Overall Objectives
(Sans Titre)
knowledge representation
semantics of knowledge representation
ontologies
semantic web
content representation
context
knowledge transformation
ontology alignment
property preservation
multimedia document adaptation
semiotics
RDF
RDF Path
OWL
XSLT
Transmorpher
Alignment API
OLA
We consider that, in future information systems, formalized knowledge will be massively exchanged. In this communication process, the computer can complement its medium and memory roles by formatting, filtering, classifying, consistency checking or generalizing knowledge. These
manipulations can be thought of as transformations. This approach is developing with the generalized use of standardized exchange languages (
XML) in network communication as well as more precisely defined languages (
RDFand
OWL) on the semantic web. Consequently, the users must ask for more guarantees from the performed transformations.
Exmo's goal is the development of theoretical and software tools for enabling interoperability in formalized knowledge exchange.
Our main approach is the investigation of the properties that transformations must satisfy. Among these properties are content or structure preservation, traceability, or, conversely, confidentiality. We want to contribute to a "general theory of transformations" grounded on transformation
properties rather than on transformations themselves.
On the one hand, we aim at showing that this approach can be applied to various contexts depending on languages, properties and transformations. We study information preservation modelled by lattices or scenario preservation in adapted multimedia documents. On the other hand, we
investigate more precisely semantic properties when transforming from one knowledge representation language to another. The main question is: are the consequences of the initial representation still consequences of the transformed one? For helping answering it, we establish relationships
between several kinds of semantic properties, known or new. We also investigate more restricted cases (family of languages, patterns) that can use these relationships for establishing properties more quickly.
Our medium term objectives are (1) the opportunity to analyse transformations in terms of a set of composition operators, (2) the isolation of useful properties and the study of their behavior regarding these operators and (3) the capability to prove these properties on this kind of
transformations.
On a longer term, we want to explore properties that we call semiotic, i.e., which concern the interpretation of the communicated representation by a human user. This goal should require an analysis of the extra-semantic rules that govern the choice of subsets of models.
Anticipated applications are in transformation system engineering (in which the information system is seen as a transformation flow) and semantic web infrastructure.
Scientific Foundations
Knowledge representation semantics
We usually work with semantically defined knowledge representation languages (like description logics
We consider a language
Las a set of syntactically defined expressions (often inductively defined by applying constructors over other expressions). A representation (
r) is a set of such expressions. An interpretation function (
I) is inductively defined over the structure of the language to a structure called interpretation domain (
D). This expresses the construction of the "meaning" of an expression in function of its components. A formula is satisfied by an interpretation if it fulfills a condition (in general being interpreted over a particular subset of the domain). A model of a set of
expressions is an interpretation satisfying all these interpretations. An expression (
) is then a consequence of a set of expressions (
r) if it is satisfied by all of its models (noted
r
).
A computer must determine if a particular expression (taken as a query, for instance) is the consequence of a set of axioms (a knowledge base). For that purpose, it uses programs, called provers, that can be based on the processing of a set of inference rules, on the construction of models
or on more classical programming. These programs are able to deduce theorems (noted
r
). They are said to be sound if they only find theorems which are indeed consequences and to be complete if they find all the consequences as theorems. However, depending on the language and its semantics, the decidability (i.e., the ability to create sound and complete provers) is not
warranted. Even for decidable languages, the algorithmic complexity of provers can prohibit their exploitation.
To solve this problem a trade-off between the expressivity of the language and the complexity of its provers has to be found. These considerations have led to the definition of languages with limited complexity - like conceptual graphs and object-based representations - or of modular
families of languages with associated modular prover algorithms - like description logics.
Exmomainly considers languages with well-defined semantics (such as
RDFand
OWLthat we contributed to define), and defines the semantics of some languages such as multimedia specification languages, in order to establish the properties of computer manipulations of the representations.
Transformations and properties
A transformation (
) is an algorithmic manner to generate a representation (
(
r)) from another one (
r), not necessarily in the same language. We will focus on transformations made by composition of more elementary transformations for which we only know the input, output and assumed properties.
A transformation system is characterized by a set of elementary transformations and a set of composition operators. A transformation flow is the composition of elementary transformation instances whose input/output are connected by channels. A transformation flow is itself a
transformation. More precisely, our work concerns syntactic transformations of
XML("eXtensible Markup Language") structures encoding knowledge representation languages. We take advantage of the
XSLTtransformation language ("
XMLStyle Language Transformations"
The design of information systems as transformation flows requires the ability to express such flows and to determine their properties. A property is a boolean predicate about the transformation (e.g., "preserving information" is such a predicate - it is true or false of a transformation -
which is satisfied if there exists an algorithmic means to recover
rfrom
(
r)).
We consider more closely preservation properties that can allow the preservation (or anti-preservation) of an order relation between the source representations and the target representations. For instance, one can identify:
Syntactic properties: based on the organization of syntactic elements, like the completion (
(
r)
r, in which
denotes structural subsumption between representations);
Semantic properties: based on model and consequence concepts, like consequence preservation (
);
Semiotic properties: based on the interpretation as signs of the manipulated objects, like interpretation preservation (let
be the interpretation rules and
L,
ibe the interpretation of individual
i,
).
Our goal is to study transformations based on transformation properties rather than on representations or transformation structures. This does not deal only with semantics but considers various properties (e.g., content or structure preservation, traceability, and confidentiality).
However, we more specifically address semantic properties. We also consider properties of transformation systems (given a transformation system, is information preservation decidable and at what cost?). We try to characterize, given a particular type of property, which transformations leave
them invariant and what is the action of composition operators.
Ontology alignment
When different conceptual models (called ontologies) for representations are used, it is necessary to identify their correspondences before designing a transformation (or another use). This task is called ontology matching and its result is an alignment. It can be described as follows:
given two ontologies, each describing a set of discrete entities (which can be classes, properties, rules, predicates, etc.), find the relationships (e.g., equivalence or subsumption) that hold between these entities.
An alignment between two ontologies
Oand
O'is a set of correspondences
<e, e', r, n>in which
e∈Oand
e'∈O'are the entities between which a relation is asserted by the correspondence (e.g., formulas, terms, classes, individuals);
ris the relation asserted to hold between
eand
e'. Nothing is said about the relation but that it must apply to the pair of entities. For instance, this relation can be a simple set-theoretic relation (applied to entities seen as sets or their interpretation seen as sets), a relationship defined in representation languages
(such as subsumption), a fuzzy relation, a probabilistic distribution over a complete set of relations, a similarity measure, etc.
nis a degree of confidence in this particular correspondence. The degree of confidence belongs to a structure
<D, ≤, ⊤, ⊥>such that
Dis the set of degrees, ≤ is an order on
D×Dsuch that whatever
d∈D,
⊥≤d≤⊤. This structure is applicable to a wide number of measures (e.g., boolean lattice, fuzzy degrees, probabilities, any lattice). The confidence degree can be computed in many ways, including users' feedback or log analysis.
A variety of methods from the literature may be used for this task, basically they perform pair-wise comparison of entities from each of the ontologies and select the most similar pairs. Most matching algorithms provide correspondences between named entities, more rarely between compond
terms. The relationships are generally the equivalence between these entities. Some systems are able to provide subsumption relations as well as other relations in the support language (like incompatibility or instanciation). Confidence measures are usually given a value between 0 and 1 and
are used for expressing preferences between two correspondences.
Application Domains
Two application contexts motivate and spur our work: transformation system engineering (§
Transformation system engineering
Computerization and networking lead organizations to exchange information in machine-readable form. E-commerce generates a continuous flow of such documents. As the transmitted information is neither addressed nor adapted to all the members of the organization, it is necessary to transform
document structure and content. Similarly, web sites are generated from databases or primary funds and e-commerce documents are applied various transformations before goods are shipped. Additionally, the Object Management Group Model-Driven Architecture (
MDA) considers the future of software development as a composition of transformations from (platform independent) domain models to other (platform dependent) models in function of platform description models. This is considering any implementation as adaptation.
Interoperability requirements have led to the definition of the structured document representation language
XMLwhich helps handling the syntax of documents straightforwardly. Other languages such as
XSLTor Omnimark, enable the implementation of standalone transformations.
However, this view of transformations is only partial and local. It seems unavoidable that, in the future, we will have to deal with complex transformation flows automating the combination of transformations, some of which coming from external sources. This will require the global
understanding of the behavior of the flow of transformations. This calls for real "transformation system engineering" which should address the following issues:
the lack of global consideration of transformations: they are processed in relation with other transformations;
the need to consider the properties of transformations and especially their semantic properties: this will require the semantic analysis of the transformations;
the design of transformation flows from external resources (as it is in software engineering): this will require the ability to consider the properties of imported transformations.
Transformation system engineering will require tools, methodologies and formal methods. As a matter of fact, it will be necessary to check that a particular transformation system does not export sensitive information or that the transformation process terminates. For that purpose, the
transformation flow must be expressed in a parsable way and the expected properties of the flow must be expressed (see §
Semantic web technologies
Internet technologies support access and sharing of knowledge, often difficult to access in a documentary form, in the organizations. However, the technologies quickly reach their limits: web site organization is expensive and full-text search inefficient. Content-based information search
is becoming a necessity. Content representation will enable computers to manipulate knowledge on a more formal ground and to carry out similarity or generality search... Knowledge representation formalisms are good candidates for expressing the content.
The vision of a "semantic web"
Our work aims at enhancing content understanding, including the intelligibility of communicated knowledge and formal knowledge transformations. The semantic web idea is essentially based on the notion of ontology (that can be quickly described as conceptual schemes or knowledge bases).
Even if a standard knowledge representation language emerges, it will still be necessary to import and exchange ontologies in such a way that the semantics of their representation language is taken care of. We work on finding the correspondences between various knowledge representation
languages and ontologies (see §
Software
Exmo's work can be implemented in software: we have proposed an
APIfor expressing ontology alignment (§
Alignment API: manipulating ontology alignments
Jérôme
Euzenat
Contact
Jérôme
Pierson
Antoine
Zimmermann
We have designed a format for expressing alignments in a uniform way
The
APIitself is a
Javadescription of tools for accessing the common format. It defines four main interfaces (Alignment, Cell, Relation and Evaluator) and proposes the following services:
Storing, finding, and sharing alignments;Piping alignment algorithms (improving an existing alignment);Manipulating (thresholding and hardening);Generating processing output (transformations, axioms, rules);Comparing alignments.It will be extended with web service and agent communication languages in order to be used as an alignment service on the semantic web.
The
APIcan be used for producing transformations, rules or bridge axioms independently from the algorithm which produced the alignment. The proposed
APIimplementation is based on the
OWL-APIand offers floating point measures between 0 and 1. It features:
a base implementation of the interfaces with all useful facilities;a library of sample aligners;a library of renderers (
XSLT, SWRL, OWL, C-OWL, SEKTmapping language);a library of evaluators (precision/recall, generalized precision/recall, precision/recall graphs and weighted Hamming distance);a parser for the format.
To instanciate the
API, it is sufficient to refine the base implementation by implementing the
align()method. Doing so, the new implementation will benefit from all the services already implemented in the base implementation.
The
Alignment APIhas been used for the processing of the
[EON2004 Ontology Alignment Contest]and
[Ontology Alignment Evaluation Initiative 2005]. It is used in the
[people's portal alignment tool]at
DERIInnsbruck and used or output by a number of alignment tools (among which
[OLA]that we develop jointly with the University of Montréal or
[CMS]from University of Southampton).
This year we have specified extensions of this
APIfor dealing with more expressive languages
The
Alignment APIis freely available since december 2003 under the
LGPLlicence at
[http://co4.inrialpes.fr/align]
Transmorpher : expression and processing of
XMLtransformation flows
Jérôme
Euzenat
Contact
In order to prove or check the properties of transformations, it is necessary to have a representation of these transformations. The
XSLTlanguage enables the expression of a transformation in
XMLbut is relatively difficult to analyse. In order to overcome this problem, we have designed and developed in collaboration with the
Fluxmediacompany, the
Transmorpherenvironment.
Transmorpherenables the definition and processing of generic transformations of
XMLdocuments. It aims at providing
XSLTextensions in order to:
Describe straightforwardly simple transformations (e.g., removing elements, replacing attribute names, merging documents, applying regular expression substitution);
Composing transformations by connecting their (multiple) input and output;
Applying transformations until closure;
Calling external transformation engines.
Transmorpherdescribes the transformation flows in
XML. Input/output channels carry the information, mainly
XML, from one transformation to another. Transformations can be other transformation flows or elementary transformations.
Transmorpherprovides a set of abstract elementary transformations (including their execution model) and one default instantiation of each. Such elementary transformations include external call (e.g.
XSLT), dispatcher, serializer, query engine, iterator, merger, generator and rule sets.
Rule sets describe transformations in a language simpler than
XSLT. It is currently processed by transformation to
XSLT(extended by regular expression substitution).
Transmorpherprovides a set of documented
Javaclasses (which can be refined or integrated into another software) and a transformation flow processing engine. A transformation flow can be described by programming in
Javaor providing an
XMLdescription.
Transmorpherhas been registered by the "Agence de Protection des Programmes" (
APP) under the IDDN.FR.001.430009.000.R.P.2003.000.30620. It is freely available, under the
GPLlicence, since June 2001 at
[http://transmorpher.inrialpes.fr].
An extension of
Transmorphertowards safer transformations consists of attaching assertions to the transformations in a transformation flow in order to tell if a property is assumed, proved or to be checked. This will allow experiments on proving properties of compound transformations.
New Results
The results in 2005 are mainly related to semantic adaptation of multimedia documents (§
Semantic adaptation of multimedia documents
Sébastien
Laborie
Contact
Jérôme
Euzenat
Nabil
Layaïda
WAMteam
When a multimedia document is played on platforms with limited resources (e.g., a mobile phone that can only display one image at once or an interactive display without keyboard), it is necessary to adapt the document to the target device. In order to assess the meaning of adaptation, we
have defined a semantic approach
Adapting amounts to finding this subset of models or, when it is empty, finding a compatible execution as close as possible to the initial execution. For that purpose, we proposed to express the set of possible interpretation by a resolved relation graph. Each relation of this graph could
be a temporal, spatial, spatio-temporal relation... This approach has been applied to the temporal and spatial dimensions based on Allen and
RCCalgebras respectively.
This year, we have followed on our work on semantic adaptation for the
SMIL2.0 language
We have extended this framework to the spatial dimension. Our goal was to find a qualitative spatial representation that computes in a reasonable time a set of adaptations close to the initial document satisfying the adaptation constraints. We have studied different qualitative spatial
representations. Some are very expressive and precise, like the directional representation, but the computational cost of adapting is high. Others, like topological representation, can be adapted quickly, but offer poor expressiveness. Thus, we have introduced a new spatial representation
which trades off between expressiveness and speed. We have compared experimentally each spatial representation on
SMILexamples.
Moreover, we have proposed a method to adapt
SMILdocuments in the spatio-temporal dimension. We have first focussed our framework on multimedia documents without animations (i.e., every multimedia object appears and disappears at the same place). We showed that chaining temporal and spatial adaptation independently would
yield non minimal solutions. Thus, we proposed to use a spatio-temporal representation and show how to calculate proximity between spatio-temporal relations. Then, we have considered multimedia documents with animation. This consists of adding multiple spatial relations to the spatio-temporal
representation and taking into account the property of continuity when the adaptation is applied.
Algorithms for graph-based knowledge representation languages
We are interested in various knowledge representation languages whose formulas can be represented by graphs and where the truth value of a formula in a model-theoretic interpretation can be expressed by some kind of graph homomorphism. Such languages naturally encompass conceptual graphs
or
RDF, but also propositional logics, constraint networks or positive Datalog.
Graph homomorphisms and entailment
Jean-François
Baget
Contact
Faisal
Al-Khateeb
Proofs of soundness and completeness results for entailment in both the conceptual graphs and
RDFformalisms have been unified into a larger framework
R, we can build a logic
Lwhose logical consequence relation is
R, and where the interpretations are formulas of the language (we call such a logic a self-interpreted logic).Any preorder on a countable set of objects is a suborder of graph homomorphism.
Any conjunctive logic having a countable set of formulas is thus equivalent to a self-interpreted logic where both truth values and entailment are expressed via a graph homomorphism. However, the third item above being non constructive, such an equivalent logic cannot be automatically
generated. We have instead adopted another approach, where the semantics of a language is described via graph homomorphisms.
Consider a logic, i.e., a set of formulas, a set of interpretations, and a relation expressing which interpretations are models for a logic.
Both formulas and model-theoretic interpretations can be encoded as graphs (respectively formula graphs and interpretation graphs)Assume a graph homomorphism, called a projection (a mapping of nodes fulfilling particular properties) such that:
The interpretation
Iis a model of the formula
fif and only if the graph of
fprojects into the graph of
I;For every satisfiable formula
f, there exists a model
Iof
fsuch that the graph of
Iprojects into the graph of
f.
In that case, the projection considered is a sound and complete entailment mechanism for the logic involved.
Through this framework, we have characterized a class of logics, including
RDFor conceptual graphs, whose semantics can be expressed via a graph homomorphism. The main interest of this approach is to automatically generate a decision procedure for the entailment relation from the description of the semantics (as graph homomorphism computation).
Path RDF as a query language for RDF
Faisal
Al-Khateeb
Contact
Jean-François
Baget
Querying
RDFgraphs can be reduced to computing entailment and a popular way to compute entailment is to find a projection from an
RDFgraph to another. Another way to query
RDFgraphs is to query for path expressed by regular expressions holding between nodes. The two kinds of queries do not identify the same set of queries (the former allows for full graphs branching and cycling as queries, the latter allows for indetermined lengths of
paths).
We have introduced Path
RDFas a simple extension of
RDFin which relations can be replaced by regular path expressions over relations. This allows to express that there is a path combining relations in a particular way between two resources. For instance, one can query that there is transportation means by any combination of
bus or train between two cities. We have provided the syntax and semantics of Path
RDFand provided an extension of the projection operations between graphs that can be used between Path
RDFgraphs.
By using the above-mentioned framework, we have shown that the new projection is sound and complete for computing entailment (and hence answering queries) when a query in Path
RDFis evaluated against an
RDFgraph. This projection is also sound when used between two Path
RDFgraphs (hence computing query containment). In general, this projection operation is not complete. However, we have shown that if the query is grounded (there is no blank node), then the projection is complete as well.
We are working towards extending Path
RDFprojection to wider classes and to take into account
RDFSand
OWLfragments.
Ontology alignment
In the context of the Knowledge web network of excellence, we are developing activities for structuring the research on ontology alignment at the European level. This has led to numerous activities and results.
Benchmarking
Jérôme
Euzenat
Contact
Marc
Ehrig
Karlruhe Universität
Like last year
In order to evaluate the performance of ontology matching algorithms it is necessary to confront them with test ontologies and to compare the results. The most prominent criteria are precision and recall originating from information retrieval. However, it can happen that an alignment be
very close to the expected result and another quite remote from it and they both share the same precision and recall. This is due to the inability of precision and recall to measure the proximity of the results. To overcome this problem, we have proposed a framework for generalizing
precision and recall
This framework replaces the intersection of retrieved and expected sets by some expression of a similarity between these sets. This similarity must satisfy some constraints (preserving the results obtained by precision and recall, being maximized by the size of the largest set) that
guarantee that precision/recall is one of its instances.
We have instantiated the framework with three different measures and have shown in a motivating example that the proposed measures are prone to solve the problem of rigidity of classical precision and recall. We are now working on providing measures that are semantically justified.
Towards an alignment algebra
Antoine
Zimmermann
Contact
Jérôme
Euzenat
We have designed a theoretical framework to formally define ontology alignments and operations that can be done with them. This framework, based on category theory, considers ontologies and ontology alignments as first order objects independently of the representation language, as
opposed to entity based definitions. Indeed, if ontologies are objects in a category, and morphisms correspond to ontology refinements, then an ontology alignment is a categorical relation between two ontologies, i.e. a pair of morphisms with the same domain. So, if
f: A→Oand
f': A→O'are both ontology refinements, then
<A, f, f'>is an alignment of
Oand
O',
Aapproximates ontologies
Oand
O'and it describes a part of the knowledge that is common to them. We call this structure a V-alignment. With such a characterization, an algebra has been designed which describes what is ontology merging, alignment comparison, composition, union and intersection using well known
categorical constructions.
Concrete categories of ontologies exist in the literature, but fail to express complex alignments, e.g. alignments expressing subsumption relation between concepts of two ontologies. To solve this problem, we investigated two approaches: defining more complex categories or improving the
structure of our categorical alignments. On the one hand, more complex categories enhance the expressivity of the alignments, but the categories lose some interesting properties such as the existence of the merge for any V-alignment and pair of ontologies. On the other hand, simpler
categories, such as the category of theories and theory morphisms in institution theory, can describe complex alignments if the alignment structure is more elaborate. In collaboration with the University of Karlsruhe, we introduced W-alignments: a structure having an additional ontology
containing bridge axioms relating to
Oand
O'by two V-alignments
Similarity for ontology alignment
Jérôme
Euzenat
Contact
Petko
Valtchev
Université de Montréal
In order to be able to align ontologies written in
OWL-Lite, we developed an algorithm (
OLA), adapted from a method for measuring object-based similarity.
OLArelies on a universal measure for comparing the entities of two ontologies that combines in a homogeneous way all the knowledge used in entity descriptions: it deals successfully with external data types, internal structure of classes as given by their properties and
constraints, external structure of classes as given by their relationships to other classes and the availability of individuals. This is an improvement over other methods that usually take advantage of only a subpart of the language features.
The proposed method does not only compose linearly individual methods for assessing the similarity between entities, it uses an integrated similarity definition that makes them interact.
OLAis based on the definition of a distance between entities of two ontologies as a system of equations that has to be solved in order to extract an alignment.
These equations are parameterized by a number of weigths corresponding to the respective importance of different components of ontologies. This year we introduced a preprocessing step which considers the ontologies to align and evaluate the availability of the corresponding features in
order to choose the corresponding weights
This work is developed in collaboration with the University of Montréal.
Dynamic context management in pervasive computing
Jérôme
Pierson
Contact
Jérôme
Euzenat
In a pervasive computing environment, the environment itself is the interface between services and user. Using context information coming from sensors, location technologies and aggregation services, applications adapt their run-time behaviour to the context in which users evolve (e.g.,
physical location, social or hierarchical position, current tasks as well as related information). These applications have to deal with the dynamic integration in the environment of new elements (users or devices), and the environment has to provide context information to newly designed
applications. We study and develop a dynamic context management system for pervasive application. It must be flexible enough to be used by heterogeneous applications and to run dynamically with new incoming devices.
We have designed an architecture in which context information is distributed in the environment. Each device or service implements a context manager component in charge of maintaining its local context. It can communicate with other context manager components: some of them are context
information producers, some of them are context information consumers and some of them are both. We have defined a simple protocol to allow a consumer to identify and determine the producer for the information it needs. Context manager components express their context information using a
OWLontology, and exchange
RDFtriples with each other. The openness of ontology description languages makes possible the extension of context descriptions and ontology matching helps dealing with independently developed ontologies. Thus, this architecture allows the introduction of new components and new
applications without interrupting what is working.
We contributed the development of a demonstrator application for pervasive computing called ambient instant messaging. An appartment is divided in three zones. In each zone, there are different user interface devices. An instant messaging application (such as Jabber,
MSN, or
ICQ) is accessible to users in each room with different modalities and interfaces corresponding to their locations and activities. We worked on a centralized management system for context information used in this demonstrator. It takes information from distributed sensors in
the environment and from a location system, and it provides user "chatting" activity and communication status to applications.
The next step will be to use an alignment service in order to be able to match application needs in terms of context information with that provided by the other devices.
This work is developed in collaboration with France Telecom R&D.