Knowledge based applications require linguistic, terminological and ontological resources. These applications are used to ful ll a set of tasks such as semantic indexing, knowledge extraction from text, information retrieval, etc. Using these resources and combining them for the same application is a tedious task with di erent levels of complexity. This requires their representation in a common language, extracting the required knowledge and designing e ective large scale storage structures o ering operators for resources management. For instance, ontology repositories were created to address these issues by collecting heterogeneous ontologies. They generally o er a more e ective indexing of these resources than general search engines by generating alignments and annotations to ensure their interoperability. However, these repositories treat a single category of resources and do not provide operations for reusing them. The aim of this research is building a large repository of knowledge resources. This repository is a collection of heterogenous resources represented in di erent languages and o ers a set of operations to generate new resources based on the existing ones.
Knowledge extraction and representation is a widely explored research problem.
Most of the proposed solutions to this problem are based on the usage of
auxiliary knowledge resources [
In this paper, we present a model and a taxonomy of abstract operations for managing and extracting knowledge from resources. We consider the possibility of combining these operators to perform complex processes such as semantic enrichment or generating a new resource by merging some other resources. 2
A central point of our approach is to build a repository of knowledge resources. This repository should o er the possibility to store and integrate heterogenous knowledge resources and organize their usage in common context. It should also o er operators for managing and combining these resources. For this we have proposed a three steps methodology: { propose a method and a formalism allowing to represent heterogeneous terminological, linguistic and ontological knowledge resources; { de ne the major representation languages by means of the repository's concepts (Resource, Entity, Relation, etc.); { de ne a set of operations performed on these resources to generate new resources bases on some criteria; { propose multiple implementations per operator depending on the resource type and the representation language; { implement a resources repository to study and resolve scalability problems that arise by evaluating the usability of such a system.
Our approach is not focused on a particular domain, it aims to represent di erent resources from diverse domains and manipulate them using di erent operations. We distinguish two categories of resources. The rst category is about autonomous resources like ontologies, corpora or terminologies. These resources are widely used in multiple applications of knowledge management. The second one represents enrichment resources like annotations or alignments. They link two or more autonomous resources and they result from the application of a process on autonomous resources. 3
For managing heterogeneous resources in large knowledge repositories we need to resolve the problem of resources representation and storage at rst and then address the problem of de ning and implementing resources management operators (collected from existing approaches and classi ed by type such as alignment operators, annotation services, translation mechanisms, etc.).
Some large repositories have been created to o er a more e ective indexing for
knowledge resources than common search engines. For example, Swoogle1
indexes more than 10 000 ontologies; DAML repository2 provides search based
on ontology components (classes, properties, . . . ) or metadata (URI, funding
source, . . . ); BioPortal3 has similar searching and browsing tools [
A repository containing heterogeneous types of knowledge resources is needed. Hence, multiple languages for representing these resources are required. For this purpose, it is necessary to develop a set of knowledge resources operators that can import, export and process these resources while keeping a trace of their origin (the provenance of the resources, for example externally imported or generated from the combination of multiple ones). 3.2
There are many models for knowledge representation, but they usually focus
on one or two aspects only: ontological, terminological, lexical, textual,
documentary, etc. It is more di cult to nd models representing various aspects of
knowledge or resources of di erent kinds. For the integration of heterogeneous
resources, [
In our case, we need to preserve the originality of all resources and treat them within their original context and representation language. This is why we propose a meta-model treating a resource as an entity in the repository. Each resource can have di erent derivations which are also resources represented in di erent languages. 3.3
In the context of mapping linguistic and ontological resources, [
Since there exist many di erent (and incompatible) ways to express knowledge in resources (from formal logic to semi-formal or natural languages). Moreover, the same resource may be involved in processes that can only handle speci c representation formalisms. For instance, an ontology alignment algorithm might be implemented for OWL ontologies, while another algorithm might be about resources in a WordNet-like model. It can be the same for other processes like automated text annotation, multilingual text alignment, word sense disambiguation, etc.
We have proposed a MOF-based model9 to unify the representation of
heterogeneous resources in a common formalism [
Depending on the user's needs, a resource in the repository can be represented di erently using multiple languages, each language uses a subset of the resource's entities and link them in a di erent way compared to another language (for example, a class hierarchy representation links the concepts of an ontology using the subClassOf relation which leads to a di erent derivation of this resource, otherwise a semantic network representation of that resource will lead to the use of another set of relations). Table 1 shows some example of languages that have been described in the current version of the repository.
For example, to represent an ontology we can focus on the hierarchy of classes if we need it in a task of classi cation. We can also represent the same ontology by focusing on axioms and complex expressions using logics if we need it for a reasoning task. 5
The aim of a resources repository is not only to collect heterogenous knowledge resources but especially to o er instruments for reusing them. In order to formalize the de nition of processes over these resources, we have de ned a set of generic primitive operations. We represented then an abstract class of operators in the repository's ontology in order to manage multiple implementations for each operator and to represent restrictions about each implementation. We de ne a process as a sequence of operators applied on resources' derivations. By means of processes descriptions we managed to construct a process dictionary that stores each instance of a process and apply it each time there is an 10 http://cui.unige.ch/isi/onto/tok/OWL Doc/ evolution in the involved resources. Therefore, we must develop a subsequent meta-operators. The de nition of these operators depends on the treatment of the resources. 5.1
These are the basic construction operators for representations. The abstraction and rei cation operations create the resources in the repository and map them to their original derivation in the repository (representation of the resource in its original language). Language mapping operations creates new derivations in other languages.
Importation or abstraction We denote by iRL the import operation that produces an instance of a resource R in the resources repository and by creating the content of the resource in its original languageL. This operation can be followed by a derivation which produces a derivation of the resource in a representation language.
Exportation or rei cation We denote by eRL the export operation that transforms a derivation of a resource R expressed in a language L and its metadata into an external le in a certain formalism related to the derivation's language. Rei cation is generally used at the end of a process (sequence of operations) to produce the new resource. Consequently this operator can have as much instances as the possible combinations from the representation languages implemented in the repository (for example OWL, UML, DL, Graphs, etc.) to the possible required formats (txt, xml, rdf, ttl, n3, etc.).
Derivation This abstract operator is used to create new representations of a resource in di erent languages (represented already in the repository). For instance, an UML class diagram could be derived into a Class diagram representation, then mapped to WordNet-like lexical ontology model (by dropping all the associations except part-of and subclass ). Since a derivation may \forget" information, in general L2L1 is not the inverse of L1L2 . It is not always necessary to preserve the entire contents of a resource when deriving a new representation of its content (this can be compared to generating a view in the relational approach). In particular, if the representation language is less expressive than the original language it is obvious that some knowledge will be lost. 5.2
The enrichment operations generate new alignments or annotations on existing
resources. They are generally based on sophisticated algorithms (more precisely
heuristics) and use auxiliary resources like lexical ontologies.
Alignement Alignment allows to express explicitly the correspondences
between resources [
The signature of this operator is :
OpAlign : L1; L1; [L2; : : :] ! (L1; Lal)
Lal is a language that includes the alignment relations used to represent the correspondences (v; , etc.), OpALIGN is the operator used for the alignment.
A typical example of the need for simpli ed languages is the ontology
alignment task. Most of the current alignment algorithms can align ontologies
represented in OWL language, but they do not take advantage of all the semantics
expressed in such ontologies [
Annotation The annotation operator is used to describe elements of a resource R1 in terms of a resource R2, this description is through adding a set of relationships between entities of these resources according an annotation language.
The signature of this operator is:
OPAnn : L1; L2 ! L1; L2; Lann where L1 is the language of the resource's derivation to annotate and L2; : : : are the languages of the resources' derivations that serve as reference in the annotation. Lann is the annotation language. For example, word sense disambiguation is a kind of annotation operation. Starting from a natural language text and a reference lexical ontology (and possibly other resources), it produces a set of correspondences between the text words and their meanings (the concepts of the ontology). 5.3
These operations are intended to produce new resources' derivations by selecting and combining entities of one or more resources.
Selection This type of operation selects entities from a resource's derivation to generate a new resource's derivation in the same language. This ltering is speci ed by a boolean function applied on each entity. The computation of the ltering function for a resource entity may depend on other entities from the same resource or others entities associated to it by means of annotations or alignments. In addition, the selection may generate a natural alignment between entities of the original and new resource's derivations. Each selected entity is associated to its original entity.
The signature of a selection operation is of the form
OpSel : L1 ! L1 where L1 is the language of the input resource and the resulting selection.
For instance, in a description logic ontology, this operator can select individuals in the ABox (Assertional Box), leaving the TBox (Terminological Box) untouched (as in a database selection) or it can select a subset of the TBox, and hence drop the ABox entities that depend on unselected TBox concepts or roles (as in a database projection).
Composition Composition operations may be applied on alignments and annotations. It is an operator that generates new derivation of the composed resources in the same language.
The composition of two alignment resources (from S1 to S2 and from S2 to S3 results in a new alignment resource from S1 to S3. The semantics (relation type) of the resulting alignment depends on the relation types of the given alignments. If A1 and A2 have the same relation type R and R is transitive, then A1 A2 has type R.
Merge The idea of the merge operation is to build a new resource by taking
all the entities of two given resources [
The signature of the merge operator has the form:
OpMerge : L1; L1; [Lal] ! (L1)[Lal]
This operator takes as parameters a list of resources represented in the same language and uses auxiliary resources such alignments between them. Merging two alignments or annotations can occur only if they are about a common resource. First, for each resource Ri to merge, we must consolidate and merge all correspondences whose source is Ri and represented in the same alignment language Lal. A multiple inputs and outputs alignment resource is constructed and represented within the language Lal. Both the set of resources to merge and the constructed alignment provide required ingredients for the merge. 6
Our main objective is to build a large repository for integrating heterogeneous resources represented in di erent languages. We have identi ed three major steps for implementing this repository. First we have de ned an upper level model for representing knowledge resources and dealing with di erent representation languages. Then we have de ned a set of abstract operators having multiple implementations in order to combine the content of the repository and generate new resources from existing ones. We will focus on de ning examples and a set of use cases in order to validate this approach and nally address the scalability issues. To ensure the usage of the repository by means of knowledge representation and resources management operators we are currently focusing on the following issues: (1) de ne a model for each processing task using resources, these tasks models should be the result of a re ection on a set of use cases; (2) de ne and implement a set of heuristics for the automatic detection of entity mappings to construct alignments between resources during the execution of any task.
For the third part of this research we will focus on the experimentation and the implementation of the repository. An implementation of a prototype is intended to prove the research results and de ne software requirements by studying the available technologies and APIs that can be used. For instance, we should address the following issues: { evaluation and study of RDF storage approaches must be driven to select the best storage API to use for storing knowledge resources especially focus on the scalability issues; { for the sake of generality we should investigate the possibilities for providing resources management operators using web services; { de ne the interface that should be used for the repository's portal and the de ne the criteria of accessibility and user pro les.