On Overview of KRL, a Knowledge Representation Language

ion. There is no commitment in KRL as to what sorts of concepts in the domain should be represented at what level of categorization. The specific three types are based on psychological studies (Rosch, 1975; Rosch & Mervis, 1975) that human reasoning makes extensive use of a layered system of categories. The choice of whether a particular prototype (such as Person or Visit) should be basic, abstract, or a specialization depends on the way in which descriptions are built up and used in matching. What is provided is a mechanism by which the careful use of levels can result in achieving many of the efficiency benefits of semantic marker mechanisms and classification trees. Individuals. The KRL matcher and other primitive mechanisms for building SThis includes much of the work on semantics associated with transformational grammar (e.g., Katz & Fodor, 1964), A1 formalisms such as conceptual dependency (Schank, 1975a), and most , forms of case grammar (see Bruce, 1975, for a summary). Ernis tree corresponds to a simple generalization hierarchy, as discussed in Winograd (1975a). 12 ' D. G. BOBROW AND T. WINOGRAD descriptions assumes that different individuds are different unique entities in the world being modeled. For example, no individual can match (in a simple sense) a different individual. An inconsistency is signaled whenever there is an attempt to use pointers to two different individuals as descriptors in a single description. However, there are no built-in assumptions about how individuality should be assigned. The definition of what, should constitute anindividual within .a domain is relative to a particular set of reasoning purposes. As a simple example, in the, air travel domain, a particular flight (including date) could be an individual, with the flight number as a property (filling a slot), or, alternatively, each flight in the schedule (by number) could be treated as an individual, with a particular flight instance represented as a manifestation (see below). Manifestations. Often it is useful to group together a set of descriptions which belong to some individual. There are three main cases in which we anticipate this need, and units with category type manifestation can be used for all of them: 1. Further specified individuals: A manifestation can be used to provide a single memory unit (for purposes of retrieval and content-dependent description) containing a set of descriptions belonging to an individual within one context. For example, we might separate out the physical properties of an object for which we also have functional or historical descriptions, or the description of some person as a scientist from the description of that person as a friend. 2. Contingent properties: An individual can be described using timedependent descriptions without creating a separate manifestation. However, it is often useful to collect a set of descriptions which are true at some time (or in some hypothesized world) and treat them as time-independent descriptions of a manifestation which represents the individual at that particular time. 3. Ghosts: A representation must enable us to describe entities whose unique identity is not known. There are many cases in which we may know many properties of some object without knowing which of the known objects in our world it is. Such objects have at times been called "formal objects" (Sussman, 1975) and "ghosts" (Minsky, 1975). A standard detective story plot involves knowing that one of the people in a house is a murderer, knowing many properties of the murderer, and not knowing which individual it is. The unit used to represent the murderer is a manifestation that has no' associated individual. Relations and propositions. An abstract relationship, such as the relative magnitude of two numbers, can be described using the ideas of slots and description we have used so far. There is a unit, with category class relation, which represents the relationship (or predicate) as an abstract mapping; a proposition unit represents each instantiation of the relationship. The truth value of a proposition is specified explicitly rather than being determined as an implicit consequence of its existence in the data base. AN OVERVIEW OF KRL 13 2.4 The Family of Descriptors Each descriptor in a description is an independent characterization of the object associated with the description. The variety of descriptor types corresponds to the notion of natural description discussed above. Each descriptor type is intended to express a different mode of describing conceptual objects. The syntax of descriptors depends on key words (such as a , the, from, which) based on analogy with simple English phrases. They are mnemonic indicators for a set of precisely defined structures within the formalism. This set of descriptors was not designed with the goal of boiling everything down to the smallest possible set of mechanisms. On the contrary, it is based on an attempt to provide a simple and nahlral way of stating information conceptualized in di.fferent ways. There is a great deal of overlap. For example, the notion of "bachelor" might be represented in any of the following ways: There could be a prototype unit for Bachelor, with an individual described as (a Bachelor with. . .) Bachelorhood could be represented .indirectly by having a prototype for MalePerson and Adult, and a predicate for IsMarried, and using the description: {(a MalePerson) (an Adult) (NOT (which IsMarried))) There could be a unit representing a Marriage with slots for the malepartner and fernaleparmer, and a description: {(a MalePerson) (an Adult) (NOT (the malepartner from (a ~ a n i a ~ e ) ) ) ) There could be a one-place predicate, IsBachelor. The predicate definition might (but need not) include a special procedural test for bachelorhood. An individual would then be described using the predication (which IsBachelor) These KRL forms are described in general in Fig. 3. No one of these forms is automatically primary. All of them could coexist, and be defined in terms of each other. The system provides the necessary reasoning mechanisms to interrelate the different forms in which essentially equivalent information could appear, and the hope is that additional knowledge (especially procedural knowledge) which is best stated with respect to any one form can be represented directly. Our intuition leads us to believe that prototypes and perspectives will most often -serve as the fundamental organizing representation, with the others serving to provide secondary information. In order to demonstrate the different uses of these descriptors, we present here a more extended example. It is based on a hypothetical system that acts as a travel assistant, making reservations and computing costs of trips. As with the example above, this is greatly oversimplified and is not intended as a careful 14 D. G. BOBROW AND T. WINOGRAD . , ' Descriptor name: direct pointer Format: a unit name, number, string, or quoted LISF? object' Use: A pointer to units, or to data directly in the description. Provides a unique identifier (this includes using a proper name like "Boston") Examples: Block1 7, PaloAlto, 356, "a string", (QUOTE (A PIECE (OF LIST) STRUCTURE))) + i * m + + + + r + + * + + + Descriptor name: perspective Format: (a prototype with identifier1 = fillerl ... identifier, = filler,) Use: ,.Assigns an object to membership in a category (such as "city"). A ,comparison of the current object with the "prototype", with slots further specifying this object Examples: (a Trip with destination = boston airline = TWA) *************** Descriptor n a r k specification Format: (the slotspecifier from view targetDescription) Use: Specifies the current object in terms of its role in a perspective of prototype: "view". States a role in a complex object or event (e.g., the "destination" of a particular trip) Examples: (the actor from Act (a Chase with quarry = {Car22 (a Dodge))) *************** Descriptor name: predication Format: (which predicateName . predicateArgs) Use: Describes a relationship in which the object is a participant (being "to the North of Providence"). Defined in terms of a specification. A way of specifying an object in terms of a relation and arguments; allows special procedural attachment. Examples: (which Owns (a Dog)) (which IsBetween Block17 (a Pyramid)) *************** Descriptor name: logical boolean Format: (OR . booleanArgs) or (XOR . booleanArgs) or (NOT booleanArg) Use: Simple logical connectives. A description is an implicit AND of descriptors, thus AND is not needed Examples: (OR (a Dog) {(a Cat)(which hascolor Brown))) (NOT (a Pet with owner = (a Student))) * * * * * * * * * * * * * * * FIG. 3 Different descriptor types in KRGO (a partial list). AN OVERVIEW OF KRL Descriptor name: restriction Format: (theone restrictionDesc) Use: Marks the enclosed description as being sufficient to .refer to 'a unique object in context Examples: (theone {(a Mouse)(which Owns (a Dog))} * * * * * * * * * * * * * * * Descriptor name: selection Format: (using selectionDesc selectFrom selectionPatternl selectionl ... ~electionpattern~ selection, otherwise defaultSelection) Use: This is a declarative form corresponding to CASE or SELECT statements in programming languages. Examples: (using (the age from Person ThisOne) selectFrom (which isLessThan 2) Infant (which isAtLeast 12) Adult otherwise Child) Descriptor name: set specification Format: one of: (SetOf setElementDescription), (In . setDescription), (Items . elements), (Notltems . elements), (Allltems . elements), (ListOf . elements), (Sequence . elements), Use: These descriptors allow specification of partial information about sets, sequences and lists. Describes an object in terms of membership in a set, or a set in terms of the objects it contains Examples: (SetOf {(an Integer)(which hasFactor 2))) ... all elements are even numbers (Items 2 4) ... at least 2 and4 are in this set (Allltems ,2 4 64 {(an Integer)(which hasFactor 3))) ... a four element set (Notltems 5 1 ) ... 51 is not in this set (In {(Setof (an Integer)) (Items 2 5 8) (Notltems 4))) ... describes an object in a set of integers which contains at least 2 5 8, and not 4 *****.**.*.****