In the past, due to the lack of storage capacity and netDue to the information growth, distributed environments work bandwidth, especially in most desktop computers, dyare offered as a feasible and scalable solution where Peer- namic sharing of information from end user machines was to-Peer networks have become more relevant. They bring prohibitively costly. As a consequence, networks of commany advantages as high flexibility for peers to join or leave puters were mostly reduced to set of powerful connected the network dynamically, scalability, autonomy and high re- servers. In this configuration, it is relatively simple to know silience against peer failures. However, the use of propri- which servers are available and which information is availetary interfaces within the network and the requirement that able where to whom. This is also the typical architecture peers must implement them to join makes P2P networks un- in business coalitions where several companies share their able to interact with other systems and environments, iso- assets within a network of e.g. partners. lating the network as a whole. In this paper, we report on a On the other hand, with the boom of Web-based filesolution based on a proxy-based architecture and semantic sharing services (e.g., Napster, Gnutella, Morpheus), peermappings in order to allow the sharing of content between to-peer (P2P for brevity) networks have become more relpeers within a P2P network with content from other systems evant. The advantages of the P2P approach include: high outside the network. flexibility for peers to join or leave the network dynamically, scalability (recently it was shown that for really large networks, a hybrid solution with super-peers scales better [13]), Categories and Subject Descriptors autonomy as peers do not relinquish control over their reH.3.3 [Information Systems]: Information Storage and sources and high resilience against peer failures. The main Retrieval-Information Search and Retrieval ; C.2.4 [Computer disadvantages are that P2P networks require constant manSystems Organization]: Computer-Communication Net- agement, as peers join and leave continuously (producing an works-Distributed Systems extra load on the network and may slow response times during search) and the use of proprietary interfaces within the network (what usually means that only peers implementGeneral Terms ing such interfaces can query for or provide content in the Distributed Environments, Interoperability network). Obviously, peers must implement specific P2P network interfaces in order to join them. This means an extra effort
Often, learning object providers do not want to abandon control over their resources to a common server, even among the members of a coalition. The same concern about abandoning control also often applies to individuals, who may not want to give away their content to any centralized repository. Distributed environments have shown to be a feasible solution for interconnection, integration and access to large amounts of information that deal with this issue. P2P networks are an example of the impact this distribution of information might have in the sharing of information. In such networks, peers can offer various services to the user that range from search and delivery to personalization and security services. In addition, they present a solution to the information growth where every learning resource provider offers its information but does not loose the control over it.
The Edutella P2P network [
It is important to note that we consider in this paper only the sharing of metadata about LO’s. While this metadata is typically available, the learning object itself might not be. Therefore, we do not deal with negotiations for the actual use of LO’s by users here.
Admittedly, providing transparent access to all available repositories would be easy if all players would use the same metadata profile, query language, storage layer and communication protocol. However, this is not going to happen in the very near future due to the lack of a standard and the proprietary solutions adopted by most of them.
In the following, we explain what requirements LO’s repositories must satisfy in order to achieve interoperability and which are the assumptions within our network.
Common Communication Protocol and Interface
Repositories provide different access methods and
interfaces, over, among others, Web Services, different
Remote Procedure Call methods, HTTP forms or even
other appropriate solutions. In order to be able to
communicate to each other, it is needed that they agree
on a common protocol and a common interface. In this
paper, we built on the methods specified in the
Simple Query Interface [
At the lower levels of data management, metadata is stored in different kinds of repositories, such as relational databases, RDF repositories, file systems, XML stores, etc. On top of this lower level, repositories expose their content through different search and query languages. Some examples are SQL, XQuery, QEL or CQL. In our system we have several wrappers implemented in order to provide access to the most common repositories (relational databases, RDF repositories, RDF files, etc...). For all them, the wrapper receives a query in QEL and transforms it into the local query language.
Although IEEE LOM [
It is important to notice that although we assume the configuration described above it could be perfectly possible to use a different query language than QEL, a different communication protocol than Web Services and a different interface than SQI though our implementations currently do not support it. 4.
P2P networks are dynamic networks where peers can act as server and client indistinctly and peers might freely join and leave the network over the time. Obviously, peers must implement the specific P2P network interface in order to connect to it. This means an extra effort for systems willing to connect which already have a different query interface. This barrier makes P2P networks unable to interact with other systems and environments.
In order to solve this problem, we based our solution in proxies that are used to connect peers in a P2P network with the “outside” world. This proxies bridge two systems with different capabilities by means of implement the protocol and/or interface supported by each system respectively. This way, a proxy is able to to forward requests and responses from one system to another.
Nowadays, many systems provide their services/resources via Web Services and therefore we implemented proxies able to bridge the proprietary JXTA/Edutella protocol1 and interface into a Web Service protocol based on the Simple Query Interface.
Taking the P2P network as a reference, there are two
different desirable scenarios [
Client EDUTELLA NETWORK
Consumer
P2P 1Query 1 Web Service
Query 4 Web Service Answer SQI Consumer
Proxy we would like to offer the content of a P2P network on a web site. The first solution would be to make the web server join the P2P network. However, the load of the server would increase considerably and even some problems could arise in case the server wants to provide content from more than one network (it would need to join all of them). A cleaner solution (and the one we follow in this paper) is to forward the query from the Web Site to the P2P network by means of e.g. Web Services and retrieve the answer with the same mechanism. 2. An external provider wants to offer content to the P2P network. We assume that providers that have already implemented a Web Service based interface will not want to spend time and money in developing the proprietary interface of the network. In contrary, they would like to reuse the one they have which would also ease its administration (as only one interface needs to be maintained).
According to these two scenarios, there are two different types of proxies. The former scenario requires the so-called “consumer proxy” and the latter the so-called “provider proxy” (names are assigned according to the role they play). A consumer proxy acts as a mediator between an external client that wants to query the network and the P2P network itself. A provider proxy acts a mediator in order to provide the content of an external provider into the P2P network. 4.1
As described above in scenario 1, in some cases it is needed to be able to query a P2P network without the need of joining it. A consumer proxy is a peer which is part of the P2P network (and therefore it is able to send queries and receive the answers from it) and which is also able to receive requests and send responses using a different protocol and interface. This way, an external client is able to query the P2P network through the proxy.
In our implementation, a consumer proxy mediates between the Edutella/JXTA and Web Service protocols. As depicted in figure 1, it is responsible for 1. Receiving queries from external clients via SQI 2. Forwarding the query to the Edutella network using the JXTA/Edutella interface 3. Collecting the results sent from peers within the network using the JXTA/Edutella interface 4. Forwarding those results to the requester system via
SQI
This simple mechanism allows any system to query the content of the Edutella P2P network without needing to implement its specific interface.
In addition, the proxy can return the results to the client application • Asynchronously. The results are sent to the client as soon as they arrive to the proxy. This is the typical mechanism in distributed environments as not all the results are generated at once but they must be gathered from the different systems in the network. • Synchronously. The results are gathered at the proxy and sent together to the client. Although this is not the intuitive way for a distributed environment it could be desirable in some scenarios (e.g., in mobile devices we do not want our device to receive a new message everytime a new result arrives to the proxy but better ask for new results in a proactive manner). 4.2
In order to fulfil our scenario 2 a second type of proxy has been developed. This provider proxy is a peer connected to the P2P network which also is able to send requests and receive responses by means of a different protocol and interface. Therefore, it is able to forward queries to external providers and receive their answers providing their content to the network.
As in the the case of consumer proxies, our provider proxy mediates between the Edutella/JXTA and Web Service protocol. As depicted in figure 2, it is responsible for 1. Receiving queries from peers in the network using the
JXTA/Edutella interface 2. Forwarding them to the external provider via SQI 3. Receiving the results from the external provider via
SQI 4. Sending them back to the peer that sent the query using the JXTA/Edutella interface
Due to the asynchronous nature of a P2P network, it is possible for the provider proxy to receive the results from the external provider in a synchronous (e.g., in case the external provider is a relational database) or asynchronous (e.g., if the external provider is another distributed environment) way.
S3 S5
S4
In previous sections, there have been described some of the basis for interoperability, namely common protocol and interfaces (or the use of proxies as presented in section 4) and common query language (or the use of appropriate wrappers). Although these elements ensure that two systems are able to talk to each other it does not guarantee that they will be able to understand each other due to the possible use of different schemas/ontologies.
Nowadays, there is a big effort on standardization of
domain ontologies. For example, Dublin Core [
In a distributed network we can distinguish among several
integration possibilities:
• If no virtual and unified schema is assumed in the
network, then systems within the network must provide
pairs of mappings between each two systems. Then the
distributed network can be seen as a directed graph (as
shown in figure 3) in which each arrow represents an
available mapping from one node to another. Then,
they can be applied transitively in order to infer new
mappings which were not explicitly defined. This is
specially useful in P2P networks as it is usually not
possible to enforce a unique and common schema.
Authors in [
UNION SELECT S2.attr1, S2.attr2, S3.attr3, S4.attr4
FROM S2, S3 WHERE S2.attr1 = S3.attr1 S1(attr1, attr2, attr3, attr4)
S2(attr1, attr2)
S3(attr1, attr3, attr4)
In order to provide semantic interoperability in our network, we have developed a module which transforms a query q1 into a query q2 according to the set of mappings specified. This module is intended to work on pairs of mappings without a unified schema or in Local As View integration approaches.
QEL, the language we use in our network, is based on datalog. In addition to standard datalog constructs, QEL includes some built-in predicates. Taking into account that in our network only metadata (in RDF) is queried and exchanged, the most important one is
qel:s(Subject, Predicate, Object)
which according to the QEL specification [
?- qel:s(X, dc:title, ’Artificial Intelligence’). will return all the resources which title is “Artificial Intelligence”. Other useful built-in predicates are qel : like(X, Y ) (“used to determine whether an RDF literal or URI contains a string as a substring”), qel : lessT han(X, Y ) and qel : greaterT han(X, Y ) which are used to compare two RDF literals.
Given this short introduction to the language, let us present the following simple query that we will use for our examples in the rest of the section: @prefix qel: <http://www.edutella.org/qel#>. @prefix dc: <http://purl.org/dc/elements/1.1/>. @prefix lom: <http://ltsc.ieee.org/2002/09/lom-rights#>. ?- qel:s(X, dc:title, Title), qel:s(X, dc:description, Description), qel:s(X, dc:creator, Creator), qel:s(X, lom:cost, Cost), qel:s(X, dc:subject, Subject).
This query retrieves all the resources with title, description, creator and subject attributes from Dublin Core and the cost from LOM. The first lines of the query with prefix “@” define the namespaces.
Given such a query, we identified the following
requirements
• The query specifies a property2 that does not exist in
the source but the source has an equivalent property
which could be used instead of. For example, if one
data source has its own schema where it uses the
property “abstract” instead of the property “description”
from the Dublin Core standard.
• The query specifies a property and one value according
to a specific taxonomy and the source uses a different
taxonomy. For example, if the query searches for
resources with dc:subject following the ACM
classification [
In order to satisfy these requirements we developed a
module that performs two types of mappings and one extra
2In the paper we will use property and attribute indistinctly
3Here we assume that only conjunctives queries are sent.
Edutella and QEL support disjunctive queries but we will
omit them here because of simplicity
transformation: property mapping, property-value mapping
and default value transformation (see table 1 for the whole
list of mappings and [
A property mapping specifies how one property in the query must be reformulated. Its general syntax is
(X, p1, Z) ← (X, p2, Z)
When the mapping module receives a query that contains the triple qel : s(X, p1, Z) it rewrites it into qel : s(X, p2, Z). Using our example query and taken into account the requirement in which the source does not contain the property “dc:description” but “own:abstract” (where own stands for their local namespace), it is possible to define the following mapping:
(X, dc : description, Z) ← (X, own : abstract, Z)
This mapping is currently a 1-to-1 mapping, that is, there
is only one triple at each side of the mapping (separated by
the left arrow) but it is also possible to specify 1-to-2, 2-to-1
and 2-to-2 mappings (see table 1). For example, suppose
the author in the source is encoded using the property full
name from the vcard ontology [
The mapping described above assumes that one property is completely mapped into another one. However, mapping can be brought to the granularity of values. A propertyvalue mapping applies only when a query contains not only a specific property but also a specific value for that property and then both of them map into other (possibly the same) property and value. Its syntax is
(X, p1, v1) ← (X, p2, v2)
For example, assume that our example query uses the ACM classification in the property “dc:subject” and our source does have the property “dc:subject” but annotated with the Dutch Basic Classification taxonomy. We could use several mappings of the form (X, dc : subject,0 Sof tware/P rogramming Languages0) ← (X, dc : subject,0 Computer Science/P rogramming Languages0) to specify how the different values from the ACM taxonomy map into the Dutch Basic Classification.
In the same way as the property mapping, it is possible to extend this 1-to-1 to 2-to-1, 1-to-2 and 2-to-2 mappings.
Property and property-value mappings provide rules which define how some triples are reformulated into another triples.
The way default values work is a bit different. The properties especified in default values do not exist in the source repository and therefore they must be removed (not just reformulated) in the new query.
Following this approach, when a query is received by our
mapping module, if there exists in the query any occurrence
Mapping type
1-to-1 property mapping
1-to-1 property-value-value mapping
2-to-1 property mapping
2-to-1 property-value mapping
1-to-2 property mapping
1-to-2 property-value mapping
2-to-2 property mapping
2-to-2 property-value mapping
Default value
of a property specified in the default values, this ocurrence is in the form of television programs, radio shows, instructional
temporarily removed. This way, the query is sent to the local material, web sites, physical distributions on CDs and videos
repository without that property (otherwise the query would etc. The educational resources are expressed using the
RDFfail) and a resultset is returned. However, this resultset still binding of IEEE/LOM, qualified Dublin Core and some
exdoes not contain the default values that were requested (the tensions specific to UR. For storage solution we use
SCAMproperties previously removed) and therefore they must be [
For example, following with our example query, suppose out. Queries in the QEL language is directly translated to
that our source repository does not have the property “lom:cost” SQL and passed to the database. We use a library called
but all the resources in the repository are free of charge. We GETSQL4 for doing the translation. GETSQL can handle
can then define the following default value disjunction, rules, most of the constraints, outer join, and
also the retrieval service [
(lom : cost ←0 N o0) The peer does two kinds of semantic mappings: This way, any triple in the query referring to the property “lom:cost” would be removed before the query is sent to the repository and added subsequently to the returned resultset together with the default value “No”.
Using the elements described previously in this paper,
it has been possible to bring interoperability to Edutella
providers (Media Library, Nature and Technology and
Confolio System), external consumers (ARIADNE) and
external providers (ULI and ARIADNE) within the context of
the EU Network of Excellence Prolearn [
In this section we introduce some of the existing peers which are currently connected directly to the Edutella P2P network. All of them are mapped into a common schema (LAV approach). In addition, all the modules used by the peers (including the mapping modules) make use of the highly configurable java library provided in the Edutella project and that it is available from the Edutella CVS (at http://edutella.jxta.org/). Some of them, like the mapping modules, are implemented in a way that they can also be used in environments where the Edutella P2P network is not present. 6.1.1
Media Library - Swedish Educational Radio and Television
The media library is a joint project between the KMR
group-[
Nature and Technology - Swedish National Agency for Education
This peer provides educational resources provided by Swedish highschool teachers to be shared. The educational resources are of several kinds, e.g. experiments, articles, projects, tests. The resources are expressed with simple Dublin Core and a few extra properties. The classification is done via the SLI community’s taxonomy. The storage solution is SCAM and the GETSQL library is used to translate the QEL queries to SQL here as well. The mappings provided are similar as for the media library with the difference that here we translate from qualified to simple Dublin Core. 6.1.3
Confolio system - portfolios hosted at Royal Institute of Technology
The confolio system is another application built on top of SCAM. It provides users with a directory structure where 4GETSQL is expanded to ’Generic Edutella query language Translation to SQL’ and is available from Edutella CVS. Its modular design allows the database and database schema to be changed easily.
QI
S KPS KPS
ARIADNE KNOWLEDGE
POOL
KPS Media Library
Nature and Technology SQI
Ariadne Provider Proxy
Confolio System
Edutella User WebServiceSQI
ULI Provider Proxy they can upload files, store links, or just pure metadata e.g. events, persons, books, concepts. The confolio system allows you to define what kind of metadata that you want to provide for a specific type of resource (multiple types on a resource results in the union of the metadata). A consequence of this flexibility is that there is no specific schema since it depends on what people are doing with the system, on the other hand there is a lot of reuse of Dublin Core, IEEE/LOM etc. 6.2
In its aim to facilitate both academic education and corporate training, the ARIADNE Foundation supplies its members with tools and methodologies for producing and reusing learning objects. The core of these services is a distributed network of repositories that replicate content and metadata. Doing so, each node contains a copy of all metadata instances. The LO’s however can only be replicated to other servers if no download restrictions apply to them. This infrastructure, also known as the Knowledge Pool System, enables the ARIADNE user community to transparently manage learning objects.
The ARIADNE Knowledge Pools offers a client-server approach, where applications can query the ARIADNE knowledge pool through a web services layer. As metadata is replicated in this distributed network, there is no need to federate queries in ARIADNE. However, in order to provide these applications with access to a bigger pool of learning objects, a federated search layer has been built on top of different SQI targets (Edutella, ARIADNE, Merlot, Celebrate) enabling applications to search beyond the borders of the ARIADNE knowledge pool.
Technically, ARIADNE contains metadata information of LO’s stored in a relational databases, and the results returned while answering queries is delivered as XML instances following the LOM standard. This information is transformed by means of XSLT stylesheets in order to conver it into RDF.
Currently, ARIADNE is connected to Edutella as
external consumer (Edutella content is offered in ARIADNE)
and as provider (the content of ARIADNE is offered in
Edutella) [
The course repositories ULI [
Though the courses usually differ in the kind and amount of learning materials they use, their use of learning resources is surprisingly homogeneous. The average course is divided in 6 to 7 units or knowledge modules which themselves can be split into 3 to 7 learning resources. This leads to an average number of about 35 learning resources per course, with a learning resource being the slides of the lecture, a video or any other set of pages dealing with on subject.
Technically, ULI repositories are based on RDF files with Dublin Core and LOM metadata and they are currently accessed by means of the RDQL query language. In addition, mappings and default values have been specified in order to convert to a global schema. 7.
In [
[
In this paper, we showed how by means of proxies and semantic mapping it is possible to connect a P2P network like Edutella with other systems outside the network. These proxies provides the necessary mediation between the different protocols and intefaces and semantic mappings overcome the problem of schema heterogeneity. Both together allow external systems to query and provide content in the network avoiding the isolation of P2P networks from the rest of the world.
In this paper we have focused the interoperability problem on search. However, although this is of course the most important service, there are still some other issues that must be researched. One of the main topics we plan to research on in the future is distributed ranking algorithms. Currently, a lot of research has been done around web ranking (e.g., on the Web) and merging of ranking lists (e.g., on meta-search engines). However, the former assumes that relationships of the form of links exist among resources in different repositories and the latter assumes that there exist overlapping in the content different repositories offer and rank. Unfortunately, this does not apply in a P2P network and the only existing measures are based on trust/reputation of the peers.
In addition, a challenge for the Local As View mappings
described in this paper is how they would work in
combination with Edutella Retrieval. Edutella Retrieval [
The authors thank Stefaan Ternier for useful discussions and his work in connecting the ARIADNE system. The research of Olmedilla was partially supported by the IST project ELENA (http://www.elena-project.org). The research of Palmer was partially supported by the IST Network of Excellence PROLEARN (http://www.prolearn-project.org/).