56 From: AAAI Technical Report SS-99-01. Compilation copyright © 1999, AAAI (www.aaai.org). All rights reserved. A Comprehensive Approach to Argumentation Philip N. Judson and Jonathan Vessey LHASA Ltd School ofChemistry University ofLeeds Leeds LS29JT OK judson@dircon.co.uk jonv@chcin.lccds.ac.uk Abstract (save in the case of a conclusiveargument in oppositionto the first, whichis the specialcaseof contradiction). A modelhasbeendeveloped whichimplements thekey principles of theLogic ofArgumentation - a method of reasoning based onthehuman process ofreaching conclusions Reasoning about Toxicity bycomparing arguments forandagainst propositions. The modelis beingusedto improve reasoningaboutthe potential Three examplesillustrate inter-relationships between toxicityof novelcompounds. argumentswhicha modelfor reasoning about poltntial toxicity needsto incorporate. Introduction It is generally thought that chemicals showing carcinogenicity in rats should not automatically be There is a needto makepredictions about toxicities of expected to be similarly toxic to other species (including novel substances, perhaps even before they have been man)if they promoteperoxisomeproliferation, because synthesised. Experienced toxicologists aremoderately this mechanismis peculiar to the rat. Of course, a good atdoing this, even when theavailable information is chemicalmaybe carcinogenicfor different reasons, andso veryuncertain. Howcancompultrs assist them? it is not valid to state, or to imply,that a compound that is carcinogenicto rats but causesperoxisome proliferation is Oneapproach seeksto imitalt humanexperts in safe to man- or eventhat the evidencearguesin favourof knowledge-based expert systems (KBS). There havebeen safety. A reasoning model needs to be able to deal somesuccesses in theuseof KBSin toxicology. For correctly with evidencethat undercutsan argument,rather example, theDEREK system isreported tobequilt good thannegatesit, like this. atpredicting thepotential ofchemicals asskin scnsitizers (Deardenet al. 1997). Current KBSnevertheless fall Certain substructural features in chemicals are long wayshort of the reasoningcapabilities of a human commonly associated with carcinogenicity. In some cases, being. this connection isnotstrictly corrcc% thereality being that positive Amestest results havebeenlinked to the presence Thelogic of argumentation(LA)(Krauseet al. 1995) of the sub-structuralfeatures. TheAmes test is an indirect wasused as the basis of a prototypeimprovedsystemfor indicator of potential carcinogenicitybut this distinction predictingtoxicity (Foxet al. 1996),but with a restricted has been lost somewherealong the line between the model.This paper outlines a morecompletemodel. original researchersand, ultimately, the popularpress. A computerreasoning system should construct the overall LAformalists the waythat people reach decisions by prediction of carcinogenicityfromthe basic steps, but it identifying the argumentsfor and against a proposition, shouldbe able to report its reasoningin full, so that a user and then weighing them against each other. Some is properlyinformed. arguments relate directly to the decisionor propositionof interest. Other argumentsmaybe about the grounds of Electrophiles with reactivitics within a particular the direct ones, or about the reliability of the arguments range cause skin sensitization. Theymustbe sufficiently themselves.So the reasoningprocess can be represented reactive to combine with proteins, leadingto the raising of by a tree graph. antibodies, but not so reactive that, becomingcompletely Arguments in a chain ofreasoning in a complicated boundto the skin surface, they never reach the site of action. Suchreactivity can be assumedwith reasonable domain arcrarely conclusive. Indeed, if thereis a confidence for a substancethat has certain substructural conclusive argument fororagainst a proposition, then the features (for example,an alpha-betaunsaturatedaldehyde, matter is decidedand all other argumentsare irrelevant 57 whichacts as a Michaelacceptor). Substancescontaining the right features do not all provokesensitization, for a variety of reasons. Oneis thought to be the failure of a substanceto penetrate the skin becauseof its physicochemicalproperties. Acomputersystem needsto be able to modifyits predictions about a chemicalthat meetsthe criteria for electophilicity, after taking into account physicalpropertiesthoughtto affect skin penetration. An Outline of the Model The modelwehave develol~dwill be described in detail elsewhere (Judson and Vessey, in press), and only outline is given here. Figure 1 illustrates the main elementsof a tree for reasoningin LA. "’:y B" /f’: Figure l" o,:o f 1 - A simpleI~re~~ Thecapital letters at the nodesof the graph represent assertions and the arcs represent argumentsthat connect them.Thegraphis a directed one, fromleR to right. So, for example,the tree indicates that the likelihood of C maybe deducedfromthe likelihood of A. Thelower case letters represent degrees of persuasiveness of the argumentsthemselves.So, for example,"a" is howlikely C wouldbe if Awereproventrue (or howunlikely, if the truth of Aarguesfor the falsenessof C). Reasoning along the chain from A to E is like the prediction of carcinogenityfroma substructural feature associated with positive Amestest results. Step A to C might be "if such-and-suchsub-structural feature is present in the molecule,a positive Amestest result is probable".Step C to E mightbe "if an Ames test result is positive, then carcinogenicity is plausible". A more complete implementationwouldfurther sub-divide the chain to include the implicit steps from "Amestest positive’ to "genotoxic" and from "genotoxic" to "carcinogenic". Thepersuasivenessof an argumentcan be subject to other influences, as in the case of predicting toxicity in man fromobservedtoxicity in the rat, wherethe argumentis weakened if peroxisome proliferation is also seen. This is the purposeof arcs leading towardsthe values of other arcs, suchas the oneleadingfrom"F" to "a". In some domains, there may be a case for allowing argumentsof the kind "If Ais plausible then C is true", whichwouldrequire the notion of a threshold value for the argument linking A and C. "a’" represents this threshold value. Theconceptis includedfor completeness but it will not be discussedfurther in this paper. In LAan argumentfor a proposition is consideredto haveno direct effect onbelief that the propositionis false, and vice versa. For example, observed toxicity to an animal maysupport the proposition that a chemicalmay be toxic in man:absenceof observedtoxicity in the animal does not supportthe propositionthat the chemicalis nontoxic in man;there is simplya lack of evidenceto support the proposition that it is. This does not mean that negative reasoning is excluded,only that it mustbe explicit. Toassert that the absenceof toxicity in laboratoryanimalsmakestoxicity in manunlikely,it is necessaryto write a rule suchas, "If no toxic effects are observedin laboratoryanimalsthen it is doubtedthat the chemicalwill be toxic in man". In the implementationof the modelwhichwefavour, if there are several argumentsin supportof a conclusion, the level of belief in the conclusionis equal to the level conferredby the strongest argument.For example,it does not matter howmanyindependentargumentssupport the viewthat a conclusionis plausible, it remainsplausible. Nonumberof reasons for thinking somethingis plausible shouldmakeit a certainty, to use an extremeillustration. If the weightsof argumentfor and against a conclusion are equal, belief in the conclusionis classedas equivocal. This is distinguished from the case wherethere is no evidence either way, which is classed as open. For example, given the information that a chemical shows carcinogenicity in the rat but also causes peroxisome proliferation, it maybe appropriateto class the likelihood that the chemicalwill be carcinogenic in manas open, whereastoxicological concernwouldbe equivocal for a chemicalwith the right structural features for a skin sensitizer but with physical properties on the borderline for skin penetration. Giventhe statement "if A is true then the level of belief in C is a", what should happenif A is less than true? In the model,the level of C becomeswhicheveris the weaker- the level of belief in A or the value of a. Supposefor illustration, that ’probable’ meansmore 58 certain than ’plausible’ and ’certain’ meanscompletely certain. Giventhe statement "If A is certain then C is probable": C can never be predicted to be more than probable on the basis of evidence about A; if A is probable,thenso is C; ifA is plausible,then C is also only plausible. As commented above,if Ais provenfalse, nothing can be said about C on the basis of this evidence.It is not valid to concludethat C is also false, becauseit maybe true for someother reason. Moregenerally, if evidence about the truth of Afalls anywhere in the realmof doubt rather than belief, nothingcan be concludedaboutC. Theseparationof argumentsfor andagainst raises the possibility of contradiction.Theoretically,there mightbe an argumentthat proves a conclusion to be true and an independentargumentthat proves it to be false. Sucha situation can arise in practice, too - as the result of erroneous observations, for example, or because of genuinescientific contradiction that has not yet been resolved. In the implementation of the model which we are using, contradiction dominatesamongargumentsrelating directly to the sameproposition,unless it is over-ruledby an independentargumentthat the proposition is true or false: it is usual scientific practice whenconfrontedwith apparentcontradictionto seek out alternative evidenceto resolvethe difficulty. Contradictionis not transmittedas contradiction fromone level to the next in the reasoning process,but is interpretedas the openstate. Giventhat "if A is true then C is plausible", for example,contradiction in Ameansthat Ais either true or false. If it is true then C is plausible. Asexplainedabove,if Ais false then C is open,whichincludesthe possibility of its beingplausible. So, overall, Cis open. The Basic Model The central componentsof the modelare the methodsfor aggregating argumentsrelating to the sameproposition (e.g. the combinationof degrees of belief in E conferred via the arcs fromC andDin figure I) and for determining the degreeof belief conferredby each argument(e.g. the belief transmittedto C fromA). Theyare as follows. Combining Arguments Theaggregatevaluefor a node,T, is determinedby: T = Resolve[Max{For(C~x, Cb,y.... }, Max{ Against(C~.x, Cb.y.... }1 where: Resolve[] returns the single value whichis the resolution of any pair of opposingforces in a resolution matrix(see below);the sets For andAgainstare the sets argumentssupportingand opposingT, respectively; C~,x, Cb,y.... are the forcesof thosearguments; Max{...} returns the memberof the set upon whichit operates whichis highest in an ordered list For or Against, as appropriate (see below). Reasoning along a Chain Thevaluetransmittedalongan arc in the decisiontree for the statement, "if a is Xothen C is x," is determinedas follows: if Xois exclusive then if a = Xothen C=x else C = open endif else if x ~ Subset(-,a)then C = open else ifa = Min{a,Xo}then C -- Min{a,x} else C=x end where:Subset(~a)is the set of terms For or Againstwhich does not contain value ’a’; Min{a,x} is the term ranking lowerin the orderedlist, For or Against,to whicha and x belong.Theterm ’exclusive’ is explainedbelow. The Resolution Matrix For the equationaboveto operateit is necessaryto define a matrixfor the resolutionof pairs of levels of belief. The matrixwhichweare using is shownin Table1. OrderedSets For and Against Weare usingthe followingsets: For Against certain impossible contradicted contradicted probable improbable plausible doubted equivocal equivocal open open The operations described aboveare only valid for terms that can be ordered and resolved unambiguously. For the 59 purposes of this implementationof the modelthe terms are defined as follows (note that in somecases the definitions are morerestrictive than the meaningsof the wordsin common usage)" certain there is proofthat the propositionis true probable there is at least one strong argumentthat the proposition is true and there are no argumentsagainst it plausible the weight of evidence supports the proposition equivocal there is an equalweightof evidencefor and against the proposition improbable there is at least one strong argumentthat the proposition is false and there are no arguments that it is true impossiblethere is proof that the propositionis false Themembers of the sets aboveare categorisedas follows: Exclusiveterms:contradicted, equivocal,open. Non-exclusiveterms: impossible, improbable, doubted, plausible, probable,certain. Conclusion This paper outlines a modelwhichwebelieve can be used for reasoning under uncertainty in any domain. Our current interest is in using it to improvethe performance of knowledge basedsystemsfor toxicology. Acknowledgements Wewishparticularly to thank JohnFoxfor his vision in promotingthe basic concepts of LAand Paul Krausefor his advice and commentsduring the developmentof the modeldescribedhere. contradicted there is apparently proof that the propositionis both true andfalse open there is no evidence that supports or opposesthe proposition Exclusivity Whensometerms from the abovesets are attached to a given argumentor proposition, they implythe terms below themin the list. For example,if somethingis known to be certain, then the answerto the question"is it plausible?" is "yes". Someterms excludeall others. For example,the state of knowledgeabout somethingcannot be open and simultaneouslyhave someother value. The latter terms are describedas’exclusive’. References Dearden,J. C.; Barratt, M. D.; Benigni, R.; Bristol, D. W.; Combos,R. D.; Cronin, M. T. D.; Judson, P. N.; Payne, M. P.; Richard, A. M.; Tichy, M.; Worth,A. P.; and Youriek,J. J. 1997. The Development and Validation of Expert Systemsfor Predicting Toxicity. ATLA25:223252. P. N. and Vessey, J. A Comprehensive Approach to Argumentation. Submitted to the IMAJournal of Mathematics Appliedin Businessand Industry, 1999. Judson, P. J.; Ambler,S. J.; Elvang-Geransson,M.; and Fox, J. 1995. A Logic of Argumentationfor Reasoning UnderUncertainty. ComputationalIntelligence 11(1): 113-121. Krause, Table 1 The resolution matrix impossible contradicted improbable doubted equivocal open certain contradicted probable contradictedimpossible impossible certain contradicted contradicted certain contradicted equivocal certain contradicted plausible certain contradicted probable certain contradicted probable plausible impossible contradicted doubted equivocal plausible plausible equivocal impossible contradicteA improbable doubted equivocal equivocal open impossible contradicted improbable doubted equivocal open