Content uploaded by Brendan Flanagan
Author content
All content in this area was uploaded by Brendan Flanagan on Feb 28, 2019
Content may be subject to copyright.
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
1
Knowledge Map Creation for Modeling Learning Behaviors in
Digital Learning Environments
Brendan Flanagan1*, Rwitajit Majumdar1, Gökhan Akçapınar1,2,
Jingyun Wang3 and Hiroaki Ogata1
1 Kyoto University, Japan
2 Hacettepe University, Turkey
3 Kyushu University, Japan
* flanagan.brendanjohn.4n@kyoto-u.ac.jp
ABSTRACT: There has been much research that demonstrates the effectiveness of using
ontology to support the construction of knowledge during the learning process. However, the
widespread adoption in classrooms of such methods are impeded by the amount of time and
effort that is required to create and maintain an ontology by a domain expert. In this paper,
we propose a system that supports the creation, management and use of knowledge maps at
a learning analytics infrastructure level, integrating with existing systems to provide modeling
of learning behaviors based on knowledge structures. Preliminary evaluation of the proposed
text mining method to automatically create knowledge maps from digital learning materials is
also reported. The process helps retain links between the nodes of the knowledge map and
the original learning materials, which is fundamental to the proposed system. Links from
concept nodes to other digital learning systems, such as LMS and testing systems also enable
users to monitor and access lecture and test items that are relevant to concepts shown in the
knowledge map portal.
Keywords: Knowledge map; concept-based analytics; concept maps; knowledge extraction;
1 INTRODUCTION
It has been well documented that learners can benefit from the use of maps to represent the key
concepts of knowledge (Lee et al., 2012). Ausubel (1963; 1968) defined the effective assimilation of
new knowledge into an existing knowledge framework as the achievement of “meaningful learning”,
by which knowledge maps can serve as a kind of scaffold to help learners to organize knowledge and
structure their own knowledge framework (Novak et al., 2006). However, the process of creating and
maintaining these maps often involves a domain expert manually creating the knowledge map based
on their experience and previous knowledge (Wang et al., 2017).
To support the creation and use of knowledge maps by teachers and learners, we propose a
knowledge map system that integrates with existing digital learning environments and learning
analytics infrastructure. To assist in the creation of knowledge maps from digital learning materials,
we propose a process for extracting key concepts from unstructured text to generate knowledge
structures. Maps that have been generated are stored in a Knowledge Map Store (KMS) and an
authoring system is provided for teachers to create, edit, and manage stored knowledge maps before
publishing. The Knowledge Portal provides visualizations of knowledge maps with attributes
determined from the analysis of learning behavior event log data from existing learning analytics
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
2
infrastructure. In the final section of this paper, we outline the anticipated cases in which the system
will be utilized by both teachers and students to monitor individual and group knowledge states.
There are many previous researches into the generation and use of ontologies, concept maps, and
knowledge maps in education to show and create knowledge frameworks. Association rules and other
data mining techniques have been used to construct concept maps based on the results of test and
quizzes to show the relation between knowledge that was tested (Hwang, 2003; Tseng et al., 2007;
Chen et al., 2010; Chen et al., 2013). While this technique is applicable to the structured format of
tests, it is difficult to apply similar techniques to unstructured text that is contained in digital learning
materials.
Figure 1: An overview of how the proposed system would integrate with existing LA infrastructure.
2 SYSTEM OVERVIEW
In this section, we provide an outline of the proposed Knowledge Map system, how it integrates with
existing LA infrastructure, and how stakeholders will interact with the system. Fig. 1 shows an
overview of the system with the main components consisting of:
• Existing user facing LA infrastructure, such as: LMS, Digital Learning Material Reader, Testing system,
etc.
• LRS and Analytics Processor.
• Knowledge Extraction Processor.
LRS
KMS
Knowledge
Extraction
Event
Logs
Contents
Analysis
LMS
Digital
Learning
Material
Reader
Te st in g
System,
etc…
KM
Authoring
Knowledge Portal
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
3
• KMS (Knowledge Map Store) and a teacher facing Knowledge Map Authoring portal.
• User facing Knowledge Portal.
The existing user facing infrastructure, such as: LMS, Reader, and Testing system serve as an
interaction event sensor and also as a source of learning material contents that are sent to the
Knowledge Extraction Processor. Recent implementations of LA platforms often utilize an LRS and
Analytics Processor as a pipeline for storing and processing event statement data about the use of
user facing learning systems (Chatti et al., 2017; Flanagan and Ogata, 2017). We use this existing
pipeline to provide information to augment the visualization of knowledge structures representing
the underlying learning materials, lecture attendance, and past academic achievement.
Figure 2: Hierarchy of node attributes based on event log analysis.
The main hierarchy of node attributes based on analytics is shown in Fig. 2, where each level is linked
to important stages in the formal learning process: lecture attendance, reading learning materials,
confirming acquired knowledge through the answering of tests, and attaining a credit for having
satisfied the requirements of a course. The most basic form of effort by a learner is to attend a lecture
in which learning material related with the concept node was covered. When a learner actively reads
the learning material the concept node is attributed as Read. If a learner has correctly answered a test
item relating to the concept, then the node is given the Answered attribute. Finally, the if the student
passes the course then the Credit attribute is assigned.
The Knowledge Extraction Processor analyzes learning content data from the LMS, Reader, and Testing
system. In the present paper, we focus on the extraction of knowledge maps from PDF contents that
have been uploaded to the digital learning material reader. The results of this process are then stored
in the KMS. Teachers are able to manage knowledge maps stored in the KMS through a teacher facing
authoring portal.
3 KNOWLEDGE MAP EXTRACTION FROM CONTENTS
Course curriculum in K-12 education is often well structured and defined by government level
organizations that regulate education. However, higher education often is less regulated with the
course curriculum being decided by the teacher. In Japanese universities, teachers in charge of courses
are busy and course contents are often finished close to when a lecture is due to start, allowing little
time to create knowledge maps manually.
Credit
Answered
Read
Attended
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
4
The authoring section of the proposed system automatically analyzes contents uploaded by teachers
to support the generation of knowledge maps. As a part of the authoring process, the system requires
the map to be checked by the teacher before being used by students. The teacher is also able to edit
the automatically generated knowledge map to add, remove, or alter required sections.
A knowledge map can be thought of as a graph of key points that are contained within the digital
learning material contents that it represents. The relation between nodes of this graph are expressed
as a weighted edge representing the strength of the relation between two key points that are in the
contents. In this paper, we use a process based on a method previously proposed by Flanagan et al.
(2013) as shown in Fig. 3.
Figure 3: The process used to extract a knowledge map from digital learning material contents.
The lecture slides are usually written in Japanese with sections also in English. The text is extracted
from lecture slides PDF files using pdfminer
1
and parsed with MeCab (Kudo, 2006) to separate
individual words and parts-of-speech (POS) from a sentence using morphological analysis. Key concept
terms are extracted by selecting the longest sequences of nouns and conjugate particles in a sentence.
These were then indexed using the GETAssoc
2
search engine to form a co-occurrence matrix of terms.
The link between the concept terms and the sections of the learning material are also included as an
attribute in the search engine so relevant learning resources can be retrieved. The final step of the
process involves minimizing the complexity of the co-occurrence graph using a minimum spanning
tree algorithm to select the strongest concept term relations. In this implementation a thesaurus of
technical terms in Japanese and English was used to guide the knowledge map generation process
with hierarchical selection.
Table 1: Learning materials for the evaluation.
Lecture
Pages
Concepts (Gold Standard)
Max Concepts (Proposed)
1
30
12
125
2
32
10
153
3
45
6
222
We conducted a preliminary experiment using the proposed method in a university course on
Information Science. A knowledge map that includes the concepts of three lecture learning materials
was created manually by the course teacher and used as the gold standard for evaluation as shown in
Table 1. Knowledge maps were automatically generated for each lecture with the strongest relation
calculated using the SMART weight as described in Salton (1983). The precision/recall evaluation when
comparing generated maps to the gold standard is shown in Fig. 4 with maximum precision of 0.72 at
1
https://euske.github.io/pdfminer/
2
http://getassoc.cs.nii.ac.jp
Extract Key
Concepts
Generate
Co-occurrence
Graph
Minimize
Graph
Complexity
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
5
a threshold of 11 nodes for each generated map. As the threshold is increased the precision decreases,
however the evaluation shows a majority of correct nodes are extracted at low thresholds. The
generated knowledge maps would require some manual editing by a teacher before use in order to
represent the same structure as the gold standard, and therefore is an ongoing topic of research.
Figure 4: Plot of precision recall of proposed method
4 KNOWLEDGE MAP STORE AND AUTHORING
A centralized storage system of curated knowledge maps is fundamental to the analysis of knowledge
accumulated over long-time spans. At the center of the proposed system, a KMS (Knowledge Map
Store) acts as an LRS would for a conventional LA platform, collecting data about learning materials
from disparate tools and systems to reduce information silos. This could enable the cross referencing
and merging of knowledge maps from separate courses, learning materials, and even educational
institutions if a KMS is deployed at the inter-institutional level.
The key data that a KMS should store are:
• The structure of knowledge maps that have been generated automatically by the system or
created manually by teachers using the authoring interface.
• Links from the concept nodes of a knowledge map to related lecture schedule, learning materials,
test items, and learner academic achievement records.
We are proposing that the structure of the knowledge map and links to learning materials/test items
should be stored using a standards-based RDF storage service.
The proposed system has an authoring portal to facilitate the creation and management of knowledge
maps by teachers. Automatically generated maps are initially stored as a draft and are not publicly
available until the course teacher has confirmed the structure and its link to learning materials/test
items. Maps can be edited to remove irrelevant nodes and add nodes that are required to cover the
concepts in the course. A search function similar to the proposed knowledge map extraction process
can be used to support the linking of relevant sections of learning materials and tests items to
manually added nodes.
Knowledge maps can also be related with global concepts in the KMS to support large scale knowledge
mapping across multiple courses. This feature is intended to facilitate the analysis of prior learner
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
precision
recall
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
6
knowledge, thus allowing a teacher or learner to view what concepts learners have and have not
acquired. There is also potential to apply the results of the analysis to recommend learning materials
that should be studied to fill in knowledge gaps before attending a course.
5 KNOWLEDGE PORTAL
Once a knowledge map has been published with the authoring tool, it is available for use by students
and teachers in the Knowledge Portal. The visualization interface for the proposed Knowledge Portal
is based on a web-based open source ontology visualization system called WebVOWL (Lohmann et al.,
2014). The interface of the proposed system is shown in Fig. 4 with the main knowledge map
visualization on the left, and the right frame displays detailed information about the attributes of the
selected node with relation to relevant learning materials. At the top of the right frame the use is
given an overview of the percentile rank for each of the attributes: attend, read, and answer. It will
also show if a credit has previously been attained in relation to the node concept. The user is able to
follow the links to study learning materials or confirm their knowledge by a test item on the node
concept. The visualization also features a filter to select specific nodes/relations and reduce the
complexity of the knowledge map using varying degrees of edge collapse.
Figure 4: The user interface of the proposed system.
Additional functionality supports the augmentation of the base map structure with analytics results
as visual attributes of nodes as was shown in Fig 3, to give uses visual ques to the overall knowledge
state.
Figure 5: Node visual augmentation definition.
Lecture 1
Concept B:
Concept A
define
define
associated
Concept C
Concept D
Concept B
Test 1.5
C
A
R
A
R
A
Question 3
Read
Viewed
Interacted
Recall Mastered
Credit
Not
Answered Answer ed
Attended
No Attrib
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
7
The outline color of a node represents the level of a learner’s effort with the relevant learning
materials that describe the node concept, and the fill color of the node relates to the learner’s
knowledge level of the node concept as shown in Fig 5. The degree of coloring in both the outline and
fill are displayed to represent the percentile rank of achievement when compared to the whole
student cohort. If there is no or very low percentile rank of event data for Attendance/Read/Credit
and Answer relating to a node concept, then the outline and fill are displayed as grey.
6 USES OF THE KNOWLEDGE PORTAL
The following section outlines different cases in which the knowledge portal could be used to guide
both teaching and learning. An overview of the four main cases is shown in Fig. 6.
Figure 6: Different cases in which Knowledge Graphs could inform learns and teachers.
The first case is that of a learner confirming their current knowledge state for the support of self-
regulated learning. It is intended that the learner could use the knowledge portal for the monitoring
and planning of their learning by searching for concepts that they have not yet studied and following
the links to appropriate learning materials to reading and test items to confirm their knowledge.
The second case enables the learner to reflect on how their knowledge has evolved over a period of
time short or long, such as a student’s knowledge at: t1 = elementary school, t2 = high school, t3 =
undergraduate university. This could also be used to help students find possible gaps in their
knowledge that occurred in the past, and enable the revision of learning materials to resolve
knowledge gaps.
The final two cases deal with comparing the knowledge state of groups of learners. For a student, this
can enable them to compare their own knowledge to that of the broader student cohort and find
possible areas in which their knowledge is lacking. The learner can then study to improve their
1
Knowledge State (KS)
time →
Compare
KS over Time
Compare KS to Other Students
Group KS Overview
t1t2t3
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
8
knowledge state by working on specific concepts by reading learning materials and testing themselves
with linked resources.
Teachers can also benefit from using the proposed knowledge map system to get an overview of the
current knowledge state of all of the students in their course. The individual knowledge maps of all of
the students are merged into a single aggregated knowledge map. An example use of this would be
to check the prior knowledge of students before they attend a lecture, or checking the degree to which
students have previewed concepts and the related learning materials to an upcoming class. The
teacher then can adjust the lecture to either skip concepts that have been adequately learnt, or focus
on concepts that require revision or greater explanation. It is expected that this case will be of
particular use when managing courses with large numbers of students.
At a global knowledge map level, the relation between courses could provide insight into what parts
of the knowledge map are important and central knowledge to a subject, and highlight what parts are
difficult for students to understand and could be incorporated as a filter feature in the knowledge
portal. This can be utilized in two different ways: for teachers it gives them an understanding of what
knowledge is difficult to understand and may require more thorough explanation, and for students it
allows them to see the knowledge that is central to the course and what areas they should pay
attention to as it has been difficult for past students.
Knowledge map analysis could also be used in the recommendation of contents both inside and
outside the course to learners based on their achievement and focus. Under achieving students may
benefit from the recommendation of learning materials that cover concepts that they have yet to
master. On the other hand, outperforming students may be interested in exploring extra learning
materials outside of the course to expand their knowledge beyond that which would be traditionally
offered.
7 CONCLUSIONS
In this paper, we proposed a system to support the creation, management and use of knowledge maps
in digital learning environments at a learning analytics infrastructure level. In particular, we proposed
processes for the automatic extraction, authoring, storing and use of knowledge maps by students
and teachers. For the automatic extraction process, we proposed a text mining method for generating
knowledge maps from digital learning materials and conducted a preliminary experiment to evaluate
its effectiveness. A key feature of the method is the ability to link extracted concept nodes directly to
specific parts of the learning materials from which they were extracted. These links are used to provide
not only a reference for users to the original materials, but also as a method of associating learning
behavior logs collected in existing system and mapping the analysis of these logs directly onto the
knowledge map. This provides feedback to the user about the current learning behavior state overlaid
on a knowledge structure.
In future work, the use of the knowledge portal to increase learner knowledge awareness and group
formation by knowledge map clustering should be investigated.
Companion Proceedings 9th International Conference on Learning Analytic s & Knowledge (LAK19)
Creative Comm ons License, Attribution - NonCommerci al-NoDerivs 3.0 Unported (CC B Y-NC-ND 3.0)
9
ACKNOWLEDGMENTS
This work was supported by JSPS KAKENHI Grant Number 16H06304.
REFERENCES
Ausubel, D. P. (1963). The psychology of meaningful verbal learning, New York: Grune and Stratton.
Ausubel, D. P., Novak, J. D., & Hanesian, H. (1968). Educational psychology: A cognitive view (Vol. 6).
New York: Holt, Rinehart and Winston.
Chatti, M. A., Muslim, A., & Schroeder, U. (2017). Toward an open learning analytics ecosystem. In Big
data and learning analytics in higher education (pp. 195-219). Springer, Cham.
Chen, S. M., & Sue, P. J. (2013). Constructing concept maps for adaptive learning systems based on
data mining techniques. Expert Systems with Applications, 40(7), 2746-2755.
Chen, S. M., & Bai, S. M. (2010). Using data mining techniques to automatically construct concept
maps for adaptive learning systems. Expert Systems with Applications, 37(6), 4496-4503.
Flanagan, B., Yin, C., Inokuchi, Y., & Hirokawa, S. (2013). Supporting interpersonal communication
using mind maps. The Journal of Information and Systems in Education, 12(1), 13-18.
Flanagan, B., & Ogata, H. (2017). Integration of Learning Analytics Research and Production Systems
While Protecting Privacy, Proceedings of the 25th International Conference on Computers in
Education (ICCE2017), 333-338.
Hwang, G. J. (2003). A conceptual map model for developing intelligent tutoring systems. Computers
& Education, 40(3), 217-235.
Kudo, T. (2006). Mecab: Yet another part-of-speech and morphological analyzer.
http://taku910.github.io/mecab/.
Lee, J. H., & Segev, A. (2012). Knowledge maps for e-learning. Computers & Education, 59(2), 353-364.
Lohmann, S., Link, V., Marbach, E., & Negru, S. (2014). WebVOWL: Web-based visualization of
ontologies. In International Conference on Knowledge Engineering and Knowledge
Management (pp. 154-158). Springer, Cham.
Novak, J. D., & Cañas, A. J. (2006). The theory underlying concept maps and how to construct them.
Technical report IHMC CmapTools 2006-01 Rev 01-2008. Florida Institute for Human and
Machine Cognition.
Salton, G., & McGill, M. (1983). Introduction to modern information retrieval. McGraw-Hill, New York.
Tseng, S. S., Sue, P. C., Su, J. M., Weng, J. F., & Tsai, W. N. (2007). A new approach for constructing the
concept map. Computers & Education, 49(3), 691-707.
Wang, J., Flanagan, B., & Ogata, H. (2017). Semi-automatic construction of ontology based on data
mining technique. In 6th IIAI International Congress on Advanced Applied Informatics (IIAI-
AAI) (pp. 511-515). IEEE.