PaperlayerView exact revision on arXiv

Source-linked AI summary

Complex Claim Verification with Evidence Retrieved in the Wild

Jifan Chen, Grace Kim, Aniruddh Sriram, Greg Durrett, Eunsol Choi

arXiv:2305.11859v2cs.CL

TL;DR

Prior fact-checking approaches often lack evidence, rely on fact-checker-curated evidence, or use unrestricted retrieval, unlike the evidence available when an emerging claim is checked. This paper retrieves pre-claim web evidence through a multi-stage pipeline and finds that it improves veracity judgments, although the evidence may not cover every claim aspect.

  • Problem

    Prior fact-checking methods use no evidence, fact-checker-curated evidence, or unrestricted retrieval, rather than modeling evidence available when claims emerge.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    The pipeline decomposes claims into subquestions, retrieves pre-claim web documents and relevant spans, summarizes evidence, and classifies veracity.

    Source: §1 · p. 1, §3.3 · p. 4

  • Results

    Retrieved web evidence improves final fact-checking accuracy, while human evaluation finds summaries mostly faithful and helpful but evidence often incomplete.

    Source: §1 · p. 2, §8 · p. 9

  • Takeaways & Limitations

    The authors conclude that real-world fact-checking needs human involvement because web search often cannot surface all information needed to verify a claim.

    Source: §8 · p. 9, Limitations and Future Directions · p. 10

    First-stage retrieval bottlenecks performance: 36.0% of questions are unanswerable with the most faithful summaries, including cases where relevant information is absent from the web except on fact-checking sites.

    Source: Limitations and Future Directions · p. 10

Abstract

from arXiv · show

Evidence retrieval is a core part of automatic fact-checking. Prior work makes simplifying assumptions in retrieval that depart from real-world use cases: either no access to evidence, access to evidence curated by a human fact-checker, or access to evidence available long after the claim has been made. In this work, we present the first fully automated pipeline to check real-world claims by retrieving raw evidence from the web. We restrict our retriever to only search documents available prior to the claim's making, modeling the realistic scenario where an emerging claim needs to be checked. Our pipeline includes five components: claim decomposition, raw document retrieval, fine-grained evidence retrieval, claim-focused summarization, and veracity judgment. We conduct experiments on complex political claims in the ClaimDecomp dataset and show that the aggregated evidence produced by our pipeline improves veracity judgments. Human evaluation finds the evidence summary produced by our system is reliable (it does not hallucinate information) and relevant to answering key questions about a claim, suggesting that it can assist fact-checkers even when it cannot surface a complete evidence set.

1 Introduction

The paper studies automated fact-checking with evidence retrieved from the web under restrictions intended to reflect what is available when a claim emerges. Its pipeline decomposes claims, retrieves and summarizes evidence, and uses the summaries for veracity classification.

1 Introduction

  • Prior approaches either use human-curated evidence or retrieve without restrictions, unlike this study’s realistic temporal retrieval setting.The evaluation uses CLAIMDECOMP, a dataset of 1,200 real-world complex political claims with veracity labels.Source: §1 · pp. 1–2
  • The pipeline decomposes claims into subquestions, retrieves relevant documents and text spans, generates claim-focused summaries, and classifies veracity from those summaries.Retrieval is restricted to documents authored before the claim and excludes fact-checking websites.Source: §1 · p. 1, Figure 2 · p. 2
  • On CLAIMDECOMP, retrieved web evidence improves veracity classification over fact-checking without evidence, and decomposition improves raw-document retrieval over using the claim alone.The human study also finds summaries mostly faithful and helpful, though retrieved evidence often covers only some aspects of a claim.Source: §1 · p. 2
  • Human evaluation finds claim-focused summaries mostly faithful and helpful, but retrieved evidence rarely covers every aspect of a claim.The authors identify finding sufficient raw evidence in the wild as a core challenge for automated fact-checking.Source: §1 · p. 2

2 Background and Motivation

Prior fact-checking work often uses claims alone, fact-checker-provided justifications, or unconstrained retrieval, which may include fact-checking websites. This paper emphasizes realistic retrieval and evidence assistance because truth-label accuracy alone does not capture the complexity of fact-checking.

2 Background and Motivation

  • Claim-only systems may rely on surface-level linguistic patterns, while evidence-based systems have often assumed fact-checker justifications or unconstrained retrieval.Unconstrained retrieval can include fact-checking websites, and prior web-question retrieval work used humans to filter irrelevant results.Source: §2 · p. 2
  • The proposed realistic setting restricts retrieval to evidence available when a claim was made, a challenging condition in which many claims are not checkable.The authors emphasize evidence retrieval as assistance for human fact-checkers and suggest reusing the task setting and evaluation.Source: §2 · p. 2
  • Fact-checking accuracy is only a proxy because the task is subjective and complex, and needed information is not always available on the web.The paper argues that best-effort information and evidence can help humans assist in judgments from error-prone AI systems.Source: §2 · pp. 2–3

3 Methodology

The methodology turns complex claims into questions, retrieves and filters web documents, selects claim-relevant spans, summarizes documents, and classifies veracity. Its example pipeline produces evidence that can inform a judgment even when its verdict does not match PolitiFact.

3 Methodology

  • The five pipeline stages are claim decomposition, raw-document retrieval, fine-grained retrieval, claim-focused summarization, and veracity classification.Source: §3 · p. 3

3.1 Claim Decomposition

  • Claims are decomposed into 10 yes/no questions using a few-shot prompt to a large language model.Source: §3.1 · p. 3

3.2 First-stage Retrieval

  • First-stage retrieval filters out post-claim and fact-checking-site pages; timestamped and untimestamped URL sets show little overlap.The system retrieves 10 documents per question; about one-third of URLs cannot be scraped.Source: §3.2 · p. 3, Table 1 · p. 4

3.3 Second-stage Retrieval

  • Second-stage retrieval uses BM25 to select relevant text spans, expands and merges overlapping spans, then retains the top-ranked documents.This reduces lengthy retrieved documents to material more focused on the claim.Source: §3.3 · p. 4

3.4 Claim-Focused Summarization

  • A large language model summarizes each retrieved document with respect to the claim, and human evaluation found few-shot prompting worked better.The zero-shot prompt instructs the model not to judge the document’s stance.Source: §3.4 · p. 5

3.5 Veracity Classification

  • A trained DeBERTa-large model classifies claim-summary pairs into six veracity labels.The classifier is trained with five random seeds, and the model with the best development-set performance is evaluated on the test set.Source: §3.5 · p. 5

3.6 Final Pipeline

  • In an example, the retrieved and summarized documents shed light on the claim from complementary perspectives, even though the final judgment differs from PolitiFact’s.The decomposition also produces some redundant questions.Source: §3.6 · p. 5
  • A pilot found that summarizing all documents together often produced unreliable verdicts, while ChatGPT performed worse than the fine-tuned classifier.Source: §3.6 · p. 5

4 Experimental Setup

The experiments evaluate end-to-end veracity prediction on CLAIMDECOMP, using fact-checker labels and comparing evidence-based systems with claim-only and justification-based baselines.

Data and task

  • The evaluation runs the complete pipeline for claim veracity prediction, using fact-checker-assigned labels as its target.Source: §4 · p. 5

Data

  • CLAIMDECOMP contains 1,200 complex PolitiFact claims, split into 800 training, 200 development, and 200 test examples, with six-way labels.The dataset also provides expert-written justification paragraphs and annotated subquestions.Source: §4 · p. 5

Evaluation metrics

  • The study reports accuracy, mean absolute error, Macro-F1, and soft accuracy, which counts adjacent-label errors as correct on the six-point scale.Source: §4 · p. 5

Comparison systems

  • Claim-only classification is the lower-bound baseline, while appending a label-excluded human justification provides an oracle upper-bound comparison.Source: §4 · p. 5

5 Automatic Evaluation: Claim Veracity

The automatic evaluation compares retrieval conditions and pipeline-stage choices, then tests how first-stage retrieval changes over time. Retrieval and summarization improve veracity classification over claim-only models, while constraints and retrieval stability expose important limits.

5.1 Constrained vs. Unconstrained Search

  • Retrieval-based models significantly improve on claim-only models across nearly all four metrics in nearly all settings, including constrained retrieval.Table 2 reports veracity classification performance under different retrieval constraints.Source: §5.1 · p. 6, Table 2 · p. 6
  • Adding temporal or non-fact-checking-site constraints dramatically reduces performance, consistent with unconstrained search retrieving post-claim fact-checks containing synthesized evidence.The authors argue future retrieval research should use a constrained setting.Source: §5.1 · pp. 5–6

5.2 Stage Ablations

  • Replacing generated subquestions with the original claim for first-stage web search notably decreases performance.The authors suggest multiple subquestions help search across separate aspects of a claim; gold subquestions performed worse than predicted ones, possibly because the system predicts more questions.Source: §5.2 · p. 6, Table 3 · p. 7
  • Second-stage retrieval query variants yield only slight, statistically insignificant performance differences, even when using human-written justifications.The authors attribute this to expanding retrieved text by a context window of ±150 words, which can make retrieval precision less important.Source: §5.2 · p. 6, Table 3 · p. 7
  • Feeding retrieved text directly to the classifier without claim-focused summarization gives the worst performance across all metrics.The authors suggest long inputs can exceed DeBERTa’s context window or make relevant information harder for the classifier to identify; zero-shot and few-shot prompts have similar classification results.Source: §5.2 · p. 6, Table 3 · p. 7

5.3 Stability of First-stage Retrieval

  • URL-set Jaccard similarity declines across retrieval rounds, but the change does not significantly affect veracity-assessment efficacy.The authors caution that growing differences in documents returned by the Bing Search API challenge consistent benchmarking and advocate a public document-set benchmark.Source: §5.3 · p. 7, Table 4 · p. 7

6 Human Evaluation of Summaries

Human evaluation examines whether claim-focused summaries are faithful to their source documents and useful for answering claim-related questions. Few-shot summaries are mostly factual, but comprehensive coverage of complex claims remains difficult.

6.1 Faithfulness Evaluation

  • Few-shot summaries substantially reduce hallucinations, and 89% are rated factual or containing only minor errors, making them good enough as classifier evidence.The study evaluates 200 document-summary pairs; annotator agreement is reported as Fleiss Kappa 0.30, with disagreements often reflecting subjectivity about factual-error severity.Source: §6.1 · p. 8, Table 5 · p. 8

6.2 Comprehensiveness Evaluation

  • Zero-shot summaries yield more answerable questions than few-shot summaries, but the authors suggest this may reflect unsupported information imputed by hallucinations.Annotators distinguish answerable, partially answerable, and unanswerable subquestions.Source: §6.2 · pp. 8–9
  • Few-shot summaries partially address over 60% of gold subquestions derived from PolitiFact justifications, despite the system lacking access to those often subtle factors.The evaluation uses human-annotated yes/no subquestions as a proxy for summary comprehensiveness.Source: §6.2 · pp. 8–9

6.3 Combined Evaluation

  • Thirty-six of 50 claims contain at least one unanswerable question, while only 1 of the 14 fully addressable claims has a major summary error.The claim-level analysis considers faithfulness and comprehensiveness together.Source: §6.3 · p. 9, Table 7 · p. 9

7 Related Work

Related work spans retrieval-augmented NLP, question decomposition, and automated fact-checking. The paper distinguishes its setting from approaches that assume a fixed corpus, fact-checker-provided evidence, or unconstrained web retrieval.

Retrieval augmented models

  • Retrieval augmentation has been studied for question answering, generation, language modeling, and dialogue, often assuming access to a fixed corpus.For real-world fact-checking, the authors note that no such corpus exists and describe their use of the Bing Search API as following WebGPT (Nakano et al., 2021).Source: §7 · p. 9

Question decomposition

  • Question decomposition supports complex question answering and retrieval, and Chen et al. (2022a) showed it can also aid evidence retrieval for complex-claim verification.The passage also cites concurrent work on fact-checking generated text and Wikipedia.Source: §7 · p. 9

8 Conclusion

The conclusion presents a pipeline for realistic automated fact-checking of complex political claims using raw web evidence. It reports improved accuracy from integrating retrieved evidence, while emphasizing incomplete evidence coverage and the need for human involvement.

8 Conclusion

  • The pipeline checks complex political claims by retrieving raw web evidence and improves final fact-checking accuracy by integrating that evidence.The authors report promising results on CLAIMDECOMP.Source: §8 · p. 9
  • Web search often fails to surface every piece of information needed to verify a claim, motivating a human-in-the-loop fact-checking system.The conclusion frames real-world evidence retrieval as a continuing challenge.Source: §8 · p. 9

Limitations and Future Directions

The pipeline is bottlenecked by first-stage retrieval, and some claims require evidence unavailable on the web outside fact-checking sites. The work covers English-language political claims and proposes a human-in-the-loop system as a future direction.

Limitations and Future Directions

  • 36.0% of questions were unanswerable using the most faithful claim-focused summaries, with failures tied to unavailable web information or inadequate generated subquestions.Some claims require talking to or emailing specific people, which is beyond this work's scope; stronger question-generation or decontextualization models may help in other cases.Source: Limitations and Future Directions · p. 10
  • A proposed human-in-the-loop system would let fact-checkers rerun searches and update summaries when retrieved documents are unsatisfactory.The authors also envision learning from fact-checker feedback, but developing such systems is outside this work's scope.Source: Limitations and Future Directions · p. 10
  • The study addresses only English-language political claims, leaving other languages and claim types such as image- or meme-based social-media claims for future work.The authors suggest the decomposition and retrieval approach could play a role in systems for these other claim types.Source: Limitations and Future Directions · p. 10

A Experimental Details

The appendix records retrieval and prompting details, classifier settings, pipeline compression measurements, and examples of summary errors. Its examples distinguish minor errors that preserve interpretation from more serious errors that alter it.

A.1 List of Websites being Filtered

  • The retrieval filter excludes fact-checking URLs, PDFs, and videos; examples show generated questions for oil-export, Texas-ballot, and police-training claims.The listed sites include the Washington Post fact-checker, AP fact-check, and Reuters fact-check pages.Source: §A.1 · pp. 13–14

A.2 Question Generation Prompt and Deduplication

  • Generated subquestions are deduplicated by exact string match.The paper shows a few-shot question-generation prompt in Figure 6.Source: §A.2 · p. 14, Figure 6

A.3 Question-focused Summarization Prompt

  • The appendix provides zero-shot and few-shot prompts for generating claim-focused summaries.The prompts ask for a concise summary of document information relevant to the claim.Source: §A.3 · p. 14, §A.4 · p. 14, Figure 8 · p. 15

B Information Compression through the Pipeline

A.4 Hyperparameters of Veracity Classifier

  • The classifier settings specify a maximum sequence length of 512, a 3e-5 initial learning rate, Adam with linear decay, and MAE for dev-model selection.Training used an NVIDIA-A6000.Source: §A.4 · p. 14

C.1 Examples of Unfaithful Summaries

  • Minor summary errors preserve the original document's interpretation, whereas major and completely wrong errors alter it.Figure 12 presents one example of each error category, with mismatches marked in red.Source: §C.1 · p. 14

C.2 Recruiting Process

The appendix describes recruitment for the faithfulness and comprehensiveness studies and includes the claim-scoring prompt used with summaries. It notes that ChatGPT may rely heavily on prior knowledge rather than use the supplied summary effectively.

C.2 Recruiting Process

  • The faithfulness study selected 15 of 31 MTurk workers who passed a five-example qualification test and provided reasonable rationales.Workers received $3 for the qualification test and $2 for a HIT containing four document-summary pairs.Source: §C.2 · p. 15
  • The comprehensiveness study selected 17 of 28 MTurk workers who passed a ten-example qualification test and provided reasonable rationales.The qualification test required more than 7/10 correct answers; the task payment is only partially specified in the passage.Source: §C.2 · p. 15
  • The ChatGPT prompt asks for a 0-to-100 truthfulness score using document summaries, but ChatGPT may rely heavily on prior knowledge instead of using them effectively.The prompt defines 0 as “pants-on-fire” and 100 as true.Source: §C.2 · p. 16

D Using LLMs as a Veracity Classifier

ChatGPT is prompted to score claim truthfulness from summaries, with scores mapped onto the dataset's six labels. On the reported test-set comparison, adding summaries lowers ChatGPT performance unlike DeBERTa, and few-shot prompts do not improve its performance.

D Using LLMs as a Veracity Classifier

  • ChatGPT explains its reasoning and assigns a 0-to-100 truthfulness score, which is mapped to discrete labels using training-set label percentiles.The score endpoints are 0 for false and 100 for true; examples are ranked and mapped to the label distribution of the training set.Source: D Using LLMs as a Veracity Classifier · p. 16
  • Adding summaries causes a notable performance drop for ChatGPT, although its claim-only performance is slightly better than DeBERTa's.Table 9 reports test-set veracity-classification performance for ChatGPT prompts; the passage gives no numerical values.Source: D Using LLMs as a Veracity Classifier · p. 16, Table 9 · p. 17
  • Few-shot prompts did not yield better performance than the zero-shot prompt.Source: D Using LLMs as a Veracity Classifier · p. 16

Related papers

Loading 2305.11859v2…