The workshop would not be possible without contributions of many persons and institutions. We are very thankful to the organisers of the ESWC 2011 conference for providing us with an opportunity to organize the workshop, for their excellent collaboration, and for looking after many important logistic issues. We are also very grateful to the members of the program committee for their commitment in reviewing the papers and assuring the good quality of the workshop program. We also thank the authors for their invaluable contributions to the workshop by writing, revising and presenting their papers. Of course, great appreciation of her time and expertise goes to our keynote speaker Vania Dimitrova. We also want to express our strong gratitude to the publishers of CEUR for publishing the Linked Learning 2011 workshop proceedings, to the European Commission (EC) and the EC-funded research project mEducator for sponsoring the best paper award and to the EasyChair developers for supporting the submission and review process.
May 2011, Stefan Dietze, Mathieu d'Aquin, Dragan Gasevic, Miguel-Angel Sicilia
Soeren Auer, University of Leipzig, Germany Panagiotis Bamidis, Aristotle University of Thessaloniki, Greece Charalampos Bratsas, Aristotle University of Thessaloniki, Greece Dan Brickley, W3C & Free University of Amsterdam, The Netherlands Vania Dimitrova, University of Leeds, UK John Domingue, The Open University, UK & Semantic Technologies Insitute International, Austria.
Nikolas Dovrolis, Democritus University of Thrace, Greece Marek Hatala, Simon Fraser University, Canada Jelena Jovanovic, University of Belgrade, Serbia Eleni Kaldoudi,Democritus University of Thrace, Greece Tomi Kauppinen, University of Münster, Germany Carsten Keßler, University of Münster, Germany Effie Lai-Chong Law, Leicester University, UK & ETH, Zurich, Switzerland Nikos Manouselis, Greek Research and Technology Network, Greece Dave Millard, University of Southampton, UK Evangelia Mitsopoulou, St George's University London, UK Wolfgang Nejdl, L3S Research Center, Germany Mikael Nilsson, Royal Institute of Technology, Sweden Carlos Pedrinaci, The Open University, UK Davide Taibi, Institute for Educational Technologies, Italian National Research Council, Italy.
Vlad Tanasescu, University of Edinburgh, UK Fridolin Wild, The Open University, UK Martin Wolpers, Fraunhofer FIT.ICON, Germany
Hong Qing Yu, The Open University, UK
Fouad Zablith, Miriam Fernandez and Matthew Rowe Knowledge Media Institute (KMi), The Open University Walton Hall, Milton Keynes, MK7 6AA, United Kingdom
ff.zablith, m.fernandez, m.c.roweg@open.ac.uk Abstract. The aim of this paper is to introduce the current e orts toward the release and exploitation of The Open University's (OU) Linked Open Data (LOD). We introduce the work that has been done within the LUCERO project in order to select, extract and structure subsets of information contained within the OU data sources and migrate and expose this information as part of the LOD cloud. To show the potential of such exposure we also introduce three di erent prototypes that exploit this new educational resource: (1) the OU expert search system, a tool focused on nding the best experts for a certain topic within the OU sta ; (2) the Buddy Study system, a tool that relies on Facebook information to identify common interest among friends and recommend potential courses within the OU that `buddies' can study together, and; (3) Linked OpenLearn, an application that enables exploring linked courses, Podcasts and tags to OpenLearn units. Its aim is to enhance the browsing experience for students, by detecting relevant educational resources on the y while reading an OpenLearn unit. 1
Introduction
The explosion of the Linked Open Data (LOD) movement in the last few years
has produced a large number of interconnected datasets containing information
about a large variety of topics, including geography, music and research
publications among others. [
The movement is receiving worldwide support from public and private sectors
like the UK1 and US2 governments, international media outlets, such as the
BBC [
Despite its relevance and the support received in the last few years, very few pieces of work have either released or exploited LOD in the context of education. One of these few examples is the DBLP Bibliography Server Berlin,4 which provides bibliographic information about scienti c papers. However, education is principally one of the main sectors where the application of the LOD technologies can provoke a higher impact.
When performing learning and investigation tasks, students and academics have to go through the tedious and laborious task of browsing di erent information resources, analysing them, extracting their key concepts and mentally linking data across resources to generate their own conceptual schema about the topic. Educational resources are generally duplicated and dispersed among different systems and databases, and the key concepts within these resources as well as their inter and intra connections are not explicitly shown to users. We believe that the application of LOD technologies within and across educational institutions can explicitly generate the necessary structure and connections among educational resources, providing better support to users in their learning and investigation tasks.
In this context, the paper presents the work that has been done within The Open University (OU) towards the release and exploitation of several educational and institutional resources as part of the LOD cloud. First, we introduce the work that has been done within the LUCERO project to select, extract and structure subsets of OU information as LOD. Second, we present the potential of this data exposure and interlinking by presenting three di erent prototypes: (1) the OU expert search system, a tool focused on nding the best experts for a certain topic within the OU sta ; (2) the Buddy Study system, a tool focused on exploiting Facebook information to identify common interests among friends and recommend potential courses within the OU that `buddies' can study together, and; (3) Linked Open Learn, an application that enables exploring linked courses, Podcasts and tags to OpenLearn units.
The rest of the paper is organised as follows: Section 2 presents the state of the art in the areas of LOD within the education context. Section 3 presents the work that has been done within the LUCERO project to expose OU data as part of the LOD cloud. Sections 4, 5 and 6 present example prototype applications that consume the OU's LOD for Expert Search, Buddy Study and Linked OpenLearn respectively. Section 7 describes the conclusions that we have drawn from this work, and section 8 presents our plans for future work. 2
Related Work
While LOD is being embraced in various sectors as mentioned in the previous
section, we are currently witnessing a substantial increase in universities adopting
4 http://www4.wiwiss.fu-berlin.de/dblp/
the Linked Data initiative. For example, the University of She eld's
Department of Computer Science5 provides a Linked Data service describing research
groups, sta and publications, all semantically linked together[
With the increase of the adoption of LOD publishing standards, the exchange of data will be much easier, not only within one university, but also across the LOD ready ones. This enables, for example, the comparison of speci c quali cations o ered by di erent universities in terms of courses required, pricing and availability. 3
The Open University Linked Open Data The Open University is the rst UK University to expose and publish its organizational information in LOD.11 This is accomplished as part of the LUCERO project (Linking University Content for Education and Research Online)12, where the data extraction, transformation and maintenance are performed. This enables having multiple hybrid datasets accessible in an open way through the online access point: http://data.open.ac.uk.
The main purpose of releasing all this data as part of the LOD cloud is that members of the public, students, researchers and organisations will be able to easily search, extract and, more importantly, reuse the OU's information and data. 3.1
Creating the OU LOD Detailed information about the process of LOD generation within the OU is available at the LUCERO project website.12 We brie y discuss in this section 5 http://data.dcs.shef.ac.uk 6 http://prism.talis.com/manchester-ac 7 https://mmb.ilrt.bris.ac.uk/display/ldw2011/University+of+Bristol+data 8 http://ldfocus.blogs.edina.ac.uk/2011/03/03/university-buildings-as-linked-datawith-scraperwiki 9 http://data.ox.ac.uk 10 http://www.lodum.de 11 http://www3.open.ac.uk/media/fullstory.aspx?id=20073 12 http://lucero-project.info the steps involved in the creation of Linked Data. To achieve that, the main requirement is to have a set of tools that generate RDF data from existing data sources, load such RDF into a triple store, and make it accessible through a web access point.
Given the fact that the OU's data repositories are scattered across many departments, using di erent platforms, and subject to constant update, a wellde ned over ow needs to be put in place. The initial work ow is depicted in Figure 1, and is designed to be e cient in terms of time, exibility and reusability. The work ow is component based, and the datasets characteristics played a major role in the implementation and setup of the components. For example, when the data sources are available in XML format, the XML updater will handle the process of identifying new XML entities and pass them to the RDF extractor, where the RDF data is generated, and ready to be added to (or removed from) the triple store. Finally the data is exposed to the web, and can be queried through a SPARQL endpoint.13
The scheduler component takes care of initiating the extraction/update
process at speci c time intervals. This update process is responsible for checking
what was added, modi ed, or removed from the dataset, and accordingly
applies to the triple store the appropriate action. Having such a process in place
is important in the OU scenario where the data sources are continuously
changing. Another point worth mentioning is the linking process that links entities
coming from di erent OU datasets (e.g., courses mentioned in Podcast data and
library records), in addition to linking external entities (e.g., course o erings in
a GeoNames de ned location14). To achieve interlinking OU entities,
independently from which dataset the extraction is done, we rely on an Entity Named
System, which generates a unique URI (e.g., based on a course code)
depending on the speci ed entity (this idea was inspired from the Okkam project15) .
Such unique URIs enable a seamless integration and extraction of linked entities
within common objects that exist in the triple store and beyond, one of the core
Linked Data requirements [
The Data Data about the OU courses, Podcasts and academic publications is already available to be queried and explored, and the team is now working to bring together educational and research content from the university's campus information, OpenLearn (already available for testing purposes) and library material. More concretely, data.open.ac.uk o ers a simple browsing mechanism, and a SPARQL endpoint to access the following data: 13 http://data.open.ac.uk/query 14 http://www.geonames.org 15 http://www.okkam.org
Fig. 1. The LUCERO Work ow
{ The Open Research Online (ORO) system16, which contains information
about academic publications of OU research. For that, the Bibliographic
Ontology (bibo)17 is mainly used to model the data.
{ The OU Podcasts,18 which contain Podcast material related to courses and
research interests. A variety of ontologies are used to model this data,
including the W3C Media Ontology,19 in addition to a specialised SKOS20
representation of the iTunesU topic categories.
{ A subset of the courses from the Study at the OU website,21 which provides
courses information and registration details for students. We model this data
by relying on the Courseware,22 AIISO23 and GoodRelations ontologies [
Furthermore, there are other sources of data that are currently being processed. This includes for example the OU list of provided publications, the 16 http://oro.open.ac.uk 17 http://bibliontology.com/speci cation 18 http://podcast.open.ac.uk 19 http://www.w3.org/TR/mediaont-10 20 http://www.w3.org/2004/02/skos 21 http://www3.open.ac.uk/study 22 http://courseware.rkbexplorer.com/ontologies/courseware 23 http://vocab.org/aiiso/schema library catalogue, and public information about locations on the OU campus (e.g., buildings) and university sta . 4
The OU Expert Search Expert search can be de ned as the task of identifying people who have relevant expertise in a topic of interest. This task is key for every enterprise, but especially for universities, where interdisciplinary collaborations among research areas is considered a high success factor. Typical user scenarios in which expert search is needed within the university context include: a) nding colleagues from whom to learn, or with whom to discuss ideas about a particular subject; b) assembling a consortium with the necessary range of skills for a project proposal, and; c) nding the most adequate reviewers to establish a program committee.
As discussed by Yimam-Seid and Kobsa [
OU researchers, students and lecturers constantly produce a plethora of documents, including for example conference articles, journal papers, thesis, books, reports and project proposals. As part of the LUCERO project, these documents have been pre-processed and made accessible as LOD. The purpose of this application is therefore to exploit such information so that OU students and researchers can nd the most appropriate experts starting from a topic of interest.24 4.1
Consumed Data This application is based on two main sources of information: (a) LOD from the Open Research Online system, and (b) additional information extracted from the OU sta directory. The rst information source is exploited in order to extract the most suitable experts about a certain topic. The second information source complements the previous recommended set of experts by providing their corresponding contact information within the OU. Note that sometimes, ex-OU members and external collaborators or OU researchers may appear in the ranking of recommended experts. However, for those individuals, no contact information is provided, indicating that those experts are not part of the OU sta .
As previously mentioned, the information provided by Open Research Online contains data that describe publications originating from OU researchers. In particular, among the properties provided for each publication, this system exploits the following ones: a) the title, b) the abstract, c) the date, d) the authors and, e) the type of publication, i.e., conference paper, book, thesis, journal paper, etc. 24 The OU Expert Search is accessible to web15.open.ac.uk:8080/ExpertSearchClient OU sta at: http://kmi
To exploit this information the system performs two main steps. Firstly when the system receives the user's query, i.e., the area of expertise where a set of experts need to be found (e.g., \semantic search"), the system uses the title and abstract of the publications to nd the top-n documents related to that area of expertise. At the moment n has been empirically set to 10.
Secondly, once the top-n documents have been selected, the authors of these documents are extracted and ranked according to ve di erent criteria: (a) original score of their publications, (b) number of publications, (c) type of publications, (d) date of the publications and, (e) other authors of the publication.
The initial score of the publications is obtained by matching the user's keyword query against the title and the abstract of the OU publications. Publications that provide a better match within their title and abstract against the keywords of the query are ranked higher. This matching is performed and computed using the Lucene25 text search engine. Regarding the number of publications, authors with a higher number of publications (among the top-n previously retrieved) are ranked higher. Regarding the type of publication, theses are ranked rst, then books, then journal papers, and nally conference articles. The rationality behind this is that an author writing a thesis or a book holds a higher level of expertise than an author who has only written conference papers. Regarding the date of the publication, we consider the `freshness' of the publications and continuity of an author's publications within the same area. More recent publications are ranked higher than older ones, and authors publishing in consecutive years about a certain topic are also ranked higher than authors that have sporadic publications about the topic. Regarding other authors, experts sharing a publication with fewer colleagues are ranked higher. The rationality behind this is that the total knowledge of a publication should be divided among the expertise brought into it, i.e., the number of authors. Additionally we also consider the order of authors in the publication. Main authors are considered to have a higher level of expertise and are therefore ranked higher.
To perform the rst step (i.e., retrieving the top-n documents related to the user's query) we could have used the SPARQL endpoint and, at run-time, searched for those keywords within the title and abstract properties of the publications. However, to speed the search process up, and to enhance the querydocument matching process, we have decided to pre-process and index the title and abstract information of the publications using the popular Lucene search engine. In this way, the fuzzy and spelling check query processing and ranking capabilities of the Lucene search engine are exploited to optimise the initial document search process.
To perform the second step, once the top-n documents have been selected, the rest of the properties of the document (authors, type, and date) are obtained at run-time using the SPARQL endpoint.
Finally, once the set of authors have been ranked, we look for them in the OU sta directory (using the information about their rst name and last name). If the author is included in the directory, the system provides related information about 25 http://lucene.apache.org/java/docs/index.html the job title, department within the OU, e-mail address and phone number. By exploiting the OU sta directory we are able to identify which experts are members of the OU and which of them are external collaborators, or old members not further working for the institution.
Without the structure and conceptual information provided by the OU LOD, the implementation of the previously described ranking criteria, as well as the interlinking of data with the OU sta directory, would have required a huge data pre-processing e ort. The OU LOD provides the information with a negrained structure that facilitates the design of ranking criteria based on multiple concepts, as well as the interlinking of information with other repositories. 4.2
System Implementation The system is based on lightweight client server architecture. The back end (or server side) is implemented as a Java Servlet, and accesses the OU LOD information by means of HTTP requests to the SPARQL endpoint. Some of the properties provided by the LOD information (more particularity the title and the abstract of the publications) are periodically indexed using Lucene to speed-up and enhance the search process by means of the exploitation of its fuzzy and spell checker query processing, and ranking capabilities. The rest of the properties (authors, date, and type of publications) are accessed at run time, once the top-n publications have been selected.
The front end is a thin client implemented as a web application using only HTML, CSS and Javascript (jQuery).26 The client doesn't handle any processing of the data, it only takes care of the visualisation of the search results and the search input. It communicates with the back-end by means of an HTTP request that passes as a parameter the user's query and retrieves the ranking of authors and their corresponding associated information by means of a JSON object. 4.3
Example and Screenshots In this section, we provide an example of how to use the OU expert search system. As shown in Figure 2, the system receives as a keyword query input \semantic search", with the topic for which the user aims to nd an expert. As a result, the system provides a list of authors (\Enrico Motta", \Vanessa Lopez ", etc), who are considered to be the top OU experts in the topic. For each expert, if available, the system provides the contact details (department, e-mail, phone extension) and the top publications about the topic. For each publication, the system shows its title, the type of document, and its date. If the user passes the cursor on the top of the title of the publication, the summary is also visualised (see the example in Figure 2 for the publication \Re ections of ve years of evaluating semantic search systems "). In addition the title of the publication also constitutes a link to its information in the open.ac.uk domain. 26 http://www.jquery.com
Fig. 2. The OU Expert Search system 5
Buddy Study The Open University is a well-established institution in the United Kingdom, offering distance-learning courses covering a plethora of subject areas. A key factor in enabling learning and understanding of course materials is support for students, provided in the form of an on-hand tutor for each studied module, where interactions with the tutor are facilitated via the Web and/or email exchanges. An alternative method of support could be provided through peers, in a similar manner to a classroom environment, where working together and explanations of problems from disparate viewpoints enhances understanding.
Based on this thesis, Buddy Study27 combines the popular social networking platform Facebook with the OU Linked Data service, the goal being to suggest learning partners { so called `Study Buddies' { from a person's social network on the site together with possible courses that could be pursued together. 5.1 Buddy Study combines information extracted from Facebook with Linked Data o ered by The Open University, where the former contains `wall posts' { messages posted publicly on a person's pro le page { and comments on such wall posts, while the latter contains structured, machine-readable information describing courses o ered by The Open University. 27 http://www.matthew-rowe.com/BuddyStudy
Combining the two information sources, in the form of a `mashup', is performed using the following approach. First the user logs into the application { using Facebook Connect { and grants access to their information. The application then extracts the most recent n wall posts and the comments on those posts { n can be varied, thereby a ecting the later recommendations. Given the extracted content, cleaning is then performed by removing all the stop words, thus reducing the wall posts and comments to their basic terms.
A bag of words model is compiled for each person in the user's social network as follows: for each wall post or comment posted by a given person all the terms are placed in the bag, maintaining duplicates and therefore frequencies. This model maintains information of the association between a user and his/her social network members in the form of shared terms. A bag of words model is then compiled for each OU course in a similar manner: rst we query the SPARQL endpoint of the OU's Linked Data asking for the title and description for each course. For the returned information, stop words are removed and the title and description { containing the remaining terms { are then used to build the bag of words model for the course.
The goal of Buddy Study is to recommend study partners to support course learning. Therefore we compare the bag of words model of each person with the bag of words model of each course, recording the frequency and terms that overlap. The user's social network members are then ranked based on the number of overlapping terms { the intuition being that the greater the number of common terms with courses, the greater the likelihood of a course being correlated with the user. Variance of n will therefore a ect this ranking, given that the inclusion of a greater number of posts will increase the number of possible study partners, while smaller values for n will yield more recently interacted with social network members. Variance of this parameter is provided in the application.
The application is not nished yet; we still need to recommend possible courses that could be studied with each possible study buddy. This is performed in a similar fashion, by comparing the bag of words model of the social network member with the model of each course, counting the frequencies of overlapping terms for each course, and then ranking accordingly. Due to space restrictions, and to avoid information overload, we only show the top-10 courses. For each social network user, and for each course that is suggested, Buddy Study displays the common terms, thereby providing the reasons for the course suggestion.
If for a moment we assume a scenario where Linked Data is not provided by the OU, then the function of Buddy Study could, in theory continue, by consuming information provided in an alternative form. However, this application forms the prototype upon which for future work { explained in greater detail within the conclusions of this paper { is to be based. Such advancements will utilise concepts for study partner recommendation rather than merely terms, the reasoning behind this extension is to alleviate the noisy form that terms take. By leveraging concepts from collections of terms, recommendations would be generated that are more accurate and better suited to the user in question. Without Linked Data, this is not possible. The application is live and available online at the previously cited URL. It is built using PHP, and uses the Facebook PHP Software Development Kit (SDK)28. Authentication is provided via Facebook Connect,29 enabling access to Facebook information via the Graph API. The ARC2 framework30 is implemented to query the remote SPARQL endpoint containing The Open University's Linked Data, and parse the returned information accordingly. 5.3
Example and Screenshots To ground the use of Buddy Study, Figure 3 shows an example screenshot from the application when recommending study partners for Matthew Rowe { one of the authors of this paper. At this rank position in the results, the possible study mate is shown together with the courses that could be studied together. The courses are hyperlinked to their resource within the OU Linked Open Data service, and in the proceeding brackets the terms that correlate with the courses are shown. In this instance the top-ranked course is identi ed by the common terms `API' and `Info'. The Open University o ers a set of free learning material through the OpenLearn website.31 Such material cover various topics ranging from Arts32, to Sciences and Engineering.33 In addition to that, the OU has other learning resources published in the form of Podcasts, along with courses o ered at speci c presentations during the year. While all these resources are accessible online, connections are 28 https://github.com/facebook/php-sdk 29 http://developers.facebook.com/docs/authentication 30 http://arc.semsol.org 31 http://openlearn.open.ac.uk 32 OpenLearn unit example in Arts: http://data.open.ac.uk/page/openlearn/a216 1 33 A list of units and topics is available at: http://openlearn.open.ac.uk/course not always explicitly available, making it hard for students to easily exploit all the available resources. For example, while there exists a link between speci c Podcasts and related courses, such links do not exist between OpenLearn units and Podcasts. This leaves it to the user to infer and nd the appropriate and relevant material to the topic of interest.
Linked OpenLearn34 is an application that enables exploring linked courses, Podcasts and tags to OpenLearn units. It aims to facilitate the browsing experience for students, who can identify on the spot relevant material without leaving the OpenLearn page. With this in place, students are able, for example, to easily nd a linked Podcast, and play it directly without having to go through the Podcast website. 6.1
Consumed Data Linked OpenLearn relies on The Open University's Linked Data to achieve what was previously considered very costly to do. Within large organizations, it's very common to have systems developed by di erent departments, creating a set of disconnected data silos. This was the case of Podcasts and OpenLearn units at the OU. While courses were initially linked to both Podcasts and OpenLearn in their original repositories, it was practically hard to generate the links between Podcasts and OpenLearn material. However, with the deployment of Linked Data, such links are made possible through the use of coherent and common URIs of represented entities.
To achieve our goals of generating relevant learning material, we make use of the courses, Podcasts, and OpenLearn datasets in data.open.ac.uk. As a rst step, while the user is browsing an OpenLearn unit, the system identi es the unique reference number of the unit from the URL. Then this unique number is used in the query passed to the OU Linked Data SPARQL endpoint (http://data.open.ac.uk/query), to generate the list of related courses including their titles and links to the study at the OU pages.
In the second step, another query is sent to retrieve the list of Podcasts related to the courses fetched above. At this level we get the Podcasts' titles, as well as their corresponding downloadable media material (e.g., video or audio les), which enable users to play the content directly within the application. Finally the list of related tags are fetched, along with an embedded query that generates the set of related OpenLearn units, displayed in a separate window. The user at this level has the option to explore a new unit, and the corresponding related entities will be updated accordingly. The application is still a prototype, and there is surely room for further data to extract. For example, once the library catalogue is made available, a much richer interface can be explored by students with related books, recordings, computer les, etc. 34 http://fouad.zablith.org/apps/openlearnlinkeddata We implemented the Linked OpenLearn application in PHP, and used the ARC2 library to query the OU Linked Data endpoint. To visualise the data on top of the web page, we relied on the jQuery User Interface library,35 and used the dialog windows for displaying the parsed SPARQL results. The application is operational at present, and is launched through a Javascript bookmarklet, which detects the OpenLearn unit that the user is currently browsing, and opens it in a new iFrame, along with the linked entities visualised in the jQuery boxes. 6.3
Example and Screenshot To install the application, the user has to drag the applications' bookmarklet36 to the browser's toolbar. Then, whenever viewing an OpenLearn unit, the user clicks on the bookmarklet to have the related entities displayed on top of the unit page. Figure 4 illustrates one arts related OpenLearn unit, with the connected entities displayed on the right, and a running Podcast selected from the \Linked Podcasts" window. The user has the option to click on the related course to go directly to the course described in the Study at the OU webpage, or click on linked tags to see the list of other related OpenLearn units, which can be browsed within the same window.
Fig. 4. Linked OpenLearn Screenshot 35 http://www.jqueryui.com 36 The bookmarklet is available at: http://fouad.zablith.org/apps/openlearnlinkeddata, and has been tested in Firefox, Safari and Google Chrome
Conclusions In this section we report on our experiences when generating and exploiting LOD within the context of an educational institution. Regarding our experience on transforming information distributed in several OU repositories and exposing it as LOD, the process complexity was mainly dependent on the datasets in terms of type, structure and cleanliness. Initially, before any data transformation can be done, it was required to decide on the vocabulary to use. This is where the type of data to model plays a major role. With the goal to reuse, as much as possible, already existing ontologies, it was challenging to nd the adequate ones for all our data. While some vocabularies are already available, for example to represent courses, it required more e ort to model OU speci c terminologies (e.g., at the quali cations level). To assure maximum interoperability, we chose to use multiple terminologies (when available) to represent the same entities. For example, courses are represented as modules from the AIISO ontology, and at the same time as courses from the Courseware ontology. Other factors that a ected the transformation of the data are the structure and cleanliness of the data sources. During the transformation process, we faced many cases where duplication, and information not abiding to the imposed data structure, hampered the transformation stage. However, this initiated the need to generate the data following well-de ned patterns and standards, in order to get easily processable data to add to the LOD.
Regarding our experiences exploiting the data, we have identi ed three main advantages of relying on the LOD platform within the context of education. Firstly the exposure of all these material as free Web resources have open opportunities for the development of novel and interesting applications like the three presented in this paper. The second main advantage is the structure provided by the data. This is apparent in the OU Expert Search system, where the di erent properties of articles are exploited to generate di erent ranking criteria, which when combined, provide much stronger support when nding the appropriate expertise. Finally, the links generated across the di erent educational resources have provided a new dimension to the way users can access, browse and use the provided educational resources. A clear example of this is the exploitation of LOD technology within the OpenLearn system, where OpenLearn units are now linked to courses and Podcasts, allowing students to easily nd in a single site, all the information they are looking for.
We believe that universities need to evolve the way they expose knowledge, share content and engage with learners. We see LOD as an exciting opportunity that can be exploited within the education community, especially by interlinking people and educational resources within and across institutions. This interlinking of information will facilitate the learning and investigation process of students and research sta , enhancing the global productivity and satisfaction of the academic community. We hope that, in the near future, more researchers and developers will embrace LOD approach, by creating new applications and learning from previous experiences to expose more and more educational data in a way that is directly linkable and reusable. The application of Linked Data within the OU has opened multiple research paths. Regarding the production of Linked Data, in addition to transforming the library records to LOD, the LUCERO team is currently working on connecting the OU's Reading Experience Database (RED)37 to the Web of Data. Such database aims to provide access and information about reading experiences around the world. It helps the readership for books issued in new editions for new audiences in di erent countries to be tracked. Its publication as LOD is an interesting example about how the integration of Linked Data technology can open new investigation paths to di erent research areas, in this case humanities.
Regarding the consumption of LOD, we envision, on the one hand, to enhance the three previously mentioned applications and, on the other hand to generate new applications as soon as more information is available and interconnected. As example of the former, for the Buddy Study application we plan to extend the current approach for identifying common terms between social network members and courses to instead utilise common concepts. At present the use of online messages results in the inclusion of abbreviated and slang terms, resulting in recommendations that are generated from noise. By instead using concepts, we believe that the suggested courses would be more accurate and suitable for studying. As an example of the latter, we aim to generate a search application over the RED database, able to display search results on an interactive map and link them not just to relevant records within the RED database, but also with relevant objects of the LOD cloud.
37 http://www.open.ac.uk/Arts/reading Q3 - Who succeeded to { Charles VII the Victorious } as ruler of France ? SELECT DISTINCT ?kingHR ?successorHR WHERE { ?x <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/class/yago/KingsOfFrance> . ?x <http://dbpedia.org/property/name> ?kingHR . ?x <http://dbpedia.org/ontology/successor> ?z . ?z <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/class/yago/KingsOfFrance> . ?z <http://dbpedia.org/property/name> ?successorHR } LIMIT 30
Q3 uses the YAGO ontology to ensure that the resource retrieved is indeed a king of France. Out of 30 results, one was incorrect (The three Musketeers). The query generated duplicates because of the multiple labels associated to each king. The same king was named for instance Louis IX, Saint Louis, Saint Louis IX. Whereas deduplication is a straight forward process in this case, the risk of inconsistent naming patterns among options of the same item is more difficult to tackle. An item was indeed generated with the following 3 options: Charles VII the Victorious, Charles 09 Of France, Louis VII. They all use a different naming pattern, with or without the king’s nickname and with a different numbering pattern.
Q4 - What is the capital of { Argentina }? With feedback SELECT ?countryHR ?capitalHR ?pictureCollection WHERE { ?country <http://dbpedia.org/property/commonName> ?countryHR . ?country <http://dbpedia.org/property/capital> ?capitalHR . ?country <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/class/yago/EuropeanCountries> . ?country <http://dbpedia.org/property/hasPhotoCollection> ?pictureCollection } LIMIT 30
The above question is a variation of Q1. It adds a picture collection from a distinct dataset in the response feedback. It uses the YAGO ontology to exclude countries outside Europe and resources which are not countries. A feedback section is added. When the candidate answers the item, he then receives a feedback if the platform allows it. In the feedback, additional information or formative resources can be suggested. Q4 uses the linkage of the DBpedia dataset with the Flickr wrapper dataset. However the Flickr wrapper data source was unavailable when we performed the experiment.
Q5 - Which category does { Asthma } belong to? SELECT DISTINCT ?diseaseName ?category WHERE { ?x <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Disease> . ?x <http://dbpedia.org/property/meshname> ?diseaseName . ?x <http://purl.org/dc/terms/subject> ?y . ?y <http://www.w3.org/2004/02/skos/core#prefLabel> ?category } LIMIT 30
Q5 aims to retrieve diseases and their categories. It uses SKOS and Dublin Core properties. The Infobox dataset is only used to find labels. Labels from the MESH vocabularies are even available. Nevertheless, the SKOS concepts are not related to a specific SKOS scheme. Categories retrieved range from Skeletal disorders to childhood. For instance, the correct answer to the question on Obesity is childhood. 4.2 The publication of items on the TAO platform The TAO platform3 is an open source semantic platform for the creation and delivery of assessment tests and items. It has been used in multiple assessment contexts, including large scale assessment in the PIAAC and PISA surveys of the OECD, diagnostic assessment and formative assessment.
We imported QTI items generated for the different item models in the platform, in order to validate the overall Linked Data based item creation streamline. Figure 4 presents an item generated from Q1 (Figure 3) imported in the TAO platform.
Figure 4 - Item preview on the TAO platform 3 http://www.tao.lu
The experimentation of the streamline was therefore tested with SPARQL queries which use various ontologies and which collect various types of variables. It raised two types of issues for which future work should find relevant solutions: the quality of the data and the relevance of particular statements for the creation of an assessment item. 5.1 Data quality challenges
In our experiment, the chance that an item will have a defective prompt or a defective correct answer is equal to the number of defective variables used for the item creation. Q1 uses the most challenging dataset in terms of data quality. 7 out of 30 questions had a defective prompt or a defective correct answer (23,33%).
The chance that an item will have defective distractors is represented by the following formula, where D is the total number of distractors, d(V) is the number of defective variables and V is the total number of variables:
We used 2 distractors. Among the items generated from Q1, 10 items had a
defective distractor (33,33%). Overall, 16 out of 30 items had neither a defective
prompt nor a defective correct answer nor a defective distractor (53,33%).
As a comparison, the items generated from unstructured content (text) that are
deemed usable without edit were measured between 3,5% and 5% by Mitkov et al.
[18] and between 12% and 21% by Karamanis et al. [
The labels The choice of the label for each concept to be represented in an item is a challenge when concepts are represented by multiple labels (Q4). The selection of labels and their consistency can be ensured by defining representation patterns or by using datasets with consistent labeling practices.
Inaccurate statements Most statements provided for the experiment are not inaccurate in their original context but they sometimes use properties which are not sufficiently precise for the usage envisioned (e.g., administrative capital). In other cases, the context of validity of the statement is missing (e.g., Leopoldville used to be the capital of a country called Congo). The choice of DBpedia as a starting point can increase this risk in comparison to domain specific data sources provided by scientific institutions for instance. Nevertheless, the Semantic Web raises similar quality challenges as the ones encountered in heterogeneous and distributed data sources [19]. Web 2.0 approaches, as well as the automatic reprocessing of data can help improve the usability of the Semantic Web statements. This requires setting up a traceability mechanism between the RDF paths used for the generation of items and the items generated. Data linkage Data linkage clearly raises an issue because of the reliability of the mechanism on different data sources. Q3 provided 6 problematic URIs out of 30 (i.e., 20%). Q4 generated items for which no URI from the linked data set was resolvable since the whole Flickr wrapper data source was unavailable. This clearly makes the generated items unusable. The creation of infrastructure components such as the SPARQL Endpoint status for CKAN4 registered data sets5 can help provide solutions to this quality issue over the longer run.
Missing inferences Finally, the SPARQL endpoint does not provide access to inferred triples. Our streamline does not tackle transitive closures on the data consumer side (e.g., through repeated queries), as illustrated with Q3. Further consideration should be given to the provision of data including inferred statements. Alternatively, full datasets could be imported. Inferences could then be performed in order to support the item generation process.
Different strategies can therefore be implemented to cope with data quality issues we encountered. Data publishers can improve the usability of the data, for instance with the implementation of an upper ontology in DBpedia. However, other data quality issues require data consumers to improve their data collection strategy, for instance to collect as much information as possible on the context of validity of the data, whenever it is available. 5.2 Data selection The experiment also showed that the Linked Data statements should be selected. The suitability of an assessment item for a test delivered to a candidate or a group of candidates is measured in particular through such information as the item difficulty. 4 http://www.ckan.net 5 http://labs.mondeca.com/sparqlEndpointsStatus/index.html The difficulty can be assessed through a thorough calibration process in which the item is given to beta candidates for extracting psychometric indicators. In low stake assessment, however, the evaluation of the difficulty is often manual (candidate or teacher evaluation) or implicit (the performance of previous candidates who took the same item). In the item generation models we have used, each item has a different construct (i.e., it assesses a different knowledge). In this case, the psychometric variables are more difficult to predict [20]. A particular model is necessary to assess the difficulty of items generated from Semantic Web sources. For instance, it is likely that for a European audience, the capital of the Cook Islands will raise a higher rate of failure than the capital of Belgium. There is no information in the datasets, which can support the idea of a higher or lower difficulty. Moreover, the difficulty of the item also depends on the distractors, which in this experiment were generated on a random basis from a set of equivalent instances. As the generation of items from structured Web data sources will become more elaborated, it will therefore be necessary to design a model for predicting the difficulty of generated items. 6 Conclusion and future work The present experimentation shows the process for generating assessment items and/or assessment variables from Linked Data. The performance of the system in comparison with other approaches shows its potential as a strategy for assessment item generation. It is expected that data linkage can provide relevant content for instance to propose formative resources to candidates who failed an item or to illustrate a concept with a picture published as part of a distinct dataset. The experimentation shows the quality issues related to the generation of items based on such a resource as DBpedia. It should be noted that the measurements were made with a question which raises particular quality issues. It can be easily improved as shown with other questions. Nevertheless the Linked Data Cloud also contains datasets published by scientific institutions, which may therefore raise less data accuracy concerns. In addition, the usage model we are proposing is centered on low stake assessment, for which we believe that the time saved makes it worthwhile having to clean some of the data, while the overall process remains valuable.
Nevertheless, additional work is necessary both on the data and on the assessment
items. The items created demonstrate the complexity of generating item variables for
simple assessment items. We aim to investigate the creation of more complex items
and the relevance of formative resources which can be included in the item as
feedback. Moreover, the Semantic Web can provide knowledge models from which
items could be generated. Our work is focused on semi-automatic item generation,
where users create item models, while the system aims to generate the variables.
Nevertheless, the generation of the items from a knowledge model as in [
We will focus our future work on the creation of an authoring interface for item models with the use of data sources from the Semantic Web, on the assessment of item quality, on the creation of different types of assessment items from Linked Data sources, on the traceability of items created, including the path on the Semantic Web datasets which were used to generate the item, and on the improvement of data selection from semantic datasets.
Acknowledgments. This work was carried out in the scope of the iCase project on computer-based assessment. It has benefited from the TAO semantic platform for eassessment (https://www.tao.lu/) which is jointly developed by the Tudor Research Centre and the University of Luxembourg, with the support of the Fonds National de la Recherche in Luxembourg, the DIPF (Bildungsforschung und Bildungsinformation), the Bundesministerium für Bildung und Forschung, the Luxemburgish ministry of higher education and research, as well as OECD. Presented at the 2009 Ninth IEEE International Conference on Advanced Learning Technologies (ICALT), Riga, Latvia. (2009) 13. Xu, Y., Seneff, S. Speech-Based Interactive Games for Language Learning: Reading, Translation, and Question-Answering. Computational Linguistics and Chinese Language Processing Vol. 14, No. 2, pp. 133-160. (2009) 14. Lai, H., Alves, C., & Gierl, M. J. Using automatic item generation to address item demands for CAT. In Proceedings of the 2009 GMAC Conference on Computerized Adaptive Testing. (2009) 15. Gierl, M.J., Zhou, J., Alves, C. Developing a Taxonomy of Item Model Types to Promote Assessment Engineering. Journal of Technology, Learning, and Assessment, 7(2). (2008) 16. Sarre, S., Foulonneau, M. Reusability in e-assessment: Towards a multifaceted approach for managing metadata of e-assessment resources. Fifth International Conference on Internet and Web Applications and Services. (2010) 17. Suchanek, F. M., Kasneci, G., & Weikum, G. Yago: a core of semantic knowledge. In Proceedings of the 16th international conference on World Wide Web (pp. 697–706). (2007) 18. Mitkov, R., An Ha, L., & Karamanis, N. A computer-aided environment for generating multiple-choice test items. Natural Language Engineering, 12(02), 177– 194. (2006) 19. Foulonneau, Muriel, Cole, Timothy W. Strategies for reprocessing aggregated metadata. European Conference on Digital Libraries. Lecture notes in computer science 3652 , 290-301 (2005) 20. Bejar, I. I., Lawless, R. R., Morley, M. E., Wagner, M. E., Bennett, R. E., & Revuelta, J. A feasibility study of on-the-fly item generation in adaptive testing.
Educational Testing Service. (2002) A Mobile and Adaptive Language Learning Environment based on Linked Data
Davy Van Deursen1, Igor Jacques2, Stefan De Wannemacker2, Steven Torrelle1, Wim Van Lancker1, Maribel Montero Perez2, Erik Mannens1, and Rik Van de Walle1 Abstract. The possibilities within e-learning environments increased dramatically the last couple of years. They are more and more deployed on the Web, allow various types of tasks and ne-grained feedback, and they can make use of audiovisual material. On the other hand, we are confronted with an increasing heterogeneity in terms of end-user devices (smartphones, tablet PCs, etc.) that are able to render advanced Web-based applications and consume multimedia content. Therefore, the major contribution of this paper is an adaptive, Web-based e-learning environment that is able to provide rich, personalized e-learning experiences to a wide range of devices. We discuss the global architecture and data models, as well as how the integration with media delivery can be realized. Further, we give a detailed description of a reasoner, which is responsible for the adaptive selection of learning items, based on the usage environment and the user pro le.
Keywords: Adaptive, Language Learning, Mobile, Web-based 1 The last years, the use of e-learning environments has increased spectacularly, not only in formal educational settings, but also in working and private environments. At the same time, the possibilities within these e-learning environments increased dramatically: learning environments for instance have become easier and more pleasant to use, they allow various types of tasks and ne-grained feedback, and they can make use of audiovisual material. Moreover, while e-learning environments were traditionally o ered as applications on stand-alone computers, nowadays they are more and more being rendered over the Internet. It is clear that these evolutions are related to technological evolutions, and the wide availability of fast multimedia computers and internet access.
Next to the fact that e-learning environments are more and more deployed over the Web, we are confronted with an increasing heterogeneity in terms of end-user devices that are able to connect to the Web and consume multimedia content. Therefore, personal devices such as tablet PCs and smartphones could be used as learning devices, next to traditional desktop and laptop devices. Also, the role of personalization within e-learning environments has become more and more important. Personalization can be applied both at the learning level (i.e., adjust learning sessions according to the learner's capabilities) and at the environmental level (i.e., adjust the rendering of the learning environment according to the characteristics of the usage environment).
The above described challenges are exactly the ones that are currently tackled in the IBBT MAPLE project (Mobile, Adaptive & Personalized Learning Experience3), which aims to make adaptive mobile e-learning possible. Therefore, in this paper, we present a Web-enabled e-learning environment that is able to o er personalized learning sessions on any device, primarily focused on language learning making optimal use of digital multimedia. In order to realize such an environment, we need the following key components: { a common, machine-understandable data model that is independent of usage environments and is able to express both learning content and metadata about the learning content; { a logging framework that allows to capture the behaviour and performance of the learner on a detailed level; { a reasoner that is able to select learning items based on the learner's capabilities and behaviour; { a media delivery platform taking into account usage environment characteristics and restrictions.
In the remainder of this paper, we provide an overview of the architecture of our adaptive e-learning platform. Further, we discuss the above described key components in more detail. Finally, we discuss related work, draw a number of conclusions, and discuss some future work. 2
MAPLE platform In order to o er a highly adaptive e-learning platform that can also deal with (mobile) multimedia delivery, we designed the architecture that is depicted in Fig. 1. Two major parts can be distinguished: the e-learning platform and the media delivery platform. The e-learning platform relies on two RDF stores, i.e., a store for learning exercises and a store for learner pro les. The learning items store is lled through the learning item ingest service. More details regarding the creation of learning items and the data model according to which they are modeled are provided in Section 3. Further, the learner pro le store is build up, based on the learners' actions and preferences (see Section 3.5). The reasoner is responsible for selecting the most adequate exercise, based on the learner's pro le and environment and the available learning items. Detailed information 3 http://www.ibbt.be/en/projects/overview-projects/p/detail/maple-2 media ingest service learning item ingest service media store learning item
DB regarding the reasoner is provided in Section 4. Finally, the learning endpoint is the communication point between learner devices and the e-learning platform.
The media delivery platform corresponds to NinSuna4, which is a
metadatadriven media adaptation and delivery platform [25]. At its core, format-independent
modules for temporal selection and packaging of media content are present.
Almost all existing media delivery channels are supported by NinSuna: RTSP,
RTMP, HTTP progressive download, and HTTP adaptive streaming. Moreover,
native support for Media Fragments 1.0 [24] is provided, which enables the
delivery of media fragments (i.e., temporal or track fragments) in a standardized
way [
The presented e-learning platform exposes its data (i.e., learning content and accompanying media resources) as linked data. More speci cally, it follows the guidelines regarding the publication of linked data5: use dereferencable HTTP URIs as names for things, provide useful information using the standards (RDF, SPARQL), and include links to other URIs. Hence, within our platform, the learning items and learner pro les are available through a SPARQL endpoint, while the metadata of the media resources are published as RDF URIs. This way, services such as the reasoner and the ADTE can rely on the linked data and can start reasoning over it. 4 http://ninsuna.elis.ugent.be 5 http://www.w3.org/DesignIssues/LinkedData.html
A typical e-learning scenario using this architecture is then as follows: (1) the learner logs in into the Web-based e-learning application using its mobile device, which contacts the learning endpoint of the e-learning platform; the end point approaches the reasoner which provides a personalized overview of the available courses; (2) based on the course selected by the learner, the reasoner selects an exercise from the learning item store, taking into account the learner pro le and the available exercises within that course; (3) when the selected exercise contains media content (audio, video, or images), the ADTE of NinSuna is contacted in order to select the media resource version that ts best for the current device; (4) the learning endpoint renders the selected exercise in HTML and sends the response to the learner; (5) when the learner is solving the selected exercises, his/her answers and his/her behaviour in terms of clicks and timing is logged and sent back to the elearning platform; (6) the received answers and behaviour information are used to update the learner's pro le.
In the next sections, more detailed information regarding a number of components in the architecture is provided. 3
Data Models and Instance Generation A number of di erent data models need to be developed in order to structure and de ne the content used on the e-learning platform. More speci cally, we need the following data models: { model for the learning items and their metadata (e.g., question, possible answers, di culty level); { model for the learning domain; { model for the metadata of the media resources (e.g., bit rate); { model for the learner pro le; { model for the logging.
In the following subsections, we provide more information regarding these different models and how they are populated. Note that all ontologies are modelled in OWL and published online.
Model for learning items and their metadata The model for learning items consists of two ontologies: one for the learning items themselves6 and one for their metadata7. An example instance of a learning 6 http://multimedialab.elis.ugent.be/organon/ontologies/maple/content 7 http://multimedialab.elis.ugent.be/organon/ontologies/maple/llomp
Listing 1.1. Representing a learning item and its metadata in RDF (in Turtle). 1 @prefix mplc : <http :// multimedialab . elis . ugent . be / organon / ontologies / maple / content # >. @prefix llomp : <http :// multimedialab . elis . ugent . be / organon / ontologies / maple / llomp # >. @prefix xsd : <http :// www . w3 . org /2001/ XMLSchema # >.
@prefix dc : <http :// purl . org / dc / terms /> .
15
20
25
30
35
40
item modelled according to our model is shown in Listing 1.1. We explain and
illustrate both ontologies based on this example. The model is heavily based
on the Learning Object Metadata (LOM, [
We describe not only the metadata of learning items, but also the exercises themselves. This way, they are formally represented, independent of any rendering. Moreover, they can be easily integrated with their metadata and corresponding media resources. Also, the reasoner (Section 4) will not only rely on the learning item metadata, but also on the items themselves (e.g., this type of exercise is preferred by the learner). For the moment, six mplc:exerciseTypes are supported (focussed on language learning): { Multiple Choice: given a number of answers, the learner has to choose exactly one answer; { Multiple Response: given a number of answers, the learner has to choose one or more answers; { Fill Gaps: given a text with some gaps, the learner needs to ll in missing text in text boxes; { Dropdown: same as Fill Gaps, but instead of free text elds, the learner can choose between a number of prede ned answers; { Click on Text: given a text, the learner needs to click/tab on one or more words; { Click on Zone: given an image or video, the learner needs to click/tab at one or more regions within the image or video.
Note that media elements can also occur within the rst ve types of exercises. For instance, a movie can be played followed by the question to solve. Only the last type (Click on Zone) uses multimedia in an interactive way as described in [19].
In Listing 1.1, a multiple choice exercise is used as example (line 9). A link to a movie fragment is provided via the mplc:media property (line 10), which takes as value a Media Fragment URI (see Section 3.3). The mplc:task description (line 11) provides the question or task in multiple languages (based on the level of the learner, the reasoner can choose if the language is presented in the native language of the learner or not). Further, the mplc:answerSpace (line 13) corresponds to the zone where the learner can enter its answers. Within such an answer space, mplc:input is provided (line 14), where each mplc:answer corresponds to one possible answer. In case of a multiple choice type, each answer corresponds to a mplc:Choice. It contains information such as `is this possible answer the correct one?', `how much does the learner score when (s)he selects this one?', and the possible answer itself. LOM-speci c elements such as llomp:lifeCycle (line 40) and llomp:educational (line 34) are present as well.
As a part of the aforementioned LOM extensions, we added the learning component property to the educational component. Since the MAPLE project focusses on language learning, we extended this learning component property with speci c support for language learning. The learning component is split up into three separate subcomponents: target language, theme and language component. The latter component can have one or more of the following subproperties: { knowledge property: vocabulary, pronunciation, etc.; { skill property: reading, listening, writing or speaking We also de ned a hierarchical structure for the range of the knowledge property based on which the exact knowledge URIs can be deduced. This was done in a language-independent way extendable with language-speci c elements. As the
Listing 1.2. Representing a learning component in RDF (in Turtle). 1 @prefix lang : <http :// kuleuven - kortrijk . be / itec / ext / ontologies / itec_elearning_ontology / languagecomponent /# >. @prefix llomp : <http :// multimedialab . elis . ugent . be / organon / ontologies / maple / llomp # >. 5 10 <http :// ninsuna . elis . ugent . be / rdf / resource / maple / learningComponent_40001 > a llomp : LearningComponent ; llomp : theme " agriculture " ; llomp : targetLanguage "en - UK " ; llomp : languageComponent [ a llomp : LanguageComponent ; llomp : knowledge <http :// kuleuven - kortrijk . be / itec / ext / ontologies / itec_elearning_ontology / languagecomponent / grammar / partsOfSpeech / substantive > ; llomp : skill lang : writing .
] . skill and knowledge property exists next to each other, it is possible to specify the subject of an exercise very accurately. In Listing 1.2 an example instance of a learning component can be found. The exercise in this instance trains the writing skill of substantives related to agriculture.
Within the MAPLE project, we use learning items from Televic Education (TEDU)8. Currently, TEDU stores their learning items and accompanying metadata in a SQL store. Through XML feeds, the store can be accessed from outside. Hence, we implemented a converter taking as input the XML feeds and producing RDF learning items according to the above described model.
Model for the learning domain The learning items are not physically arranged into courses. Which learning objects belong together is determined by the metadata, namely the learning component within the educational component of each item. The domain model consists of two type of relations: prerequisite and hierarchical relations. In the project, the domain model is supposed to be simple. It is a three level hierarchical model in which the items are rst distinguished by their target language, secondly by their theme, and thirdly by their language component. Additionally, there exist prerequisite requirements between the language components, expressing one language component depends on the knowledge of another. The reasoner will take into account these prerequisites when determining what courses are available for the learner. 3.3
Model for media metadata
To model media resources, we rely on the W3C Media Annotations ontology [
Listing 1.3. Representing a learner pro le in RDF (in Turtle). 1 @prefix itec : <http :// kuleuven - kortrijk . be / itec / ext / ontologies / itec_elearning_ontology # >. @prefix foaf : <http :// xmlns . com / foaf /0.1/ >. @prefix mplc : <http :// multimedialab . elis . ugent . be / organon / ontologies / maple / content # >. 10 15 20 25 already contains mappings to many other existing metadata formats. Further, the ontology also provides support for Media Fragment URIs.
Model for the learner pro le
In order to steer the decision making of the reasoner, an up-to-date learner
pro le is required for each of the learners in the learning system. This pro le
holds pro ciency score estimations for each of the appropriate learning subjects.
Each of these values is supplemented with a reliability parameter, namely the
variance of the estimator. As we focus on language learning, the pro ciency scores
are expressed on a continuous scale based on the discrete European Language
Levels [
The properties in the model will be caught either automatically either by means of preference setting. The learner's favourite learning item types can be edited through a preference menu and the learner's pro ciency scores will be updated by a module of the reasoner. Additionally, the ontological model sup
Listing 1.4. Representing a logging abstract in RDF (in Turtle). 1 @prefix itec : <http :// kuleuven - kortrijk . be / itec / ext / ontologies / itec_elearning_ontology # >. @prefix learners : <http :// kuleuven - kortrijk . be / itec / instances / maple / learners # >. @prefix log : <http :// kuleuven - kortrijk . be / itec / instances / maple / logging # >. @prefix maple : <http :// ninsuna . elis . ugent . be / rdf / resource / maple / >. 10 15 20 25 ports properties like motivation, learning style, learner strategy, and cognitive ability's, but currently these are not used in the MAPLE e-learning platform.
Model for logging the learner's activity Finally, we developed a model for describing logging information. For instance, the model is able to express information such as the start and stop of a learner session or the learner's course selection. Once the learner has chosen a course, a learning session is initiated in which the reasoner successively selects a new learning item, each time resulting in a learning item session which lasts for the time the learner interacts with the item. During such an item session a learner can give an answer, request a hint, or change his mind by changing his answer. All these interactions are logged by the system. This results in a huge amount of information which is consumed in two ways. Firstly, a part of the logging information is used at run-time by the reasoner. For instance, a score attained by the learner will a ect the pro ciency score of a learner's pro le through the functionality of the reasoner's pro ciency manager. Secondly, after runtime, the logged information will be used as input for statistical research tracing how certain interactions of the learner give information about the learning process. In Listing 1.4 an example instance can be found. The learner and learning session, and the session of the item are respectively interconnected by the itec:hasSubSession and the itec:hasItemObjectSession relation.
These resulting triples are partially generated in the core of the reasoner, e.g. the start en stop of the learner and the learning sessions. The low level interactions concerning one speci c exercise are generated at the client and sent back to the reasoner which processes the logging and stores it in the learner pro le RDF store. 4 The reasoner, introduced in Section 2, is a crucial component within the MAPLE learning system architecture as it is responsible for the adaptive learning item selection. If a learner logs in, the reasoner will rst of all provide a short list of courses from which the learner can choose. As the reasoner is aware of the learning goals for each learner through the learner pro le model, only courses that contribute to the not yet attained learning goals can be selected. Next, once the learner has chosen a course, the reasoner will start up a learning session and will successively decide on the exact exercise to deliver to the learner.
The reasoner takes into account the learner pro le as well as some real time
environmental properties. For the environmental adaptivity, both the screen
capacity and connection quality of the user's device are sources of adaptivity. In
case the screen size is too small, the reasoner will avoid the use of exercises with
media. A slow network connection will also result in avoiding media exercises.
For the learner pro le adaptivity, there are two main policies which can steer
the decision process. The rst one is based on a theory stating that the
exercise di culty needs to be increased each time a learner has answered a series of
four exercises correctly. Similarly, when four consecutive exercises are answered
incorrectly, it should go down [
To ful l the aforementioned tasks, the architecture of the reasoner (shown in Fig. 2) consists of six modules, supplemented by a facade for communicating with the learning endpoint. The six reasoner modules are the Learner manager, Environment manager, Learning task decision manager, Sequence manager, Logging manager and Pro ciency manager. We elucidate the functionality of these modules by means of the following example.
Learner profile
DB Learning item
DB
Environment manager Sequence manager Learner manager
Learning task decision manager Proficiency manager
Logging mananger
Reasoner Fig. 2. The reasoner architecture
Facade
Learning endpoint
Suppose a learner's initial pro le was set by a teacher thereby providing the learning goal \Achieve the B2 level for the French verb form imparfait" and also providing an estimation for the learner's initial level, namely A2, for \the French verb form imparfait". When the learner logs in, the Learner manager produces a learner session. Consequently, the Learning task decision manager loads the learner's learning goals in order to compose a three-level tree representation of all courses relevant for this learner, as explained in Section 3.2. This tree is sent to the learning endpoint which produces a representation such that the learner can navigate through the tree. Let us assume that the learner rst selects `French' followed by the theme `General' and nally the language component `Imparfait'. Besides, the learner opens the preferences menu and sets the dropdown exercise type as his favourite one.
Next, the Learning task decision manager composes a learning task object which is send to the Sequence manager. Here the learning task is sequencing the items (exercises and feedback) with the rst policy of adaptivity, starting from level A2, having as a stop criterion the achievement of the level B2, and taking into account the learner's preferred exercise types and environmental properties. Subsequently, the Sequence manager loads the sequencer necessary for the learning task. To this end, the sequencer makes use of the Environment manager, which is an access point for information on the current connection quality and the screen size of the device of the learner. At this point, the sequencer can successively decide on the id of the next item and passes its choice to the learning endpoint, which automatically generates a visual representation and makes use of the delivery platform in case media are present.
Once the learner nishes the exercise or has read the feedback in case of a feedback item, the logging information about the interactions of the learner with the item are sent back to the Logging manager of the reasoner. The latter sends this information as a speci c logging object to a couple of observer objects which all have di erent functionalities. For instance, there is an observer writing these logs to the learner pro le RDF store. Another observer warns the sequencer when for example four exercises have been consecutively answered correctly and yet another sends the learner's score to the Pro ciency manager together with the level, di culty and the learning subject of the answered exercise. The Pro ciency manager keeps the pro ciency scores up to date. Prior to every decision of the sequencer, the stop criterion is tested based on a pro ciency that is retrieved from the Pro ciency manager. If this criterion is reached, the sequencer sequences a special concluding feedback item announcing the end of the learning session to the learner. 5
Related Work
The architecture of the reasoner builds further on existing proposals for generic
learning system architectures such as in [20]. These architectures however have
mostly been designed having an adaptive hypermedia learning system in mind.
Even though most systems currently developed are based on providing learner
control based on adaptive links, e.g. [
The ontology for the learner pro le is a compact non-exhaustive synopsis of
the most common learner characteristics found in literature [
Our realizations in this project largely replace the functionality of the
restrictive SCORM standard [
Gang et al. proposed a framework for mobile learning in [
Conclusions and Future Work In order to exploit the possibilities of Web-based e-learning environments, we proposed an e-learning architecture that is able to provide rich, personalized e-learning experiences to a wide range of devices. We discussed the various data models used within the e-learning framework. Moreover, we provided details of the reasoner, a crucial component allowing to select learning items based on the usage environment and the learner pro le.
Future work consists of exploiting the possibilities of the Semantic Web even more by linking learning items to the Linked Open Data cloud. Further, data models could be optimized and linked to upcoming e orts (e.g., how to represent the life cycle of a learning item as provenance information on the Web). Also, more detailed domain models should be investigated. Regarding the reasoner, future work consists of taking into account more information obtained from the logging framework, as well as investigating how error-speci c feedback could be generated (e.g., link frequently occurring errors to answers).
Acknowledgments The research activities as described in this paper were funded by Ghent University, the Interdisciplinary Institute for Broadband Technology (IBBT, 50% co-funded by industrial partners), the Institute for the Promotion of Innovation by Science and Technology in Flanders (IWT), the Fund for Scienti c ResearchFlanders (FWO-Flanders), and the European Union.
References