A Framework for Content Preparation to Support Open-Corpus Adaptive Hypermedia

A Framework for Content Preparation to Support
Open-Corpus Adaptive Hypermedia

Killian Levacher, Éamonn Hynes, Séamus Lawless, Alexander O'Connor,
Vincent Wade

Centre for Next Generation Localisation, Knowledge & Data Engineering Group,
School of Computer Science and Statistics, Trinity College, Dublin, Ireland

{Killian.Levacher, hynese, Seamus.Lawless, Alex.OConnor,
Vincent.Wade}@cs.tcd.ie

Abstract. A key impediment for enabling the mainstream adoption of
Adaptive Hypermedia for web applications and corporate websites is the
diculty in repurposing existing content for such delivery systems. This
paper proposes a novel framework for open-corpus content preparation,
making it usable for adaptive hypermedia systems. The proposed frame-
work processes documents drawn from both open (i.e. web) and closed
corpora, producing coherent conceptual sections of text with associated
descriptive metadata. The solution bridges the gap between information
resources and information requirements of adaptive systems by adopting
state-of-the-art information extraction and structural content analysis
techniques. The result is an on-demand provision of tailored, atomic in-
formation objects called slices. The challenges associated with open
corpus content reusability are addressed with the aim of improving the
scalability and interoperability of adaptive systems. This paper proposes
an initial architecture for such a framework in addition to reviews of
associated technologies.

Key words: Open Corpus Content, Adaptive Hypermedia, Statistical
Content Analysis, Structural Content Analysis

1 Introduction
The increasing pervasiveness of the internet is fundamentally changing how peo-
ple author, interact with and consume content. There are early signs of a shift
in the way digital content is created, from the linear authoring and publication
of material towards the aggregation and re-use of existing content from various
disparate sources.
Adaptive Hypermedia Systems (AHS) have traditionally attempted to de-
liver dynamically adapted and personalised presentations to users through the
sequencing of recongurable pieces of information.
While the eectiveness and benets of such systems have been proven in
numerous studies [1][2], a major obstacle to their widespread adoption relates to

the restriction of suitable content available to provide such adaptivity in terms
of volume, granularity, style and meta-data.

One of the key impediments in oering a wide range of information objects is,
for example, the considerable amount of manual eort involved in creating these
resources. Information object content consists of either pre-existing documents
[3], or of content created by small groups of individuals [4], intentionally authored
for a particular system [5]. This generates both scalability and interoperability
issues, which prevents the mainstream adoption of adaptive systems.

Metadata standards such as LOM [6] attempt to provide generic content
packaging methods with the aim of enabling the interoperability of information
objects within AHSs. Although they implement usage-agnostic principles, their
development is very time consuming due to the complexity involved [7]. Fur-
thermore, they also require AHSs to comply with a specic content structure in
order to make full usage of these resources.

In parallel with these developments, a wealth of information is now accessible
on the web, in digital repositories and as part of library initiatives. However, due
to its heterogeneity, this information is not suited to direct usage by AHSs in its
present form. It is available in several languages and within large documents with
very limited meta-data. These objects are also generally very coarse-grained and
correlated with unnecessary noise such as navigation bars, advertisements etc.
Automated content preparation techniques are therefore necessary to provide
scalable solutions to AHS specic content requirements.

Our goal is to provide AHSs with the ability to benet from the wealth
of knowledge already accessible on the web and in digital libraries by bridging
the gap between these information resources and the information requirements
of adaptive systems. We propose the creation of a service that can tailor open-
corpus content for use in AHSs without requiring such systems to adopt a generic
content structure. This system would hence resolve scalability issues in content
authoring by providing AHSs with a large array of information objects. More-
over, the inclusion of open corpus information would provide up-to-date content
in a wide variety of styles and structures.

We aim to combine several disciplines such as Information Retrieval, Adap-
tive Web and Information Extraction in order to provide an on-demand informa-
tion object service. This service will harvest open corpus information, segment
it structurally and automatically augment it with successive layers of internal
metadata. Any AHS could then provide this service with a request that includes
requirements for an information object which would in turn fetch and tailor this
content to its specication.

This paper is structured as follows: Section 2 will present the key challenges
addressed within this research in relation to AHS information requirements as
well as structural and statistical text analysis. The framework we envisage de-
veloping is presented in detail in section 3, along with the overall workow pro-
duced. Section 4 will present the application area of such a technology and how
we intend to evaluate it. Finally, section 5 will conclude and present the road

map ahead. This work is applicable to any type of content, however some of the
examples presented in this paper will be based on adaptive e-learning systems.

2 Key Challenges and Related Work

Throughout the evolution of adaptive systems, technologies introduced have
moved from original all in one solutions, towards increasing modularity whereby
AHSs become consumers of domain, pedagogical and user models [8]. This has
the eect of leaving such systems with a core adaptive engine capable of deal-
ing with a wider range of loosely coupled models that are integrated as desired.
Content, however, is still very tightly coupled to these engines and as a result
strongly impedes the general adoption of these systems.
Most AH systems, up until recently have operated in closed document spaces
with content specically authored for their usage [4], hence obstructing interop-
erability, both by accepting only a narrow eld of content, and by motivating the
generation of content in a highly-specic format. As a result adaptive systems
also encounter scalability issues due to the limited amount of content available
to adapt on, arising from the small amount of manual contributions available.
Open corpus content is increasingly seen as providing a solution to these is-
sues [9]. However, most systems incorporating this type of content have, for the
moment, mainly focused on linking it with internal content as alternative ex-
ploration paths. Those incorporating it fully into their system [10][11] require
manual mark-up of such content with specic meta-data. Schemas such as LOM
(in the area of e-learning) and IMS packagings, attempt to provide usage-agnostic
solutions, however they require a lot of development eort thus prohibiting scal-
ability [7]. Moreover, AHSs must comply with specic content structures so as to
avail of these resources. In order to leverage the full potential of open corpus re-
sources, adaptive systems need to incorporate and correlate this type of content
fully into existing systems using structure-agnostic and automated approaches.
Our approach, on the other hand, moves away from the packaging model
altogether by considering content as a pluggable feature of adaptive systems to
the same extent as user or domain models. A service providing open corpus con-
tent automatically in a form and format customised for each requesting adaptive
system decouple this feature from current systems, disentangling it from content
provision and hence providing a solution to both scalability and interoperability
issues.
Fully integrating information harvested over the web within adaptive presen-
tations, on the other hand, generates new issues. Web pages are usually written
as stand-alone content without any re-composition purposes. In addition to the
main content, they usually present navigation bars, advertisement banners and
irrelevant features from various sources that must be removed prior to proper
re-use.
Structural segmentation will therefore be an initial fundamental requirement
in tailoring open corpus content for use within adaptive systems. A large propor-

tion of pages harvested will need to be stripped of any redundant information
that might prevent its adequate re-use in adaptive systems.
The problem of structural segmentation has already been addressed from
dierent perspectives mainly with the purpose of tailoring existing content for
dierent displays [12]. Most of the techniques up until recently have focused
on analyzing the DOM structure of a HTML page as a basis for segmentation.
Several attempts, for instance, focus on removing templates from pages by iden-
tifying common DOM sub-trees [13] or using isotonic regression [14]. Machine
Learning algorithms are also used to decide which pair of tags should coincide
with suitable segmentation points. However, it has become increasingly popu-
lar to separate any formatting style from within HTML tags using Cascading
Style Sheets (CSS) and Javascript, thus increasing the heterogeneity of rendered
content from similar HTML trees [15]. For this reason, such techniques do not
appear to be an adequate solution when dealing with a large set of documents
as diverse as the World Wide Web.
Vision-based techniques, using entropy reduction [12], or techniques such as
VIP algorithms [16] on the other hand, partition a page based on its rendering.
These have the benet of covering a wider range of pages regardless of their
HTML trees. But their usage within systems dealing with large volume of data is
questionable since rendering must be performed prior to any analysis, inevitably
producing delays. Moreover, they completely ignore DOM features altogether
which do provide structural clues where rendered surfaces appear similar [12].
This part of our research will therefore seek to provide a solution to this
paradigm that scales both in terms of processing speed as well as structural
segmentation precision within a wide heterogeneous range of pages. The proce-
dure will additionally need to incorporate the notion of granularity in order to
keep cohesive chunks of text together to provide meaningful input to subsequent
semantic analysis.
Within the statistical analysis component of the framework, it is important
to extract concepts with high precision/recall and eciently represent these con-
cepts in such a way as to make them as interoperable and re-usable as possible.
Brusilovosky et al. [9] discuss the lack of re-usability and interoperability between
adaptive systems and how current adaptive hypermedia systems are restricted
to closed corpora by the nature of their design. This framework architecture
is based on the idea that open corpus adaptive hypermedia is feasible: the ef-
fort will be in moving away from tables of document-term frequencies towards
a richer semantic representation of data that is uncomplicated, easy-to-use and
highly interoperable.

3 Framework for Content Preparation
The framework proposed in this research is divided into three separate com-
ponents as shown in gure 1. Each part of the framework executes a specic
task on the open corpus content. A specic layer of meta-data is subsequently
appended, enriching it with structural and semantic clues for further analysis at

each step. The rst two stages of the pipeline are executed pre-emptively prior
to any client request while the intelligent slicing task is executed at run time.
The concept of a slice is an abstract notion representing a stand-alone piece
of information, originally part of an existing document, extracted and segmented
to full a specic information request. A slice can be atomic or composed of
other slices. It possesses its own set of properties (or meta-data) and can inherit
those of others. A slice is virtual in the sense that it only exists temporarily to
full a client request. It is specic to each information request and represents a
subjective perspective on a particular piece of a document and its description.
The degree of complexity of a slice will match the requirements of the system
which requests it.
The rest of this section will present in detail each stage of the framework.
The structural analysis of open corpus content will initially be described followed
by the statistical analysis of these resources. Finally, the overall work-ow be-
tween this framework and AHSs clients will be outlined using intelligent slicing
technology.

Fig. 1. Content analysis methodology

3.1 Structural Analysis

In order to provide the initial structural segmentation needed to create slices,
this research will attempt to combine benets of both DOM- and vision-based
segmentation techniques. The process will initially divide pages in atomic blocks
using relevant HTML tags similar to [15]. Following this extensive break down,
the page will be reconstructed, aggregating relevant blocks together with similar
inherent textual properties. Changes in text ow (short phrases in navigation
bars vs. long sentences in text), for instance, will be considered as segment
delineations.

The notion of text density, originating from the eld of computer vision will
also be applied to these pages using the ratio between the number of words and
the space occupied on the page [15] as an additional hint for section divisions.
Furthermore, the use of linguistic analysis techniques such as style boundary
detection will be considered as additional boundary indication.

3.2 Statistical Analysis

In this section we discuss methods for statistically analysing text and extracting
information such that it can be used as part of the wider framework; it serv-
ing as a substrate for the incorporation of several dierent statistical analysis
techniques. In order for the proposed framework to be able to interact naturally
with information seekers, it must cope with inherent ambivalence, impreciseness
or even vagueness of requests and deal with them in a structured and systematic
fashion. For these reasons, statistical methods oer a wealth of neat techniques
for dealing with the underlying complex characteristics of language.
Work currently underway includes looking at methods for the automatic
extraction of concepts from text. The vision is to extract and create a conceptual
layer over the text which is amenable to ecient searching and logical reasoning.
We move away from working only on closed-corpus texts and go some way to
addressing the huge challenge that is the open domain problem [9] by making
our algorithms as generalised as possible and not restricted to any particular
domain. One method for identifying concepts from text is to use supervised
learning methods. We use HMMs (Hidden Markov Models) to extract coherent,
relevant passages from text [17]. This method is particularly suited to extraction
from expositional texts as opposed to highly expressive literary texts (i.e. poetry,
literary prose, etc.) which have high entropy vocabularies and diverse language
use. Jiang et al. report very high precision/recall compared to other methods and
their work ts in nicely with the idea of extracting concepts/coherent relevant
passages of text.
The basic idea is as follows: given some a priori concept that is to be extracted
from a text, several passages identied as being associated with that particular
concept are used as training data for the HMM classier. Once a reasonable
amount of training data has been collected (enough data such that the classier's
train and test error converges towards the desired performance), the HMM can
then make soft decisions on a text stream and identify coherent relevant passages
of varying length with very high accuracy.
The topology of the HMM is shown in gure 2 below: the states B1, B2 and
B3 are the background states, state E is an end state and state R is a relevant
state. The 5 state HMM structure allows the system to emit a relevant symbol,
move to a background state and then re-emit a relevant symbol at some later
point before reaching the E state. The output observation probabilities can be
extracted directly from the training corpora, and the objective is to learn the
transition probabilities so as to maximise the likelihood of observing a relevant
passage given a concept.

Fig. 2. HMM topology

On a synthesised corpus consisting of a number of article abstracts (coherent
relevant passages) stitched together, with each abstract having an associated
label, very high precision and recall values were reported [17]. It remains to be
seen how well the HMM method performs on non-synthetic corpora and work is
currently underway to explore this.
Several extensions to the work of Jiang et al. have also been explored, includ-
ing the incorporation of anaphora to resolve pronouns and augment the presence
of important entities such as people and places. In addition to the uni-gram lexi-
cal features, more elaborate query expansion methods and semantic analysis has
been looked at to extract more elaborate features for the classier to work from.
The incorporation of this additional functionality is expected to yield higher
performance in terms of precision and recall on the synthetic document set and
more reliable performance on real corpora. Instead of just a single-dimension
vector with just one feature (i.e. a stemmed word), the word is represented
as a slightly higher dimensional vector (i.e. the stemmed word with additional
features) that helps improve the overall accuracy of the classier.

Word Sense Clustering Another method used to extract conceptual enti-
ties from text is word sense clustering. Lexical resources such as WordNet [18]
provide ne-grained sense descriptions for a huge number of words in the En-
glish language. However, such ne-grained detail is not always required: it is
possible to cluster together word senses and come up with a looser word sense
that abstracts away a lot of the detail and is more compatible with the frame-
work as described in this paper. Snow et al. [19] describe a supervised learning
method for merging word senses using a SVM-based (Support Vector Machine)
clustering algorithm. This algorithm yields senses at a slightly higher level of
granularity and is more compatible with corporation into a conceptual layer
representation. Use of latent semantic indexing techniques is another powerful
and elegant method for dealing with synonymy and polysemy whilst at the same
time, automatically generating concepts, or clusters of word senses [20]. In la-
tent semantic indexing, each document is represented as a vector of terms, with

each element of the vector having an associated weight; some function of the
term frequency. Dimensionality reduction techniques are then applied to each of
these document vectors, resulting in a set of vectors that point in the direction
of semantically similar words. Each term in a document can then be projected
onto the reduced-dimensionality semantic space, where the dominant vector(s)
correspond to the concept that the term is most associated with. Probabilistic
latent semantic indexing techniques [21] provide an even more accurate means
of mapping from terms to concepts and such techniques will be used extensively
for the statistical analysis component of the framework.
Representing content in the form of a semantic map is thought to mimic the
way the human mind classies linguistic categories. While there is no direct link
between semantic maps and the representation of concepts in the human mind,
it is an objective starting point for investigation of the eciency of the semantic
map representation on human cognition.

3.3 Content Preparation Work-Flow

The framework architecture presented in gure 3 belongs to the core functional
unit of a web service that delivers on-demand slices to adaptive systems.
As a content provider, the framework will rst of all require domain specic
data gathered previously by pre-emptive crawls in areas of interest to adaptive
engine consumers. We envisage as an initial step to use the OCCS harvesting
system [22], developed previously, as a content gathering tool in order to boot-
strap this framework with a renewable wide range of content. The system will
be content harvester agnostic, hence any other content gathering system could
be plugged in if required, such as private repositories for instance.
Once suitable content is available, the framework described previously will
rstly remove any redundant content, then hand over only relevant information
to our semantic analysis unit, and then create a conceptual map of the data.
These two steps will produce a large amount of internal meta-data, both from a
semantic and structural point of view.
At this stage, the service will be ready to receive content requests from a
diverse range of consumers through conceptual queries which the service will at-
tempt to map as closely as possible to the internal meta-data gathered previously
on its content. The intelligent slicing unit will be responsible for matching re-
quests to available data within the system and slicing information objects to the
nest semantic grain possible matching the query. Semantic maps constructed a
priori will be used to produce a personalised slice for the adaptive systems. Se-
mantic technologies (OWL and SPARQL) will form the underlying mechanism
navigating through pre-emptive semantic mappings.

4 Evaluation and Roadmap
This section provides a brief summary of what are the main benets of this
framework in comparison to traditional systems. The limitations of this archi-
tecture are also discussed along with possible solutions. This will result in a

Fig. 3. Content analysis architecture

roadmap which focuses on an initial empirical evaluation of critical components,
its implementation and the overall evaluation of the framework.

The content preparation framework discussed in this paper aims at bridging
the gap between an increasingly large amount of open-corpus resources available
on the web and AHS information requirements. The automated content slicing
approach proposed oers a solution to the authorship scalability issues currently
impeding the mainstream adoption of adaptive web technologies.

This method removes any content packaging interoperability issues by local-
ising slices to the requirements of the AHS, instead of the AHS complying to a
specic format. Moreover, the pipelined architecture of this framework enables
the incorporation of additional plug-able meta-data generation services to the
system, thus removing the need for unscalable hand-made meta-data.

However, this novel content preparation solution raises new challenges. At
a component level, several technological challenges exist. To the best of our
knowledge, no existing structural segmentation approach currently fullls all of
our requirements. For this reason, an initial evaluation of existing algorithms will
be conducted, resulting in the combination of several techniques. Additionally,
the semantic analysing step of this framework seen as the critical stage of this
pipeline. An empirical evaluation of this step will therefore be needed so as to
measure the quality and size of the meta-data created by this component. This
process is critical since it will subsequently determine methods available to match
slices with content requests. As a result, the content precision of slices generated
will depend on the successive layers of the automated meta-data available. It is
possible, in addition, that this precision could be improved by the use of an a
posteriori manual crowd sourcing renement of the slices.

As part of the Centre for Next Generation Localisation (CNGL), an initial
implementation of this framework within a personalised multi-lingual customer
care (PMCC) system is planned. A number of framework level evaluations will be
undertaken in partnership with our industrial partners. An empirical evaluation
of automatic meta-data generation performance will be conducted in order to
estimate the degree of reliance towards automatic slice generation. These initial

steps will be necessary to ultimately provide a reliable foundation for the PMCC
application envisaged.

5 Discussion
This paper described how restrictions in the provision of suitable content for
Adaptive Hypermedia Systems currently impedes their full mainstream adoption
due to interoperability and scalability authoring issues. The decoupling of con-
tent infrastructures from Adaptive Systems along with the use of automatically
tailored open corpus resources was presented as a solution to this predicament.
Consequently, the architecture for a content provision service that lls the
gap between open corpus information resources and Adaptive System informa-
tion requirements, was outlined. To the best of our knowledge, this framework
represents the rst attempt to produce a pipelined, client agnostic, information
preparation service that produces virtual content slices for adaptive systems.
Other innovations in the framework architecture include the modular and plug-
gable open corpus analysers, which produce successive layers of internal meta-
data information. The combination of these components and how they integrate
coherently to produce the framework is also an aspect under investigation.
An initial prototype of this framework is currently under development and
aims towards integrating fully with a customer care information system that is
currently being deployed as part of the Centre for Next Generation Localisation
project.

Acknowledgements This research is supported by the Science Foundation
Ireland (grant 07/CE/I1142) as part of the Centre for Next Generation Locali-
sation (www.cngl.ie) at Trinity College, Dublin. The authors wish to thank the
anonymous reviewers for their comments.

References
1. Brusilovsky, P.: Adaptive Navigation Support in Educational Hypermedia: An
Evaluation of the ISIS-Tutor. Journal of Computing and Information Technology
(1998)
2. Hook, K.: Evaluating the utility and usability of an adaptive hypermedia system.
International Conference on Intelligent User Interfaces (1997)
3. Henze, N., Nejdl, W.: Adaptivity in the KBS Hyperbook System. 2nd Workshop
on Adaptive Systems and User Modeling on the WWW (1999)
4. Dieberger, A., Guzdial, M.: Coweb - Experiences with Collaborative Web Spaces.
Interacting with Social Information Spaces (2002)
5. De Bra, P.: Teaching Hypertext and Hypermedia through the Web. Journal of
Universal Computer Science 2 (1996)
6. IMS Global Learning Consortium: Learning Object Metadata (LOM) Information
Model 1.2.2 (2009)

7. Farrell, R.G., Liburd, S.D., Thomas, J.C.: Dynamic Assembly of Learning Objects.
In: Proceedings of the 13th International World Wide Web Conference on Alternate
Track Papers Posters. (2004) 162169
8. Conlan, O., Wade, V., Bruen, C., Gargan, M.: Multi-model, Metadata Driven Ap-
proach to Adaptive Hypermedia Services for Personalized eLearning. In: Adaptive
Hypermedia and Adaptive Web-Based Systems. Springer-Verlag (2002) 100111
9. Brusilovsky, P., Henze, N.: The Adaptive Web. In: Open Corpus Adaptive Edu-
cational Hypermedia. Springer (2007) 671696
10. Henze, N., Lenski, W., Wette-Roch, E.: Adaptation in Open Corpus Hypermedia.
Journal of Articial Intelligence in Education 12 (2001) 325350
11. Carmona, C., Bueno, D., Guzman, E., Conejo, R.: SIGUE: Making Web Courses
Adaptive. In: Adaptive Hypermedia and Adaptive Web-Based Systems. (2002)
12. Baluja, S.: Browsing on Small Screens: Recasting Web-Page Segmentation into
an Ecient Machine Learning Framework. In: Proceedings of WWW-2006, ACM
(2006) 3342
13. Vieira, K., da Silva, A.S., Pinto, N., de Moura, E.S., Cavalcanti, J.M., Freire,
J.: A Fast and Robust Method for Web Page Template Detection and Removal.
In: Proceedings of the 15th ACM International Conference on Information and
Knowledge Management, ACM (2006) 258267
14. Chakrabarti, D., Kumar, R., Punera, K.: Page-Level Template Detection via Iso-
tonic Smoothing. In: Proceedings of the 16th International Conference on World
Wide Web, ACM (2007) 6170
15. Kohlsch, C., Nejdl, W.: A Densitometric Approach to Web Page Segmentation.
In: Proceedings of the 17th ACM Conference on Information and Knowledge Man-
agement, ACM (2008) 11731182
16. Cai, D., Yu, S., Wen, J.R., Ma, W.Y.: Extracting Content Structure for Web Pages
Based on Visual Representation. Springer (2003)
17. Jiang, J., Zhai, C.: Extraction of Coherent Relevant Passages using Hidden Markov
Models. ACM Transactions on Information Systems (TOIS) 24 (2006) 295319
18. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press (1998)
19. Snow, R., Prakash, S., Jurafsky, D., Ng, A.Y.: Learning to Merge Word Senses.
In: EMNLP. (2007)
20. Deerwester, S., Dumais, S., Furnas, G., Landauer, T., Harshaman, R.: Indexing
by Latent Semantic Analysis. Journal of the American Society for Information
Science 41 (1990) 391407
21. Hofmann, T.: Probabilistic Latent Semantic Indexing. In: 22nd Annual Inter-
national ACM SIGIR Conference on Research and Development in Information
Retrieval. (1999)
22. Lawless, S., Hederman, L., Wade, V.: Occs: Enabling the Dynamic Discovery,
Harvesting and Delivery of Educational Content from Open Corpus Sources. In:
Eighth IEEE International Conference on Advanced Learning Technologies, IEEE
Computer Society (2008) 676678

A Framework for Content Preparation to Support Open-Corpus Adaptive Hypermedia

More Related Content

What's hot (9)

Viewers also liked (8)

Similar to A Framework for Content Preparation to Support Open-Corpus Adaptive Hypermedia (20)

Recently uploaded (20)

A Framework for Content Preparation to Support Open-Corpus Adaptive Hypermedia