Content uploaded by Marco Ronchetti
Author content
All content in this area was uploaded by Marco Ronchetti
Content may be subject to copyright.
UNIVERSITY
OF TRENTO
DEPARTMENT OF INFORMATION AND COMMUNICATION TECHNOLOGY
38050 Povo – Trento (Italy), Via Sommarive 14
http://www.dit.unitn.it
MOBILE ELDIT: CHALLENGES IN THE TRANSITION FROM AN
E-LEARNING TO AN M-LEARNING SYSTEM
Anna Trifonova, Judith Knapp, Marco Ronchetti and Johann Gamper
January 2004
Technical Report # DIT-04-009
.
Mobile ELDIT: Challenges in the Transition
from an e-Learning to an m-Learning System
Anna Trifonova*, Judith Knapp+, Marco Ronchetti*, Johann Gamper°
* University of Trento – Sommarive 14 – 38050 Povo (Trento) – Italy
Tel. +39-0461882033 Fax +39-0461882093 E-mail {Marco.Ronchetti, Anna.Trifonova}@dit.unitn.it
+ European Academy of Bolzano – Drususallee 1 – 39100 Bozen – Italy
Tel. +39-0471055092; Fax +39-0471055089; E-mail Judith.Knapp@eurac.edu
° Free University of Bolzano – Dominikanerplatz 3– 39100 Bozen – Italy
Tel. +39-0471315666; Fax +39-0471315649; E-mail Johann.Gamper@unibz.it
Abstract: This paper presents ‘Mobile ELDIT’ (m-ELDIT), a system under development which
goal is to offer access from PDAs to the learning materials of ELDIT – an adaptable language
learning platform. The ELDIT system, which is developed at the European Academy of Bolzano,
consists of a learner’s dictionary and a text corpus. Exercises, a tandem system and a tutor module
are planned. The system works online and the adaptation of the content is done dynamically
according to interactions with the learner. We analyze the requirements for adaptation and
transformation of the data, the user interfaces and also the architecture of the ‘Mobile ELDIT’ . We
discuss the necessity of specific user modeling in order to provide both online and offline access to
the learning materials from mobile devices .
Introduction
Mobile learning is a field that recently has attracted the interest of lots of researchers in the learning domain
(Trifonova & Ronchetti 2003b). Basically m-learning is considered any form of studying, teaching or learning that is
delivered through a mobile device or in a mobile environment. In general by mobile device we mean PDAs and
digital cell phones, but more generally we might think of any device that is small, autonomous and unobtrusive
enough to accompany us in every moment and can be used for educational purposes.
We analyzed different ways to apply mobile devices in education (Trifonova & Ronchetti 2003a) and
discussed that a mobile learning system should have three main functionalities – “Context Discovery”, “Mobile
Content Management and Presentation Adaptation” and “Packaging and Synchronization”. First of all the “Context
Discovery” should discover context information which is important for the learning situation, like the devices’
capabilities and limitations (software and hardware) or other information about the infrastructure; user/device
location and relevant environmental information; temporal information; preferences and etc. This data should be
used by the “Mobile Content Management and Presentation Adaptation” for adapting the content for the specific
device and user needs. Finally the “Packaging and Synchronization” should take care of selecting the content that
will be needed by the user during the offline usage of the system and upload it. It is best that the entire process is
done automatically, thus is important that the user activities are tracked and feedback to the algorithm to improve its
work. In our current work we want to develop a mobile version of an existing online language learning system,
called ELDIT, which is designed to satisfy the specific needs of the bilingual region of South Tyrol in Italy. ELDIT
is currently adaptable and will be adaptive to user behavior, needs and preferences. As mobile devices (PDAs,
smart-phones, etc.) are becoming more popular a useful possibility would be to access the system from such devices.
In order to provide such functionality we started to build the Mobile ELDIT (or m-ELDIT) system. Beside the
adaptation of the content for the specific needs of mobile devices the main difference of m-learning and e-learning is
the connectivity. While the content of ELDIT is generated on the fly upon a user request, thus requiring always ‘on’
connection, mobile devices often have periods of disconnection, either intentionally (when the connection is too
expensive) or not (when no infrastructure is provided). Facing this problem Mobile ELDIT should be able to hoard
the content needed for the offline usage.
Hoarding is a technique for selecting a set of documents to be uploaded and used when disconnected.
Related terms are caching and prefetching, though they are more often used when considering online conditions and
web performance. Caching is a technique for keeping content that has been requested by one user available on the
nearest server for a certain amount of time so other requestors can access it faster. Prefetching on the other hand is a
technique which tries to guess what will be needed to the client in the near future, cache it and this way improve the
clients’ experience. Different schemes of caching and prefetching are proposed and the goal is to reduce network
traffic, to minimize access latency, bottlenecks, servers’ workload, etc. in the WWW world. Although the goal of
hoarding content for offline use is little shifted from the one of Web caching, some of the techniques can be reused.
However, while in the online case one can balance between the accuracy of the cached set and the added traffic, in
the situation we consider a much higher accuracy is required, and, as an additional constraint, the memory is limited.
The rest of the paper is organized as follows: Section 2 gives a description of ELDIT; Section 3 gives more
detailed description of the problems we are facing for supporting the use of the system from personal digital
assistants (PDAs) and an overview of the architecture that we consider; in Section 4 we discuss the user modeling
that is needed for adapt ation and automatically selecting content for offline usage. Related work, Conclusions and
References follow.
The ELDIT System
South Tyrol is a bilingual (German and Italian) province located in the north-east of Italy. Although both
languages, German and Italian, are official languages, only few people consider themselves truly bilingual. Citizens
are entitled to use their mother tongue in dealings with the public administration including judicial authorities.
Therefore, passing the so-called exam in bilingualism is a prerequisite for employment in the public sector.
The main scope of the ELDIT project (http://www.eurac.edu/eldit) is to create an innovative electronic
language learning system for the population of South Tyrol to prepare for the exams in bilingualism. However, the
system has been designed in a very general way, such that everybody interested in learning the German or Italian
languages can profit from it.
Figure 1 shows the main modules of the ELDIT system. Based on the material for preparation of the exams
in bilingualism, we have developed an electronic learner's dictionary and a text corpus. The dictionary is especially
designed to reduce the burden of vocabulary acquisition in foreign language learning. The text corpus contains all
the texts of the exams in bilingualism. Each word is annotated with lemma and part-of-speech and is linked to the
corresponding dictionary entry, which facilitates a quick dictionary access for unknown words. Furthermore, we will
implement simple quizzes that can be generated automatically out of the existing data set, a tandem module for
collaboration, and an adaptive tutor which guides the learner through an individual vocabulary acquisition process
which alternates between vocabulary learning and applying the vocabulary on a suitable text.
Figure 1: ELDIT Architecture (main modules)
Figure 2: Screenshot of the dictionary entry
of the German word “Haus”
A dictionary entry is presented to the user in two frames (see figure 2). The left -hand frame shows the
lemma of the word and a list of different word meanings, each of which is described by a definition, an example
sentence, and an optional translation equivalent in the other language. The right -hand frame is organized in several
tabs and shows additional, semantic and syntactic information such as word combinations, related words, linguistic
difficulties, etc. The linguistic difficulties are also indicated by a kind of footnote numbers and shown in a small
window on the place where they occur.
One of the main design guidelines in the ELDIT project was to consider pedagogical and psycholinguistic
learner demands. In order to respond to these learner demands we need to store rather detailed information about the
language learning material and resources (Gamper & Knapp, 2003). Our learning material shows some important
characteristics which differ from traditional systems: the data are semi-structured and highly interlinked and have to
be annotated at a very fine-grained level of detail. In fact, we have to encode information at the level of single
words and even below. This level of detail is needed in order to support the language learner as much as possible,
and at the same time it allows reusing the learning material for several purposes.
Last but not least, very fine analyses about user preferences and user behavior can be carried out. In the
ELDIT system adaptable and adaptive features are distinguished. Adaptable features allow for the manual, a-priori
customization of the system. Adaptive features cover the aspect that the system adapts automatically to the user,
based on assumptions about the user as well as on observations about the user’s interaction with the system.
Currently, only adaptable features are implemented, but not yet enabled; the implementation of the adaptive features
is future work.
Adopting a rapid prototyping approach, we were seeking for a simple, yet expressive language to
implement our data model, which at the same time is robust to frequent changes and facilitates the knowledge
engineering process. XML shares these properties and turned out to be a good choice for the implementation
(Gamper & Knapp, 2003).
m-ELDIT: The Problem and the Proposed Architecture
Design Goals and Functionalities
The objective of the Mobile ELDIT project is the development of a mobile version of the ELDIT system
introduced in the previous section. In the study we assume that the m-ELDIT user is a self-motivated learner,
preparing for the bilingual exam, who doesn’t need any supervisory control of studying process. As mentioned
earlier ELDIT could be used by the students both as learners’ dictionary and for preparing for the bilingual
examination. Commercial dictionaries for mobile devices are available on the market. The development of such a
dictionary itself is not an interesting research topic. The only issue that is open here, considering the ELDIT
dictionary, is the compression of the data so that it can fit in the device memory because the ELDIT dictionary itself
contains big quantity of data which is hard to store on the nowadays still limited device’s memory. This is because
of the fine-granularity of the data, where the annotations are at the level of single words and even lower, in order to
support the adaptive generation of the pages, scalability and robustness to frequent changes and updates . Our
assumption is that only part of the whole ELDIT content can fit into the memory and the amount of memory may
vary on different devices. Our system aims to automatically select and upload the content that will be needed by the
user during the next period of disconnected usage of the system. The content will be adapted to the device
characteristics in advance. The decision on the hoarding set (what content should be uploaded) must be more precise
then in the general hoarding or the web prefetching case (see related work section), thus we need more efficient
prediction of the user future needs. Predicting what words an arbitrary user will search for in a dictionary is in
practice an unfeasible task. In this sense we find it suitable to provide access to mobile users only to text modules of
ELDIT and the dictionary entries related with them (words). The prediction can benefit from some preferences of
the learning domain, which are:
- The search space is much more limited than in the whole web case
- Semantic information can be available through the metadata
- Behavior of generic users can be analyzed so as to extract most likely paths to be followed
- Behavior of the particular user (preference, learning style, etc.) could contribute finding an optimal strategy.
As mentioned earlier for offering different services to mobile users, including access of learning content, the system
should support three main functionalities – “Context Discovery”, “Mobile Content Management and Presentation
Adaptation” and “Packaging and Synchronization”. The content accessed from mobile devices should be especially
designed or automatically adapted for the limited device capabilities. The presentation of learning materials is an
important issue and should be carefully designed. If the content will be accessed through a standard web-browser on
the PDA then it should not contain incompatible elements, for example scripts.
The Architecture
In our system we have decided to introduce a separate server (we call it m-ELDIT server) which will take
care of the three functionalities needed to support mobile users’ access (see figure 3). As discussed the ELDIT data
(both word entries and texts) are XML files. The presentation logic is separated from the data itself, thus different
presentations can be generated according to the needs. The special formatting could be done either on the client
device, either on an intermediate server (often called transcoding server) or inside the LMS. The mobile devices
have much more limited processing power and memory limitations and also the battery power should be considered,
thus a probably better choice is that the server does this task. The client request for a page (figure 4) gives the
context information needed to generate the pages, i.e. the HTTP version, the device display resolution, support of
colors, etc. Using the HTTP request information the ‘Content Redesign Engine’ should produce from the ELDIT
XML files the proper presentation pages.
The ‘User Behavior Analyzing Engine’ is responsible for retrieving useful information about the users’
learning styles and preferences that will be used later by the ‘Hoarding Engine’ in finding the hoarding set. Although
ELDIT has its own user models collected during the usage of the system they do not contain all the information
needed for the mobile scenario. For example by analyzing the web server log files we can discover similarities
between learning paths of different users, but also the differences. The user models will be discussed in the next
section. The ‘Hoarding Engine’ is the main module through which the mobile user interacts with mobile ELDIT.
Based on the user request’s context data it should ask the ‘Content Redesign Engine’ for properly formatted pages
that should be pushed to the device’s cache memory. The hoarding algorithms should take as input the output from
the ‘User Behavior Analyzing Engine’ (i.e. the user models with the similarities and the differences of the particular
user with the common users’ behavior and the current user preferences and learning history) and additional
information about the learning content itself. Then it should predict which path the user is most likely to follow and
assign weights to the learning objects depending on how important they are for the next user session. The objects
(the redesigned pages in our case) with higher weights should be uploaded to the device first; afterwards the
materials with smaller weights should be uploaded until the device’s available cache is filled. The module should be
able to analyze how successfully the previous hoarding was done and try to improve further prediction. Possible
methodologies for evaluation are discussed further in the paper.
Eldit Server
Web Interface
Word entriesText corpusUser model
m-Eldit Server
Mobile Client
User Interface / Web Browser
Content Redesign
Engine
Searcher Tandem
Client-side proxy
Mobile User Models
User Behavior
Analyzing Engine
Cached PagesTracking Data
Server Logs
Hoarding
Engine
Redesigned Content
Tutor etc.
Figure 3: m-ELDIT architecture
GET
http://www.science.unitn.it/~foxy/mELDIT/Texts/Gener
al.056.html
HTTP/1.1
Accept: */*
UA-OS: Windows CE (POCKET PC) - Version 3.0
UA-color: color16
UA-pixels: 240x320
UA-CPU: ARM SA1110
UA-Voice: FALSE
UA-Language: JavaScript
Accept-Encoding: gzip, deflate
User-Agent: Mozilla/2.0 (compatible; MSIE 3.02;
Windows CE; PPC; 240x320)
Host: www.science.unitn.it
Proxy-Connection: Keep-Alive
Figure 4: HTTP Request from a mobile device
(iPAQ Pocket PC)
In addition a small proxy on the client device is responsible for receiving the browser requests and
retrieving the content from the server or from the local store (‘cached pages’ in the figure above) when there is no
connection available at the moment. The client-side proxy could also seamlessly upload the content that will be used
in the future, based on the prediction done in the ‘Hoarding Engine’. Uploading might be done on a special user
request, where the user might also be given an option of setting different parameters, e.g. provisioned disconnection
time, expected duration of time in which the system will be used offline, preferred by the user topic and etc.
Different other options could be foreseen, like the proxy might be aware of the “cost” of the connection and behave
in different ways according to that, i.e. synchronizing the cache when the ‘cheap’ connection is available (internet
through LAN or cradle) and using only the cached content whenever possible on ‘expensive’ connections. It should
take care not to return errors if a requested page is not in the cache, but some meaningful message, e.g. “The page is
not available at the moment. It can be provided on the next synchronization”. Another functionality of the proxy
should be the tracking of the user activities and feedback to the mobile LMS. Thus the mobile user models should be
aware of the user needs and adapt accordingly. The mobile LMS should be responsible for calculating and updating
the user models, which will differ from the user models in a standard LMS.
User Modeling
The hoarding algorithm in our system has the role of deciding which texts the user would prefer to study
during the next offline session(s) and upload them with the associated words. Every text consists of a small set of all
the words in ELDIT. Even considering only the words that construct the text not all of them will be requested by the
user. Moreover depending on the word the user might be interested in seeing only the translation of the word and
example or might want more detailed information of it (as shown on figure 2).
User behavior observation and capturing of common access patterns will help predicting the objects (LO)
to be needed next with a certain confidence. Both the comparative analysis on all the users of the system and the
analysis of particular user behavior can be very important.
A possible modeling criterion might be the level of language knowledge of the user (i.e. beginner,
intermediate, proficiency, etc.). In this way an analysis on what words are never viewed by users with similar
language skills as the currently observed user would lead to excluding those words from the hoarding set and thus
reducing the memory used. And also vice versa – if a word (used in a text) is viewed very often by users classified
in the same category as the current user the algorithm should include it in the hoarding set with big priority.
One of the factors that might be considered in the user learning style analyses is the learner interest in the
current material. The ELDIT text corpus contains texts about different subjects both for the Italian and German
languages (e.g. Food and drinks; Economic and industry; Tourism, countries and cities; Art; Family; etc.). A useful
measure for this might be the time spent on reading a concrete text page. If the student is interested he/she will
probably spend more time on reading and reviewing unknown words. In such a way the users with common interests
can be grouped.
Individual user models should store information about the learner and his/her behavior. The individual user
models in ELDIT are a separate XML file (for every user) that contain information about user identification - login
and password; log-in and log-out times; number of words reviewed; total number of clicks during the session;
number of clicks onto the tabs (currently collocations, idiomatic expressions, derivations and compound words, and
linguistic characteristics explained in footnotes). Recording and using information about user’s native language,
language interests, proficiency, learning style and other information is planned. Some of the data currently collected
in the ELDIT user profiles might be useful for the analyses together with the web server log files, e.g. log-in and
log-out times might help us connect sequences of requested pages (recorded in the log files) with particular users.
We should point out that the learning style of the user may change, depending on the task but it might also
develop over time for the same task. Thus the individual students’ learning styles should be handled in a flexible
way. The recommendations should be taken as ‘current’ and dynamic changes should appear in the user model. The
students might have different interests while exploring the same system (ELDIT) in mobile context, and while using
a PDA for accessing the content they might develop different learning style. Thus the mobile LMS should keep
separate user models.
Methodology of Evaluation of the System
There are different aspects in which such a system should be evaluated. One is connected with the
presentation of such learning materials on PDAs and how convenient and handy users find such a system. In this
aspect the evaluation might be done by questionnaire or oral interrogation with the users. Another aspect is the
accuracy of the offline use supporting sub-system. As the decisions for the automated hoarding are taken based on
analyses of server log files we can use the techniques applied in machine learning evaluations. Typically it consist of
dividing a data set into a training set and a test set, using the former to learn the model, and the latter to evaluate the
model’s performance. This methodology has been applied to the predictive user models developed to date
(Zukerman & Albrecht, 2001).
An often used metric in the evaluation of caching proxies is the hit ratio. Hit ratio is calculated by dividing
the number of hits by the total number of uploaded predictions (cache size). In hoarding systems a more often used
measure is the miss ratio - a percentage of accesses for which the cache is ineffective. Kuenning & Popek (1997)
defined a ‘miss cost’ as a main difference in the evaluation of a caching and a hoarding system. In
caching/prefetching systems the misses in the prediction reflect as a time penalty as the missing content should be
retrieved from the web. This defers from the mobile case where with unavailable internet connection a miss in the
hoard might be fatal. In order to quantify this measure Kuenning & Popek (1997) demand a user rating on every
miss, using 4 different impact values:
0. The computer is completely unusable as a result of a miss. No future work can be done
1. The current task cannot proceed. Work can continue on a less desirable task
2. The current work will proceed, but the activity on that task will change in some way as a result of the hoard
miss
3. The hoard miss will cause little or no trouble
They also define the ‘time to first miss’ - a simple count between the start of the disconnected operation and the first
hoard miss. This evaluation criterion can be used only on real-use of a system (and its hoard part). It is strongly
connected with the hoarding size. Another measurement is the ‘miss-free hoard size’, defined as the minimum
amount of disc space that a particular hoarding system would require to allow a complete disconnection period to
take place without any misses.
Related Work
There is a lot of ongoing work in the area of content adaptation for mobile devices and device independent
representation of web content. In this context different approaches are proposed for describing device capabilities
(HTTP Request Header, CC/PP, UAPROF, etc.). Also different architectural approaches are developed for using the
information of devices’ capabilities and adapting the content accordingly. The adaptation could be server-based
(XML/XSLT, Cocoon, Axkit), proxy-based (AvantGo, Palm Web Clipping) and client-based (XHTML/CSS). A
comprehensive review of the current device independence technologies and activities can be found in Butler (2001)
or on the W3 Consortium web site (www.w3.org). Adapting the content through transcoding servers or proxies is
one of the most often used techniques. Different transcoding techniques can be applied: for simply translating from
one presentation language to another (e.g. WAP-HTML-WAP), for reducing the contents size, for satisfying
bandwidth or screen capabilities of the devices, to adapt the structure of the content in a more appropriate way and
etc. Joshi (2000) argues about the appropriateness of direct client-server architecture for mobile hosts accessing
multimedia web content and the use of proxy-based architectures. He proposes a combined system as a better
solution. As the content negotiation is one of the hot topics nowadays the author supposes that in the future more
and more servers will take care for providing an adapted content for the concrete device. Meanwhile a proxy is
needed to negotiate the content on the account of the device and/or to transcode the multimedia content if the server
doesn’t do it. The prototype system that is discussed there concentrates on the transcoding of images and video and
supposes that the server will provide alternative versions for active components (e.g. CGI alternative for JavaScript
functionalities), thus they don’t provide a solution for transcoding this type of content. Moreover the author
mentions that after this automated transcoding the resulted pages might be sometimes visually unpleasing. Although
the author mentions the problem of synchronicity, which is important when mobile hosts are used, he does not
consider an offline browsing. It is assumed that the disconnections are temporal and the user will not work offline
but the system should be able to support the continuation of the work after reconnection from the same point where
the disconnection appeared.
One of the things that drastically differ m-learning from e-learning is the offline usage. We believe that
delivering content for offline usage is an important issue as still mobile devices are often disconnected because of
the lack of access in certain places but also because of the high prices in most of the cases, thus our intention is to
support both online and offline access to data. Operation in a disconnected or intermittent connection situation is a
common problem in mobile computing. A disconnection should be distinguished from connection failure as it is
either intentional or is foreseeable, thus some preparation can be done by the interested applications. Nevertheless
only some of the transcoding proxies take care also for caching web pages for offline usage (e.g. AvantGo).
A similar problem to our one (off-line access to data) is treated in the offline browsing of web content. A
quick review of the available offline browser utilities (like www.avantgo.com, www.httrack.com,
www.webstripper.net, etc.) shows that generally during the online periods the user selects sites that should be
uploaded for later off-line usage and entire sites are dumped to the local storage or the user specifies the depth of the
links to be cached. In situation where mobile devices are considered the memory limitations should be taken into
account and thus it is preferable that only the content that will be needed during the next usage of the system is
downloaded.
The preliminary downloading of the data that will be needed in the future, called caching, perfetching or
hoarding is often considered in the Internet world. Wang (1999) has presented a survey of the state-of-the-art
techniques and elements of Web caching systems. These techniques include Prediction-by-Partial-Matching,
analyses of users’ access patterns, provided by the servers, prediction of the user’s future Web accesses by analyzing
his or her past Web accesses, etc. Although some of these techniques are useful for predicting the content needed
also in m-learning domain still they aim at different goal s – reduction of bandwidth consumption, of access latency,
server workload and etc. They explore the case of the Web where the search space is much bigger, the users are
numerous and have different interests thus the prediction accuracy is quite low comparing to what is needed in our
scenario, but could be compensated by the fact that the internet connection is permanent. The learning scenario has
characteristics that expose some additional information to be considered and thereby possibility to improve the
existing solutions. In our opinion a more efficient strategy can be defined for hoarding by taking advantage of the
peculiarity of the m-learning scenario and the semantic data that is kept in the LMS.
The idea of hoarding for disconnected devices in distributed file systems has been first described by Kistler
& Satyanarayanan (1992), although they do not consider mobile devices. They propose the Coda File System to
explore the use of caching of data not for improving performance but for increasing the availability. They propose
architecture for hoarding and for keeping the coherence of the utilized files. The initial system was based on client-
server architecture which tracks the local file modifications and saves a ‘Client Modification Log’. The project has
lately evolved into UbiData project (Helal et al. 2002, Zhang et al. 2003) and the direction it took is in double-
middleware architecture for ubiquitous data (file) access. They introduce incremental hoarding, where the idea is to
use a version control system to maintain object differences and they also study the automatic data selection problem.
A metadata server is included to store the ‘users’ mobile profile’ which keeps a list of user files that are considered
‘interesting’. They define a “hybrid priority” metric for choosing the hoarding set of files. The “hybrid priority” is
calculated by taking into account the recency of use, the frequency of access and the active periods of the file use
and the algorithm also considers upper space limit of memory. The reported effectiveness of their filtering algorithm
is more than 84% (Zhang et al. 2003).
Facing the hoarding problem for mobile computing in disconnected mode an interesting solution has been
proposed in SEER (Kuenning & Popek 1997; Kuenning et al. 1997). The authors were also inspired by the work on
the Coda system but go in different direction. They defined a new measure, “semantic distance”, between individual
files by observing the user activities and propose an algorithm for automatic hoarding of projects for mobile
computers. With “semantic distance” the author tries to quantify the user's intuition about the relationship between
files in the same project. For this different measuring criteria are used – “temporal semantic distance”, “sequence-
based semantic distance”, “lifetime semantic distance”, directory membership, filename conventions and hot links,
and are combined to assign weights to documents and take decisions for hoarding them in an automatic way
(automatic periodic hoarding). The approach met some unpredictable behavior in the real-world system, which
appeared because of the way the operating systems and some often used programs work (like the “find” operation
under Unix). Recent experimentations with the same system (Kuenning et al. 2002) showed surprising findings – the
complex clustering methods that are used in the system work in most of the cases worse then a LRU (least recently
used) algorithm enhanced with some heuristics. This shows us that the research field is still open for work.
Conclusions
In the paper we had presented a mobile language learning system, called Mobile ELDIT (m-ELDIT), which
aims to support people on the move, interested in taking the bilingualism exams in South Tyrol. In a previous paper
(Trifonova & Ronchetti 2003b) we have discussed that mobile learning in general applies best to activities with
certain characteristics: (1) short, for filling the gaps of waiting time; (2) simple and with added value; (3) delivered
just in time/place. Language learning fits very well in this frame. We are building our mobile learning system on top
of an existing adaptive language learning platform, thus the users will have the possibility to work both from their
desktop PC and while away from their homes and offices on their PDAs. We have pointed out that m-ELDIT should
take care for redesigning the requested materials for the specific characteristics of the device used and also predict
and upload the content that will be used during the offline periods. We have proposed an architecture for the system,
which consists of a small module on the client device (client -side proxy) and three modules on the mobile server
side – “Content Redesign Engine”, “User Behavior Analyzing Engine” and “Hoarding Engine”. The hoarding
algorithm should take advantage of the user models from the ELDIT system, but also separate analyses are needed
for the access patterns to the system from mobile devices.
References
Butler M. (2001). Current Technologies for Device Independence: HP Labs Technical Reports HPL-2001-83
Gamper J., Knapp J. (2003). A Data Model and its Implementation for a Web-Based Language Learning System. In
Proceedings of the Twelfth International World Wide Web Conference (WWW2003)
Helal A., Khushraj A., Zhang J. (2002). Incremental Hoarding and Reintegration in Mobile Environments. In
Proceedings of Symposium on Applications and the Internet (SAINT)
Joshi A. (2000). On proxy agents, mobility, and web access. Mobile Networks and Applications, 5 (2000) p.233–241
Kistler J., Satyanarayanan M. (1992). Disconnected Operation in the Coda File System, ACM Transactions on
Computer Systems, Vol. 10, No. 1
Kuenning G., Popek G. (1997). Automated Hoarding for Mobile Computers. In Proceedings of the 16th ACM
Symposium on Operating Systems Principles (SOSP -16), San-Malo, France, pp.264-275
Kuenning G., Reiher P., Ma W., Popek G. (2002) Simplifying Automated Hoarding Methods. In Proceedings of 5th
ACM International Workshop on Modeling, Analysis and Simulation of Wireless and Mobile Systems
(WiM’02), Atlanta, Georgia, USA, pp.15-21
Kuenning G., Reiher P., Popek G. (1997). Experience with an Automated Hoarding System. Personal Technologies,
1, 3, September 1997, pp.145-155
Trifonova A., Ronchetti M. (2003a). A General Architecture for M-Learning. In Proceedings of the Second
International Conference on Multimedia and ICTs in Education (m-ICTE 2003), Badajoz, Spain
Trifonova A., Ronchetti M. (2003b). Where is Mobile Learning Going?. In Proceedings of E-Learn 2003
Conference, Phoenix, Arizona, USA
Wang J. (1999). A Survey of Web Caching Schemes for the Internet. ACM Computer Communication Review, 25(9)
Zhang J, Helal A., Hammer J. (2003). UbiData: Ubiquitous Mobile File Service. In Proceedings of the ACM
Symposium on Applied Computing (SAC), Melbourne, Florida
Zukerman, I., Albrecht D. W. (2001). Predictive Statistical Models for User Modeling. User Modeling and User-
Adapted Interaction, 11, pp. 5-18
