Using Linked Data as a basis for a Learning Resource Recommendation System

Resource List Management Systems (RLMS) allow the electronic publication of course reading lists. Aside from electronic access, existing systems in this area provide little utility for teachers and learners above and beyond the traditional paper based reading lists. Our vision is that resource lists could in actual fact become Open Educational Resources that can be shared, re-mixed and re-used across institutions and borders. This paper introduces how we used linked data to architect a RLMS to meet this vision. However, in implementing this system, questions arose around the provenance, sustainability, licensing and reliability of today’s linked data cloud. This paper documents the steps we took to address these critisms in our implementation. The paper goes on to discuss how the ecosystem of learning data managed by this application opens the way for future work, which involves leveraging typed relationships between learning goals, educational resources and system actors to provide recommendation-like services for academics creating new content. 1 What is Linked Data When Sir Tim Berners-Lee originally expressed his vision for the Semantic Web, he was imagining a Web of Data[1]: I have a dream for the Web [in which computers] become capable of analyzing all the data on the Web the content, links, and transactions between people and computers. A Semantic Web, which should make this possible, has yet to emerge, but when it does, the day-to-day mechanisms of trade, bureaucracy and our daily lives will be handled by machines talking to machines. The intelligent agents people have touted for ages will finally materialize. Berners-Lee would later go on to define some of the properties of this Web of Data, and in doing so coined the term ‘Linked Data’, which simply refers to a set of best practices for publishing and connecting structured data on the Web. Berners-Lee expressed these a set of simple rules[2]: 1. Use URIs as names for things 2. Use HTTP URIs so that people can look up those names 2 Nadeem Shabir, Chris Clarke 3. When someone looks up a URI, provide useful information 4. Include links to other URIs, so that they can discover more things This vision of a Web of Data, combined with these principles for linking data gave rise to the The Linking Open Data Project which is a communityled effort to create openly accessible, and interlinked, RDF Data on the Web. When this community began its efforts, in 2007, there were only a handful of these connected, and openly accessible, sets of data. We can contrast this with how the Linked Open Data graph looks today, as an increasing number of data providers have begun publishing data using these principles, leading to the creation of a Web of Data that already contains billions of statements: Fig. 1. The Linked Open Data Cloud as of March 2009 2 Our vision for Resource List Management Usually organised around a set of study topics, resource lists contain details of books, journal articles, web pages and audio visual content which, along with annotations provided by the teacher, guide students to discover relevant subject material required to complete assignments or other course assessments. These lists are extensively used within Higher Education (HE), and can also be found in use at further and secondary education establishments. 1 http://linkeddata.org/ 2 http://www4.wiwiss.fu-berlin.de/bizer/pub/lod-datasets_2009-03-05.html Linked Data as a basis for Learning Resource Recommendations 3 In the past 10 years, several specialist online systems have emerged to assist in the management of such lists, such as Sentient Discover, Talis List, LORLS and Blackwell Reading Lists online. Such systems offered an online representation of the paper-based lists, as well as providing tools for the library to assist in stock acquisition. However, current Resource List Management Systems (RLMS) have provided limited extended functionality for the teacher over and above that of the paperbased solution save online access for students this may explain their limited adoption, and the continued proliferation of paper handouts. Some allow linking to journal articles via institutional link resolvers, and for items the library physically holds, most allow linking to the library catalogue. However, these systems are simply signposting solutions, providing none of the added services that users of Web 2.0-like systems might expect, such as recommendation services, rich user interface metaphors or the integration or in-lining of the resources themselves, including full text, into the list. Our research showed that often teachers construct these lists in a style which reflects both the chronological order and/or the major topic areas the course unit covers. Thus the structure of these lists, and the relative position of resources on it tells us something about their intended usage and how they relate to each other. We also know that in authoring these lists, teachers are either explicitly or implicitly influenced by similar works by their peers. An example of explicit influence is the teacher that seeks out similar syllabi when trying to author their own. An example of implicit influence is where peers discuss the availability and quality of educational resources, which may later lead to their use (or not) within the classroom. In essence, the latter could be described as a professional variant of the water cooler effect. This made us consider the impact of creating a system that enhances the authoring process of the lists by making it possible to formally harness the existing work of peers, thus supporting either the creation of derived works (with appropriate attribution), or using them as the basis for content suggestions for authors of new lists within comparable subject areas. In the development of our new system, Talis Aspire, our vision was to create a system which would allow resource lists themselves to be considered and operated on as Open Education Resources (OERs). This means that they can be re-used, remixed, shared and collaborated on easily, supporting the notion that open access to knowledge is in the interests of all. 3 http://www.sentientdiscover.co.uk 4 http://www.talis.com/list 5 https://lorls.lboro.ac.uk/ 6 http://www.readinglists.co.uk 7 http://www.wordspy.com/words/watercoolereffect.asp 8 http://www.talis.com/aspire 9 http://en.wikipedia.org/wiki/Open_educational_resources 4 Nadeem Shabir, Chris Clarke 3 Benefits of a Linked Data approach A key objective that makes our vision workable is that a user should be able to easily discover appropriate content (or have it recommended) to re-use and remix. It follows then that resource lists must be richly and homogeneously described in order that lists from different authors are comparable. In addition, combination with datasets outside the system boundary become important in the creation of a data ecosystem which supports discovery and recommendation or in other words, ease of discovery of new relationships. The nature of RDF-based systems, such as those that underpin the datasets on the linked data web, make it easy to re-combine graphs of data from multiple sources, allowing these new relationships to be discovered. We concluded early that not only would the system have to merge resource metadata from multiple and incompatible sources, but that each individual customer implementation of the system should be able to publish the resulting resource lists in a way that could be re-combined at a later date to enable reuse, remixing and sharing of data within a multi-institution ecosystem. Without the resulting scale that combining data from multiple institutions provided, any discovery or recommendation features within a particular subject domain would be of limited use. Our experience with RDF and specifically linked data indicates suitability for richness of description, standardised publication, interlinking and interoperability between disparate sets of data. By settling on linked data principals, as described in an earlier paper by Clarke[10], the team were able to unify not only the description of resources using shared ontologies such as Bibliographic Ontology, Resource List Ontology, SIOC and FOAF, but also on how the resource lists were to be published, allowing them to be combined with other data sources at a later date. This approach is supported by one of the challenges that a recent JISC-funded report[11] suggests semantic technologies can address: Information in UK HE/FE institution seems to be fragmented and in formats that makes it often inaccessible. Discovery of relevant information over a large number of sources needs to be supported. Information that is publicly available on the institutions Web pages is not available in machine processable formats making it difficult to compare programmes of study, syllabuses or research angles. 10 http://bibliontology.com/ 11 http://vocab.org/resourcelist/ 12 http://rdfs.org/sioc/spec/ 13 http://xmlns.com/foaf/0.1/ Linked Data as a basis for Learning Resource Recommendations 5 4 Leveraging the data ecosystem to support the discovery and recommendation of content in an Open Education context Given that one could interconnect resource metadata used to construct resource lists between departments, schools and even institutions, one could discover which modules cite the textbook Financial Accounting and Reporting (Elliot & Elliot). Unifying the description of those modules, one could discover if the textbook was largely being cited on 1st year Business Studies courses, or if it was actually a core text on most MBA programmes. What resources usually appear alongside Elliot & Elliot on resource lists? Combine with this knowledge about how students actually use the text, (for example, do they purchase it or do they ignore it) and multiply this knowledge across all disciplines and resources used for learning and it is conceivable that one could create advanced, context-aware recommendation systems for a multitude of use cases. For example, when building the list, teachers can use the system to help them predict the impact of including a text on a particular resource list. They can discover, and aggregate, which resources are routinely included by their peers, or how the majority of students choose to use resources. If they choose to include an item which is not held by the library, the system could suggest similar items that are held. Additionally, it is conceivable that they could browse a repository of lists in a related subject area, licensed as OERs, and use them as a basis for their own work. Our work to date has focused on seeding an ecosystem of resource list data as a basis for future work in developing discovery and recommendation functionality we describe above. To date we have six UK HEIs using the system with plans to expand to a further twenty five during 2009/10. It is our assumption that this is around the lower limit required before the functionality described is viable for the end user. What follows is the set of techniques we intend to blend together to realise the above scenarios, thus unlocking the potential of the ecosystem of data we have built. 4.1 Explicit hierarchical classification Chandrasekaran stated that hierarchical classification was one of the generic tasks that must be addressed when designing expert systems[13]. To build effective recommendation systems, it is important to know the context of educatorselected resources for example, a list for level one students for a module entitled Introduction to Clinical Psychology, is unlikely to have much in common with data from lists around the topic of Organisational Behavior at the same institution, so similarity between resources should be weighted much lower than those from equivalent level one Clinical Psychology courses delivered at other institutions. 6 Nadeem Shabir, Chris Clarke The system uses the AIISO ontology to organise lists into a tree hierarchy at the institutional level, allowing one to trace the module, programme, department and institution that a list and its resources belong to. However, when taking a view across the whole ecosystem, one cannot use this mechanism to map equivalent lists at different institutions. By augmenting the AIISO descriptions with data from the Joint Academic Coding System (JACS) we now have a basis for comparison at the course level. Mixing in the level of each academic programme gives us further data to complete our hierarchical classification system. In addition, it is our intention to ask teachers to optionally indicate their principal subject areas in their profile. This allows other users to locate their profile and subscribe to updates about resources they have recently pulled into the system (their bookmarks) and lists they have made available under OERcompatible licenses. The aim is to replicate the watercooler effect inside the application teachers can follow the implicit recommendations made by their peers an example being the inclusion of a resource on a resource list. 4.2 Deriving similarity without explicit classification Where no explicit classification of a resource list is made, it is possible to derive similarity between a set of given lists by analyzing the pattern of resource usage (inference). The results can be used to populate a similarity index to aid the discovery of lists in a related topic area, or to recommend individual resources within a particular domain. A very naive example is as follows: List 1 at location A contains a resource X. List 2 at location B contains a resource Y. List 3 at location C contains resource X and states Y as an alternative should X be unavailable. By merging data from all locations, we can discover new relationships between list 1 and list 2, even though they contain no shared resources and thus no direct links, even in the merged result set. We can suggest to the author of list 1 that Y could be a relevant resource, and given a high density of equivalent resources, we could make an assertion that lists 1 and 2 are potentially similar to a teacher pursuing a discovery use case. The power of linked data here is that although the sophistication of the inference algorithm can be increased, the merging of datasets across institutions remains trivial. 4.3 Wisdom of crowds By allowing teachers on-mass to re-mix, reuse and share lists to create derivative works, we introduce the wisdom of crowds into the system. The level of an individual teacher’s impact in his subject area could be formulated as a product of the quantity of derivative works in the ecosystem. In essence, we enable the most attributed content to float to the top of the pile. 14 http://vocab.org/aiiso/ 15 http://www.hesa.ac.uk/dox/jacs/JACS_complete.pdf Linked Data as a basis for Learning Resource Recommendations 7 When combined with the ability to follow the actions of others, we provide the context for wider social network functionality within the system.