A metadata profile to establish the context of small learning objects: the slicing book approach (1).

In order to re-use small learning objects their correct technical and semantic context must be preserved. The paper discusses this problem based on the experience of re-engineering a large corpus of pre-existing texts into learning objects of fine granularity.

WHAT IS THE CONTEXT OF A LEARNING OBJECT?

There are a number of definitions of what a learning object is. All of these definitions emphasize the reusability of learning objects. Only few of them require that learning objects must be self-contained in some sense. In our view learning objects, which are literally self contained, that is, do not require any prerequisites, do not exist. To learn from a learning object, the learner needs the ability to integrate the information contained in this object into his/her view of the world. This requires establishing connections of the new information with existing knowledge. But that is only possible if the learning object is somehow related to other information in a way that can be exploited to connect it also with the learner's prior knowledge.

In his speech "Food for Thought" March 3rd 2005 at the Prolearn workshop in Leuven Wayne Hodgins promoted the use of learning objects, which are so small that they could not be made any smaller without loosing their re-usability. The following article describes some of the context related problems encountered when dealing with such small objects.

In fact there are good reasons to handle small learning objects. An example of a small re-usable learning object is a single line exercise from a textbook which may be re-used in an online test. The smaller the learning objects are, the more flexibly they can be combined and re-used. However, this does not come without a price. When learning objects are small, it is likely that an application will need many of them, and all of them have to be described by metadata, which may be costly to achieve.

But on the other hand, when small learning objects are embedded in larger structures, it may be possible to attach many of the required metadata to the fewer but larger aggregations of learning objects, which can be seen also as learning objects themselves, and to let this metadata be inherited by the small learning objects which make up the aggregation, thus saving a lot of the costs of manual metadata assignment.

It is also worth noting that there may not be a general optimal granularity of learning objects. A beginner trying to understand a certain subject may need the complete story with all arguments enclosed while an expert may want to lookup and reuse just the few small learning objects which are relevant for his/her actual case.

Small learning objects almost always are too small to contain all relevant information which must be understood in order to learn from them. Consider the following example.</p> <pre> If x and y are negative integers

* what is the sign of x* y? * what is the sign of x + y? </pre> <p>We may consider this as a self contained learning object if we presume that the learner knows what are negative integers, addition, and multiplication.

When we consider instead just the question "what is the sign of x+y?" taken from this exercise it may not be reused without the information that x and y denote negative integers. Re-establishing this context is necessary for a proper reuse.

Let us slightly modify this example and look at it's source code, say from an HTML page:</p> <pre> If <i>x</i> and <i>y</i> are negative integers, <i>f</i> and <i>g</i> are the functions introduced in <a href="definition2p1">Chapter 2, Definition 1</a>, then <ul>

<li> what is the sign of <i>f(x,y)</i>?</li> <li> what is the sign of <i>g(x,y)</i>?</li> </ul> </pre> <p>We observe that each of the questions in this example has two strictly different kinds of context:

* To produce a technically correct document containing any of the list items, the <ul> ... </ul> context must be restored. We call this the technical context of the learning object in question.

* To produce a semantically correct document containing any of the list items, the introductory sentence and the required definition from Chapter 2 must be included. We call this the semantic context of the learning object in question.

We shall explain below how open specifications can be profiled to allow restoration of the technical as well as of the semantic context of learning objects for Slicing Book Technology.

BASICS OF SLICING BOOK TECHNOLOGY

Slicing Book Technology (Dahn, 2000, 2001) consists basically of the following four steps.

1. Existing documents are sliced into reusable learning objects.

2. The learning objects are augmented with metadata

3. Declarative rules describe which content should be presented to users depending on the user profile and the users's learning objectives (learning scenario)

4. On request of the user an automated inference system combines available information about the user and metadata of the available learning objects with the rules describing the requested documents in order to recommend a selection of learning objects to be presented to the user and possibly also to recommend an available style for presentation. The document is generated, possibly after further modification requested by the user, on the fly, and delivered to the user over the network.

We mention that Slicing Book Technology includes all elements of Semantic Web Technology, except that it is not assumed that the repository of objects is distributed. Slicing Book Technology can be applied to all well structured documents. Such documents are frequent in natural sciences, but also in some fields of social sciences, legal documents, or technical documents. Also documents in public administration or internal documents in companies frequently have a standard prescribed structure. Especially XML documents can be prepared for the application of Slicing Book Technology with a high degree of automation.

During the years 2000-2003, the European project Trial-Solution investigated the particular issues which emerge when this technology is applied to a number of books which have been previously created by different authors without regard to the technology. The Trial-Solution sliced books library contained 25,000 learning objects from 5,000 pages of text.

During the project a complete workflow was established, which covered automated and manual reengineering, metadata maintenance, and delivery. The tools developed in the Trial-Solution project exchanged sliced books in the form of IMS Content Packages. To meet the needs of the technology, the IMS Metadata DTD of that time had to be considerably modified. Current versions of IMS and IEEE LOM Metadata Schema and the new IMS VDEX Vocabulary Definition Exchange specification offer more possibilities for profiling.

We will explain how these possibilities are applied to map the Trial-Solution Metadata to a profile of current IMS Metadata.

RESTORING THE TECHNICAL CONTEXT--THE HIERARCHICAL TREE FORMAT

The Hierarchical Tree Format is a simple format for the representation of sliced books. It is conceptually fundamental for Slicing Book Technology, though it is technically frequently replaced by other representations for efficiency reasons. So this format is not intended to prescribe any particular kind of implementation.

The Hierarchical Tree representation of a sliced book consists of a tree. The upper levels of this tree are in a one-to-one correspondence with the items in the table of contents of the original integrated preexisting document. Depending on the intention of reuse and on the possible reengineering efforts, this hierarchy may be further refined. Thus for the example given there would be a node in the tree for the group of questions and this node has for each of the questions a separate child node.

The content of the document to be sliced is distributed into files which are attached to these nodes. Usually this slicing process can be done with a high degree of automation. After this process each node can have a number of files attached. Usually there will be a file to be included into a document before any content taken from the subtree rooted at this node (a start file) and a file (the end file) to be included after all content from that subtree. For example the start file of the example above would contain:</p> <pre> If <i>x</i> and <i>y</i> are negative integers, <i>f</i> and <i>g</i> are the functions introduced in <a href="definition2p1">Chapter 2, Definition 1</a>, then <ul> </pre> <p>while the end file would contain just</p> <pre>

</ul> </pre> <p>The leaves of the Hierarchical Tree will have attached content files, in our example the list items.

Before we continue we note that each learning object represented by a node of the tree may consist of a number of files, which are to be used in different roles (start or end file or also a variant suited for a particular presentation mode). The metadata application profile will have to take this into account.

The important feature of the Hierarchical Tree Format is, that for any selection of nodes of the tree the required technical context is recreated whenever the smallest subtree containing the selected nodes and rooted in the root of the complete tree, is traversed in a depth first manner and files attached to it's nodes are used according to their role. We note that all information necessary to reconstruct the technical context of a learning object are contained in the nodes on the path leading from the root of the Hierarchical Tree to the directory representing this object. Therefore we can state the following:

Path Inclusion Principle: Whenever a learning object is extracted from a sliced book for reuse, all information on the path leading from the root of the Hierarchical Tree to the node representing this object must be extracted too.

Quoting parts from a sliced book can be more precise than quoting a conventional text since the smallest reusable part can be referenced directly. To support a human readable form of such quotation, the concept of a sourcereference has been introduced and added to the metadata definition. The sourcereference of a node in the Hierarchic Tree is an integer, which denotes the position of this node among all nodes having the same parent directory. This sourcereference in combination with the path inclusion principle and the sourcereferences found on the path to an object assigns a unique sequence of numbers to each learning object, which describes exactly its position in the original work. Sourcereferences are also necessary to determine the sequence in which the nodes of a Hierarchical Tree have to be traversed.

RESTORING THE SEMANTIC CONTEXT--RELATIONS

Unlike for the technical context, restoring the semantic context requires domain expertise. In Slicing Book Technology this expertise is captured in assigning relations between learning objects and in defining declarative rules which specify how these relations are to be used to reconstruct an appropriate semantic context. The following set of relations has been selected as a minimal independent set. Other relations are inferred at runtime as needed or precalculated.

* To understand object A, the content of object B must be understood. We note that this is closely related to the surmise relation between topics discussed in Heller, Steiner, Hockemeyer, & Albert (2006, this issue).

* Each document containing object A must include object B.

* To understand object A, some part of the content of object B must be understood but A may be understood without understanding all of B.

* Object B contains an explanation for object A.

* Object B is a translation of object A.

Besides these, External relations are established between slices and external objects like external HTML pages.

The rules that are used to construct the semantic context also make use of other information that must be encoded in metadata. These are:

* key phrases taken from a structured thesaurus, and

* information about the semantic type of an object, for example whether it is an exercise or an example.

DEFINING METADATA

The Hierarchical File Tree Format can be encoded easily as an IMS Content Package, mapping the hierarchy to the default organization of the package. This gives the possibility to add metadata to the items of the default organization representing the directories of the file tree as well as to the individual files. Now the Path Inclusion Principle previously stated implies that with each item of a content package representing a sliced book also all metadata on the path to this item is extracted.

This gives the possibility to distribute the relevant metadata on this path and to avoid duplications. For example, information about the author and about the terms of reuse need to be kept only once at the top item. This leads to the following:

Metadata Inheritance Principle: Metadata can be inherited downward in the default organization of a content package. Inheritance must be declared.

PROFILING SPECIFICATIONS

The structure of sliced documents can be defined completely using the IMS Content Packaging specification. However it was not possible to capture all the metadata described above using one of the usual metadata schemas. Hence the IMS Metadata Specification was profiled. This IMS metadata profile consisted, in accordance with the forthcoming IMS Application Profiling Guidelines, of a description of the deviations from the IMS Metadata Specification.

The application profile suggested for supporting Slicing Book Technology is a mild extension of the IMS Metadata Specification 1.2.2, that is, the specification is only extended at the defined extension points. The relation and classification elements are optionally extended by information specifying the author of the metadata assignment, the status (whether it is authorized by some metadata maintenance authority), a declaration of inheritance and by a human readable description of the metadata assignment. The kind of a relation element is to be taken from a fixed flat vocabulary.

For the learningresourcetype element the use of a domain specific multilingual flat vocabulary is prescribed. The technical element is extended by an integer denoting the sourcereference as explained earlier and by a resource element, which in turn is extended by information about the role of the resource for a specific requested form of output.

Key phrases are described as taxons where the source refers to a multilingual hierarchical VDEX vocabulary. The aforementioned use of extension points for the relation, classification, and technical elements requires the addition of schemas for three new namespaces, which are referenced by the application profile. Note that the requirement of usage of a correct vocabulary is an additional constraint, which cannot be described as a modification of a single XML schema.

The problem that now remains to be solved is to handle multiple files that make up a learning object where these files have different roles. Unfortunately, the IMS Metadata location element does not have an extension point. Therefore, to modify the base specification as little as possible, it is proposed to let location point to a directory in the content package and to let the values of the resource element point to the files which make up this learning object.

No other changes of the IMS Metadata specification are required.

The Path Inclusion Principle and the Metadata Inheritance Principle constitute an intimate connection between the structure and the metadata. In contrast with the Content Packaging and Metadata Specification and their profiles they do not pose restrictions on the structure of documents but rather on allowed uses of sliced documents. Therefore they may be rather seen as specifications, which prescribe a certain behaviour of services using these sliced documents--the Path Inclusion Principle sets a condition for further disaggregation and the Metadata Inheritance Principle is relevant for services which search for the small learning objects inside sliced documents.

BENEFITS OF METADATA FOR SMALL LEARNING OBJECTS

While Slicing Book Technology has been mostly applied to add value to preexisting textbooks it is by no means restricted to texts. For example, it has been used within the German project In2Math to create on the fly interactive pdf documents where the content of some parts of was created dynamically on the server taking the user input and the user profile into account. One of the simplest applications of this is the dynamic random generation of examples or exercises adequate for the knowledge and interests of a particular user. More complex applications realized in In2Math permit the user to call an interface to an interactive system such as a computer algebra system, theorem prover, or note taker and to incorporate the output of this system as a new learning object into new course material.

Even if only a flat list of combinable learning objects with metadata and relations between them is available, the technology can be applied since the rule based inference engine used in Slicing Book systems can make use of the metadata and relations to produce content packages or documents for a particular learning purpose. As an application of the same technology in a different field, a congress planner was developed by Slicing Information Technology GmbH, which created for participants of the Dutch Kennis-congress a selection of the many concurrent sessions according to the participant's interest.

For historic reasons it may be worth mentioning that the Trial-Solution project in 2001 decided to develop its own metadata DTD since the available metadata specifications did not allow an encoding of the required information explained earlier. Only recently with the availability of extension points in the specifications and with the support for structured vocabularies in the VDEX specification the mapping to an application profile of the IMS Metadata specification previously described became feasible.

This mapping was developed within the European project Telcert. A moderate collection of 500 small learning objects from the Trial-Solution project now serves within the Telcert project as a set of test data for a content test system which is to be automatically adapted to application profiles of eLearning specifications.

References

Dahn, I. (2000). Symbiose von Buch und Internet. Proceedings of Learntec 2000, (pp. 551-558), Karlsruhe, Germany

Dahn, I. (2001). Automatic textbook construction and web delivery in the 21st century. Journal of Structural Learning and Intelligent Systems, 14(4), 401-413

Heller, J., Steiner, C., Hockemeyer, C., & Albert, D. (2006). International Journal of E-Learning, 5(1).

Note

(1) Work supported in part by the European Commission within the project TELCERT (Contract Nr. 507128)

INGO DAHN

University Koblenz-Landau--Koblenz, Germany

[e-mail]