Developing HARNESS: Towards a hybrid architecture for LKIF

tags: LKIF

We have a couple of related technologies:

Description logic based tools mostly used for ontology modeling, which are decidable and have efficient reasoners
Rule-based tools used for modeling (legal) norms, which are much more expressive but less tractable

This paper describes an effort to get the benefits of both by modeling norms as well as ontology in OWL 2 DL.

Notes

Executive Summary

NOTER_PAGE: (3 . 0.21722846441947566)

prototype for experimenting with normative legal reasoning using OWL 2 DL

NOTER_PAGE: (3 . 0.3625468164794008)

by starting with an ontology cast in OWL-DL, the rule formalism has to respect the semantic constraints under which the OWL reasoner has delivered its rea- soning results: the rules have to be ‘DL-safe’

NOTER_PAGE: (3 . 0.5910112359550562)

hybrid archi- tecture where only a small subset of LKIF rules could be used

NOTER_PAGE: (3 . 0.6576779026217229)

we could use the same DL reasoner (Pellet) to handle both the ontology and the legal norms for assessing legal cases.

NOTER_PAGE: (4 . 0.14307116104868914)

even if all deontic reasoning can be handled by a DL reasoner, we will also need rules for more complex legal reasoning tasks.

NOTER_PAGE: (4 . 0.7235955056179776)

Introduction

NOTER_PAGE: (9 . 0.13468)

Legal assessment: a core task in legal reasoning

NOTER_PAGE: (9 . 0.322826)

fundamental problems in aligning a DL (Description Logic) reasoner with a rule engine

NOTER_PAGE: (9 . 0.4052434456928839)

combines reasoning with an ontology and a normative knowledge base, both expressed in OWL 2 DL

NOTER_PAGE: (9 . 0.43820224719101125)

DL reasoners are monotonic and deductive, and it is generally assumed that (legal) norms require some form of non-monotonicity

NOTER_PAGE: (9 . 0.4868913857677903)

norms express generic situations

NOTER_PAGE: (9 . 0.6838624338624338)

A case is a description of an actual situation

NOTER_PAGE: (9 . 0.6970899470899471)

Motivation for developing HARNESS

NOTER_PAGE: (10 . 0.13468)

Carneades has been used to run LKIF-Rules and LKIF-Argument

NOTER_PAGE: (10 . 0.2552910052910053)

translate OWL-DL into rules

NOTER_PAGE: (10 . 0.3134920634920635)

only a specific subset of DL expressivity can be used

NOTER_PAGE: (10 . 0.34788359788359785)

hybrid architecture using different formalisms in which different subtasks are handled by different reasoners

NOTER_PAGE: (10 . 0.4384920634920635)

DL based representations sound, complete and efficient reasoners, which are not (yet) available for rule based approaches.

NOTER_PAGE: (10 . 0.7235449735449735)

ontologies have a special status in reasoning

NOTER_PAGE: (11 . 0.27579365079365076)

In OWL, expressiveness is sacrificed for decidability to enable sound and complete reasoning

NOTER_PAGE: (11 . 0.66005291005291)

Hybrid architecture

NOTER_PAGE: (12 . 0.13468)

In the hybrid approach there is a strict separation between the ordinary predicates, which are basic rule predicates and ontology predicates, which are only used as constraints in rule antecedents.

NOTER_PAGE: (12 . 0.45039682539682535)

starts from an ontology expressed in a DL formalism and searches for compatible rule formalisms.

NOTER_PAGE: (12 . 0.576058201058201)

the same reasoner can handle the rules and the ontology at the same time, provided the rules are specified as DLP (Descriptive Logic Programming) rules.

NOTER_PAGE: (12 . 0.6243386243386243)

DLP is now even defined in the OWL2 proposal as a special ‘profile’: OWL 2 RL

NOTER_PAGE: (12 . 0.6865079365079365)

the DLP rules are not very expressive

NOTER_PAGE: (12 . 0.7208994708994708)

it is the rules that do the work and the ontology provides semantic constraints

NOTER_PAGE: (12 . 0.7996031746031745)

the Case Description is an A-Box in DL terms

NOTER_PAGE: (13 . 0.6395502645502645)

extended case description is submitted to the rule engine that applies norms.

NOTER_PAGE: (13 . 0.7308201058201058)

Theoretical and practical problems in a hybrid architecture

NOTER_PAGE: (13 . 0.795527)

content of the A-Box has to be trans- lated in a format digestible for the rule engine.

NOTER_PAGE: (13 . 0.8419312169312169)

the rules that are complementary to OWL-DL should be ‘DL-safe’, meaning that they should be also constrained in expressiveness similar to OWL-DL

NOTER_PAGE: (14 . 0.6164021164021164)

DL-safe rules are still more expressive than DLP rules.

NOTER_PAGE: (14 . 0.7552910052910052)

DL-safe is rather restrictive.

NOTER_PAGE: (14 . 0.8068783068783069)

we want to cover as much as possible the full range of LKIF’s KR formalisms in one rational architecture that allows ontologies cast in OWL to be used in rule based legal reasoning. The requirement that the rule formalism should be DL-safe already limits the constructs in LKIF rules drastically.

NOTER_PAGE: (15 . 0.14351851851851852)

rules enable constructs that are difficult or even impossible to model in OWL 2 DL.

NOTER_PAGE: (15 . 0.22685185185185183)

OWL does not know about variables.

NOTER_PAGE: (15 . 0.24404761904761904)

Searching for an appropriate rule companion for OWL is also a long time concern for the W3C committee on rules for the Semantic Web. An initial candidate, SWRL has been abandoned and after about four years of work, the current proposal is an interchange format rather than a specific formalism: RIF

NOTER_PAGE: (15 . 0.3498677248677248)

modeling knowledge in OWL-DL is not as easy as in rules

NOTER_PAGE: (15 . 0.5496031746031745)

Exploiting OWL for reasoning with norms

NOTER_PAGE: (15 . 0.678538)

legal norms could also be represented as classes in OWL.

NOTER_PAGE: (15 . 0.7261904761904762)

norms to be expressed in LKIF- rules were modeled in OWL using the LKIF-Core ontology that contained the definitions of LKIF rules.

NOTER_PAGE: (15 . 0.7731481481481481)

classify descriptions of individuals (a case description) as belonging to a norm class

NOTER_PAGE: (16 . 0.18386243386243384)

classifier can classify the norms in a subsumption hierarchy which is the basis for discovering which norms are exceptions to which other norms

NOTER_PAGE: (16 . 0.2876984126984127)

exceptions are discovered by the reasoner

NOTER_PAGE: (16 . 0.33796296296296297)

one can use DL non-safe rules as long as the effects do not modify anything stated (constrained) by the OWL reasoner.

NOTER_PAGE: (16 . 0.4722222222222222)

Pellet is a monotonic reasoner that cannot handle the inconsistencies inherent in ‘exceptions’ to norms.

NOTER_PAGE: (16 . 0.6071428571428571)

identity-of-individuals problems

NOTER_PAGE: (16 . 0.7136243386243386)

If not, we need also variables, i.e. rules.

NOTER_PAGE: (16 . 0.7453703703703703)

OWL is part of LKIF in the first place because legal ontologies are usually written in OWL. In practice these ontologies are used for information retrieval purposes, while the legal knowledge systems developed use some rule formal- ism.

NOTER_PAGE: (16 . 0.7718253968253967)

Exploring the power and limitations of OWL

NOTER_PAGE: (18 . 0.13468)

use OWL 2 DL to express norms

NOTER_PAGE: (18 . 0.37896825396825395)

Classification and norm application

NOTER_PAGE: (18 . 0.69298)

Problems in identity and identifying individuals

NOTER_PAGE: (23 . 0.786685)

A general introduction to the problem

NOTER_PAGE: (23 . 0.821611)

The identity problem boils down to the fact that in some cases we would like to talk about things at the level of individuals, instead of at the level of concepts, which are actually sets of individuals.

NOTER_PAGE: (24 . 0.3683862433862434)

the course which the course book belongs to, is the same course that the student is enrolled in.

NOTER_PAGE: (24 . 0.4484126984126984)

When using a rule formalism we can easily model this example just by using variables

NOTER_PAGE: (24 . 0.4808201058201058)

Using conjunctive queries

NOTER_PAGE: (25 . 0.217802)

Definition and usage

NOTER_PAGE: (25 . 0.414687)

very similar to the body (condition) of SWRL rules

NOTER_PAGE: (25 . 0.5628306878306878)

Inferences with conjunctive queries

NOTER_PAGE: (26 . 0.716736)

Application in HARNESS

NOTER_PAGE: (27 . 0.70888)

A GUI for prototyping HARNESS

NOTER_PAGE: (28 . 0.337133)

An experimental test domain: a library regulation

NOTER_PAGE: (30 . 0.550783)

Conclusions, extensions and exploitation

NOTER_PAGE: (35 . 0.13468)

normative reasoning can be performed with a class-based, monotonic and deductive DL reasoner.

NOTER_PAGE: (35 . 0.3941798941798942)

this assessment task plays a minor but essential role in legal problem solving.

NOTER_PAGE: (35 . 0.5079365079365079)

ontologies in legal applications are not really used for reasoning

NOTER_PAGE: (35 . 0.5906084656084656)

a serious limitation in the interpretation of cases remains: it is not always possible to keep track of individuals.

NOTER_PAGE: (36 . 0.2275132275132275)

the HARNESS approach is not only capa- ble of checking consistency and generating correct solutions, but also in generating all cases that can be distinguished by a regulation: usually that is some orders of magnitude than one can imagine

NOTER_PAGE: (38 . 0.455026455026455)

Notes

Executive Summary

prototype for experimenting with normative legal reasoning using OWL 2 DL

by starting with an ontology cast in OWL-DL, the rule formalism has to respect the semantic constraints under which the OWL reasoner has delivered its rea- soning results: the rules have to be ‘DL-safe’

hybrid archi- tecture where only a small subset of LKIF rules could be used

we could use the same DL reasoner (Pellet) to handle both the ontology and the legal norms for assessing legal cases.

even if all deontic reasoning can be handled by a DL reasoner, we will also need rules for more complex legal reasoning tasks.

Introduction

Legal assessment: a core task in legal reasoning

fundamental problems in aligning a DL (Description Logic) reasoner with a rule engine

combines reasoning with an ontology and a normative knowledge base, both expressed in OWL 2 DL

DL reasoners are monotonic and deductive, and it is generally assumed that (legal) norms require some form of non-monotonicity

norms express generic situations

A case is a description of an actual situation

Motivation for developing HARNESS

Carneades has been used to run LKIF-Rules and LKIF-Argument

translate OWL-DL into rules

only a specific subset of DL expressivity can be used

hybrid architecture using different formalisms in which different subtasks are handled by different reasoners

DL based representations sound, complete and efficient reasoners, which are not (yet) available for rule based approaches.

ontologies have a special status in reasoning

In OWL, expressiveness is sacrificed for decidability to enable sound and complete reasoning

Hybrid architecture

In the hybrid approach there is a strict separation between the ordinary predicates, which are basic rule predicates and ontology predicates, which are only used as constraints in rule antecedents.

starts from an ontology expressed in a DL formalism and searches for compatible rule formalisms.

the same reasoner can handle the rules and the ontology at the same time, provided the rules are specified as DLP (Descriptive Logic Programming) rules.

DLP is now even defined in the OWL2 proposal as a special ‘profile’: OWL 2 RL

the DLP rules are not very expressive

it is the rules that do the work and the ontology provides semantic constraints

the Case Description is an A-Box in DL terms

extended case description is submitted to the rule engine that applies norms.

Theoretical and practical problems in a hybrid architecture

content of the A-Box has to be trans- lated in a format digestible for the rule engine.

the rules that are complementary to OWL-DL should be ‘DL-safe’, meaning that they should be also constrained in expressiveness similar to OWL-DL

DL-safe rules are still more expressive than DLP rules.

DL-safe is rather restrictive.

we want to cover as much as possible the full range of LKIF’s KR formalisms in one rational architecture that allows ontologies cast in OWL to be used in rule based legal reasoning. The requirement that the rule formalism should be DL-safe already limits the constructs in LKIF rules drastically.

rules enable constructs that are difficult or even impossible to model in OWL 2 DL.

OWL does not know about variables.

Searching for an appropriate rule companion for OWL is also a long time concern for the W3C committee on rules for the Semantic Web. An initial candidate, SWRL has been abandoned and after about four years of work, the current proposal is an interchange format rather than a specific formalism: RIF

modeling knowledge in OWL-DL is not as easy as in rules

Exploiting OWL for reasoning with norms

legal norms could also be represented as classes in OWL.

norms to be expressed in LKIF- rules were modeled in OWL using the LKIF-Core ontology that contained the definitions of LKIF rules.

classify descriptions of individuals (a case description) as belonging to a norm class

classifier can classify the norms in a subsumption hierarchy which is the basis for discovering which norms are exceptions to which other norms

exceptions are discovered by the reasoner

one can use DL non-safe rules as long as the effects do not modify anything stated (constrained) by the OWL reasoner.

Pellet is a monotonic reasoner that cannot handle the inconsistencies inherent in ‘exceptions’ to norms.

identity-of-individuals problems

If not, we need also variables, i.e. rules.

OWL is part of LKIF in the first place because legal ontologies are usually written in OWL. In practice these ontologies are used for information retrieval purposes, while the legal knowledge systems developed use some rule formal- ism.

Exploring the power and limitations of OWL

use OWL 2 DL to express norms

Classification and norm application

Problems in identity and identifying individuals

A general introduction to the problem

The identity problem boils down to the fact that in some cases we would like to talk about things at the level of individuals, instead of at the level of concepts, which are actually sets of individuals.

the course which the course book belongs to, is the same course that the student is enrolled in.

When using a rule formalism we can easily model this example just by using variables

Using conjunctive queries

Definition and usage

very similar to the body (condition) of SWRL rules

Inferences with conjunctive queries

Application in HARNESS

A GUI for prototyping HARNESS

An experimental test domain: a library regulation

Conclusions, extensions and exploitation

normative reasoning can be performed with a class-based, monotonic and deductive DL reasoner.

this assessment task plays a minor but essential role in legal problem solving.

ontologies in legal applications are not really used for reasoning

a serious limitation in the interpretation of cases remains: it is not always possible to keep track of individuals.

the HARNESS approach is not only capa- ble of checking consistency and generating correct solutions, but also in generating all cases that can be distinguished by a regulation: usually that is some orders of magnitude than one can imagine

Notes that link to this note