argumentation and reasoned actiontbc/publications/eca_lisbon_proceedings... · isbn...

Argumentation and Reasoned Action

Proceedings of the 1st European Conference on Argumentation,

Lisbon 2015

Volume II

Edited by

Dima Mohammed and

Marcin Lewiński

© Individual author and College Publications 2016 All rights reserved. ISBN 978-1-84890-212-1 College Publications Scientific Director: Dov Gabbay Managing Director: Jane Spurr http://www.collegepublications.co.uk Original cover design by Orchid Creative www.orchidcreative.co.uk Printed by Lightning Source, Milton Keynes, UK All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form, or by any means, electronic, mechanical, photocopying, recording or otherwise without prior permission, in writing, from the publisher.

D.Mohammed&M.Lewiński(eds.)(2016).ArgumentationandReasonedAction:Proceedingsofthe1stEuropeanConferenceonArgumentation,Lisbon,2015.Vol.II,163-180.London:CollegePublications.

163

2

UsingAbstractDialecticalFrameworkstoArgueaboutLegalCases

LATIFAAL-ABDULKARIM

DepartmentofComputerScience,UniversityofLiverpool,[email protected]

KATIEATKINSON


TREVORBENCH-CAPON


Recent work has shown how to map factor hierarchies forlegal reasoning intoAbstractDialectical Frameworks (ADFs),bydefiningacceptanceconditionsforeachnode.Inthispaperwe model as ADFs bodies of case law from various legaldomains,rewritethemaslogicprograms,comparetheresultswith previous legal reasoning systems and proposeimprovements by increasing the scope of reasoningdownwardstofacts.

KEYWORDS: ADF, case based-reasoning, factors, legalreasoning

1.INTRODUCTIONA recent development in computational argumentation has beenAbstract Dialectical Frameworks (ADFs) by Brewka and Woltran(2010).ADFscanbeseenasageneralisationofstandardArgumentationFrameworks (AFs) (Dung, 1995) in which the nodes representstatementsratherthanabstractarguments,andeachnodeisassociatedwithanacceptanceconditionthatdetermineswhenanodeisacceptablein terms of whether its children are acceptable. Thus links in AFsexpressonlyonerelationship,namelydefeat,butADFscanrepresenta

LatifaAl-Abdulkarimetal.

164

varietyofattackandsupportrelations.Inconsequence,whereasnodesinanAFhaveonlythesingleacceptanceconditionthatalltheirchildrenare defeated, nodes in ADFs can have different acceptance conditionsspecificallytailoredforeachnode.InAl-Abdulkarimetal.(2014)itwasargued that ADFs are very suitable for representing factor basedreasoningwithlegalcasesasfoundintheCATOsystem(Aleven,1997)andasformalisedinPrakkenandSartor(1998).

The key idea inAl-Abdulkarim et al. (2014) is that the abstractfactorhierarchyofCATO(Aleven,1997) (anextractgiven inFigure1)correspondsdirectlytothenodeandlinkstructureofanADF,orrather(sincethelinksarelabelled“+”or“−”)aPrioritisedADF(PADF)Brewkaetal.(2013)whichpartitionslinksintosupportingandattackinglinks,andsocorrespondstothelabelsonthelinksinthefactorhierarchy.ToexpressCATO’sfactorhierarchyasanADF,acceptanceconditionsneedto be supplied for each of the nodes. Finally the logical model of IBP(Brüninghaus&Ashley,2003)canbeusedtotiethevariouspartsofthefactorhierarchytogethertosupplydecisionsforparticularcases.InAl-Abdulkarim et al. (2014) it was suggested that the acceptanceconditions couldbeexpressedasPrologprocedures.These could thenbe used directly to form a Prolog program that could be executed toclassifycases(astowhichsidetheyaredecidedfor)representedassetsoffactors.

In this paper we will evaluate this approach. We will use USTrade secrets as the domain, allowing us to use the analysis of CATOwhich will permit direct comparison with CATO, IBP, the varioussystemsusedascomparators inBrüninghaus&Ashley(2003)andtheAGATHA system of Chorley and Bench-Capon (2005). We will firstlyconsideraquantitativeanalysis,intermsofperformanceandhoweasilytheprogramcanberefinedtoimproveperformance,andthenconsiderthe system in terms of the transparency of its outputs, the relation tocase decision texts, and the relation to formal frameworks forstructured argumentation such as Prakken (2010). Finally we willconsiderwhetherthemethodcanbereadilyappliedtootherdomains,by briefly describing an application of the ADF approach to Popov vHayashiandrelatedcasesasmodelledinBench-Capon(2012).2.BACKGROUNDInthissectionwewillrecapitulatetheessentialsofADFs(Brewkaetal.,2013),CATO(Aleven,1997)andIBP(Brüninghaus&Ashley,2003).

Abstractdialecticalframeworks

165

2.1AbstractDialecticalFrameworksAnADFisdefinedinBrewkaetal.(2013),as:

Definition1:AnADFisatupleADF=<S,L,C>whereS isthesetofstatements(positions,nodes),LisasubsetofS×S,asetoflinks,andC={Cs∈S}isasetoftotalfunctionsCs:2par(s)→{t,f}, one for each statement s. Cs is called the acceptanceconditionofs.

In a Prioritised ADF, L is partitioned into L+ and L−, supporting andattacking links, respectively. Although the acceptance conditions areoftenexpressedaspropositionalfunctions,thisneednotbethecase:allthat is required is thespecificationof conditions for theacceptanceorrejectionofanodeintermsoftheacceptanceorrejectionofitschildren.2.2CATOCATO(Aleven,1997),whichwasdevelopedfromRisslandandAshley’sHYPO (1990), takes as its domain US Trade Secret Law. CATO wasprimarilydirectedatlawstudents,andwasintendedtohelpthemformbetter case based arguments, in particular to improve their skills indistinguishingcases, andemphasisinganddownplayingdistinctions.Acore ideawas to describe cases in terms of factors, legally significantabstractions of patterns of facts found in the cases, and tobuild thesebase-level factors into an hierarchy of increasing abstraction, movingupwardsthroughintermediateconcerns(abstractfactors)toissues.AnextractfromthefactorhierarchyisshowninFigure1.

Figure1–CATOAbstractFactorHierarchyfromAleven(1997)


166

Each factor favours either the plaintiff or the defendant. The programmatchesprecedent caseswitha current case toproducearguments inthree plies: first a precedent with factors in common with the caseunderconsiderationiscited,suggestingafindingforoneside.Thentheother side cites precedents with factors in common with the currentcase but a decision for the other side as counter examples, anddistinguishes the cited precedent by pointing to factors not shared bythe precedent and current case. Finally the original side rebuts bydownplayingdistinctions,citingcasestoprovethatweaknessesarenotfatal anddistinguishing counter examples. CATOused twenty-six baselevelfactors(thereisnoF9),asshowninTable1.

Table1-BaseLevelFactorsinCATO


167

Thereis,however,nosinglerootforthefactorhierarchyaspresentedinAleven(1997):ratherwehaveacollectionofhierarchies,eachrelatingto a specific issue. To tie them together we turn to the Issue BasedPrediction(IBP)systemofBruninghausandAshley(2003).2.3IssueBasedPredictionInIBP,whichisfirmlybasedonCATO,theaimisnotsimplytopresentarguments, but to predict the outcomes of cases (find forplaintiff/defendant).Toenable this, the issuesofCATO’shierarchyaretied together using a logical model derived from the Uniform TradeSecretActandtheRestatementofTorts.ThemodelisshowninFigure2.

Figure2–IBPLogicalModelfromBrüninghaus&Ashley(2003)

Nowconsider the factorhierarchy,partofwhich isshown inFigure1.WecannowregardthisasanADFbyformingthesetSfromtheissues,intermediateconcernsandbaselevelfactors,L+fromthelinkslabelled“+” and L− from the links labelled “−”. Using the complete factorhierarchygiveninFigures3.2and3.3of(Aleven,1997)wewillhaveanADFwhichhasas its leafnodes thebase level factorsofCATO.This isdescribedintabularforminTable2.


168

TherootsofCATO’shierarchiescorrespondtothe leavesof the logicalmodel: we can therefore form them into a single ADF by using thisstructure.TherelevantadditionstotheADFneededtointegratetheIBPmodel are shown in Table 3 (note that F124 is not discussed inBrüninghaus&Ashley,2003).

Table3-IBPLogicalModelasADF

IBP used 186 cases, 148 cases analysed for CATO and 38 analysedspecifically for IBP. Unfortunately, these cases are not all publiclyavailableandsowewillusethe32casesharvestedfrompublicsourcesbyAlisonChorleyandusedtoevaluateherAGATHAsystem(2005).Aspart of the evaluation in Brüninghaus and Ashley (2003) nine othersystemswere also considered to provide a comparison.Most of thesewere different forms of machine learning system, but programsrepresenting CATO and HYPO were also included. IBP was the bestperformer:resultsreportedinIBP(Brüninghaus&Ashley,2003),NaiveBayes (the best performer of the ML systems), CATO, HYPO and a

Table2CATOasADF


169

versionofIBPwhichusesonlythemodel,withnoCBRcomponent,areshowninTable41.

Direct comparison with AGATHA is hampered by the fact thatevaluation in AGATHA was directed towards evaluating the differentheuristicsandsearchalgorithmsusedinthatsystem,andsonoversioncan be considered “definitive”, and, of course,many fewer caseswereusedintheexperiments.However,typically27-30ofthe32(≈84−93%)cases were correctly decided by the theories produced by AGATHA(Chorley&Bench-Capon,2005).3.ACCEPTANCECONDITIONSWe now supply acceptance conditions for each node, to supply theelementsofCrequiredfortheADF.WewillrelyonlyonthedefinitionsofthefactorsinAleven(1997).Wedonotuseprecedentsatthisstage;asAlevenremarks:

for certain conflicts, it is self evident how they should beresolved. For example, the fact that plaintiff’s product wasunique in the market (F15) arguably supports a conclusionthat plaintiff’s information (which is used to make theproduct)isnotknownoutsideplaintiff’sbusiness(F106),butnot if it isalsoknown, forexample,thatplaintiffdiscloseditsinformation in a public forum (F27). Common sense dictatesthatinthosecircumstances,theinformationisknownoutsideplaintiff’sbusiness.Itisnotnecessarytolooktopastcasestosupport that point. CATO’s use of link strength enables aknowledgeengineertoencodeinferenceslikethis.(p.47).

FromTables2and3wecanseethatwehaveeighteennodestoprovidewith acceptance conditions. One (F124) has only a single supporting

1NoexplanationforusingadifferentnumberofcasesforCATOandIBPmodelisgiveninBrüninghausandAshley(2003).

Table4-Resultsfrom[12]


170

child: thus the acceptance conditionwill be Parent←→Child.Wewillwrite this (and the other acceptance conditions) as a set of tests foracceptanceandrejection,tobeappliedintheordergiven,whichallowsus to express priority between them. The last test will always be adefault.Wechoosethisformofexpressionbecausewefinditeasiertoread in many cases, because it corresponds directly to the defeasibleruleswithprioritiesusedinformalismssuchasASPIC+(Prakken,2010),andbecause it isdirectlyusableasPrologcode.ThuswewriteParent←→Childas

AcceptParentifChild.RejectParent.

Where NOT is required we use negation as failure. The tests areindividually sufficient and collectivelynecessary, ensuring equivalencewith the logical expression (see Clark, 1978). Six nodes (F201, F203,F105, F108, F114 and F124) have only supporting links: these can bestraightforwardlyrepresentedusingANDandOR.WefollowedtheIBPmodel for the twonodes taken fromthatmodel (F201andF203),andused OR for the other four. The most complicated wasInfoMisappropriated(F201):

AcceptInfoMisappropriatedifF114ANDF112.AcceptInfoMisappropriatedifF110.RejectInfoMisappropriated.

Fivenodeshaveonesupportingandoneattacking link.Thesearebestseenasforminganexceptionstructure:accept(reject)theparentifandonly ifsupporting(attacking)childunlessattacking(supporting)child.Notethattheexceptionmaybethesupportingortheattackingchild:inthe former case thedefaultwill be reject, and in the latter thedefaultwillbeaccept.Thus:

AcceptParentifSupportAND(NOTAttack).RejectParent.RejectParentifAttackAND(NOTSupport).AcceptParent.

ForF110,F120andF121theattackingchildistheexception,whileforF122 and F123 the supporting links are the exceptions. This leavessevennodes.ForF200weregard theattacking linkasanexception tothecasewheretheconjunctionofthesupportinglinksholds.


171

For F104 and F112 we see the supporting links as offeringdisjointwaysofacceptingtheparent,andtheattackingchildasawayofestablishingthatthefactorisnotpresent.Wedefaulttoyesbecauseinmanycases thereareno factors foreithersidepresentrelating to thispoint.Wetakeitthatthisfactorwasoftensimplyacceptedonthefactsand uncontested, and so there was no discussion on the point. A fulldescription of all the truth conditions is given in Al-Abdulkarim et al.(2015).4.PROLOGPROGRAMTheProlog2programwas formedbyascending theADF, rewriting theacceptance conditions as groups of Prolog clauses to determine theacceptability of each node in terms of its children. The tests wererestatedusing theappropriatesyntax,withsomereporting to indicatewhether the node is satisfied (defaults are indicated by the use of“accepted that”), and some control to call the procedure to determinethenextnode,andtomaintaina listofaccepted factors.ExamplescanbefoundinAl-Abdulkarimetal.(2015).

Each of the tests in the acceptance condition is applied in aseparateclause,usingthesetoffactorscurrentlyidentifiedaspresentinthe case, before proceeding to the next factor,with the current factoraddedtotheapplicablefactorsif it isaccepted.Toallowcompletionofthedatabase(Clark,1978),afinalclauseisaddedtocatchanycasenotcovered by any of the preceding clauses. These defaults may favoureither side. In some cases, the default is accept because few casedescriptions related to these abstract factors, although they are a sinequanonforanyclaim.Ourbeliefisthattheseaspectswereuncontestedand so the factorswerenot explicitlydiscussed in the trial, and sodonot appear in the CATO analysis. Where we felt it was clear that thefactorneededtobeexplicitlyestablished,thedefaultwasreject.

The above demonstrates that it is a straightforward andreasonablyobjectiveprocess to transforma factorbasedanalysissuchas is found in (Aleven, 1997) to an executable program via an ADF.Although judgement was sometimes required to form the acceptanceconditions,wewouldsuggestthatsuchjudgementswerenotdifficulttomake. Moreover if there are difficult choices, the effect of thealternativescanbecomparedonasetoftestcases.Overalltherelativelysmallnumberof factors relevant toparticularnodesgreatly simplifiesthetask.

2 Prologwas used because of its closeness to the acceptance conditions, andmadetheimplementationquick,easyandtransparent.


172

5.RESULTSWecannowruntheprogramonthecases.Werepresentthecasesasalistofbase-levelfactors.Forexample,theBoeingcase3isrepresentedascase(boeing,[f4,f6,f12,f14,f21,f1,f10]).

givingoutput:1?-go(boeing).acceptedthatdefendantisnotownerofsecreteffortsmadevisavisoutsiderseffortsmadevisavisdefendanttherewasaconfidentialityagreementdefendantwasonnoticeofconfidentialitytherewasaconfidentialrelationshipacceptedthattheinformationwasusedquestionablemeanswereusedacceptedthattheinformationwasnotavailableelsewhereacceptedthatinformationisnotknownacceptedthattheinformationwasneitherknownnoravailableacceptedthattheinformationwasvaluablenotacceptedthattheinformationwaslegitimatelyobtainedimpropermeanswereusedeffortsweretakentomaintainsecrecyinformationwasatradesecretatradesecretwasmisappropriatedfindforplaintiffboeing[f200,f201,f203,f102,f110,f104,f111,f112,f114,f115,f121,f122,f123,f4,f6,f12,f14,f21,f1,f10]decisioniscorrect

Theinitialprogramcorrectlyclassified25outofthe32cases(78.1%).Whilealltenofthecaseswonbythedefendantwerecorrectlyclassified,seven of the 22 cases won by the plaintiff were not. The figure forcorrect answers is remarkably close to the 77.8% reported for theversion of CATO used in (Brüninghaus and Ashley, 2003), which, ofcourse,usesexactlythesameanalysisofthedomainandcasesthatwehaveadoptedhere.Thusasafirstconclusionwecantentativelysuggestthat executing the analysis in Aleven (1997) as an ADF produces verysimilar results to those obtainable using the original CATO program(albeitweareusingasmallersetofcases).Wecannowinvestigatehowtheinitialprogrammightbeimproved. 3TheBoeingCompanyv.SierracinCorporation,108Wash.2d38,738P.2d665(1987).


173

Thewronglypredictedcaseswere:

case(spaceAero,[f8,f15,f18,f1,f19]).case(televation,[f6,f12,f15,f18,f21,f10,f16]).case(goldberg,[f1,f10,f21,f27]).case(kg,[f6,f14,f15,f18,f21,f16,f25]).case(mason,[f6,f15,f21,f1,f16]).case(mineralDeposits,[f18,f1,f16,f25]).case(technicon,[f6,f12,f14,f21,f10,f16,f25]).

Examination of the cases showed that five of the seven had F16(ReverseEngineerable)presentandthatthesecasesweretheonlycasesfoundfortheplaintiffwithF16present.Theprobleminthesefivecasesis that theprogram finds for thedefendantbecause the information isavailable elsewhere (F105). This is established by the presence ofReverseEngineerable and is unchallengeable. Examination of the ADFshowsthatF16isimmediatelydecisive:ifthatfactorispresent,thereisno way the plaintiff can demonstrate that the information is a tradesecret. Goldberg4 also fails through F105 (information known oravailable),sincedisclosureinapublicforum(F27)issufficienttodenythe information trade secret status. It would appear that we couldsignificantly improveperformancebyrefining thisbranch toallowtheplaintiff some way to defend against, in particular, F16. See Al-Abdulkarimetal.(2015)fordiscussionofthetextsofthedecisionsandpossible refinements. The refined program can equal or better theperformanceofanyoftheexistingsystems.5.1DiscussionBy using the ADF we can readily explain the points at which theacceptance conditions do not concur with the decisions taken in theactualcases.Wecanthenreturntotheoriginaldecisionsandusethemtodeterminepossiblerefinementstotherepresentation.Insomecases,the problem seems to lie with the attribution of the factors. Suchmatterswerecontestedintheactualcase,andascribingthepresenceorabsenceofparticular factorsrequires interpretationof thecaseby theanalyst. The interpretation cannot be disputed without descending tothe level of facts as advocated by Atkinson et al. (2013) and Al-Abdulkarimetal.(2014).Additionofthefactlayerhasbeenthesubjectofworksubsequenttothatreportedhere.Otherdecisionssuggestthatwemaywishtomodifythedescriptionoffactorsintendedtoguidethe

4Goldbergv.Medtronic,686F.2d1219(7thCir.1982).


174

analyst. Adding or removing a factor to or from a particular caseprovides a local solutionwhichwill solve a problemwith a particularcase. Our results, however, indicated a general problem which wasapplicable to several cases: the dominant affect of F16, reverseengineerable.ItseemedcleartousthatthepresenceofF16shouldnotbyitselfbesufficientforafindingforthedefendant.Againthedecisionsthemselves suggested severalpossiblewaysof arguingagainstF16: inparticular the use of restricted materials and the uniqueness of theproduct.EitherorbothoftheseexceptionscouldbeincorporatedintheADF without adversely affecting any of the test cases, but we wouldneedtohaveareasonablylargesetofnewcasesinordertoevaluatethedifferentsolutionsandtoguardagainstoverfitting.

Finallyitshouldbeconcededthatthedecisionsthemselvesmaybeerroneous.Assumingthatthereareleastsomepoordecisionswhichwe would not wish to serve as precedents, we should be willing totolerateacertainnumberofdivergencesfromourresults.

Tosummarise:- Simply translating the analysis of (Aleven, 1997) into an ADF andexecuting the resultingprogramgave results almost identical to thosefound for CATO in the IBP experiments reported in Brüninghaus andAshley(2003).Notethatthisisachievedwithoutneedforbalancingofproandconfactorscentraltoexistingcasebasedreasoningsystems.-Thereasonsforthe“incorrect”decisionscanbereadilyidentifiedfromtheoutputandtheADF,aswesawfromthediscussionof thewronglydecidedcasesabove.- Examination of the texts of the decisions readily explained why theresults diverged, and suggested ways in which the analysis could beimproved,eitheratthecaselevelbychangingthefactorsattributed,orat the domain level by including additional supporting or attackinglinks.

From this we conclude that use of ADFs provides goodperformance, and has a number of positive features from a softwareengineering(anddomainanalysis)standpoint,whichwouldenabletheADFtoberefinedandperformance improved.Wealsobelievethatwedoneedtoincludeafact layertopermit increasedtransparencyintheascriptionoffactorstocases.6.QUALITYOFEXPLANATIONSAs the Prolog program proceeds it reports on the acceptability orotherwiseofthevariousabstractfactorsandtheresolutionofissues.Asshown above, this provides an excellent diagnostic for divergent


175

decisions, but how does it measure up the actual decisions found incases?

Of course, without facts, we will not be able to follow thedecisionveryclosely.Butconsiderareorderingof theelementsofourdecision for, sayBoeing.Wealsoomit someelements, and add a littlelinking text. Recall too that we wrote the program used thus far to“decide”thecases:inaversiontosupplyexplanationwewouldwanttocustomise the text reports to indicate theparticular clausebeingusedforanodebygivingthebaselevelfactorsused.Belowiswhatadecisionmight look like: we show the current program output in boldface,possible clause-specific customisations in italics and linking text inordinaryfont.

We find forplaintiff. The informationwas a trade secret:effortsweretakentomaintainsecrecy,sincedisclosures tooutsiderswererestrictedandthedefendantenteredintoanon-disclosureagreementandothersecuritymeasureswereapplied.The information was unique. It is accepted that theinformation was valuable and it is accepted that theinformationwasneitherknownnoravailable.A trade secret was misappropriated: there was aconfidential relationship since the defendant entered into anon-disclosure agreement and it is accepted that theinformationwasused.Moreover improper means were used since the defendantusedrestrictedmaterials.

Thisseemstohavethemakingsofareasonableexplanation.Therearetwoproblems:itdoesnotindicatewhatthedefendantcontended,sincetheclausesoftheprogramwhichwerenotreacheddonotfeatureinthereport,and,ofcourse, thefactsonwhichthefindingarebasedarenotpresent. None the less, we find the output a distinct improvement onprevious work such as Chorley and Bench-Capon (2005). We believethattheoutputfromthecurrentprogramcouldbereadilyusedtodriveaprogramof thesortenvisagedbyBranting (1993), and that thiswillbecomeevenmoreusefulwhenwehaveaddedafactlayertoallowtheexplanationoftheattributionoffactors.7.APPLICATIONTOASECONDDOMAINIntheabovewehaveconsideredtheapproachwithrespecttoasingledomain. If the approach is to be of general significance, however, itneedstobeapplicabletootherdomains.Thissectiondescribesafurtherexercise designed to show that the approach is more generally


176

applicable.Wewillapplythemethodtoadomainwhichhasoftenbeenusedasanillustrationoffactorbasedreasoning:thewildanimalscasesandPopov v Hayashi. Thewild animals caseswere introduced into AIand Law in Berman and Hafner (1993) and extended to the baseballcase of Popov in Wyner et al. (2007). We will use the factor-basedanalysisofBench-Capon(2012)asourstartingpoint.

Briefly the wild animals cases concern plaintiffs chasing wildanimalswhentheirpursuitwasinterruptedbythedefendant.Postwaschasingafoxforsport.Keeblewashuntingducks,YoungfishandGhenawhale, all in pursuit of their livelihoods. Popov v Hayashi concerneddisputedownershipof abaseball (valuablebecause it hadbeenhit byBarryBondstobreakahomerunrecord).Popovhadalmostcompletedhis catch when he was assaulted by a mob of fellow spectators andHayashi (who had not taken part in the assault) ended up with thebaseball when it came free. The wild animals cases were cited whenconsidering whether Popov’s efforts had given him possession of theball.

Thirteen, base-level, factors are identified in Bench-Capon(2012). The first task is to form them (together with appropriateabstract factors) into a factor hierarchy, to use as the node and linkstructureofourADF.This factorhierarchy isshowninFigure3:someadaptations have been made; for example, we include a factor Res(ResidenceStatus)toindicatetheattachmentoftheanimalstotheland,since it appears to make a difference whether they are therepermanently, seasonally, habitually, occasionally, or whatever. ThenodesandlinksaregiveninTable5.

Figure3-FactorHierarchy/ADFforPopov


177

Wenowsupplyacceptanceconditionsfortheninenon-leafnodes.

1. DecideforPlaintiffifNOT(NoBlame)AND(OwnershipOR(RightToPursueANDIllegalAct))

2. Ownershipif(OwnsLandANDResident)ORConventionORCapture

3. CaptureifNOT(NotCaught)OR(VerminandHotPursuit)4. RightToPursueifOwnsLandOR(HotPursuitANDPMotiveAND

(NOT(better)DMotive))5. PMotiveifPLivingOR(PSportORPGain)AND(NOTDLiving)6. DMotiveifNOTMaliceAND(DLivingORDSportORDGain)7. IllegalActifTrespassORAssault8. TrespassifLegalOwnerANDAntiSocial9. AntiSocialif(NuisanceORImpolite)AND(NOTDLiving)

The only real controversy here is with the determination of Right toPursuewhen both the plaintiff and the defendant have goodmotives.Essentiallywewanttosaythatifthelandisnotownedbyoneofthem,the right to pursue is given to the party with the better motive. Theremainder seem fairly uncontroversial. The acceptance conditions caneasilybeexpressedasPrologproceduresandthenembeddedincodeaswasdoneforCATO.Wecannowexecutetheprogram.Runningthecasefor Young v Hitchens produces the output (note that the programabbreviatesfactornames):

1?-go(young).theplaintiffhadnotcapturedthequarrytheplaintiffdidnotownthequarryplaitiffhasgoodmotivedefendanthasgoodmotive

Table5-PopovasADF


178

plainiffdidnotownthelandplainiffhadarighttopursuethequarrydefendantcommittednoantisocialactsdefendantcommittednotrespassnoillegalactwascommitteddonotfindfortheplaintifffindforthedefendantyoung[rtToPursue,dMotive,pMotive,nc,hp,imp,pliv,dliv]

WeproducecorrectresultsfromallfivecasesdiscussedinBench-Capon(2012),andonthisbasiswebelievethattheADFrepresentationcanbeusedtoencapsulatetheknowledgeofthedomain.Wecannotevaluateitas a decision making program since there are insufficient casesavailable, but this suggests that the method can be appliedstraightforwardly to a second domain to construct an executableprogram. Ingeneralwebelieve that themethodcanbeapplied toanydomainforwhichfactorbasedreasoningintheCATO(orHYPOorIBP)style is appropriate. This has encouraged us sufficiently to attempt toapply the method to a larger scale problem in the domain of the USautomobileexception to the fourthamendmentrule forwhich there isnoacceptedanalysisintofactorsavailable,sowethatneedtostartfromthe case decision texts: we will also incorporate a fact layer in thisdomain.Thisisthesubjectofthenextstageofourproject.8.CONCLUDINGREMARKSIn this paper we have evaluated an approach to reasoning with legalcases described in terms of factors using Abstract DialecticalFrameworks,asdescribedandadvocatedinAl-Abdulkarimetal.(2014).Wefindthat:- The success of the implementation depends to a large extent on thequality of the analysis. Making a direct translation of the analysis ofCATO(Aleven,1997)yieldsasuccessratealmostidenticaltothatfoundforCATOin(Brüninghaus&Ashley,2003).Thisisacreditable78.1%ofthecasesdecided“correctly”.- The ADF does, however, provide very transparent output thatidentifies precisely where the outcomes suggested by theimplementation diverge from the actual outcomes. Now reading theoriginal decision texts suggests one of four solutions. These are, inascendingorderofdivergencefromtheoriginalanalysis:

1. Removingafactorwronglyattributedtothecase2. Addingafactorwronglyomittedfromthecase3. Modifyinganacceptancecondition:e.g.changingthepriorities


179

4. Modifying the ADF: e.g. adding a supporting or attacking nodefor theproblemnode.Often several of thesemodifications canpotentiallysolvetheproblem,andthechoiceismadeaccordingto the context provided by the other divergent cases we aretryingtoaccommodate.

-TheADFapproachprovidesagoodwayofusingasetoftestcasestorefineaninitialanalysis.- The output from the program provided good diagnostics and areasonable explanation of the outcome. Our output does, however,currently lack the citations and facts which are prominent in actualdecisions.-Themethodemphasisesreasoningwithportionsofprecedents,ratherthanwholecases.Webelievethatthisdoescorrespondtolegalpracticeasmanifestinrealdecisions.-Themethodcanbeappliedtodifferentdomains.Webelievethatanydomain forwhich factorbased reasoningwouldbeappropriatewouldbeamenabletothismethod.

Wefindallofthisencouraging.Thenextimportantstepwillbetoextendthemethodtothe fact level,soas topermitargumentaboutthe ascriptionof factors, and tobe able to groundour explanations inthe particular facts of a case. Once the method has been extended toinclude the facts of particular cases at the lowest level of the ADF, aprogramtopresenttheoutput ina formresemblingthetextsofactualdecisionscanalsobeconsidered.

REFERENCESAl-Abdulkarim,L.,Atkinson,K.,&Bench-Capon,T. (2014).Abstractdialectical

frameworksforlegalreasoning.ProceedingsofJurix2014,61–70.Al-Abdulkarim,L.,Atkinson,K.,&Bench-Capon,T.(2015).EvaluatingtheUseof

AbstractDialecticalFrameworkstoRepresentCaseLaw.ProceedingsoftheFifteenthInternationalConferenceonArtificialIntelligenceandLaw,156–160.

Aleven, V. (1997). Teaching case-based argumentation through a model andexamples.PhDthesis,UniversityofPittsburgh.

Ashley, K. (1990). Modelling Legal Argument: Reasoning with Cases andHypotheticals.Cambridge,MA.:MITPress.

Atkinson,K.,Bench-Capon,T.,Prakken,H.,&Wyner,A.(2013).Argumentationschemes for reasoning about factorswith dimensions.Proceedings ofJURIX2013,39–48.

Bench-Capon, T. (2012). Representing Popov vHayashiwith dimensions andfactors.ArtificialIntelligenceandLaw,20(1),15–35.


180

Berman, D., & Hafner, C. (1993). Representing teleological structure in case-based legal reasoning: The missing link. Proceedings of the FourthInternationalConferenceonArtificialintelligenceandLaw,50–59.

Branting, L. K. (1993). An issue-oriented approach to judicial documentassembly.Proceedingsofthe4thICAIL,228–235.ACM.

Brewka,G.,Strass,H.,EllmauthalerS.,Wallner,J.,&WoltranS.(2013).Abstractdialectical frameworks revisited. Proceedings of the Twenty-ThirdinternationaljointconferenceonArtificialIntelligence.AAAIPress.

BrewkaG.,&WoltranS. (2010).Abstractdialectical frameworks. InPrinciplesofKnowledgeRepresentationandReasoning:ProceedingsoftheTwelfthInternationalConference.

Brüninghaus, S.,&Ashley,K. (2003). Predicting outcomes of case-based legalarguments.InProceedingsofthe9thICAIL,233–242.

Chorley, A., & Bench-Capon, T. (2005). Agatha: Using heuristic search toautomate the construction of case law theories.Artificial IntelligenceandLaw,13(1),9–51.

Clark,K.L.(1978).Negationasfailure.Logicanddatabases,293–322.Springer.Dung,P.M.(1995).Ontheacceptabilityofargumentsanditsfundamentalrole

innonmonotonic reasoning, logicprogramming,andn-persongames.ArtificialIntelligence,77,321–357.

Prakken,H.(2010).Anabstractframeworkforargumentationwithstructuredarguments.ArgumentandComputation,1(2),93–124.

Prakken, H., & Sartor, G. (1998). Modelling reasoning with precedents in aformaldialoguegame.ArtificialIntelligenceandLaw,6(2-4),231–287.

Wyner, A. Z., Bench-Capon, T. J.M.,&Atkinson, K. (2007). Arguments, valuesand baseballs: Representation of popov v. hayashi. JURIX2007, 151–160.

argumentation and reasoned actiontbc/publications/eca_lisbon_proceedings... · isbn...

Documents