Use and reuse of legal ontologies in knowledge engineering and information management

Transcription

1 Use and reuse of legal ontologies in knowledge engineering and information management Joost Breuker and Radboud Winkels University of Amsterdam Department of Computational Jurisprudence (LRI) P.O. Box 1030 NL 1000 BA Amsterdam, the Netherlands April 2003 Abstract The research reported here covers a long period of time over a number of mid- size (European) projects. In this article we will pursue and discuss the views and results related to the development of a core ontology that identifies the main concepts that are typical, and preferably exclusive for law. These are only a few, those related to normative knowledge (deontic terms) and to notions about legal responsibility. We have also identified terms which are not exclusive, but still very typical and well elaborated for law: document, document-structure, role, etc. are central terms in law and may be grounded in a core ontology that imports these notions from a high level and simple foundational ontology. The vast majority of terms or concepts found in legal sources refers to common sense, albeit a special and often more restricted version of common sense knowledge. Of course we can never represent all common sense knowledge, so we will have to resort to foundational ontologies. For various reasons we cannot use existing ones, so we are developing our own: LRI Core. For specific legal areas, domain ontologies are built that hook into LRI Core. This combination of ontologies has been used in several projects, mainly for advanced information retrieval purposes. First of all, the ontologies can be used as a vocabulary for indexing and tagging legal sources. Secondly, they can be used to expand terms used in search queries in various ways. Finally, they can be used for disambiguation and clustering of return sets. 1 Introduction This article presents insights acquired during more than a decade of development, use and reuse of legal ontologies at LRI-UvA. 1 These ontologies were/are constructed in various projects concerned with the development of legal knowledge systems and legal information management. These insights provide a framework rather than a methodology for modeling legal knowledge and reasoning. Results and experiences in the projects are used to illustrate this framework. Although this framework is operationalized by reusable legal knowledge system architectures (e.g. ON-LINE [Valente et al., 1999], TRACS [Breuker & Den Haan, 1991], [Haan, 1996]), the focus of this article is on discussing the conceptual views on legal knowledge and reasoning. In the projects a large variety of legal domains have been analyzed and modeled: traffic regulations; tax, criminal, and administrative law; international treaties on trade, and safety at sea. To enable reuse of these ontologies in a library and to abstract the common legal denominators of these legal domains, much effort has been spent in finding a unifying view on legal domains. In this paper 1 LRI is an acronym for Laboratorium voor Rechtsinformatica (legal informatics) ; the name is now changed to: computational jurisprudence, indicating more precisely that the research focusses on legal knowledge representation and artificial legal reasoning and information management. 1

2 we will discuss two proposals for a legal core ontology: FOLaw [Valente et al., 1999] and more recently: LRI-Core. FOLaw has proven to be a good modeling support for building legal knowledge systems, because it reflects an understanding of types of knowledge and dependencies in legal reasoning. However, when applied to ontology based management of legal information services, the support is very limited: in Section 5 we will explain why and what views have lead us to develop this new LRI-Core ontology. An ontology describes how some domain is committed to a particular view: not so much by the collection of the terms involved but in particular by the way these terms are structured and defined. This structure tells us what a domain is about. It does not come as a surprise that for instance, medical domains are about malfunctions. These malfunctions are often diseases, i.e. processes; they are classified in (multiple) taxonomies, and associated with sets of typical symptoms, and with treatments ( [Clancey, 1985, Patil, 1988]; more recently: [Rector, 2002]). In the next section (Section 2) we will make such an analysis for legal domains. The article is further structured as follows. We will give a short overview of ontological assumptions that can be found in legal theory. Next we will introduce our functional ontology of law and see that it is rather an epistemic framework than a detailed ontology. From there on we will explore what is in a true ontology of law; what are typical and unique elements of law? When we have identified them, we will see them put to use in several applied R&D projects. Finally, we will discuss the current state of our work and draw some conclusions. 2 What is law about? An ontology makes explicit the views one is committed to in modeling a domain. Modeling is taken here in the broad sense that includes the notion of understanding. A major and typical problem from jurisprudence (legal theory) occurs already in the use of the term law. In the title of this section we avoided any commitment to whether we mean the law or laws. Indeed, the problem of what counts as the unit of law is already one of the fundamental ones questions in jurisprudence and is called the individuation problem: Classifying laws in logically distinct categories has always been one of the major tasks of legal philosophy... The classification of laws presupposes a solution to the more fundamental problem of the individuation of laws, i.e., an answer to the question What is to count as one complete law? [Raz, 1972, page 825] There are two extreme views. The first one takes all legally valid statements in legal sources (legislation, precedence law, etc) as a whole: the law. The assumption is that in principle the individual statements in this whole are or should be coherently organized. This is the predominant view in jurisprudence and legal philosophy (see also Section 3). Whether this coherence is an actual concern for the legal system (i.e., the law should be the object of proper knowledge management), or whether it is genetically built-in by the constraints provided by fundamental, natural legal principles, is a long and classical debate in legal theory. As [Van Der Velden, 1992] points out, the latter view takes the notion of coherence beyond what he sees as linguistic or semantic coherence. It is this kind of coherence we are concerned with here. However, the other extreme takes all legally valid statements as being individual laws. In extremo this view is incorrect, if only because it presupposes some legally valid statement that covers the legal validity of an individual statement. This view is not a view that is shared with jurisprudence. Jurisprudence is in the first place concerned with justifying law, so legal scholars will not easily take validity statements in law as a side issue (see e.g. [Kelsen, 1991]). However, in legal knowledge engineering this alternative perspective is a far more fruitful one. The validity problem is presumed and the emphasis is on the coherent modeling of actual law. The coherence is not to be found in the collection of legal sources themselves but in the worlds (domains) where the statements in these 2

3 sources refer to. These statements are normative statements about behaviour. They qualify some kinds of situations as disallowed. 2 The collection of normative statements is not aimed at describing all possible situations in a domain, but only those that have normative relevance. What is possible is assumed to be known (or to be found out) by the agents to which the law is addressed. This means that the collection of individual statements about some legal domain does not provide a full description of the domain, neither that coherence is to be found in the legal statements. The coherence has to be found in modelling the possible behaviours in the domain by reconstructing what is assumed by the legal statements. In the next section we will illustrate this principle for the domain of traffic. 2.1 Distinguishing normative and world knowledge: TRACS Can you develop a computer program that can check if the new traffic regulation (RVV-90) are consistent and complete? This apparently innocent question, posed by a government agency concerned with traffic safety, SWOV 3 triggered a decade of research at our institute. The question is highly similar to the verification of software. In fact, [Kowalski, 1990] has argued that drafting legislation is similar to programming in Prolog. In one respect he is mistaken: the nature of a (Prolog) program is very different from a regulation, as can be easily assessed by some close reading of a regulation. Different from a normal text or a computer program, the individual articles have no internal coherence. Their coherence is only the result of understanding to what kind of situation(s) each statement refers. The individual articles in the RVV-90 all refer to situations in traffic, but there is no discourse structure that connects the statements. Therefore, it will be hard to assess whether a regulation is complete. Complete may mean two things: whether there are no gaps and whether it covers the traffic domain completely. If we take the traffic regulation as a normal text, then the gaps are obvious. The traffic regulation is aimed at fostering safety and optimizing traffic flow. Although the physical conditions of roads and cars and the psychological states of the drivers are the major factors to control safety, the normal way to prescribe or advice behaviour is by the use of guidelines: coherent texts that describe safe conduct. In fact, the view of the SWOV was that the RVV-90 should be rewritten or paraphrased in such a way that it could line out procedures of safe conduct as for instance used in teaching children about traffic behaviour. Similarly, one will not find in administrative law, for instance, a procedural description of what steps one must take to appeal to decisions. This is not (only) due to the very conventional (almost text-fetishist) way the code for drafting codes regulates this format, but also because guidelines would not convey the true nature of law: by restricting possible behaviour by sanction. Although the gaps are textual ones, one may assume that any agent, familiar with the domain, can easily recognize the restricted situations. So, we should use (non-legal) domain experts to help us to reconstruct this assumed world. However, this does not complete the picture. The legal drafter may not have had in mind a complete or completely detailed view on the domain. He may have had some model of the domain in mind (hopefully a consistent and valid one). However, this model is not explicit. It perspires by the use of special, legal definitions and by abstractions implicit in the situation descriptions that go with the norms. For instance, traffic codes contain qualitative abstractions of the physical world of roads. There is a limited number of speeds that are relevant; the angle of crossings is of no importance and visibility, etc, is assumed. In summary, we can expect gaps: intended by the nature of law, by the way norms are expressed and by the abstractions taken by the legal drafter in modeling a domain. However, these gaps may be filled by reconstructing this model, which is what we do to some extend in understanding some regulation by reading it. However, there is no formal way to establish whether a regulation fits the scope of the domain. For instance, the RVV-90 does not talk about children, while they are the most endangered species in the traffic jungle. The question is whether these kind of gaps are accidental (and non-intended), or due to the limited role of law, or whether there simply is no real gap. One may argue that children have insufficient control or understanding to obey the law. Also: they are subsumed (implied) in the RVV-90 under pedestrian, etc. There are indirect ways to establish this covering or completeness. 2 Allowed is usually the default of all other situations. Note that we use here a binary qualifier, where usually the classical three deontic operators P, O and F are used. In Section 4 we will give a detailed explanation. 3 Stichting Wetenschappelijk Onderzoek Verkeersveiligheid; Foundation for Research on Traffic Safety. 3

4 The first one is to use constraints (requirements, goals). One of the aims of the RVV-90 is to avoid collisions, i.e., if there is a collision, at least one participant should have trespassed at least one norm. The other one is to compare the same situation under two different codifications. For instance, the legislator has indicated in which respects the RVV-90 is aimed to be different (an improvement) from the previous version. At the end of this section we will elaborate on this issue. If it is difficult to assess the completeness of a regulation, consistency is another problem, as can be illustrated by the first normative statements in the RVV-90: Art 3.1 Vehicles should keep as much as possible to the right Art 3.2 Two bicyclists may ride next to each other Article 3.1 describes an obligation and 3.2 a permission to a subset of drivers: the drivers of bicycles. In fact, it is not so easy to see how 3.2 is an exception to 3.1. It requires some complex spatial reasoning ( right, next to ) to see that the left-hand bicyclist violates article 3.1, because she leaves the right bicyclist between her and the side of the lane. 4. This exception is a normative conflict and therefore a logical inconsistency. This inconsistency can be repaired by applying meta-rules, such as the principles that provide priority to the more recent, the higher and the more specific norm. 5 Exceptions are intended. They are used to limit overgeneralizations of more generic norms, in the same way as we still want to classify penguins under birds despite the fact that they lack one of the essential characteristics of birds: that they can fly. 6 As we will see in Section there is a way to get rid of normative exceptions without affecting the normative qualifications intended by the code. We may be able to handle exceptions as not really inconsistencies in law, but the problem remains that we cannot distinguish on formal grounds between intended and non-intended exceptions. In fact in the TRACS project [Haan & Winkels, 1994, Breuker et al., 2000] we have found out that nonintended exceptions are often hidden in the implicit normative structure of a regulation. We should add that normative, or conflicts of disaffirmation are not the only kinds of inconsistencies that may occur in regulations; [Lindahl, 1992] also distinguishes compliance conflicts. These conflicts are between mandatory norms (obligations and prohibitions) which are not jointly realizable. For instance, a bus driver may be obliged to keep the time-table, but at the same time speed-limits may prohibit him to be able to comply with the time-table norm. This all is to say that in legal knowledge engineering one rather sees the law as a set of unrelated normative statements about some world of social activities and it is this world that is in the first place the objective of modeling. This has strong implications for the construction of legal (core) ontologies, because it means that such a core ontology would either be almost empty, or would rather cover common concepts in social worlds. This is the key question that will be pursued through the major part of this article. However, the next section contains a short description of how these insights have guided us to build a system for testing (traffic) regulations Verifying regulations: the TRACS system The basic idea in testing a regulation is to submit (all possible) situations that may occur in some domain to a regulation applier which can draw the conclusion whether the situation contains one or 4 In fact, the exception is even more subtle because the left-hand bicyclist is also held to keep still as much as possible to the right, i.e., the exception is not a licence for this bicyclist to take any position in the lane next to the right-hand bicyclist. What about three bicyclists riding next to each other? (For the (legal) answer see the last footnote of this paper). 5 The various meta principles such as Lex specialis derogat legis generalis, etc. Note that the recency principle establishes the currently valid law. 6 This is the typical example to illustrate the need for non-monotonic reasoning. However, this analogy does not necessarily imply that normative reasoning with exceptions is really non-monotonic. Of course the conflict resolution implied by the leges speciales suggests so, but a closer look at what the conflict resolution means is that there is no retraction of beliefs of what is the case in the world. The case remains unchanged. It is only that there are conflicting normative conclusions. The typical view in jurisprudence and AI & Law is that the conflict is resolved by putting priorities to applicable, conflicting norms: not to withdraw norms or facts about the case. Article 3.a is still valid for the left-hand bicyclist: she still has to keep as much as possible to the right. Despite this observation, most researchers assume that normative reasoning is non-monotonic. 4

5 Legal Source KB World Model (e.g. traffic) KB Regulation Applier (prod. system) situation description applicable legal norms (NL-) Dialogue User Figure 1: Architecture of TRACS for testing regulations [den Haan, 1996] more violations of norms. For the RVV-90 the regulation applier plays the role of a traffic policeman. The next step is to assess whether these outcomes are in accordance with the goals or requirements of the legislator, for instance by comparing these with earlier or other (draft) regulations. As in general the number of possible situations is far too large to assess the outcomes by hand. Some automatic pre-processing is required that focusses on interesting subsets: e.g. those situations for which various regulations give different outcomes. One may ask the question whether there is not a more direct way to test a regulation: for instance by comparing the normative statements from one regulation with those of another one. We will come to this question in Section Fig 1 presents an overview of the TRACS architecture (see also [Haan, 1996]). There are two knowledge bases, corresponding to the distinction between normative and world knowledge. The WORLD MODEL represents the domain of law, i.c. the traffic world, while the LEGAL SOURCE contains the normative statements. These normative statements are represented using terms from the WORLD MODEL, i.e. one may say that both knowledge bases share the same ontology. The upper part of this ontology is presented in Figure 2. The core concepts in traffic are actions that make or change movements with vehicles. Movements are made by traffic participants: mostly drivers using various kinds of vehicles. Movements are changes of positions and positions may be defined as some part of the road, or relative as an orientation with respect to other participants. These changes are qualitative ones, i.e., the distances and directions are qualitative; not quantitative ones. If quantities occur in the RVV-90, e.g., as maximum speeds, they are treated in the same way as landmarks in qualitative physics. Besides the focussing on the actions of drivers and the qualitative abstraction of space, the specific views of the legislator on the meaning of terms are expressed by definitional articles. The first two articles of the RVV-90 contain terms and definitions for traffic participants, vehicles and lanes. In general, the definitions given by the legislator are restrictions on the common sense use of terms. 7 Besides articles that have a normative sense, one will find articles that contain definitions. The distinction between definitional and normative knowledge in legal reasoning is very important [Kowalski & Sergot, 1985, Breuker & Den Haan, 1991, Valente, 1995]. In TRACS this definitional knowledge (ontology) is implemented in TRACS as part of the T-Box of the LOOM knowledge representation system [MacGregor & Bates, 1987]. In principle the working of this architecture is very simple. The situation generator contains a 7 This is in general the case when in legislation definitions are found. However, for reasons of (political) prudence, the legislator may also make the meaning of a term less specific, leaving it to the interpretation of courts to see how that works out in specific cases. 5

6 TRAFFIC_TERM PHYSICAL_OBJECT vehicle motor_vehicle car passenger_car bicycle road part_of_road lane tram_lane drive_lane street_refuge round_about crossing path bicycle_path traffic_sign (M) SPACE SPATIAL_ORIENTATION TRANSVERSAL forward backward ORTHOGONAL left right SPATIAL_POSITION next ROLE driver pedestrian CHANGE_OF_STATE PROCESS MOVING ACTION driving (controlling_moving) changing_direction turning_left reversing (M) changing_speed driving_off accelerating stopping composite_actions parking overtaking special_manoeuvres reversing (M) signalling MENTAL_OBJECT SIGN (M) traffic_sign (M) regulation (M) traffic_regulation RVV-90 Figure 2: Excerpt of the ontology of traffic terms in, and for, RVV-90; upper case terms are not in RVV-90; (M) stands for multiple classification 6

7 Figure 3: Example of output of TRACS: the tram should not ride on the tramway [den Haan, 96] kind of grammar to compose traffic situations based on the ontology of the RVV-90. This grammar contains in the first place the rules for configuring legal roads from parts of road. For each kind of road, different kinds of positions for different kinds of traffic participants (vehicles) are generated. In principle this way of generating is almost unmitigated combinatorial. If we do it this way, the RVV-90 implies more than 2.5 billion different situations [Haan & Breuker, 1996]. However, quite a number of these situations are redundant, due to symmetry (e.g. at crossings, reciprocal spatial relationships, no-influence (parked car, etc)). An important constraint is the fact that no article in the RVV-90 implies some history. All rules apply to a kind of situation or situation change. The policeman does not have to follow traffic participants to assess violations. This reduces the total number of situations to about a hundred thousand; still a very large number. We do not know what the exact number is. The situation generator was never completed. When we tried out the prototype, we found problems for almost twenty percent of the hand fed situations (about eighty). Problems are outcomes which are obviously not intended. For instance, Figure 3 is a screen-dump from TRACS. It says that the tram is not allowed to run on the tram-lane. This is due to the fact that Article 10.1 states that all vehicles except those mentioned in Articles 5 to 8 should use the drive-lanes. However, the tram is not mentioned in these articles, and the tram-lane is not a drive-lane Compiling out normative structures in law Although the prototype demonstration of TRACS was convincing to the ministry, due to changes in policy no operational system was built. In the meantime we were concerned with the question whether it was possible to compare regulations in the same domain in a more direct way, given that the world ontologies could be easily unified. For instance, it appeared that the traffic code of Minnesota had a largely similar ontology to the Dutch RVV-90. Minor differences were found in parts of road and in vehicle classification. However, the normative structures look very different. Under the perspective of lex specialis normative structures are subsumption lattices of the generic situation descriptions of the normative statements in a regulation. As stated above, a norm is to be viewed as an abstract, but partial description of a situation about which a deontic qualifier ( operator ) indicates whether the situation is forbidden (F), obliged (O) or permitted (P). Permissions are invariably exceptions to mandatory norms (O, F, see [Lindahl, 1992, Kelsen, 1991]). Exceptions are necessary to correct 7

8 overgeneralized mandatory norms. In this view, the meaning of permissions is context dependent. 8 Therefore, if we remove all exceptions by correcting the overgeneralizations, we obtain a normative structure containing only mandatory deontic qualifiers, i.e. we can go from a three (or rather two-anda-halve valued) to a two valued deontics, which removes the need for conflict resolution and simplifies deontic reasoning to tractable algorithms [Valente et al., 1999, Winkels & Haan, 1995]. We call this the normative compiling-out of a regulation. A compiled-out regulation contains far more and also more complex normative statements as one may easily see when we compile out Article 3 of the RVV-90: 9 Art x.1 All vehicles are obliged to keep to the right as much as possible Art x.2 For all motor-vehicles, horses, horses-and-carriage,..., it is forbidden to ride next to each other Art x.3 It is forbidden for more than two bicycles to ride next to each other The complexity in terms of number of situational propositions per norm increases drastically if we would also take into account the exceptions on the position of the road for overtaking, two lane traffic, etc. It is obvious that a compiled out normative KB would make the reasoning much faster. Although a compiled out version and the original version of a regulation are semantic paraphrases, the compiled-out version is no longer isomorphic, which is a problem in justifying the conclusions in terms of the officially legal, original version [Bench-Capon & Coenen, 1991]. Here we are interested in the question whether there is a more direct way to compare two regulations. Indeed, if the ontologies are aligned, compiled-out regulations may be easily compared by sorting the situation descriptions by same or matching propositions. However, a closer inspection of what compiling-out entails shows that it is the analytic counterpart of (syntatic) situation generation. In compiling-out, like in situation generation, situation descriptions should include the most specific levels that the domain ontology contains. It can be shown that the main difference between the two approaches is that the situation generator makes also explicit all situations that are implied by the ontology, but have no deontic qualification: the gaps in a regulation. This complement of the compiled-out regulation is not difficult to add anyway. One may attribute a default normative value to this complement about which the law is silent [Valente et al., 1999]. This value is usually defined as what is not forbidden is allowed. 10 In summary, our experiences in and reflections about the TRACS project have made us aware that an ontology which describes the concepts in a world to which a regulation applies is pivotal in legal reasoning. However, the world ontology contains concepts which are mostly not typical for law, but are largely common-sense, or at least a large amount of common-sense concepts are assumed. Even if a concept in a world ontology is a legal one, or one refined by legislators, then it is not a high level generic concept of law. This makes the construction of a foundational or core ontology in a bottom-up way rather implausible. A foundational or core ontology is a high level ontology aimed at organizing domain ontologies, for instance in a library of domain ontologies. Another, probably as important role, is its support in analyzing new legal domains. In a top-down analysis of concepts in law, the first insights may be found in the literature on jurisprudence and legal philosophy. 8 Contrary to the view in (standard) deontic logics, where permission is a first class term. 9 Close reading of the original article 3 of the RVV along with the intended meaning means that only bicycles may ride in pairs, but also this pair has to keep to the right as much as possible. In fact, it shows that Art.3.b of the RVV is not a direct exception to Art.3.a (as the legislator wants us to believe), but to an assumed view that the left-hand bicyclist is not completely to the right. However, the exception is in fact to an implicit prohibition: that no vehicles should ride next to one-another. 10 For domains for which rules are made to enhance safety the default is usually the opposite. Nothing is allowed unless prescribed (obligatory). There is a subtle difference between explicit permissions in regulations and this default allowance [Valente et al., 1999]. See also below (Section 3 on this difference between strong and weak permissions according to [Kelsen, 1991] 8

9 3 Ontological Assumptions in Legal Theory Legal theories usually contain elements of an ontology (ontological assumptions), but they are normally framed under some specific theoretical goal which lies beyond the ontology itself. For instance, Hart s theory intends to explain how legal systems evolve; Kelsen s goal was to demonstrate the difference between laws and morals; yet both propose very specific views on what competence is behind legal phenomena, and what primitive concepts are used to represent law. To extract the ontological view proposed by these authors is therefore a difficult task, and one which we do not intend to pursue in depth. Nevertheless, we present below a brief account of the ontological views contained in the works of some major legal theorists and philosophers. We will emphasize the primitive concepts they have proposed to represent legal knowledge, which can be directly translated into an ontology. Kelsen In his last work ( General Theory of Norms [Kelsen, 1991]), Kelsen proposed four basic types of norms: command, empower, permit and derogate. Commanding norms command (prohibit, obligate) a certain behavior. Empowering norms give to some individuals the power to posit and apply norms under certain restrictions. Permitting norms refer to what he called the positive sense of permission. Kelsen argued that we may permit behavior in the sense that this behavior is neither prohibited nor commanded, in which case we have a negative (sometimes called weak) permission. In contrast, permitting norms use a positive sense of permission, in which behavior is actively allowed. Kelsen sees permitting norms as always an exception to a command, i.e. it occurs when there is a commanding norm about a certain behavior, and this command is then derogated (and thus allowed) by the permitting norm. This restriction to permitting norms, in our view, is not a necessary one: it may be useful to permit a certain behavior in order to prevent that other (lower, unspecified) commanding norms impose restrictions on this behavior. Kelsen points out that the difference between commanding and prohibiting on the one hand, and permitting (positively) and empowering on the other, can be mapped onto the distinction between observing/violating and applying a norm. Only a norm which commands (or prohibits) a certain behavior can be observed or violated. Norms that permit or empower cannot be violated or observed, only applied or not. Finally, derogating norms repeal of the validity of [...] another norm. Kelsen stresses that a derogating norm does not repeal another norm, but its validity: a derogated norm still exists, but it is no longer valid. Derogation can occur in two different situations: when a conflict exists between norms, and independently of such conflict. In the first case, a certain norm enters in conflict with another and is derogated, meaning that it loses the winning norm is to be applied, while the derogated one is not. In the second case, the norm-positing authority sees a norm to be undesirable and wants to repeal it, and does so by means of creating a derogating norm. Hart The Hartian distinction between primary and secondary rules (norms) has become a quasistandard in legal theory. Hart s distinction, carefully detailed in his Concept of Law [Hart, 1961], draws a line between a first level which refers to human behaviour and a second, meta-level of the first, which contains knowledge about primary norms. These secondary rules may belong to three types: (i) rules of adjudication, that can be used to determine authoritatively whether a certain primary rule has been violated or not; (ii) rules of recognition which define, directly or indirectly, which rules are the valid ones, and can therefore be applied; (iii) rules of change, which define how rules are to be made, removed or changed. These distinctions point out three functions of secondary norms: to provide support for solving conflicts (adjudication), to specify the limits of the legal system (recognition) and to specify how the legal system can change in time (change). Bentham Bentham s theory is divided in two parts [Bentham, 1970]. The first is a logic of imperation which uses four basic operators: commanded, prohibited, non-commanded and permitted. 11 These are in fact interdefined, resulting in only one of the four as primitive. The second part is a logic of obligations and rights in which he defines three primitive concepts: obligation, right to a service 11 A logic of imperation an idea also mentioned by Austin [Austin, 1954] was later developed in more detail by Hofstadter [Hofstadter & McKinsey, 1939], but it is presently considered to be superseded by deontic logics. 9

10 and liberty. These are also interdefined based on obligation, which Bentham sees as an obligation someone has to the effect that something (some state of affairs) occurs. Therefore, we are left with only two atomic concepts: commanded and obligation. Hohfeld Hohfeld s theory is considered a landmark in American jurisprudence [Hohfeld, 1919]. An interesting (and unusual) aspect of Hohfeld s theory is that rights and other positional concepts that represent legal relations are considered primitives. There are two groups of interrelated legal relations or positions. The first group is composed by right, duty, no-right, privilege and has a strong normative flavour. These concepts are closely related to Bentham s concept of right, obligation and liberty. The second group consists of power, liability, disability immunity. These concepts are more closely related to legal competences and legal responsibilities. 3.1 Legal Theory in AI and Law There are a number of studies in AI & Law which have used ontological assumptions drawn from legal theory in the manner we propose in this paper. Allen and Saxon (e.g. [Allen & Saxon, 1991, Allen, 1997]) have developed a language for legal relations (LLR) in which they have transformed the Hohfeldian ontological primitives into about forty relations ( cascading propositions ), taking the notion of duty as primitive. Hamfelt [Hamfelt & Barklund, 1990] proposed and implemented a representation of legal knowledge in which Hart s primary and secondary rules were mirrored in meta-levels of a logic programming formalism. [Valente, 1995] has formalized normative legal knowledge mainly on the basis of legal theoretical concepts from Kelsen and Hart. This is also the perspective on legal normative knowledge we take here. There is also work which has an ontological flavour, but which has not been based on legal theory. For example, McCarty s Language of Legal Discourse [McCarty, 1989] can be seen as an ontology of Law, where his modalities play the role of knowledge categories and are linked together with a formal (logical) presentation. Also, the research in deontic logics as a basis for normative reasoning sometimes uses ontological assumptions from or is applied to legal theory see for instance [Alchourrón & Bulygin, 1971, Alchourrón & Bulygin, 1981]. Kelsen s view on norms as descriptions of an ideal world can be seen the basis of deontic logic. Deontic logic provides interpretations for the terms obligation, prohibition and permission. Indeed some form of deontic logic is often proposed as a formalism for automated normative reasoning. Because (standard) deontic logic is intractable, and gives rise to pseudo-paradoxes, all kind of extensions and simplifications have been developed. In fact, [Valente, 1995] has shown that deontic logic does not make the necessary distinction between the normative status of a situation and the normative operator of a norm. By making this distinction a much simpler and tractable inference mechanism has been formally defined and implemented, as we discussed in the previous section. Van Kralingen ([vankralingen, 1995]) uses the theory of Brouwer ([Brouwer, 1990]) as a starting point for what he sees as the core elements of law: norms, actions and (legal) concepts. 4 Knowledge typing and dependencies in legal reasoning: ON-LINE and FOLaw Although the combination of world and normative knowledge make up the reasoning and representation of reasoning in relatively concrete legal domains, this is only part of the story. In the mid-90ies, Andre Valente constructed a core ontology that distinguished also other types of knowledge [Valente, 1995, Valente et al., 1999]. This core-ontology, called FOLaw, served a number of purposes. The first one was to distinguish the various types of knowledge in legal reasoning, and in particular those types that are typical for legal reasoning. Related to this role it also explained the dependencies between these types of knowledge in legal reasoning. This typing and its dependencies could easily be translated into an architecture for legal reasoning: ON-LINE. The second role is 10

11 META-LEGAL KNOWLEDGE legally qualified behaviour description agent(s) that caused the behaviour NORMATIVE KNOWLEDGE RESPONSIBILITY KNOWLEDGE classified behaviour (allowed/disallowed) agent(s) that are responsible for the behaviour REACTIVE KNOWLEDGE CAUSAL KNOWLEDGE WORLD KNOWLEDGE (Legal Abstract Model) CREATIVE KNOWLEDGE legal reaction DEFINITIONAL KNOWLEDGE abstract legal agency LEGAL SYSTEM COMMONSENSE KNOWLEDGE social behaviour (states of affairs, stories) UNIVERSITY OF AMSTERDAM MINISTERY OF EDUCATION SOCIETY PHILIPS JOOST BREUKER Figure 4: FOLaw: a functional ontology for law [Valente, 1995] the typical core-ontology role: to organize and index libraries of domain ontologies and to support knowledge acquisition to construct new ontologies. The ontology is a functional ontology. This means that the roles the legal system plays as part of, and to some extent controlling, society are taken as point of departure. Although the role of legal sources within this legal system and its functions is emphasized, a legal-sociological view is taken rather than a perspective from (the) law itself, as in most legal theoretical literature. There is a secondary notion of functional involved: the dependencies between the types of knowledge indicate the roles that types of knowledge play in the reasoning. These two views on functional are not independent. One may see the reasoning as to some extent simulating, or re-ifying, the social roles, in the same way as reasoning about physical systems consists to a large extent of simulating physical processes. We will give here a summary description of FOLaw. Figure 4 provides the comprehensive picture of dependencies of the various types of knowledge in legal reasoning. At the same time it also expresses the role of the legal system as controlling the actual social behaviour of individuals and organizations in society. 4.1 FOLaw s types of knowledge in legal reasoning The major types of knowledge we have distinguished are normative knowledge, world knowledge, responsibility knowledge, reactive knowledge, creative knowledge and meta-legal knowledge Normative Knowledge Normative knowledge is the most typical category of legal knowledge, to such an extent that to many authors normative and legal are practically the same thing. Normative knowledge is seen in the legal theory literature as having two functions: prescribing behaviour and 11

12 defining a standard of comparison for the social reality. The basic conception of norm used in the ontology is largely derived from [Kelsen, 1991]. A norm expresses an idealization: what ought to be the case (or to happen), according to the will of the agent that created the norm. Meta-legal Knowledge Normative systems are defined on the basis of individual norms: e.g. individual articles in a regulation. The difference between the standard defined by the normative system on the one hand and the standards defined by the norms on the other hand is fundamental to understanding the role of normative knowledge in law, and it is accounted for by knowledge about individual norms. This distinction is captured by defining the categories of primary norms and meta-legal knowledge. Primary norms are entities that refer to human behaviour, and give it a normative status. This normative status is, in principle, either allowed (legal, desirable, permitted) or disallowed (illegal, undesirable, prohibited). However, each norm refers only to a few types of behaviour, in the sense that it is only able to provide a status if applied to some types of cases. For the remaining types of cases, the norm is said to be silent. There may be a difference between the normative status given by a single norm and the one given by the normative system. Individual norms may conflict: if that is intended then some norm may be an exception to another norm. In order to solve these normative conflicts, meta-legal knowledge is applied. Typical conflict resolution is provided by meta-legal rules that state for instance that the more specific rule should be applied rather than a more general one: lex specialis derogat legis generalis expresses this age old wisdom in law. Meta-legal knowledge is not only used for solving conflicts between norms. Another function is to specify which legal knowledge is valid. Validity is a concept which can be used both for specifying the dynamics of the legal system and its limits. A valid norm is one that belongs to the legal system. World Knowledge By its very nature, law deals with behaviour in the world. Therefore, it must contain some description of this behaviour. For instance, in order to describe how the world should (ought to) be, norms must describe how things are or can be. In addition to adopting a category of legal knowledge which describes the world, we propose that this knowledge constitutes a structured model of some domain of law. For instance in a traffic act, the world of traffic is assumed to operate in a certain way, and the legislator may pose constraints norms on this behaviour. A car may drive on all sides of a road, but the legislator (in most countries) obliges us to take the right hand side. Implicitly, the legislator has some model in mind of how traffic operates and can operate. The behaviour in this world has to be modeled. This model is a generic model: how things (may) work, are done, or may be done in general if there are no normative limitations. In principle, the legislator has to foresee all possible types of situations and label these as allowed (desirable) or not. Thus, the term legal abstract model or LAM is used as a synonym for world knowledge when its model character is to be stressed. The legal abstract model is an interface between the real world and the legal world. Its role is to define a model of the real world which is used as a basis to express normative and other categories of legal knowledge. The bulk of the LAM consists of definitions of concepts that represent entities and relations in the world. Apart from describing the world, what is behavioral reasoning used for in law? We propose that this description of possible and relevant behaviours is built around the concept of cause, in order to allow the assignment of responsibility of an agent for a certain case. Causal knowledge, however, refers or uses a static description of the world (e.g. to model world states). Accordingly, we propose that the world model is actually composed of two related types of knowledge: definitional knowledge and causal knowledge. The definitional knowledge is used by the normative knowledge to describe the ideal world they define. The causal knowledge is used by the responsibility knowledge to describe who or what have caused a given state of affairs, and can thus be considered responsible for it. Responsibility Knowledge It is commonplace to state that cause and responsibility are important concepts in law, but it is not equally simple to provide a more concrete account of how these concepts 12

13 enter the realm of law. We see responsibility knowledge as a category of legal knowledge that has as a function to assign or limit the responsibility of an agent over a given (disallowed) state of affairs i.e. to (dis)establish a link between the violation of a norm and an agent which is to be considered responsible (accountable, guilty, liable) for this violation. This responsibility link may be established by a causal connection between the agent and the disallowed behaviour, but this is not the only way to establish responsibility. Responsibility is the intermediary concept between normative and reactive knowledge, since a reaction can only occur if the agent is held responsible for a certain norm violation. Responsibility knowledge plays the role of linking causal connections with a responsibility connection i.e. that connection which makes an agent accountable for a norm violation and possibly subject to legal reactions (see also Section 4.1). As [Hart & Honore, 1985] point out, however, responsibility does not have any implication as to the type of factual connection between the person held responsible and the harm that is, causal connections are only a non-tautologous ground or reason for saying that [an agent] is responsible [Hart & Honore, 1985, pag. 66]. In principle, commonsense says that one is only responsible for what one causes. Naturally, there is an a priori link between these two relations: in principle, all agents are legally responsible (and only responsible) for that which they cause. But this causal connection is not always necessary or always sufficient for establishing responsibility in a legal context. The role of responsibility knowledge is exactly to interfere with this prima facie connection between causing and being responsible. This interference is made so that legal systems extend responsibility [or] cut it off in ways which diverge from the simpler principles of moral blame [Hart & Honore, 1985, pag. 67]. This mechanism has rather practical motives. Given the innumerable problems in establishing, proving and reasoning with causal connections, the assignment of legal responsibilities which bypasses these connections to some extent is used to give more precision in situations where the use of the commonsense or moral concept of responsibility by the law can lead to inconsistencies or undesired results, or when there is a practical interest (based on an implicit of explicit policy) that a frontier should be drawn so that it becomes easier to define what are the limits of responsibility under certain circumstances. There are two basic mechanisms which are used in responsibility knowledge. First, the law may establish a responsibility connection independent of a causal connection i.e. a responsibility assignment. This can be seen in a rule used in e.g. French, German or Brazilian law, by which parents are held responsible for the the damage done by their children even if there is no specific causal link between their attitudes or actions and the damage. That is, the parents are held responsible even though they have not necessarily caused the damage. Second, the law may limit the responsibility of an agent under certain circumstances, disregarding some possible causal connections i.e. a responsibility restriction. For instance, in England a man is not guilty of murder if the victim dies more than one year after the attack, even if the death was a consequence of this attack. Other well-known factors that may influence the establishment of responsibility connections in law are knowledge and intention. We refer here further to the work of [Lehmann, 2003] who has worked out the relation between (physical and agent) causation and the various notions of responsibility of law; Lehmann has specified this in a foundational ontology that defines the notions of causality and causation. Reactive Knowledge To reach the conclusion that a certain situation is illegal (based on normative knowledge), and that there is some agent to blame for it (responsibility knowledge) would be probably useless if the legal system could not react towards this agent. That knowledge that specifies which reaction should be taken and how is what we call reactive knowledge. Usually this reaction is a sanction, but in some situations it may be a reward. The penal codes, which are usually a fundamental part of legal systems of the Romano-Germanic tradition, contain basically responsibility and reactive knowledge only. Creative Knowledge A legislator may indirectly create some entity that did not exist before in the world, using what we call creative knowledge. It is usually stated in imperative terms, designating an agency that previously did not exist as part (or not) of the reality from a certain point of the time on. 13

14 The creative function has a somewhat exceptional (or even abnormal) status if compared to the other ones. In this case, the law not only wants to classify or to react over certain agents that already exist in the real world, but attempts to create a new agent. 4.2 In search of ontological foundations of law We have used the framework of FOLaw as a lead for fundamental research [Lehmann & Breuker, 2001, Lehmann, 2003], and as the basis for practical applications and architectures for legal reasoning (e.g., ON-LINE [Valente et al., 1999]). The CLIME project 12 was aimed at the construction of a legal information server. The try-out domain were international rules for safety and environmental care at sea, and the rules for ship classification (certification): in total about 15,000 different articles. This CLIME information server has two modes of operation. The first mode is typical information retrieval, where keywords (in phrases) are matched against terms in the rules. A large ontology (over 3,500 concepts) allows the elaboration of the keyword-terms by implied terms. The second, more expensive and experimental mode is in fact a question answering one. The CLIME system assesses whether a case, e.g. results of the inspection of a ship, or legal questions during the design of a ship, complies with the rules or not. The applicable (violated or potentially violated) articles provide the justification and focus for the answer. An overview of CLIME and an evaluation of its results can be found in [Winkels et al., 2002]. Other applications of the FOLaw framework (annex architecture) are reported in [Muntjewerff & Breuker, 2001] (PROSA, a training system for solving legal cases); in the KDE project 13 the ontologies of CLIME have been re-used [Jansweijer et al., 2000]. These results about the use and re-use of FOLaw also show its limitations. In developing the legal domain ontologies it turns out that the major effort in modeling is in the world knowledge. This is not surprising, given that the initial analysis about the content of legislation as in the TRACS process already revealed that world knowledge is the driving force in legal reasoning systems. More theoretically, it appears that law does not have its own ontological foundation. When legal philosophers discuss the ontological assumptions in law and legal reasoning, it is invariably about normative knowledge, and sometimes also about responsibility knowledge [Hart & Honore, 1985]. This is different from other knowledge based fields of practice like medicine [Van Heijst et al., 1997] or engineering [Borst et al., 1997], which have abstract ontological foundations in notions about physics, mathematics, etc. Jurisprudence and legal philosophy are primarily concerned with the justification of law and legal systems, rather than the explanation of the working of law and its relation to social reality. This is not to blame jurisprudence. The explanation of social reality is the concern of sociology, political and management science, etc. Explanations are models and these are grounded in ontological commitments. However, justification which is derived from the term ius (law) is the domain of epistemology; the study of what we can know and believe. Epistemology is about reasoning, argument and evidence, while ontology is concerned with modelling and explaining the world. Therefore, it is no surprise to see that core ontologies about law are rather epistemic frameworks [Van Kralingen et al., 1999, Hage & Verheij, 1999]. [Lehmann, 2003] and [Mommers, 2002], who make the same classical philosophical distinction between epistemology and ontology, construct ontologies that mix both ontological and epistemological entities. Therefore FOLaw is rather to be viewed as an epistemological framework then as a core structure of legal ontologies. It can easily be re-written as a high level CommonKADS inference structure. An inference structure consists of inference functions (e.g., abstract, compare, etc.) and roles. The knowledge types distinguished in FOLaw are static roles : input of generic knowledge to feed the inferences: so putting inference names in the actual positions of the types of knowledge (and taking the types as inputs for these functions) is sufficient to make it a formal CommonKADS inference structure: the dependencies between the types (now: inferences) are the static roles. FOLaw is therefore the inference structure of a very general problem solving method (PSM) for solving legal problems and 12 CLIME was an European project (IST 25414, ): see 13 KDE, for Knowledge worker Desktop Environment is a European IST project (IST 28678, ); see 14

15 can be added to the CommonKADS library of PSM s [Breuker & Van De Velde, 1994]. 14 This is also in accordance with the practical applications developed (re-)using FOLaw. 5 Ontologies for reuse: LRI-Core If FOLaw is not sufficiently detailed and is rather an epistemic framework, there is a need for a new approach. FOLaw has shown us two distinctive sets of concepts that are typical for law. (1) normative terms (and their definitions and axioms), and (2) responsibility terms (liability, guilt, causation, etc), which confirms what we have found in legal theory. An ontology of normative terms has been worked out by [Valente, 1995]. A foundational ontology that relates responsibility issues to agent and to physical causation has been constructed by [Lehmann, 2003]. The question is now whether the third type of knowledge that is pivotal for law world knowledge may contain still elements (concepts) that are typical, but maybe not exclusive for legal domains. For instance, law relies heavily on documents (authentication, referable evidence, etc.). Also, the law is aimed at social activities. That means that notions of role, social position, and other social relationships and activities, in particular communicative ones, play a dominant role in domains of positive law. 15 However, these typical, but not exclusive concepts for law are founded deeply in common sense. That means that for modeling and understanding some legal domain we should be able to include notions about agents, actions, processes, time, space, etc, i.e., some foundational ontology appears to be indispensable on top of a core ontology of the typical legal terms, because the concepts of law are spread over almost the full range of common sense. We could not simply start with one of the currently available upper-ontologies (e.g. [Sowa, 2000], the CYC upper ontology or in particular the IEEE-Standard Upper Ontology (SUO) that is under development ( and because their focus is rather on describing the physical and formal-mathematical world: not the social/communicative world which is more typical for law. Besides this lack of sufficient covering, we did not agree about the physical part anyway. Our analysis of causation in law has made us aware that the notion of process as a change of (physical) attributes should be separated completely from the notion of event: i.e. the things that actually occur. Events are understood as to be explained by instances of processes [Lehmann & Breuker, 2001]. Also, the kinds of processes distinguished in these ontologies are an arbitrary, heterogeneous set, ranging from change-of-state (which may be viewed as the definition of event) to coloring. Although actions (i.e. as intentionally-caused-processes in SUMO 16 ) are identified, and distinguished from processes (i.e. as natural process ), there is no other guiding principle to classify processes. In LRI-core the principle applied is the notion of what can happen to an object and the answer is in three categories: (1) it can change its properties (e.g. change of colour, form, substance (eg rusting), etc.); (2) it can change position (move), and (3) it may be created or destroyed. Moreover, SUMO does not know about force or energy, i.e. it does not consider objects to consist of matter and energy, which is necessary to explain that processes consume or produce energy. The types of objects are in fact as heterogeneous as processes but completely unrelated to one another. For instance, organic objects are distinguished from inorganic ones, but that has no relation of what can happen to these objects. In summary, SUMO appears as a too limited and a too arbitrarily structured collection of terms to guide the process of knowledge acquisition for domains that assume these terms to be known. What makes SUMO particularly unsuitable for covering legal documents is that there is no notion of document or any medium for communication. What comes closest is the concept of 14 It is curious to note that both the first author of this paper and Andre Valente have worked on the design and content of this CommonKADS PSM library, but have not noticed this close formal correspondence. 15 In this respect, the traffic domain is not so typical for law. Although the traffic code is directed towards agents (drivers, pedestrians) the aim is the control over physical movement rather than over social relationships. 16 SUMO stands for Suggested Upper Merged Ontology, i.e. a merger of a number of proposed upper ontologies. It appears that the IEEE upper ontology initiative has taken a step back by preparing first a logical framework (IFF, for Information Flow Framework) that should support SUMO [Kent, 2003]. However, it is not clear how or whether SUMO will return to this framework. 15

16 representation, but that is only viewed as a physical entity. 17. It may be the case that SUMO, as a mixture of various upper ontologies that may have been more consistent and systematic, is a bad example, despite its claim that it reflects the best combination of what has been published so far. We may see it as the latest technology supported version of an already three thousand years lasting debate. Reflection on the ontological or metaphysical basis of the world has revealed more disagreement than probably any other subject: not only in the world of philosophy and physics, but also in the recent world of ontological engineering. The discussion list of the IEEE-SUO shows a growing divergence rather than convergence over the last year. The critics on the CYC upper ontology have never stopped (e.g. [Sowa, 2000]). Why shouldn t we take the obvious lessons from these experiences? There are several reasons. The first one is that this ontology is not meant to carry the full weight of a complete and comprehensive common sense knowledge base (CYC) or to be the basic interlingua for the Semantic Web (IEEE- SUO). LRI-Core is rather meant to support knowledge acquisition for legal domain ontologies. It needs therefore an extensive covering, going from physical worlds to mental ones, but there is not the requirement that this breadth should have the same thickness : only for those concepts that are typical for law we need more dense and axiomatic definitions [Hayes, 1985]. Therefore, we think we can do with no more than about concepts, in particular those concepts that give both a large covering and can be aligned in a more or less systematic fashion, so that the debate can be about its organizational principles, rather than about individual concepts. Second, working in large committees may not be as productive as in smaller scale projects. Instead of obtaining agreement, the first aim is to experiment with this ontology in the e-court and E-POWER projects. 18 Finally, we have, of course, also learned, i.e. imported, from these experiences. However, the major reason to develop this ontology is that we really need such a support if we want to avoid making (legal) domain ontologies from scratch: if we can have some reuse from some abstract structure we may finally arrive at some convergent coherence at the highest level of domain ontologies. Of course, we can only claim this ontology to be (pragmatically) valid if experiences show this convergence. Up till now we have only experiences in applying it to criminal law, to (legal) document structures, and to tax regulations in respectively the e-court project, the MetaLex standard 19 and the E-POWER project. In fact, developing one, coherent upper ontology may be not be a feasible solution at all. The SUO work group has moved from developing the SUMO ontology directly to preparing a framework, based on category theory, to provide the semantics for the upper ontology (see This framework is a formal, articulate meta-ontology, containing by itself already about 700 terms to formally describe theories, languages, logic and models, suggesting that an upper ontology may consist of such components. SUO s intention is to anchor in the first place any upper ontology for the semantic web on solid grounds. This meta-ontology knows three levels: top, upper and lower. Whether this will provide the instruments to have a better basis to reach consensus is doubtful. It should be noted that the development of the Semantic Web is foreseen not to be dependent on agreed upon upper standards. The predominant view and the technology enablement presupposes local ontologies in OWL/RDF that are well agreed upon amongst communities and that naturally grow and get debugged by practice: the formal basis provided by OWL should be sufficient. A more interesting approach, for us, and probably also for enabling the emergence of the Semantic Web, is taken in the WonderWeb European project. Instead of a huge formal, mathematical apparatus, a simple but formal methodology is used: OntoClean [Guarino & Welty, 2002]. The foundational ontology under development is called DOLCE 20 [Gangemi et al., 2002]. In fact it is intended to be one of the foundational ontologies developed in WonderWeb. The question now is whether we could 17 It is a physical representation of a mental entity: except for icons the correspondence between the physical appearance and the mental entity it refers to is arbitrary. It should be noted that this mental entity may often also refer to some physical object or process, but we should keep all these references disentangled. Keeping track of references is however, not the job of an ontology, but the ontology should enable the right distinctions 18 e-court is a European 5th Framework IST( ) project, see E-Power is also a European 5th framework IST ( ) project, see 19 MetaLex is an open XML standard for legal documents we developed; see and Section For Descriptive Ontology for Linguistic and Cognitive Engineering. 16

17 re-use DOLCE instead of (further) developing LRI-core. In this case the answer is not so simple. Both were developed in parallel, without being aware of one-another s work. However, looking at the results now we still prefer LRI-core, but this preference is clearly debatable, and hopefully with less divergent results as the discussions on foundational ontologies so far. A major difference between LRI-core and DOLCE is in two respects. The first one is that DOLCE takes as a (the) major distinction the difference between enduring vs. perduring entities (continuants vs. occurents). This distinction is certainly the oldest in metaphysical philosophy and still its handling is one of the most hotly debated issues in the development of foundational ontologies. We do not believe in this distinction in the same way. This distinction (only?) applies to the physical world. However, whether something changes (endurance), and even loses identity, or whether something stays as it is (perdurant) is not some matter of time perspective, but of viewing the components of the physical world. The distinction between endurants and perdurants coincides with that of processes and objects. When we take a closer look at what s in an object or in a process, we see that it is some mixture of matter and energy. In other words we see the basic ontological concepts as unrelated to the time/change perspectives in the perdurant/endurant distinction, but as matter and energy distributions that give rise to (causes) processes and (changes of) objects in actual situations. 21 Moreover, we have acquired a remarkable accuracy in identifying these physical processes and objects and in predicting their consequences in actual situations. This knowledge is particularly evolved for those kind of processes that explain the most apparent and frequent changes in the environment. Apparent refers to our perceptual apparatus (sensory ranges, grain-size); frequent to importance. It explains why we have a high alertness and accuracy level for movement. Thermodynamics may come next, while our understanding of chemical and biological processes is more scattered. Indeed we live in an incredibly stable physical world. This intuitive knowledge we probably share with many animals. [Pinker, 1994] In this evolutionary view on knowledge acquisition, that may have lead to hard-wired instincts, the evolution of knowledge about the mental world may have been bootstrapped from our understanding of the physical world. [Lakoff & Núñez, 2000] show how metaphorization based on a number of instincts (which we share with primates and some other high level mammals) about the physical world may explain the origins and development of such abstract constructs as mathematics. There are many suggestions that we may see the mental world as a metaphor of the physical world. Thoughts (= object) and content (= substance) are prone to energetic processes (= motivation, emotion). Like the temperature of objects, the natural energy states of thoughts may vary largely. If emotions have a thermodynamic flavour, the cognitive processes themselves have a movement flavour (cf. the terms motivation and emotion). Storing and retrieving memories are an example; our intentional stance has direction, and communication is typically cast in terms of transferring information. This view is maybe not incompatible with the perspective taken in DOLCE; DOLCE has also a cognitive bias and takes current common sense as its objective, but it is not (yet) apparent what that means for the ontology. Mental and social categories are distinguished, and there is a physical and non-physical view on agents, but the mental world is not specified other than being non-physical (see Figure 5). 22 Therefore, at the moment DOLCE is insufficiently developed to support the mental world. If there is a difference in flavour, this is due to the fact that in DOLCE there is a linguistic and communicative view that is taken as such, while in LRI-Core an evolutionary and metaphorical perspective is taken. The major principles from which LRI-core is constructed are the following: Objects and processes are the primary entities of the physical world. In objects energy and matter are distributed, so that objects participate in processes, while processes transfer or transform energy. The participation of objects may change some of the quantities or qualities (transforma- 21 This is not a revisionary view on physics [Strawson, 1959] in the light of relativity theory where an equality relation between mass and energy is discovered. It gives an account that objects are packed with energy. A hot stone (thermo-energy) and a falling stone (kinetic energy) are examples in mind. 22 In DOLCE one may find quite a number of negated categories (which implies that the categories should be read as exclusive ones) but it is hard to see how these negative categories ( cognitive artefacts ) are based upon human perception, cultural imprints and social conventions [Massolo et al., 2002, p.8]. 17

18 Figure 5: Taxonomy of basic categories in DOLCE (from Masolo et al, 2002) tion) or may change position (transfer (movement, emission, etc)), or their existence, both of the material nature and composition of objects and the energy types and distribution in processes. Mental entities behave largely analogous to physical objects. In fact, one may argue that the mental world consists largely of metaphors of the physical world. 23 A typical mental object is concept, and mental processes affect mental objects. This reflects our folk psychology which assumes e.g. that if one is informed about some fact, this fact is stored in memory. Whether this fact is believed or not is an epistemological issue. Facts of belief and knowledge are mental objects consisting of concepts. Communication proceeds via physical objects (documents, sounds) and actions (talk, reading) which represent mental objects (information). The mental and the physical world overlap in concept of agent. It is ambiguous because agent is classified as both a physical object and a mental object. 24 Social organization and -processes (e.g. communication) are composed of roles that are performed by agents that are identified as individual persons. The law associates norms to roles. For instance, the traffic regulation provides norms for traffic participants (or its subclasses, eg pedestrian or driver of a motor vehicle). 25 However, when it comes to solving legal cases, the responsibility is with the individual who performed a role. 23 [Lakoff & Núñez, 2000] present a convincing account of the primacy of conceptual schemas about physical processes that are metaphorized to conceptualize arithmetic, respectively full mathematics. 24 The multiple view evades the classical mind-body problem; cf. [Strawson, 1959]. 25 There is an interesting exception and that is that criminal law does not address roles but individuals, in the same way as morals do. Probably, from all domains of law, criminal law is (still) closest to morals. 18

19 Figure 6: LRI-Core in Protege-2000 Time and space have also an ambiguous status. Related to occurrences, they provide positions of events and situations. However, as physical entities they provide the qualities of extension (size, life-cycle) of objects and processes (field, duration). LRI-Core, containing about 200 concepts, is still under development, but has definitions for most of the anchors that connect the major categories used in law (person, role, action, process, procedure, time, space, document, information, intention, etc.). This is the major purpose of this ontology. It should not only provide some framework to get a coherent view on a particular legal domain ontology, but also it allows inheritance of well defined terms e.g. for verifying the domain ontology. LRIcore is written in DAML+OIL/RDF using Protégé. As soon as translators become available it will be translated in OWL. In Figure 6 the major structure is presented. Validating an upper ontology is not simple. The representation tools (DAML+OIL with FACT) enable consistency checking (verification) but the real proof is in actual uses. The main intended use is supporting knowledge acquisition for (legal) domains, but a real test of its semantics should be whether it enables natural language understanding of descriptions of simple events as in legal cases. What is meant here is that a system (artificial agent) should be able to infer e.g. the consequences of a document that is created (start-of-life-cycle) that is modified or extended etc. Of course, this is ultimately what the Semantic Web is about and we can be sure that the life cycle of a foundational ontology is not very likely to arrive in that stage: like so many theories it will be rather replaced by better argued for proposals far before it has been undergone comprehensive empirical testing. In fact, our contribution is in this respect rather in making explicit the principles than in its applied use in the various projects. This role of LRI-Core of as a super-structure for a legal domain ontology is presented in Figure 7 for some of the major terms in criminal law. 19

20 6 Criminal law Figure 7: LRI-Core as a super-structure of a legal domain The e-court project is aimed at the semi-automated information management of documents produced during a criminal trial: in particular the transcriptions of hearings. The structure of this type of document is determined by the debate/dialogue nature of these hearings, but also by specific, local court procedures. Besides tagging its structure, it is also important to identify (annotate) content topics of a document. These vary from case descriptions (e.g., in oral testifying) to topics from criminal law (e.g., in the indictment). The case descriptions have a strong common-sense flavour and of course we do not intend to develop here a comprehensive common-sense ontology. 26 Therefore we are currently developing an ontology that covers Dutch criminal law, whose major structure we will discuss below. As the e-court solutions are aimed to work for most European countries, in principle we have to develop such an ontology for every jurisdiction that intends to use e-court. This Dutch ontology will be the framework for ontologies of Italian and of Polish criminal law. How much of this framework will be reusable and whether it is easy to map terms from these jurisdictions to one another is difficult to predict. This mapping is more complex than the mapping of the vocabularies of different languages (cf EURO-Wordnet), because criminal law has been the traditional moral concern of nations, and will be the last kind of legislations to get harmonized in the European tuning of legislation. It will be largely an empirical enterprise. Studies of comparative criminal law are rather concerned with characterizing the major principles and histories that make up similarities and differences between criminal legislation, but are hardly concerned with a definitional comparison of the terms used in the actual legislation. Modelling and inference tools may here be very useful, but they are still unknown to legal scholars. See [Boer & vanengers, 2003] about the role and construction of such tools in comparing legislation. Moreover, the focus in comparative law is rather on differences than on similarities: these are probably assumed. 26 The legal professionals who are the intended users are in the first place interested in the legal aspects of the case. Also, criminal law contains already terms of criminal actions, means and objects. However, we also intend to experiment here with natural language ontologies (vocabularies) like Wordnet. 20

21 agent person natural person juristic-person company association foundation collection-of-agents group organization public Ministry-of-Justice courts-by-jurisdiction criminal-court administrative-court courts-by-level cantonal-court court-of-appeal Supreme-court Figure 8: Agents in Dutch Criminal Law (OCL.NL) (excerpt) Because it is not apparent where divergences between the concepts of criminal law between legal systems will be found, we want to ground, or anchor, the ontology of criminal law at a very abstract level in the LRI-Core. There will be little debate about the fact that criminal actions are physical or symbolic ones; that a verdict is a mental qualification represented by a document; that a person who is being accused will perform in court the role of defendant, and that persons as agents can perform both physical and mental activities etc. These anchor-points are not only useful to attach the legal sub-classes and composites. They provide a checklist and, more importantly, they foster the understanding that legal concepts most often imply several perspectives. This is not the same as the classical ambiguities that terms may have. For instance, the most important aspect of a legal document is its content, but the law often requires that the document is physically unique as well, and that it has a prescribed structure or elements. By multiple classification (and inheritance) these points of view can be easily combined and distinguished. Moreover, during the modeling the related points of view may suggest additional classifications and can be used for consistency checking. We can illustrate the use of anchors in the LRI-Core ontology with parts of the ontology for Dutch criminal law (OCL.NL). In Figure 8 the boldface terms are terms from LRI-Core. LRI-Core knows about the distinction between a person as a lifetime identity and roles that a person may perform. Roles are taken by persons who are agents. Agents are both physical and mental objects. However, roles are only behavioural specifications or requirements: they are not agents or persons [Guarino & Welty, 2002]. Agents or persons can perform roles, but they cannot be roles. We need this perspective to be able to understand what is meant by the generic statement that drivers of vehicles should keep to the right : drivers are roles that can perform actions 27. However, we should also be able to interpret a statement from a case description that says that Alexander Boer did not keep to the right of the A-5 with his car, in such a way that Alexander Boer is a natural person that acted in the role of driver and performed the actor-role in the keeping (= driving) action. In Figure 9 a selection of typical legal roles in (criminal) law is presented. One may observe the fact that these roles are not the roles to which (substantial) criminal law refers to. As stated before, 27 To be precise, there are two kinds of roles involved here: the role of a person to play driver and the actor-role where the driver performs the drive action. These latter roles are roles of actions, while the former roles are roles of agents. 21

22 role social-function public-social-function jurisdiction public-prosecution criminal-investigation forensic-investigation social-role legal-role juridical-role judicial-role judge judge-presiding prosecution-role public-prosecutor defense-role defense-counselor defendant principal-defendant accessory-defendant offender convict witness clerk-of-court lawyer the-regent the-state public-servant owner-of-goods owner-of-rights/duties creditor debtor Figure 9: roles and functions in Dutch Criminal Law (OCL.NL) (excerpt) 22

23 mental-object juridical-mental-object legal-norm judicial-mental-object complaint accusation judicial-decision verdict conviction acquit final-verdict juridical-qualification deontic-qualification deontic-legal-role-attribute right duty authority deontic-modalities-of-norms permission obligation prohibition reasoning-object evidence testimony eye-witness-testimony forensic-evidence problem-solving-role solution problem problem-solving-method argumentation-roles debate-argument-role accusation-position defense-position mental-process/action internal-mental-processes reasoning communicative-mental-action testifying interrogating argument dialogue dialogical argument dispute judicial-dispute mental-state legal-mental-state sane mental-incapacity mentally-handicapped Figure 10: Mental objects, processes and states in Dutch Criminal Law (OCL.NL) (excerpt) 23

24 criminal law is addressed to persons instead of roles, and is somewhat an exception in the legal world. However, as soon as persons are object of criminal procedures, they get roles in these procedures. In LRI-Core we distinguish between social roles and social functions. Social functions are external roles of organizations. Social roles make up the functional internal structure of an organization. In these figures we cannot show multiple classification, nor other relations between classes than subsumption. For instance, an organization has social functions and has-as-parts social roles. This is not the only view on the composition of an organization. The hierarchy of authority is another one, but this hierarchy maps onto the roles: authority is a mental entity: to be precise a deontic-legal-role-attribute (see Figure 10). Figure 10 gives in a nutshell some of the major categories of the mental world. To some extent, the mental world contains many metaphors of the physical world, but it is in no way a direct mapping. It provides a vocabulary of the folk (naive) psychology and sociology we apply when thinking about and modeling the mental world. We have to model mental worlds in order to understand one-self, but more importantly to interpret and understand the actions and expressions of others. Note that there is no need to ground this ontology on a biological-physical basis (reductionism). Mental objects are as much real and first class citizens as physical objects. We avoid commitments to mind-body issues by having both a physical view and a mental view on agents. 28 In this representation of the mental world we have skipped some views. Some mental legal objects, such as accusation are in fact (illocutionary) acts. In legal discourse an accusation is really treated as an object, i.e. it is the (content; sometimes the literal surface structure) text that is referred to. However, its meaning is indeed the act of accusation, so it should inherit properties of mental objects and those of (illocutionary) mental actions. Legal procedures may objectify or reify these actions. Many objects of the mental world are reifications of epistemological roles. Terms like reason, evidence, explanation, problem, dispute etc. come from the vocabulary of reasoning methods and are concerned with assessing the (trust in) the truth of (new) beliefs. As stated in the Introduction, law is particularly concerned with terms that act like handles to come to grips with justifying legal decisions. In fact, one may see even terms like obligation, prohibition, etc. to objectify the imperatives of (illocutionary) discourse. The statement that vehicles should keep to the right is reified as an obligation. 29 An important distinction between physical and mental actions is that for the latter the effects may not be confined to the mental world. Thinking, memorizing etc, are actions that concern only the inner mental world, but talking, pointing, writing etc. are communicative acts that transfer mental objects via symbols to the physical world. If we want to describe these communicative acts including their intended effects we have to add illocutionary acts. When communicative acts interact, we use the term dialogue. Communicative actions are modeled in the first place as mental actions, mediated by symbolic and physical representation. One may argue that this still contains the flavour of epistemology because it states a theory about how we acquire information, but it says nothing about how we justify this information. The hard core of the OCL.NL consists of actions. There are two major types: the criminal actions themselves (called offences ). These are of course the actions executed by the person who is successively acting as suspect, defendant, and eventually convict (if true and proven...). Ontheother side, the convict may be at the receiving end of the punishment actions, that are declared by the legal 28 At some level of granularity this may pose problems. For instance, folk psychology holds that the container of mental objects -the mind- coincides with a part of the body; the brain. Aside from the fact that this anatomical correspondence is neurologically and physiologically incorrect the nervous system is highly influenced by non-neural physiological processes this ontological commitment is not required at all to describe mental and physical activities. It does not matter to model mental actions by reference to body-parts. Traditionally, we think with our brain and feel with our heart, but this (fading-out) notion of folk psychology may be replaced by a more embodied mind view that is emerging in cognitive science. However, for the purpose of interpreting mental actions we do not need some commitment to physical location of processing. Note also that we are talking here about folk psychology; not about the metaphysical question whether there is even such a problem. 29 There is a strong tendency in law to do away even with these verbs of modality. In the Dutch instructions for drafting legislation, it is advised to avoid these verbs and put the statements as factual description, i.e. this article is expressed in the final version of the traffic code as: drivers of vehicles keep to the right as much as possible. In terms of discourse this trespass of common sense modality rules conveys a certain arrogance...: thecommand is presented as a fact. 24

25 action criminal-action offence felony offence-against-the security-of-the-state taking-life-of-regent theft-and-stripping offences-against-human-life murder manslaughter deception misdemeanour lesser-offences-related-to-public-order lesser-offences-related-to-public-moral punishment principal-punishment imprisonment for-life for-a-determinate-period detention additional-punishment disqualification deprivation-of-a-right deprivation-of-an-immunity deprivation-of-a-privilege Figure 11: Criminal actions in Dutch Criminal Law (excerpt) 25

26 system etc. Crime and punishment are the keys to criminal law that is synonym to penal law. 6.1 Use of ontologies in legal information retrieval The ontologies of criminal laws are to be used in e-court to support the information retrieval of information contained in the hearing session documents. Criminal law is only part of the discourse in these sessions, but an important part. Another part consists of descriptions of what has happened for which only full blown common sense (CYC) or superficial but extensive ontologies like Wordnet may play a role in information retrieval. Thus far we focus on the criminal legal terms, because the primary type of users are legal professionals. In e-court, two user modes of search are used: basic and advanced. The basic search mode allows meta-data and/or keyword search by specifying values for one or more meta-data fields and/or keywords. The advanced search mode includes possibilities to use linguistic weights and quantifiers with the keywords, to select the language of the query and the searched documents; to choose particular document sections of interest. In this subsection we describe the specific additional information management functions that are supported by ontologies Annotation and XML tagging of legal documents In information management the emphasis has been on archiving and retrieving documents by their formal, syntactic characteristics. These structures are abstracted in meta-data: RDBM schemas, DTDs for XML-tags, XML-Schemata, etc. This works fine as long as the structures are rather fixed and the occurrence of parts ( sections ) is easy to identify in an automatic way. The criminal trial hearing documents in e-court are not the typical kind of documents that are handled by information systems. Hearing documents reflect in the first place oral, often spontaneous dialogue from the court room. The role of ontologies in indexing the e-court hearing documents is threefold: The first role is an indirect one: the ontologies provide the structured vocabulary for metadata descriptions and maintain consistent use and semantic distinctions. The XML-Schemata only provide syntactic, structural information, but the ontologies (expressed in RDF/S) enable semantic coherence and verification. Although we may design DTDs (or XML-Schemata) in advance to capture dialogue-turns, phasic structuring, and argument-roles, most of these cannot be identified and tagged in an automatic way in the documents themselves. In most cases this can only be performed by a human agent, e.g. the transcriber who is capable of understanding what is (legally) going on in the hearing. This identification process is supported by a browser that gives options for annotating/tagging (in a context sensitive way) to structure the hearing transcripts. In such a way we may obtain multiple structuring of the hearing documents that increases the search options of the user. The identification of dialogue-turns can be (almost) fully automated by the use of simple voice-recognition devices that have only to distinguish voice characteristics of the participants in the dialogue. The e-court system indexes all documents. A number of these indexed terms correspond with terms of the ontologies. In this way we can link documents automatically with some semantics, i.e. one may gather what the document is about, which is functionally equivalent to (XML)- tagging the document with these terms For pragmatic reasons we have provisionally opted for this solution, although XML-tagging is the method to be used on the web. 26

27 6.1.2 Query expansion The set of keywords used in a query can yield unsatisfactory results because the actual use of terms in a document may not correspond to what the user has in mind. This is obvious in the use of synonyms. However, also more abstract terms may be used to denote a more specific object: e.g. killing (synonym: manslaughter) for murder. A reference to a murder may be missed because in the document the terms killing and manslaughter are used. The reverse may also be relevant in information retrieval. The user may search for the weapon that is used in a particular criminal case, but may not know what kind of weapon exactly was used. By browsing a taxonomy of weapons (e.g. as part of an ontology of terms in criminal law) she may specify the query further. We have observed that users of legal information retrieval systems have a tendency to underspecify the cases they present. They do not provide all potentially relevant facts, and they use terms that are too general. Therefore the system may miss potential exceptions in the set of norms, and the user may deduce a wrong normative assessment of her case. Having a user interface that explicitly allows for specification of used terms may help. Another solution is to have the system return potential exceptions as well as norms that exactly match the case at hand ([Winkels et al., 1999]). In both search modes (basic and advanced) the ontology repository is consulted for subsumed or subsuming terms with respect to the keywords given. Expansion by Subsuming Classes By adding terms for searching that are superclasses 31 of the already specified terms, the search is directed also to the more general, abstract terms. In searching documents that contain regulations (laws, statutes, contracts) where applicable provisions are often formulated in generalized and abstract terms, this IR strategy is in fact the only one to avoid false negatives (i.e. missed applicable provisions). In the CLIME project (IST ) this strategy has been implemented by the University of Amsterdam as part of the MILE demonstrator and it has been used to determine the applicability of norms on the basis of an ontology of about 3,500 terms ([Winkels et al., 2002]). Expansion by Subsumed Classes The example of the search for a weapon above shows the problem when the user is searching for a subclass of a term she may well know. There are two possibilities. The user may allow all subsumed terms to participate as keywords in the search (which may lead to an explosive return of candidates) or she may have already restricted the set of possible documents and have a look at those weapons that occur as indices of these documents. In fact, the example is typical for the kind of searches where one is looking for additional, very specific information that should answer a question. In those cases, the user usually has specific cases in mind: even has the document already retrieved but has to find the exact information. In CLIME, the conceptual retrieval part of the legal information server could extend the query using any relation between the search terms and others in the domain ontology. As a result, the return set could be very large and we had to devise special mechanisms for filtering and ordering it ([Winkels et al., 2000]. Disambiguation of a keyword term is another role of ontologies in IR. Classical ambiguity consists of terms that have different meanings but the same orthography. Except for orthographic coincidences, most ambiguous terms in fact share meaning, besides their differences. Disambiguation occurs in the context of use and is a matter of degree. There may be little ambiguity in the term car as an isolated term, but there is little overlap in what it implies between the mechanic s and the salesman s view of cars, even if they work for the same company. In ontologies persistent, but context (role) dependent ambiguity is represented as multiple classification. Except for disambiguation and selective use of terms of subsumed classes, the additional terms are added as disjunctive keywords to the query set, which means that the set of documents that is returned the result set may have increased exponentially. One may find more correct returns, but one must be prepared for a large amount of false positives: the classical problem of information overload we try 31 One may also include wholes from part-of hierarchies. 27

28 to avoid and for which the major web stakeholders (at least the W3C) see the solution in the semantic web technology. It appears there is not a free lunch at the web, nor at e-court that seeks the same solutions. There are two methods to cope with this problem. The first one is to have the user refine his query. However, this is often a problem because the user may not have enough information to do this. A second solution consists of (re)organizing the result set. The typical problem in (WWW) information search is that the number of returned documents may be unmanageably large and heterogeneous. The cause of much heterogeneity is the fact that a term may have multiple senses/views. In particular, the legal (criminal) domain is full of multiple views as we explained in Section 6, so we expect that disambiguation may occur by not only matching the indices of the returned documents with the keywords, but also have a second filtering/clustering where we also match indices with associated terms in the ontologies, i.e. the value(-classe)s and other related terms in the ontologies. 7 Ontologies and the structure of legal documents In knowledge and information management, there is a large variety of types of document-structures. If we look at texts documents may range from narrative texts (stories, histories, case descriptions, testimony) via non-narrative texts (reports, articles, handbooks, instructions) to fully pre-structured filled-in forms. Legal documents cover this full range. Moreover, the actors in the legal domain deeply believe in the universal adequacy of textual expression. One may say, that legal practioners are textfetishists. There is a good reason: the concern with evidence makes recording a basic requirement. However, the bad side is that written texts (on paper) are still the almost exclusive trustworthy media of recording and communication. In our bureaucratic society, documents are the universal basis of managing organizations. However, there are some types of documents which are almost unique to law and which play an important part in e-court. The first ones are regulations (codes, legislation, contracts, etc.). They have already played a major role in other projects we are/were involved (E- POWER, CLIME) but they are also the object of international initiatives to arrive at standards for describing the structure of legislation by the use of XML tags. Ontologies play here the role of providing dictionaries for the tagging framework. The framework we developed for the LeXML initiative, as described in the next section, will also be the basis for tagging documents that contain criminal law in e-court in order to be able to associate references to criminal issues with the appropriate articles. In the E-POWER project, this framework is both used to support drafting tax-regulations, as well as for linking it with other documents and also legal reasoning systems that refer to regulations. 7.1 Capturing the structure of regulations: MetaLex Regulatory documents function as the source of law in the continental system. Case law and legal doctrine is of importance, but codified norms in the form of rules and regulations are the prime source. Text is the main, if not only, way to represent them 32. This representation is necessary to pinpoint what valid law at a certain moment in time is. It is also useful for communicating norms (typically legal professionals are the intended audience, not the general public), and for referencing and (thereby) justifying decisions (the backing in legal argumentation). Regulatory documents are strange documents. They are never read from cover to cover; each article presents a separate discourse. One can read the contained articles in any order, the resulting discourse structure, the message, is supposed to be the same. Each of those articles plays an independent role as instrument in certain (epistemic) acts. The containing document is a special purpose container that posits the article in the legal system and provides a position, an identity, by which it can be unambiguously referenced. The legal source as an object with a unique identity and history is of course not the same as some paper or electronic copy of it. A laborious process in both legal publishing and decision making is determining what the contents of legal sources are at some point or interval in time. Changes can be announced in separate documents and publishers keep track of 32 Some norms may be captured in hardware like a speed delimiter on trucks. 28

29 all documents from certain publication channels to be able to reconstruct what the form of an organic law is at some time point. Each element, each sentence of a document can go through a complex lifecycle. It exists in some time-interval, it may be active or inactive, and its scope of application may be extended or limited in contained time-intervals. To keep track of this lifecycle each element must have an object identity by which it can be referenced so that it can be positioned in time and in relation to other document elements. For purposes of representation we distinguish three different viewpoints on the meaning of legal documents: Form A legal document can usually be recognized and classified by certain phrases and formulas. Formal requirements on structure and phrasing mostly reflect considerations of consistency of language and ease of access for the reader, but it also provides a context for the interpretation of the content of the document. Structural requirements are defined in XML schemas where appropriate. Role Although we may look at the phrases and formulas in a written decision to classify a document as a law, we know that it is not the structure of the document that makes it a law, but the role the document plays in the activities of public persons and bodies most importantly the activities that produced the document. Information about the document of this nature is captured in RDF statements about the document. Content We also classify documents depending on what its content means: It represents a type of decision. If it is just a public decision its meaning is limited to a particular occurrence or case. If it is a norm or policy its meaning extends to general class of occurrences or cases and it postulates a value theory for making and judging decisions. This is captured in RDF statements about this content: relating acts, norms, agents to (parts of) the document. The MetaLex XML schema limits itself as much as possible to the form of legislative documents. Since the standard is supposed to be independent of language and jurisdiction, a guiding principle in designing it has been: when in doubt, leave it out. Regulations come from many legislators (for instance, municipal, water authority, provincial, national, EU, and international for a civil servant in the Netherlands) and special-purpose software to support decision making processes is affected by and may have to manipulate legislation that conforms to many different standards for legislative drafting and is delivered in as many different formats. The XML schema for regulations groups together articles in parts, and subdivides articles in subparts, bottoming out in sentences. Articles are self-contained discourses in the sense that they can be read and understood without reference to nearby articles to resolve anaphoric references [Boer et al., 2002]. To keep track of versions MetaLex provides a number of attributes for every structural XML element in the document that can be identified, selected, and thus changed; the date-publication of an element is the time the element is officially published or announced. The date-enacted, the time the content becomes applicable in decision making, is always later than or the same as date-publication, but before date-repealed, the time the content becomes inapplicable in decision making. Between date-enacted and date-repealed the element and its content is active, and outside this interval it is inactive. The date-version attribute represents the date the correctness of the content and other dates of the XML element was last verified. The XML document looses its value as a normative reference as time progresses and the time-interval between date- version and today increases. 8 Discussion and conclusions The research reported here covers a long period (about fifteen years) over a number of mid-size (European) projects. In this article we have only pursued and discussed the views and results related to the development of a core ontology that identifies the main concepts that are typical, and preferably exclusive for law. This guided tour has revealed that in fact the law has only a few exclusively typical 29

30 concepts: probably only those related to normative knowledge (deontic terms) and to notions about legal responsibility [Lehmann, 2003]. Already at the start of our investigations in the TRACS project that was concerned with traffic regulations we found that by far the majority of terms referred to the common-sense world of traffic; the only exception were the already mentioned deontic terms. In dividing up legal knowledge in world knowledge, normative knowledge and responsibility knowledge 33 where world knowledge will take about 99% percent of the terms used in a legal domain. Does this mean that law is a typical common-sense domain? The answer is yes and no. The no is explained by the fact that the legal world knowledge (LAM, in terms of Valente s FOLaw [Valente, 1995]) is a filtered and adapted version of what may have started as a common-sense. It should be noted that we have not divided up law in the way it is universally conceived and taught in law, where the first distinction is between private and public law, and public law covers such legal domains as administrative law and criminal law. In each of these domains of law one will find concepts that may have evolved from common-sense, but which have received a typical and exclusive meaning in law. Moreover, as in all domains of professional practice new concepts may have been developed. The understanding and use of the current state of these legal-domain specific legal concepts is for instance what legal education is about. A major reason for the evolution of legal terms in specific domains of law is in the first place due to the fact that the power of law is limited: it cannot command the physical world (so what is desirable in law is always a subset of what is possible). Therefore, only those concepts and relations are object of law that can be affected by human conscious (and individual) intervention. A second reason is that common sense terms may get refined and redefined in such a way that they correspond better with principles of law. A third reason is that law transpires the views and goals of the recent and current politics. Therefore it appears that within legal domains much common-sense is filtered, cleaned and transformed into a layer only understandable and usable for legal professionals. However, this is only part of the story. The legal system needs a close correspondence between the common sense view of the world and its terminology. This is not only required in modeling and legislation, but also in interpreting cases. Cases are accounts of what has happened and they are cast in narrative discourse: events that are connected by causes and intentions (reasons). Therefore, a mapping between legal and common-sense concepts has to be maintained. All this is to say that a core ontology for law in general has problems in covering this large area of world knowledge and has to resort in a first approach to common-sense foundational ontologies: the worst option for professional fields. The situation is as bad as it looks. We have also identified terms which are not exclusive, but still very typical and well elaborated for law: document, document-structure, role, etc. are central terms in law and may be grounded in a core ontology that imports these notions from a still high level and simple foundational (common-sense) ontology. This does not make up a nice, evenly distributed core ontology for law. However, some core ontologies proposed which give this flavour have been identified as being rather epistemological frameworks than as real ontologies. FOLaw is a clear example, and although epistemological frameworks may have important support roles in constructing legal reasoning services, they provide little support for the modeling of the bulk of concepts in a legal domain. This insight was acquired in the CLIME, KDE, E-POWER and e-court projects: all of these projects are rather concerned with legal information serving, and only in the CLIME project the question answering mode required legal problem solving capabilities. Therefore, the need to capture also high level concepts of world knowledge that are typical in legal domains became more important and we started the development of LRI-Core. In information retrieval there is less need for deep semantics (inference): the focus is on the indexing of terms and simple, direct implications like subsumption. At first sight it may seem that the use of ontologies in legal information management is not paralleled by the effort in creating high level well represented and cleaned ontologies. The descriptive logics basis of the representation system (DAML+OIL/OWL) may look like overkill. One may object that such a relatively heavy apparatus with constrained expressiveness is not really necessary. Most of the information retrieval and storage 33 We leave reactive and creative knowledge aside as being indeed exclusive for law, but further of little ontological or epistemological interest. 30

31 functions can be supported by relatively simple lexicons. In fact, we use now in e-court such a lexical approach as built-in in the latest version of Oracle to start experimenting with a prototype as soon as possible. However, there are various reasons to use a more richer, formally well grounded knowledge representation formalism. All these reasons are related to the fact that these formalisms allow trustworthy, proven reasoning methods. So why do we need these? In the first place to verify the consistency of the ontologies created. Informal modeling in ontologies does not give any check on errors other than some kind of visual inspection. For ontologies larger than 200 terms this becomes unmanageable. We do not advocate here strict and formal modeling in a kind of straightjacket, but our experiences show that particularly in designing the basic framework for an ontology, consistency checking plays a very important diagnostic role. In later stages of knowledge acquisition, the consistency checking rather gets the role of tracing local errors and mistakes. Still, consistency checking and classification facilities cannot replace a good understanding of the domain. The second reason is that lexicons do not allow for multiple classification and inheritance. Legal domains are riddled with multiple view terms and these facilities are therefore certainly required. Another reason is that we do not only need the terms as they occur in some classification lattice, but also their attributes, values and relations with other terms. This is required for the disambiguation and for clustering of returned documents. In fact, one can see these values as additional query information that enables further distinctions. For instance, for the attributevalues for the car salesman are prices, accessories, etc: for the mechanic they are parts and part-numbers, while both may have to refer to the same models and versions of types of cars. References [Alchourrón & Bulygin, 1971] C. Alchourrón and E. Bulygin. Normative Systems. Springer- Verlag, Wien, [Alchourrón & Bulygin, 1981] [Allen & Saxon, 1991] [Allen, 1997] [Austin, 1954] C. Alchourrón and E. Bulygin. The expressive conception of norms. In R. Hilpinen, editor, New Studies in Deontic Logic, pages D. Reidel, L. Allen and C. Saxon. More IA needed in AI: Interpretation assistance for coping with the problem of multiple structural interpretations. In Proc. of the Third International Conference on AI and Law, pages 53 61, Oxford, L. Allen. The language of legal relations, RLL: useful in a legal ontology toolkit? In P.R.S. Visser and R. Winkels, editors, Legal Ontologies: proceedings of the ICAIL-97 workshop, pages ACM, New York, J. Austin. The Province of Jurisprudence Determined. Weidenfeld and Nicolson, London, edited by H.L.A. Hart. [Bench-Capon & Coenen, 1991] T. Bench-Capon and F. Coenen. Exploiting isomorphism: development of a kbs to support british coal insurance claims. In M. Sergot, editor, Proceedings of the third International Conference on AI and Law, pages 62 69, New York, Acm. [Bentham, 1970] J. Bentham. An Introduction to the Principles of Morals and Legislation. Athlone Press, London, Edited by J. Burns and H. Hart. 31

32 [Boer & vanengers, 2003] A. Boer and T. van Engers. A knowledge engineering approach to comparing legislation. In M. Wimmer, editor, Proceedings of the Conference on Knowledge Management in E-Government. Springer LNAI 2645, [Boer et al., 2002] A. Boer, R. Hoekstra, and R. Winkels. Metalex: Legislation in XML. In T. Bench-Capon, A. Daskalopulu, and R. Winkels, editors, Legal Knowledge and Information Systems, pages 1 10, Amsterdam, IOS Press. Proceedings of JURIX 2002: The Fifteenth Annual Conference. [Borst et al., 1997] [Breuker & Den Haan, 1991] P. Borst, J. M. Akkermans, and J. L. Top. Engineering ontologies. International Journal of Human Computer Studies, 46(2/3): , Joost Breuker and Nienke Den Haan. Separating world and regulation knowledge: where is the logic? In Proceedings of the third international conference on AI and Law, Oxford, Acm. [Breuker & Van De Velde, 1994] Joost Breuker and Walter Van De Velde, editors. CommonKADS Library for Expertise Modeling: reusable problem solving components. IOS-Press/Ohmsha, Amsterdam/Tokyo, [Breuker et al., 2000] [Brouwer, 1990] J.A. Breuker, E. Petkov, and R. Winkels. Drafting and validating regulations: the inevitable use of intelligent tools. In S.A. Cerri and D. Dochev, editors, Artificial Intelligence: methodology, systems and applications, pages Springer, P.W. Brouwer. Samenhang in Recht: een analytische studie (coherence in law). Wolters-Noordhoff, [Clancey, 1985] W.J. Clancey. Heuristic classification. Artificial Intelligence, 27: , [Gangemi et al., 2002] A. Gangemi, N. Guarino, C. Masolo, A. Oltramari, and L. Schneider. Sweetening ontologies with DOLCE. In A. Gomez-Perez and V.R. Benjamins, editors, Proceedings of the EKAW-2002, pages Springer, [Guarino & Welty, 2002] N. Guarino and C. Welty. Evaluating ontological decisions with OntoClean. Communications of the ACM, 45(2):61 65, [Haan & Breuker, 1996] N. Den Haan and J.A. Breuker. Constructing normative rules. In R. Van Kralingen, editor, Proceedings of JURIX-96, pages 41 56, [Haan & Winkels, 1994] N. Den Haan and R.G.F. Winkels. The Deep Structure of Law. In H. Prakken, A.J. Muntjewerff, and A. Soeteman, editors, Legal Knowledge Based Systems: The Relation with Legal Theory Seventh International Conference on Legal Knowledge-based Systems Legal Knowledge-Based Systems, JURIX 1994, pages Koninklijke Vermande, [Haan, 1996] N. Den Haan. Automated Legal Reasoning. PhD thesis, University of Amsterdam,

33 [Hage & Verheij, 1999] [Hamfelt & Barklund, 1990] [Hart & Honore, 1985] J. Hage and B. Verheij. The law as a dynamic interconnected system of states of affairs: a legal top ontology. International Journal of Human Computer Studies, 51: , A. Hamfelt and J. Barklund. Metaprogramming for representation of legal principles. In Proc. of Meta 90, pages , H. Hart and T. Honore. Causation in the Law. Oxford University Press, New York, second edition edition, [Hart, 1961] H. Hart. The Concept of Law. Clarendon Press, Oxford, [Hayes, 1985] [Hofstadter & McKinsey, 1939] [Hohfeld, 1919] [Jansweijer et al., 2000] [Kelsen, 1991] [Kent, 2003] [Kowalski & Sergot, 1985] [Kowalski, 1990] P.J. Hayes. The second naive physics manifesto. In J.R. Hobbs and R.C. Moore, editors, Formal Theories of the Common Sense World, pages Ablex Publishing Corp., Norwood, A. Hofstadter and J. McKinsey. On the logic of imperatives. Philosophy of Science, 6(4): , W. Hohfeld. Fundamental Legal Conceptions as Applied in Legal Reasoning. Yale University Press, Edited by W.W. Cook, fourth printing, W. Jansweijer, E. Van Der Stadt, J. Van Lieshout, and J.A. Breuker. Knowledgeable information brokering. In B. Stanford and P.T. Kidd, editors, E-bussiness;: key issues, applications and technologies. IOS-Press, Hans Kelsen. General Theory of Norms. Clarendon Press, Oxford, R. Kent. The iff foundation for ontological knowledge organization. In Knowledge Organization and Classification in International Information Retrieval, Cataloging and Classification Quarterly. The Haworth Press Inc., Binghamton, New York, R. Kowalski and M. Sergot. Computer representation of the law. In Proceedings of IJCAI-85, pages , Los Altos, CA, Morgan Kaufmann. R.A. Kowalski. Programming in the year In J.W. Lloyd, editor, Computational Logic: symposium proceedings, Brussel 1990, pages , Berlin, Springer Verlag. [Lakoff & Núñez, 2000] George Lakoff and Rafael Núñez. Where Mathematics Comes From. Basic Books, [Lehmann & Breuker, 2001] [Lehmann, 2003] J. Lehmann and J.A. Breuker. On defining ontologies and typologies of objects and processes for causal reasoning. In A. Pease, C. Menzel, M. Uschold, and L. Obrst, editors, Proceedings of IEEE Standard Upper Ontology, pages 31 36, Menlo Park, AAAI- Press. J. Lehmann. Causation in Artificial Intelligence and Law: a modelling approach. PhD thesis, University of Amsterdam,

34 [Lindahl, 1992] [MacGregor & Bates, 1987] [Massolo et al., 2002] [McCarty, 1989] [Mommers, 2002] [Muntjewerff & Breuker, 2001] [Patil, 1988] L. Lindahl. Conflicts in Systems of Legal Norms: A Logical Point of View. In Bob Brouwer, Ton Hol, Arend Soeteman, Willem Van Der Velden, and Arie De Wild, editors, Coherence and Conflict in Law, pages 39 64, Deventer, Boston / Zwolle, Kluwer Law and Taxation Publishers / W.E.J. Tjeenk Willink. Proceedings of the Third Benelux-Scandinavian Symposium in Legal Theory, Amsterdam, January 3-5, R. MacGregor and R. Bates. The LOOM knowledge representation language. Technical Report ISI/RS , ISI, University of Southern California, C. Massolo, S. Borgo, A. Gangemi, N. Guarino, A. Oltramari, and L. Schneider. The WonderWeb foundational ontologies: preliminary report. Technical Report Deliverable D17, version 2, ISTC-CNR (Italy), T. McCarty. A language for legal discourse I. basic structures. In Proc. of the Second International Conference on AI and Law, pages , Vancouver, Acm. L. Mommers. Applied legal epistemology: building a knowledgebased ontology of the legal domain. PhD thesis, University of Leiden, A. Muntjewerff and J. A. Breuker. Evaluating PROSA, a system to train legal cases. In Johanna Moore, Carol Redfield, and Lewis Johnson, editors, Artificial Intelligence in Education, FAIA-Series, pages , Amsterdam, IOS-Press. ISSN: , ISBN: R.S. Patil. Artificial intelligence techniques for diagnostic reasoning in medicine. In H.E. Shobe and AAAI, editors, Exploring Artificial Intelligence: Survey Talks from the National Conferences on Artificial Intelligence, pages Morgan Kaufmann, San Mateo, California, [Pinker, 1994] S. Pinker. The language instinct. Harper Collins, [Raz, 1972] [Rector, 2002] [Sowa, 2000] Joseph Raz. Legal principles and the limits of law. The Yale Law Journal, 81: , A. Rector. Medical informatics. In F. Baader, D. Calvanese, D.L. McGuinness, D. Nardi, and P.F. Patel-Schneider, editors, Description Logic Handbook, pages Cambridge University Press, franconi/dl/course/dlhb/dlhb- 10.pdf. John F. Sowa. Knowledge Representation: Logical Philosophical, and Computational Foundations. Brooks Cole Publishing Co, Pacific Grove, CA, [Strawson, 1959] P.F. Strawson. Individuals: an essay in descriptive metaphysics. Routledge,

35 [Valente, 1995] [Valente et al., 1999] [Van Der Velden, 1992] [Van Heijst et al., 1997] [Van Kralingen et al., 1999] [vankralingen, 1995] [Winkels & Haan, 1995] [Winkels et al., 1999] A. Valente. Legal knowledge engineering: A modelling approach. IOS Press, Amsterdam, The Netherlands, A. Valente, J.A. Breuker, and P.W. Brouwer. Legal modelling and automated reasoning with ON-LINE. International Journal of Human Computer Studies, 51: , W. Van Der Velden. Coherence in law: a semantic and deductive explanation. In Bob Brouwer, Ton Hol, Arend Soeteman, Willem Van Der Velden, and Arie De Wild, editors, Coherence and Conflict in Law, pages , Deventer, Boston / Zwolle, Kluwer Law and Taxation Publishers / W.E.J. Tjeenk Willink. Proceedings of the Third Benelux-Scandinavian Symposium in Legal Theory, Amsterdam, January 3-5, G. Van Heijst, A. Th. Schreiber, and B.J. Wielinga. Using explicit ontologies for kbs development. International Journal of Human- Computer Studies, 46(2/3): , R. Van Kralingen, P. Visser, T. Bench-Capon, and van den Herik H. A principled approach to developing legal knowledge systems. International Journal of Human Computer Studies, 51: , R. W. van Kralingen. Frame-based Conceptual Models of Statute Law. PhD thesis, University of Leiden, The Hague, The Netherlands, R.G.F. Winkels and N. Den Haan. Automated legislative drafting: Generating paraphrases of legislation. In Icail-95, editor, Proceedings of the Fifth International Conference on AI and Law, pages , College Park, Maryland, Acm. R.G.F. Winkels, D. Bosscher, A. Boer, and J.A. Breuker. Generating exception structures for legal information serving. In Th.F. Gordon, editor, Proceedings of the Seventh International Conference on Artificial Intelligence and Law (ICAIL-99), pages , New York (NY), ACM. [Winkels et al., 2000] R.G.F. Winkels, D. Bosscher, A. Boer, and R. Hoekstra. Extended conceptual retrieval. In J.A. Breuker, R. Leenes, and R.G.F. Winkels, editors, Legal Knowledge and Information Systems (JURIX-2000), pages 85 98, Amsterdam, IOS Press. [Winkels et al., 2002] Radboud Winkels, Alexander Boer, and Rinke Hoekstra. CLIME: lessons learned in legal information serving. In Frank Van Harmelen, editor, Proceedings of the European Conference on Artificial Intelligence-2002, Lyon (F), Amsterdam, IOS-Press. 35