Designing a Belief Function-Based Accessibility Indicator to Improve Web Browsing for Disabled People

0.61 MB·English

by  Jean-Christophe Dubois

Designing a belief function-based accessibility
indicator to improve web browsing
for disabled people
Jean-ChristopheDubois,YolandeLeGallandArnaudMartin
5
1
0
2

n
a AbstractThepurposeofthisstudyistoprovideanaccessibilitymeasureofweb-
J
pages,inordertodrawdisableduserstothepagesthathavebeendesignedtobeac-

0
cessibletothem.Ourapproachisbasedonthetheoryofbelieffunctions,usingdata
2
whicharesuppliedbyreportsproducedbyautomaticwebcontentassessorsthattest

] thevalidityofcriteriadefinedbytheWCAG2.0guidelinesproposedbytheWorld
C WideWebConsortium(W3C)organization.Thesetoolsdetecterrorswithgradual
H degreesofcertaintyandtheirresultsdonotalwaysconverge.Forthesereasons,to
. fuseinformationcomingfromthereports,wechoosetouseaninformationfusion
s
c framework which can take into account the uncertainty and imprecision of infor-
[ mation as well as divergences between sources. Our accessibility indicator covers

fourcategoriesofdeficiencies.Tovalidatethetheoreticalapproachinthiscontext,
1
v weproposeanevaluationcompletedonacorpusof100mostvisitedFrenchnews
2 websites, and 2 evaluation tools. The results obtained illustrate the interest of our
9 accessibilityindicator.
7
4
0
.
1 1 Introduction
0
5
1 TheWebconstitutestodayanessentialsourceofinformationandcommunication.
: Whileusershaveagrowinginterestintermsofsocial,culturalandeconomicvalue,
v
i andinspiteoflegislationsandrecommendationsoftheW3Ccommunityformaking
X
websitesmoreaccessible,itsaccessibilityremainshardlyefficientforsomedisabled
r orageingusers.Actually,makingwebsitesaccessibleandusablebydisabledpeople
a
isachallenge[8]thatsocietyneedstoovercome[1].
Jean-ChristopheDubois,YolandeLeGall,ArnaudMartin
DRUID-IRISA,UniversityofRennes1,rueEdouardBranly,22300Lannion,FRANCE
e-mail:[email protected]
e-mail:[email protected]
e-mail:[email protected]
1
2 Jean-ChristopheDubois,YolandeLeGallandArnaudMartin
To measure the accessibility of a webpage, several accessibility metrics have
been developed [16]. Evaluations are based on the failure to comply with the rec-
ommendations of standards, using automatic evaluation tools. They often give a
finalvalue,continuousordiscrete,torepresentcontentaccessibility.However,the
fact remains that tests on accessibility criteria are far from being trivial [2]. Eval-
uationreportsofautomaticassessorscontainerrorsconsideredascertain,butalso
warningsorpotentialproblemswhichareuncertain.Moreovertherearedifferences
betweenassessorevaluations,evenforerrorsconsideredascertain.
This work provides a new measure of accessibility and an information fusion
framework to fuse information coming from the reports of automatic assessors al-
lowingsearchenginestore-ranktheirresultsaccordingtoanaccessibilitylevel,as
someuserswouldlike[10].Thisaccessibilityindicatorconsidersseveralcategories
ofdeficiencies.Ourapproachisbasedonthetheoryofthebelieffunctionsadapted
totakeintoaccountthedefectsofaccessibilitygivenbyseveralautomaticassessors
seenasinformationsources,theuncertaintyoftheirresults,aswellasthepossible
conflictsbetweenthesources.
Inthesections2and3wewillgiveadescriptionofaccessibilitytoolsbasedon
a recent standard and of data provided in their reports. In the 4th section, we will
describe the principles of our indicator and develop how we implement the belief
functions.Inthe5thpart,wewillpresentanexperimentbeforeconcluding.
2 Defectdetectionofwebpageaccessibility
Variousaccessibilitystandardsproposerecommendationsforimprovingaccessibil-
ity of webpages. The Web Content Accessibility Guidelines (WCAG 2.0) [3] pro-
posedbytheW3Cnormalizationorganization,constitutesaninternationalreference
in the field. These guidelines cover a wide range of disabilities (visual, auditory,
physical,speech,cognitive,etc.)andseverallayersofguidanceareprovided:
• 4overallprinciples:perception,operability,understandability&robustness;
• testablesuccesscriteria:foreachguideline,testablesuccesscriteriaareprovided.
Everycriterionisassociatedtooneofthe3definedconformancelevels(A,AA
andAAA),eachrepresentingarequirementofaccessibilityforusers.
Severalautomaticaccessibilityassessors,basedonvariousaccessibilitystandards,
havebeendeveloped[5]forITprofessionals.Theirlimitsdependontheautomatic
tests. Because it is at present not possible to test some criteria about the quality
of some pages, some assessor results are given with ambiguity. Consequently, the
existing automatic assessors look for the criteria which are not met and give the
defectsaccordingto3levelsofvalidity:thenumberoferrors,whichareestimated
certain, the number of likely problems (warnings) whose reality is not guaranteed
and the number of potential problems (also called generic or non testable) which
leadstoacompleteuncertaintyonthetestedcriterionaccessibility.
Designingabelieffunction-basedaccessibilityindicatortobrowsewebpages 3
Finally, even though the results obtained by different assessors match for some
testedcommoncriteria,resultscandiffer,evenforerrorsconsideredascertain.
3 Proposedaccessibilityindicator
Afterarequest,theindicatorhastosupplyinformationdescribingtouserstheac-
cessibility level of each webpage proposed by a search engine. Presented simulta-
neouslywiththesepages,theindicators’informationcovertwoaspects:
• the accessibility for categories of deficiencies: as previously proposed for ac-
cessibility estimation [6] we use 4 major categories: visual, hearing, motor and
cognitivedeficiencies,asdefinedby[15].Theyarecalled“deficiencyframes”;
• thelevelofaccessibilityforeachdeficiencyframe.
Collectingresultsfromseveralassessorshasallowedustobenefitfromeachoftheir
performance. In addition, it strengthens accessibility evaluation for similar results
and manages conflicts in case of disagreements. Automatic assessors check a set
of criteria which correspond to many deficiencies. As our accessibility evaluation
variesforeverydeficiencyframe,ourmethodconsistsinselectingtherelevantcri-
teria for each deficiency frame and then balancing each criterion to consider the
difficultiesmetbyusersincaseoffailure.Thisweightingisbasedonthecriterion
conformance level (A, AA, AAA), which corresponds to decreasing priorities (A:
most important, etc.). The errors and problems detected for every criterion of the
accessibilitystandardaffecttheaccessibilityindicatoroftheWebcontenttestedac-
cording to the deficiency frame the criterion belongs to, its weighting within the
frame, the number of occurrences when it is analyzed as a defect in the webpage
andthedefect’sdegreeofcertainty(error,likelyorpotentialproblem).
4 Defectdetectionandaccessibilityevaluation
AftercollectingwebpageUniformResourceLocators(URL )selectedbyasearch
p
enginefromarequest,theseaddressesaresuppliedtotheaccessibilityassessorsand
successivelyforeachpage,wedetectaccessibilitydefects,thenestimateaccessibil-
itylevelbydeficiencyframeforeachassessor,beforefusingthedatabydeficiency
frameandtakingthedecisionforeverydeficiencyframe[7].
4 Jean-ChristopheDubois,YolandeLeGallandArnaudMartin
4.1 Assessorevaluationsofselectedpages
EachURL issubmittedtotheaccessibilityevaluationtestsbyeachassessorithat
p
testsallthecriteriakoftheWCAG2.0standard,andthefollowingdataarecollected
byafilterthatextractstherequireddataforeachdeficiencyframe:
• Ne :errorsobservedforacriterionkbyanassessori;
k,i
• Nc :correctcheckpointsforacriterionkbyanassessori;
k,i
• Te :teststhatcaninduceerrorsforacriterionkbyanassessori;
k,i
• Nl :likelyproblemsdetectedforacriterionkbyanassessori;
k,i
• Tl :teststhatcaninducelikelyproblemsforacriterionkbyanassessori;
k,i
• Np :potentialproblemssuspectedforacriterionkbyanassessori;
k,i
• Tp :teststhatcaninducepotentialproblemsforacriterionkbyanassessori;
k,i
• T :totaltestsbyanassessori,with:
i
T =∑(Ne +Nl +Np +Nc ) (1)
i k,i k,i k,i k,i
k
4.2 Accessibilityindicatorlevelofthepages
Tomodelinitialinformationincludinguncertainties,thereliabilityoftheassessors
seenasinformationsourcesandtheirpossibleconflicts,weusethetheoryofbelief
functions[4][13].Ourobjectiveistodefineifawebpageisaccessible(Ac)ornot
accessible(Ac)andtosupplyanindicationbydeficiencyframe.Consequently,these
questionscanbehandledindependentlyforeverydeficiencyframeΩ ={Ac,Ac}.
h
Wecanconsidereverypowerset2Ωh ={0/,Ac,Ac,Ω}.
The estimation of the accessibility Ac for a deficiency frame h and a source i
(assessor) is estimated from the number of correct tests for each of the criteria k
occurringinthisframe,andfromtheirconformancelevelrepresentedbyα :
k
∑k(Nkc,i∗αk)
E(Ac) = (2)
h,i
T
i
TheestimationofthenonaccessibilityAcforadeficiencyframehandasource
iisestimatedfromthenumberoferrorsforeachofthecriteriak occurringinthis
frame,andfromtheα coefficient.Aweakeningβecoefficientisalsointroducedto
k i
modelthedegreeofcertaintyoftheerror:
∑k(Nke,i∗αk∗βie)
E(Ac) = (3)
h,i Te
k,i
The estimation of the ignorance Ω for a deficiency frame h and a source i is es-
h
timated from the number of likely and potential problem for each of the criteria k
Designingabelieffunction-basedaccessibilityindicatortobrowsewebpages 5
occurringinthisframe,andfromtheα coefficient.Theweakeningcoefficientsβl
k i
p
orβ arealsousedtomodelthedegreeofcertaintyoftheproblem:
i
∑k(Nkl,i∗αk∗βil+Nkp,i∗αk∗βip)
E(Ω = (4)
h,i ∑ (Tl +Tp)
k k,i k,i
Themassfunctionsofthesubsetsof2Ωh arecomputedfromtheestimations:
E(Ac)
h,i
m(Ac) = (5)
h,i
E(Ac) +E(Ac) +E(Ω)
h,i h,i h,i
E(Ac)
h,i
m(Ac) = (6)
h,i
E(Ac) +E(Ac) +E(Ω)
h,i h,i h,i
E(Ω)
h,i
m(Ω) = (7)
h,i
E(Ac) +E(Ac) +E(Ω)
h,i h,i h,i
Inaddition,thesourcereliabilitycanbemodeled[11]withaδ coefficient,which
i
constitutesabenefitwhensomeassessorsaremoreefficientthanothers:

 mδi(Ac)h,i=δi∗m(Ac)h,i
mδi(Ac)h,i=δi∗m(Ac)h,i (8)
mδi(Ω)h,i=1−δi∗(1−m(Ω)h,i)
4.3 Mergingassessorresultsanddecision-making
Once the masses for each assessor have been obtained, a fusion of the results is
conducted by deficiency frame, using the conjunctive rule [14], to combine them
and give information in the form of a mass function. These rule properties, which
strengthencommonresultsandmanageconflictsbetweensources,areparticularly
relevant in this context, to deal with divergences between assessor evaluations. To
calculate the final decision D (URL ) for a page by deficiency frame, we use the
h p
pignisticprobability[14].
Thereareseveralwaysofpresentingtheaccessibilityindicatortousers.Tovisu-
alizethedeficiencyframes,existingspecificpictogramsareeffective.Topresentthe
accessibilitylevelwediscretizethedecisioninto5levels(verygood,good,moder-
ate,badorverybadaccessibility)usingthresholdsandvisualizeditbyan”arrow”:
• ifD <S ,theWebcontentaccessibilityisverybad(↓),
h 1
• ifS <D <S ,theWebcontentaccessibilityisbad((cid:38)),
1 h 2
• ifS <D <S ,theWebcontentaccessibilityismoderate(→),
2 h 3
• ifS <D <S ,theWebcontentaccessibilityisgood((cid:37)),
3 h 4
• ifS <D ,theWebcontentaccessibilityisverygood(↑).
4 h
6 Jean-ChristopheDubois,YolandeLeGallandArnaudMartin
5 Experiments
Tovalidateourapproach,wepresentheretheresultsobtainedonasetof100news
Websites, among the most visited ones, all referenced by the OJD 1 organization
whichprovidescertificationandpublicationofattendancefiguresforwebsites.We
test their homepages, following a study [12] concluding that their usability is pre-
dictiveofthewholesite.WechosetwoopensourceassessorsAChecker,(source1)
[9],andTAW(source2)fromwhichweextractautomaticallytheaccessibilitytest
results. Weight and threshold values given in Table 1 were previously empirically
definedfromWebpages2assumedtobeaccessible.
α1;α2;α3 A,AA,AAAconformancelevels 1;0.8;0.6
Weightings βe;βl;βp Certaintylevelsoferrorsorproblems 1;0.5;1
i i i
δ1;δ2;α3 ACheckerandTAWreliabilities(sources) 1;1
Thresholds S1;S2;S3;S4 Accessibilityindicatorlevels 0.6;0.7;0.8;0.9
Table1:Constantvaluesforouraccessibilitymetric.
TheresultsofthesesourcesaresummarizedinFigure1forthe3levelsofcer-
tainty defects. The box plots present how their defects are distributed: minimum
andmaximum(whiskers),1st (bottomboxplot)and3rd quartiles(topboxplot)and
average(horizontalline).Weobservesimilaritiesbetweentheassessors’resultsfor
theerrorsdetectedascertain,butalsohugedifferencesforthelikely(warnings)and
potential(nontestable)problems.Thenumberoflikelyproblemsisalmostnullfor
ACheckerandthepotentialoneremainsalwaysthesameforTAW.
Fig.1:Resultsofautomaticassessors.
Thedetecteddefectsaretakenintoaccountinouraccessibilityindicatorresults
presented in Figure 2. The mass function values of accessibility m(Ac) for the 2
sources, TAW and AChecker, and the fusion result are visualized for 3 deficiency
1OJD:http://www.ojd.com/Chiffres/Le-Numerique/Sites-Web/Sites-Web-GP
2SiteslabeledbyAccessiweb:http://www.accessiweb.org/index.php/galerie.html
Designingabelieffunction-basedaccessibilityindicatortobrowsewebpages 7
framesamongthe4,andgloballyforalldeficiencies.Firstly,wecanseethatm(Ac)
is not evenly distributed between the 2 sources: their distributions of errors (Fig-
ure 2) are comparable even if there is a larger range for AChecker; however the
mass function of accessibility is smaller for AChecker compared to TAW. This is
duetothemorenumerouspotentialproblems(nontestablecriteria)detectedbythe
ACheckerassessor,increasingsubstantiallythedenominatorinthecomputationof
m(Ac)(Eq.5).Bytheway,thevaluesofE(Ω)andconsequentlyofm(Ω),aremore
important,astheβp weightforpotentialproblemsis2timeshigherthanβl forthe
i i
likelyproblems(warnings).Wecanalsonoticethatthefusionresultobtainedbythe
conjunctiverulestrengthensthemassfunctionsofthe2assessors.
Fig.2:Accessibilityindicatorresults.
In this corpus, visual and cognitive deficiencies have a higher impact on con-
tent accessibility than the motor ones. This is logical for news websites, as their
homepages include a large number of images. By the way, the motor indicator is
lessimpacted,inparticularbythelackofalternativesforimages,usefulforvisual
andcognitivedeficiencies.Finally,weobserveasimilaritybetweenthevisualand
globalindicators,asaround80%ofallthecheckpointsconcernvisualdeficiencies
andalsobecausethesecontrolsareproperlytakenintoaccountbyassessors.
Decision
Webcontent(URLp) Visual Motor Cognitive Global
LeParisien.fr 0.972↑ 0.989↑ 0.974↑ 0.971↑
Famili.fr 0.769→ 0.924↑ 0.838(cid:37) 0.766(cid:37)
Arte.tv 0.701→ 0.718→ 0.717→ 0.686(cid:38)
LePoint.fr 0.630(cid:38) 0.725→ 0.673(cid:38) 0.627(cid:38)
Table2:Examplesofdetailedaccessibilityresultsbydeficiencyframe.
In Table 2 are presented detailed results for several sites with significant indi-
catorresultdifferences.Forexamples,LePoint.frandArte.tv,respectively19th and
33th mostconsultedwebsitesinFrance,obtainonly0.627and0.686fortheglobal
result,whereasLeParisien.fr,ranked12th,reaches0.971.ForFamily.frweobserve
8 Jean-ChristopheDubois,YolandeLeGallandArnaudMartin
differences between the deficiencies, nevertheless focus on accessibility generally
benefitsalldeficienciesonthewholecorpus.
6 Conclusion
We present an indicator estimating webpage accessibility levels for distinct cate-
gories of deficiencies, in order to supply easily understandable accessibility infor-
mationtousersonpagesproposedbyasearchengine.Ourmethodbasedonbelief
function theory fuses results from several automatic assessors and considers their
uncertainties. An accurate modelization of the assessor characteristics and of the
impactofdefectguidelinecriteriaonaccessibilityisproposed.Anexperimentper-
formed on a set of 100 news websites validates the method, which benefits from
each of the assessor performances on specific criterion tests. Our future research
willfocusontheimplementationofauser’spersonalweightingtobalancetheim-
portanceofcriteria.
References
1. AbascalJ.,ArrueM.,FajardoI.,GarayN.andTomhasJ.Theuseofguidelinestoautomati-
callyverifywebaccessibility.UniversalAccessintheInformationSociety,3(1):71-79,2004.
2. BrajnikG..Comparingaccessibilityevaluationtools:amethodfortooleffectiveness.Univer-
salAccessintheInformationSociety,3(3-4):252-263,2004.
3. BrajnikG..WebAccessibilityTesting:WhentheMethodistheCulprit.InternationalConfer-
enceonComputersHelpingPeoplewithSpecialNeeds,ICCHP,156-163,2006.
4. Brajnik G.. Validity and Reliability of Web Accessibility Guidelines. ACM SIGACCESS
ConferenceonComputersandAccessibility,ASSETS,131-138,2009.
5. Caldwell B., Cooper M., Guarino Reid L. and Vanderheiden G.. Web Content Acces-
sibility Guidelines (WCAG) 2.0, W3C Recommendation, December 11th 2008. DOI:
http://www.w3.org/TR/WCAG20/.AccessedJun2010.
6. DempsterA.P..UpperandLowerprobabilitiesinducedbyamultivaluedmapping.Annalsof
MathematicalStatistics,38:325-339,1967.
7. DuboisJ.-C.,LeGallY.andMartinM..SoftwareEngineeringSystemfortheAnalysisof
AccessibilityData.PatentNo.1451562,FiledFebruary26th2014.
8. EuropeanDisabilityStrategy2010-20:ARenewedCommitmenttoaBarrier-FreeEurope.
COM636final.,2010.
9. GayG.andQiLiC..AChecker:open,interactive,customizable,webaccessibilitychecking.
InW4AInternationalCrossDisciplinaryConferenceonWebAccessibility,(33),ACMPress,
2010.
10. IvoryM.Y.,YuS.andGronemyerK..Searchresultexploration:apreliminarystudyofblind
andsightedusers’decisionmakingandperformance.CHIExtendedAbstracts,1453-1456,
ACMPress,2004.
11. MartinA.,OsswaldC.,DezertJ.andSmarandacheF..Generalcombinationrulesforqualita-
tiveandquantitativebeliefs,JournalofAdvancesinInformationFusion,3(2):67-82,2008.
12. NielsenJ.andTahirM..HomepageUsability:50WebsitesDeconstructed.NewRidersPub-
lishing,2001.
13. ShaferG..Amathematicaltheoryofevidence.PrincetonUniversityPress,1976.
Designingabelieffunction-basedaccessibilityindicatortobrowsewebpages 9
14. Smets P.. Belief Functions: the Disjunctive Rule of Combination and the Generalized
BayesianTheorem.InternationalJournalofApproximateReasoning,9:1-35,1993.
15. VanderheidenG.C.andVanderheidenK.R..Accessibledesignofconsumerproducts:guide-
linesforthedesignofconsumerproductstoincreasetheiraccessibilitytopeoplewithdis-
abilitiesorwhoareaging.Madison-WI,TraceR&DCenter,1991.
16. VigoM.andBrajnikG..Automaticwebaccessibilitymetrics:whereweareandwherewe
cango.InteractingwithComputers23(2):137-155,Elsevier,2011.