PaperlayerView exact revision on arXiv

Source-linked AI summary

Fact or Fiction: Verifying Scientific Claims

David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, Hannaneh Hajishirzi

arXiv:2004.14974v6cs.CL

TL;DR

Scientific literature grows too quickly for researchers and the public to reliably assess claims, while scientific verification requires domain knowledge and specialized reasoning. The paper formalizes a claim-verification task, builds SCIFACT and VERISCI to retrieve and classify evidence with rationales, and demonstrates reasonable efficacy on COVID-19 claims. It positions the dataset and models as resources for research on retrieving and reasoning over specialized scientific evidence.

  • Problem

    Rapidly expanding literature makes scientific claims difficult to evaluate, and verification requires domain knowledge, directional reasoning, and specialized statistical interpretation.

  • Method

    The paper constructs SCIFACT from 1,409 expert-annotated claims and develops VERISCI to retrieve abstracts, select rationale sentences, and predict SUPPORTS, REFUTES, or NOINFO labels.

  • Results

    VERISCI was judged plausible for 23 of 36 COVID-19 claims, indicating it can retrieve and classify evidence for many real-world claims.

  • Takeaways & Limitations

    SCIFACT provides a challenging testbed and the released models and data support research on scientific fact-checking with specialized domain knowledge.

  • Takeaways & Limitations

    Scientific claim verification still requires incorporating background information and assessing the strength and provenance of evidence sources, especially beyond the research literature.

Abstract

from arXiv · show

We introduce scientific claim verification, a new task to select abstracts from the research literature containing evidence that SUPPORTS or REFUTES a given scientific claim, and to identify rationales justifying each decision. To study this task, we construct SciFact, a dataset of 1.4K expert-written scientific claims paired with evidence-containing abstracts annotated with labels and rationales. We develop baseline models for SciFact, and demonstrate that simple domain adaptation techniques substantially improve performance compared to models trained on Wikipedia or political news. We show that our system is able to verify claims related to COVID-19 by identifying evidence from the CORD-19 corpus. Our experiments indicate that SciFact will provide a challenging testbed for the development of new systems designed to retrieve and reason over corpora containing specialized domain knowledge. Data and code for this new task are publicly available at https://github.com/allenai/scifact. A leaderboard and COVID-19 fact-checking demo are available at https://scifact.apps.allenai.org.

1 Introduction

The paper introduces scientific claim verification to help evaluate rapidly changing research claims, especially during public-health crises. It builds SCIFACT and baseline systems that retrieve evidence, select rationales, and classify claim–abstract relations, including a COVID-19 case study.

  • Motivation: Rapidly expanding scientific literature makes it difficult to track current findings, creating a need for automated claim-verification tools.The challenge is especially acute during COVID-19 because findings appear quickly and decisions may rely on incomplete or outdated information.
  • Challenges: Scientific claim verification requires background knowledge, directional reasoning, and interpretation of specialized statistical language.The paper illustrates these requirements with troponin, higher-versus-lower relationships, and p < 0.001.
  • Dataset: SCIFACT contains 1,409 expert-annotated scientific claims paired with supporting or refuting abstracts and rationale annotations.Claims are reformulated from naturally occurring citation sentences, while citation links indicate documents likely to contain relevant evidence.
  • Baseline model: VERISCI retrieves claim-related abstracts, selects rationale sentences with a BERT-based model, and labels abstracts SUPPORTS, REFUTES, or NOINFO.The baseline benefits from training on claims from domains including Wikipedia and politics.
  • COVID-19 case study: Expert assessment judged VERISCI evidence plausible for 23 of 36 COVID-19 claims from the CORD-19 corpus.The case study evaluates whether retrieved evidence abstracts contain reasonable rationales and labels.
  • Contributions: The paper contributes a formalized task, an annotated dataset, baseline models, and a real-world COVID-19 verification demonstration.It identifies substantial opportunities for improvement across the modeling pipeline.

2 Background and task definition

The paper defines scientific claim verification as finding research-literature evidence that supports or refutes a claim. SCIFACT focuses on claim–abstract relations with expert-justified rationales rather than assigning a global truth label across a corpus.

  • Task definition: Scientific claim verification identifies research-literature evidence that SUPPORTS or REFUTES a given scientific claim.The task evaluates claim veracity against a scientific corpus.
  • Scientific claims: A SCIFACT scientific claim is an atomic, verifiable statement about one aspect of a scientific entity or process.Valid claims can be verified from a single source, whereas opinions and compound statements are excluded or split.
  • Caveat: The paper cautions that its model is a research prototype and must not be used for medical decisions.
  • Claim sources: SCIFACT claims are naturally derived from citation sentences, unlike synthetic claims created by mutating Wikipedia sentences in FEVER.This construction keeps discussed topics representative of the research literature.
  • Scope: Assigning a global truth label for a fixed scientific corpus would require systematic review by expert teams.The paper therefore studies the simpler problem of labeling individual claim–abstract pairs SUPPORTS or REFUTES.
  • Rationales: Each SUPPORTS or REFUTES relation must include at least one rationale consisting of minimal sentences that imply the claim to a domain expert.Rationales support interpretable predictions by identifying the sentences necessary for each decision.

3 The SCIFACT dataset

SCIFACT is an expert-annotated dataset of scientific claims and evidence abstracts, constructed from citation contexts across biomedical research and labeled with supporting or refuting rationales.

  • Dataset overview: 1,409 scientific claims are verified against a corpus of 5,183 abstracts.
  • Data source and corpus construction: SCIFACT samples highly cited articles from curated journals spanning basic science and clinical medicine, then expands the corpus with co-cited abstracts.The seed set requires at least 10 citations; co-cited articles are added when citances reference articles outside the seed set.
  • Claim writing: Claims are written from source citances without exposing annotators to the cited abstracts, producing naturally occurring claims tied to specific scientific findings.Annotators write up to three claims and skip citances that do not state specific findings.
  • Claim negation: An NLP expert creates refuting examples by reversing claim effects without relying on obvious negation triggers.The claim-only model’s poor performance suggests this process did not introduce severe artifacts.
  • Annotation and quality: Annotators label each cited abstract as SUPPORTS, REFUTES, or NOINFO and mark rationale sentences; evidence appears in 63% of cited abstracts.Independent re-annotation produced Cohen’s κ of 0.75 for labels and 0.71 for rationales.

4 The SCIFACT task

SCIFACT frames scientific claim verification as retrieving evidence abstracts, assigning SUPPORTS, REFUTES, or NOINFO labels, and selecting rationale sentences that justify those decisions.

  • Task formulation: The task takes a scientific claim and an abstract corpus, then labels every abstract as SUPPORTS, REFUTES, or NOINFO.
  • Evidence representation: Evidence abstracts are those that support or refute a claim, and each evidence abstract may contain one or more annotated rationales.
  • Task formulation: A system must retrieve evidence abstracts and predict both a relation label and rationale sentences for each retrieved abstract.Predicted rationale sentences may cover multiple gold rationales in one collection.
  • Task evaluation: Evaluation covers abstract-level and sentence-level performance in both retrieved-evidence and oracle-abstract settings.Abstract-level metrics assess evidence identification and labeling, while sentence-level metrics assess rationale selection and labeling.
  • Task evaluation: Abstract-level evaluation limits predictions to three rationale sentences and measures micro-F1 over correctly labeled and rationalized evidence abstracts.

5 VERISCI: Baseline model

VERISCI is a pipeline baseline that retrieves abstracts, selects rationale sentences, and predicts claim relations using separate model components.

  • Pipeline architecture: VERISCI retrieves candidate abstracts, selects rationale sentences, and predicts SUPPORTS, REFUTES, or NOINFO labels for each abstract.
  • Abstract retrieval: Abstract retrieval ranks corpus abstracts by TF-IDF similarity to the claim.
  • Rationale selection: Rationale selection scores each abstract sentence after encoding it with the claim using a BERT-style language model.
  • Label prediction: Label prediction encodes the predicted rationale sentences with the claim and applies a three-output classifier for SUPPORTS, REFUTES, and NOINFO.NOINFO is predicted when no rationale sentences are selected.
  • Baseline analysis: A Longformer model encoding entire abstracts was not competitive with the pipeline setup, likely because relevant information was harder to identify in full abstracts.

6 Experiments

The experiments evaluate VERISCI’s components, training data, encoders, inputs, and full-pipeline performance on SCIFACT. Results show benefits from in-domain training, complementary component errors, and science-specific reasoning challenges.

  • Experimental setup: The experiments compare training datasets, sentence encoders, and model inputs for rationale selection and label prediction on SCIFACT.The claim-only model cannot select rationales, while the abstract-only model lacks the claim during prediction.
  • Pipeline components: Training first on FEVER and then fine-tuning on SCIFACT achieves the best label-prediction performance.In 36 of 44 relevant cases, the SCIFACT-only model predicts NOINFO, suggesting improved recognition of non-domain-specific entailment cues after pretraining.
  • Pipeline components: SCIFACT-only training produces the best rationale-selection results, while additional out-of-domain data provides little benefit.The missed rationales generally contain science-specific vocabulary.
  • Pipeline components: RoBERTa-large performs best for label prediction, whereas SciBERT has a slight edge for rationale selection.The claim-only and abstract-only models perform poorly, indicating that prediction benefits from access to the relevant inputs.
  • Full task: Roughly 70% end-to-end classification accuracy is implied in the oracle-abstract setting when gold rationales are provided.Oracle abstract rows provide gold evidence abstracts, while oracle rationale rows provide gold rationales as input.
  • Full task: 47% relative improvement occurs on open SentenceSelection+Label, and 28% on open AbstractLabel+Rationale, over FEVER alone.The three pipeline components make similar contributions to overall model error; oracle rationale and retrieval replacements each add roughly 20 F1 points in the reported comparisons.
  • Robustness: 10,000 bootstrap-resampled test sets indicate that observed performance differences are statistically robust rather than attributable to random dataset variation.The robustness analysis computes standard deviations for the reported performance metrics.
  • Full task: The baseline requires science background, directionality, numerical reasoning, cause-and-effect reasoning, and coreference to verify difficult claims.These reasoning types are identified from claims misclassified by the modeling baseline.

7 Related work

Related work places scientific claim verification among fact-checking, rationalized NLP, citation contextualization, biomedical question answering, and automated evidence synthesis. The paper distinguishes SCIFACT by requiring verification of rewritten scientific claims against evidence.

  • Fact checking: Existing fact-checking datasets primarily address political news, social media, and web misinformation, while scientific claim verification previously lacked an established dataset.The paper calls the task claim verification to emphasize helping researchers interpret scientific findings rather than countering disinformation.
  • Rationalized NLP: SCIFACT belongs to rationalized NLP because models must justify predictions with rationales drawn from source documents.ERASER includes SCIFACT among benchmark datasets for evaluating rationalized NLP models.
  • Scientific NLP: Unlike SCIFACT, citation contextualization identifies relevant spans for citations without rewriting them into atomic claims, making verification more difficult.BioMedSumm reported 21.7% inter-annotator agreement among expert annotators.
  • Question answering: Biomedical question answering datasets such as BioASQ and PubMedQA contain 855 and 1,000 yes/no questions, respectively.These tasks, like claim verification, require knowledge-intensive matching between queries and supporting text.
  • Evidence synthesis: The authors position claim-verification systems as potential components of future automated evidence-synthesis frameworks.Evidence synthesis seeks to automate systematic-review processes, including extracting PICO snippets and inferring clinical-trial outcomes.

8 Conclusion and future work

The paper concludes that scientific claim verification, SCIFACT, and VERISCI support tracing and assessing scientific claims, including real-world COVID-19 claims. Future work must address richer background reasoning and evidence provenance across broader sources.

  • Conclusion: SCIFACT and VERISCI are released to support scientific claim verification and research on tracing the sources and veracity of scientific claims.The authors connect these abilities to the COVID-19 pandemic and the broader reproducibility crisis.
  • Conclusion: VERISCI can be trained and deployed with reasonable efficacy on real-world COVID-19 claims.The conclusion describes this as a result of the paper’s models and resources.
  • Future work: Future systems should incorporate background information, reason about scientific processes, and assess the strength and provenance of evidence sources.The authors identify these as promising research directions for scientific claim verification.
  • Future work: Evidence provenance becomes especially important when verifying claims against sources beyond the research literature, such as social media and news.The paper frames this as a challenge for future work rather than a demonstrated capability of the current system.

A.2 Training the RATIONALESELECTION module

The rationale-selection module is trained across several datasets and transformer encoders with tuned optimization settings. The training procedure uses dataset-specific learning rates, epochs, and rationale inputs for NOINFO examples.

  • Training setup: The experiments train rationale-selection models on SCIFACT, FEVER, and UKP Snopes, including FEVER pretraining followed by SCIFACT fine-tuning.The models use SCIBERT, BioMedRoBERTa, RoBERTa-base, and RoBERTa-large as sentence encoders.
  • Optimization: SCIFACT training uses a 1e-5 learning rate for the transformer base and 1e-3 for the linear layer.For FEVER plus SCIFACT, the learning rate is 1e-5 for the entire model during pretraining and fine-tuning.
  • Optimization: For FEVER and UKP Snopes models, the transformer-base learning rate is 5e-6 and 1e-5, respectively.The corresponding linear-layer rates are 5e-5 for FEVER and 1e-4 for UKP Snopes.
  • Rationale inputs: NOINFO training uses the k most similar abstract sentences as rationales, with k sampled uniformly from {0, 1}.This procedure is used when training on claim/cited-abstract pairs labeled NOINFO.
  • Compute: Rationale-selection training takes about 150 minutes on SCIFACT, 120 minutes on UKP Snopes, and 700 minutes on FEVER on a single Nvidia P100 GPU.The reported runs use 20 SCIFACT epochs, 5 UKP Snopes epochs, and 3 FEVER epochs.
  • Hyperparameters: Learning rate, batch size, and epoch count are manually tuned using development-set F1, with batch sizes of 64 and 256 tested.Epochs are stopped after model convergence.

B Statistical analysis

The paper quantifies uncertainty in its main results with bootstrap resampling and reports corresponding test- and development-set statistics. It also summarizes the journal distribution of cited abstracts.

  • Uncertainty estimation: 10,000 bootstrap-resampled test sets quantify uncertainty for metrics computed on the 300-claim test set.Each resampled set contains 300 claims sampled uniformly with replacement.
  • Statistical conclusion: The conclusion that SCIFACT training improves performance remains robust to the uncertainties reported in the bootstrap analyses.This conclusion is based on the uncertainty estimates summarized for the main results.
  • Reported statistics: Table 6 reports bootstrap means and standard deviations for the test-set metrics, while Table 7 reports development-set metrics.Table 6 standard deviations are shown as subscripts, and sampling variability can slightly change reported means.
  • Corpus statistics: Table 8 counts cited abstracts by journal, with co-cited abstracts from journals outside the curated set grouped under “Other”.The “Other” category therefore includes documents from journals not in the predefined journal set.

C Dataset collection and corpus statistics

The dataset combines citation-derived scientific claims with evidence abstracts, distractors, and expert-supervised annotation procedures. Claims are rewritten into atomic statements, while verification annotations identify support, refutation, or rationale evidence.

  • Corpus statistics: Cited abstracts are counted by source journal, with co-cited articles from journals outside the curated set categorized as “Other”.The journal breakdown is reported in Table 8.
  • Corpus construction: Distractor abstracts enlarge the corpus while preserving topical similarity and reducing retrieval shortcuts based on lexical overlap.They are co-cited in the same articles as evidence abstracts, but were omitted from the source citances used to generate claims.
  • Claim writing: Citances are rewritten as atomic factual claims by removing irrelevant material and expanding acronyms such as CVD to cardiovascular disease.Figure 4 illustrates this conversion for a citance concerning atherosclerotic cardiovascular disease and IL-6.
  • Evidence annotation: A claim may use multiple rationales from one abstract when each rationale independently provides sufficient evidence for verification.Figure 5 shows this annotation pattern.
  • Annotation quality control: Student claim writers were trained and monitored by experts, while verification annotators passed a quiz and underwent reliability review before peer annotation.Low-quality claims were rewritten, and graduate students generally reviewed undergraduate verification annotations.
Loading 2004.14974v6…