Description Logic as Core Machinery for the Automatic Generation of Multilingual Technical Documents

level. Below such a general, linguistically motivated, classi cation we have been modelling a level which we call the middle model. This level contains knowledge about physical and functional objects (e. g. electrical objects, connections etc.) and their potential states (e. g. power-supplied, tightened etc.) in the domain of technical products. Additionally, this level includes the principles which describe how these states can be achieved and how they in uence other objects which are related together to form a complex whole. Finally, the domain model represents the lowest level and contains knowledge about particular technical devices in a concrete application (e. g. the toaster domain). 1.2 Representation Language Our experience is that a representation system only supporting description logics is not able to master all the requirements above. The integration of additional programming paradigms in the representation system is therefore a necessary condition, in order to realize adequate models. In our opinion a reasonable representation system must supply an inference engine which successfully integrates frame{like knowledge, as well as rule{ like knowledge combined with a generalization of the conception of object{oriented method dispatching. We are using the knowledge representation tool called Loom [MacGregor and Bates, 1987], a descendent of the KL-ONE family and based on Lisp. Loom is a good candidate for our representation purpose because it o ers more than just a very powerful description logic. Besides this Loom is already used (although not really exploited) for the upper model of Penman. In addition Loom provides powerful deductive support with forward{ and backward{chaining, including both strict and default reasoning and automatic consistency checking. It also o ers procedural programming, a full rst{ order query language, production rules, multiple knowledge bases, and object-oriented methods [MacGregor, 1991b]. In addition, the empirical analysis of six terminological representation systems in [Heinsohn et al., 1994] showed Loom to be the most expressive and fastest one.3 However, Loom's inference algorithm is incomplete [MacGregor, 1991a]. This deliberate conceptual decision of Loom's developers is not surprising since it has been shown that determining subsumption between terms, which is needed for the classi cation process, is NP-hard [Domini et al., 1991] or even undecidable [Schmidt-Schau , 1989] for reasonably expressive languages. 3The representation systems were: BACK, CLASSIC, KRIS, LOOM, MESON and SB-ONE. 2 Extending the Modelling Capacities For any adequate modelling of physical systems, it is important to identify the relevant phenomena, and to identify just the appropriate level of detail to model each phenomenon. Hence, a knowledge engineer must have the a priori information about the inferences to be drawn in his particular application. Adequate models, therefore, incorporate abstractions and approximations that are well suited to the problem solving task [Nayak, 1995] and to the expressiveness, performance and completeness of the representation system. Moreover, the representation should conform to the crucial modelling quality factors of object{oriented systems: reusability, understandability and extensibility [Meyer, 1988]. To identify the relevant phenomena and the appropriate level of detail, we focused our objects from a functional and structural perspective, we will not address here (but cf. [Liebig and R osner, 1996]). 2.1 Part{Whole Relations The part{whole relation plays a fundamental role in the description of complex objects. This relation can be found in many di erent domains. As in Franconi's [Franconi, 1993] proposal, we model the part{whole relationship as a re exive, anti{symmetric and transitive relation : 8x: (x; x): 8x; y: (x; y)^ (y; x)! x = y: 8x; y; z: (x; y)^ (y; z)! (x; z): As in Sattler's [Sattler, 1995] concept language P for engineering applications, our part{whole relations are direct part{whole relations in the sense that they must satisfy the immediate inferior de nition.4 As noted in [Artale et al., 1995], this constraint should a ect the ABox reasoning process in order to discard non{intended models. In contrast to Sattler's application, which lacks ABox reasoning, we use data{driven rules in order to conform to the above de nitions of immediate inferiors and anti{symmetry. These rules observe the actual knowledge base and cause a warning whenever an assertion violates a particular de nition. As discussed in [Artale et al., 1995], part{whole relations cannot simply be modelled by ordinary attributes like price or color. The representation formalism should take their speci c meaning and their transitivity rules into account. To illustrate this, consider the example taken from [Artale et al., 1995]: \an arm is part of a musician, the musician is part of an orchestra, but it would sound a bit strange to state that the arm is part of the orchestra". The inacceptability of this inference is due 4Given a partially ordered set P , we say that a is an immediate inferior of b if a < b and there does not exist an x 2 P such that a < x < b. to a mixing of two di erent interpretational meanings of the the part{whole relation. Winston, Cha n and Herrmann [Winston et al., 1987] proposed a distinction between six kinds of specialized part{whole relations to overcome such problems: Component/Integral{Object, Member/Collection, Portion/Mass, Stu /Object, Feature/Activity and Place/ Area. These distinctions allow more suggestive reasoning mechanisms along the part{whole relationships, as far as one single kind of relation is involved.5 All speci c part{whole relationships are modelled as specializations of a general part{whole relation. The transitive version of this relation does, therefore, hold between parts and wholes which have chains of di erent kinds of part{whole relations between them. To determine transitive part{whole relationships we have de ned additional relations capturing the transitive closure of each of the relations above.6 The most important part{whole relationship in our context is between integral{objects which have a structure, and their components which have a speci c functionality and which are separable. Part{whole relations can be also de ned as essential. Instances in the range of essential relations are then automatically classi ed as essential{parts. This allows for the automatic determination of relevant components in order to support the knowledge engineer in building up his domain. Knowledge about essential parts is available at concept level, so even when there are no instances created, the user can ask for the essential parts of a speci c device (example in [Liebig and R osner, 1996]). To continue in this direction, we propose automatic generation of all essential parts when instantiating the respective whole. In order to have a generic function for this purpose, it is necessary to generalize the conception of object{oriented method dispatching to concepts and relations. 2.2 Functional Aspects The representation we propose combines structural and functional information about a complex object (e. g. a device). As discussed in [Keuneke, 1991], functional structuring is useful for problem{solving mechanisms, which must often decompose the device's function into the functicion of the components. The functional speci cation describes the device's goals at a level of abstraction that is of interest at the object level [Keuneke, 1991]. 5However, the composition of di erent part{whole relations can have relevant meanings too, but not in all cases (see composition table in [Sattler, 1995]). 6The de nition of the transitive closure includes the ltering of the re exivity property of the part{whole relations in order to prevent cyclic paths. For a detailed discussion see [Liebig and R osner, 1996]. We use a model based on a structural organization enriched with functional components. The function of a device is its intended purpose, which is achieved by behaviours [Keuneke, 1991]. Our model represents behaviour as the causal sequence of transitions of partial states/predicates. Determining the actual states depends on the individual abstract object. An electrical control appliance has, for example, the status \on" if all of its components are in a status in which they close the underlying electrical circuit. The control appliance of electrical devices usually consists of a set of switches. These di erent kinds of switches (e.g. binary switch, tune etc.) can be adjusted by using a generic action. This action must be applicable to all kind of switches and should therefore use object{oriented method dispatching, generalized to the purposes of description logics. In order to capture the essentials of complex objects we need to express \vertical" and \horizontal" relationships and constraints. 2.3 Vertical Inheritance Vertical dependencies can be di erentiated into relationships between the existence of the whole and the existence of certain parts, and relationships between the properties of a whole and the properties of its parts [Simons, 1987]. Existential dependencies have been addressed by proposing relations as essential. The class of relationships between the properties of a whole and the properties of its parts (and vice versa) can be di erentiated into [Artale et al., 1995]: (a) Properties which the parts inherit from the whole. (b) Properties which the whole inherits from its parts. (c) Properties of the parts which are systematically related to properties of the whole (These are not yet captured in our model). We describe the rst two varieties in turn: (a) The location is a property which parts inherit from the whole (with respect to a certain granularity). In order to inherit properties of this kind we have written macros expressing Franconi's [Franconi, 1993] left and right distributive quanti ers for relations from his language ALCS: C:R(a; b) i 8x:( (a; x) ^ C(x)) ! R(x; b) C:R(a; b) i 8x:( (b; x) ^ C(x)) ! R(a; x): The operators and express the left and right distributive readings. They can be quali ed by a qualication predicate C, which was omitted in our macros (formally we assume C >). These macros extend the description logic by adding backward{chaining implication rules to the knowledge base. Figure 1 shows the inheritance of a property represented by the relation R. ttttttJJJJJJJJJJ......................^...........................................^..................... ............ ............ ............ ............ ............ ............baxR(a; b)Figure 1: The thin lines indicate the part{whole relations( ); the dashed line represents the relation R which isinherited (dotted arrows) left{distributive to all parts ofa (displayed for x only).Since we also express properties in form of concepts,we extend the quali ed plural quanti ers with a dis-tributive reading for concepts:D(a) i 8x: (a; x):! D(x)(b) Properties which the whole inherits from its partsare, in the domain of a household toaster, \on" or\power{supplied". In order to inherit these prop-erties upward, we have written macros which ex-pand to backward{chaining implication rules, similar toFranconi's quali ed plural quanti ers, but based on theinverse of the part{whole relation.2.4 Horizontal InheritanceHorizontal relationships are composed of constraintsamong parts which characterize the integrity of thewhole. Although they are important for capturing thenotion of a whole, they nd little attention in currentmodelling formalisms [Artale et al., 1995].To determine the energized or hot parts of a device,for example, it is necessary to model the electrical con-nections of the components of the device. In order torecognize an electrical object as energized, the reason-ing mechanism must nd those objects which have anelectrical connnection to an internal or external powersupply. Additionall, the status of any switch on this pathmust be taken into account. The determination of thispath along the electrical connection between objects isdirectly encoded in Lisp as a depth{ rst search exten-ded with additional ltering procedures.Horizontal and vertical inheritance often in uence eachother in order to create complex dependencies. For ex-ample, the temperature status of the heating wire de-pends on the horizontal constraint of being energized.The temperature status of the toaster, therefore, is in u-enced by the temperature of the heating wire in a verticalmanner.2.5 Temporal AspectsIn a rst version of the knowledge base switching thetoaster \on" or \o " does immediatly in uence the tem-perature status of the heating wire and the toaster.However, this does not model the real world adequatelyto be able to derive knowledge about hot objects. Forexample, the heating wire does not become hot until ithas been energized for a little while. Analogously, thewire remains hot for a few minutes after it is no longerenergized.To capture these regularities, one will need a descrip-tion logic with temporal extensions. Loom's extensionfor temporal concepts and relations allows us to makefactual assertions about role llers and instances thathold only over speci ed intervals, rather than being uni-versally true.A drawback of the current version of Loom: it wasnot possible to express some relevant temporal depen-dencies in concept or relation de nitions directly. Wewere forced to write production rules in order to detectrelevant temporal changes in the knowledge base. Thecorresponding explicit temporal assertions were placedin the action parts of these rules.3 Concluding RemarksAs a guideline for structuring complex objects, we pro-pose a model using a physically oriented organizationenriched with functional extensions. The transitive part{whole relation and the ability to inherit properties verti-cally and horizontally play key roles in this structuring.This skeleton can be used as a basis for a large varietyof di erent modelling purposes, not only in the domainof technical products. Imagine, for example, the mod-elling of a structured organisation: their administrationprocesses and the inheritance of competences and tasks.We have enriched the representation language in orderto simply enable upwards or downwards inheritance ofconcepts or relations.In order to reject unintended models, we use ABoxreasoning. For this purpose we needed production rulessuitable for triggering concepts and relations.Object{oriented methods, gerneralized for conceptsand relations, were needed for generically applicable cre-ation or adjust methods.In our opinion, description logic can only build the coreof a programming envionment which ful lls the require-ments that came along with the multiple uses (text gener-ation, qualitative simulation etc.) of the knowledge to be represented. Additional paradigms like data{driven pro-gramming and object{oriented programming are needed.Indeed, theses paradigms have to be carefully integratedinto the description language without missing importantclasses of inferences.Suggestions for future work include a more conformedintegration of temporal extensions into the model. Sincemany properties are time{dependent, it seems inadequateto write production rules for every particular temporalconcept or relation in order to assert temporal facts.References[Artale et al., 1995] Alessandro Artale, Enrico Franconi,Nicola Guarino, and Luca Pazzi. Part{whole rela-tions in object-centered systems: An overview. Data& Knowledge Engineering { North{Holland, Elsevier,December 1995.[Bateman et al., 1990] John A. Bateman, Robert T.Kasper, Johanna D. Moore, and Richard A. Whitney.A general organization of knowledge for natural lan-guage processing: the penman upper model. Technicalreport, ISI, 1990.[Domini et al., 1991] F. Domini, M. Lenzerini, D. Nardi,and W. Nutt. The complexity of concept languages.In J. F. Doyle, R. Files, and Erik Sandewall, editors,Principles of Knowledge Representation and Reason-ing, Proceedings of the Second International Confer-ence (KR '91), pages 151 { 162, Cambridge, MA, April1991. Morgan Kaufmann Publishers, Inc., San Fran-cisco, CA.[Franconi, 1993] Enrico Franconi. A treatment of pluralsand plural quanti cations based on a theory of collec-tions. Minds and Machines, 3:453 { 474, 1993.[Heinsohn et al., 1994] Jochen Heinsohn, Daniel Kuden-ko, Bernhard Nebel, and Hans-Jurgen Pro tlich. Anempirical analysis of terminological representation sys-tems. Arti cial Intelligence, 68(2):367 { 398, 1994.[Keuneke, 1991] Anne M. Keuneke. Device representa-tion, the signi cance of functional knowledge. IEEEEXPERT, pages 22 { 25, April 1991.[Liebig and Rosner, 1996] Thorsten Liebig and DietmarRosner. Modelling of reusable product knowledge interminological logics: a case study. In Proceedings ofthe First International Conference on Practical As-pects of Knowledge Management, | to appear |,1996.[MacGregor and Bates, 1987] R. MacGregor and R.Bates. The LOOM knowledge representation lan-guage. Technical report, ISI/RS-87-188, ISI, Uni-versity of Southern California, 1987.[MacGregor, 1991a] Robert M. MacGregor. Inside theLOOM Description Classi er. SIGART Bulletin,2(3):88 { 92, June 1991.[MacGregor, 1991b] Robert M. MacGregor. Using a de-scription classi er to enhance deductive inference. InProceedings of the Seventh IEEE Conference on AIApplications, pages 141 { 147, 1991.[Mann, 1983] William C. Mann. An overview of the pen-man text generation system. In Proceedings of the Na-tional Conference on Arti cial Intelligence, pages 261{ 265. AAAI, August 1983.[Meyer, 1988] B. Meyer. Object{oriented Software Con-struction. Prentice Hall, New York, 1988.[Nayak, 1995] P. Pandurang Nayak. Automated Mod-eling of Physical Systems. Number 1003 in LectureNotes in Arti cial Intelligence. Springer Verlag, Ber-lin, 1995.[Rosner et al., 1996] Dietmar Rosner, Bjorn Ho ing,and Knut Hartmann. From natural language docu-ments to sharable product knowledge. In Proceed-ings of the First International Conference on PracticalAspects of Knowledge Management, | to appear |,1996.[Sattler, 1995] Ulrike Sattler. A concept language for anengineering application with part{whole relations. InA. Borigida, M. Lenzerini, D. Nardi, and B. Nebel,editors, Proccedings of the International Workshop onDescription Logics, pages 119 { 123, Rome, Italy, June1995.[Schmidt-Schau , 1989] M. Schmidt-Schau . Subsump-tion in KL-ONE is undecidable. In H. J. Levesqueand R. Reiter, editors, Principles of Knowledge Re-presentation and Reasoning, Proceedings of the FirstInternational Conference (KR '89), pages 421 { 431,Toronto, ON, May 1989. Morgan Kaufmann Publish-ers, Inc., San Francisco, CA.[Simons, 1987] Peter Simons. Parts: A Study in Onto-logy. Clarendon Press, Oxford, 1987.[Stede and Rosner, 1994] Manfred Stede and DietmarRosner. Generating multilingual documents froma knowledge base: The TECHDOC project. InCOLING-94, Proceedings, Kyoto, 1994.[Winston et al., 1987] Morton Winston, Roger Cha n,and Douglas Herrmann. A taxonomy of part-wholerelations. Cognitive Science, 11:417{444, 1987.