In this paper, we present a full system that provides a personalized educational service by making use of the new ePub3 eBook standard. The system has been developed in the framework of the iRead Project [7] which aims to develop tools for personalized learning of reading skills. Based on a structural representation of the task \learning to read in a speci c language" which captures linguistic expert knowledge, a pro le (for each user) which re ects the users competence in mastering the reading skill is created and maintained. Based on the user's pro le as well as her reading history, a new personalized interactive activity eBook which contains educational content (ePub3 embeddable mini-games) appropriate for the current mastering level of the speci c user is created on demand. We present in detail the main concepts behind this \authorless" activity book creation: domain modeling and user pro ling, learner pro le adaptation, personalized content selection.
iRead is the acronym for the \Infrastructure and integrated tools for
personalized learning of reading skills" European Union Horizon-2020 funded project [
Consider a typical paper activity-book that aims to help a learner develop her reading skill. Such a book may be full of activities of the form \which (out of the given two or three options) is the correct way to spell a given word", \ ll the ending in the following words", or \underline the spelling errors in the following text". We have all grown up with such books containing activities and exercises given as practice material. One characteristic of these activity books is that they are of a static nature in many aspects: They can be used only once, unless lled with a pencil and then erased. No matter how many times the reader goes over the book she always sees the same material. Usually the only feedback provided for a erroneous response by a reader is just the correct answer with no clues on what went wrong. Finally, the learning material is limited in size; she may need to practice more but no more activities are available.
Besides the above limitations and having decided to buy such an activity book, we face a more serious question: \which is the best book to buy?". The answer to this question requires \to know" the intended reader. In our example, the parent has to select the book that matches the mastery level of the child's reading skill. But, is there a variety of alternatives available to a parent? The reading activity-books are all classi ed to a small set of categories (based on age or school-year) and the parent has to select one of them. Usually, there is little hope to locate a book composed of the right mixture of spelling activities: a (small) set of reinforcing activities on the items the child has mastered, a large set of activities on the items the child has made partial progress towards mastering, and a (small) set of activities on the \harder" items that are to follow. From the above discussion, it is obvious that we would like to move from the classical paper activity-books that are static and non-personal to activity eBooks that are interactive and personalized.
Besides the typical interactivity of an eBook (e.g., turning pages, adding
bookmarks, changing the font-size), the latest IDPF eBook speci cation, ePub3
[
One particular feature being present under the new ePub3 standard is the ability of an interactive eBook to communicate with a (remote) server and store to it the reader's responses/answers, i.e., the reader's reading/playing history, for further processing. Based on the user performance on speci c activities (information present in the user's reading history), we can build a learner pro le which re ects the current level of mastering of the reading skill. In turn, by also utilizing our knowledge on the learning material/activities the learner has been exposed to through her reading so far, we can prepare in an automatic way (i.e., author-less ) a new personalized activity book which aims to advance the current level of the reading skill.
In order to be able to automatically capture the mastering level of the reading skill, we must identify the features that are important in mastering the reading skill and the relations among them. These features and their interrelationships constitutes a domain model, the construction of which require incorporating linguistic expert knowledge. Having identi ed the features present in a domain model, we can estimate the mastering level, also referred as competence, of a particular learner for each feature. This array of competence-values constitutes the learner's pro le. So, assuming that small (large) competence values correspond to low (high) level of mastery, the goal of any reading skill acquisition program is, through interactive literacy activities, to bring the learner's competence for each domain feature to a level that it can be considered as being mastered. Our ability to update the learner's competence values based on the correctness of her responses to interactive activities is referred to as learner pro le adaptation. Of course, all the above assume that we are also equipped with a way to measure the learner's competence.
In iRead, we have implemented six JavaScript literacy mini-games which, when complemented with appropriate literacy content, they form activities capable of supporting the learning of multiple domain features. Each page of an iRead interactive eBook is devoted to a speci c domain feature and utilizes exactly one of the developed mini-games. In principle, the content of an iRead eBook is developed automatically and is always relevant to a speci c learner. The selection of features to be targeted, the mini-games to be utilized in the activities and the literacy material to be used, is based on the learner's pro le, the learner's reading/playing history, a set of language speci c content selection rules and a set of language speci c resources (dictionaries). The size (number of pages) of each eBook is typically user de ned.
The rest of the paper is structured as follows: In Section 2, we provide an overview of the architecture of iRead's personalized eBook system and of the supported user work- ow. In Sections 3, we examine the domain/learner models employed by iRead's eBook system. In Section 4, we describe how the iRead eBook system utilized learner pro ling in order to produce personalized activity eBooks. We conclude in Section 5. 2
iRead eBook System Architecture In this section we review the architectural design of the iRead eBook system4. We start our description with a brief presentation of a typical supported user work- ow which highlights the services provided to its users. 2.1
We assume that the learner is a child and that some actions are taken on her behalf from her parent. From the system's point of view, these two entities are not distinguishable; they are both treated as users of the system operating the same account. The iRead eBook system supports the following user work- ow: 1. The user [parent ] creates an account in the iRead eBook server. 2. The user [parent ] chooses the type of activity eBook to be created by selecting one of the available language domain models (currently supported: English, Greek). A pro le is created and initialized for the speci ed domain model. 3. On the eBook creator application, referred to as eBook synthesizer, the user [parent ] speci es eBook attributes (number of pages; one activity per page, which mini-games to include, etc). Default values are available. 4. Based on the learner's pro le and reading/playing history, the eBook synthesizer selects appropriate literacy content and generates a new activity eBook. 5. The user [parent ] downloads the generated eBook from the iRead eBook server to her tabloid. 6. (a) The user [learner ] opens and reads (plays) the activity eBook on her favorite ePub3 compliant eBook reader. During play, reading/playing actions (referred to as progress) are locally saved. (b) Locally saved data (progress) is synced to the iRead eBook server. (c) The learner-pro le is updated according to the newly received progress. 7. The user [parent ] can login to the iRead eBook server, to see her [learner ] overall progress, saved eBooks etc.
Step-6 is repeated as long as the learner reads the eBook. Step-6a (data sync)
and Step-6b (pro le update) are transparent to the user.
4 For a description of the architecture of the complete iRead system see [
part of Figure 1 and it consists of the following parts:
{ eBook reader. Any commercial eBook reader running on an android device
should work. Our interactive activity eBooks have been successfully tested
on the following readers: ePub Reader for Android [
The save/sync server app is also used to synchronize it's local database data with the eBook remote server's database which containing the progress of all users. A unique id embedded in every eBook upon construction, associates it with the user the eBook was created for.
Figure 1 and it consists of the following parts: { User interface web service. It handles user account creation, account management, access to services and data etc. It is not shown in Figure 1. { eBook Synthesizer. It is the component responsible for the generation of personalized eBooks. Through the eBook synthesizer web interface the user [parent] speci es the characteristics of the interactive eBook to be created (e.g., topic [domain model], number of pages, which mini-games). Typically, the eBook's content is generated automatically based on the learner's model and her playing/reading history, however, the user also has the option to manually specify the activities that populate the eBook's pages. Then, the eBook synthesizer requests, trough an appropriate API, related content (according to the user's options) from the content provider. Based on the received content, it composes and generates the interactive eBook (according to the ePub3 standard). Finally, the eBook synthesizer informs the user-data database about the eBook-id of the generated eBook, associating it to the user who requested it. { eBook progress synchronizer. It is the component which communicates with the save/sync server (running on the user's android device) in order to sync the learner's progress. In this way, a complete reading/playing history for each is maintained at the eBook server which, in turn, can be used to update the learner's pro le. { User-data database. It is used to store user-account info, the generated eBooks, and playing/reading history for all users. { Content provider. The content provider, is responsible for generating the content to be included in a new personalized eBook. In doing so, it also maintains (stores and updates) the learner's pro les. It employs a local Model/Resources database which stores domain-models (topics), related resources and mini-games. Upon request, the content provider generates appropriate educational content and forwards it to the eBook synthesizer system. The automatic content selection is based on the user's pro le and playing/reading history. Learner pro le maintenance and update is the task of the user pro ling and adaptivity component (see Section 4).
The content provider can also utilize external source in order to receive user
modeling services and/or relevant resources. In Fig. 1, the iRead project[
Domain models constitute a central component of iRead which inform all
subsystems (NAVIGO Game, AMIGO Reader, Text Classi cation, Personalized
eBooks) and determine the individualized pedagogical design used for each child.
The role of each domain model is to (a) specify the linguistic structures, referred
to as domain features, that must be mastered when learning to read as well as
progression schemes, in the form of prerequisites, indicating the order of
teaching of those features, and (b) provide the basis for the formulation of individual
learner models, which enable the iRead system to record each learners individual
strengths and weaknesses. The linguistic information included in domain models
was especially selected to address all the linguistic sub-skills that are relevant
or necessary for reading development and covers awareness of speech sounds
(phonology), spelling patterns (orthography), word meaning (semantics),
grammar (syntax) and patterns of word formation (morphology) [
Domain models are represented and stored within the iRead infrastructure as weighted directed acyclic graphs (DAGs for short). Vertices of the graph correspond to language features while directed edges indicate prerequisites. Naturally, in such a representation, there can be no directed cycle, while the non-negative weight on the edges allows to express the fact that two di erent features can be both prerequisites of a third feature, but one of them is more important than the other. In the case where a language feature has prerequisites, the sum of the weights of its incoming edges must be equal to 1.
Using weighted DAGs allows to e ciently instantiate learner models, most commonly referred to simply as pro les, and store a learners progress. Pro les are instantiations of domain models, which can be personalized to each learner via their interactions with the interactive educational components (smart literacy games and interactive eBooks). Pro les are also represented as weighted DAGs (in fact, a copy of the corresponding domain model graph), where each vertex is equipped with an additional attribute, called competence. Competence is a nonnegative integer value showing the current mastering of the underlying language feature. When a learner exercises a particular language feature, the value of the competence can increase or decrease, depending on her performance. The value of competence cannot vary arbitrarily: for every language feature we de ne its # of Features (vertices) # of Prerequisite relations (edges)
/c/-c 6 7 0.3 /k/-c 5 7 0.2 /s/-ce 0 /s/-s 8 7 Fig. 2. Domain modeling: Unlocked and locked features. Unlocked features are shown in green, locked in yellow. Unlocked edges are drawn black, locked red. We assume that the unlock value of all vertices is equal to 0.75 and the competence threshold value for unlocking is 5. In the leftmost case, the bottom feature is locked as the sum of the unlocked incoming edges is equal to 0.5 (i.e., smaller than 0.75). In the middle con guration, the corresponding sum is equal to 0:8 (i.e., greater than 0.75) and becomes unlocked, while in the rightmost con guration, although the sum of the unlocked incoming edges is less than the unlock value of 0.75, the competence of the feature is 5, equal to the threshold value. Thus, the feature becomes unlocked. minimum and maximum value. By default, and unless otherwise speci ed within the domain model, we consider the minimum value to be equal to zero, and the maximum value equal to ten.
For a particular learner, not all language features are \available" for practicing. It is clear that a learner can practice a language feature if she has mastered the feature's prerequisites. In order to express the \availability" of a language feature, we equipped both vertices and edges of the graph with the unlock value attribute. The availability status of a vertex (or edge) can be either locked or unlocked, i.e. available or non-available respectively. An edge becomes unlocked if the competence of its source is higher than the unlock value of the edge, otherwise it is locked. Vertices can be also unlocked in three ways: (i) if a vertex has no prerequisites it is, by default, unlocked, or (ii) if the sum of the weights of their incoming unlocked edges is higher than the unlock value of the vertex, or (iii) if the competence of the feature is higher than a competence threshold value. As an example, consider the con gurations shown in Figure 2: vertices are drawn as rectangular shapes and their label shows the corresponding language feature and the competence of the learner (e.g. the bottom vertex has competence 0 in the rst two con gurations and competence 5 in the third one). For the sake of simplicity, we assume that the unlock value of all vertices is equal to 0.75 and the competence threshold value for unlocking is equal to 5. The unlock value for the edges is denoted within a yellow circle, while the weight of each edge is also present. An immediate consequence of the above unlocking procedure is that a language feature can become unlocked even if the learners competence on the prerequisites cannot unlock incoming edges. Hence, the sub-graph of the unlocked vertices and edges, is not necessary a connected graph. This property can prove very useful in the case where the initialization of a pro le is based on an undertaken test; the domain model contains too many language features and it is almost impossible to test all of them. Still, the available unlocked vertices will allow the learner to practice di erent areas of the domain model in parallel. 4
The ability of the iRead eBook System to provide personalized interactive activity books is based on its learner modeling capabilities. By maintaining learner pro les it can choose the proper language features to focus on and, in turn, choose appropriate activities relevant to these features to be included in the eBook. Finally, by consulting the playing/reading history of the learner, it can select from the available language resources the appropriate (for the speci c learner) linguistic material to be used in conjunction with the chosen activities.
In the architecture of the iRead eBook system, services related to learner pro ling and adaptivity as well as services responsible for maintaining domain models and related linguistic resources, are part of the content provider subsystem (see Fig. 1). Even though these services completely follow the logic employed in the iRead infrastructure, we have decided to provide a separate implementation as part of our eBook server. This allows for the incorporation of models (and their relevant resources) which are not covered by iRead (e.g., relevant to mathematics) and the production of personalized activity eBooks in the corresponding elds. However, as indicated in Fig. 1, it is possible to obtain this functionality from external resources, as the iRead infrastructure in our case, provided it is available. Actually, in our implementation, we utilize the iRead dictionary services by directly connecting to the iRead infrastructure.
Figure 3 provides a view of the architecture focusing on the components of the content provider as well as the ow of information that this relevant to pro le adaptation and content selection, the topic of this Section. During the presentation, we should always keep in mind the typical work- ow of a learner: (a) she downloads an activity eBook, (b) she read it, i.e., she plays the activities in it and, when nishes reading/playing, (c) she orders a new activity eBook (more precisely, her parents do so). Also, keep in mind that when the learner is online, the save/sync app that runs on the tablet sends the learner's progress to the eBook server. 4.1
When a new eBook is being ordered for a speci c learner, we have to make sure that the learner's pro le in accurate and re ects her reading skill state. If it is the rst time that we generate an eBook for the learner in question, we have to initialize a new pro le for her. This can be done by either adopting one of prede ned initial pro les that t the characteristics of the learner (e.g., age, school year, an informal classi cation) or by giving her a literacy test on the features of the domain model (currently, this option is not implemented in iRead). If we already have a pro le for the user, we have to make sure that the pro le is accurate. Assuming that the pro le was last updated during the creation
eBook
save eBook Server
Synthesizer production
get content get playing
history and previous
eBooks Usage Data get playing history
Content Creation get pro le get resources User Pro les
Models update pro le Adaptivity
Resources games words sentences of the last book ordered for the learner, we now have to re-evaluate it. It is the pro le re-evaluation that guarantees that the activities and content included in the eBook are appropriate for the learner, i.e., the eBook is personalized.
The learner's responses for each activity in the eBook are all recorded in her reading/playing history. So, after reviewing the performance data for a speci c activity, we should be able to decide whether the learner improved her competence on the feature targeted by the activity5. Typically, if a user has a high success rate over the few last played activities targeting the speci c language feature, the competence is increased by one unit. In iRead (where the reading skill is improved by playing the NAVIGO literacy game, and the feature competence takes integer values in the range [0 : : : 10]) we focus on the last three played activities and we request a success rate of 75% in all of them. In our activity eBook system we have adopted a lighter approach, that is, we allow feature competence to only take the values zero (0) and one (1) and we only consider the last two (2) played activities. The rational for doing so is based on our desire to (theoretically) allow the learner to master all language features in a reasonable amount of eBook pages. Assuming the user provides only correct answers, the English reading skill can be mastered by reading 558 pages6. In any case, these values can be easily adjusted. Note that, during reevaluation it makes sense to also reduce the feature competence if a learner provides too many wrong responses. This allows the learner to work again on improving her skills on language features she has not practiced recently and has forgotten them.
Updating the competence values for all features in the current reading/playing history of the learner completes the reevaluation process. However, we are not ready yet to proceed to the next stage; the selection of personalized content. Before we do that, we have to examine whether the changes in competence for some features results to unlocking more language features (which can be the 5 Each activity aims to improve the competence of the learner in one speci c language feature. It tries to achieve that by asking the learner to play an appropriate minigame. 6 The English model has 279 features. Playing 2 activities per feature requires 2 279 = 558 eBook pages in total. topic of new activities). This process is triggered by the end of the reevaluation process. For each feature with an improved competence value, we check whether the feature that depend on it (i.e., they come immediately after it in the DAG representing the domain model) satisfy the conditions for unlocking them. 4.2
Each page of the activity eBook contains one activity. An activity is de ned to be a triplet consisting of (a) a language feature, (b) a type of activity7, and (c) a mini-game, together with appropriate educational content (e.g., words, sentences). Six mini-games used in the NAVIGO games have been adopted for use in the iRead eBook system and developed by iRead partners Patakis and PickaTale. A list of them together with a brief description is given in Table 2.
Cleomatchra The learner connects pairs of strings to form full words.
PickaTale [
Perilous paths The learner selects a string (word or sentence) over three possible
options, to complete a sentence or answer a question. PickaTale [
Remove the runes The learner selects all words from a given list that have a speci c
property. Patakis [
Sliceophagus The learner splits compound words to stems. PickaTale [
xed placement. PickaTale [
Raft rapid re Words appear sequentially on the learner's screen. The learner must
\hit" all words that have a prede ned property. Patakis [
Information about each activity is stored in a database and is accessed during the selection of personalized content. Note that the process of selecting the content required in order to play an activity may be quite elaborate. For example, we may be required to identify several words relevant to a feature (e.g., words starting with \b" as in \bag") but also distractor words relevant to another feature (e.g., words starting with \d" as in \dog"). These rules are speci ed as a collection of lters which are employed during content selection.
The get activity service is responsible for determining the activities to be played. The selection of the activities is made by focusing only to those relevant to unlocked features in the domain model. Emphasis is given to unlocked features with competence below their threshold value. The reading/playing history database is also consulted. We want to avoid repetitions, we want to use as 7 Possible activity types for literacy games are: accuracy, automaticity and blending. many mini-games (relevant to the feature) as possible. The aim is to get a good mixture of activities so that we build an engaging eBook.
Having decided on which activities to include in the eBook, it remains to select content for each activity. In the case of iRead eBooks, this basically means selecting words and passing them through the lters mentioned above. Specialized dictionaries are used to complete this task. The dictionaries are specialized in the sense that each word is pre-processed in advance so that word selection is speed up and the application of the lters is made easier. An additional requirement is that the words have been carefully selected so that they cover all language features and they are appropriate to be presented to children.
The get content service is responsible for returning content for each activity. The service identi es the words in the dictionary that are relevant to the activity's feature and pass all the lters. The word's di culty (an attribute of each word stored in iRead's dictionary) is taken into account in the nal word selection (the smaller the learner competence the easiest the words we select). In addition, we also consult the reading/playing history in order to avoid words that the user has seen too many times, to replay words that the user usually get wrong and, in general, get a good mix that avoids boring repetition.
In addition to the fully automatic activity selection made by get activity we also provide a mechanism for the semi-automatic way to select the content of an activity eBook. It is semi-automatic in the sense that a teacher/parent can only specify the activity that appears in each page of the eBook8. The rational behind providing such a service is the need to allow teachers to build activity eBooks that re ect the speci c teaching plan they use in class. 5
We presented the iRead eBook system that provides an integrated personalized learning service aiming to develop the learner's reading skills by engaging her in activities delivered through interactive eBooks. The system has been implemented in the framework of the iRead H2020 EU research project. There are several threads of research that are worth investigating: { Experiment with di erent methods to select activities and content for each activity. { Expand the available resources for the reading skill acquisition. More minigames, more elaborate activities, larger dictionaries, more languages. { Develop domain models for learning skills in other elds (besides \reading").
Mathematics may be an option.
{ Consider potential privacy issues that may arise due to the use of pro ling,
for example, in the case that an eBook is addressed to a learner belonging in
a speci c group such as children with dyslexia. This question may be more
relevant for services deployed and use in Europe due to the GDPR [