PaperlayerView exact revision on arXiv

Source-linked AI summary

AVeriTeC: A Dataset for Real-world Claim Verification with Evidence from the Web

Michael Schlichtkrull, Zhijiang Guo, Andreas Vlachos

arXiv:2305.13117v3cs.CL

TL;DR

Existing automated fact-checking datasets often use artificial claims or provide inadequate, superficial, or temporally leaked evidence. AVeriTeC addresses these gaps with a real-world dataset, structured question-answer evidence, justifications, and an evaluation pipeline, while its metric cannot detect unsupported abstractive answers.

  • Problem

    Existing automated fact-checking resources often use purpose-made claims, lack evidence annotations, or include evidence published after the investigated claims.

  • Method

    AVeriTeC combines real-world claims with web evidence, question-answer pairs, textual justifications, and a multi-component baseline for retrieval, veracity prediction, and justification generation.

  • Results

    4,568 claims and κ = 0.619 establish AVeriTeC as a benchmark with annotated evidence, reasoning justifications, and an accompanying evaluation scheme.

  • Takeaways & Limitations

    The dataset supports multi-step claim verification against the open web while exposing retrieval and question-generation as important challenges.

  • Takeaways & Limitations

    The evaluation metric cannot detect answers hallucinated by abstractive models when they are unsupported by source documents.

Abstract

from arXiv · show

Existing datasets for automated fact-checking have substantial limitations, such as relying on artificial claims, lacking annotations for evidence and intermediate reasoning, or including evidence published after the claim. In this paper we introduce AVeriTeC, a new dataset of 4,568 real-world claims covering fact-checks by 50 different organizations. Each claim is annotated with question-answer pairs supported by evidence available online, as well as textual justifications explaining how the evidence combines to produce a verdict. Through a multi-round annotation process, we avoid common pitfalls including context dependence, evidence insufficiency, and temporal leakage, and reach a substantial inter-annotator agreement of $κ=0.619$ on verdicts. We develop a baseline as well as an evaluation scheme for verifying claims through several question-answering steps against the open web.

1 Introduction

Existing automated fact-checking datasets often use artificial claims or provide insufficient, context-dependent, or temporally leaked evidence. AVeriTeC addresses these limitations with real-world claims, structured evidence annotations, and multi-round quality controls.

  • Fact-checking helps limit misinformation, but limited manual resources motivate automated fact-checking as an assistive tool.
  • Existing datasets often contain purpose-made claims, lack evidence annotations, or include evidence published after the investigated claims.
  • AVeriTeC combines real-world claims with web evidence, question-answer decomposition, and textual justifications explaining how evidence supports or refutes verdicts.
  • AVeriTeC addresses context dependence through claim normalization, evidence insufficiency through repeated annotation and blind quality control, and temporal leakage through pre-claim evidence and temporal splits.
  • 4,568 examples from 50 fact-checking organizations achieved free-marginal κ = 0.619 and support a baseline for multi-step web verification.

2 Related Work

Prior fact-checking resources commonly lack evidence, intermediate reasoning annotations, or safeguards against context dependence and temporal leakage. AVeriTeC builds on question-answer decomposition while targeting these gaps with real-world, retrieved evidence.

  • Real-world claims are valued because their inclusion by professional fact-checkers indicates checkworthiness, but prior datasets often lack evidence or intermediate-step annotations.
  • Some real-world datasets annotate evidence superficially, allowing evidence published after the claims and other temporal leaks.
  • Dataset comparisons distinguish claim source and type, context independence, evidence sufficiency, temporal leakage, and whether evidence retrieval is required.
  • Question-answer decomposition is promising, but earlier datasets either used questions mainly for context or did not verify evidence sufficiency and used post-claim evidence.

3 Annotation Structure

AVeriTeC represents fact-checking through linked question-answer evidence, verdicts, and textual justifications. Its four-way label set captures conflicting evidence or misleading cherry-picking alongside support, refutation, and insufficient evidence.

  • Each claim includes question-answer pairs representing evidence, a veracity label, and a textual justification describing how the evidence supports that label.
  • Questions can have multiple answers and refer to previous questions, supporting disagreement representation and multi-hop reasoning.
  • Answers must link to web sources unless marked unanswerable, and evidence pages are cached in the Internet Archive.
  • AVeriTeC adds Conflicting Evidence/Cherrypicking to the usual supported, refuted, and not-enough-evidence classes.
  • Textual justifications explain how evidence leads to verdicts, including disagreements and logical reasoning beyond the retrieved evidence.
  • Metadata covers claim speaker, publisher, publication date, location, claim type, and fact-checking strategy.

4 Annotation Process

AVeriTeC uses a multi-phase annotation pipeline that normalizes claims, generates web-supported questions and answers, and independently checks evidence sufficiency. Temporal safeguards restrict evidence and organize dataset splits, but the leakage guarantee remains approximate.

  • Starting from 8,000 articles, the pipeline removes 537 duplicates and 802 paywalled or dead articles before annotation.
  • Annotators extract claims and metadata, generate evidence-backed questions and answers, assign a temporary verdict, and obtain an independent verdict and justification.
  • Claim normalization enriches extracted claims with necessary context and excludes speculative and multimodal claims.
  • The search interface restricts documents to those published before the claim date, and fact-checking articles themselves cannot serve as evidence.
  • Temporal-leakage prevention is approximate because claim dates may be unavailable, publication dates may be misidentified, and fact-checking can follow claims by several days.

5 Dataset Statistics

AVERITEC uses temporally separated splits and provides multiple question-answer annotations per claim, with substantial verdict agreement but an imbalanced label distribution.

  • AVERITEC is split into training, validation, and test data temporally.
  • Claims have on average 2.60 questions, while questions have 1.07 answers on average.Answers are 53% extractive, 26% abstractive, and 17% boolean; 4% of questions are unanswerable.
  • 53% of answers are extractive, followed by 26% abstractive and 17% boolean answers.
  • The dataset is somewhat unbalanced because journalists tend to select false or misleading claims, leaving most claims refuted.
  • κ=0.619 measures substantial inter-annotator agreement on verdict annotations from re-annotating 100 claims.A second annotator set repeated question-and-answer generation after claim extraction and normalization, using Randolph’s free-marginal multirater κ for the unbalanced dataset.

6 Evaluation

The evaluation scores open-web evidence retrieval with approximate matching and conditions veracity and justification scores on retrieving correct evidence. It also recognizes alternative evidence paths as a limitation and supports extensions for multiple reference strategies.

  • AVERITEC evaluates retrieved evidence by agreement with gold evidence, awarding veracity and justification credit only after correct evidence retrieval.Unlike closed-source benchmarks, the open-web setting prevents reliance on exact source matching.
  • Hungarian METEOR optimally matches generated question-answer sequences to references using a pairwise scoring function.The metric uses METEOR for approximate similarity and generalizes exact-match evidence recall when the pairwise function is exact match.
  • Evaluation computes Hungarian METEOR separately for questions and for concatenated questions and answers.This reflects that asking the right questions is itself a retrieval subtask, since good questions can function as search queries without good answers.
  • Alternative evidence formulations can both establish a claim, but the evaluation may penalize systems for choosing a valid path unlike the reference.The authors recommend human evaluation because judging evidence sufficiency was already part of annotation.
  • The metric can extend to multiple reference sets by selecting the best-matching score across alternative questioning strategies.This extension would support annotations representing different valid ways to question a claim.
  • Hungarian METEOR scores between independently generated annotations are 0.28 for questions and 0.22 for questions plus answers.The low scores highlight the difficulty of automatic evaluation for this task.

7 Experiments

The experiments evaluate a multi-stage AVeriTeC verification pipeline against no-search, gold-evidence, and ChatGPT variants. Retrieval remains the main bottleneck, while stance detection and evidence support create additional failure modes.

  • Baseline: The baseline combines search-question generation, Google retrieval, evidence reranking, question generation, veracity prediction, and justification generation.Its components include BLOOM, BM25, BERT-large, and BART-large models.
  • Evaluation: The evaluation compares the full AVERITEC pipeline with no-search, gold-evidence, and ChatGPT systems across veracity labels and macro-average F1.Table 4 reports supported, refuted, conflicting evidence/cherrypicking, and not enough evidence categories.
  • Evaluation: λ = 0.25 is recommended because it requires important information while allowing paraphrasing and partial evidence.The baseline performs decently at λ = 0.2 and λ = 0.25 but poorly at higher evidence cutoffs.
  • Retrieval: 9/20 development examples contained useful evidence from claim- and generated-question searches, compared with 16/20 when gold questions were used.The gap highlights the difficulty of retrieval and query/question generation.
  • Error analysis: The veracity model frequently misclassified conflicting evidence/cherrypicking, achieving 10% precision with gold evidence because contextual questions were treated as refuting.The any-different-stance rule produced many false positives.
  • Comparisons: ChatGPT outperformed the baseline in pure question generation but received a lower AVERITEC veracity score because generated answers often failed to match gold answers or hallucinated.ChatGPT cannot produce source-backed answers comparable to the baseline.

8 Limitations

AVeriTeC’s evaluation and dataset scope have important boundaries. The metric does not verify answer-source consistency, and the dataset is English-only and inherits biases from fact-checking articles.

  • Evaluation metric: The evaluation metric does not ensure that generated answers are consistent with their source documents, allowing unsupported hallucinated answers to evade detection.Only 53% of gold answers are fully extractive, so abstractive models are expected to be used.
  • Dataset scope: All fact-checking sources and claims are in English, despite geographic coverage of regions around the world.The dataset therefore does not provide multilingual source coverage.
  • Dataset scope: Because claims are taken directly from fact-checking articles, AVeriTeC is subject to biases present in those articles, including documented source-selection bias.The cited example concerns journalists rating claims by male sources as more credible than claims by female sources.
  • Temporal scope: Evidence availability dates rely on Google’s best estimates, so some documents may not have been available when their associated claims were published.The temporal restriction generally avoids leakage but may have exceptions.

9 Ethics Statement

The dataset’s labels and justifications are evidence-bound rather than truth-telling tools, and remain exposed to annotator, journalist, search, and source biases.

  • AVeriTeC is not intended for truth-telling or automated content moderation.Its labels and justifications concern only evidence recovered by annotators.
  • 85 answers rely on flagged sources, while the source list is incomplete and may contain false positives.The authors warn that the dataset may include misleading examples and cause harm if treated as authoritative.
  • The dataset was not anonymised because its claims concern public figures and events drawn from public journalistic data.References to people and events are retained because they are important for fact-checking claims.

10 Conclusion

The paper concludes by presenting AVERITEC as a real-world fact-checking benchmark built around decomposed evidence, justifications, and quality-controlled annotations.

  • 4,568 claims form AVERITEC, each paired with question-answer evidence, a veracity label, and a textual justification.The question-answer pairs decompose the fact-checking process, while justifications explain how evidence supports the label.
  • κ = 0.619 measures substantial inter-annotator agreement after the multi-step annotation process.The process is designed to provide evidence sufficiency and avoid temporal leakage.
  • The paper introduces and analyses a baseline and evaluation scheme, establishing AVERITEC as a new benchmark.

A Dataset Access

The appendix documents dataset access and baseline prompt procedures, including adversarial question generation, evidence-conditioned prompts, and implementation details.

  • Dataset Access: The dataset and baseline are released on GitHub, while the test split is withheld from public release at submission time.Training and development splits are released; reviewers may request the test split privately.
  • Dataset Access: Annotators were paid through Appen, using a team-developed annotation platform whose code is available upon request.The paper reports a total crowdworker cost of C40,835 before the passage ends.
  • Baseline Models: Search-question generation retrieves 10 similar training claims with BM25 and prompts a model using the claim and speaker in an adversarial template.The adversarial tone was found crucial for retrieving useful results beyond claim-only search.
  • Baseline Models: Evidence-question generation retrieves 10 similar training question-answer pairs using BM25 between answers and evidence lines.Including the claim reduced performance because the model generated claim-related rather than evidence-related questions.
  • Baseline Models: The no-evidence baseline uses retrieved similar claims to prompt justification generation, while adversarial prompting provided no performance improvement.
  • Baseline Models: The baseline fine-tunes BERT-large components and uses a single Nvidia A100 GPU for training.The evidence reranker uses corrupted claim, question, and answer negatives; the supplied passage reports BERT-large’s 340 million parameters.

E.2 Stance Detection

The stance-detection appendix describes a BERT-large classifier trained from annotated claims and contrasts evidence examples used in the broader baseline pipeline.

  • Stance Detection: The stance detection model uses BERT-large with a text-classification head, fine-tuned with Adam at learning rate 0.001 and batch size 128.The model has 340 million parameters and is trained on examples constructed from the training set.
  • Stance Detection: One evidence example links a Facebook post about Rachel Levine to a satire-site label through a question identifying the website.
  • Stance Detection: Another example turns evidence about the wording “complete 57 mega dams” into a question about the Kenyan Government manifesto.
  • Stance Detection: The justification-generation model uses BART-large with two-beam search and a maximum output length of 100 tokens.It is fine-tuned with Adam at learning rate 0.001 and batch size 128.

F Dataset statistics

The dataset analysis reports evidence-source diversity and describes prompts, examples, and baseline evaluation across claim types.

  • Dataset analysis: The dataset analysis includes statistics on evidence modalities, claim types, fact-checker strategies, and baseline performance by claim type.These analyses are presented in Tables 5–9.
  • Evidence sources: 2,989 domains supplied evidence, spanning social media, fact-checking organizations, official sources, news media, and some non-textual sources.Twitter was the most frequent domain at 3%, followed by africacheck.org at 2.5%.
  • Example reasoning: An example claim is marked refuted because its answer and source identify the underlying story as an April Fool’s joke.The example concerns a married Florida couple allegedly selling fake golden tickets to heaven.
  • Prompt examples: The examples illustrate prompts for generating justifications and for producing evidence and verdicts with ChatGPT.One example concerns the claim that Florida mayors cannot discuss gun reform.

I Data Statement

AVeriTeC is constructed from real-world fact-checking claims and uses structured question-answer evidence, metadata, and textual justifications to support verification. Its annotation process addresses context, evidence sufficiency, and label consistency while documenting dataset scope and provenance.

  • Data Statement: 8,000 fact-checking texts were processed, with duplicates, paywalled or dead articles, unsuitable modalities, and unresolved label disagreements removed.The source material came from the Google FactCheck Claim Search API and included claims selected across a defined two-year interval.
  • Data Statement: The dataset contains English-language material from 50 fact-checking organizations, with expected variation across several varieties of news-domain English.Annotator demographic details were not fully available, and speaker demographics were not analyzed.
  • J.1 Introduction: AVeriTeC is designed around real-world claims and multi-step evidence retrieval represented as question-answer pairs.Decomposition breaks complex claims into components that can be answered separately before reaching a verdict.
  • J.1 Introduction: The dataset adds textual justifications explaining how retrieved answers are combined to reach verdicts and expose system reasoning.This lets users follow retrieval and verification steps rather than seeing only a final label.
  • J.3.4 Labels: The four labels are supported, refuted, not enough evidence, and conflicting evidence or cherry-picking.Annotators distinguish absent supporting evidence from evidence that directly contradicts a claim.
  • J.5.1 Overview: Quality-control annotators assign labels and write justifications using the claim, metadata, and question-answer pairs without web search or unsupported background knowledge.Justifications should describe the available information used, or identify what is missing when no verdict can be reached.
Loading 2305.13117v3…