luoojason's picture
Add files using upload-large-folder tool
daa6e2d verified
|
Raw
History Blame Contribute Delete
115 kB
Fact or Fiction: Verifying Scientific Claims
DavidWadden†∗ ShanchuanLin† KyleLo‡ LucyLuWang‡
MadeleinevanZuylen‡ ArmanCohan‡ HannanehHajishirzi†‡
† UniversityofWashington,Seattle,WA,USA
AllenInstituteforArtificialIntelligence,Seattle,WA,USA
{dwadden,linsh,hannaneh}@cs.washington.edu
{kylel,lucyw,madeleinev,armanc}@allenai.org
Abstract
Claim
We introduce scientific claim verification, a Cardiac injury is common in
new task to select abstracts from the re- critical cases of COVID-19.
searchliteraturecontainingevidencethatSUP-
PORTS or REFUTES a given scientific claim, Fact-checker
----- ---------- --- ------- ---------- ------ --- --- --- ------------ --- --- ---
Corpus
and to identify rationales justifying each de-
------- -------- ---------- ----- ------------ -------- --- --- --- --------- --- -------- ---
Decision: SUPPORTS
cision. To study this task, we construct SCI-
More severe COVID-19 infection
FACT, a dataset of 1.4K expert-written scien-
------------ --------- ------ ------------------------ --- ------ --- --- ------------------------- --- ----------------------- --- -----
is associated with higher mean
tific claims paired with evidence-containing
troponin (SMD 0.53, 95% CI 0.30
abstracts annotated with labels and rationales. to 0.75, p < 0.001)
WedevelopbaselinemodelsforSCIFACT,and Rationale
------------------------------------- --- ---- ------------- --- ---------- ------ --------------- --- ------ --------- ----------- ---
demonstrate that simple domain adaptation
Figure 1: A scientific claim, supported by evidence
techniquessubstantiallyimproveperformance
compared to models trained on Wikipedia or identifiedbyoursystem. Tocorrectlyverifythisclaim,
the system must possess background knowledge that
--------- ------ ------ ------- ----------- --------- ---------- ------------ ------- ---------- ------- ---------- ----
political news. We show that our system is
troponin is a protein found in cardiac muscle and that
able to verify claims related to COVID-19 by
identifying evidence from the CORD-19 cor- elevated levels of troponin are a marker of cardiac
injury. In addition, it must be able to reason about di-
---- --------------- --- -------- ---- ------- ------- ------------ --- ------- ------- ------------ ---
pus. Our experiments indicate that SCIFACT
will provide a challenging testbed for the de- rectionalrelationshipsbetweenscientificprocesses: re-
placing higher with lower would cause the rationale
--- --- --- --- --- --- ------- ------ ---- ----------- ----- ------------- ---
velopmentofnewsystemsdesignedtoretrieve
to REFUTE theclaimratherthan SUPPORT it. Finally,
--- --- --- --- --- --- --------- ------------------ --- --- ------- --- --------
andreasonovercorporacontainingspecialized
domain knowledge. Data and code for this thesystemshouldinterpret p<0.001asanindication
thatthereportedfindingisstatisticallysignificant.
new task are publicly available at https://
--------------------------- --- -------- ------------- --- --------- --- --- --- --- --- --- ---
github.com/allenai/scifact. A leader-
board and COVID-19 fact-checking demo
are available at https://scifact.apps. Fact-checking – a task in which the veracity
allenai.org.
of an input claim is verified against a corpus of
--- --- --- --- --- --- ----------- ----- ------- -------- ------- --------- -----
documents that support or refute the claim – has
1 Introduction
beenstudiedtocombattheproliferationofmisin-
Due to rapid growth in the scientific literature, it formationinpoliticalnews, socialmedia, andon
is difficult for researchers – and the general pub- the web (Thorne et al., 2018; Hanselowski et al.,
2019). However,verifyingscientificclaimsposes
licevenmoreso–tostayuptodateonthelatest
findings. Thischallengeisespeciallyacuteduring new challenges to both dataset construction and
publichealthcriseslikethecurrentCOVID-19pan- effectivemodeling. Whilepoliticalclaimsareread-
ilyavailableonfact-checkingwebsitesandcanbe
demic,duetotheextremelyfastrateatwhichnew
findingsarereportedandtherisksassociatedwith verifiedbycrowdworkers,annotatorswithexten-
makingdecisionsbasedonoutdatedorincomplete sive domain knowledge are required to generate
andverifyscientificclaims.
information. As a result, there is a need for auto-
------------ --- --------- ----- --------- --------- --- --- --- --- --- --- ---
matedtoolstoassistresearchersandthepublicin In addition, NLP systems for scientific claim
evaluatingtheveracityofscientificclaims. verificationmustpossessadditionalcapabilitiesbe-
yond those required to verify factoid claims. For
--- --- --- --- --- --- ---------- -------- --- --------- ------- ------- ---
∗WorkperformedduringinternshipwiththeAllenInsti-
tuteforArtificialIntelligence. instance,toverifytheclaimshowninFigure1,a
7534
Proceedingsofthe2020ConferenceonEmpiricalMethodsinNaturalLanguageProcessing,pages7534–7550,
November16–20,2020.(cid:13)c2020AssociationforComputationalLinguistics

Claim1:Lopinavir/ritonavirhaveexhibitedfavorableclinicalresponseswhenusedasatreatmentforcoronavirus. Supports:...Interestingly,afterlopinavir/ritonavir(Kaletra,AbbVie)wasadministered,β-coronavirusviralloadssignificantly decreasedandnoorlittlecoronavirustiterswereobserved. Refutes:Thefocuseddrugrepurposingofknownapproveddrugs(suchaslopinavir/ritonavir)hasbeenreportedfailedfor curingSARS-CoV-2infectedpatients.Itisurgenttogeneratenewchemicalentitiesagainstthisvirus... Claim2:Thecoronaviruscannotthriveinwarmerclimates. Supports:...mostoutbreaksdisplayapatternofclusteringinrelativelycoolanddryareas...Thisisbecausetheenvironment canmediatehuman-to-humantransmissionofSARS-CoV-2,andunsuitableclimatescancausethevirustodestabilizequickly... Refutes: ...significantcasesinthecomingmonthsarelikelytooccurinmorehumid(warmer)climates,irrespectiveofthe climate-dependenceoftransmissionandthatsummertemperatureswillnotsubstrantiallylimitpandemicgrowth. Table1: EvidenceidentifiedbyoursystemassupportingandrefutingtwoclaimsconcerningCOVID-19. system must have the ability to access scientific CORD-19corpus(Wangetal.,2020). Expertanno- backgroundknowledge,reasonoverincreasesand tatorsjudgeretrievedevidencetobeplausiblefor 23of36claims.1 decreasesinquantitiesormeasurements,andmake Ourresultsandanalysesdemon- senseofspecializedstatisticallanguage. stratetheimportanceofthenewtaskanddatasetto In this paper, we introduce the task of scien- supportsignificantfutureresearchinthisdomain. tific claim verification to evaluate the veracity of Insummary,ourcontributionsinclude: (1)We scientific claims against a scientific corpus. Ta- introduceandformalizethescientificclaimverifi- ble 1 presents some examples. To facilitate re- cation task. (2) We develop a novel annotation search on this task, we construct SCIFACT, an protocol to generate and verify 1.4K naturally- expert-annotateddatasetof1,409scientificclaims occurringclaimsaboutscientificfindings. (3)We accompanied by abstracts that support or refute establish strong baselines on this task, and iden- eachclaim,andannotatedwithrationales(Leietal., tify substantial opportunities for improvement at 2016)justifyingeachSUPPORTS/REFUTESdeci- allstagesofthemodelingpipeline. (4)Wedemon- sion. Tocreatethedataset,wedevelopanovelan- stratetheefficacyofoursysteminareal-worldcase notationprotocolinwhichannotatorsre-formulate studyverifyingclaimsaboutCOVID-19againstthe

naturallyoccurringclaimsinthescientificliterature researchliterature.
–citationsentences–intoatomicscientificclaims.
2 Backgroundandtaskdefinition
--- --- --- --- --- --- ----------------------------- --- --- --- ---
Usingcitationsentencesasasourceofclaimsboth
speedstheclaimgenerationprocessandguarantees
As illustrated in Figure 1, scientific claim verifi-
------------------------ --- --- ------------------- --- --- -------------- --------- --- ---------------- -------
thatthetopicsdiscussedin SCIFACTarerepresen-
cationisthetaskofidentifyingevidencefromthe
tativeoftheresearchliterature. Inaddition,citation
------------------------------ --- --- ------------------- --- --- ---------------------- --- -------- --- ---------
researchliteraturethat SUPPORTS or REFUTES a
linksindicatetheexactdocumentslikelytocontain
given scientific claim. Table 1 shows the results
--- --- --- --- --- --- ---------------- ------ ----- ------- -----------
evidencenecessarytoverifyagivenclaim.
of our system applied to claims about the novel
--- --- --- --- --- --- ------------- ------- --- ------------ ---------
WeestablishperformancebaselinesonSCIFACT
coronavirus COVID-19. For each claim, the sys-
--- --- --- --- --- --- ----------- --------- --- --------------- --------
withanapproachsimilartoDeYoungetal.(2020a),
temidentifiesrelevantscientificabstracts,andla-
whichachievesstrongperformanceontheFEVER
bels the relation of each abstract to the claim as
------------------------------------------- ------------- ------ ----- --------- --- ------------------------ ----------- ---- ------------------- --------
claimverificationdataset(Thorneetal.,2018). Our
eitherSUPPORTSorREFUTES. Verifyingscientific
baseline is a pipeline system which retrieves ab-
claimsischallengingandrequiresdomain-specific
stracts related to an input claim, uses a BERT-
------- ------- ----------- ----------- --- ----- --- --- --- --- ---
backgroundknowledge–forinstance,inorderto
based(Devlinetal.,2019)sentenceselectortoiden-
identify the evidence supporting Claim 1 in Ta-
-------------- ---------- --- ---------- ------------- --- -------- ------------ ---------- ----- --------
tify rationale sentences, and labels each abstract
ble1,thesystemmustdeterminethatareductionin
as SUPPORTS, REFUTES,or NOINFO withrespect
------------ ---------- --- ------ ----------- --- --- --- --- --- ---
coronavirusviralloadindicatesafavorableclinical
to the claim. We demonstrate that our baseline
------------- --- ----------- ---- ------------ --- --- --- --- --- ---
response,eventhoughthisfactisnevermentioned.
canbenefitfromtrainingonclaimsfromdomains
ScientificclaimsIn SCIFACT,ascientificclaimis
--- --- --- --- --- --- ------------------ --- -------------------------- --- ---
includingWikipediaarticlesandpolitics.
We showcase the ability of our model to ver- anatomicverifiablestatementexpressingafinding
ify expert-written claims concerning the novel
------------------ --- ------ ---------- --- ----- --- --- --- --- ---
1Weemphasizethatourmodelisaresearchprototypeand
coronavirusCOVID-19againstthenewly-released shouldnotbeusedtomakeanymedicaldecisionswhatsoever.
7535
about one aspect of a scientific entity or process,
Corpus
source.2 N=601
--------- ----------- --- ---- -------- -------- --- --- ----- --- --- --- --- ---
which can be verified from a single For
Citing
instance,“TheR ofthenovelcoronavirusis2.5”
-------------- --- --------------------------- --- --- --- --- --- --- --- --- --- --- ---
0
is valid, but opinion-based statements like “The Seed
--------- ----------------- --- --- ---------- ---- ---- --- --- ---- --- --- --- ---
governmentshouldrequirepeopletostandsixfeet
N=140
apart to stop coronavirus” are not. Compound
-------- ----------------- --- --- -------- -------- --- --- --- --- --- --- --- ---
claimslike“Aerosolizedcoronavirusdropletscan
travelatleast6feetandcanremainintheairfor3 Co-
------------------------------------------ --- --- --- --- --- --- --- --- --- --- --- --- ---
cited
hours”shouldbesplitintotwoatomicclaims.
Claim
Claims in SCIFACT are natural – they are de- N=4,259
------ ---------- --- --- ------- ------ ------- --- ------- --- --- --- --- ---
rivedfromcitationsentences,orcitances(Nakov Cardiac injury is
common in critical
et al., 2004), that occur naturally in scientific ar- Dis-
-------------- ---- ----- --------- --- ---------- --- --- --- ---- --- --- --- ---
cases of COVID-19.
tractor
ticles. This is similar to political fact-checking
------------ --- ------- ------------ --- ------------- --- --- --- --- --- --- --- ---
datasetssuchasUKPSnopes(Hanselowskietal.,
2019),whichusepoliticalfact-checkingwebsites
Figure 2: Corpus construction. Citing abstracts are
--- --- --- --- --- --- --- ------ --- -------------------- --- ------ --------- ---
as a source of natural claims. On the other hand, identified for each seed document. A claim is written
basedonthesourcecitanceinthecitingabstract.
claimsinthepopular FEVERdataset(Thorneetal.,
------------------ --- --- ------------------------- --- --- --- --- --- --- --- --- --- ---
2018)aresynthetic,sincetheyarecreatedbyanno-
tators by mutating sentences from the Wikipedia
--------- -------- --------- --- ---- ------------- --- --- --- --- --- --- --- ---
Abstractsthatsupportorrefuteeachclaimarean-
articlesthatwillserveasevidence.
notated with rationales. We describe our corpus
---------- --- -------- --- -------- ------- ----- ------- ---- ----------- --- -------- --- ------
Supporting and refuting evidence In most fact-
creationandannotationprocess.
checkingwork,claimsareassignedaglobaltruth
labelbasedontheentiretyoftheavailableevidence. 3.1 Datasourceandcorpusconstruction
ForexampleinFEVER,theclaim“BarackObama
Toconstruct SCIFACT,weuseS2ORC(Loetal.,
---------- -------------------------------- --- --- --- --- --- ----------- --- --------------------------- --- ------ ----------- ---
wasthe44th PresidentoftheUnitedStates”canbe
2020), a publicly-available corpus of millions of
verifiedusingWikipediaasanevidencesource.
scientificarticles. Toensurethatdocumentsinour
----- ------- ---------------------------- --- --- --- --- ------------------- --- -------------------------- --- ----------- --- ------
While SCIFACT claimsareindeedverifiableas-
dataset are of high quality, we randomly sample
sertionsaboutscientificfindings,accuratelyassign-
articlesfromamanuallycuratedcollectionofwell-
ingaglobaltruthlabeltoascientificclaim(givena
regardedjournalsspanningdomainsfrombasicsci-
fixedscientificcorpus)requiresasystematicreview ence(e.g.,Cell,Nature)toclinicalmedicine(e.g.,
byateamofexperts. Inthisworkwefocusonthe
----------------- --- --- ---------------------- --- --- --- --- --- --- --- --- --- ---
JAMA,BMJ).Thefulllistofjournalsisincludedin
simplertaskofassigning SUPPORTSor REFUTES
---------------------- --- --- ---------- --- ------- --- ------------ --- ------------------------------- --- --- --- ---
AppendixC.1. Werestricttoarticleswithatleast
relationstoindividualclaim-abstractpairs. 10 citations. The resulting collection is referred
Each SUPPORTS or REFUTESrelationbetween to as our seed set. We use the S2ORC citation
claimandabstractmustbejustifiedbyatleastone
graph to sample source citances from citing arti-
---------- ----------- --- ---- ------- ---------- --- ------------------------------- --- ------------- -------- --------------- ------ -----
rationale. A rationale is a minimal collection of
cleswhichcitetheseseedarticles. Ifacitancecites
sentenceswhich,takentogetheraspremisesinthe otherarticlesnotintheseedset,werefertothese
contextoftheabstract,canreasonablybejudgedby
asco-cited articlesandaddthemtothecorpus,as
-------------------------------- --- --- --- --- ---------- --- ------------------ --- -------------------------------- ----------------------- --- --- ---
adomainexpertasimplyingtheclaim. Rationales
depictedinFigure2. Thecontentofthecitedab-
facilitatethedevelopmentofinterpretablemodels stractsencompassesadiversearrayoftopicswithin
whichnotonlyhavetheabilitytomakelabelpre- biomedicine,asshowninFigure3. Themajority
dictions,butcanalsoidentifytheexactsentences
of citances used for SCIFACT cite only the seed
--- --- --- --- --- --- --- --- -------- -------- ------- ---- ---- --------
thatarenecessaryfortheirdecisions.
article(noco-citedarticles),aswefoundininitial
annotationexperimentsthatthesecitancestended
3 The SCIFACT dataset
----- ------- ------- --- --- --- --- --- --- --- --- --- --- ---
toyieldspecific,easy-to-verifyclaims.
To expand the corpus, we identify five papers
----------- --- ------- -------- -------- ---------- --- --- ------ ----------- --- -------- ---- ------
The SCIFACT dataset consists of 1,409 scientific
claims3 verifiedagainstacorpusof5,183abstracts. citedinthesamepaperaseachsourcecitancebut
inadifferentparagraph,andaddthesetothecor-
2Requiringannotatorstosearchmultiplesourcesincreases pusasdistractorabstracts. Theseabstractsoften
cognitiveburdenanddecreasesannotationquality.
3SCIFACTiscomparableinsizetorecentscientificdatasets has1,000 questions), and informationextraction(e.g. Sci-
for tasks such as QA (e.g. PubMedQA (Jin et al., 2019) ERC(Luanetal.,2018)has500annotatedabstracts).
7536

cationmodelperformspoorly,suggestingthatthe Humans

Mice negationprocessdidnotintroducesevereartifacts.
Animals
Female
Male 3.3 Claimverification
--- ---- --- --- --- --------------------- --- --- --- --- --- ---
Middle Aged
Aged
Adult AnnotationForeachclaim,alloftheclaim’scited
--- ----- --- --- --- ------------------------------------------- --- --- --- --- --- ---
Cell Line
Signal Transduction abstracts are annotated for evidence. Annotators
RNA
Receptors areshownasingleclaim-citedabstractpair,and
--- --------- --- --- --- ------------------------------------------ --- --- --- --- --- ---
Gene Expression Regulation
askedtolabelthepairas SUPPORTS, REFUTES,or
--- --- --- --- --- --------------------- --- --- --------- --- ---------- ---
Models
Risk Factors
NOINFO. Althoughourtaskdefinitionallowsfora
--- --- --- --- --- ------- ----------------------------------- --- --- --- --- ---
Cells
Cell Differentiation
Mutation singleclaimtobebothsupportedandrefuted(by
--- ---------- --- --- --- ------------------------------------------- --- --- --- --- --- ---
Microscopy differentabstracts)–anoccurrenceweobserveon
Infant
real-world COVID-19 claims (§6.3) this never
--- --- --- ------- ------- ---------- -------- --- ------ ------ --- ----------
0.0 0.2 0.4 0.6 0.8 1.0
Fraction of evidence abstracts
occursinourdataset. Eachclaimhasasinglelabel.
--------- --------------- --------- ------- ------- ------------------------------------ --- --- ------------------------- --- --- -----
CountsforeachlabelareshowninTable2a. Over-
Figure 3: Most frequently occurring Medical Subject
all,theannotatorsfoundevidencein63%ofcited
Headings(MeSH)terms(y-axis)amongcitedabstracts.
MeSHisacontrolledvocabularyusedforindexingar- abstracts. Iftheannotatorassignsa SUPPORTS or
ticles in PubMed. Topics range from clinical trial re- REFUTES label, they must also identify all ratio-
ports(“Humans”,“RiskFactors”)tomolecularbiology nalesasdefinedin§2. Table2bprovidesstatistics
(“CellLine”,“RNA”). onthenumberofsentencesperrationale,thenum-
berofrationalesperclaim/abstractpair,andthe
discusssimilartopicstotheevidencedocuments, number of evidence abstracts per claim. No ab-
stracthasmorethan3rationalesforagivenclaim,
increasing the difficulty of abstract retrieval and
---------- -------------- ----------- --------- --- --- --- --- --- --- --- ---
andallrationalesconsistofatmostthreesentences.
makingourmetricsmoreaccuratelyreflectthesys-
RationalesinSCIFACTaremutuallyexclusive. 28
--- --- --- --- --- ---------------------------------------- --- --- --- --- --- ---
tem’sperformanceonalargeresearchcorpus.
rationalescontainnon-contiguoussentences.
3.2 Claimwriting The verifiers included three NLP experts, five
lifescienceundergraduates,andfivegraduatestu-
AnnotationAnnotatorsareshownasourcecitance
dents studying life sciences. Annotators verified
--- --- --- --- --- -------------- --- -------------- --- ---------- --- --------
inthecontextofanarticle,andareaskedtowriteup
claimsthattheydidnotwritethemselves. Annota-
--- --- --- --- --- ------------------------------------ --- --- --- --- --- -------
tothreeclaimsbasedonthecontentofthecitance;
tionguidelinesareprovidedinAppendixD.
seeAppendixC.2foranexample. Thisresultsin
--------------------------- --- --- ------------- --- ------- ------ --- ------------ --- ------- ---------
SCIFACT claims are verified against abstracts
naturalclaimsbecausetheannotatordoesnotsee
rather than full articles since (1) abstracts can be
--- --- --- --- --- ----------- ---- -------- ----- --- --------- ------
thecitedarticle’sabstract–thecitedabstract–at
annotatedmorescalably,(2)evidenceisfoundin
the time of claim writing. Annotators are asked
-------- ----------------- ---------- --- ----- --- --- --- --- --- --- ---
theabstractinmorethan60%ofcases,and(3)pre-
toskipcitancesthatdonotmakestatementsabout
viousattemptsatfull-documentannotationsuffered
specificscientificfindings.
fromlowannotatoragreement(§7).
The claim writers included four experts with
--------- ---------------- --- ------------ ---- --- --- --- --- --- --- ---
backgroundinscientificNLP,fifteenundergradu- QualityWeassign232claim-abstractpairsforin-
dependent re-annotation. The label agreement is
------------- ------------------ --- -------- -------- --------- -------------- --- --- ----- --------- ---
ates studying the life sciences, and four graduate
0.75Cohen’sκ,comparablewiththe0.68Fleiss’
students(doctoralormedical)inthelifesciences.
Detailedinformationontheannotatortrainingpro- κ reported in Thorne et al. (2018), and 0.70 Co-
hen’sκreportedinHanselowskietal.(2019).
cess can be found in Appendix C.3. The claim- To
-------- -------------------- --- -------- ------ ------- --------- ---------- --- --- ----- ---------
measure rationale agreement, we treat each sen-
writinginterfaceisshowninAppendixD.
tenceaseitherclassifiedas“partofarationale”or
Claim negation Unless the authors of the source
-------------- ------ ----------- ------ ------ --- --- --- --- --- --- ---
“notpartofarationale”andcomputesentence-level
citance were mistaken, cited articles should pro-
--------------- ------------------ ----- --------------- -------- ---------- --------------------------- --- ---- --- --- ---
agreement. TheresultingCohen’sκis0.71.
vide supporting evidence for the claims made in
a citance. To obtain examples where an abstract
4 The SCIFACT task
REFUTESaclaim,anNLPexpertwrotenegations
of existing claims, taking precautions not to bias TaskFormulationTheinputstoourtaskareasci-
thenegationsbyusingobviouskeywordslike“not” entificclaimcandacorpusofabstractsA. Allab-
(Schusteretal.,2019;Gururanganetal.,2018). In stractsa ∈ Aarelabeledasy(c,a) ∈ {SUPPORTS,
§6.1, we demonstrate that a “claim-only” verifi- REFUTES, NOINFO } with respect to a claim c.
7537

Fold SUPPORTS NOINFO REFUTES All goldevidenceabstractforc,and(2)Thepredicted label is correct: y(c,a) = y(c,a). It is correctly Train 332 304 173 809 (cid:98) Dev 124 112 64 300 rationalized if,inaddition,thepredictedrationale Test 100 100 100 300 sentencescontainagoldrationale,i.e.,thereexists All 556 516 337 1409 somegoldrationaleR i (c,a) ⊆ S(cid:98)(c,a). (a)DistributionofclaimlabelsinSCIFACT. LikeFEVER,whichlimitsthemaximumnumber ofpredictedrationalesentencestofive, SCIFACT 0 1 2 3+ limitstothreepredictedrationalesentences. Over- Citedabstractsperclaim - 1278 86 45 all performance is measured by the micro-F1 of Evidenceabstractsperclaim 516 830 37 26 theprecisionandrecalloverthecorrectly-labeled Rationalesperabstract - 552 290 153 Sentencesperrationale - 1542 92 11 andcorrectly-rationalizedevidenceabstracts. We refertotheseevaluationsasAbstract and (b)Evidencecountsatvariouslevelsofgranularity.Forexam- Label-Only ple,Column2oftherow“Rationales/abstract”indicatesthat Abstract Label+Rationale ,respectively. 290claim/abstractpairsaresupportedby2distinctrationales. Sentence-levelevaluationmeasuresperformance Table 2: Statistics on claim labels, and the number of inidentifyingindividualrationalesentences. Un- evidenceabstractsandrationalesperclaim. like the abstract-level metrics, this evaluation pe- nalizesthepredictionofextrarationalesentences. A predicted rationale sentence s(c,a) is cor- (cid:98) The abstracts that either SUPPORT or REFUTE c rectlyselected if(1)Itisamemberofsomegold arereferredtoasevidenceabstractsforc,denoted rationaleR (c,a),(2)allothersentencesfromthe i as E(c). Each evidence abstract a ∈ E(c) is an- same gold rationale R (c,a) are among the pre- i notated with rationales. A single rationale R i is dicted S(cid:98)(c,a), and (3) y (cid:98) (c,a) (cid:54)= NOINFO 4. It is acollectionofsentences{r (c,a),...,r (c,a)}, 1 m correctly labeled if, in addition, the abstract a is wheremisthenumberofsentencesinrationaleR . i correctlylabeled: y(c,a) = y(c,a). (cid:98) We denote the set of all rationales as R(c,a) = Overallperformanceismeasuredbythemicro- {R (c,a),...,R (c,a)}. 1 n F1oftheprecisionandrecallofcorrectly-selected Given a claim c and a corpus A, the system andcorrectly-labeledrationalesentences,denoted mustpredictasetofevidenceabstractsE(cid:98)(c). For Sentence and Sentence . Selection-Only Selection+Label each abstract a ∈ E(cid:98)(c), it must predict a label Forsentence-levelevaluation,wedonotlimitthe y (cid:98) (c,a), and a collection of rationale sentences numberofpredictedrationalesentences,sincethe S(cid:98)(c,a) = {s (cid:98)1 (c,a),...,s (cid:98)(cid:96) (c,a)}. Note that al- evaluationpenalizesmodelsthatover-predict. thoughthegoldannotationsmaycontainmultiple separaterationales,tosimplifythepredictiontask 5 VERISCI: Baselinemodel weonlyrequirethemodeltopredictasinglecol- Wedevelopabaseline(referredtoasVERISCI)that lectionofrationalesentences;thesesentencesmay takes a claim c and corpus A as input, identifies encompassmultiplegoldrationales. evidenceabstractsE(cid:98)(c),andpredictsalabely (cid:98) (c,a) TaskEvaluationWeevaluatethetaskattwolevels andrationalesentencesS(cid:98)(c,a)foreacha ∈ E(cid:98)(c). of granularity. For abstract-level evaluation, we Followingthe“BERT-to-BERT”modelpresented assessthemodel’sabilitytoidentifytheabstracts inDeYoungetal.(2020a);Soleimanietal.(2019), thatsupportorrefutetheclaim. Forsentence-level VERISCIisapipelineofthreecomponents: evaluation, we evaluate the model’s performance

  1. ABSTRACTRETRIEVAL retrieves k abstracts atidentifyingthesentencessufficienttojustifythe withhighestTF-IDFsimilaritytotheclaim. abstract-levelpredictions. Weconductevaluations
  2. RATIONALESELECTION identifies rationale in both the “Open” FEVER-style (Thorne et al., sentencesS(cid:98)(c,a)foreachabstract.
  1. setting where the evidence abstracts must
  1. LABELPREDICTIONmakesthefinallabelpre- beretrieved,andthe“Oracleabstract” ERASER- dictiony(c,a). (cid:98) style (DeYoung et al., 2020a) setting where the Rationale selection Given a claim c and ab- goldevidenceabstractsE(c)areprovided. stract a, we train a model to predict z (cid:44) i Abstract-level evaluation is inspired by the 4Condition(3)eliminatesrationalesentenceswhichwere FEVERscore. Givenaclaimc,apredictedevidence identifiedbytherationaleselector,butprovedinsufficientto abstracta ∈ E(cid:98)(c)iscorrectlylabeled if(1)aisa justifyafinalSUPPORTS/REFUTESdecision 7538

1[a isarationalesentence] for each sentence a RATIONAL-SELECT. LABEL-PRED.

i i
in a. For each sentence, we encode the concate- Trainingdata P R F1 ACC.
nated sequence w = [a ,SEP,c] using a BERT-
------------------- ---------- ----- --------------------- -------- ------- ------- ------------- --- ---- --------- --- ----
i i
FEVER 41.5 57.9 48.4 67.6
style language model and predict a score z˜ =
i UKPSnopes 42.5 62.3 50.5 71.3
SCIFACT 73.7 70.5 72.1 75.7
σ[f(CLS(w ))],whereσ isthesigmoidfunction,
i FEVER+SCIFACT 72.4 67.2 69.7 81.9
f isalinearlayerand CLS(w i )isthe CLS token
from the encoding of w . We train the model on Sentenceencoder P R F1 ACC.
i
pairs of claims and their cited abstracts and min- SCIBERT 74.5 74.3 74.4 69.2
BioMedRoBERTa 75.3 69.9 72.5 71.7
------------------- --- --- ------------ --- ------- --------- ------------- --- ---- --------- --- ----
imize cross-entropy loss between z i and z˜. i For
RoBERTa-base 76.1 66.1 70.8 62.9
eachclaim,weusecitedabstractslabeledNOINFO,
RoBERTa-large 73.7 70.5 72.1 75.7
------- ---------------- --- --------- --- ----------- --------- ------------- --- ---- --------- --- ----
as well as non-rationale sentences from abstracts
Modelinputs P R F1 ACC.
labeled SUPPORTS and REFUTES as negative ex-
Claim-only - - - 44.5
amples. To make predictions, we select all sen- Abstract-only 60.1 60.9 60.5 53.3
tencesa withz˜ > tasrationalesentences,where
------- ------ ----------------------------- --- --- --- --- ---------------------------------------------- --- --- --- --- ---
i i
Table3: Comparisonofdifferenttrainingdatasets, en-
t ∈ [0,1]istunedonthedevset(AppendixA.1).
coders, and model inputs for RATIONALESELECTION
---------------- -------- --------- ------ ------------- --- ---------- -------------------- ----- ---------- ------------------ ------ -------
Label prediction Sentences identified by the ra-
and LABELPREDICTION, evaluated on the SCIFACT
tionale selector are passed to a separate BERT-
devset. Theclaim-onlymodelcannotselectrationales.
based model to make the final labeling decision.
----------- ----- ----- -------- -------------- -------------- --------- --- --- --- --- --- ---
Given a claim c and abstract a, we concatenate
the claim and the predicted rationale sentences VERISCI, and (4) discuss some modeling chal-
lengespresentedbythedataset.
u = [s (c,a),...s (c,a),SEP,c]5, and predict
--------- ------------- ---------------- -------------- --- ------ ------- --- --- --- --- --- ---
(cid:98)1 (cid:98)(cid:96)
y˜(c,a) = φ[f(CLS(u))], where φ is the softmax
6.1 Pipelinecomponents
function,andf isalinearlayerwiththreeoutputs
------------- --- ------------------------------ --- --- --- --- --- --- --- --- --- ---
representingthe{SUPPORTS, REFUTES, NOINFO We examine the effects of different training
datasets,sentenceencoders,andmodelinputson
} labels. We minimize the cross-entropy loss be-
--------- ----------- --- --- ------------- --- -------- --- --- --- --- --- ---
tweeny˜(c,a)andthetruelabely(c,a). the performance of the RATIONALESELECTION
Wetrainthemodelonpairsofclaimsandtheir and LABELPREDICTION modules. The RATIO-
citedabstractsusinggoldrationalesasinput. For NALESELECTIONmoduleisevaluatedonitsability
toselectrationalesentencesgivengoldabstracts6.
cited abstracts labeled NOINFO, we choose the
--------------- --- ------- ------- --- --------- --- --- --- --- --- --- ---
k sentences from the cited abstract with high- TheLABELPREDICTIONmoduleisevaluatedonits
est TF-IDF similarity to the claim as input ra- 3-waylabelclassificationaccuracygivengoldratio-
nalesfromcitedabstracts. Citedabstractslabeled
--------- --- ----------- --- ------ --- --------- ------------------------ --- --- --------------------- --- ---
tionales. For prediction, we use the predicted
rationale sentences S(cid:98)(c,a) as input and predict NOINFOareincludedintheevaluation. Theseab-
yˆ(c,a) = argmaxy˜(c,a). NOINFO is predicted stractshavenogoldrationalesentences;asin§5,
weprovidethek
forabstractswithnorationalesentences. mostsimilarsentencesfromthe
abstractasinput(moredetailsinAppendixA).
Weexperimentedwithalabelpredictionmodel
whichencodesentireabstractsviatheLongformer
Training Data We train on (1) FEVER, (2) UKP
--- --- --- --- --- --- --- ------------- --- ----- ------------- --- -------
(Beltagy et al., 2020), and makes predictions us- Snopes,(3)SCIFACT,and(4)FEVERpretraining
ingthedocument-level CLS token. Performance followedby SCIFACTfine-tuning. RoBERTa-large
wasnotcompetitivewithourpipelinesetup,likely
(Liuetal.,2019)isusedasthesentenceencoder.
because the label predictor struggles to identify
------- --------- --------- --- --------- --- -------- ---------------- --- ------------ ------- --- ------
Sentence encoder We fine-tune SCIBERT (Belt-
relevantinformationwhengivenfullabstracts.
agyetal.,2019),BioMedRoBERTa(Gururangan
etal.,2020),RoBERTa-base,andRoBERTa-large.
6 Experiments
SCIFACTisusedastrainingdata.
Inourexperiments,we(1)analyzetheperformance
Model Inputs We examine the performance of
--- --- --- --- --- --- --- ------------ --- ------- --- ----------- ---
ofeachindividualcomponentofVERISCI,(2)eval- “claim-only” and “abstract-only” models trained
uatefulltaskperformanceinboththe“Oracleab-
onSCIFACT,usingRoBERTa-largeasthesentence
stract”and“Open”settings,(3)presentpromising
encoder. Theclaim-onlymodelmakeslabelpredic-
results verifying claims about COVID-19 using
----------------- --- ------ ----- -------- --- ----- ------------------ --- ------------------ --- --- ------
6Our FEVER-trained RATIONALESELECTION module
5WetruncatetherationaleinputifitexceedstheBERT achieves79.9sentence-levelF1ontheFEVERtestset,virtu-
tokenlimit.cisnevertruncated. allyidenticalto79.6reportedinDeYoungetal.(2020a).
7539
Sentence-level Abstract-level
Selection-Only Selection+Label Label-Only Label+Rationale
Retrieval Model P R F1 P R F1 P R F1 P R F1
Oraclerationale 1 100.0 80.5 89.2 89.6 72.2 79.9 90.1 77.5 83.3 90.1 77.5 83.3
2.1 3.0 2.4 2.4
--- --- --- --- --- --- --- --- --- --- --- --- ---
Oracle
43.8 37.2 66.3 51.8
--- --- --- --- ---- --- ---- --- --- ---- --- --- ----
abstract Zero-shot 2 42.5 45.1 2.0 36.1 38.4 2.3 86.9 53.6 3.1 67.9 41.9 3.4
VERISCI 3 76.1 63.8 69.4 66.5 55.7 60.6 87.3 65.3 74.7 84.9 63.5 72.7
2.6 3.1 2.8 2.9
--- --- --- --- --- --- --- --- --- --- --- --- ---
Oraclerationale 4 100.0 56.5 72.2 3.3 87.6 49.5 63.2 3.7 88.9 54.1 67.2 3.2 88.9 54.1 67.2 3.2
Open Zero-shot 5 28.7 37.6 32.5 2.3 23.7 31.1 26.9 2.3 56.0 42.3 48.2 3.3 42.3 32.0 36.4 3.3
VERISCI 6 45.0 47.3 46.1 3.0 38.6 40.5 39.5 3.0 47.5 47.3 47.4 3.1 46.6 46.4 46.5 3.1
Table 4: Test set performance on SCIFACT, according to the metrics from §4. For the “Oracle abstract” rows,
thesystemisprovidedwithgoldevidenceabstracts. “Oraclerationale”rowsindicatethatthegoldrationalesare
providedasinput. “Zero-shot”indicateszero-shotperformanceofaverificationsystemtrainedonFEVER. Addi-
tionally,standarddeviationsarereportedassubscriptsforallF1scores. SeeAppendixBforstandarddeviations
onallreportedmetrics.
tionsbasedontheclaimtextalone,withoutaccess inordertoidentifyrelevantevidence.
to evidence abstracts. The abstract-only model
--------------------------------------------- ---------- ----------------- --- ----- --- -------- --- --- --- --- --- ---
selectsrationalesentencesandmakeslabelpredic- 6.2 Fulltask
tionswithoutaccesstotheclaim.
Experimental setup Based on the results from
--- --- --- --- --- ------------ --- ----- ----- --- ------ ------- ----
ResultsTheresultsareshowninTable3. For LA- §6.1,weusetheRATIONALESELECTIONmodule
trainedon SCIFACT only,andthe LABELPREDIC-
--- --- --- --- --- --------- --- ------- ----------- --- ------------ --- ---
BELPREDICTION,thebestperformanceisachieved
by training first on the large FEVER dataset and TIONmoduletrainedonFEVER+SCIFACTforour
thenfine-tuningonthesmallerin-domainSCIFACT finalend-to-endsystem VERISCI. AlthoughSCIB-
trainingset. TounderstandthebenefitsofFEVER ERTperformsslightlybetteronrationaleselection,
pretraining,weexaminedtheclaim/evidencepairs usingRoBERTa-largeforbothRATIONALESELEC-
wherethe FEVER+ SCIFACT-trainedmodelmade TION and LABELPREDICTION gavethebestfull-
correctpredictionsbuttheSCIFACT-trainedmodel pipeline performance on the dev set, so we use
RoBERTa-largeforbothcomponents. FortheAB-
-------- ------------- ------------ --- -------- ------------------------------- --- --- --- --- --- --------- ---
did not. In 36 / 44 of these cases, the SCIFACT-
trainedmodelpredictsNOINFO. Thuspretraining STRACTRETRIEVALmodule,thebestdevsetfull-
on FEVER appears to improve the model’s abil- pipeline performance was achieved by retrieving
thetopk = 3documents.
---------------- ------- ---------- ------------- --- ------- --- ------------- --- --- --- --- ---
ity to recognize textual entailment relationships
betweenevidenceandclaim–particularlyrelation-
Model comparisons We report performance of
--------------- ---------------------- --- --- ---- ----- ----------- --------- --- ------- ----------- ---------- ---
ships indicated by non-domain-specific cues like
three model variants. For the “Oracle rationale”
“isassociatedwith”or“hasanimportantrolein”. setting,the RATIONALESELECTION moduleisre-
For RATIONALESELECTION, training on SCI- placedbyanoraclewhichoutputsgoldrationales
FACT alone produces the best results. We exam- forcorrectlyretrieveddocuments,andnorationales
inedtherationalesthattheSCIFACT-trainedmodel forincorrectretrievals. The“Zero-shot”settingre-
portsthezero-shotgeneralizationperformanceof
identified but the FEVER- trained model missed,
---------- -------------- ------- ----- ------- --- --- --- --- --- --- --- ---
and found that they generally contain science- a model trained on FEVER (the results on UKP
specific vocabulary. Thus, training on additional Snopeswereslightlyworse). VERISCIreportsthe
performanceofourbestsystem.
out-of-domaindataprovideslittlebenefit.
RoBERTa-large exhibits the strongest perfor- Results The results are shown in Table 4. In the
mance on label prediction, while SCIBERT has oracleabstractsetting,theabstract-levelF1scores
areroughlycomparabletolabelclassificationaccu-
a slight edge on rationale selection. The “claim-
-------- ----------------- ---------- --- ------- --- --- --- --- --- --- --- ---
only”modelexhibitsverypoorperformance,which racies,andtheAbstract scoreinRow
Label+Rationale
providessomereassurancethattheclaimnegation 3impliesanend-to-endclassificationaccuracyof
roughly70%,givengoldabstracts.
proceduredescribedin§3.2doesnotintroduceob-
viousstatisticalartifacts. Similarly,thepoorperfor- Access to in-domain data during training
manceofthe“abstract-only”modelindicatesthat clearly improves performance. Despite the
themodelneedsaccesstotheclaimbeingverified small size of SCIFACT, training on these data
7540

Reasoningtype Example Claim: Rapamycinslowsaginginfruitflies. Science Evidence: ...feedingrapamycintoadultDrosophilaproduceslifespanextension... background GoldVerdict: SUPPORTS Reasoning: Drosophilaisatypeoffruitfly. Claim: Inhibitingglucose-6-phospatedehydrogenaseimpairslipogenesis Evidence: ...suppressionof6PGDincreasedlipogenesis Directionality GoldVerdict: REFUTES Reasoning: Adecrease(notincrease)inlipogenesiswouldindicatelipogenesisimpairment. Claim: Bariatricsurgeryimprovesresolutionofdiabetes. Evidence: StrongassociationswerefoundbetweenbariatricsurgeryandtheresolutionofT2DM, Numerical withaHRof9.29(95%CI6.84-12.62)... reasoning GoldVerdict: SUPPORTS Reasoning: AHR(hazardratio)thatisgreaterthan1with95%confidenceindicatesimprovement. Claim: Majorvaultprotein(MVP)functionstodecreasetumoraggression. Evidence: KnockoutofMVPleadstomiR-193aaccumulation...inhibitingtumorprogression Causeand effect GoldVerdict: REFUTES Reasoning: Knockingout(removing)MVPinhibitstumorprogression→MVPincreasestumor aggression. Claim: Lowsaturatedfatdietshaveadverseeffectsonthedevelopmentofinfants Evidence: Neurologicaldevelopmentofchildrenintheinterventiongroupwasatleastasgoodas... Coreference thecontrolgroup GoldVerdict: REFUTES Reasoning: Theinterventiongroupinthisstudywasplacedonalowsaturatedfatdiet. Table 5: Reasoning types required to verify SCIFACT claims which are classified incorrectly by our modeling baseline. Wordscrucialforcorrectverificationarehighlighted. leads to relative improvements of 47% on sultsindicatethattheobserveddifferencesinmodel open Sentence , and 28% on open performancearestatisticallyrobustandcannotbe Selection+Label Abstract Label+Rationale overFEVERalone(Row6vs. attributedtorandomvariationinthedataset. Row5). Thethreepipelinecomponentsmakesimi- 6.3 VerifyingclaimsaboutCOVID-19 larcontributionstotheoverallmodelerror. Replac- ingRATIONALESELECTIONwithanoracleleads Weconductexploratoryexperimentsusingoursys- toaroughly20-pointriseinSentence Selection+Label tem to verify claims concerning COVID-19. We F1(Row6vs. Row4). ReplacingABSTRACTRE- taskedamedicalstudenttowrite36COVID-related TRIEVALwithanoracleaswellleadstoagainof claims. For each claim c, we used VERISCI to roughly20morepoints(Row4vs. Row1). predictevidenceabstractsE(cid:98)(c). Theannotatorex- Nearly all correctly-labeled abstracts are sup- amined each (c,E(cid:98)(c)) pair. A pair was labeled portedbyatleastonerationale. Thereisonlyatwo- plausibleifE(cid:98)(c)wasnonempty,andatleasthalfof point difference in F1 between Abstract Label-Only theevidenceabstractsinE(cid:98)(c)werejudgedtohave and Abstract in the oracle setting reasonablerationalesandlabels. For23/36claims, Label+Rationale (Row 3), and a one-point difference in the theresponseofVERISCIwasdeemedplausibleby open setting (Row 6). The differences between ourannotator,demonstratingthatVERISCIisable Sentence andSentence are to successfully retrieve and classify evidence in Selection-Only Selection+Label larger,causedbyexampleswherethemodelfinds manycases. TwoexamplesareshowninTable1. theevidencebutfailstopredictitsrelationshipto Inbothcases,oursystemidentifiesbothsupporting theclaim. Weexaminethesein§6.4. and refutingevidence. Weevaluatethestatisticalrobustnessofourre- 6.4 Erroranalysis sultsbygenerating10,000bootstrap-resampledver- sions of the test set (Dror et al., 2018) and com- TobetterunderstandtheerrorsmadebyVERISCI, puting the standard deviation of all performance weconductamanualanalysisoftestsetpredictions metrics. Table4showsthestandarddeviationsin whereanevidenceabstractwascorrectlyretrieved, F1score. Uncertaintiesonallmetricsforboththe butwherethemodelfailedtoidentifyanyrelevant devandtestsetcanbefoundinAppendixB.There- rationalesorpredictedanincorrectlabel. Weiden- 7541

tifyfivemodelingcapabilitiesrequiredtocorrect knowledgeintensivetaskswhichrequireanunder- thesemistakes(Table5providesexamples): standingoftherelationshipbetweenaninputquery andrelevantsupportingtext.

Science background includes knowledge of
domain-specificlexicalrelationships. Automated evidence synthesis (Marshall and
Directionalityrequiresunderstandingincreasesor Wallace, 2019; Beller et al., 2018; Tsafnat et al.,
decreasesinscientificquantities. 2014;Marshalletal.,2017)seekstoautomatethe
Numericalreasoninginvolvesinterpretingnumer- processofcreatingsystematicreviewsofthemed-
icalorstatisticalfindings. icalliterature7 –forinstance,byextractingPICO
Cause and effect requires reasoning about coun- snippets (Nye et al., 2018) and inferring the out-
comesofclinicaltrials(Lehmanetal.,2019;DeY-
terfactuals.
Coreferenceinvolvesdrawingconclusionsusing oungetal.,2020b). Wehopethatsystemsforclaim
contextstatedoutsideofarationalesentence. verificationwillserveascomponentsinfutureevi-
dencesynthesisframeworks.
7 Relatedwork
8 Conclusionandfuturework
Fact checking and rationalized NLP models
------------- ------- -------- ------------ -------- ---------- ------- ---------------------------------------- ---- ------- --------------- -------- -----------
Claim verification allows us to trace the sources
Fact-checking datasets include PolitiFact (Vla-
andmeasuretheveracityofscientificclaims. These
chos and Riedel, 2014), Emergent (Ferreira and
abilities have emerged as particularly important
Vlachos, 2016), LIAR (Wang, 2017), SemEval
2017Task8RumorEval(Derczynskietal.,2017), in the context of the current pandemic, and the
broader reproducibility crisis in science. In this
---------------------- ------ ------- ------ ----------- ------ ------ -------- --------------- --- -------- ------------- -------
Snopes (Popat et al., 2017), CLEF-2018 Check-
article, we formalize the task of scientific claim
That! (Barro´n-Ceden˜o et al., 2018), Verify (Baly
et al., 2018), Perspectrum (Chen et al., 2019), verification,andreleaseadataset(SCIFACT)and
models (VERISCI) to support work on this task.
----- ------- --- ----------- --- ------- ------ ------ --------- --- ---------- ---- -------------
FEVER (Thorne et al., 2018), and UKP Snopes
Ourresultsindicatethatitispossibletotrainmod-
(Hanselowski et al., 2019). Hanselowski et al.
------------------------------ --- ------- ------ ----------- ----------- ------ ------- ---------- ------------- --- --- -----------
els for scientific fact-checking and deploy them
(2019)providesathoroughreview. Toourknowl-
edge, thereare noexistingdata setsfor scientific with reasonable efficacy on real-world claims re-
latedtoCOVID-19.
claimverification. Werefertoourtaskas“claim
------------------ --- --- ------------------------ --- --- --- --- --- --- --- --- ---
Scientificclaimverificationpresentsanumber
verification”ratherthan“fact-checking”toempha-
sizethatourfocusistohelpresearchersmakesense ofpromisingavenuesforresearchonmodelscapa-
bleofincorporatingbackgroundinformation,rea-
ofscientificfindings,nottocounterdisinformation.
soning about scientific processes, and assessing
--- --- --- --- --- --- --- ------ ----- ---------- ---------- --- -------------
Fact-checkingisoneofanumberoftaskswhere
the strength and provenance of various evidence
------- ----------- --- ---------- --- ---------- ------- ------------ --- ---------- --- ---------- --------
a model is required to justify a prediction via ra-
tionalesfromthesourcedocument. TheERASER sources. Thislastchallengewillbeespeciallycru-
cialforfutureworkthatseekstoverifyscientific
dataset (DeYoung et al., 2020a) provides a suite
------- -------- --- ------- ------ -------- ------- --- --- --- --- --- ---
claimsagainstsourcesotherthantheresearchlit-
of benchmark datasets (including SCIFACT) for
------------ --- -------- ---------- --- -------- --- --- --- --- --- --- ---
evaluatingrationalizedNLPmodels. erature–forinstance,socialmediaandthenews.
We hope that the resources presented in this pa-
--- --- --- --- --- --- --- ------- ---- ------------- --- --------- -----------
RelatedscientificNLPtasksThecitationcontex-
perencouragefutureresearchontheseimportant
tualization task (Cohan et al., 2015; Jaidka et al.,
----------- ---- ------ --- ---------- ------ ------- --- --- --- --- --- ---
challenges,andhelpfacilitateprogresstowardthe
2017)istoidentifyspansinaciteddocumentthat
broadergoalofscientificdocumentunderstanding.
arerelevanttoaparticularcitationinacitingdoc-
ument. Unlike SCIFACT, these citations are not Acknowledgments
re-written into atomic claims and are therefore
---------------------- ---- ------ ------------------------ --- ------- --------- ----------------- --- ------------- --- ----------------- --------
This research was supported by the ONR MURI
moredifficulttoverify. Expertannotatorsachieved
N00014-18-1-2670, ONR N00014-18-1-2826,
verylow(21.7%)inter-annotatoragreementonthe
DARPAN66001-19-2-4031,NSF(IIS1616112),
BioMedSummdataset(Cohenetal.,2014),which
contains314citationsreferencing20papers. Allen Distinguished Investigator Award, and the
Sloanfellowship. WethanktheSemanticScholar
--- --- --- --- --- --- --- ---------------- --- ------------------------- --- --- ---
Biomedicalquestionansweringdatasetsinclude
teamatAI2,UW-NLP,andH2labatUWforhelp-
BioASQ(Tsatsaronisetal.,2015)andPubMedQA
(Jin et al., 2019), which contain 855 and 1,000 fulcommentsandfeedback.
“yes/no”questionsrespectively(Guetal.,2020).
7https://www.cochranelibrary.com/
Claimverificationandquestionansweringareboth- about/about-cochrane-reviews
7542

References Jay DeYoung, Eric Lehman, Ben Nye, Iain James Marshall, and Byron C. Wallace. 2020b. Evi- Ramy Baly, Mitra Mohtarami, James Glass, Llu´ıs dence inference 2.0: More data, better models. In Marquez,AlessandroMoschitti,andPreslavNakov. BioNLP@ACL. 2018. Integratingstancedetectionandfactchecking inaunifiedcorpus. InNAACL. RotemDror,GiliBaumer,SegevShlomov,andRoiRe- ichart.2018. Thehitchhiker’sguidetotestingstatis- Alberto Barro´n-Ceden˜o, Tamer Elsayed, Reem ticalsignificanceinnaturallanguageprocessing. In Suwaileh, Llu´ıs Marquez i Villodre, Pepa ACL. Atanasova, Wajdi Zaghouani, Spas Kyuchukov, Giovanni Da San Martino, and Preslav Nakov. Bradley Efron and Robert Tibshirani. 1993. An intro- 2018. Overviewoftheclef-2018checkthat! labon ductiontothebootstrap. automaticidentificationandverificationofpolitical claims.task2: Factuality. InCLEF. William Ferreira and Andreas Vlachos. 2016. Emer- gent: a novel data-set for stance classification. In Elaine Beller, Justin Clark, Guy Tsafnat, Clive Elliott NAACL. Adams, Heinz Diehl, Hans Lund, Mourad Ouzzani, Kristina Thayer, James Thomas, Tari Turner, J. S. Yu Gu, Robert Tinn, Hao Cheng, M. Lucas, Xia,KarenA.Robinson,andPaulPGlasziou.2018. Naoto Usuyama, Xiaodong Liu, Tristan Nau- Makingprogresswiththeautomationofsystematic mann, Jianfeng Gao, and Hoifung Poon. 2020. reviews: principles of the international collabora- Domain-specific language model pretraining for tionfortheautomationofsystematicreviews(icasr). biomedical natural language processing. ArXiv, SystematicReviews,7. abs/2007.15779. Suchin Gururangan, Ana Marasovic, Swabha IzBeltagy,KyleLo,andArmanCohan.2019. Scibert: Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, A pretrained language model for scientific text. In and Noah A. Smith. 2020. Don’t stop pretraining: EMNLP. Adapt language models to domains and tasks. In Iz Beltagy, Matthew E. Peters, and Arman Cohan. ACL. 2020. Longformer: Thelong-documenttransformer. Suchin Gururangan, Swabha Swayamdipta, Omer ArXiv,abs/2004.05150. Levy, Roy Schwartz, Samuel R. Bowman, and Taylor Berg-Kirkpatrick, David Burkett, and Dan Noah A. Smith. 2018. Annotation artifacts in nat- Klein. 2012. An empirical investigation of statisti- urallanguageinferencedata. InNAACL. calsignificanceinnlp. InEMNLP. AndreasHanselowski, ChristianStab, Claudia Schulz, Zile Li, and Iryna Gurevych. 2019. A richly anno- Sihao Chen, Daniel Khashabi, Wenpeng Yin, Chris tated corpus for different tasks in automated fact- Callison-Burch,andDanRoth.2019. Seeingthings checking. InCoNLL. fromadifferentangle: Discoveringdiverseperspec- tivesaboutclaims. InNAACL. Kokil Jaidka, Muthu Kumar Chandrasekaran, Devan- shuJain,andMin-YenKan.2017. Thecl-scisumm Arman Cohan, Luca Soldaini, and Nazli Goharian. shared task 2017: Results and key insights. In 2015. Matching citation text and cited spans in BIRNDL@JCDL. biomedical literature: a search-oriented approach. InNAACL. QiaoJin,BhuwanDhingra,ZhengpingLiu,WilliamW. Cohen, and Xinghua Lu. 2019. Pubmedqa: A Kevin Bretonnel Cohen, Hoa Trang Dang, Anita dataset for biomedical research question answering. de Waard, Prabha Yadav, and Lucy Vanderwende. InEMNLP. 2014. Tac 2014 biomedical summarization track. https://tac.nist.gov/2014/BiomedSumm/. Eric Lehman, Jay DeYoung, Regina Barzilay, and By- ron C. Wallace. 2019. Inferring which medical Leon Derczynski, Kalina Bontcheva, Maria Liakata, treatments work from reports of clinical trials. In RobProcter,GeraldineWongSakHoi,andArkaitz NAACL. Zubiaga.2017. SemEval-2017task8: RumourEval: Determining rumour veracity and support for ru- Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. mours. InSemEval. 2016. Rationalizingneuralpredictions. InACL. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and YinhanLiu,MyleOtt,NamanGoyal,JingfeiDu,Man- KristinaToutanova.2019. Bert:Pre-trainingofdeep dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, bidirectional transformers for language understand- Luke Zettlemoyer, and Veselin Stoyanov. 2019. ing. InNAACL. Roberta: A robustly optimized bert pretraining ap- proach. ArXiv,abs/1907.11692. Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, EricLehman, CaimingXiong, RichardSocher, and KyleLo,LucyLuWang,MarkNeumann,RodneyKin- Byron C. Wallace. 2020a. Eraser: A benchmark to ney,andDanielS.Weld.2020. S2ORC:TheSeman- evaluaterationalizednlpmodels. InACL. ticScholarOpenResearchCorpus. InACL. 7543

Yi Luan, Luheng He, Mari Ostendorf, and Hannaneh Andreas Vlachos and Sebastian Riedel. 2014. Fact Hajishirzi. 2018. Multi-task identification of enti- checking: Task definition and dataset construction. ties, relations, and coreference for scientific knowl- In ACL Workshop on Language Technologies and

edgegraphconstruction. InEMNLP. ComputationalSocialScience.
IainJamesMarshall,Joe¨lKuiper,EdwardBanner,and Lucy Lu Wang, Kyle Lo, Yoganand Chandrasekhar,
RussellReas,JiangjiangYang,DarrinEide,Kathryn
ByronC.Wallace.2017. Automatingbiomedicalev-
-------------------- --- -------------- ----------------------- ---- --- --- ------------ ------- ----- ------ ------------- -------
Funk, Rodney Kinney, Ziyang Liu, William. Mer-
idencesynthesis: Robotreviewer. ACL.
rill, Paul Mooney, Dewey A. Murdick, Devvret
Iain James Marshall and Byron C. Wallace. 2019. Rishi, Jerry Sheehan, Zhihong Shen, Brandon Stil-
Toward systematic review automation: a practical son,AlexD.Wade,KuansanWang,ChristopherWil-
helm, Boya Xie, Douglas M. Raymond, Daniel S.
-------- ----- ------- -------- --- ----- ----------- ---------- ---- ------- --- -------- ---------
guide to using machine learning tools in research
Weld,OrenEtzioni,andSebastianKohlmeier.2020.
synthesis. SystematicReviews,8.
---------- -------------------- --- --- --- --- --- -------- --- -------- ---- -------- --------
Cord-19: The covid-19 open research dataset.
ArXiv,abs/2004.10706.
Preslav I Nakov, Ariel S Schwartz, and Marti Hearst.
--------- ------ ----- ----------- --- --------- ------- --- --- --- --- --- ---
2004. Citances: Citation sentences for semantic WilliamYangWang.2017. “liar,liarpantsonfire”: A
analysis of bioscience text. In SIGIR workshop on
-------- ------------- --- ----- -------- -------- --- ---------------------------------------- --- --- --- --- ---
newbenchmarkdatasetforfakenewsdetection. In
SearchandDiscoveryinBioinformatics.
ACL.
Benjamin Nye, Junyi Jessy Li, Roma Patel, Yinfei Thomas Wolf, Lysandre Debut, Victor Sanh, Julien
Yang, Iain James Marshall, Ani Nenkova, and By- Chaumond, ClementDelangue, AnthonyMoi, Pier-
ronC.Wallace.2018. Acorpuswithmulti-levelan- ricCistac,TimRault,R’emiLouf,MorganFuntow-
notationsofpatients,interventionsandoutcomesto icz, and Jamie Brew. 2019. Huggingface’s trans-
support language processing for medical literature.
------- -------- ---------- --- ----------- --- ----------- -------- ---------------- --- ------- -------- --------
formers: State-of-the-art natural language process-
InACL.
ing. ArXiv,abs/1910.03771.
Kashyap Popat, Subhabrata Mukherjee, Jannik
----------------------------- ------------------------------------ ---------- ------- ----------- ---------- --------- --- --- --- --- --- ---
Stro¨tgen, and Gerhard Weikum. 2017. Where the
truth lies: Explaining the credibility of emerging
claimsonthewebandsocialmedia. InWWW.
Tal Schuster, Darsh J. Shah, Yun Jie Serene Yeo,
Daniel Filizzola, Enrico Santus, and Regina Barzi-
lay.2019. Towardsdebiasingfactverificationmod-
els. InEMNLP.
Amir Soleimani, Christof Monz, and Marcel Worring.
--------------- ------------------------------------ -------- --------- --- --------- -------- --- --- --- --- --- ---
2019. Bert for evidence retrieval and claim verifi-
cation. InEuropeanConferenceonInformationRe-
trieval.
James Thorne, Andreas Vlachos, Christos
------------------- -------------------------- ----------- ---------- ------------ -------------- ---------- --- --- --- --- --- ---
Christodoulopoulos, and Arpit Mittal. 2018.
Fever: a large-scale dataset for fact extraction and
verification. InNAACL.
Guy Tsafnat, Paul P Glasziou, Miew Keen Choong,
Adam G. Dunn, Filippo Galgani, and Enrico W.
Coiera. 2014. Systematic review automation tech-
nologies. SystematicReviews,3:74–74.
George Tsatsaronis, Georgios Balikas, Prodromos
Malakasiotis, Ioannis Partalas, Matthias Zschunke,
Michael R. Alvers, Dirk Weissenborn, Anastasia
Krithara, Sergios Petridis, Dimitris Polychronopou-
los, Yannis Almirantis, John Pavlopoulos, Nico-
las Baskiotis, Patrick Gallinari, Thierry Artie`res,
Axel-Cyrille Ngonga Ngomo, Norman Heino, E´ric
Gaussier,LilianaBarrio-Alvers,MichaelSchroeder,
IonAndroutsopoulos,andGeorgiosPaliouras.2015.
An overview of the bioasq large-scale biomedical
----------- --- ------ ------ ----------- --- ---------- --- --- --- --- --- ---
semanticindexingandquestionansweringcompeti-
tion. InBMCBioinformatics.
7544
A Modelimplementationdetails A.3 Trainingthe LABELPREDICTION
module
AllmodelsareimplementedusingtheHuggingface We adopt similar settings as we used for the RA-
Transformerspackage(Wolfetal.,2019). TIONALESELECTION moduleandonlychangethe
learningrateto1e-5forthetransformerbaseand
1e-4forthelinearlayerformodelstrainedon SCI-
--- --------------------- --- --- --- ------------- --- --- --------------------------------------- --- --- --- -------------- ----
A.1 Parametersforthefinal VERISCIsystem
FACT,FEVER,andUKPSnopes. Whentrainingon
claim/citedabstractpairslabeledNOINFO,weuse
Forthe ABSTRACTRETRIEVALmodule, VERISCI
------ ------------------------ --- --- --- --- ------- --- --- --- --- --- --- ---
thek
retrieves the top k = 3 documents ranked by TF- sentencesintheabstractwithgreatestsimi-
laritytotheclaimasrationales(§5). k issampled
--- ---------- --- ----- ------- -------- --- --------- --------------------------------- --- --- --- ----------- ---
IDF similarity using unigram + bigram features.
from{0,1}withuniformprobability.
Theseparametersaretunedonthe SCIFACTdevel-
---------------------------- --- ------ ----------- --- ------------- --- ------ ----------------------------- --- --- --- --- ---
opmentset. A.4 Additionaltrainingdetails
When making predictions using the RATIO-
AllmodelsaretrainedusingasingleNvidiaP100
NALESELECTIONmoduledescribedin§5,wefind GPUonGoogleColabortoaryProplatform.8
For
that the usual decision rule of predicting zˆ = 1
-------- --- ------------------------------- ----- ---- ------------- ------- ------ ---------------------------------------- --- ----- ---------- -------------- ---
i theRATIONALESELECTIONmodule,ittakesabout
when ≥ 0.5 works well for models trained on
i 150 minutes to train on SCIFACT for 20 epochs.
SCIFACT. However,formodelstrainedonFEVER
120minutesonUKPSnopesfor5epochs,and700
andUKPSnopes,weachievebetterperformance minuteson FEVER for3epochs. Forthe LABEL-
by tuning the classification threshold t, such that
--- ------ ------------------ --- ----- ----------- ------- -------- ---------- ----------------------------- --- -------------- ----------- ---
PREDICTION module,ittakesabout130minutes
= 1 when z˜ t, on the SCIFACT dev set.
i i to train on SCIFACT for 20 epochs, 160 minutes
The best threshold was t = 0.025 when training
onUKPSnopesfor5epochs,and640minuteson
on FEVER, and t = 0.75 when training on UKP
--- ------ --- --- --------- -------- --- ------ --- --- --- --- --- ---
FEVERfor3epochs.
Snopes.
A.5 Hyperparametersearch
--- --- --- --- --- --- --- --- ------------------------ --- --- --- --- ---
Thelearningrate,batchsize,andnumberofepochs
A.2 Trainingthe RATIONALESELECTION
--- ----------- --- ------------------ --- --- --- --- ----------------------------------- --- --- --- ------ ---
arethemostimportanthyperparameters. Weper-
module
formmanualtuningandselectthehyperparameters
that produce the highest F1 on the development
--- ---------- --- ---- ------- -------- ----- ---- --------------------------------------------- --- ------- ----- --------------- ---
We experiment with various learning rates when
set. Forthelearningrate,weexperimentwith1e-3,
training SCIBERT,BioMedRoBERTa,RoBERTa-
base,andRoBERTa-large. Belowwedescribethe 1e-4, 5e-5, 1e-5, and5e-6. Forbatchsize, weex-
settingfortrainingRoBERTa-large. perimentwith64and256. Thenumberofepochs
arecutoffafterthemodelconverges.
Formodelstrainedon SCIFACT,weuseanini-
------------------ -------- ------ ------- ------------------- --------------- ----- ------ --------------------- --------------- --- ------ ---------------- ---
tial learning rate of 1e-5 on the transformer base
B Statisticalanalysis
and 1e-3 on the linear layer. For FEVER + SCI-
We assess the uncertainty in the results reported
FACT,thelearningrateissetto1e-5fortheentire
modelforpre-trainingon FEVER andfine-tuning inthemainresults(Table4)usingasimpleboot-
on SCIFACT. Weuseabatchsizeof256through strapapproach(Droretal.,2018;Berg-Kirkpatrick
et al., 2012; Efron and Tibshirani, 1993). Given
-------- --- ------------ --- --- ------------ --- -------- ------------- ----- --- ----------- ------ -----
gradient accumulation and apply cosine learning
ratedecayover20epochstofindthebestperform- our test set with n = 300 claims, we gener-
test
ingmodelonthedevset. ate n = 10,000 bootstrap-resampled test sets
boot
byresampling(uniformly,withreplacement)n
Formodelstrainedon FEVER,wesetthelearn- test
------------------ --- --- --- -------------------- --- --- --- --------------------- --- --- ------------------------ --- ----
claimsfromthetestset. Foreachresampledtestset,
ingrateto5e-6forthetransformerbaseand5e-5
wecomputethemetricsinTable4. Table6reports
--- ---------- ------ --- ---------- ------- --- ------ ---------------------------- --- --- --- ------------- ---
for the linear layer. For models trained on UKP
themeanandstandarddeviationofthesemetrics,
Snopes,wesetthelearningrate1e-5forthetrans-
computedoverthebootstrapsamples. Table7re-
----------------------------------- --- --- --- --- --- --- ------ -------------------------------- --- --- ------------------------- --------- ---
formerbaseand1e-4forthelinearlayer. Wefind
portsdevsetmetrics. Ourconclusionthattraining
thattheselearningrateshelpthemodelsconverge.
onSCIFACTimprovesperformanceisrobusttothe
We only train the model for 3 epochs on FEVER
--- ---- --------- ----- --- -------- --- ----- --- --- --- --- --- ---
uncertaintiespresentedinthesetables.
and 5 epochs on UKP Snopes because they are
------------ -------- --- ------- ------ --------- --- -------- ----------------------------------- --- --- --- --- ---
larger datasets and the models converged within
earlyepochs. 8https://colab.research.google.com/
7545

Sentence-level

Selection-Only Selection+Label
Retrieval Model Row P R F1 P R F1
Oraclerationale 1 100.0 80.5 89.2 89.6 72.2 79.9
0.0 3.3 2.1 2.7 3.7 3.0
Oracle
Zero-shot 2 42.6 45.2 43.8 36.2 38.4 37.2
-------- --------------- --- ----- ---- ---- ---- ---- -------
abstract 2.2 3.2 2.0 2.5 3.0 2.3
VERISCI 3 76.2 63.9 69.4 66.5 55.7 60.6
2.9 3.6 2.6 3.4 3.7 3.1
Oraclerationale 4 100.0 56.6 72.2 87.6 49.5 63.2
0.0 4.0 3.3 3.5 3.9 3.7
Open Zero-shot 5 28.7 37.6 32.5 23.8 31.1 26.9
2.3 3.4 2.3 2.3 3.1 2.3
VERISCI 6 45.0 47.4 46.1 38.5 40.6 39.5
3.0 3.8 3.0 3.0 3.6 3.0
(a)Sentence-levelresults.
Abstract-level
Label-Only Label+Rationale
--------- --------------- --- ---- ---------- ---- ---- --------------- -------
Retrieval Model Row P R F1 P R F1
90.1 77.5 83.3 90.1 77.5 83.3
Oraclerationale 1 2.2 2.8 2.4 2.2 2.8 2.4
Oracle
abstract Zero-shot 2 86.9 2.9 53.6 3.4 66.3 3.1 67.9 3.9 41.9 3.2 51.8 3.4
VERISCI 3 87.3 2.6 65.3 3.2 74.7 2.8 84.9 2.8 63.5 3.2 72.6 2.9
--- --------------- --- ---- -------- -------- -------- -------- ------------
Oraclerationale 4 88.9 2.7 54.1 3.5 67.2 3.2 88.9 2.7 54.1 3.5 67.2 3.2
Open Zero-shot 5 56.0 3.9 42.3 3.4 48.2 3.3 42.3 4.0 32.0 3.2 36.4 3.3
VERISCI 6 47.5 3.3 47.3 3.5 47.4 3.1 46.6 3.3 46.4 3.5 46.4 3.1
--- ------- --- ---- -------- -------- -------- -------- ------------
(b)Abstract-levelresults
Table 6: Test set results as in Table 4, reporting mean and standard deviation over 10,000 bootstrap samples.
Standarddeviationsarereportedassubscripts. SomemeansreportedhereareslightlydifferentfromTable4due
tosamplingvariability.
Sentence-level
Selection-Only Selection+Label
--------- --------------- --- ----- -------------- ---- ---- --------------- -------
Retrieval Model Row P R F1 P R F1
Oraclerationale 1 100.0 81.9 90.0 91.4 74.9 82.3
0.0 3.2 1.9 2.5 3.6 2.9
Oracle
Zero-shot 2 40.7 48.1 44.0 36.1 42.6 39.0
-------- --------------- --- ----- ---- ---- ---- ---- -------
abstract 2.1 3.4 2.1 2.5 3.4 2.5
VERISCI 3 79.4 59.0 67.7 71.4 53.0 60.8
2.7 3.6 2.8 3.5 3.6 3.3
Oraclerationale 4 100.0 58.4 73.7 90.2 52.7 66.4
0.0 4.3 3.4 3.3 4.3 3.9
Open Zero-shot 5 28.6 38.5 32.8 24.8 33.4 28.4
2.0 3.6 2.3 2.2 3.4 2.4
VERISCI 6 52.5 43.8 47.7 46.9 39.2 42.6
3.5 3.7 3.2 3.7 3.6 3.2
(a)Sentence-levelresults.
Abstract-level
Label-Only Label+Rationale
--------- --------------- --- ---- ---------- -------- -------- --------------- ------------
Retrieval Model Row P R F1 P R F1
Oraclerationale 1 91.4 2.2 76.1 3.0 83.0 2.5 91.4 2.2 76.1 3.0 83.0 2.5
Oracle
Zero-shot 2 88.9 2.8 58.3 3.7 70.4 3.2 69.2 3.9 45.4 3.5 54.8 3.5
--- --------- --- ---- -------- -------- -------- -------- ------------
abstract
VERISCI 3 91.0 2.3 67.4 3.3 77.4 2.7 85.2 2.9 63.2 3.5 72.5 3.1
---- --------------- --- ---- -------- -------- -------- -------- ------------
Oraclerationale 4 91.0 2.6 53.1 3.8 67.0 3.4 91.0 2.6 53.1 3.8 67.0 3.4
Open Zero-shot 5 52.7 41.6 46.5 43.6 34.4 38.4
3.7 3.7 3.4 3.7 3.5 3.3
VERISCI 6 55.4 47.5 51.0 52.6 45.1 48.5
3.7 3.6 3.3 3.7 3.6 3.3
(b)Abstract-levelresults
Table7: DevsetresultsasinTable4,reportingmeanandstandarddeviationover10,000bootstrapsamples.
7546
Journal Count Source citance
BMJ 60
"Future studies are also warranted to evaluate
the potential association between WNT5A/PCP
Blood 8
----- --- --- --- --- --- --- --- --- --- --- --- ---
signaling in adipose tissue and
CancerCell 8 atherosclerotic CVD, given the major role that
---------- --- --- --- --- --- --- ----------------------------------------------- --- --- --- --- ---
Cell 51 IL-6 signaling plays in this condition as
revealed by large Mendelian randomization
CellMetabolism 10 studies 44, 45 ."
-------------- --- --- --- --- --- --- -------------- --- --- --- --- ---
CellStemCell 41
Claim
Circulation 12
----------- --- --- --- --- --- --- --- --- --- --- --- ---
IL-6 signaling plays a major role in
Immunity 33
-------- --- --- --- --- --- --- --- --- --- --- --- ---
atherosclerotic cardiovascular disease.
JAMA 79
---- --- --- --- --- --- --- --- --- --- --- --- ---
MolecularCell 27 Figure 4: A claim written based on a citance. Mate-
MolecularSystemsBiology 5 rialunrelatedtothecitationisremoved. Theacronym
“CVD”isexpandedto“cardiovasculardisease”.
Nature 29
----------------- --- --- --- --- --- --- --- --- --- --- --- ---
NatureCellBiology 26
NatureCommunications 19 ten. Theseretrievalswouldbeincorrectlymarked
NatureGenetics 8
-------------- --- --- --- --- --- --- --- --- --- --- --- ---
wrongbyourevaluationmetrics.
NatureMedicine 89
-------------- --- --- --- --- --- --- --- --- --- --- --- ---
Distractorabstractsasdefinedin§3.1havetwo
NatureMethods 1
-------------------- --- --- --- --- --- --- --------- ------- --------- ---------- ------------- ------------
qualities that make them a good addition to the
NucleicAcidsResearch 10
SCIFACT corpus: (1) They are cited in the same
PlosBiology 36
articles as our evidence abstracts, meaning that
PlosMedicine 38
theyoftendiscusssimilartopicsandincreasethe
Science 7 difficulty of abstract retrieval methods based on
---------------------------- --- --- --- --- --- --- ------------------------------------------- ---------------------------------- -------------------------- --- ---------- ---------
ScienceTranslationalMedicine 2
lexicalsimilarity. (2)Theauthorsofourcitances
TheLancet 22
were aware of the distractor abstracts, and chose
Other 120 nottomentiontheminthecitancesusedtogenerate
claims. Thismakesthemunlikelytobeasourceof
Total 741
falsenegativeretrievals.
Table8: Numberofciteddocumentsbyjournal. Some
------- -------------------------------- --- --- --- --- ---- --- ------------------ --- --- --- ---
C.2 Annotationexamples
co-citedarticles(§3.1)comefromjournalsoutsideour
curatedset;theseareindicatedby“Other”. Converting citances to claims Figure 4 shows
an example of a citance re-written as a claim.
--- --- --- --- --- --- --- ----------- --- ------------ ---------- ------------ -----------
The citance discusses the relationship between
C Datasetcollectionandcorpusstatistics
“atheroscleroticCVD”and“IL-6”,andcitestwopa-
C.1 Corpus pers(44and45)asevidence. Toconverttoaclaim,
---------- --- --- --- --- --- --- ------------------------ --- --- --- ------------------ ---
theacronym“CVD”isexpandedto“cardiovascu-
Source journals Table 8 shows the number of
------ -------- ----- ------- --- ---------- --- --- --- --- --- --- ---
lardisease”,irrelevantinformationisremoved,and
citedabstractsfromeachofourselectedjournals.
theclaimiswrittenasanatomicfactualstatement.
The “Other” category includes “co-cited” (§3.1)
----------- -------- --------- -------- ---------- --------- ------ ------------------ --- --- ---------------------- --- ---
abstracts that came from journals not among our
Multiplerationales Figure5showsaclaimsup-
pre-definedset.
ported by two rationales from the same abstract.
--- --- --- --- --- --- --- ------ ------ ---------- ---- -------- ---------
Thetextofeachrationaleonitsownissufficientto
Distractor abstracts In §3.1, we mention how
----------- ---------- ------- ---------- --- --------- ---- --------------- --------------------------- --- --- --- ---
we increase the size of the corpus by adding dis- entailtheclaim.
tractor abstracts. The reason why we do not use
C.3 Annotatorsandqualitycontrol
theentiretyofalargeresearchcorpuslikeS2ORC
asourfact-checkingcorpusisthatdoingsowould Claimwriting Studentclaimwritersattendedan
introducemanyfalsenegativeretrievals: abstracts in-person training session where they were intro-
containingevidencerelevanttoagivenclaim,but ducedtothetaskandreceivedin-personfeedback
notmentionedintheclaim’ssourcecitance. This fromthefourexperts. Followingtraining,student
canoccureitherbecausethecitanceauthorssimply annotatorscontinuedwritingclaimsremotely. The
were not aware of these abstracts, or because the expertannotatorsmonitoredclaimsforqualitydur-
abstractswerepublishedafterthecitancewaswrit- ing the remote annotation process, and provided
7547

https://scifact.s3-us-west-2.amazonaws. Claim Antibiotic induced alterations in the gut com/doc/evidence-annotation-instructions. microbiome reduce resistance against pdf. Clostridium difficile

Decision: SUPPORTS
Antibiotics can have significant and long-
lasting effects on the gastrointestinal tract
microbiota, reducing colonization resistance
against pathogens including Clostridium
difficile.
Rationale 1
Our results indicate that antibiotic-mediated
alteration of the gut microbiome converts the
global metabolic profile to one that favours
C. difficile germination and growth.
Rationale 2
Figure5:Aclaimsupportedbytworationalesfromthe
same abstract. The text of each rationale on its own
-------------- --- ---- ------- --------- ----------
providessufficientevidencetoverifytheclaim.
feedbackwhennecessary;low-qualityclaimswere
returnedtotheannotatorsforre-writing. Asafinal
------------------------------------- ---------- --- ------ -------------- --------
check, all submitted claims were proofread (and
edited if necessary) by an undergraduate whose
claimsweredeemedespeciallyhigh-qualitybythe
expertannotators.
Claim negations As mentioned in §3.2, an ex-
--------------- --- --- --------- -------- ------
pertannotatorwroteclaimnegationstointroduce
caseswhereanabstractREFUTESaclaim. Thean-
---------------------------------- --- --- --- --- ------
notatorskippedclaimsthatcouldonlybenegated
by adding obvious triggers like “not”. The ma-
--------- ------- --------- -------- ----------- -----------
jority of claim negations involved a reversal of
effectdirection;forinstance“Ahighmicroerythro-
cytecountprotectsagainstsevereanemia”canbe
negatedas“Ahighmicroerythrocytecountraises
vulnerabilitytosevereanemia”.
Claimverification Annotationswereperformed
----------------------------- ------- ------------- ------------------------ -------------- ----------
remotelythroughawebinterface. Annotatorswere
required to pass a 10-question “quiz” before an-
notatingtheirownclaims. Afterpassingthequiz,
subsequentsubmissionswerereviewedbyanNLP
expertuntilthatexpertdeemedtheannotatorreli-
able. Approvedannotatorswerethenassignedto
revieweachothers’submissions. Ingeneral,grad-
----------------------------- --- --- --- --------------- ---
uatestudentswereassignedtoreviewannotations
fromundergraduates.
D Annotationinterfacesandguidelines
Weshowascreenshotoftheclaimwritinginterface
inFigure6,andtheclaimverificationinterfacein
Figure7. Thecompleteannotationguideforclaim
------------ ---------------------------------- --------- --- ------------- ----
verification is available at the following URL:
7548

Figure 6: The claim-writing interface. The citation sentence is highlighted in blue on the top left. Additional context is provided on bottom left. The right side shows two claims that could be written based on this citation sentence. 7549

Figure7: Theevidencecollectioninterface. 7550