlines and audience time constraints can force subjective triage, introducing bias despite attempts to use objective criteria.The paper calls for systems that interact with users in real time as their needs evolve.
- Dataset Artefacts & Biases: Models can exploit dataset artefacts instead of learning fact-checking, with FEVER predictions largely driven by indicative claim words.Adversarial claims and resilient systems are presented as responses to this problem.
- Multimodality: Large-scale annotated datasets containing evidence beyond metadata would support multimodal fact-checking across text, tables, images, audio, and video.Most existing work has focused on text, although some studies incorporate multimodal information.
- Multilinguality: Multilingual verification remains open because claims and evidence may use different languages, while adding other-language training data has not improved performance.The paper notes a multilingual dataset spanning 25 languages and calls for better cross-language resource alignment.
- Faithfulness: Generated justifications may be unfaithful, creating risks when they are used to convince users that model predictions are valid.The paper states that faithfulness is difficult to evaluate and that better evaluation techniques are needed.
- From Debunking to Early Intervention and Prebunking: Most automated fact-checking reacts after misinformation appears, whereas early detection and prebunking could support intervention before misinformation spreads widely.The paper reports that prebunking has been shown to be more effective than post-hoc debunking.
7 Conclusion
The paper unifies automated fact-checking research within a framework covering four core components and uses it to survey datasets, models, and future challenges.
- 7 Conclusion: The framework comprises claim detection, evidence retrieval, verdict prediction, and justification production.It unifies task formulations and methodologies across different research efforts.
- 7 Conclusion: The survey provides an extensive overview of existing datasets and modelling strategies and identifies vital challenges for future research.