Towards More Flexible and Common-Sense Reasoning About

TowardsMore Flexible and Common-SensicalReasoning about Beliefs

Gees C. Stein John A. Barnden Computing Research Lab Computing Research Lab & Computer Science Dept. & Computer Science Dept. NewMexico State University NewMexico State University Las Cruces, NM88003-8001 Las Cruces, NM88003-8001 gstein@crl, nmsu. edu j barnden@cri. nmsu. edu

Abstract in the future. Someimportant research problemswithin reasoning about beliefs are howto deal withincomplete, inaccurate or uncertain Belief-Reasoning Issues Addressed beliefs, whento ascribe beliefs and howto makereasoning In this section we discuss aspects of several of the crucial about beliefs morecommon-sensical. We present two systems issues mentioned in the call for papers for the workshop. that attack such problems.The first, called CaseMent,uses These aspects are addressed by one or both of the systems we case-basedreasoning to reason about beliefs. It appears will be describing. that using cases as opposedto roles has advantageswith respectto reusability, context-sensitivityof reasoning,belief Kinds of Reasoning Done by People ascriptionand introspection. Although the programis still in an initial stage, the first results fromthe existingimplementation Classical logics rely on deductive inferences and most systems look promising. The second system, ATI’-Meta,addresses that reason about beliefs do the same (Creary, 1979; Haas, the fact that metaphoricaldescriptions of mentalstates play 1986; Hadley, 1988; Lakemeyer, 1991; Levesque, 1984; an important role in the coherence of mundanediscourse Maidaet al., 1991). In particular, believers are implicitly or (including stories). ATI’-Meta’sreasoning modulehas explicitly cast as performing deduction. However,people do advancedprototype implementation. It incorporates powerful not only use deductionbut also induction, abduction, analogy- facilities for defeasibleand uncertain reasoning. Both systems based reasoning and other forms of plausible reasoning. They use simulative reasoning and thereby reduce the amountof havethe ability to infer general rules to explain someparticular explicit self-knowledgethey needin ascribing,to other agents, beliefs and reasoningssimilar to their own. examples(induction) or generate plausible explanations for observedfacts (abduction --- see, for example,Hobbs et al., 1993). While deduction produces logically sound results, Introduction induction and abduction produce plausible results, which do The call for papers addressed the issues of what is needed not necessarily have to be true. Ideally, systems that reason for story understanding, the way mental states should be about beliefs should be able to ascribe abduction, deduction, represented, the reusability of representations, introspection, induction, analogy-basedreasoning etc. to people. and the type/amountof explicit self-knowledge.In this paper two systems will be discussed that reason about beliefs and Incomplete, Incorrect or Uncertain Information address certain aspects of these issues. The two systems are Ideally, reasoning happenswith all the relevant information based on different reasoning techniques and focus on different available, and with all the information being correct and problems of reasoning about beliefs. The first system uses free of uncertainties. In reality people often have to reason case-based reasoning in order to grapple with somecommon- with incomplete, incorrect and uncertain information. They sensical aspects of reasoning about beliefs. The secondsystem generate defeasible conclusions, conclusions that seemlikely focuses on other common-sensicalaspects in concentrating to be true but later mightturn out to be false. A systemthat on discourse contexts that contain metaphorical descriptions reasons about beliefs has to deal with these issues at both the of mentalstates. system level and within a belief space. Becauseagents might Althoughboth systems are currently confined to reasoning have incompleteor uncertain beliefs it is not enoughfor the about beliefs as opposedto other mental states, we plan to system itself to knowhow to reason with its ownincomplete, expandto other mental states in the future. uncertain beliefs; it has to ascribe such reasoning to other Section 2 discusses those issues from the workshop’s agents as well. Only a minority of research on belief has call for papers that our systems address to a significant endowedsystems with this sort of capability. Oneexample is extent. Section 3 discusses the general technique of case- the work of Cravo & Martins (1993), on a version of SNePS based reasoning, which is the basis of our first system, that uses default reasoning. CaseMent.Section 4 sketches CaseMentitself, and section 5 shows why case-based reasoning helps with some of the Shades of Belief issues of section 2. Section 6 sketches our ATT-Metasystem, Most systems do not deal with different kinds of beliefs. and then section 7 discusses its benefits with regard to the However,beliefs can be of different varieties or "shades". issues of section 2. Section 8 concludes, and briefly points For example, they can be held with different degrees of out how CaseMentand ATT-Metacould be brought together consciousness, they can be more or less active (i.e., more

127 or less capable of affecting reasoning and action), and they (ii) if other people similar to P believe B, then ascribing B can be momentaryor long-lasting. The particular shade of a P is reasonable. belief can be implied by its context. Take for instance "Helga (iii) if other people similar to P believe somethingsimilar believes that Werner is intelligent," in a context where subject matter to B, then ascribing B to P is reasonable. Helga is not just in love with Werner, but really worships Thus, analogy might be a good way for a system to improve him, and where Werner is not in fact intelligent. Then, its belief ascription. Of course, a suitable notion of similarity plausibly, Helga’s belief is a kind of "wishful thinking" would need to be developed. belief: something halfway between wishing and believing. Contrast this with "Marybelieves olive oil is very healthy," Introspection: Knowledge of Own Mental given that Mary has done a lot of research on howolive Processes oil affects health and that Maryis a well-respectedscientist. It is useful for an intelligent systemto be able to view other Mary’s belief is presumably not wishful thinking. Other agents as reasoning in muchthe sameway as it itself reasons. mental states also come in various shades. For instance, This view is a useful default, in the absence of specific hoping involves a context-sensitive mixture of expectation evidence about how another agent is reasoning. Now, a and desire. particularly effective way of embodyingthis strategy in a Levesque(1984) and Lakemeyer(1991) makea distinction system is simulative reasoning. This technique has often been between "explicit" and "implicit" beliefs, mainly to solve suggested in AI, under various names(see, e.g., Attardi the logical omniscienceproblem. It is commonin philosophy Simi, 1994; Ballim & Wilks, 1991; Chalupsky, 1993; Creary, to distinguish "tacit" beliefs as a special case (see Richard 1979; Dinsmore, 1991; Haas, 1986; Konolige, 1986; Moore, (1990) for discussion). However,these distinctions only 1973); it is also closely related to the Simulation Theory scratch the surface of the issue of shadesof belief (or of other in philosophy and psychology (Davies & Stone, in press). mentalstates). Simulative reasoning works as follows. Whenreasoning about another agent X, the system in effect pretends it is Ascribing Beliefs agent X by using X’s beliefs as if they were its own, and Tobe able to reason about other agents’ beliefs a systemhas to applying its own reasoning mechanism to them. Results makeassumptions about what the agents believe. Normally, of such pretences are then wrappedinside an "X believes relatively few if any of the beliefs of an agent are knownwith that" layer. Simulative reasoning has various advantages over the alternative methodsthat would otherwise be needed certainty, and the system has to "ascribe" beliefs: assume (such as reasoning explicitly about inference steps taken that the agent has particular beliefs. by X). Onesalient benefit is that it does not require the Peopledo a similar thing: they can interact even with some- system to have explicit information about its own reasoning one totally unknownby assumingthat this person has certain abilities in order to to be able to ascribe themto other agents. beliefs. People appear often to make a default assumption This advantage is especially strong in a system that has a that other people have beliefs similar to their own.By default variety of reasoning methods--- e.g., induction, abduction, everybodyis expectedto believe that the earth is round. At the analogy-based reasoning, as well as deduction --- or where sametime a person might have certain beliefs that he or she an individual methodis procedurally very complex--- e.g., does not ascribe to others: an astronomer will not generally analogy-based reasoning. (See Haas, 1986, and Barnden, in assumethat non-astronomershave the same beliefs as (s)he press, for further discussion and other advantages.) Under aboutstars. simulative reasoning, the system’s reasoning methods are A system has to knowwhich of its beliefs it can ascribe to automatically and implicitly ascribed to other agents. other agents and which not. The same is relevant for nested Nevertheless, there are complications. Simulative reason- beliefs: the system has to reason about agent A deciding ing should be defeasible. That is, the conclusionsit provides whether agent B can be expected to have one of A’s beliefs about another agent’s beliefs should be capable of defeat by or not. Whether to ascribe a belief or not should depend other evidence. Also, it should be gracefully integrated with at least on the contents of the belief and the nature of the non-simulative reasoning about other agents’ beliefs. Non- agent involved, as we see in the astronomer example. Now, simulative reasoning is still needed, because someinferences someexisting systems go a useful distance towards handling about an agent’s beliefs are not dependent on the agent’s this issue. For example, BaUim&Wilks (1991)’s ViewGen reasoning--- e.g., there couldbe a rule that says that an agent system has a liberal, default method for belief ascription whobelieves the earth is flat probably has manyother incor- between agents. The default is that everything the system rect beliefs --- and becausesimulative reasoning is relatively believes can be ascribed to another agent, unless contradicted weakat producing negative belief conclusions (conclusions by knownbeliefs of the agent. (Similarly in nested cases.) of the form: X fails to believe Q). However,the default can be bypassed for a particular belief by specifying that the belief is only possessed by particular Context-Sensitivity agents or types of agents. Varieties of context-sensitivity were touched upon in section However,existing systems do not take account of context 2.3 and 2.4. In addition, the context of a reasoning problem in moresophisticated but clearly useful ways.For instance, it can have a important influence on the outcomeof the problem. is plausible that, whentrying to ascribe belief B to person P, People under stress might cometo one conclusion about a a person can use the following heuristics: certain problem, while those people in relaxed circumstances (i) if B is similar in subject matterto an existing belief of P’s, might come to a very different conclusion. Or, a doctor then ascribing B to P is reasonable. reasoning about a patient’s problemwill systematically try

128 to avoid mistakes. Whenthe same doctor is reasoning about and places her believing in the recesses of that space; that whatto eat (s)he can afford to be muchless ’scientific’ in the is, roughly speaking, the representation "pretends" that the reasoning. Thus, context plays an important rule in reasoning sentence is LITERALLYtrue. Although our system can in about beliefs, in several ways. Currently, systems do not somecases partially translate such metaphor-imbuedinternal makemuch use of context whenreasoning about beliefs. representations into metaphor-freeones, there are manycases whenthis is not possible and yet the metaphor-imbuedrepre- Story Understanding, Mental States and Metaphor sentations can be used in inference processes that yield useful In real speech and text, including stories, mental states are results. This point is clarified below. often described metaphorically. This is true even whenthe discourse is not particularly literary. The following are some Reusability of Knowledge examples of common, mundane forms of common-sense, Finally there is the issue howknowledge about beliefs could metaphoricaldescription of mental states/processes: be reused. Often systems try to focus on a specific domain John was leaping from idea to idea. within which reasoning about beliefs occurs. ’Belief’ bases Georgeput the idea that he was a racing driver have to be developedtogether with reasoning rules or other into Sally’ s mind. reasoning mechanisms.It would be helpful if this knowledge His desire to go to the party was battling with the could be easily reused by other systems. Not muchattention knowledgethat he ought to work. has been paid to this issue though. Peter hadn’t brought the two ideas together in his mind. Martin had a blurred view of the problem. Case-Based Reasoning Veronicacaught hold of the idea and ran with it. The first system that will discussed, CaseMent,has as its Onepart of Sally was convinced that Mike was cheating main motivation that it tries to be a more common-sensical on her. and realistic system with respect to the way people reason For further examples and discussion of metaphors of mind about beliefs. This system uses case-based reasoning (CBR) see, for example, Belleza (1992), Bruner & Feldman(1990), as opposedto rule-based reasoning to reason about beliefs. Cooke and Bartha (1992), Fesmire (1994), Gallup In this section wegive a short description of CBRin general; Cameron (1992), Gentner and Grudin (1985), Gibbs we present the CaseMentsystem in the next section. O’Brien (1990), Jakel (1993), Johnson (1987), Lakoff, CBRis an approach that tries to solve a problem by penson and Schwartz (1991), Lakoff & Johnson (1980), using knowledge/experience with previous cases that have Larsen (1987), Richards (1989), Sweetser (1987, 1990), similarities with the problemat hand (Riesbeck and Schank, Tomlinson (1986). 1989; Kolodner, 1993). A case can be seen as a previous Metaphors of mind can make a crucial difference to the experience. Often people like judges, doctors, use their understandingof discourse, including stories. In particular, previous experience with similar cases whentrying to solve metaphorical descriptions often provide important informa- a problem. If a certain approach in a previous similar case tion about what inferences the characters in a story have was successful, it might work again. If no rules are known made. For instance, if someonehas not "brought" two ideas to solve a problem,the only wayleft to deal with it mightbe "together" in his or her mind, then it can be assumedthat looking at similar knownand tried cases. (s)he has not drawneven the obvious conclusions from them. The main steps involved in CBRare finding relevant cases, Notice also that the last examplesentence aboveimplies that, analyzing howthey differ from the problemat hand, choosing probably, someother part of Sally did not have the mentioned one or more best cases, and adapting solutions to fit the conviction. It is our contention that description of complexor current problem. Becausecases that are slightly different subtle mental states often requires metaphor, so that under- from the problemstill can be used by adapting solutions, a standing of metaphoricaldescriptions of mindis an important case can be used for more than one problem. Dependingon consideration for the workshop.However, the matter has not how’smart’ the system is, what kind of analogies it can find been intensively addressed in AI other than in our ownwork betweena problemand a different case, and howgood it is at on the AT’l’-Meta system (Barnden, 1988, 1989a,b, 1992; adapting solutions, a case can be used in manyways. Because Barndenet al. 1994a,b, and forthcoming). a case does not have to matcha problemexactly it is relatively easy to deal with incomplete and incorrect information. Logical Form of Mental-State Representations A case-based reasoning system’s performance is improved Ourattention to metaphoricaldescription of mentalstates, in in several ways. Cases are added which gives the system the case of our ATr-Metasystem, has had profound conse- more experience to build on. The system gets feedback from quencesfor the style of internal logical representation that its proposedanswers. Not only correct answersare stored, but we use in that system. Usually, metaphorresearchers tacitly the system also keeps track of its failures, to prevent similar assumethat a system that interprets metaphorical sentences mistakes in the future. Cases inherently provide morecontext- must construct internal meaningrepresentations that cash out sensitivity than rules through storing moreinformation about the metaphors in objective, non-metaphorical terms. (There individual situations. are exceptions, notably Hobbs, 1990.) However,we take an CBRseems to be especially suited for those ’non-exact’ opposite tack, for reasons explained in Barndenet al. (in domainsthat do not have clear rules that tell what to do in press). That is, the meaningrepresentation for the sentence a certain situation. If the knowledgeabout certain problems "In the recesses of her mind Veronica believed that Sally is known but very complex and time consuming to use, was wrong" takes Veronica’s mind to be a physical space case-based reasoning can improve a system’s performance.

:t29 Case-based reasoning systems have been developed for such First the system can consider its owncases, that might be different tasks as creating recipes, diagnosingheart failures, about Annette’s beliefs, or about someoneelse’s beliefs that classifying hearing disorders, settling resource disputes and seemto apply to the question and its context. This can be seen designing physical devices (Kolodner, 1993). Case-based as metareasoningwith cases. Next it will look at Annette’s reasoning has been used to metareason about a system’s cases and see if any are relevant. This is where simulative reasoning process. Examples include improving reasoning reasoning comes in. What happens is that the system will by detecting and repairing reasoning failures (Cox, 1995) consider Annette’s cases as they were its own. It will apply and supporting the reasoning process using introspection its own case based reasoning to these cases. In other words, (Oehlmann& Edwards, 1995). To our knowledge,it has never simulative reasoning in CaseMentamounts to ascribing its been applied to common-sensicalreasoning about beliefs. CBRto other agents. Anexample of a nested case is: The CaseMentSystem: Using CBRfor Belief case_part (c3, Reasoning believes_case(annette, al) , i) The current CaseMentsystem is implemented in Prolog and still in its initial stage. First a short description of the basic case_part(al,not-gossips (wife) , i) system will be given to give a feeling howthe system works. case_part(al,married_to (jones ,wife) , 2) The next section will discuss the expectedadvantages of using If the query is whether "believes(Annette,gossips(wife))" case-based reasoning. is true or not, and the context contains "married_to(jones, wife)", both Annette’s case (al) and the system’s case (c2) A case in CaseMentis an experience that might or might will be considered. Both cases mentiona belief about Jones’ not involve beliefs. Whenbeliefs are involved, they can be wife gossiping. (c2) mentionsJones’ belief and (al) contains nested. Anagent might have a set of cases about a variety of Annette’s belief. As the question is about what Annette things. He can have an experience about cooking a chicken of believes, the system will consider (al) a morerelevant case 5 pounds for 2 hours which cameout very good. This would and use it to say that Annettebelieves that Jones’s wife does be one case. Another case might be the experience that his not gossip. neighbor Mr. Jones believes that womenalways gossip. The In general more than one case might seem relevant in an fact that Mr. Jones has a wife that gossips a lot could also be episode of CBR.The system has to be able to comparecases part of that case. In the current system, these two cases, (cl) to each other with respect to their relevance to the query. and (c2), wouldbe represented as follows: This will happenwithin simulative reasoning, and also at the case_part(cl, chicken(food) , i) metareasoninglevel. If the question is about (nested) beliefs, case_part(cl,weight (food, 5_ibs) , 2) both simulative reasoning and metareasoning might give an case_part(cl, cooked(food,2_hrs) , 3) answer. The system then has to comparethe cases used for case_part(cl, test (food,good) , 4) both reasoning methods and decide which one seems to fit the query best. case_part( c2, believes( j ones, woman(X) -->gossips(X)) , Advantages of CBRfor Belief Reasoning case_part(c2,married_to (jones,wife) , 2) Giventhe basic idea of the system, the expected advantages case_part(c2, gossips(wife) , 3) of using case-based reasoning to reason about beliefs will be case_part(c2, because (i, [2,3] ) , 4) discussed, with special attention to the issues listed in section The third argumentof a case part is a unique numberthat can 2. be used to relate to case parts to each other. In case (c2) the As was said earlier, case-based reasoning is especially 4th case part says that Jones believes that all womengossip suited for those domainsthat are not very well understoodand because his wife gossips. Right nowthe only way to relate for which it is very hard to define precise rules. Reasoning parts of cases to each other is the ’because’predicate. about beliefs seems to fit that description. People might Usingthe first case the systemmight decide that a turkey of try to be reasonable but sometimes they are not. If not similar weightprobably needs 2 hours of cookingif it realizes impossibleit is at least very hard to define rules that tell us that turkey and chicken are similar. Using the second case howdifferent people having different sets of beliefs reason and reasoning about another manwho believes that women about beliefs in different circumstances. It is easier to gossip, it might assumethat perhaps this manalso has a wife describe examplesof people having certain beliefs and making that gossips a lot. In general a query consists of the question particular inferences: to create cases. itself and contextual information. The answer right nowis Cases do not only have to describe people’s logically either ’yes’, ’no’, ’have insufficient information’ or ’have correct inferencing, but can also store mistakes people make. unresolvably contradicting information’, together with one or Cases can capture different kinds of reasoning, mistakes in morecase(s) on which the answer is based. reasoning or inferences not made. An example might be a The system does not only have its owncases but can also case where someonebelieves that if the car’s gas tank was contain other agent’s cases, in other wordscase bases can be just filled with gas it has to be full. But becausehe finds it nested. Supposeagent Annette holds her ownset of cases and emptyhe reasons that the tank was not filled. Of course this the system has a question about Annette’s beliefs, together is not necessarily true, there mightbe a leak in the tank. with relevant information. case_part( c4, believes( joe, filled(tank)-->full (tank)) , is done to. A general case about eating pizza in a restaurant case_part(c4, believes( j oe, can probably be ascribed to most agents. A case that is about not-full(tank)) , specialist medical informationshould not be ascribed to other case_part( c4, believes( joe, agents unless that agent is familiar with the medical field. not-filled(tank)) , Ascription of a case can also happen within belief spaces, case_part(c4, filled(tank) , 4) for instance ’does agent X think that his case C is common case_part(c4, leaks(tank) , 5) enoughto be ascribed to agent Y? case_part(c4, sleepy(joe) , 6) A similar problem is the reverse: suppose an agent has a case_Ioart(c4, in_hurry(joe), 7) certain case. Thesystem does knowabout this, but will it use case_part(c4, because (3, [2, i] ) , 8) this case to solve a problem?It mightif the systemtrusts this The example above is an example of a person making a particular agent with respect to this particular case. A similar reasoning mistake. One example does not mean that this question can be asked within several belief layers: ’does agent reasoning mistake is a general one. Context is very important. X trusts agent Y enough with respect to agent Y’s case C Within the context of being under a lot of pressure people that it will use case C?’. If the case C is a medicalcase and might act or reason differently than under very relaxed cir- the agent Y is a well respected doctor, it seemsreasonable to cumstances.While rules try to be as general as possible, cases ’trust’ Y and X can use case C as it were its own. are meant to store relevant context. Because cases can be Thethird related problemoccurs in the following situation: used not only for exactly matchingqueries but also for ’close’ suppose the system tries to reason about agent X’s belief, ones, the context does not restrict their usage but makesit given a certain context. If finds a case C similar to the possible to use context whenavailable. problem,but it is a case within agent Y’s case base. This case Becausea case does not have to match exactly to be used, can only be used if it is likely that X mighthave had a similar case-based reasoning can deal with inaccurate information or experience. So the question is whetheragent X is similar to even missing information. The case (cl) about chicken might agent Ywith respect to the case C. be used for turkey too if the system knowsthat chicken and Although this feature is not implementedyet, the system turkey are similar. Case (c4) might still be used to reason will use case-based reasoning to decide what it should do about another person, Sally, even if it is not knownwhether in any of the three situations described above. Cases might Sally is sleepy or not. The system can assumethose missing mention situations of certain agents (not) believing other parts of a case if otherwise it matches the input. Together agents. Or they might mention certain facts that are not with an answer the system also tells on which case(s) it believed by certain people. Whentrying to ascribe a case C based. These cases implicitly showhow certain/uncertain the to person P, cases that P already has can be checked. If P answeris by their degree of similarity to the input problem. has a very similar case C’ it is probablyacceptable to ascribe Beliefs in different kind of circumstances might have case C to P. If P has a contradicting case it does not make different intensity, or level of consciousness, or they might sense to ascribe case C to P. Another possibility is to look influence behaviorin different ways.Our feeling is that there at other agents whosecase bases contain the case C. If these are different shades of belief (section 2.3). However,it is hard agents are very similar to agent P the system might ascribe to makethe differences explicit and to define whenthey are case C to P. The combination of both previous examples is present. Becausecases keep the context of a situation they also possible: another agent has a case C’ similar to case C. If implicitly can deal with different kinds of beliefs. Consider the cases are similar enoughand the agents seemsimilar with cases partially described as follows: respect to these cases, case C can be ascribed to P. Not only agent P’s cases and other agent’s cases can be used: meta case_part(c5,worships(helga, werner),l) cases about agent P and other agents can also contain useful case_part( c5, in_love_with(helga, information, for instance about the beliefs someonedoes not werner),2) have. case_part(c5, believes (helga, Concerning the issue of introspection, most belief- intelligent(werner)) , 3) reasoning systems use explicit rules to define positive and case_part( c5, not- intelligent(werner) , 4 negative introspection. A case-based reasoning systems can base introspection on cases it has about agents’ introspection, case_part (c6, done_research(marian, and hereby makeintrospection more context sensitive. swimming), 1 ) case_part(c6, does (marian,swimming) , 2 Because CBRis combined with simulative reasoning in CaseMent,the characteristics described aboveare true not just case_part(c6, believes (marian, healthy(swimming) ) , 3 for the system’sreasoning but also for other agents’ reasoning. Not only the system can handle incomplete information, Bothof these cases mentionbeliefs, but it is plausible that agents will handle incomplete information the same way. By they involve different kinds of beliefs. The belief in (c5) default, an agent will be assumedto ascribe cases to other could well be "wishful thinking." Accordingly, other parts agents similar to the way the system does. The system does of (c5) and (c6) mightinvolve distinctively different types not need explicit information about howagents (including facts about Helga and Marian. itself) can reason, which makes the system more efficient, If the system has a certain case/experience, whencan it flexible and modifiable. ascribe this to another agent? As pointed out in section 2.4, It is not only with respect to the above issues that CBR this should dependon the case and on the agent the ascription seems promising. It also has somecharacteristics that make

131 it suitable for reuse. First there is the point that a case can brother’s recipe had said to fry the mushroomsfor one hour be used in manyways. If the CBRsystem is ’smart’ enough that it can not just detect surface similarities but also more (la) She did this even though she believed the recipe to abstract similarities, cases can be used in creative ways. A be wrong. case can present an experience from several points of view (lb) She did this even though in the recesses of her mind she and can therefore be used in different ways. The creative use believed the recipe to be wrong. of cases makesit easier to reuse cases in different domains. (le) She did this even though she thought, "The recipe’s After all, people often use their experiencesin creative ways wrong’ ’. across domains. Another important characteristic that makesit easier to (la) is a non-metaphorical continuation, (lb) contains reuse cases is the fact that they store the context of their MINDAS PHYSICALSPACE metaphor and (lc) uses the main contents. Rules are often tailored for a specific domain, IDEAS AS INTERNALUTIERANCES metaphor. In all handcrafted in such a waythat they interact in the right way examplesthere is the disparity betweenVeronica’s following with each other. Using rules in different domainsand having the recipe and her believing it to be incorrect. (lc) gives to interact with newrules might cause problems. The context the impression that Veronica consciously believed that the a rule should be used in is not knownanymore, and the system recipe was wrong, so a reasonable assumption would be that might use it whenit does not apply. Becausea case contains she had a special reason to follow the wrong recipe. For moreof the context for whichit was meantto solve problems example,perhaps she was ordered to follow it. By contrast, a it has less danger of doing the wrong thing in a different reasonableinterpretation for (lb) is that Veronica’sbelief domain.It will still fit those situations it wasmeant for best only minimallyinvolved in her conscious reasoning, so that becauseof its context. she did not very consciouslybelieve that she followeda recipe Finally a word about possible applications this kind of that was wrong. Continuation (la) does not have a metaphor system would suit especially well. Anyapplication where that can guide us towards a certain interpretation, but it cases are available or can be extracted should be fine. A seems a reasonable (default) assumption that she believes specific example would be a computer aided instruction it consciouslyand thus has a special reason just as in (lc). system where a student interacts with a system to gain However,the pressure towards this explanation is less strong knowledge. For each student the system can keep track of than in the case of (lc). what is learnt and what mistakes were made.When interacting Apart from the two metaphors mentioned above, other with a student the system will use this knowledgeto know important ones we have considered are MINDPARTS AS what it has to explain and what it can assume as known. PERSONSand COGNIZINGAS SEEING. Under the former, Previous mistakes would help to understand new mistakes an agent’s mind is viewed as being broken up into separate and even anticipate mistakes. Not only cases of the student "sub-persons"which have their ownmental states, emotions, could be used, the system might also use other students’ etc., and are often cast as communicatingwith each other in experiences to anticipate and understand. natural language. A vivid exampleof the use of the metaphor is "One part of Veronica was vociferously insisting that ATT-Meta the recipe was wrong." An example of COGNIZINGAS The second system we present in this paper, ATT-Meta, SEEINGis "Veronica only dimly saw that the recipe was is rule-based. It combines metaphor-based reasoning and wrong." simulative reasoning to reason about discourse coherence in ATr-Metais unusual with respect to other systems for certain specialized ways. In mundanetext and speech, mental reasoning about mental states, in that it copes with metaphor- states are often described with the aid of metaphors, and ical descriptions. ATr-Metais also unusual with respect to the metaphorical descriptions can convey information that other systems that address metaphor(whether the metaphors is crucial for understanding the discourse. Being able to are about mental states or not) in that ATr-Metadoes not reason with mental metaphors helps a system to understand (in general) translate metaphoricaldescriptions into internal, characters that might forget certain things, might not make non-metaphoricalversions, but instead represents them liter- the obvious conclusions, might act in certain waysand might ally (see section 2.8). For examplethe internal representation makecertain inferences. Crucial information bearing on those of (lb) itself casts Veronica’s mindas a physicalspace that has points is often communicatedvia metaphors. recesses. This is done because it is hard to choose metaphor- ATr-Metaconsists of two parts. Thefirst part (still under free representations that can be used in the sameways as the construction) has as input the discourse, and produces the metaphor-imbued ones. Also, keeping the metaphor makes logical representation of the discourse. The latter is the the system moreflexible with regard to novel manifestations input for the reasoning part whichhas an advancedprototype of a metaphor.(Novelty to the system, not necessarily to the implementation (demonstrated at the 1994 meeting of the language community.)Suppose, for instance, the system is Association for Computational Linguistics). ATl’-Metais told that described in moredetail in Barndenet al. (1994a, 1994b). "One part of John was vociferously insisting that Sally To see the importance of metaphors in mental-state de- was being unfair, but another was calmly resigned to this scription, consider the following discourse fragment(1), with idea." possible continuations(1 a-c): This sentence directly contributes premises saying that a subpersonof John was [literally] vociferously insisting that (1) Veronica was preparing for her dinner party. Her Sally was being unfair, but another subpersonwas [literally]

132 calmly resigned to this idea. Now,the defeasible inferences itative certainty factors to the conclusions it draws. These from these premises could include: factors result fromcertainty factors attached to rules in the the first subpersonbelieved that Sally was being unfair, system’s knowledgebase and from the certainty factors of and did not like her being unfair; and the propositions the rules work on. Conclusionsthat are not the second subpersonbelieved that Sally was being unfair absolutely certain can be defeated by contrary conclusions but did not mind this being the case. from other lines of reasoning. Such defeats arise partly with Through very general "transfer rules" that map from the aid of a specificity-based heuristic. In addition, agents’ metaphorical representations to metaphor-freeones, the sys- beliefs themselveshave qualitative certainty factors. Thefac- tem could then infer that John (as a whole)believed that Sally tors at different levels are independent.Thus, it is possible for was being unfair, and to some extent disliked and and to ATT-Metato be certain that agent X believes P as a default; someextent did not mindher being unfair. The point is that alternatively, ATI’-Metamight assumeby default that agent the system does not need to have any transfer rules concern- Xis certain that P. Certainty factors that are "within agents’ ing insisting, vociferousness, calmnessand resignation, that minds" are manipulated during simulative reasoning in just is, concerningthe intermediate notions used in the chain of the same way as the certainty factors at the system’s own reasoning. level. ATr-Meta uses a combination of simulative and non- Onkinds of reasoning and self-knowledge: by virtue of using simulative reasoning to reason about other agents’ beliefs simulative reasoning, AT’l’-Metaascribes its owncomplex, (see section 2.5. for whysuch a combinationis necessary). sophisticated default/uncertain reasoning method to other To avoid logical omniscience,results of simulative reasoning agents, without needing to describe that reasoning explicitly are at most default results. Metaphoricalinformation can be to itself. It can also ascribe its simulative reasoning and its used to downgradethese results further. For instance, if agent metaphor-basedreasoning to other agents, to any degree of X believes P and believes Q, and R follows from P and Q, nesting. See Barnden et ah (1994b) for some discussion then the system would normally conclude (as a default) that the significance of various types of nesting. X believes R. However,if the input discourse implies that P Unlike CaseMent, ATT-Metahas no special handle on and Q are "far apart" in X’s mind, then the system retracts context-sensitivity and belief ascription. the conclusion. Thus, ATI’-Meta’ssimulative reasoning can be modulatedby metaphorical reasoning. Conclusion ATT-Metaand the Issues of Section 2 The two systems discussed both contribute to improving reasoning about beliefs, making it more common-sensical, On logical-form: the system uses metaphor-imbuedinter- realistic and flexible, although they focus on different issues nal representations (as well as morestandard, metaphor-free and are based on different techniques. ones). Asa result, its representation style is, in part, funda- Aq’T-Metawas the first of the two systems to be developed mentally different from that of other mental-state reasoning and uses a complicated combination of simulative reasoning, systems. metareasoning, metaphorbased reasoning, multi-valued logic Onintrospection: the system can in principle take a metaphor- and truth maintenance.Making sure that rules interact in a ical view of its ownmental states/processes (although we have desirable way becomes hard, and it is not always easy to not actually madethe system do this yet). Moreusefully, the predict how the system will reason. Even for relatively system can cope with speakers whosay things like "Part of simple questions the system has to do a large amountof work. me wants to go to the party." That is, the system can deal With this in mind it was thought that a system based on with metaphor-basedintrospection by other agents. cases instead of rules might result in a less complicated, Onshades of belief’, by paying attention to metaphors, ATr- easier to understandsystem for reasoning about mental states. Metarepresents a far richer variety of mentalstates than other Initial results look promisingbut still a great deal of work systems are capable of. In particular, examples like (lb) has to be done before final conclusions can be made. The and (lc) above illustrate that AT’l’-Metaaddresses different current case-based system, CaseMent, does not focus on shades of belief. mental metaphors. Future work might include trying to give On reusability: when reasoning about a metaphorically- CaseMentsimilar metaphoricalreasoning capabilities to those described mental state, ATI’-Metauses general knowledge AqT-Metademonstrates. about the source domainof the metaphor.E.g., if the metaphor On the other hand, there is room for rules in a case- is MINDAS PHYSICALSPACE, ATI’-Meta uses general based system that reasons about mental states. There has knowledgeabout physical space (and physical objects). This been a considerable amount of work in the CBRcommunity knowledge could also be used for other, non-metaphorical on integrating case-base and rule-based reasoning --- see, purposes, and is thus more reusable than the knowledgethe e.g., Kolodner (1993). Currently, our case-based system system would need to have if it were to get the same effect essentially free of rules, but that is largely in order to simplify by reasoning entirely with metaphor-free representations of the research issues. mental states. In fact, in developing ATr-Meta we have madea conscious effort to ensure that our knowledgebases for domainslike physical space are useful for reasoning in References those domains as such, rather than just for metaphorical Attardi, G. & Simi, M. (1994). Proofs in context. In reasoning about minds. Doyle, E. Sandewall & P. Torasso (Eds), Principles of Onuncertainty and incompleteness: ATT-Metaattaches qual- Knowledge Representation and Reasoning: Proceedings

133 of the Fourth International Conference, pp.15--26. San ligence, Tokyo, pp. 176-181. Los Altos, CA: Morgan Mateo, CA: Morgan Kaufmann. Kaufmann. Ballim, Afzal and Yorick Wilks (1991). Artificial Believers: Davies, M. and Stone, A. (Eds) (in press). Mentalsimulation. The Ascription of Belief Hillsdale, NewJersey: Lawrence Oxford: Blackwell. ErlbaumAssociates, Publishers. Davis, E. (1990). Representations of CommonsenseKnowl- Barnden, J.A. (1988). Propositional attitudes, commonsense edge. San Mateo, California: MorganKaufmann Publish- reasoning, and metaphor. In Procs. lOth Annual Con- ers, Inc. ference of the Cognitive Science Society, Hillsdale, N.J.: Dinsmore, John (1991). Partitioned representations. Dor- LawrenceErlbaum. pp.347--353. drecht: Kluwer AcademicPublishers. Barnden, J.A. (1989a). Belief, metaphorically speaking. Fesmire, S.A. (1994). Aerating the mind: the metaphor In Procs. 1st Intl. Conf. on Principles of Knowledge mental functioning as bodily functioning. Metaphor and Representation and Reasoning (Toronto, May1989). San SymbolicActivity, 9 (1), pp.31--44. Mateo, CA: Morgan Kaufmann. pp.21--32. Gallup, G.G., Jr. & Cameron,P.A. (1992). Modality specific B arnden, J.A. (1989b). Towardsa paradigmshift in belief rep- metaphors: is our mental machinery"colored" by a visual resentation methodology.J. Experimental and Theoretical bias? Metaphorsand SymbolicActivity, 7 (2), pp.93-- 101. Artificial Intelligence 2, 1989, pp. 133-- 161. Gentner, D. & Grudin, R. (1985). The evolution of mental Barnden, J.A. (1992). Belief in metaphor: taking common- metaphors in psychology: A 90-year perspective. American sense psychologyseriously. ComputationalIntelligence, 8 Psychologist, 40 (2), pp. 181-- 192. (3), pp.520--552. Gibbs, R.W., Jr. & O’Brien, J.E. (1990). Idioms and Barnden, John A. (in press). Simulative Reasoning, Com- mental imagery: the metaphorical motivation for idiomatic monsensePsychology and Artificial Intelligence. In Mental meaning. Cognition, 36 (1), pp.35--68. Simulation, edited by M. Davies and A. Stone. Haas, AndrewR. (1986). A Syntactic Theory of Belief and Barnden, J.A., S. Helmreich, E. Iverson and G.C. Stein Action. Artificial Intelligence. Volume28, pp. 245-292. (1994a). An Integrated Implementationof Simulative, Un- Hadley, Robert F. (1988). A Process-Oriented, Intensional certain and Metaphorical Reasoning about Mental States. Model of Knowledge and Belief. Proceedings of the Principles of Knowledge Representation and Reason- l Oth Annual Conferenceof the Cognitive Science Society, ing: Proceedings of the Fourth International Conference, (Montreal), pp. 354-360. Hillsdale, NJ: LawrenceErlbaum pp.27--38. San Mateo, CA: MorganKaufmann. Associates. Barnden, J.A., S. Helmreich, E. Iverson and G.C. Stein Hobbs, Jerry R., Mark E. Stickel, Douglas E. Appelt and (1994b). CombiningSimulative and Metaphor-BasedRea- Paul Martin (1993). Interpretation as abduction. Artificial soning about Beliefs. Procs. 16th Annual Conference of Intelligence 63, pp. 69-142. the Cognitive Science Society (Atlanta, Georgia, August Hobbs, J.R. (1990). Literature and cognition. CSLILecture 1994), pp.21--26. Hillsdale, N.J.: LawrenceErlbaum. Notes, No. 21, Center for the Study of Language and Barnden,J.A., Helmreich,S., Iverson, E. & Stein, G.C. (forth- Information, Stanford University, Stanford, CA. coming). Artificial intelligence and metaphors of mind: Jakel, O. (1993). The metaphorical concept of mind: mental Advantages of within-vehicle reasoning. Metaphor and activity is manipulation. Paper No. 333, General and The- Symbolic Activity. Acceptedwith revisions encouraged. oretical Papers, Series A, Linguistic Agency,University of Belleza, F.S. (1992). The mind’s eye in expert memorizers’ Duisburg, D-4100 Duisburg, Germany. descriptions of remembering. Metaphor and Symbolic Johnson, M. (1987). The body in the mind. Chicago: Chicago Activity, 7(3 & 4), pp.119--133. University Press. Bruner, J. & Feldman, C.F. (1990). Metaphors of con- Kolodner, Janet (1993). Case-Based Reasoning. San Mateo, sciousness and cognition in the history of psychology. In CA: MorganKaufmann Publishers, Inc. D.E. Leary (Ed.), Metaphorsin the History of Psychology, Konolige, K. (1986). A Deduction Modelof Belief. London, pp.230--238. NewYork: CambridgeUniversity Press. U.K.: Pitman. Chalupsky, Hans (1993). Using hypothetical reasoning as Lakemeyer, Gerhard (1991). On the Relation between Ex- methodfor belief ascription. J. Expt. Theor. Artif IntelL, plicit and Implicit Belief. Proceedings of KR ’91, pp. Volume5, pp. 119- 133. 368-375. San Mateo, California: Morgan KaufmannPub- Cooke, N.J. & Bartha, M.C. (1992). An empirical investiga- lishers, Inc. tion of psychological metaphor. Metaphor and Symbolic Lakoff, G., Espenson, J. & Schwartz, A. (1991). Master Activity, 7 (3 & 4), pp.215--235. metaphor list. Draft 2nd Edition. Cognitive Linguistics Cox, M.T. (1995). Representing Mental Events (or the Lack Group, University of California at Berkeley, Berkely, CA. Thereof). In M. Cox & M. Freed (Eds.), Proceedings of Lakoff, G. & Johnson, M. (1980). Metaphors we live by. the 1995 AAAI Spring Symposiumon Representing Mental Chicago: University of ChicagoPress. States and Mechanisms. MenloPark, CA: AAAIPress. Larsen, S.F. (1987). Rememberingand the archaeology Cravo, Maria R. and Jo~to P. Martins (1993). SNePSwD: metaphor. Metaphorand Symbolic Activity, 2 (3), 187-- newcomer to the SNePSfamily. J. Expt. Theor. Artif 199. Intell., Volume5, pp. 135 - 148. Levesque, Hector J. (1984). A Logic of Implicit and Explicit Creary, Lewis G. (1979). Propositional attitudes: fregean Belief. Proceedings of the National Conference on Arti- representation and simulative reasoning. Proceedings of ficial Intelligence, pp. 198-202. Los Altos, CA: Morgan the 6th International Joint Conferenceon Artificial Intel- Kaufmann.

134 Maida, Anthony S., Jacques Wainer and Sehyeong Cho (1991). A syntactic approachto introspection and reasoning about the beliefs of other agents. FundamentaInformatica XV, pp. 333-356. Mamede, Nuno J. (1993). SNePSR-- A SNePS with re- sources. Journal Expt. Theor. Artificial Intelligence, Volume5, pp. 199 - 213. Moore, R. C. (1973). D-SCRIPT:A computational theory descriptions. IEEE Transactions on Computers, C-25(4), 366--373 (1976). Oehlmann,R. & P. Edwards(1995). Introspection planning: Representing metacognitive experience. In M. Cox & M. Freed (Eds.), Proceedings of the 1995 AAAI Spring Sym- posium on Representing Mental States and Mechanisms. Menlo Park, CA: AAAIPress. Richard, M. (1990). Propositional attitudes: an essay on thoughts and how we ascribe them. Cambridge, U.K.: CambrdigeUniversity Press. Richards, G. (1989). On psychological language and the physiomorphicbasis of humannature. London: Routledge. Riesbeck, Christopher K. and Roger C. Schank(1989). Inside Case-based Reasoning. Hillsdale, NewJersey: Lawrence ErlbaumAssociates, Publishers. Sweetser, E.E. (1987). Metaphorical models of thought and speech: a comparison of historical directions and metaphorical mappingsin the two domains. In J. Aske, N. Beery, L. Michaelis & H. Filip (Eds), Procs. 13th Annual Meetingof the Berkeley Linguistics Society. Berkeley, CA: BerkeleyLinguistics Society, pp. 446--459. Sweetser, E.E. (1990). From etymology to pragmatics: metaphorical and cultural aspects of semantic structure. Cambridge, U.K.: CambridgeUniversity Press. Tomlinson, B. (1986). Cooking, mining, gardening, hunting: metaphorical stories writers tell about their composing processes. Metaphorand Symbolic Activity, 1 (1), 57--79.