This research aims to define an integrated methodology for modelling judgments, starting from legal texts and capturing not only structural parts but also arguments used by judges to reach conclusions. The goal is to build a complete ontology framework capable of detecting and modelling knowledge from the judgement's text. The formalized judgements provide the necessary metadata for the rule layer to enable argumentation towards the acceptance or rejection of a given interpretation. For pursuing this goal it is important to integrate the legal ontology construction with a rule formalization following legal reasoning-oriented theory and defeasible logics, just like the Carneades application (presented here) follows Walton's argumentation theory. The XML interchange uses OWL, RuleML, and the emerging LegalRuleML.
Precedent is a main element of legal knowledge worldwide: by settling conflicts and sanctioning illegal behaviours, judicial activity enforces law provisions within the national borders, therefore supporting the validity of laws as well as the sovereignty of the government that issued them. Moreover, precedents (or case-law) are a fundamental source for law interpretation and it paradoxically happens that the exercise of jurisdiction can influence the scope of the same norms it has to apply, both in common law and civil law legal systems – even if to different extents.
The goal of the present research is to define a framework for case-law semantics,
exploiting Semantic Web technologies to achieve isomorphism between the text
fragment (the only binding legal expression) and the legal rule, thus "filling the gap"
between document representation and rules modelling [
This research is based on a middle-out methodology: top-down for modelling the core ontology, bottom-up for modelling the domain ontology and the argumentation rules. The research starts from the analysis of a sample set of 27 decisions of different grade Italian case-law (tribunal, court of appeal, Cassation Court) concerning consumer law2.
While DL is very powerful and semantically rich (making OWL an acknowledged standard for the modelling of document metadata), monotonic logics are not sufficient for modeling legal rules in a verifiable way, while at the same time maintaining the structure of the legislation and regulation they are ought to represent: legislation is typically organized as general rules subject to exceptions, and arguments made by applying legal rules are often defeasible. Moreover, the application of laws depends on time, and various legal rules may conflict with each other: these conflicts are resolved using legal principles about priority relationships between rules. Metarules such as these, however, need defeasible logics in order to be properly managed.
The research relies on the previous efforts of the community in the field of legal
knowledge representation [
3
The Core Ontology (an extension of the LKIF-Core Ontology) introduces a model of the legal domain whose main elements are: legal rules, legal concepts, material circumstances, judicial claims, judicial interpretations, adjudications3. 2 The matter is specifically disciplined in Italy through the "Codice del Consumo" (Consumer Law) and articles 1341-1342 of the Civil Code. This discipline is also present in all foreign legal systems, which will allow an extension of the research to foreign decisions and laws. 3 Even though the core ontology should be domain-generic and not modeled upon a specific legal subject, the sample model was conceived only to successfully represent the interaction in the civil law subject, when contracts, laws and judicial decisions come into play.
Following this structure, the metadata taken from judicial documents are
represented in the Domain Ontology. The modeling was carried out manually by an
expert in the legal subject, which actually represents the only viable choice in the
legal domain: automatic information retrieval and machine learning techniques, in
fact, do not yet ensure a sufficient level of accuracy, even if some progress in the field
has been made, for example in applying NLP techniques to recognize law
modifications [
Such a layered ontology creates an environment where the knowledge extracted from the decision’s text can be processed and managed, in such a way as to enable a deeper reasoning on the interpretation instances grounding the decision itself. Example of this deeper reasoning include: finding relevant precedents which were not explicitly cited in the decision; validating the adjudications of the judge on the claims brought forward by the parties during the trial on the basis of applicable rules, accepted evidence and interpretation; suggesting legal rules/precedents /circumstances that could bring to a different adjudication of the claim. The layered structure of the ontology set also allows an efficacious scaling from legal concepts to factors, up to dimensions and legal principles: all these concepts can be represented in the domain ontology, and the hook of the core concepts to the LKIF taxonomy should allow semantic alignment between domain ontologies when different authors are concerned (yet, the ontology set alignment has not been tested by the present research). 4
The last part of the research relies on argumentation modeling to perform legal
reasoning on the knowledge contained in the ontologies, in order to evaluate their
capabilities. As a first test of the ontology set, the research developed a pilot case
using the Carneades Argumentation System4 [
The present research developed applications [
The so-built system is capable of creating argumentation graphs pro or con a given
legal statement: for example, it can state whether a contract clause can be judged as
inefficacious, under which norms, and following which judicial precedents and
judicial interpretations. These arguments are brought forward by the system not only
when all the premises for the argument are accepted in the knowledge base:
Carneades is in fact capable of “suggesting” incomplete arguments [
As a result of their application to the ontology set, the LKIF-Rule and Clojure rule languages showed to be unsatisfactory in: identifying a border between semantic representation and syntactic modeling; importing ontologies: in the Carneades application no distinction is made between stated and inferred knowledge, and no feedback of new knowledge into the ontology reasoner is possible; providing basic legal deontics operators to represent obligation, violation, reparation and penalty; providing defeasibility logic operators defining the hierarchy between rules; modeling temporal dimensions to represent the three axes of enforceability, efficacy, applicability; assigning IDs for single parts of the rules, necessary to reach full isomorphism between the rule and the source legal document(s); qualifying rules with metadata such as author, data of creation of the rule, jurisdiction of the rule, etc.
The critical point in importing ontologies, adding factors and writing rules was the design and management of information between the ontologies and the rules: some of the axioms already modeled in the ontologies, in fact, could better meet their potentialities if modeled as an LKIF or Clojure rules instead. The issue, anyway, should be solved with general criteria, since the two systems use different logics (description logic for OWL vs. first-order predicate logic for LKIF and Clojure). This suggests the distinction between static information (thesauri, taxonomies, administrative and procedural data) to be included in the ontology, and legal concepts (legal statuses, subsumptions, inclusion of a material circumstance into the scope of a norm) to be modeled as rules for argument evaluation.
To manage defeasibility of arguments (and thus rules) Carneades includes proof
standards [
5
The situation presented above is likely to present itself over and over again as long
as modeling decisions (i.e. the introduction of elements of defeasible logics) are taken
just for the purposes of a single application. In order to fully exploit its potentialities,
AI&Law systems (such as legal argumentation systems) should instead rely on open
and shared standards. The research community joined the efforts towards the
definition of a standard for the syntax of legal rule extending LKIF-Rules with a
modeling of temporal parameters, giving birth to the LKIF++ language [
LegalRuleML [
Modelling the legal rules in the LegalRuleML language highlighted the potentialities of this tool in achieving isomorphism and representing defeasibility and temporal parameters. In the present proposal, the <ruleInfo> section introduces detailed information on the context of the rule. This rule-centric metadata approach favours the isomorphism with the legal text during the change management and the encapsulation of all information related to the rule in a unique XML node: <lrml:ruleInfo id="ruleInfo2" appliesTo="#rule2"> <lrml:sources id="sourceBlock2"> <lrml:source element="#atom1" idRef="#art1341-com2"/> <lrml:source element="#atom2" idRef="#art1341-com1"/> <lrml:source element="#atom3" idRef="#art1341-com1"/> 5 General contents of the proposal can be found at https://www.oasis-open.org/apps/org/workgr oup/legalruleml/download.php/46379/1.2PRINCIPLES.004.doc. 6 Documents of the OASIS LegalRuleML TC are available at https://www.oasisopen.org/committees/documents.php?wg_abbrev=legalruleml. The mailing list describing the work in progress can be browsed at https://lists.oasis-open.org/archives/legalruleml/.
<lrml:source element="#atom4" idRef="#art1341com2"/> </lrml:sources> <lrml:strength iri="&dfsont;defeater"/> <lrml:jurisdiction iri="&jurisdictions;italy"/> <lrml:author idRef="#aut1"/> <lrml:times idRef="#t1"/> <lrml:creationDateTime idRef="#e1"/> </lrml:ruleInfo>7
This section represents a context data container, since it introduces metadata which can be used in multiple circumstances to classify the rules. The bond between the text fragment, the legal rule and the author of the model is ensured by a list of <sources> for every element which constitutes the rule and by the element <author/>, linking the rule and its parts to specific individuals (a fragment of text and a person respectively), identified through an IRI. In this way it is possible to explicitly refer to the source documents of each part of a rule, also when a rule takes origin from multiple documental sources, allowing a clear distinction of which part of the rule comes from which document. At the same time, different rule authors are allowed to model the same text fragment in different ways, being always clear which author modelled which rules on a certain legal document.
Inside the <ruleInfo> section, the element <strength/> defines the role of each rule in the defeasible logics dynamics (strict/defeasible/defeater). Priority relations are built through the <Overrides> element8, in the following form: <Overrides id=”ovr1”> <Rule keyref="#rule_3"/> <Rule keyref="#rule_2"/> </Overrides>9
The <ruleInfo> section contains also a <times> element, not indeed a normal attribute but rather a section introducing a whole different layer: it contains information on the time periods of the legal rule’s coming into force, efficacy, application. The representation of temporal dimension of legal rules using three axes is a crucial addition towards the automatic management of legal rules in connection 7 The proposal can be found at http://www.oasis-open.org/apps/org/workgroup/legalruleml/dow nload.php/45887/2.8isomorphism.001.doc. 8 This tag is developed jointly with RuleML (http://ruleml.org/) and in particular with the
Defeasibility RuleML TG (http://ruleml.org/1.0/defeasible.html).
9 This implementation of defeasibility should allow a better management of exceptions than in
LKIF-Rule, where exceptions had to be made explicit in the rule syntax. If an exception
presents itself in the form of two legal fragments not explicitly referring to each other but
rather disposing opposing legal consequences (i.e. efficacy vs. inefficacy) it is possible to
model these rules independently, and then create a priority relation reflecting the actual
hierarchy between the two norms. Moreover, this solution allows a relative management of
hierarchy, without the need to assign arbitrary "weight" values to each rule. The proposal can
be found at https://www.oasis-open.org/committees/download.php/46454/2.1.1defeasibility.0
06.doc. Meaningful comments by TC member Tara Athan can be found at
http://www.oasisopen.org/apps/org/workgroup/legalruleml/download.php/45888/2.1defeasibility.002.002.doc.
with their legally binding documents [
Events' IDS are recalled by the <timeBlock> element, which adds information on the event which occurs (start, end) and on the axis which is affected: <lrml:timeBlock id="t1">
<lrml:time start="#e2" refType="&lkif;#efficacy"/> </lrml:timeBlock>10 6
The project presented here represents an effort towards the acquisition of an acknowledged standard for the rule layer of the semantic web layer cake, while at the same time trying to improve the state-of-the-art of legal knowledge representation by facing its main issues: the gap between document representation and rule modeling, and the need for a shared standard in the logic layer to represent legal reasoning. The Carneades Argumentation System has been used as a test field for the ontology set and the rules design, highlighting potentialities and issues of the approach.
Under these points of view, the approach proposed by the research represents a
step towards the filling of that gap, relying on existing standards to achieve the
isomorphism between legal document and rules; at the same time, the research defines
the requirements for a reasoning engine to be capable of semantically managing
knowledge coming from the ontology and applying legal rules to it. This engine will
probably not be a closed one, embedded in some argumentation tool (as in
Carneades): It should rather consist of a set of libraries to be implemented into
existing engines in order to introduce a complete management of defeasibility and a
standard language for interaction between these rules and OWL-encoded knowledge.
The intention, in the upcoming research on this behalf, is to rely on a Drools11
application under construction by CIRSFID and on NICTA's SPINdle12 [
A particular acknowledgement is due to Tom Gordon, provider of meaningful research inputs during my stage in Berlin, FOKUS, in the last autumn. Thanks also to Adam Wyner who gave me important feedbacks for improving the readability of the paper and also, as a legal domain expert, fruitful comments on the ontology modeling choices. Thanks go also to Harold Boley and Tara Athan of the RuleML Initiative for his helpful hints. Last but not least, I am grateful to my tutor and mentor Monica Palmirani, for her guidance in this challenging research. 10 The proposal can be found at http://markmail.org/message/oy34tkzzr3r2ldhz?q=temporal+lis t:org%2Eoasis-open+list:org%2Exml+list:org%2Eebxml. 11 Drools (www.drools.org) is a production rule system based on the Rete algorithm. 12 SPINdle is an open source, Java-based defeasible logic reasoner which conducts efficient and scalable reasoning on defeasible logic theories.