Source-linked AI summary
Automated Fact Checking: Task formulations, methods and future directions
James Thorne, Andreas Vlachos
TL;DR
Automated fact checking addresses the growing need to assess online claims amid widespread information and inconsistent cross-disciplinary terminology. This survey unifies task formulations and methods across NLP and related fields, emphasizing evidence as a central distinction. It concludes that resources remain limited, complex claims require substantial reasoning, and current systems have important scope boundaries.
Problem
Online information can spread true and false content widely, while fact-checking research uses inconsistent terminology across disciplines.
Method
The paper surveys automated fact checking across NLP and related fields, unifying definitions, task formulations, datasets, modelling approaches, and evidence use.
Results
The survey identifies evidence as a distinguishing factor across task formulations and methods, including textual, structured, social, and previously checked-claim sources.
Takeaways & Limitations
The paper highlights evidence-grounded decisions as important for automated fact checking and proposes open research challenges for NLP.
Takeaways & Limitations
Current systems struggle with journalists’ complex reasoning, while available resources often constrain claims, evidence, or scale.
Abstract
from arXiv · showhide
The recently increased focus on misinformation has stimulated research in fact checking, the task of assessing the truthfulness of a claim. Research in automating this task has been conducted in a variety of disciplines including natural language processing, machine learning, knowledge representation, databases, and journalism. While there has been substantial progress, relevant papers and articles have been published in research communities that are often unaware of each other and use inconsistent terminology, thus impeding understanding and further progress. In this paper we survey automated fact checking research stemming from natural language processing and related disciplines, unifying the task formulations and methodologies across papers and authors. Furthermore, we highlight the use of evidence as an important distinguishing factor among them cutting across task formulations and methods. We conclude with proposing avenues for future NLP research on automated fact checking.
1 Introduction
The spread of online information has increased demand for fact checking, while the volume and complexity of claims motivate automating parts of the process. This survey unifies terminology, task formulations, evidence use, datasets, and modelling approaches across related research.
- Social media enables information from individuals and organizations to reach hundreds of millions, spreading both true and false content.
- False information has reached greater audiences and increased demand for fact checking of online content.
- Journalistic fact checking assesses claim truthfulness by researching evidence, understanding context, and reasoning toward a verdict.The process can take less than one hour to a few days, and automation aims to reduce the human burden.
- The survey addresses inconsistent terminology and siloed research across NLP, machine learning, knowledge representation, databases, and social media analysis.It notes that similar tasks have been called fake news detection, stance classification, and incongruent headline detection.
- The paper unifies definitions, common concepts, datasets, and modelling approaches while examining how evidence shapes NLP requirements.It discusses textual entailment and knowledge-base construction as examples of differing evidence-related requirements.
2 Fact checking in journalism
Journalistic verification and fact checking are treated as distinct but complementary processes. Verification establishes the trustworthiness of source context, while fact checking addresses a claim’s logic, coherence, and context.
- Verification often checks the source, date, and location of materials, whereas fact checking addresses a claim’s logic, coherence, and context.
- The survey adopts the verification–fact checking distinction to highlight differences between automated approaches.The text notes that the two processes are often used interchangeably but have been defined as complementary.
- Misinformation is inaccurate or incomplete information, while disinformation additionally assumes malicious intent to mislead.Disinformation is therefore treated as a subset of misinformation.
- Fact checking can help detect misinformation but cannot distinguish misinformation from disinformation.
3 Towards Automated Fact Checking
Automated fact checking varies in its inputs, evidence, outputs, and required processing. The survey emphasizes that evidence determines whether systems can merely classify claims or also retrieve and justify verdicts.
- 3.1 Inputs: Automated approaches use subject-predicate-object triples, textual claims, or entire documents as inputs.Document-level systems must first extract claims, increasing task complexity.
- 3.1 Inputs: Textual claims are typically short sentences extracted from longer passages to retain context relevant to the claim.
- 3.1 Inputs: Entity disambiguation and grounding are necessary because identical names can refer to different entities and properties.Named Entity Linking performs this task explicitly for knowledge-base triples or implicitly through textual evidence retrieval.
- 3.2 Sources of evidence: Evidence sources include social-network crowd behavior, knowledge graphs, texts, and repositories of previously fact checked claims.Crowd behavior is useful when textual sources or structured knowledge bases are unavailable.
- 3.2 Sources of evidence: Knowledge-graph topology can indicate plausibility but does not negate the truth of improbable claims.Improbable but believable claims may be more likely to become viral and need verification.
- 3.3 Output: FEVER combines claim labels with retrieval of supporting evidence from multiple Wikipedia documents and sentences.If the label is correct but the retrieved evidence is incorrect, the answer is considered incorrect.
- 3.2 Sources of evidence: Previously fact checked claims enable veracity-label prediction but limit checking to claims similar to those already in the repository.
4 Fact Checking Datasets
Fact-checking datasets range from small, heterogeneous collections to large, machine-readable resources with annotated evidence. Their scale, structure, and evidence coverage determine whether they support machine-learning development and evaluation.
- The 221-claim Vlachos and Riedel dataset links political claims to heterogeneous evidence sources, making automated evidence selection and evaluation difficult.Its limited sample size precludes developing machine-learning-based fact-checking systems.
- The Wang dataset contains 12.8K labeled Politifact claims with speaker and context metadata, supporting machine-learning training and evaluation.
- The Fake News Challenge dataset contains approximately 50K headline–body pairs derived from 300 claims and 2,595 news articles.
- The HeroX challenge released 90 labeled claims, but its missing evidence and small size required manual evaluation and prevented machine-learning training.Its broad claim types nevertheless expose varied forms of misinformation and system requirements.
- FEVER introduced 185K claims with manually annotated Wikipedia evidence sentences, enabling large-scale training of machine-learning fact-checking models.Restricting the evidence pool made evidence selection feasible to annotate at scale.
5 Methods
Automated fact-checking methods vary by input, evidence, and output formulation, with supervised learning common across tasks. Approaches include entailment, relation extraction, matching prior checks, and speaker profiling, each with scope constraints.
- Most automated fact-checking methods are supervised classifiers trained on labeled data, regardless of input type or evidence source.
- RTE-based systems classify whether textual evidence supports, opposes, or observes a claim, but generally assume that evidence is already supplied.
- Retrieval and sentence-selection stages can precede RTE when evidence must be found in large collections such as Wikipedia.
- Relation-extraction approaches identify textual patterns describing relations between entities in knowledge graphs, often using distant supervision.
- Matching claims against previously fact-checked statements reduces verification to textual similarity but handles only repeated or paraphrased claims.
- Speaker profiling can improve classification accuracy through originator credit history, but introduces bias, fails for sources without recorded histories, and raises ethical concerns.
6 Related Tasks
Related tasks provide useful signals for fact checking but do not by themselves establish truthfulness. Their limitations concern missing context, evidence, source assumptions, and the distinction between rumors, deception, and factual claims.
- Closed-world fact-checking models assume that provided evidence is true, whereas web-scale systems must handle evidence of unknown veracity.
- Common Sense Reasoning: Common-sense reasoning is needed because fact checking involves arguments requiring both explicit world knowledge and implicit warrants.
- Subjectivity and Emotive Language: Article-origin classification into Hoax, Satire, or Propaganda is subjective-language detection rather than evidence-based fact checking.
- Deceptive Language Detection: Linguistic, emotive, and syntactic cues can indicate deception, but statements still require context because truthfulness cannot be determined from these cues alone.
- Rumor Detection: Rumor detection identifies unverified reports using subjectivity and readership growth, although a claim may be true or false independently of being a rumor.
- Speaker Profiling: Speaker-profile methods face a cold-start challenge for new topics and may not capture which subjects an originator lies about.
- Click Bait Detection: Clickbait detection relies on headline features and does not use evidence, distinguishing it from evidence-based fact checking.
7 Open Research Challenges
The survey identifies open challenges in evidence, resources, rationale generation, multimodal verification, and reasoning complexity. Existing systems remain narrower than journalistic fact checking in scope and reasoning demands.
- The survey frames evidence as central to automated fact checking, covering claims expressed in text or knowledge-base triples and their need for justification.
- Open-world fact checking requires scalable information retrieval and evaluation, together with verification of the evidence itself.
- FEVER-style resources focus on short claims averaging eight words, leaving compound, complex, and document-level fact checking underdeveloped.
- FEVER justifications are restricted to selected Wikipedia sentences, unlike human fact checkers’ synthesized rationales.
- Multimodal fact checking should address evidence in images and videos, including forged visual material commonly examined in journalism.
- Journalistic fact checking currently exceeds system capabilities because it requires complex reasoning, even for apparently simple claims.