In this paper we describe a legal core ontology that is part of a generic architecture for legal knowledge systems, which will enable the interchange of knowledge between existing legal knowledge systems. This Legal Knowledge Interchange Format, is under development in the Estrella project and has two main roles: 1) the translation of legal knowledge bases written in different representation formats and formalisms and 2) a knowledge representation formalism that is part of a larger architecture for developing legal knowledge systems. A legal (core) ontology can play an important role in the translation of existing legal knowledge bases to other representation formats, in particular into LKIF as the basis for articulate knowledge serving. We describe the methodology underlying the LKIF core ontology, introduce the concepts it defines, and discuss its use in the formalisation of an EU directive.
In this paper we describe a legal core ontology that is part of a
generic architecture for legal knowledge systems, which will enable the
interchange of knowledge between existing legal knowledge systems.
This Legal Knowledge Interchange Format (LKIF), is currently being
developed in the Estrella project.1 LKIF has two main roles: enable the
translation between legal knowledge bases written in different
representation formats and formalisms and secondly, as a knowledge
representation formalism that is part of a larger architecture for developing legal
knowledge systems. These use-cases for LKIF bring us to the classical
trade-off between tractability and expressiveness, as in e.g. KIF
(Knowledge Interchange Format,
1 Estrella is a 6th European Framework project (IST-2004-027665). See also: http://www.estrellaproject.org. The views and work reported here are those of the authors.
Proposing the OWL-DL subset of SWRL as its core does not make LKIF a formalism tuned to legal knowledge and reasoning: how do we get the ‘L’ into LKIF? To “legalize” LKIF it needs to be constrained in two ways. The first is a meta-component that controls the reasoning as to gear it to typical legal tasks. For instance, legal assessment and argumentation provide control structures for legal reasoning that put specific demands on the knowledge to be obtained from a legal knowledge base. The second constraint is not specialised to legal reasoning, but to legal knowledge. Typical legal concepts may be strongly interrelated and thereby provide the basis for computing equivalencies (paraphrases) and implications. For instance, by representing an obligation as the opposite of a prohibition, a (legal) knowledge system can make inferences that are specialised to these terms. In our view, specialised legal inference should be based on definitions of concepts involved in an ontology. Concept definitions should make all necessary and sufficient interrelationships explicit; the inference engine can then generate all implied consequences. 2
A legal ontology can play an important role in the translation of
existing legal knowledge bases to other representation formats, in
particular into LKIF as the basis for articulate knowledge serving. Similar
to a translation between different natural languages, a formal,
‘syntactic’ translation may clash with the semantics implied by the original
knowledge representation. An ontology, as representation of the
semantics of terms, allows us to keep track of the use of terms in a
knowledge base. Furthermore, and more importantly, an ontology can
support the process of knowledge acquisition and modelling in legal
domains. Defining concepts like ‘norm’, ‘judge’, ‘liability’, ‘document’,
‘claim’, etc. helps to structure the process of knowledge acquisition.
Earlier experience, as in e.g.
The following sections describe the theoretical and methodological framework against which the LKIF core ontology has been developed (Section 2 and 3). Section 4 describes the different modules of the ontology, and introduces its most important concepts. Section 5 gives an example of how the ontology can be used in the formalisation of a regulation.
2 For an ontology cast in OWL-DL these inference engines are description classifiers, e.g. Pellet, http://pellet.owldl.com/
We adhere to a rather restrictive view on what an ontology should
contain: terminological knowledge, i.e. intensional definitions of
concepts, represented as classes with which we interpret the world. The
distinction between terminological knowledge (T-Box) and assertional
knowledge (A-Box) has already been around for a long time. As a rule,
terminological knowledge is generic knowledge while assertional
knowledge describes the (actual) state of some world: situations and events.
However, these asserted states can become generalised into typical
patterns related to particular situations. To be sure, if experiences re-occur
and have a justifiable structure, it might evidently pay to store these
structures as generic descriptions, because they deliver a predictable
course of events for free. Eating in a restaurant is a typical example
and it served in the Seventies to illustrate the notion of knowledge
represented by scripts
Finally, an important reason to distinguish frameworks from
ontology proper is that frameworks often imply epistemic roles which
require reasoning architectures that go beyond the services provided
by OWL-DL reasoners (e.g. meta-level reasoning). It should be noted
that frameworks are generic, i.e. they act as pre-specified patterns that
get instantiated for particular situations. We have distinguished the
following types of frameworks:
Situational frameworks Situational frameworks are stereotypical
structures of plans for achieving some goal in a recurrent context. Making
coffee may be such a plan. However, the plans may involve transactions
in which more than one actor participates. For instance, the definition
of Eating-in-a-restaurant3 shows the dependencies between actions of
clients (ordering, paying) and service personnel (noting, serving) as its
major structure. This is the internal structure of the concept, but it
usually does not make sense to create class-subclass relations between
such frame-like concepts. The Eating-in-a-restaurant is not some natural
sub-class of Eating. It refers to some typical model of how eating is
put in the context of a restaurant. We can introduce a proliferation of
all contexts of eating, such as Eating-at-home, Eating-with-family, etc.
but these contexts do not fundamentally differ, cf.
Mereological frameworks Many entities, both objects and processes often have parts: they are composites. It is tempting to include a mereological (part-of) view in the definition of a concept. For instance, defining a car as having at least three, and usually four wheels, and at least one motor. However, a full structural description of all its parts and connections goes beyond what a car essentially is. Mereological frameworks appear under a large diversity of names: structural models, configurations, designs, etc. Arguably, the distinction between a mereological framework and a defining description of a term (ontology) is sometimes be very thin. For instance, if we want to describe a bicycle as distinct from a tricycle, it is necessary to use the cardinality of the wheels as defining properties as these are central to the nature of the bicycle. On the other hand, the number of branches a tree might have hardly provides any information as to what a tree is.
Epistemological frameworks Inference structures are often represented
as epistemological frameworks of interdependencies between reasoning
steps. Typical examples are the problem solving methods (PSM) found
in libraries of problem solving components
This focus on the use of knowledge, its epistemological status (e.g.
hypothesis vs. conclusion) and the dependencies between distinct steps
in a methodology is characteristic for epistemological frameworks.
Epis4 Although the terms ‘reasoning’ and ‘inference’ are often used as more or less
synonymous, we want to reserve the term inference for making explicit what is
implicit in a knowledge base, given some inference engine.
temological frameworks can be more abstract than PSMs. For instance,
the Functional Ontology of Law, which is presented as a core ontology, is
an epistemological framework that describes the role of law as a control
system in society
The construction of LKIF followed a combination of methodologies for
ontology engineering. Already in the mid-nineties, the need for a
wellfounded methodology was recognised, most notably by
Whereas the domain of
In
For the purpose of the LKIF ontology, we have made slight
adjustments to the methodology of
This section describes how the methodology described in the previous section was applied to the development of LKIF Core. We first describe the building and clustering phase, followed by a discussion of the existing ontologies we considered for inclusion, and a description of the concepts defined in the different modules of the ontology.
The LKIF Core ontology should contain ‘basic concepts of law’. It is
dependent on the (potential) users what kind of vocabulary is aimed
at. We have identified three main groups of users: citizens, legal
professionals and legal scholars. Although legal professionals use the legal
vocabulary in a far more precise and careful way than laymen, it
appears that for most of these terms there is still a sufficient common
understanding to treat them more or less as similar
The Estrella consortium includes representatives of the three kinds of experts. Each partner was asked to supply their ‘top-20’ of legal concepts. Combined with terms we collected from literature (jurisprudence and legal text-books) we obtained a list of about 250 terms. As such a number is unmanageable as a basic set for modelling, we asked partners to assess each term from this list on five scales: level of abstraction, relevance for the legal domain, the degree to which a term is legal rather than common-sense, the degree to which a term is a common legal term (as opposed to a term that is specific for some sub-domain of law), and the degree to which the expert thinks this term should be included in the ontology. The resulting scores were used to select an initial set of 50 terms plus those re-used from other ontologies (see section 4.2), and formed the basis for the identification of clusters and the development of the LKIF Core ontology.
We expected to be able to reuse terms and definitions from existing core
or upper ontologies that contain legal terms, as e.g. listed in
The intentional nature of the core concepts for the LKIF ontology
(see e.g. sections 4.3.2,4.3.3) emphasises the distinction with other more
(meta)physically inclined top ontologies such as SUMO5, Sowa’s upper
ontology
7 www.cyc.com core norm time_modification legal_role legal_action
expression rules role action process mereology
top
time
place
1989, LLD) is a first attempt to define legal concepts in the context of
legal reasoning, using formulae and rules. Properly speaking, LLD is not
an ontology but a framework but it is a relatively rich source for legal
terms and their definitions. The Core Legal Ontology (CLO) is used
to support the construction of legal domain ontologies
The list of terms and insights from the requirements-phase resulted in
a collection of ontology modules, each of which represents a relatively
independent cluster of concepts: expression, norm, process, action, role,
place, time and mereology
Discussions, further literature study and the consideration of existing ontologies, led to an extension of the original set of clusters to 14 modules (see Figure 1), each of which describes a set of closely related concepts from both legal and commonsense domains. Nonetheless, we maintained the original views used to identify the clusters, as the explanations and justifications are still valid and applicable to the current version of the ontology. We can distinguish three layers in the ontology: the top level (Section 4.3.1), the intentional level (Section 4.3.2) and the legal level (Section 4.3.3).
The description of any legally relevant fact, event or situation requires a basic conceptualisation of the context in which these occur: the backdrop, or canvas, that is the physical world. Fundamental notions such as 8 We used both TopBraid Composer (http://www.topbraidcomposer.com) and Prot´eg´e 3.2/4.0 (http://protege.stanford.edu). owl:Thing Mental_Concept location, time, parthood and change are indispensable in a description of even the simplest legal account. The top level clusters of the ontology provide (primitive) definitions of these notions, which are consequently used to define more intentional and legal concepts in other modules. The most general classes of the LKIF ontology are borrowed from LRI Core. We distinguish between mental, physical and abstract concepts, and occurrences (Figure 2).
Mereological relations allow us to define parts and wholes, allow for expressing a systems-oriented view on concepts, such as functional decompositions, and containment (Figure 3). Furthermore, they form the basis for definitions of places (location) and moments and intervals in time.
The ontology for places in LKIF Core is based on the work of
Of the properties defined in this module, meet is the most basic as
it is used to define many of the other properties such as abut, cover,
coincide etc. See
Closely related to the theory of places of
The relations between temporal occurrences are what defines time.
Like
Place
Occurrence immediately_before, as the temporal meet relation holds only in one direction, and is asymmetric. The property is used to define other temporal relations such as before, after, during, etc.
With these classes and properties in hand, we introduce concepts of (involuntary) change. The process ontology relies on descriptions of time and place for the representation of duration and location of changes. A Change is essentially a difference between the situation before and after the change. It can be a functionally coherent aggregate of one or more other changes. More specifically, we distinguish between Initiation, Continuation and Termination changes.
Changes that occur according to a certain recipe or procedure, i.e. changes that follow from causal necessity are Processes; they introduce causal propagation. Contrary to changes, processes are bound in time and space: they have duration and take place at a time and place. We furthermore distinguish Physical_Processes which operate on Physical_Objects. Furthermore, at this level we do not commit to a particular theory of causation or causal propagation.
Legal reasoning is based on a common sense model of intelligent behaviour, and the prediction and explanation of intelligent behaviour. It is after all only behaviour of rational agents that can be effectively influenced by the law. The modules at the intentional level include concepts and relations necessary for describing this behaviour (i.e. Actions undertaken by Agents in a particular Role) which are governed by law. Furthermore, it introduces concepts for describing the mental state of these agents, e.g. their Intention or Belief, but also communication between agents by means of Expressions.
The class of agents is defined as the set of things which can be the actor of an intentional action: they perform the action and are Temporal_Occurrence
resource Process ...
Physical_Object
disjoint_with
Change Termination
Continuation
Initiation potentially liable for any effects caused by the action (see figure 5). Actions are processes, they are the changes performed by some agent who has the intention of bringing about the change. Because actions are processes they can become part of causal propagation, allowing us to reason backwards from effect to agent. Actions can be creative in that they initiate the coming into existence of some thing, or the converse. Also, actions are often a direct reaction to some other action (see figure 5).
The agent is the medium of some intended outcome of the action: an action is always intentional. The intention held by the agent, usually bears with it some expectation that the intended outcome will be brought about: the agent believes in this expectation. The actions an agent is expected or allowed to perform are constrained by the competence of the agent, sometimes expressed as roles assigned to the agent.
We distinguish between persons, individual agents such as “Joost Breuker” and “Pope Benedict XVI”, and Organisations, aggregates of other organisations or persons which acts ‘as one’, such as the “Dutch Government” and the “Sceptics Society” . Artefacts are physical objects designed for a specific purpose, i.e. to perform some Function as instrument in a specific set of actions such as “Hammer” and “Atlatl”9. Persons are physical objects as well, but are not designed (though some might hold the contrary) and are subsumed under the class of Natural_Objects. Note that natural objects can function as tools or weapons as well, the typical example being a stone, but are not designed for that specific purpose.
The notion of roles has played an important part in recent discussions on ontology (Steimann, 2000; Masolo et al., 2004; Guarino and 9 An atlatl is a tool that uses leverage to achieve greater velocity in spearthrowing, see http://en.wikipedia.org/wiki/Atlatl ...
Physical_Object
Artifact disjoint_with
member Natural_Object
Agent
Process Person
Action
Welty, 2002). Roles not only allow us to categorise objects according to their prototypical use and behaviour, they also provide the means for categorising the behaviour of other agents. They are a necessary part of making sense of the social world and allow for describing social organisation, prescribe behaviour of an agent within a particular context, and recognise deviations from ‘correct’ or normal behaviour. Indeed, roles and actions are closely related concepts: a role defines some set of actions that can be performed by an agent, but is conversely defined by those actions. Roles specify standard or required properties and behaviour (see figure 6). The role module captures the roles and functions that can be played and held by agents and artefacts respectively, and focuses on social roles, rather than traditional thematic or relational roles.
A consequence of the prescriptive nature of roles is that agents connect expectations of behaviour to other agents: intentions and expectations can be used as a model for intelligent decision making and planning10. It is important to note that there is an internalist and an externalist way to use intentions and expectations. The external observer can only ascribe intentions and expectations to an agent based on his observed actions. The external observer will make assumptions about what is normal, or apply a normative standard for explaining the actions of the agent.
10 Regardless of whether it is a psychologically plausible account of decision
making. Daniel Dennett’s notion of the Intentional Stance is interesting in this context
(cf.
Role
Function disjoint_with
Social_Role
The expression module covers a number of representational
primitives necessary for dealing with Propositional_Attitudes (viz. (
Communicated attitudes are held towards expressions: propositions
which are externalised through some medium. Statement, Declaration,
and Assertion are expressions communicated by one agent to one or
more other agents. This classification is loosely based on Searle (cf.
When propositions are used in reasoning they have an epistemic role, e.g. as Assumption, Cause, Expectation, Observation, Reason, Fact etc. The role a proposition plays within reasoning is dependent not only on the kind of reasoning, but also the level of trust as to the validity of the proposition, and the position in which it occurs (e.g. hypothesis vs. conclusion). In this aspect, the expression module is intentionally left under-defined. A rigourous definition of propositional attitudes relates them to a theory of reasoning and an argumentation theory. The argumentation theory is supplied by an argumentation ontology. The theory of reasoning depends on the type of reasoning task (assessment, design, planning, diagnosis, etc.) LKIF is used in, and should be filled in (if necessary) by the user of LKIF .
Evaluative_Attitudes express an evaluation of a proposition with respect to one or more other propositions, they express e.g. an evaluation, a value statement, value judgement, evaluative concept, etc. I.e. only ...
evaluatively_comparable Creation
Evaluative_Proposition
evaluates Evaluative_Attitude medium
Medium
Document
...
...
Speech_Act
Promise
Statement_In_Writing
Assertion
the type of qualification which is an attitude towards the thing being
evaluated, and not for instance the redness of a rose, as in
Legally relevant statements are created through public acts by both
natural and legal persons. The legal status of the statement is dependent
on both the kind of agent creating the statement, i.e. Natural_Person vs.
a Legislative_Body, and the rights and powers attributed to the agent
through mandates, assignments and delegations. At the legal level, the
LKIF ontology introduces a comprehensive set of legal agents and
actions, rights and powers (a modified version of
The Norm is a statement combining two performative meanings: it is
deontic, in the sense that it is a qualification of the (moral or legal)
acceptability of some thing, and it is directive in the sense that it
commits the speaker to bringing about that the addressee brings about
the more acceptable thing (cf.
A norm applies to (or qualifies) a certain situation (the Qualified situation), allows a certain situation – the Obliged situation or Allowed qualifies Mental_Object Permission Prohibition
Obligation equivalent_class situation – and disallows a certain situation – the Prohibited or Disallowed situation, see Figure 8. The obliged and prohibited situation are both subsumed by the situation to which the norm applies. Besides that they by definition form a complete partition of the case to which the norm applies, i.e. all situation to which the norm applies are either a mandated case or a prohibited situation. This is true of the obligation and the prohibition: they are simply two different ways to put the same thing into words. The permission is different in that it allows something, but it does not prohibit anything. The logical complement of the mandated situation is here an opposite qualified situation, about which we know only that it cannot be obliged.
The LKIF ontology not only provides a theoretical understanding of the legal domain, but its primary use in practice is as a tool to facilitate knowledge acquisition, exchange and representation: i.e. to formalise pieces of existing legislation. We evaluated the use of the ontology by formalising the EU Directive 2006/126 on driving licences,11, a relatively straightforward regulation, in which at least two types of normative statement are recognisable—definitional and deontic.
An example of a definitional statement from the EU directive is:
with a maximum design speed of not more than 45 km/h. 11 The text is available on-line at http://eur-lex.europa.eu/. The mereo module of the ontology along with a qualified cardinality restriction (available with OWL 1.1) allows us to express that AM vehicles have two or three wheels:
AM ⊑ 2composed_of.Wheel ⊔ 3composed_of.Wheel.
Modelling ‘design speed not more than 45 km/h’ is more challenging as it requires us to represent the rather common sense domain of speeds, distances etc. Of course, one could introduce the datatype property designSpeed and require its value be expressed in km/h.This choice, however, would not make justice of the conceptual complexity involved in ‘design speed not more than 45 km/h’, which contains reference to several notions: unit of measurement, number, designed speed, and a no-more relation. In fact, ‘design speed not more than 45 km/h’ can be rendered by imposing an linear ordering relation less-than on the different (instances of the) subclasses of the class DesignSpeed.12 The ordering allows us to define the class of those DesignSpeeds with a value not exceeding some N45—i.e., ∀less-than.DesSpeed-km-h-45.
Let us now look at an example of a deontic statement :
Art. 4(2) expresses an obligation whose logical form can be rendered by the implication:
If x is driving a AM vehicle, then x must be at least 16 years old. To fix some terminology, the antecedent is the context to which the obligation applies; the consequent (minus the deontic operator must) is the content of the obligation itself (what the obligations prescribes it ought to be the case). Consistently with this analysis, the LKIF ontology defines obligations as classes (see Section 4.3.3).
In our case, art. 4(2) allows the situation DriverAM ⊓ DriverOlderThan16 and forbids DriverAM ⊓ ¬DriverOlderThan16. Suppose that the classes DriverOlderThan16 and DriverAM have already been defined.13 To model the obligation that drives of AM vehicles must be at least the 16 years older, we introduced the obligation-type class MinAgeAM as follows:
12 The ordering is linear—i.e, reflexive, antisymmetric, transitive and total—since it mirrors the ordering of the natural numbers. For whenever n ≤ m, we have that DesignSpeed-km-h-n(a) less-than DesignSpeed-km-h-m(b), with a, b instances.
13 The class DriverOlderThan16 can be defined by using a more-than ordering relation, roughly along the same lines as the class ∀less-than.DesSpeed-km-h-45. The class DriverAM can be easily defined. MinAgeAM ⊑ ∀allows.(DriverAM ⊓ DriverOlderThan16).
MinAgeAM ⊑ ∃allows.(DriverAM ⊓ DriverOlderThan16).
MinAgeAM ⊑ ∀disallows.(DriverAM ⊓ ¬DriverOlderThan16).
MinAgeAM ⊑ ∃disallows.(DriverAM ⊓ ¬DriverOlderThan16).
Other deontic operators, such as permission or prohibition, can be
accounted in an alike manner (see
The representation of art. 4(2) suggests the LKIF ontology be augmented with a module taking care of quantities, units of measurement, numbers, fractions, mathematical operations, and the like. This is crucial not only for the EU Directive 2006/126, in which most definitional statements contain quantitative features of vehicles (e.g., power, cylinder capacity); quantities and calculations play a central role in any legislative text. Note, however, that the LKIF ontology can only provide a purely terminological account, without being able to do mathematical computations. This is unavoidable, given that OWL is a purely logical language. We are currently investigating whether we can import an existing OWL ontology dealing with measurements, such as PhysSys/SUMO or from the Ontolingua server15.
As LKIF Core was developed by a heterogeneous group of people, we
specified a number of conventions to uphold during the representation
of terms identified in the previous phases (See
In retrospect, this convention turned out to pose severe problems for existing OWL-DL and OWL 1.1 reasoners as their performance is significantly affected by the generic concept inclusion axioms (GCI): 14 See e.g. the SEAL project, http://www.leibnizcenter.org/project/ current-projects/seal
15 See http://www.ksl.stanford.edu/software/ontolingua/ axioms where the left-hand side of a subClassOf statement is a complex class definition. These axioms are abundant when defining classes as equivalent to e.g. someValuesFrom restrictions and in combination with lots of inverse property definitions, this creates a large completion graph for DL reasoners16 As result of these findings, the LKIF ontology has undergone a significant revision since its initial release.
Using LKIF Core in practice, as e.g. in the traffic example, points to the traditional knowledge-acquisition bottle-neck: for any formal representation of any domain, one needs to build formal representations of adjoining domains. As has been said, this can be largely overcome by including specialised foundational or domain ontologies in a representation based on the LKIF ontology provided that the quality of these ontologies is sufficient. Depending on availability we might consider providing a library of ‘compatible’ ontologies to users of LKIF Core. This will be of especial use when the ontology vocabulary will be adopted for expressing the LKIF vendor models that will be developed within ESTRELLA.
With respect to coverage of the legal domain, the purpose of the study outlined in Section 4.1 is more ambitious than only the selection of the most basic terms for describing law, but time and effort constraints make it that we could only consider a small pool of referents. The list of terms will be subjected to a more rigourous empirical study, whereby we will consult a group of legal professionals (taking courses in legal drafting), and law students. These empirical studies are planned in the sideline of ESTRELLA. By applying statistical cluster analysis, we might be able to identify the properties of the scales used (e.g. are they independent?) and whether the statistical clusters have some resemblance to the clusters we have identified based on theoretical considerations. The results of this analysis will be used to evaluate the ontology compared to the requirements we identified in the previous chapters.
The LKIF ontology is available online as separate but interdependent OWL-DL files, and can be obtained from the ESTRELLA website at http://www.estrellaproject.org/lkif-core. This website also provides links to online documentation and relevant literature.