PERSONALIZED LEARNING PATHS BASED ON

WIKIPEDIA ARTICLE STATISTICS

Lauri Lahti

Helsinki University of Technology (TKK), Department of Computer Science and Engineering, Finland

Keywords: Semantic navigation, Intelligent tutoring system, Concept map, Wikipedia, Ranking.

Abstract: We propose a new semi-automated method for generating personalized learning paths from the Wikipedia

online encyclopedia by following inter-article hyperlink chains based on various rankings that are retrieved

from the statistics of the articles. Alternative perspectives for learning topics are achieved when the next

hyperlink to access is selected based on hierarchy of hyperlinks, repetition of hyperlink terms, article size,

viewing rate, editing rate, or user-defined weighted mixture of them all. We have implemented the method

in a prototype enabling the learner to build independently concept maps following her needs and

consideration. A list of related concepts is shown in a desired type of ranking to label new nodes (titles of

target articles for current hyperlinks) accompanied with parsed explanation phrases from the sentences

surrounding each hyperlink to label directed arcs connecting nodes. In experiments the alternative ranking

schemes well supported various learning needs suggesting new pedagogical networking practices.

1 INTRODUCTION

To enhance the quality and efficiency of automated

information processing there is a need for new

computational methods. The methods used for

representing and modifying information with the

computers should be made more compatible with the

methods that are naturally used by humans to adopt

and associate meanings. Thus developing illustrative

adaptive computational methods that can be used in

knowledge management in natural language and

with intuitive visualizations should have a high

priority in research agenda.

We suggest that information available in

structured online knowledge bases can serve as a

valuable resource for computer-assisted learning in

respect to the key concepts of curriculum and

semantic relations linking them. We propose a new

adaptive method to assist individual learning of core

network of semantic relations in curriculum and later

expanding this network further. We consider that

still today the most valuable environment for

learning is social collaboration in everyday life. We

do not aim to challenge this traditional method but to

complement it in a fruitful manner. We propose a

new semi-automated method for generating

personalized learning paths from the Wikipedia

online encyclopedia by following inter-article

hyperlink chains based on various rankings that are

retrieved from the statistics of the articles. A

learning path describes a structure of actions a

learner has to perform in order to attain a

competence or a competence profile (Janssen et al.,

2008). In our proposal the learning paths are

represented with concept maps. We are interested in

methodology related to semantic navigation,

intelligent tutoring systems and content-based

filtering.

Our method aims to ensure that an optimal

coverage of concepts becomes provided to the

learner. Especially our method aims to enable active

explorations of conceptual relations taking into

account various diverse perspectives that are based

on different individual backgrounds and preference.

Besides in education, the method can be used in

various professions to support creative problem

solving to analyse various perspectives and to

acquire distant associations for inspiration. The

method can be further applied to model cultural

dependencies in conceptual structures and how they

can be flexibly exploited in cultural exchange for

better mutual understanding.

2 BACKGROUND

Among many competing learning theories

constructivism has remained widely supported. In

110

Lahti L. (2010).

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS.

In Proceedings of the 2nd International Conference on Computer Supported Education, pages 110-120

DOI: 10.5220/0002800901100120

 SciTePress

brief, it states that humans generate knowledge and

meaning from their experiences. Holmes et al.

(2001) suggested an expanded definition of social

constructivism that could fully address the synergy

between advances in information technology and

virtual environments. One general challenge comes

from the long-lasting debate if semantic structures of

natural language are independent of syntactic

structures or not (Peregrin, in press). Transferable

learning that enables applying previously acquired

training successfully for novel future events can be

achieved through the learner being exposed to the

learning material in a variety of contexts (Schmidt &

Bjork, 1992). Designing learning activities can

exploit the notion that people typically predict

upcoming words in fluent discourse (Van Berkum et

al., 2005). There is evidence that concept-oriented

reading instruction increases reading comprehension

and engagement (Guthrie et al., 2004).

Collaboratively maintained web sites, wikies,

have been actively adopted as new educational

environments with an assumption to support

constructive learning process. However, typical use

of wikies may enhance merely student engagement,

but not performance on assessment (Neumann &

Hood, 2009). A leading wiki site, the Wikipedia

online encyclopedia, provides an extensive coverage

of factual knowledge from various domains of life

and is actively used as a resource by students and

educators. Despite the concerns of accuracy, missing

reference and vandalism, the content has been

shown to be relatively reliable and up-to-date

(Chesney, 2006). The content can be added and

edited collaboratively by anyone but some parts are

more protected to prevent vandalism and consistent

rewriting. General usage patterns for various

Wikipedia editions have been analyzed (Reinoso et

al., 2009) showing a ratio of 620 reading operations

per one saving operation for articles in the English

edition.

It has been estimated that in English lexicon

there are well over 54 000 word families and an

educated adult native speaker knows around 20 000

of them (Nation & Waring, 1997). However, the

most frequent 3000 to 5000 word families typically

cover around 90 % of ordinary text and even more

of spoken language of a language. Mastering just

this fraction can already provide a strong basis for

comprehension thus allowing efficient further

learning from the context. On the other hand, the

Wikipedia online encyclopedia currently contains

over 3 million articles in English thus greatly

exceeding the average vocabulary of an educated

adult. This motivates us to propose exploiting the

Wikipedia as a valuable resource for developing

methods that assist people in adoption of conceptual

relations and learning in general.

There have been attempts to develop methods to

generate personalized representations from large

knowledge resources to serve for example in

education. After introducing our method and initial

experiment, we will discuss about some related work

in Section 5.

3 METHOD

3.1 Ranking Hyperlinks based on

Article Statistics

Depending on the Wikipedia article, the amount and

type of hyperlinks that it holds varies a lot. The more

hyperlinks exist, the more alternative learning paths

can be provided to the learner although making it

also harder to choose one of them through

comparison. A fundamental computational challenge

is to identify the most promising hyperlinks and

indicate them to the learner. For natural language

processing applications, various confidence

measures have been developed to estimate the

probability of correctness of the outputs (Gandrabur

et al., 2006).

We propose that the hyperlinks can be shown in

a list that supports ranking based on various criteria

taking into account different perspectives provided

by each hyperlink’s target article and depending on

varied preferences among learners. Obviously, using

many parallel measures for ranking hyperlinks can

enhance possibility to systematically differentiate

alternative rankings but unfortunately also increases

computational complexity. It could be also possible

to perform deep searches in the network and based

on them conclude the most promising direction to

traverse the next hyperlink. However, we wanted to

minimize the cost of searches in the network and

decided to evaluate only those articles that can be

reached within a distance of one hyperlink step.

Since we aim at developing methods that can be

used even with modest technological resources, we

want to consider now simple measures only. As

computational power constantly grows we expect

taking increasingly complex measures into use in the

future.

We propose that many statistical features about

the hyperlink’s target article can be retrieved as

useful indicators about the perspectives that the

target article represents in relation to the current

article. Thus, one can use alternative ranking

principles to sort target articles of hyperlinks in the

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS

111

current article. This enables getting target articles of

hyperlinks to be promoted in varying order of

preference, depending on to which statistical

features have been given priority in ranking. The

highest-ranking hyperlinks with each alternative

ranking can provide different perspectives for

exploration in the article network for a learner.

In our work we decided to generate ranking of

hyperlinks with such simple features that would be

as much as possible motivated by the main

functionality of the Wikipedia. We concluded by

naming five key functions of the Wikipedia and

corresponding measurable features for ranking. They

are: adding new content (article size), editing

content (editing rate), providing cross-linking

(hierarchy of hyperlinks), explaining concepts and

their relation (repetition of hyperlink terms) and

using articles as a reference (viewing rate). Each of

these five features enable relatively straightforward

ranking of hyperlinks.

For example in November 2009, Wikipedia

article about “Life” contained hyperlinks to target

articles such as “Earth” (9152 edits, size 417 499

bytes) and “Metabolism” (1478 edits, size 456427

bytes). With common statistical reasoning, one

could expect that the coverage of the article

“Metabolism” (bigger article) might be broader but

that the peer-review process of the article “Earth”

(more edits) might be more extensive. Among these

two hyperlinks, prioritizing the article size would

promote relating concept of life to metabolism and

prioritizing the number of edits would promote

relating concept of life to earth. Naturally,

cautiousness is needed in statistical evaluation

especially when having relatively low frequencies.

From the sentence surrounding the hyperlink it is

possible to parse and extract a compact explanation

phrase that depicts the semantic relation between

current article and the hyperlink’s target article. In

the previous example, one could produce two

alternative robust relation statements from the text of

article “Life” (November 2009): that forms of life

“can be found in the biosphere on earth”, and that

during life “living organisms undergo metabolism”.

For a learner to exploit the perspective that is

currently emphasized in ranking, it is

recommendable to traverse some of the high-ranking

hyperlinks. The current article, the target article

corresponding to the selected high-ranking hyperlink

and the relation statement between them can be

intuitively represented and further evaporated in a

concept map. In the concept map, nodes labelled

with article titles are connected with directed arcs

labelled with relation statements. By expanding the

concept map step by step with a preferred ranking of

hyperlinks the learner can explore and build learning

paths emphasizing desired perspectives.

3.2 Principles of Ranking

We think that the order of appearance of hyperlinks

in the article is the simplest ranking of hyperlinks to

exploit since it is inherently available in the article

text. This ranking can be assumed to suggest that the

hyperlinks in the beginning of the current article

point to articles whose titles emphasize giving a

definition of the current article. Reason for this is

that a Wikipedia article often starts with a compact

definition containing a few hyperlinks. Respectively,

the hyperlinks in the end of current article likely

point to articles whose titles emphasize giving

broader details of the current article. We will refer to

this type of ranking as the “Hierarchy of

hyperlinks”. Other rankings can be expected to rely

on alternative prioritization for hyperlinks.

Statistical features of an article can be computed

directly from the article or its revision history, or

then retrieved from the open statistics database

provided by the Wikipedia foundation. Several

specialized web sites provide an easy interface for

making queries with the statistics database.

In preliminary testing we evaluated a varied

randomized sample of 100 Wikipedia articles. We

tried to identify what kinds of target articles of

hyperlinks become typically favoured when ranking

is performed in respect to each of five features

introduced in Section 3.1. We also tried to identify

cases in which some articles become misleadingly

favoured against these expectations just noted for

each ranking. Table 1 shows our conclusion based

on the sample and describing a hypothesis about

favourable and misleading cases. We assume that on

average same kinds of tendencies appear with any

random Wikipedia article when ranking its

hyperlinks in respect to each of five features.

We propose a new method that enables to

generate personalized learning paths from the

Wikipedia online encyclopedia based on various

rankings that are retrieved from the statistics of the

articles. The method should fruitfully support

principles of constructivism and transferable

learning.

Relying on the current learning task the learner

needs to be first provided with a start concept for

exploration. The method retrieves a Wikipedia

article whose title matches with this concept. Then

the learner is provided with a list of hyperlinks in

this current article, sorted in a desired type of

ranking based on the statistics of target articles. The

list shows the title of target article for each

CSEDU 2010 - 2nd International Conference on Computer Supported Education

112

Table 1: Some favourable and misleading cases for ranking articles identified in respect to five

features.

hyperlink, accompanied with a compact relation

statement parsed from the sentence surrounding the

hyperlink. The learner is expected to evaluate the

provided hyperlinks and select intuitively, based on

personal needs and consideration, the most

promising hyperlink for extending exploration in the

article network. By selecting especially a high-

ranking hyperlink the learner can emphasize getting

a perspective specific to present type of ranking.

Each selected hyperlink progressively expands a

concept map that is shown to the learner, defining

learning paths. According to the selected hyperlink,

a new node is added to concept map, labelled with

the title of the target article. A directed arc is added

correspondingly leading from current node to new

node, labelled with the relation statement.

According to personal needs, the learner can

choose which type of ranking is used for sorting the

hyperlinks. The hyperlinks are sorted based on five

different rankings that are generated from the

statistics of the target articles. Alternative

perspectives are available based on following five

features, introduced in Section 3.1. As we have

already motivated earlier, “Hierarchy of hyperlinks”

denotes showing hyperlinks in the natural order of

increasing distance from the beginning of the article.

“Repetition of hyperlink terms” denotes showing

hyperlinks in a descending order based on how

many times the word (or group of words) forming

the title of hyperlink’s target article is mentioned in

the current article, anywhere in its full textual

content. This ordering is motivated by an

assumption that the title of target article for each

hyperlink defines a key term for current article. The

more this key term is repeated in the text of current

article, the more it seems to indicate that the

corresponding target article is highly involved in

formulating relations with the current article.

“Article size” denotes showing hyperlinks in a

descending order based on the total amount of

characters in the target article text. A motivation for

this ordering is that a bigger article size obviously

indicates more detailed content than a smaller article

size. The value of article size is approximated with

the file size in bytes that is extracted from the header

of the target article file.

“Viewing rate” denotes showing hyperlinks in a

descending order based on frequency of viewing

hyperlink’s target article by the global community.

This ordering is motivated by the assumption that an

article with a high viewing rate has a higher general

interest than an article with a low viewing rate. This

value represents total number of views per previous

full month. This information is retrieved from online

service (Wikipedia article traffic statistics, 2009)

that relies on data gathered from Wikipedia’s squid-

based cache server cluster. “Editing rate” denotes

showing hyperlinks in a descending order based on

frequency of editing hyperlink’s target article by the

global community. A motivation for this ordering is

that higher editing rates seem to indicate more

verified content than lower editing rates. The value

of editing rate is approximated with the total number

of edits for current article since its creation. This

information is retrieved from online service

(Wikipedia page history statistics, 2009) that builds

an edit history overview page for the article with the

given name.

Besides these five principal features, we still

suggest a supplementing feature that is a user-

defined weighted mixture of them all.

4 EXPERIMENT

4.1 Alternative Perspectives

A preliminary testing of the method was carried out

with a sample of 30 most frequent nouns in English

retrieved from British National Corpus (Kilgarriff,

2009). We found out that when testing the effect of

each of five features separately for ranking,

relatively different perspectives were achieved for

building personalized learning paths.

The gained perspectives can be characterized on

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS

113

Table 2: Ranking of hyperlinks of article "Life" in respect to five features.

various semantic levels, such as causality,

conjunction, temporality, quality, space and

participants. We think that there is a whole new

research domain opening in this ranking-based

exploration of wiki environments. Since many

statistical uncertainties can bring bias to

generalisation if testing does not cover very large

experimental samples we want to be cautious when

reporting our first preliminary results achieved with

relatively small samples. However, to verify overall

functionality of our computational method we feel it

to be necessary to first start with smaller samples.

To illustrate the rich varied perspectives gained

with our method Table 2 shows target articles of

high-ranking hyperlinks of Wikipedia article “Life”

(October 2009). In each major column of the table

hyperlinks are ranked based on each feature

separately. The columns “Main text” and “Only

Intro” indicate if the ranking is done for all

hyperlinks of the full article text or only for

hyperlinks mentioned in the introduction section

before the table of contents. Applying ranking only

to introduction section seems to help highlighting

fundamental relations and improves computational

performance thus decreasing delay of getting results

with the method.

We can sum all types of rankings together for

each hyperlink. Then three highest-ranking

hyperlinks in descending order for the main text are

Evolution, Earth and Organism, and for only the

introduction section Earth, Philosophy and

Organism. When ranking is done only for the

hyperlinks mentioned in the introduction section, the

promoted hyperlinks appear to be more sheared

among various perspectives than when ranking is

done for all hyperlinks of the full article text. In

addition, hyperlinks that do not belong to the

introduction section and thus are ignored from the

ranking based on the introduction section can take

high-ranking positions when ranking is based on the

full article text. This phenomenon happens with

hyperlinks Evolution and Aristotle.

We next evaluate characteristics emerging with each

feature in Table 2 against our hypothesis about

favourable cases that we outlined in Table 1. In the

following we make some brief notions from Table 2.

“Hierarchy of hyperlinks” promotes relatively

definitive and general hyperlinks. Order of

hyperlinks is the same for main text and introduction

section, thus no differences in ranking. “Repetition

of hyperlink terms” promotes for example gene and

RNA that are short words and small structural units.

“Article size” promotes hyperlinks to articles that

appear to represent major themes and relatively

biologically oriented components in life. ”Viewing

rate” promotes hyperlinks to rather diverse set of

everyday concepts belonging to life. “Editing rate”

promotes hyperlinks to articles dealing with actively

debated topics such as Evolution, Philosophy, Death

and Religion. Overall, we can conclude that these

findings match well with our previously made

hypothesis about distinctive characteristics for each

feature used in ranking of hyperlinks.

4.2 Personalized Learning Paths

To evaluate effect of the proposed method to

produce personalized learning paths we decided to

explore hyperlink chains starting from a Wikipedia

article with alternative perspectives. We produced

the learning paths in the form of concept maps by

exploring hyperlink chains following the ranking in

respect to five features described in Section 3.1. We

continued testing with the previous sample of 30

English nouns. Due to convincing but diverse

outcome we suggest extensive further evaluation

with large samples. We introduce some preliminary

results illustrated with the case of Wikipedia article

“Life”.

Figure 1 shows concept maps that we produced

for each of five . perspectives when taking into

CSEDU 2010 - 2nd International Conference on Computer Supported Education

114

Figure 1: Learning paths starting from article "Life" with five alternative perspectives: a) hierarchy of hyperlinks, b)

repetition of hyperlink terms, c) article size, d) viewing rate, and e) editing rate.

account all hyperlinks in the full article text.

Relation statements were extracted from sentences

surrounding the hyperlinks with a method

introduced in our previous work (Lahti, 2009). Due

to space constraints, we show here just short linear

versions of concept maps that rely on traversing only

the highest-ranking hyperlink at each expanding

step. However, the method enables the learner to

build concept maps with a free design in respect to

branching, interconnections and loops. When

generating Figure 1, if the highest-ranking hyperlink

pointed to an already visited article, a hyperlink with

lower rank was accepted to avoid looping. Same was

done when relation statement extraction failed to

produce a phrase, due to two special segments of the

article. Evaluating the learning paths introduced by

each of the five features gives promising results. The

learning paths seemed to offer some distinctive

perspectives corresponding to our previous

hypothesis that we outlined in Table 1.

“Hierarchy of hyperlinks” (Figure 1a) produced a

learning path that remains constantly on relatively

high level of conceptual hierarchy in the topic. This

type of learning paths could effectively introduce for

example main chapters of the curriculum.

“Repetition of hyperlink terms” (Figure 1b)

produced a path that goes through conceptual

structures of the topic across various hierarchical

levels. This type of path could suit well to learning

how the curriculum in deeper levels relies on rich

variations of some basic components. “Article size”

(Figure 1c) produced a path highlighting a collection

of the most broadly documented concepts of the

topic. This type of path could help in getting idea

about the most respected and stabilized parts of the

curriculum. “Viewing rate” (Figure 1d) produced a

path showing those concepts of the topic that get the

most attention from the general public. This type of

path could indicate which parts of the curriculum are

the most referenced ones. “Editing rate” (Figure 1e)

produced a path that offers concepts in the topic that

are actively debated by the general public. This type

of path could illustrate the parts of curriculum that

are involved in constructive criticism and

reconsideration.

We think that these preliminary results give

support for the proposed method and motivate

further research. We think that altering the use of

different rankings provided by the five features

enables the learner to explore the article network of

Wikipedia adaptively. Following his/her intuition

and consideration the learner can address current

pedagogical needs by selecting the most suitable

type of ranking and then the most promising

hyperlink. Thus, the learner himself/herself can

build personalized learning paths for various

educational purposes. Building learning paths is

naturally very sensitive for each consecutive step of

selecting the most promising hyperlink in the current

article for further exploration. Additional testing

indicated that selecting the hyperlinks only from the

introduction section of the article increased keeping

the steps of the learning path more closely related.

We think that the learner needs to make several

parallel branching learning paths to ensure sufficient

diverse coverage of the learning topic.

5 RELATED WORK

5.1 Ontology Construction

To bring structure to the meaningful content of web

pages, so called semantic web approach aims to

introduce ontologies as a formal representation for

concepts within a domain and the relationships

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS

115

between them (Berners-Lee et al., 2001). However,

many traditional ontology projects have received

criticism about being too closed, formal and hard to

update (Simperl & Tempich, 2006). Zouaq and

Nkambou (2009) proposed a method to

automatically generate a domain ontology from

plain text documents and use this ontology as the

domain model in computer-based education. They

suggested evaluating the generated domain ontology

with three dimensions: structural, semantic, and

comparative. Pirrone et al. (2005) proposed an

approach to automated learning path generation

inside a domain ontology supporting a web tutoring

system. Inspired by the knowledge space theory,

they suggest heuristics to transform an ontology in a

weighted graph where the A* search algorithm is

used to find the path.

We think that a community-driven approach,

such as wiki environments, can well support

dynamic collaboratively defined ontologies. The

Wikipedia does not have permanently fixed

categorization of its content and the relations can

sustain even radical changes to respond the changes

in the average worldview. The content providers are

asked to take care of updating the organization of the

content as well. Since previous versions can be

always reverted, it is safe to let the structure freely

slowly converge towards a consensus while

complementary contributions are gathered.

Despite uncertainties, the Wikipedia has been

considered as a promising source for ontology

construction (Haase & Völker, 2008; Hu, in press).

Every Wikipedia article describes one concept

denoted by the title of the article that has been

considered having value for general public. Each

hyperlink of this article literally shows a path to

another related concept that has been collectively

valued so much that a specific article has been

written about it as well.

With swarm intelligence, spontaneous indirect

coordination between agents can show optimal

learning paths with a form of self-organization

called stigmergy (Gutiérrez et al., 2006). Similarly,

we think that automated generation of favourable

learning paths can be effectively based on

proceeding in the conceptual network represented by

Wikipedia articles and inter-article hyperlinks.

Graph based visualizations relying on ontologies

extracted from the Wikipedia have been proposed

for education (Dicheva & Dichev, 2007; Yang et al.,

2007). We now suggest extending the use of

ontologies extracted from the Wikipedia to be

applied in building personalized learning paths. This

poses requirements to assess the quality of articles

and perspectives that they can provide.

With an aim to enhance the quality of articles,

the Wikipedia community has been labelling in a

specific review process some satisfactory articles as

“good articles” and even more professional ones as

“featured articles”. Blumenstock (2008) showed that

the featured articles can be recognized correctly with

the accuracy of 96 % using a simple heuristic that

classifies articles with more than 2000 words as

“featured” and articles with fewer than 2000 words

as “random”. Thomas and Sheth (2007) showed that

when comparing labelled good articles to other non-

stub articles having at least 50 revision milestones

they found no statistically significant difference in

convergence to a semantically stable state.

These two previous results indicate that the

maturity of an article can be measured relatively

well even with simple parameters. This seems to

support our attempt to identify few basic features of

an article that can be easily measured to create

rankings for hyperlinks, highlighting alternative

perspectives that they provide. Adoption of features

denoting “Article size” and “Editing rate” for our

method was especially inspired by these two

previous results.

Since each collaborating author of a Wikipedia

article inherently provides a different perspective,

the authors can together easily cover a rather wide

range of issues and typically produce a rich creative

output. The iterative additions, evaluations and edits

done collectively seem to be capable of assuring the

quality and coherence of the steps leading towards

the desired stabilized state of a mature article.

5.2 Accumulating Knowledge

Semantic relation analysis tries to identify semantic

units and their relations from natural language.

Previous research has strongly relied on machine

learning approach or probabilistic methods.

Semantic roles have been identified for example

from syntactic units by using various lexical

resources and predefined relations. When trying to

generate automatically favourable learning paths that

match with the learner’s needs the guidance should

have a suitable balance between constraints for

sustainability and freedom of association.

Nastase and Szpakowicz (2006) introduced an

incremental learning algorithm that effectively

mimics the way in which a human reader

accumulates knowledge and exploits it to process

new text. The algorithm assigns a semantic relation

to semantic units of text taken from a science book

with guidance from a user and builds a simple

syntactic-semantic graph surrounding the central

concept and matching it with previously analysed

CSEDU 2010 - 2nd International Conference on Computer Supported Education

116

text. Our proposed method provides an analogous

approach relying on user-driven generation of

learning paths in the form of concept maps based on

article network of the Wikipedia.

Coursey et al. (2008) argued that a combination

of keyword extraction techniques combining graph-

theoretical algorithms and methods relying on

knowledge extracted from the Wikipedia can be

successfully used to identify candidate keywords in

learning objects. They suggest using ranking

algorithm over the Wikipedia connectivity graph to

find relevant articles. Somewhat similarly, our

method exploits the titles of hyperlink’s target

articles to identify promising concepts in the

Wikipedia. We introduce ranking that enables these

concepts to be explored in learning paths,

accompanied with compact relation statements

parsed from the sentences surrounding each

hyperlink.

Serrano et al. (2009) argued that some key

regularities of written text concerning burstiness of

words, topicality and their relationship can be

modelled with two simple algorithmic techniques

that are frequency ranking with dynamic reordering

and memory effect connecting word frequencies

across different documents. They suggest that their

model enables to relate two key mechanisms that

have been assumed to affect how humans process

the lexicon: rank frequency and context diversity.

They propose using their model to study co-

evolution of content and citation structure for

example in the Wikipedia. In a resembling fashion,

our method uses rank frequency and context

diversity of the Wikipedia enabling a learner to

process lexicon to a pedagogically rewarding

structure.

There have been proposals to visually highlight

the most mutually agreed segments in a Wikipedia

article based on simple quality measures. High

survival time of a single edit does not guarantee

reliably its trustworthiness (Luyt et al., 2008). Adler

and de Alfaro (2007) proposed a measure relying on

author’s reputation that can be gained if the edits

he/she performs are preserved by subsequent

authors. It seems challenging to develop measures

taking simultaneously into account the semantics of

the article network and collective contribution

patterns coming from authors and readers. Our

method tries to address these issues.

Pavlovic (2008) introduced a model of network

computation based on Markov chains as an attempt

to extract the semantic content of the data from their

distribution among the nodes. A concept can be

identified by finding the community of nodes that

share it performing together some non-local data

processing. Pavlovic proposes ranking of paths to

extract information about the likely flow biases from

the available static information about the network

and thus to detect semantic structures in a network.

We think that our method is dealing with a same

kind of goal and that the statistics concerning

articles can be useful criteria for ranking the paths.

Haruechaiyasak and Damrongrat (2008)

proposed a method to generate a topic model from

articles in the Wikipedia Selection for Schools that

is a collection aiming to meet curricula world-wide.

Their method relies on latent Dirichlet allocation.

Based on similarity measures computed for topic

distribution profiles of the articles, the method

enables recommending related articles some of

which are not covered by the hyperlinks in a given

article. With some similarities, our method aims to

recommend well related hyperlinks for learning, We

decided to use article statistics as a measure that

enables to highlight different perspectives..

All in all, inspired by the previous work, we tried

to find measures to recommend hyperlinks that offer

pedagogically motivated exploration in the

Wikipedia article network with a preferred

perspective. To make learning paths personalized

there is a need for a method that takes into account

the learner’s preference about which hyperlink is the

most profitable to choose as the next step in the

learning path. To find a favourable chain of

hyperlinks in the Wikipedia a reasonable amount of

familiarity and continuity should be preserved while

still trying to extend knowledge of the learner.

6 DISCUSSION

The proposed method aims to suggest hyperlink

chains that offer highest pedagogic value for the

learner. An essential strength of the method is the

aim to provide a reasonable collection of alternative

hyperlink chains that maintain semantic and

educational relatedness between each step in the

chain and between parallel chains. This is based on

four key factors: collaboratively maintained initial

organization of concepts and relations (evolution of

the Wikipedia), dynamic ranking in respect to five

features supporting alternative perspectives (article

statistics), illustrations denoting previous and current

conceptual reasoning (concept maps), and letting the

learner to make the ultimate decision for next step

based on her intuition and consideration (support for

variety of personalities).

The proposed method relies heavily on extraction

and analysis of hyperlinks in Wikipedia articles

related to a chosen learning topic. Recommendable

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS

117

learning paths are represented as a gradually

expanding concept map that can be directly shown

to the learner and also applied later in various

educational purposes.

The method aims to provide a balanced trade-off

between extensive coverage and compactness in the

generated learning content. The method offers

learning paths that should enable the learner to

traverse the most essential knowledge in the least

amount of time. This traversing can be exploited as

means to adopt new knowledge or to refresh it. The

traversed learning paths become documented as

concept maps thus enabling the learner to analyse

illustratively her conceptualization concerning a

chosen topic. These knowledge structures can be

easily further edited, reused and shared with other

learners.

In contrast with many previous proposals in this

field, we have not only developed a method for

knowledge management but also implemented a

fully functional prototype that is ready to be used in

various educational contexts for many pedagogical

purposes. We do not know any previous proposal

that is similar to our work especially concerning

educational use of the Wikipedia.

We have evaluated ranking hyperlinks of the

article in respect to five different features based on

article statistics. We think that these features

correspond to the most fundamental functions of the

Wikipedia. In our experiments we found distinctive

ways to differentiate exploration of hyperlinks based

on the features preferred by the learner. Using

various rankings it is thus possible to provide

alternative perspectives to knowledge and thus

enable the learners to build independently

favourable learning paths following their personal

needs at the moment.

Concepts belonging to various domains of life

and to various abstraction levels in a certain topic

have obviously different tendencies to support the

five features. Also, features can have many hidden

correlations that should be taken into account for a

balanced use of statistics. High editing rate typically

produces high article size. Typically each single

event of editing article increases also viewing rate if

the editor wants to check the finished version. When

building learning paths, our proposed method

possibly too optimistically expects high relation

between all consecutive concepts in a traversed

chain of hyperlinks.

Hyperlinks of an article can point to target

articles that deal with topics that are opposite or

ambiguous to the title of current article.

Unfortunately, it is hard to develop general methods

that could reliably identify the exact type of relation

between target article and current article. Extracting

relation statements from the sentence surrounding

the hyperlink can also be troublesome since often

the sentence does not explicitly define the

relationship between the title of current article and

the title of target article, but instead describe

something else.

When building learning paths, a major challenge

for semantic continuity is that some measures based

on the characteristics of target article may not

indicate well the actual relatedness between current

article and target article. For example, if ranking of

hyperlinks is based on viewing rate, the target article

having the highest viewing rate is prioritized. But

this viewing rate consists of a great variety of visits

arriving to the target article through various

hyperlinks, not only from current article. Thus,

viewing rates describe just the overall distribution of

visits to individual Wikipedia articles and fail to tell

how the preference to visit a certain target article

varies depending on the current article.

There are limitations with the current method

especially since it was purposefully designed to be

simple and computationally easy. The features used

with the method could be chosen in various

alternative ways. Anyway, if the online services

used for quering statistics should become shut down

it still remains possible to retrieve statistics with

alternative implementations. The method might be

enhanced by using statistics taken from varied time

frames and making diverse temporal analysis for

views, edits, article size, etc. Articles could be

treated more equally if comparison would rely on

proportional values instead of absolute values. Thus,

view rates and edit rates could be considered

proportionally, for example in relation to article size.

The method could also somehow take advantage of

the fact that typically many changes in an article are

performed in bursts, for example after related news

has been published in the media.

Various navigational aids have been introduced

to the layout of Wikipedia articles, for example

category tags, “See also” section, naviboxes and

infoboxes. Also redirects, disambiguation pages and

“What links here” queries assist finding related

articles. However, we argue that these assistive

functions complementing each other cannot clearly

recommend the most promising hyperlinks for

further exploration. To increase efficiency of

exploration and to ensure finding the most relevant

hyperlinks, there is a need for adaptive ranking of

hyperlinks of the current article.

Since a lot of articles of the Wikipedia present

facts that have a low probability to become

constantly updated or seriously questioned, we think

CSEDU 2010 - 2nd International Conference on Computer Supported Education

118

that our method could be successfully used also off-

line. Despite of its huge coverage, the plain textual

content of English language version of Wikipedia

can be stored locally in one compressed file of 5

gigabytes. The method might use also the article

statistics from just off-line sources. We suggest that

already the statistics gathered so far can enable

creating reliable ranking that reflects conception of

global community. Relying on off-line content

would enable using the method with very low

computational costs and minimal delay.

7 FUTURE WORK

The proposed method and experiment have indicated

a promising unexplored area for research concerning

new methodology to adaptively explore the

knowledge space of the Wikipedia. We suggest that

the method we have developed for the Wikipedia

can be relatively well applied to also other

collaborative knowledge management environments

and even intellectual mental processes in human

mind.

In the future research there is a strong need for

further classifying various features that can be used

in ranking of hyperlinks that connect concepts (or

articles). It can be possible to identify individual

most favourable features for each domain of

knowledge. These specific features could enable

exploring knowledge in most coherent manner

taking into account special characteristics that are

typical for this domain.

It is also important to develop methods that can

address individual characteristics of every learner.

For each learner it could be identified what are the

features that need to be used in ranking of hyperlinks

to fulfil his/her special personal needs. For example,

preferred learning style, personality and hobbies of

the learner could be considered when setting the

ranking criteria that affect which hyperlinks become

promoted to the learner. Furthermore, it would be

advantageous that the learner could himself/herself

make adaptively consistent decisions about what

features to prioritize in ranking when exploring

varying learning contexts. In many cases, user-

defined ranking criteria should not probably support

just one perspective but instead to be a dynamic

weighted mixture of them all.

In addition, it is important to develop advanced

but still computationally sustainable analysis

methods that help to rank alternative hyperlinks and

thus to find most promising learning paths. It is

important to have such analysis methods that are not

dependent on any proprietary online service. To

effectively develop and ensure automated

knowledge management it is important to support

open access knowledge bases and open source

modules. Interfaces should be kept as interoperable

and standardized as possible to best promote

updating individual components of modular

applications or replacing them with alternatives.

Knowledge management tools should be actively

introduced for using them in ordinary life for

example in education, problem solving, decision

making, design and innovation. Research should

emphasize access for all since knowledge tools are

often most crucial for people with special needs. The

efforts should aim at providing a better quality of

life and letting the learner to excel oneself and

follow his/her personal interests.

REFERENCES

Adler, B., & de Alfaro, L. (2007). A content-driven

reputation system for the Wikipedia. Proc. 16th

international conference on World Wide Web, Banff,

Alberta, Canada, ACM Press, 261-270.

Berners-Lee, T., Hendler, J. & Lassila; O. (2001). The

semantic web. Scientific American Magazine, May

2001.

Blumenstock, J. (2008). Automatically assessing the

quality of Wikipedia articles. University of California

at Berkeley, School of Information. Technical Report

2008-021.

Chesney, T. (2006). An empirical examination of

Wikipedia's credibility. First Monday, 11(11).

Coursey, K., Mihalcea, R., & Moen, W. (2008).

Automatic keyword extraction for learning object

repositories. Proc. Conference of the American

Society for Information Science and Technology

(ASIST 2008), Columbus, Ohio, USA.

Dicheva D., & Dichev C. (2007). Helping courseware

authors to build ontologies: the case of TM4L. Proc.

13th International Conference on Artificial

Intelligence in Education, (AI-ED 2007), Los Angeles,

California, USA, IOS Press, 77-84.

Gandrabur, S., Foster, G., & Lapalme, G. (2006).

Confidence estimation for NLP applications.

Transactions on Speech and Language Processing,

3(3), 1-29.

Guthrie, J., Wigfield, A., Barbosa, P., Perencevich, K.,

Taboada, A., Davis, M., Scafiddi, N., & Tonks, S.

(2004). Increasing reading comprehension and

engagement through concept-oriented reading

instruction. Journal of Educational Psychology, 96(3),

403-423.

Gutiérrez, S., Pardo, A., & Kloos, C. (2006). Finding a

learning path: toward a swarm intelligence approach,

Proc. 5th IASTED international conference on Web-

based education, Puerto Vallarta, Mexico, ACTA

Press, 94-99.

PERSONALIZED LEARNING PATHS BASED ON WIKIPEDIA ARTICLE STATISTICS

119

Haase, P., & Völker, J. (2008). Ontology learning and

reasoning – dealing with uncertainty and

inconsistency. In da Costa, P. et al. (eds.), Uncertainty

Reasoning for the Semantic Web I. LNAI 5327, 366-

384.

Haruechaiyasak, C., & Damrongrat, C. (2008). Article

recommendation based on a topic model for Wikipedia

Selection for Schools. Proc. 11th International

Conference on Asian Digital Libraries, LNCS 5362,

339-342.

Holmes, B., Tangney, B., Fitz-Gibbon, A., Savage, T., &

Mehan, S. (2001). Communal constructivism: students

constructing learning for as well as with others. Proc.

12th International Conference of the Society for

Information Technology and Teacher Education (SITE

2001), Orlando, Florida, USA, 3114-3119.

Hu, B. (in press). WiKi’mantics: interpreting ontologies

with Wikipedia. Journal of Knowledge and

Information Systems, Springer. DOI: 10.1007/s10115-

009-0247-6

Janssen, J., Berlanga, A., Vogten, H., & Koper, R. (2008).

Towards a learning path specification. International

Journal of Continuing Engineering Education and

Lifelong Learning, 18(1).

Kilgarriff, A. (2009). BNC database and word frequency

lists. Lemmatized frequency list of British National

Corpus. URL: http://www.kilgarriff.co.uk/bnc-

readme.html

Lahti, L. (2009). Guided generation of pedagogical

concept maps from the Wikipedia. Proc. World

Conference on E-Learning in Corporate, Government,

Healthcare and Higher Education (E-Learn 2009),

Vancouver, Canada, 1741-1750. ölvlably pwjxjyt

åjtdtj, pwzyt ldltybcvtulb jtbt taj bccwt, vttbza.

Luyt, B., Aaron, T., Thian, L., & Hong, C. (2008).

Improving Wikipedia's accuracy: Is edit age a

solution? Journal of the American Society for

Information Science and Technology, 59(2), 318-330.

Nation, P., & Waring, R. (1997). Vocabulary size, text

coverage, and word lists. In Schmitt, N., & McCarthy,

M. (eds.), Vocabulary: Description, acquisition,

pedagogy. Cambridge University Press, New York,

USA, 6-19.

Nastase, V., & Szpakowicz, S. (2006). Matching

semantic-syntactic graphs for semantic relation

assignment. Proc. Textgraphs 2006 Workshop on

Graph-based Algorithms for Natural Language

Processing, New York, USA.

Neumann, D., & Hood, M. (2009). The effects of using a

wiki on student engagement and learning of report

writing skills in a university statistics course.

Australasian Journal of Educational Technology,

25(3), 382-398.

Pavlovic, D. (2008). Network as a computer: ranking

paths to find flows. Proc. Third International

Computer Science Symposium in Russia, LNCS 5010,

384-397.

Peregrin, J. (in press). The myth of semantic structure. In

Stalmaszczyk, P. (ed.), Philosophy of Language and

Linguistics. Volume I: The Formal Turn. Ontos,

Frankfurt. Germany. URL:

http://jarda.peregrin.cz/mybibl/PDFTxt/528.pdf

Pirrone, R., Pilato, G., Rizzo, R., & Russo, G. (2005).

Learning path generation by domain ontology

transformation. Proc. 9th Congress of the Italian

Association for Artificial Intelligence, LNAI 3673,

359-369.

Reinoso, A., Ortega, F., Gonzalez-Barahona, J., & Robles,

G. (2009). A quantitative approach to the use of the

Wikipedia. Proc. IEEE Symposium on Computers and

Communications (ISCC 2009), Sousse, Tunisia, 56-

61.

Schmidt, R.A., & Bjork, R.A. (1992). New

conceptualizations of practice: common principles in

three paradigms suggest new concepts for training.

Psychological Science, 3(4), 207-217.

Serrano, M., Flammini, A., & Menczer, F. (2009).

Modeling statistical properties of written text. Public

Library of Science ONE (PLoS ONE), 4(4): e5372.

Simperl, E., & Tempich, C. (2006). Ontology engineering:

a reality check. Proc. 5th International Conference on

Ontologies, Databases, and Applications of Semantics

(ODBASE2006), LNCS 4275, 836-854.

Thomas, C., & Sheth, A. (2007). Semantic convergence of

Wikipedia articles. Proc. IEEE/WIC/ACM

International Conference on Web Intelligence, Silicon

Valley, California, USA, 600-606.

Van Berkum, J., Brown, C., Zwitserlood, P., Kooijman,

V., & Hagoort, P. (2005). Anticipating upcoming

words in discourse: Evidence from ERPs and reading

times. Journal of Experimental Psychology: Learning,

Memory and Cognition, 31, 443-467.

Wikipedia article traffic statistics (2009). URL:

http://stats.grok.se/

Wikipedia page history statistics (2009). URL:

http://vs.aka-online.de/cgi-bin/wppagehiststat.pl

Yang, J., Jangwhan, H., Oh, I., & Kwak, M. (2007). Using

Wikipedia technology for topic maps design. Proc.

45th ACM Southeast Regional Conference (ACM-SE

45). Winston-Salem, North Carolina, USA, ACM

Press, 106-110.

Zouaq, A. & Nkambou, R.,(2009). Evaluating the

generation of domain ontologies in the Knowledge

Puzzle Project. IEEE Transactions on Knowledge and

Data Engineering, 21(11), 1559-1572.

CSEDU 2010 - 2nd International Conference on Computer Supported Education

120