Flexibility of Automatic Authoring for the Semantic Web Alexandra I. Cristea Faculty of Computer Science and Mathematics, Eindhoven University of Technology PO 513, 5600 MB Eindhoven, The Netherlands +31-40-247 4350 a.i.cristea@tue.nl Abstract clearer to the research community [2], and various ways to simplify the authoring process are sought. In this paper The LAOS model, a 5-layer adaptive hypermedia (AH) we show how, next to regular authoring in the LAOS authoring model, was previously shown to specify a model [13], (adaptive, adaptable) novel automatic flexible framework for (collaborative) creation of authoring methods can be used for easier, more powerful material for the semantic web. However, for adaptive AHS authoring [10]. In this way we illustrate how the behavior, an author has to design not only basic semantic LAOS model supports semantic web techniques. Some of contents (and its alternatives), but also specify the desired these automatic techniques have been tested in practice dynamics of the system, which is rather cumbersome. [12], this being beyond the scope of this paper. The Therefore, automatic authoring techniques are being remainder of the paper is organized as follows. In section researched, that aim at decreasing the authoring burden. 2 we sketch the LAOS model. Section 3 elaborates on Here we elaborate on these techniques that can be built automatic transformations and machine interpretation of based on LAOS, and show specific implementations. They the information allowed by the LAOS model; we compute exploit the LAOS structure and consist of automatic flexibility degrees and give concrete examples from transformation (interpretation) rules between different MOT. Section 4 summarizes our contributions. layers of the model (populate some layers based on the contents of others). To evaluate the effectiveness of these 2. LAOS Layered Model transformations, we have to see if and how much flexibility is lost by performing these automatic The LAOS model (figure 1, [13]), is a generalized transformations, as opposed to fully manual creation. We model for dynamic adaptive hypermedia authoring, based shall see that even with these automatic transformations on the AHAM model [26]. The model comprises five in LAOS, high flexibility can still be achieved. layers: the domain model (DM), goal and constraints model (GM), user model (UM), adaptation model (AM) 1. Introduction and presentation model (PM). The revised and extended version of these components is shortly listed in the By embracing the goals of W3C and IEEE LTTF [16] following subsections. From a semantic point of view, communities towards (ontology-based [20]) these layers represent ontologies, with exception of AM personalization and the semantic Web [23], adaptive which specifies the interpretation and behavior of the hypermedia systems (AHS) are gaining nowadays more elements of the ontologies. Their definitions are used for popularity in different communities. They respond to one the explanation of the automatic transformations and of the main goals of the semantic web, that to allow should be best used as reference for section 3. By automatic semantic processing on the web. In the case of populating these ontologies, authors can create adaptive adaptive hypermedia (AH), this is achieved by adding the hypermedia for the semantic web. necessary data for ‘intelligent’ adaptive processing of web information. Mainly, this consist of contents Table 1. Generic Definitions. alternatives and user model data, which specify which of the alternatives is relevant for which (type of) user. Definition 1. Let CM be the set of all AHS concept maps. Successful research AHS such as AHA! [14], Interbook [5], TANGOW [6], but also commercial adaptive 2.1. Domain Model (DM) systems, such as Firefly, have proven the various benefits The DM contains the basic concepts1 of the contents and customization variants of AHS. One of the big and their representation, in the form of concept hindrances that stop the wider acceptance of AHS is the attributes2. lack of powerful authoring tools [3]. Nowadays, the importance of authoring research for AHS is becoming 1 Concepts in LAOS have to have a semantic unity. Next, we show domain model components definitions. Constraint 2. Each domain concept c must be involved in at least one special link l, called hierarchical link (link to ancestor concept). Exception: root concept. 2.2. Goal and Constraints Model (GM) The GM is a constrained version of the domain model (DM) above, with constraints based on a goal. The idea is taken from the book–presentation metaphor: when designing a presentation (GM), we usually base it on some reference(s) (DM). For instance, a presentation (GM) can be based on one or more books (DM)6. The GM therefore already gives the presentation a preliminary shape. The actual presentation seen by the LAOS user however can still contain not only GM but also DM elements (e.g., for more information about a concept from the GM, other attributes of the respective DM concept, or other DM concepts related to it can be referred). This latter fact actually increases the flexibility and semantic expressivity of the created adaptive presentations, as we shall see, but, more importantly, separates links based on content relatedness (DM) from links based on presentation structure (GM). Following are the definitions for the components of the goal and constraints model. The GM is defined analogous to the DM, so the Figure 1. The five level AHS authoring model. GM set of goal and constraints maps, GM map, gl goal Table 2. Domain Model Definitions. link and ga attribute definitions are skipped here. Definition 2. Let DM be the set of all domain maps (DM ⊆CM), Table 3. Goal and Constraints Model Definitions. containing all information (resources and links between them) of the AHS relevant to the domain of the resources. Definition 7. A goal and constraints concept g∈ GMi.G is defined Definition 3. A domain map DM of the AHS is determined by the by the tuple GA≠∅ is a set of attributes; G a set of tuple ; where C a set of concepts; L a set of links and Att a sub-concepts; DMj.c∈C is the ancestor DM concept7 and DMj.c.a∈A 3 set of DM attributes (DM ∈DM). is an attribute of that concept; GMi is the name of the GM map Definition 4. A domain concept c∈DMi.C is defined by the tuple instance to whom it belongs. ; where A≠∅ is a set of DM attributes; C a set of DM sub- Constraint 3. Each goal and constraints concept g must be involved concepts; DMi the domain map instance the concept belongs to. in at least one special gl, link called prerequisite link (link to 8 Definition 5. A domain link l∈L is a tuple with S,E ancestor concept) . Exception: root concept. ⊆{DMi.ck}i,k (S,E≠∅) start and end sets of DM concept instances4, respectively; N set of link labels; W set of link weights. 2.3. User Model (UM) Definition 6. A domain attribute a∈ DMi.C.A is a tuple , where type is the name of the DM attribute; val is the value (contents) of the DM attribute. UM can be a pure overlay model, as in AHAM [25], over Constraint 1. Each domain concept is required to have a minimal the DM and GM previously defined. Another possible set of (standard) attributes5, Amin (A⊇Amin≠∅). approach is to represent the user model [11] as a concept map, so that relations are also allowed. The UM is 2 Attributes in LAOS represent different aspects (views) 5 about the same concept; e.g., a title is also an aspect of a To specify what we REALLY want the authors to fill in. 6 concept. These attributes can be specified by any standard This is why the GM layer is so dense: from one DM, (e.g. IMS [17], LOM [19], SCORM [22], etc.) or be multiple GM versions can be generated. Similarly, for one designer-defined attributes. presentation, several books can be used, so the GM-DM 3 Note that these are attributes at the level of the domain relation is a multi-multi relation. 7 map, describing it directly, and not the concepts in it. Can be void. 4 8 ck is a concept instance in an arbitrary domain DMi. GM concepts are also expected to participate in one of Please note that the generic definition allows loop links the special links called AND/OR link (link to sibling between a concept and itself. In praxis, links can be added concepts), but as there is no constraint requiring the between any concepts of the owned domain maps to any number of siblings to be above zero, this cannot be concepts of the whole space of domain concept maps. mentioned as a constraint. defined analogous to the DM, so the UM set of user →DM) 3.1. From Domain Model to itself (DM→ maps, UM user map, ul user link and ua user attribute definitions are skipped here. This section discusses the automatic (adaptive, adaptable) DM enrichment, according to its existing Table 4. User Model Definitions. structure and contents. This means that implicit Definition 8. A user concept u∈ UMi.U is defined by the tuple ; AU≠∅ is a set of UM attributes; UMi.U a set of some information retrieval technique. We have already UM sub-concepts; GMi.g/DMj.c∈G/C is the ancestor GM (or DM) concept. treated some specific DM to DM technique in [8], so this section will only shortly resume those results and extend 2.4. Adaptation Model (AM) them. In [10] we described a new, three-layer adaptation 3.1.1. DM→ →DM: by concept attribute type. The easiest model (featuring: low level assembly adaptation way to enrich the domain model is by finding language, medium level programming adaptation automatically new links between existing concepts9. language and adaptation strategies) that we are in the In [8] we have developed formulas for relatedness process of refining and populating; this however, is relations generation, for relations between concepts that beyond the scope of the present paper. share a common topic. This commonality was computed at concept attribute level, and therefore could automatically been given a type that corresponded to the 2.5. Presentation Model (PM) (name of the) attribute type. In short, we could describe10 the domain links11 found by these computations as The PM reflects only the physical properties and following: environment of the presentation; the PM provides the bridge to the actual code generation for the different If ∃c1∈DM1.C1, ∃c2∈DM2.C2 (DM1,DM2∈DM), platforms (e.g., HTML, SMIL [24]). c1∈C1, c2∈C2, two domain concepts from two possibly different domain concept maps; The PM is defined analogous to the PM, so the PM set of c1=, c2=; ∃a1∈A1, ∃a2∈A2 presentation maps, PM presentation map, pl presentation two respective attributes of these link and pa presentation attribute definitions are skipped concepts, a1=, a2=. If var1==var2 (the attributes have the here. same type) a domain link can be generated l=<{c1},{c2},{var1},{weight}> with Table 5. Presentation Model Definition. weight>0 defined as: weight=number_common_features(val1,val2). Definition 9. A presentation concept p∈ PMi.P is defined by the tuple ; PA≠∅ is a set of PM attributes; PMi.P [12] gives different implementations of the function that a set of PM sub-concepts; GMi.g/DMj.c∈G/C is the ancestor GM (or CM) concept. computes the number of common features. 3. Adaptive, Adaptable Automatic Table 6. Flexibility Index Definition. Transformations Definition 11. The mixed link flexibility index is the number of possible (bidirectional) links of mixed type that can be generated As well known, to create material for the semantic web between a selected set of concepts. is more difficult than creating material for the web of the past, because of the machine processable extra The mixed flexibility index of the links that can be information needed. Even more so is the case when generated between concepts c1 (current concept) and c2 creating machine readable semantic information for the is as follows (with shorthand notation Ai=card(Ai)): adaptive semantic web. Therefore, we look at automatic generation of some of the LAOS layers, based on 9 information from others. Moreover, we also look at the Please note that these new links can be between the flexibility index for these automatic transformations, to be concepts of the current content (concept map: e.g., able to measure how semantically expressive and course), between the current content and some other computationally flexible these automatic generations are, content created by the same author, or finally between as opposed to manual population of the layers. the current content and some other content created by a different author. Table 6. Flexibility Index Definition. 10 Notations are from the definitions in section 2. 11 Definition 10. The flexibility index is the combinatorial index This is only one of the possible ways to connect computing the number of different outcomes that can be generated by concepts – stronger versions would look at ontological a given transformation. structures. 2 mixflex(1,2) = A1 A2 ≥ Amin ; If we want to consider links that have an unequivocal IF ENOUGH (L{V (c) | c∈ topic cluster}) type12, we obtain with the above notations the following THEN NEXT(topic) flexibility index formula: Where c is a concept in a concept list, L is the List operator and V is the View operator (as defined in [13]). flex (1,2) = card ( A c1 ∩ A c 2 ) ≥ card ( A min ) = Amin Therefore, different ways of automatically creating more where Aci is the set of attributes of concept i and Amin is expressiveness within the existing domain are possible, the minimal set of obligatory attributes, as previously and there is space for more research in this direction. defined. If we consider we have C=card(C) concepts in the domain 3.1.2. DM→ →DM: by Link Type. By having an algorithm map DM, then the flexibility index between concept c1 to check the link types, it may be possible (and beneficial) and the rest of the concepts in C is given by: to create new links. flex (1,*) = ∑ j = 2 card ( A c1 ∩ A cj ) ≥ (C − 1) Amin C However, the most important contribution of link analysis The mixed flexibility index between concept c1 and the would be to compare similar concepts13 and to find that rest of the concepts would be: some attributes (or even sub-concepts) are missing. Example 2: For instance, the concept called ‘Discrete mixflex (1,*) = A1 ∑ j = 2 A j ≥ (C − 1) Amin C 2 Neuron Perceptrons’ from a Neural Networks course has Generally, the flexibility index of concept map DM is an ‘Example’ attribute, whereas the concept ‘Continuous given by the following relation: Neuron Perceptrons’ doesn’t, although they are linked flex (*,*) = ∑ i =1 ∑ j = i +1 card ( A ci ∩ A cj ) ≥ C C via their ‘Title’ attribute with a weight of 67%. In such a case, the system could look for possible C (C − 1) ≥ ∑ i =1 ∑ j = i +1 Amin = C C Amin examples via other links to this concept, or signal the 2 author about the possibly missing content item (attribute, Similarly, the mixed flexibility index of concept map DM sub-concept, etc.). is: The flexibility index14 for this link-based concept attribute C (C − 1) retrieval from the link properties between the given mixflex(*,*) = ∑i =1 Ai ∑ j =i +1 A j ≥ C C 2 Amin 2 concept c1 (current concept) and some other concept c2 can be defined as: Example 1: To give a concrete example, in the MOT flex (1,2) = card ( A c 2 − A c 1 ) ≥ 0 adaptive hypermedia authoring system, Amin ={title, If we look at all the possible connections to c1, we obtain: keywords, introduction, text, explanation, pattern, flex (1,*) = ∑ j = 2 card ( A cj − A c 1 ) ≥ 0 C conclusion} so Amin=7; in the concept map called ‘Neural Networks I’ C=card(C)=145, so flex(*,*)≥10440*7= Finally, for a whole concept map DM, the flexibility index 73080 and mixflex(*,*)≥10440*49= 511560. is: flex (*,*) = ∑i =1 ∑ j = i +1 card ( A cj − A c i ) ≥ 0 C C Please note that these are the connections implied by only Depending on the variations in attribute design between one concept map. MOT already allows inter-linking of the different concepts, this value can be large or can be concept maps, which increases this number even more. zero. Therefore, it is obvious that a great number of Please note that an extended version of the content search connections can be generated automatically, in this way could look outside the space defined by the LAOS model, making the adaptive hypermedia process easier, while such as the transition from a search within a closed space adapting towards the authoring goal. to the Web space. This is an explicit, symbolic way of linking concepts. However, this is not the only way of automatically finding concept links. Some years ago, in a different research group, we developed a sub-symbolic technique 13 for concept clustering, based on SOM networks [18]. This Similar from a link-point of view, such as concepts clustering around topics can be combined with specific sharing the same ancestor-concept, e.g., or concepts at the level-operators, as defined in [10], to write (student) user same level of the hierarchy, or concepts related with each adaptation rules of the form: other via some special link (of a given type), etc. 14 For simplicity, we use the same notation for this link- based flexibility index, as we used for the concept-based 12 meaning that the attributes that determine the link are of index, although they obviously represent different values. the same type in both concepts, as stated by the link Here, this number represents the number of potentially definition. missing items (attributes). selected attribute subset will keep the same hierarchical 3.2. From Domain - to Goal and Constraints structure as its DM source. Model (DM→→GM) Example 6: If a concept ca was a sub-concept of concept cb in the DM and we use a similar transformation as in Here we look at automatic (adaptive, adaptable) GM the previous subsection, of choosing this time the {title, generation from the DM, according to presentation text} attributes (Atransf=2); then, the generated constraints and goals (e.g., for educational purposes we La1=ca.title and La2=ca.text will be sub-concepts of can envision pedagogical strategies or pedagogical Lb1=cb.title, and the former attribute cb.text becomes techniques). This transformation represents the first step concept Lb2, which is also a sub-concept of Lb1. from information to knowledge. This was better detailed Therefore, the hierarchical link structure in DM is in [9], here are the essentials only, as follows. transformed into a new hierarchical link structure for the GM15: Lb1 ⊇ Lb2, La1, La2. 3.2.1. DM→ →GM: by Concept Attribute Type. Concept Furthermore, concepts in the GM are ordered, as attributes can be grouped into types that determine a filter opposed to concepts in the DM: Lb1 > Lb2 > La1 > La2. for the selection of the items that will appear in the goal Moreover, relations in the GM are typed; they can be and constraints model. hierarchical, as describe before, or {AND/OR}. The latter are relations between elements at the same hierarchical Example 4: E.g., for Amin ={title, keywords, introduction, depth. In the MOT GM, all elements at a certain text, explanation, pattern, conclusion} (Amin=7) as in hierarchical depth are automatically transformed into section 2, if we define Atransf⊆Amin as Atransf ={title, concepts connected via an ‘AND’ relation. However, this keywords} (Atransf=2), the transfer set from DM to GM, can then be manually altered16: we can implement a goal-constraints model representing AND(Lb1 , AND(Lb2 , La1 , La2)). the elements for the pedagogical goal “short The illustrated link-based transformation above is simple, introduction” (e.g., for a very quick overview of the as it takes into account just the hierarchical link relations whole material). in the DM; however, it is useful in order to illustrate the Example 5: If Atransf ={title, pattern} (Atransf=2) we obtain many different types of links that can be generated for the a goal-constraints model representing the elements for GM from even such a simple link sub-set. the pedagogical goal “structural presentation” (e.g., for a review of the course). 3.3. From Domain - to Adaptation Model (DM→ →AM) The flexibility degree that can be generated (showing the different ways of selecting attributes from a concept c1, This transformation represents automatic (adaptive, considering that in the goal and constraints layer the order adaptable) AM generation from the DM, according to the of concepts is important, as opposed to the domain layer), (goal, e.g., pedagogical) strategy. The adaptation model is as follows: has the role to interpret the other models: the domain -, flex (1) = ∑ i =1 c 1 P (card ( A c 1 ), i ) ≥ ∑ i =m1in P ( Am in , i ) = goal – and even presentation model. Moreover, it can card ( A ) A update these models and generate the presentation. In [10] Am in ! = ∑ i =m1in we have defined the low-level adaptation (direct A ( Am in − i )! adaptation techniques) as: where P(a,b) are permutations of a elements taken b at a a : {DM, GM, UM, AM, PM} → time. So, the flexibility degree for one single concept and {[DM], [GM], UM, [AM], PM} its extracted attributes is flex(1)≥ 13699 (for Amin =7). If Function a can furthermore be divided into a set of concepts are transformed independently, e.g., in special sub-functions: a = {update, generate} where: groups, this flexibility degree can grow significantly. update : {DM, GM, UM, AM, PM} → {[DM], [GM], UM, [AM]} 3.2.2. DM→ →GM: by Link Type. Links in the domain generate : {DM, GM, UM, AM, PM} → {PM} layer are defined (section 2.1) as either hierarchical, or of These adaptivity functions a can be written as (are other nature. These link types can be used to generate equivalent to) IF-THEN rules or Condition-Action (CA) specific links at the level of the GM model. rules as defined in [26]. The simplest way is to select for the GM model only links of a specific type (e.g., only hierarchical links). In MOT, 15 automatic transformations of hierarchical links are used to which can be regarded also as a hierarchical inclusion create a hierarchical, ordered link structure; i.e., the relation. 16 e.g., into weighted ‘OR’ relations, not further detailed here. 3 Automatic transformation from the domain model to  A Am in !  the adaptation model means to interpret the existing DM ( )3 flex (1) = ∑ i =1 C ( Am in , i ) =  ∑ i =m1in Am in  to generate adaptation rules. This can be done at the  ( Am in − i )! i!  adaptivity function level that is described above, or at a For Amin =7, flex(1)=(87)3= 658503, which is a huge higher level of adaptation language or adaptive strategies number. We obtain such a huge number because the (these levels are defined in [10]). That would mean that, events of having ‘access’ states on the left, ‘available’ instead of assigning a specific transformation for a given states on the left and ‘available’ states on the right are link type (or concept type), the same link (or concept) independent, meaning that for each state determining the could be transformed differently, according to a different attributes that appear as ‘access’-ed on the left side of the (e.g., pedagogically rooted) adaptation strategy. Here we IF all combinations of attributes with ‘available’ on the are going to refer to low-level automatic transformations left are possible, etc. So, basically, even for a very limited (CA level) and some adaptation language-level oriented situation with 2 states and only generic rule generation, a automatic transformations. great number of adaptive rules can be automatically Please note that normally the AM is supposed to work written, based on the authoring goal (inferred or not by only with the data in the GM, as this is already pre- the system). selected for presentation. 3.4. From Goal and Constraints - to Adaptation 3.3.1. DM→ →AM: by Concept Attribute Type. Attribute Model (GM→→AM) types can be used to show only specific attributes in specific conditions. These conditions can be This represents automatic (adaptive, adaptable) AM automatically deduced by the system (as in adaptivity) or generation from the GM, according to an adaptation triggered by the AHS user (adaptability). strategy or technique (e.g., based on a pedagogical Example 7: For instance, a specific automatic adaptive strategy or technique). This type of transformation is rule can express the fact that we only want to show the more natural to the design of the LAOS structure, as the ‘text’ attribute of concept c1 after the ‘title’ and existence of the GM model supposes a pre-selection of ‘introduction’ were read: the material that is to be presented to the hypermedia IF(c1.title.access=’true’ AND user, according to some (pedagogical) goal and delimited c1.introduction.access=’true’) by some (spatial, time, pedagogical, etc.) constraints. THEN c1. text.available=’true’; Please note that we wrote the condition for simplification 3.4.1. GM→ →AM: by Link Type. The GM, as said, purposes in this form, but that attribute states such as contains pre-ordered and pre-selected information from ‘access’ and ‘available’ are part of the user model17. In the DM. This structure can already be interpreted in terms order for this to be a general automatic transformation of the adaptation that is to be performed on it. For rule, for any concept C in the DM, all concepts in the UM instance, the GM allows ‘AND’ relations between that reflect the DM should have also attribute states concepts, as well as ‘OR’ relations with some weights. ’access’ and ’available’, and the following low-level rule Example 11: These can be used to express that all has to be added to the AM: concepts in a ‘AND’ relation should be read, for IF(c.title.access=’true’ AND c.introduction.access=’true’) instance: THEN c. text.available=’true’; IF ((c.name.access=’true’ OR c.contents.access=’true’) If generic rules as the one above are permitted, for each AND link(c,c2,’AND’,*)) such transformation [26] only one rule will be added. The THEN { c2.name.accessible=’true’; number of possible rules to generate is potentially c2.contents.accessible=’true’;} infinite, because it is dependent also on newly added Example 12: In a similar way, an ‘OR’ relationship can states into the UM, which can be numerous. If we be interpreted as inhibiting the reading of the other consider the case where only s=2 such states can be concepts in the same relationship18: added, as in the above example, and even more, we IF ((c.name.access=’true’ OR c.contents.access =’true’) enforce the restriction that the ‘access’ state can only be AND link(c,c2,’OR’,*) ) found on the left side, while the ‘available’ state can THEN { c2.name.accessible=’no’; appear on both sides of the rue, we obtain for the c2.contents.accessible=’no’;} flexibility degree: Example 13: A more informed version of the above would check the weight of the current concept, to see if it is 17 more precisely, part of the overlay part of the UM, because the UM can contain also other attributes such as 18 user’s prior knowledge, user’s interest, etc., that are not In such a case, an ‘OR’ relationship acts actually as a an overlay model of the DM (or GM). ‘XOR’. above some threshold, before deciding to inhibit another adding an extra condition that the interest to be changed concept: should at least be positive: IF ((c.name.access=’true’ OR c.contents.access =’true’) IF (c.interest > 0 AND LINK(c,c2,’influence’,*)) AND link(c,c2,’OR’,w) AND w>threshold) THEN { c.interest= c.interest – c2.interest;}. THEN { c2.name.accessible=’no’; c2.contents.accessible=’no’;} In such a way, various constructs can be automatically 4. Discussion and Conclusion added to the generic adaptation rules, directly by interpreting the goal and constraints model. Reducing the authoring burden has been identified as one of the major priorities in adaptive hypermedia [2] 3.5. From User - to Adaptation Model towards creating material for the semantic web. There are (UM→→AM) many ways of achieving this. In this paper we have approached the issue of improving and making AHS The user model can be a simple overlay model of the authoring easier by enumerating a number of different DM (as in [26]) or a more extended model, represented types of automatic (adaptive, adaptable) transformations also as a concept map, as defined in section 2.3. For the that can be directly performed by the adaptive first case, the user model just generates variable-value hypermedia authoring system, as shown in section 3. pairs, which can enter conditions in adaptive rules or These possible automatic authoring techniques (or which can be modified by these rules. For the second transformations) are based on the data design given by the case, not only the variable-values are important and LAOS model, which allows a concept-oriented approach interesting, but also the relationships between the for data design, analysis and usage. For exemplifying the concepts themselves, which together form the UM. transformations, we first reviewed LAOS, the five level AHS authoring model that allows a clear-cut separation 3.5.1. UM→ →AM: by Concept Attribute Type. In the of the representation levels: the domain model (DM), the Example 14: To illustrate a pure usage of UM elements goal and constrain model (GM), the user model (UM), only to generate an AM rule, we consider the same state the adaptation model (AM) and finally the presentation of ‘interest’ about a concept, which is extracted from the model (PM). overlay model of the UM. We want a rule that displays Here we have shown a glimpse on the great number of everything in the concept, if this concept is of interest to different design possibilities that these automatic the user. The conditions on the left side of the rule will be functions still allow, given the existing structure, showing part of the HM, while the resulting action on the right that the authoring capacity is not inhibited by the added side will be a part of the FM: automatic authoring functionality. The range of IF (c.interest > threshold) possibilities of outcomes was computed in the form of a THEN { c.name.available=’true’; flexibility degree, which shows also the range of the c.contents.available=’true’;} adaptivity of the final system. We have introduced and Please note that we have used for both sides concepts computed the flexibility degree offered by such from the GMw (and not the DM). transformations for different example cases, and we have Moreover, please notice that this rule is again a generic discussed the significance and extension possibilities of rule, which can be applied on all concepts in a concept some of these transformations. Although these map, therefore drastically reducing the workload. transformations have been discussed and analyzed separately (for instance, DM to AM transformation was 3.5.2. UM→ →AM: by Link Type. Link type can only be analyzed apart from GM to AM transformation, etc.), in used when the UM is itself a concept map. In this way, practice it is reasonable to expect that these we can express for instance the fact that two states in the transformations can be in parallel. The combination of UM are related. different transformations may be leading to a situation Here, however, we try to look at a different type of link where one transformation may be setting some between UM concepts. For this, let’s consider the link of restrictions on another one, but most of the time, these type ‘influence’. multiple transformations together will generate a higher Example 15: We will add a rule saying that the interest in flexibility degree. We have not extended all the examples a subject c might decrease if the user is interested in or computed the flexibility degree for all the cases, as the another subject c2. space in the paper did not permit it. Moreover, we have IF LINK(c,c2,’influence’,*) skipped some transformations, such as the ones from the THEN { c.interest= c.interest – c2.interest;} GM to the PM. Instead, we have tried to give an overview Example 16: Or if we want, for instance, to prevent of the flavor of the different possible automatic infinite loops, we limit the application of this rule by transformations, their applicability and their diversity. It is interesting to consider, for future research, the Adaptive Hypermedia and Adaptive Web-Based combination of these automatic transformations and, e.g., Systems. presentation strategies bound to specific cognitive styles. [10] Cristea, A.I., and Calvi, L. The three Layers of We expect that applying such strategies would affect Adaptation Granularity. UM’03. Springer. several layers at once. Another direction to pursue is to [11] Cristea, A.I., and Kinshuk. Considerations on compare our work with and use specifications given by LAOS, LAG and their Integration in MOT. ED- [15]; in [15], formal concept analysis is presented, that MEDIA’03. allows discovering of patterns between application data, [12] Cristea, A., De Mooij, A. Evaluation of MOT, an on one hand, and the usage of concepts, relations and the AHS Authoring Tool: URD Checklist and a special semantics given by their hierarchies, on the other hand. evaluation class, CATE'03 (International Conference on Computers and Advanced Technology in Education) 5. Acknowledgements Rhodos, Greece, IASTED, ACTA Press, ISBN 0- 88986-361-X, pp. 241-246. This research is linked to the EC project ADAPT [13] Cristea, A., De Mooij, A. LAOS: Layered WWW (101144-CP-1-2002-NL-MINERVA-MPP). AHS Authoring Model and its corresponding Algebraic Operators. In Proceedings of WWW’03, Education track. (Budapest, Hungary 20-24 May 2003). ACM. 6. References [14] De Bra, P. and Calvi, L. AHA! An open Adaptive [1] 2L690: Hypermedia Structures and Systems, Hypermedia Architecture. The New Review of Lecturer: Prof. De Bra. Hypermedia and Multimedia, vol. 4, Taylor Graham http://wwwis.win.tue.nl/~debra/2L690/ Publishers, 1998,115-139. [2] Brusilovsky, P. Developing adaptive educational [15] Ganter, B., Wille, R. Formal Concept Analysis. hypermedia systems: From design models to authoring Mathematical Foundations, 1999, Springer, ISBN: 3- tools. In: T. Murray, S. Blessing and S. Ainsworth 540-62771-5. (eds.): Authoring Tools for Advanced Technology [16] IEEE LTTF, Learning Technology Task Force. Learning Environment. Dordrecht: Kluwer Academic http://lttf.ieee.org/ Publishers, 2003. [17] IMS (Instructional Management System) [3] Brusilovsky, P. Adaptive hypermedia, User http://www.imsproject.org Modeling and User Adapted Interaction, Ten Year Anniversary Issue (Alfred Kobsa, ed.) 11 (1/2), 2002, [18] Kayama, M., Okamoto, T. and Cristea, A.I. 87-110. Exploratory Activity Support Based on a Semantic Feature Map. In Proceedings of AH’00, LNCS 1892, [4] Brusilovsky, P., Schwarz, E., Weber, G. ELM- Springer, 347-350. ART: An intelligent tutoring system on world wide web. In Proceedings of International Conference on [19] LOM standard, http://ltsc.ieee.org/wg12/ Intelligent Tutoring Systems (ITS’96) (Montreal, [20] Mizoguchi, R., Bourdeau, J. Using Ontological Canada, June 1996), 261-269. Engineering to Overcome Common AI-ED Problems, [5] Brusilovsky, P., Eklund, J., and Schwarz, E. Int. Journal of AI in Education, 11 (2), 107-121. Web-based education for all: A tool for developing [21] MOT online, http://e-learning.dsp.pub.ro/mot/ ; adaptive courseware. Computer Networks and ISDN http://e-learning.dsp.pub.ro/motadapt/ Systems, In Proceedings of Seventh International World Wide Web Conference (14-18 April 1998) 30 (1-7), [22] SCORM standard, 291-300. http://www.adlnet.org/index.cfm?fuseaction=scormabt [6] Carro, R. M., Pulido, E. Rodríguez, P. Designing [23] WC3, Semantic Web. Adaptive Web-based Courses with TANGOW .In http://www.w3.org/2001/sw/ proceedings of the 7th International Conference on Computers in Education, ICCE'99 (Chiba, Japan, [24] W3C, SMIL, Synchronized Multimedia November 4 - 7, 1999) V. 2, 697-704. Language. http://www.w3.org/AudioVideo/ [7] Cristea, A.I., Evaluating Adaptive Hypermedia [25] Wu, H., De Kort, E., De Bra, P. Design Issues for Authoring while Teaching Adaptive Systems. Proc. of General-Purpose Adaptive Hypermedia Systems. In SAC, Track ELS’04, ACM, 929-934. Proceedings of the ACM Conference on Hypertext and Hypermedia (Aarhus, Denmark, August 2001) 141-150. [8] Cristea, A.I., Automatic Authoring in the LAOS AHS Authoring Model, Hypertext 2003, Workshop on [26] Wu, H. A Reference Architecture for Adaptive Adaptive Hypermedia and Adaptive Web-based Hypermedia Applications, doctoral thesis, Eindhoven Systems. University of Technology, The Netherlands, ISBN 90- 386-0572-2. [9] Cristea, A.I., Automatic Authoring in the LAOS AHS Authoring Model, Hypertext 2003, Workshop on