A BOOTSTRAPPING METHOD FOR AUTOMATIC CONSTRUCTING OF THE WEB ONTOLOGY INSTANCES AND PROPERTIES

International Journal of Web & Semantic Technology (IJWesT) Vol.9, No.3, July 2018
DOI : 10.5121/ijwest.2018.9302 13
A BOOTSTRAPPING METHOD FOR AUTOMATIC
CONSTRUCTING OF THE WEB ONTOLOGY
INSTANCES AND PROPERTIES
Song-il CHA 1
and Myong-jin HAN 2
1
Information Science College, University of Sciences, Pyongyang, DPR Korea
2
Department of Computer Systems, Pyongsong College of Technology,
Pyongsong, DPR Korea
ABSTRACT
With the phenomenal growth of the Web resources, to construct ontologies by using existing resources
structured in the Web has gotten more and more attention. Previous studies for constructing ontologies
from the Web have not carefully considered all the semantic features of the Web documents. Hereby it is
difficult to correctly construct ontology elements from the Web documents that are increasing daily. The
machine learning methods play an important role in automatic constructing of the Web ontology.
Bootstrapping technique is a semi-supervised learning method that can automatically generate many terms
from the few seed terms entered by human. This paper proposes bootstrapping method that can
automatically construct instances and data type properties of the Web ontology, taking proper noun as
semantic core element of the Web table. Experimental result shows that proposed method can rapidly and
effectually construct instances and its properties of the Web ontology.
KEYWORDS
Ontology, Bootstrapping, Instance, Property, Seed, Web tables.
1. INTRODUCTION
Today, the Semantic Web comprises techniques that promise to dramatically improve the current
WWW and its use. With the emergence of the Semantic Web and the growing number of
heterogeneous data sources, the benefits of ontologies are becoming widely accepted [1].
Accordingly, researches for using ontologies on the Web have become active [2]. Currently, in
many cases, most Web ontologies are simpler than previous ontologies used in the design and
diagnosis. The Web ontologies define terms used as data (metadata) for explaining things of a
special domain. Manually setting ontology up would entail a lot of time, not to mention that there
are only a handful of experts available. For this reason, researchers are paying attention to
automatic transformation of the Web resources in the areas into ontologies [3, 4]. In order to
provide the necessary means to widely apply ontologies to various fields there are today many
proposals for using ontology learning and machine learning, and until now the study on domain
ontology learning has been flourishing.
Rupasingha et al. [5] proposed a Web service clustering method through calculating the semantic
similarity of Web services using ontology learning method. El Asikri et al. [6] described the
commonalities of the areas, such as the semantic web and data mining, in order to resolve
problem of extracting useful and shared knowledge, as well as solve the problem of the
interoperability between Web systems by using the ontology learning from Web content.
Rupasingha et al. [7] presented a method for calculating Web service similarity using both
ontology learning and machine learning that uses a support vector machine for similarity
calculation in generated ontology instead of edge count base method. Kumara et al. [8] proposed
clustering approach that considers the complex data type as well as the simple type in measuring

14
the service similarity. This approach used hybrid term similarity method which proposed in their
previous work to measure the similarity. Song et al. [9] reviewed the related concepts and
methods of ontology construction and extension, proposed an automatic ontology extension
method based on supervised learning and text clustering. This method used the K-means
clustering algorithm to separate the domain knowledge, and to guide the creation of training set
for Naive Bayes classifier.
Jupp et al. [10] presented Webulous that is an application suite for supporting ontology creation
by design patterns, and provided simple mechanisms for the addition of new content in order to
reduce the overall cost and effort required to develop ontologies. Peng et al. [11] proposed a
method which can learn a heavy-weighted medical ontology based on medical glossaries and
Web resources, in order to deal with heterogeneous knowledge in the medical field. Wei et al.
[12] presented a semi-automatic construction method for agricultural professional ontology from
web resources. For semi-structured web pages, this method automatically extracted and stored
structured data through a program, built pattern mapping between relational database and
ontology through human-computer interaction, and automatically generated a preliminary
ontology, finally completed checking and refining by domain experts
The Web is an enormous resource of information contained in billions of individual pages. Most
information resource on the Web is presented in the form of semi-structured or unstructured
documents, encoded as a mixture of loosely structured natural language text and template units.
Yu et al. [13] proposed a modified hierarchical concepts tree building method by applying
pruning algorithm on the graph. They used the clue words to product queries containing
hierarchical relation to get corpus rich in concepts hierarchical relation through the search engine
from Web. Vasilateanu et al. [14] proposed a semantic search engine for relevant documents in an
enterprise, based on automatic generated domain ontologies, with observing on the component for
ontology learning and population. Manvi et al. [15] focused on generating domain specific
ontology for retrieving hidden web contents. In this paper a knowledge base used in automatically
filling up search interfaces for retrieving hidden web data.
The Web tables are used mainly for structuring information, and they are the strongest means of
presenting structured information. The Table structures represent relations between data in the
table. Therefore, ontologies can be easily extracted from a table by using structural features of the
table [16]. However, understanding of table contents requires table structure comprehension and
semantic interpretation, which exceed the complexity of corresponding linguistic tasks. Previous
studies for constructing domain ontologies from the Web table are centralized to interpret table
structure.
The comparatively comprehensive and complete model for the analysis and transformation of the
tables is Hurst’s [17]. This model analyzes the tables along graphical, physical, structural,
functional, and semantic dimensions. Jung et al. [18] suggested a method for extracting table-
schemata based on table structure and heuristics. Using this method, a table is converted into a
table-schema and a triple. Chen et al. [19] employ heuristic rules to filter out non-genuine tables
from their test set and make assumptions about cell content similarity for the table recognition
and interpretation. Wang et al. [20] proposed a machine learning based approach to classify given
table entity as either genuine or non genuine. Pivk et al. [21] focused on understanding table-like
structures only due to their structural dimension and transforming the most relevant table types
into F-logic frames. Tijerino et al. [22] described the automatic generation of ontologies from the
normalized tables, which is a structure they got after normalizing table-equivalent data. Tanaka et
al. [23] proposed a method for extracting relations based on interpretations given by humans, in
order to interpret structures of each tables correctly. This method is easy to apply to tables in
various domains because it uses interpretations given by humans and generalized table structures
instead of a domain-specific knowledge base. Jung et al. [24] detected that, generally, a table

15
provides a semantic core element in a HEAD, and proposed a method for automatically extracting
domain ontology using heuristics for extracting table schemata based on semantic core element.
As specified above, most research endeavors to interpret the table by using structural
characteristics of the table. But, most Web tables are designed by humans, thus, it has a certain
limit to automatically interpret table using only structural information of the table. Though Jung
et al. [24] proposed heuristics for detecting semantic characteristics based on the location of the
table cells, they did not mention which becomes semantic core element.
Through the observation about semantic features of the table, it is found that if there are proper
nouns on the table, then they can become a semantic core element. So this paper focuses on the
proper noun extraction method, which is a pre-requirement for interpretation of table structure
based on proper nouns. That is, this paper proposes an automatic extraction method of the
instance composed of proper nouns. Bootstrapping-based semi-supervised learning method aims
to rapidly and accurately obtain brief domain ontology from the table cells consisting of proper
nouns [25, 26].
Bootstrapping method, which aims at automatically generating instances and their relations in a
given domain, is a promising technique for ontology creation. H. Davulcu et al. [27] proposed the
OntoMiner system which offers automated techniques for creating ontologies based on a small
collection of relevant Web sites. The work presented an approach for bootstrapping and
populating large, rich, and up-to-date domain ontologies that organize the most relevant concepts,
their relationships, and instances (which correspond to members of concepts). W. S. Wu et al.
[28] presented the DeepMiner system which learns domain ontologies from the source Web sites.
Given a set of sources in a domain of interest, DeepMiner first learns a base ontology from their
query interfaces. It then grows the current ontology by probing the sources and discovering
additional concepts and instances from the data pages retrieved from the sources. A. Segev et al.
[29] proposed an ontology bootstrapping process for web services. The proposed ontology
bootstrapping process integrates the results of two methods, namely Term Frequency/Inverse
Document Frequency (TF/IDF) and web context generation, and applies a method to validate the
concepts using the service free text descriptor, thereby offering a more accurate definition of
ontologies. F. Keshtkar et al. [30] presented a novel semantic bootstrapping framework that uses
semantic information of patterns and flexible match method. The work considerably enhance
based on iterative bootstrapping model which generally implies semantic drift or low recall
problem.
Through the experimental observation about semantic features of the Web tables, it is found that
if there are proper nouns on the table, then they can become a semantic core element. The author
proposes algorithms to automatically construct all instances or properties belonging to a given
class, taking few terms belonging to class composed of proper noun as the seed. A bootstrapping
method is proposed to construct ontologies with the instances and properties. The paper focuses
on the extracting instances and properties based on interpreting the table contents by using
structural and semantic characteristics of the table.
The paper is structured as follows: Section 2 presents an automatic generation method of the
instance belonging to given class from Web tables; Section 3 describes an automatic property
generation method based on proper noun extraction; Section 4 evaluates our method according to
the experimental result; Finally, Section 5 provides conclusions to our work.
2. AUTOMATIC INSTANCE CONSTRUCTION
The knowledge on the domain terminology is required in order to manually construct ontology
about products that are increasing daily such as CD / DVD, software etc., are not known already.

16
Thus, it is highly regarded to automatically build an ontology based on Internet resources of a
given area. But, in order to automatically generate semantically correct ontology, certainly, must
to be based on some clues. If to extract instances from the Web table does not depend on any clue
and uses only structural information of the table, it is difficult to determine quality of the
extracted instance. Through experimental interpreting of Web table structure, it is found that if
there are proper nouns on the table, then they can be as the semantic core element, that is, as the
instance. The semantic core element is a head cell that plays the role of the ‘pivot’ in
understanding table structure [24]. Once proper nouns are extracted from a table, table structures
can more accurately interpret focusing on proper nouns that are semantic core element of the
table. Therefore, this section describes method generating automatically other instances in the
table, taking some proper nouns such as already familiar product name as a clue. That is, this
method is an approach which extracts the rest instances in the same class by using the proper
noun extraction method, having some instances given by the user as a seed.
2.1. PROPER NOUN EXTRACTION MODEL BASED ON BOOTSTRAPPING
In this section, the proper noun extraction method employs for automatic extraction of the
instance. This method is an approach, that if there is the row or the column composed by the
proper noun in the Web table, then, considering it as instance, the proper nouns are extracted.
That is, the proper noun extraction means to extract the other terms guessed belong to the class
which the entered word (proper noun) belongs to. This section presents the proper noun extraction
method using bootstrapping.
Bootstrapping is shown as follows: firstly, generate a pattern from the document in accordance
with a small amount of seed terms, then using this pattern again extract other words from the
document, and lastly using the extracted terms create another word. A large amount of terms can
be extracted from a small amount of seed terms by a repeat of this process. To begin with we
define the fundamental notions.
A proper noun p is a noun that is the name of a specific individual, place, or object. For example,
personal name, country name, denomination name, organization name, and so on. A proper noun
set P is a set of proper nouns, i.e. p ∈P.
A seed term s is a proper noun specified artificially before proceeding with automatic learning. A
seed term set S is a set of the seed terms, i.e. s ∈S.
From the above definition, we can know that s∈P and S⊂P. For example, Table 1 shows the seed
term set which belongs to each class. Herein, first column is class name and second column is the
proper nouns which belong to the class. For example, desktop is a name of desktop class, and
Acer, Asus and Compaq are the instance (manufacturer brand) which belongs to the desktop
class.
A domain table set T is a set of genuine tables which is chosen in the Web tables of a given
domain, collected using search engine such as Google or Yahoo. In this paper, in order to obtain
genuine tables of given domain, the algorithm proposed in the previous study is used [18].
Withal, search keys for obtaining of the domain table set are a domain name (that is, a class
name) and the seed term set selected by user in the beginning.

17
Table 1. An example of the seed term belonging to a given class.
Desktop
Acer, Asus, Compaq, Dell, eMachines, Everex, Gateway, HASEE,
HP, Lenovo, Panasonic, Samsung, Sony, TCL, Toshiba, etc.
Digital Camera
Agfa, Canon, Casio, Contax, Epson, FujiFilm, HP, Kodak, Konica
Minolta, Kyocera, Leica, Nikon, Olympus, Panasonic, Pentax, Ricoh,
Samsung, Sanyo, Sigma, Sony, Toshiba, etc.
LCD TV
Akai, AOC, Axion, Benq, Casio, Dell, Diamond, Epson, Gateway,
GPX, Haier, Hewlett Packard, Hitachi, Honeywell, Hyundai, JVC,
Konka, LG, Mitsubishi, NEC, Nikon, Panasonic, Philips, Samsung,
Sanyo-Fisher, Sharp, Skayworth, Sony, Toshiba, etc.
Publishing
Company
Pearson, Reed Elsevier, ThomsonReuters, Wolters Kluwer,
Bertelsmann, Hachette Livre, McGraw-Hill Education, Grupo
Planeta, De Agostini Editore, Scholastic, Houghton Mifflin Harcourt,
Holtzbrinck, Cengage Learning, Wiley, Informa, HarperCollins,
Shogakukan, Shueisha, Kodansha, Springer Science and Business
Media, etc.
Chinese Province
Anhui, Shandong, Guangdong, Jiangsu, Hunan, Hubei, Liaoning,
Shanxi, Inner Mongolia, Tianjin, Ningxia, etc.
Country
Brazil, Canada, China, France, Germany, India, Indonesia, Italy,
Poland, Spain, Thailand, UK, Ukraine, etc.
The proper noun extraction model based on bootstrapping is shown below.
This model extracts automatically new proper nouns based on the few seed terms from the
domain table set. In this work, the model is named as IC-Model (Instance Construction-Model).
Figure 1 shows IC-Model. In the model, the dotted line arrow denotes a process for extracting a
pattern that contain the initial seed term. The extraction of the pattern from the domain table set
only needs to be determined once.
Figure 1. IC-Model
The input of the IC-Model is the domain table set T and the initial seed term set S, the output of
this model is the proper noun set. A quality of the seed term set greatly affects the accuracy of the
extracted proper nouns. Therefore, users must select more obvious and important terms for the
initial seed term set. In addition, through experimental study, it is confirmed that can increase the
accuracy of the proper noun extraction in the case of that Ns (the number of the seed terms) is
Initial seed term
Seed list
Domain table set
Table selection
& Extraction of
candidate term
Selection &
Evaluation
Proper noun set
Seed candidate set
Finish
Pattern enactment
Candidate
term set

18
three or more. The number of the seed list used in this model is always Ns. A threshold value N of
the seed list is the number of an old seed term among the seed terms entered in the each loop of
bootstrapping. In the beginning, N=Ns. Constitute the threshold value N of the seed list as greater
one than two. This means that it must contain at least two initial seed term in the seed terms used
at the every loop. That is, can effectively extract terms when N≥2. But if not, can’t guarantee
accuracy of the extracted terms. The seed list is possible combinations of initial seed terms
composed of greater one than two.
At the first loop period of bootstrapping, this model extracts terms taking Ns initial seed terms.
But, at the second loop period, it extracts possible candidate terms taking Ns-1 initial seed terms
and a new seed candidate. Likewise, at the third loop period, it extracts possible candidate terms
taking Ns-2 initial seed terms and two new seed candidates. The loop period finishes when N<2.
In the pattern production process, are instituted the TABLE tags that wrap each of the seed term
from the table. In the table selection and extraction process of the candidate term, model selects
tables that contain the seed term pattern from the domain table set and extract all rest cells which
appears with the seed terms in same row or same column of the table. In the selection and
evaluation process, model firstly, only adds the obtained candidate terms which do not overlap to
the proper noun set. Secondly, selects the seed candidates for next loop among the proper noun
set and add to a seed candidate set. Finally, evaluates the threshold value N of the seed list, and
finishes when N<2, if not repeats the loop period.
2.2. INSTANCE CONSTRUCTION ALGORITHM
This subsection proposes the detailed algorithm for acquiring the instances based on IC-Model.
The input of this algorithm is the domain table set T and three initial seeds s1, s2, s3.The output of
this algorithm is the acquired instance set P. Our goal is to rapidly and accurately generate the
domain ontology. Therefore, three seed terms are used, i.e. Ns=3. Algorithm 1 shows the instance
acquisition algorithm.
Algorithm 1. Instance construction algorithm
1: Input S,T
2: Choose s1,s2,s3∈S
3: X←{s1,s2,s3}, X0←{s1,s2,s3}, P←s1 ∪s2 ∪s3, C0←Ø
4: Create pattern which wraps S from T
5: Construct candidate term set C from the table in T by using pattern that
wraps X₀
6: //Evaluation and selection
7: // Evaluation
8: From C remove terms that don’t correspond to evaluation condition
9: From C remove all c that c∈C ∧ c∈P , it’s result also remains to C
10: P←P∪C
11: if isn’t │X│≥ 2 then
12: output P, and finish
13: end if
14: // Selection
15: if processed all elements in C0 then
16: from S select new element pair and replace X by it;
17: If C0=Ø then
18: C0←C
19: end if
20: end if
21: Select new seed candidate set X₀ from X and C0

19
22: go to 5
In algorithm 1, X is an element pair which consisted by arbitrary combination of elements in the
seed term set S, for example {s1, s2, s3},{s1, s2},{s1, s3},{s2, s3},{ s1},{ s2},{ s3}, and so on. C0 is a
candidate term set acquired when the seed term is {s1, s2, s3}. C is a candidate term set which is
consisted at the rest every loop period. X0 is a seed term set that used as input of every loop period.
In the beginning, the algorithm enters the bootstrapping process, taking {s1, s2, s3} which consist
of three terms. Then add the new acquired terms to the candidate term set C. In the second loop,
takes two old (before used) terms and a newly acquired term as a seed term. Because it can’t be
considered that the newly acquired terms are certainly right instance. This will prevent that the
accuracy is lower, and take another new elements as a seed. Fundamental steps of algorithm are
shown as follows:
Pattern enactment. In order to produce pattern, it must be found TABLE tags which wraps each
of the seeds from arbitrary table of the domain table set. For example, when {Hunan, Hubei,
Shanxi} is a seed term, it must be taken <td>Hunan</td>, <td>Hubei</td>, <td>Shanxi</td>
which is wrapped by TABLE tag as a pattern. That is, can be constituted the <td>term</td> used
for defining the table cells as a pattern.
Every table’s creator is using TABLE tag of different description form, but this algorithm
employs only <td>…</td> tag which denotes the table cell for the purpose of pattern enactment.
Withal, for a brief descriptive purpose, is considered only the case of nonuse of <td> tag attribute
such as ALIGN, VLIGN, EIDTH, HIGHT, BACKGROUND, BTCOLOR, and so on. Figure 2
shows the example of the Web table that contains above three seeds.
Figure 2. An example of a table contained the proper nouns
Acquiring the candidate term. In order to construct the candidate term set, we assume as
follows: If certain term appears with the proper noun in same row or same column of a table, then
this term also is a proper noun. Therefore, are taken the row or column of table as object of
candidate term extraction. For example, the second column of Figure 3 contains the proper noun
“Wuhan” that is known already. Withal, another terms on this column are proper nouns belonging
to “Chinese cities” class. Before extracting the candidate term, must be determined table reading
orientation, i.e. row wise or column wise [19]. For example, if the first row of the table consists of
the attribute cells, and the others are value cells, then this table is the column wise. If the table is
row wise, then take each cell of the row in which exists proper noun as candidate term, and if the
<table border>
<tr>
<th>Chinese provinces</th>
<th><p>Total population</p>
<p>(million)</th>
<th><p>Populstion density</p>
<p>(people/Km²)</th>
</tr>
<tr>
<td>Shandong</td>
<td>86.2</td>
<td>562</td>
</tr>
<tr>
<td>Guangdong</td>
<td>72.6</td>
<td>343</td>
</tr>
<tr>
<td>Jiangsu</td>
<td>68.0</td>
<td>663</td>
</tr>
<tr>
<td>Shanxi</td>
<td>29.6</td>
<td>189</td>
</tr>
<tr>
<td>Inner Mongolia</td>
<td>22.0</td>
<td>19</td>
</tr>
<tr>
<td>Tianjin</td>
<td>8.9</td>
<td>787</td>
</tr>
<tr>
<td>Ningxia</td>
<td>4.9</td>
<td>95</td>
</tr>
</table>
<tr>
<td>Hunan</td>
<td>62.5</td>
<td>306</td>
</tr>
<tr>
<td>Anhui</td>
<td>58.7</td>
<td>420</td>
</tr>
<tr>
<td>Hubei</td>
<td>55.9</td>
<td>301</td>
</tr>
<tr>
<td>Liaoning</td>
<td>39.8</td>
<td>273</td>
</tr>

20
table is column wise, then take each cell of the column in which exists proper noun as candidate
term.
Evaluation. In order to extract more accurate candidate term, must be resolved table HEAD.
Withal, must be determined candidate term extraction range in the row (or column) in which
exists proper noun, that is, must be resolved value region corresponding to the table HEAD.
Figure 4 is an example in which exists two different HEADs (1st
row and 13rd
row). For the
purpose of extracting candidate term taking above proper noun as a seed, must be restricted range
of value region, i.e. there is candidate extraction object between 2nd
row and 12nd
row. But if not,
it has possibility to extract irrelevant terms. In this example, “Mexico”, “Germany”, and so on are
irrelevant term. Therefore, in this step it is evaluated whether each term c of the candidate term
set C is a term belonging to a given class, and choose only fit terms. Withal, in the evaluation step,
evaluate whether newly extracted proper noun exists in the proper noun set P, and only add
proper noun which does not overlap to P.
Figure 3. An example ofa table which proper nouns are located in the coloum wise

21
Figure 4. An example of the table with two head
3. AUTOMATIC PROPERTY CONSTRUCTION
3.1. OBSERVATION ON THE LOCATION OF THE PROPERTIES
It is difficult to extract inclusively, if property also as well as instance does not depend on any
clue. Therefore, in this section is proposed a method for automatically constructing properties
belonging to the class using proper noun extraction approach based on bootstrapping.
In order to extract the property from the Web table, firstly, based on the proper noun extraction
method mentioned above, extract the instance belonging to the class, and take a set of these
instances as P. Then, determine three initial seed property belonging to the class, and denote them
as a1, a2, a3. Denote a set of the properties as A, i.e. a1, a2, a3 ∈A. For example, in the first row of
the table on Figure 5, attributes such as Type, Standard, Resolution, and so on, are property
belonging to a camera class, and cells (but the first cell is excepted) of the first column are
instances of a camera Model class.
l Type Standard Resolution Sensor
Frame/
Line
Rate
Interface
Light
Spectrum
CV-
S3200N
1-CCD Color
Interlaced
NTSC 758 X 486 1/2” 30 Analog Visible
CV-
S300P
1-CCD Color
Interlaced
PAL 737 X 575 1/2” 25 Analog Visible
CV-
S3300N
1-CCD Color
Interlaced
NTSC 758 X 486 1/3” 30 Analog Visible
CV-
S3300P
1-CCD Color
Interlaced
PAL 737 X 575 1/3” 25 Analog Visible
CB-
040MC
L
1-CCD Color
Progressive
SVGA 776 X 582 1/2” 60 Mini-CL Visible
CB-
040GE
1-CCD Color
Progressive
SVGA 782 X 582 1/2” 60
GigE
Vision
Visible
CB-
080GE
1-CCD Color
Progressive
XGA 782 X 582 1/3” 30
GigE
Vision
Visible
BB-
500CL
1-CCD Color
Progressive
QSXGA 1024 X 768 2/3” 15
Camera
Link
Visible
Figure 5. An example of the table that contains the several attributes

22
As shown in the figure, when the table is column wise, generally, the properties are constructed
by the cells which lie at the first row of the table. When extracting the property using the proper
noun extraction method, the seed term contains three properties and one instance which is
selected from instances collected already. In each product there are special and common attributes.
For example, not only digital camera but also computer has “resolution” attribute, that is, this is
the common attribute. When are taken only three properties as the seed property, it has possibility
which can extract even properties belonging to other class. Thus, is added one instance to the seed
term set S, in order to extract rightly the property belonging to given class.
3.2. PROPERTY CONSTRUCTION ALGORITHM BASED ON THE IC-MODEL
The property construction algorithm based on the IC-Model similar to the instance construction
algorithm (algorithm 1). The properties discussed in this subsection are Owl:DatatypeProperty
[31, 32].
Algorithm 2 shows the property construction algorithm.
Algorithm 2. Property construction algorithm
1: Input A,P,T
2: R←s1 ∪s2 ∪s3
3: Create pattern which wraps A and P from T
4: X←{s1,s2,s3}; Take arbitrary p ∈P, X0←{p,s1,s2, s3}, C0←Ø
5: Construct candidate property set C from the table in T by using pattern
that wraps X₀
6: // Evaluation and selection
7: // Evaluation
8: From C remove properties that isn’t DatatypeProperty
9: From C remove all c that c∈C ∧ c∈R , it’s result also remains to C
10: R←R ∪ C
11: if isn’t │X│≥ 2 then go to 21
12: // Selection
13: if processed all elements in C0 then
14: from A select new element pair and replace X by it
15: if C0=Ø then
16: C0←C
17: end if
18: end if
19: Select new seed property set X0 from P, X, C0
20: go to 5
21: if processed all elements in P then
22: output R, and finish
23: else go to 4
24: end if
In algorithm 2, X is a element pair which consisted by arbitrary combination of elements in the
seed property set A. C is a candidate property set which consisted at the every loop period. R is a
property set. X0 is a seed term set that it used as input of every loop period.
In the beginning, the algorithm enters the bootstrapping process, taking {s1, s2, s3} which consist
of three properties and an instance of P. In the second loop, the algorithm takes two old (before
used) properties, a newly extracted property, and an instance of P as a seed term. If all elements
of property set were processed, then replace the instance by a new element and repeat process. In
the subsection below, stepwise instantiate major steps of algorithm.

23
Pattern enactment. In order to produce pattern, must be looked for TABLE tags which wrap
each of the properties from arbitrary table of the domain table set. The HEAD tag of the table is
often denoted by <td>term</td>, <td><b>term</b></td>, <td tag attribute>term</td>,
<th>term</th>, <th tag attribute >term</th>, and so on. At the HEAD of the table, tag attribute is
used for the purpose of standing out. In order to denote HEAD cell, every table’s creator uses
table tag of different description form, but for a brief descriptive purpose, is employed only <th
tag attribute >term</th> pattern (see Figure 2).
Acquiring the candidate property. In order to construct the candidate property set, we assume
as follows: If certain term appears with the seed property in HEAD row or HEAD column of a
table, then this term also is a property. Therefore, are taken the HEAD row or HEAD column of
the table as the object of candidate property acquisition.
Evaluation. In this step, firstly, remove properties that are not DatatypeProperty from the
candidate property set. That is, if “total” or “other” appears at the HEAD of the table, then
remove it from the property candidate. Denote the term set which isn’t DatatypeProperty by
“NonwordList”, i.e. it is needed c∉NonwordList. Then, from HEAD row cell remove the element
that <th> tag’s attribute is COLSPAN. That HEAD element generally has hierarchical structure,
thus this is not DatatypeProperty that must be obtained. Withal, in this step, evaluate whether
newly acquired property exists in the property set R, and only add property which does not
overlap to R. In the case of property acquisition, also institute threshold value N of seed property
set A as greater one than two.
4. EMPIRICAL EVALUATION
4.1. EMPIRICAL EVALUATION OF THE ALGORITHM 1
Automatic instance construction method (algorithm 1) mentioned above is suited to obtain
instances, which are severely changed, such as new products. In this experiment, in order to
correctly construct the seed term set, we selected twenty publishing companies based on revenue
rank of the world’s largest publishing companies in 2008 from the Web pages. Figure 6 shows a
Web table in which second column contains the instances belonging to a “publishing company”
class, i.e. the seed term set is consisted of cells in the second column of this table.
Firstly, in order to obtain an experimental data, must be chosen initial seed belonging to
“publishing company” class. If number of the initial seed is many, then accuracy is enhanced.
Through experimental observations, it is confirmed that when take two initial seed terms, a large
amount of irrelevant terms are extracted and the accuracy was decreased. Hence, are selected
already known three initial seeds, i.e. {McGraw-Hill Education, Wiley, Reed Elsevier}. Next, in
order to obtain the domain table set, the method collects arbitrary two hundred Web tables among
around twenty thousand Web pages of the “publishing” domain acquired using search key
“publishing company”, “McGraw-Hill Education”, “Wiley” and “Reed Elsevier” of Google
search engine. After that, according to the algorithm 1 proposed in the section 2, the method
acquires the instances belonging to the “publishing company” class. In this case, must be taken
<td> McGraw-Hill Education </td>, <td> Wiley </td>, <td> Reed Elsevier </td> which
iswrapped by <td> tag and </td> tag used for defining the table cells as a pattern.
Previous researches endeavored to interpret the tables using structural characteristic, hence, it is
difficult to experimentally compare with our method. Withal, it is troublesome to rightly measure
recall and precision of the experimental results obtained by our method. Therefore, in this
experiment, we evaluated that the proposed instance construction algorithm how to effectively

24
increase the instances about every time operation. For the above purpose, we compared with
twenty instances extracted from Figure 6, i.e. watch obtaining process of correct twenty instances.
Figure 6. An example of the table of the publishing domain extracted from the Web page
In order to evaluate the experimental results, were used the recall and precision metrics as
follows:
The recall and precision metrics can be interpreted as follows:
Firstly, we discuss the recall. In the case of the simple input, the system registers singly the
correct instance depending on the input sequence. Therefore, the recall is proportional to input
time, in the case of seventeenth input only arrive at 100 percent. In the case of our proper noun
extraction, is extracted a large amount of candidate instances at first time, at second time the
recall arrives at 64 percent. However, at next time the candidate term is overlapped, and rising
speed is slow, at sixth time it arrives at 100 percent
Figure 7. The comparison of the experimental results
RANK COPANY NAME PARENT COMPANY NATIONALITY REVENUES
1 Pearson Pearson UK 5,044
2 Reed Elsevier Reed Elsevier UK/NL/US 4,586
3 ThomsonReuters
The Woodbridge
Company Ltd
Canada 3,485
4 Wolters Kluwer Wolters Kluwer NL 3,374
5 Bertelsmann Bertelsmann Germany 2,980
6 Hachette Livre Lagardere France 2,159
7
McGraw-Hill
Education
McGraw-Hill US 1,794
8 Grupo Planeta Grupo Planeta Spain 1,760
9 De Agostini Editore Gruppo De Agostini Italy -
10 Scholastic Scholastic Corp US 1,499
11
Houghton Mifflin
Harcourt
Education Media and
Publishing Group
US/Cayman Islands 1,712
12 Holtzbrinck
Verlagsgruppe Georg
von Holtzbrinck
Germany -
13 Cengage Learning Apax Partners et al. UK 1,172
14 Wiley John Wiley & Sons US 1,139
15 Informa Informa UK 1,028
16 HarperCollina News Corp US/AUS 944
17 Shogakukan Shogakukan Japan 927
18 Shueisha Shueisha Japan 902
19 Kodansha Kodansha Japan 886
20
Springer Science
And Business Media
Cinven and Candover
UK/Germany/Italy/
France
880
Recall =
Number of acquired correct instance
Total number of correct instances (1)
Total number of acquired instances
Precision =
Number of acquired correct instance
(2)

25
Through this, it is confirmed that the proper noun extraction method can increase effectively the
instance from a small amount of user input.
Next, we discuss the precision. In the case of the simple input, the precision is 100 percent. But,
in the case of our instance extraction based on IC-Model, the precision decreases and at
seventeenth time almost arrives at 94 percent. A cause that the precision does not arrive at 100
percent is that was included the irrelevant terms.
4.2. EMPIRICAL EVALUATION OF THE ALGORITHM 2
The property construction method (algorithm 2) employs the instances which extracted by the
instance construction algorithm (algorithm 1). For an experiment, are selected products which
have comparatively many attributes such as mobile phone, digital camera, LCD TV, LCD
monitor and CRT monitor. Next, in order to obtain the domain table set, we collected around ten
thousand Web pages acquired using individually search key “mobile phone”, “digital camera”,
“LCD TV”, “LCD monitor” and “CRT monitor” of Google search engine. Table 2 shows the
representative attributes belonging to each product class.
In order to illustrate property acquisition process, the experiment takes the mobile phone as an
example. Firstly, the experiment selects three attributes belonging to the mobile phone as the seed
property, i.e. {dimensions, weight, display size}. Next, extract the instances belonging to mobile
phone using the algorithm 1. Then, according to the property construction algorithm (algorithm 2),
extract the properties taking {Nokia, dimensions, weight, display size} as the seed term. In this
seed term, Nokia is an instance of the “mobile phone” class.
Next, in order to obtain more correct table set, the experiment choose again the Web pages which
contain “Nokia”, “dimensions”, “weight” and “display size” from around two thousand Web
pages acquired by search key “Mobile phone”.
When {Nokia, dimensions, weight, display size} is a seed term, it can be taken <td>Nokia</td>
and <th>dimensions</th>, <th>weight</th>, <th>display size</th> which is wrapped by HEAD
tag as a pattern. Then, according to the algorithm 2 proposed in the section 3, the experiment
acquires the properties belonging to the “Mobile phone” class.
Likewise, can be extracted properties belonging to other class. Table 3 shows total experimental
result.
Table 2. An example of the correct attribute set.

26
Table 3. Recall and precision for acquired properties.
As shown in the table 3, the average recall of our method arrives at 93 percent. Through this, it is
confirmed that proposed method can obtain almost the properties belonging to a given class.
Withal, the average precision arrives at 90 percent. To extract the properties from the Web pages
which describe only actual attributes of the product is the most ideal. But, in this experiment, did
not choose from the Web pages applied like that.
5. CONCLUSIONS
This paper has presented an automatic construction method of instances and properties belonging
to a given class from the Web table resources based on bootstrapping. Our approach was able to
acquire fast a great amount of instances and properties using bootstrapping from a small amount
of seed terms entered by human. When interpreting and straightening out table structure
according to ontology schemata, to determine accurately class, instance and property gives a
decisive impact on the use of ontologies in the future. Though this paper briefly extracted only
instances composed of proper nouns and DatatypeProperty, we consider that this approach offer a
good basis not for construction of arbitrary domain ontology fit with our intent, but for right
semantic interpretation of table structure. In our future work, a more precise method for
extracting ObjectProperty and other relationships will be developed based on extracted proper
nouns.
ACKNOWLEDGEMENTS
This work was supported in part by the science & technology development fund of University of
Sciences (100/425-63053-243/1). The authors are grateful to our colleagues at Centre for Natural
Science Research and Information Science College for their help in collecting the Web pages and
for discussing. They also would like to thank the anonymous reviewers for their careful reading
and in-depth criticisms and suggestions.
REFERENCES
[1] G. Antoniou, van F. Harmelen, (2008) A semantic web primer, (Second Edition). Cambridge, MA:
MIT Press.
[2] J. M. Serrano, (2012) Applied Ontology Engineering in Cloud Services, Networks and Management
Systems. Springer.
[3] F. Zhang, J. W. Cheng, Z. M. Ma, (2016) “A survey on fuzzy ontologies for the Semantic Web”.
Knowledge Engineering Review, Vol. 31, No 3,pp278-321.

27
[4] K. D. Mogotlane, J. V. Fonou-Dombeu, (2016) “Automatic conversion of relational databases into
ontologies: a comparative analysis of protégé plug-ins performances”. International Journal of Web &
Semantic Technology (IJWesT), Vol.7, No.3/4, pp21-40.
[5] R. A. H. M. Rupasingha, I. Paik, B. T. G. S. Kumara, T. H. A. S. Siriweera, (2016) “Domain-aware
Web Service Clustering based on Ontology Generation by Text Mining”. In Proc. of the 7th IEEE
Annual Information Technology, Electronics & Mobile Communication Conference.
[6] M. El Asikri, J. Laassiri, S. D.Krit, H. Chaib, (2016) “Contribution To Ontologies Building Using the
Semantic Web and Web Mining”. In Proc. of the International Conference on Engineering & Mis.
[7] R. A. H. M. Rupasingha, I. Paik, B. T. G. S. Kumara (2015) “Calculating Web Service Similarity
using Ontology Learning with Machine Learning”. In Proc. of the IEEE International Conference on
Computational Intelligence and Computing Research, , pp. 201-208.
[8] B. T. G. S. Kumara, I. Paik, K. R. C. Koswatte, W. H. Chen, (2014) “Ontology Learning with
Complex Data Type for Web Service Clustering”. In Proc. of the IEEE Symposium on Computational
Intelligence and Data Mining, pp. 129-136.
[9] Q. X. Song, J. Liu, X. F. Wang, J. Wang, (2014) “A Novel Automatic Ontology Construction Method
Based on Web Data”. In Proc. of the 10th International Conference on Intelligent Information Hiding
and Multimedia Signal Processing, pp. 762-765.
[10] S. Jupp, T. Burdett, D. Welter, S. Sarntivijai, H. Parkinson, J. Malone, (2016) “Webulous and the
Webulous Google Add-On - a web service and application for ontology building from templates”.
Journal of Biomedical Semantics, Vol. 7.
[11] J. Peng, Y. R. Du, Y. Chen, M. Zhao, B. Pei, (2015) “Medical ontology learning based on Web
resources”. In Proc. of the 12th Web Information System and Application Conference, pp. 116-119.
[12] Y. Y. Wei, R. J. Wang, Y. M. Hu, X. Wang, (2012) “From Web Resources to Agricultural Ontology:
a Method for Semi-Automatic Construction”. Journal of Integrative Agriculture, Vol. 11, No. 5, pp.
775-783.
[13] H. Yu, X. Q. Lv, L. P. Xu, (2015) “Use web resources to construct ontology concept hierarchy”. In
Proc. of the International Conference on Applied Science and Engineering Innovation, Vol. 12, pp.
1006-1011.
[14] A. Vasilateanu, N. Goga, E. A. Tanase, I. Marin, (2015) “Enterprise Domain Ontology Learning from
Web-Based Corpus”. In Proc. of the 6th International Conference on Computing, Communication and
Networking Technologies, pp. 112-117.
[15] Manvi, A. Dixit, K. K. Bhatia, B. Wadhwa, (2014) “Generating Domain Specific Ontology for
Retrieving Hidden Web Contents”. In Proc. of the International Conference on Information Systems
and Computer Networks, pp. 66-71.
[16] W.Gatterbauer, P. Bohunsky, M. Herzog, B. Krüpl, B. Pollak, (2007) “Towards Domain-Independent
Information Extraction from Web Tables”. In Proc. of 16th International World Wide Web
Conference (WWW2007), pp.71-80.
[17] M. Hurst, (1999) “Layout and language: Beyond simple text for information interaction - modelling
the table”. In Proc. of 2nd International Conference on Multimodal Interfaces, pp. 243–249.
[18] S.W. Jung, M.Y. Kang, H.C. Kwon, (2006) “Hybrid Approach to Extracting Information from Web-
Tables”. In Proc. of 21st International Conference on the Computer Processing of Oriental Languages
(ICCPOL2006), LNAI, vol. 4285, pp. 109-119.
[19] H.H. Chen, S.C. Tsai, J.H. Tsai, (2000) “Mining Tables from Large Scale HTML Texts”. In Proc. of
18th International Conference on Computational Linguistics (COLING2000), Saarbrucken, Germany,
pp.166-172.
[20] Y. Wang, J. Hu, (2002) “A Machine Learning Based Approach for Table Detection on the Web”: In
Proc. of 11th International World Wide Web Conference (WWW2002), pp. 7-11.

28
[21] A. Pivk, P. Cimiano, Y. Sure, (2004) “From Tables to Frames”. In Proc. of 3rd International Semantic
Web Conference (ISWC2004), LNCS, Hiroshima, Japan, Vol. 3298, pp.166-181.
[22] Y. A. Tijerino, D. W. Embley, D. W. Lonsdale, G. Nagy, (2003) “Ontology Generation From
Tables”. In Proc. of 4th International Conference on Web Information Systems Engineering
(WISE'03), Rome, Italy, pp.242-249.
[23] M. Tanaka, T. Ishida, (2006) “Ontology Extraction from Table on the Web”. In Proc. of 2006
International Symposium on Applications and the Internet (SAINT'06), pp.284-290.
[24] S.W. Jung, M.Y. Kang, H.C. Kwon, (2007) “Constructing Domain Ontology Using Structural and
Semantic Characteristics of Web-Table Head”. In Okuno, H.G., Ali, M. (eds.) IEA/AIE 2007, LNAI,
Vol. 4570, pp. 665-674.
[25] S. Abney, (2002) “Bootstrapping”. In Proc. of 40th Annual Meeting of the Association for
Computational Linguistics (ACL'02), Philadelphia, pp. 360-367.
[26] S. I. Cha, M. J. Han, (2015) “A Method to Extract Automatically Ontology Instances Using
Bootstrapping”. Bulletin of Academy of Sciences, the DPR Korea, Vol. 367, No. 1, pp.19-20, ISSN
1680-4223.(in Korean)
[27] H. Davulcu, S. Vadrevu, S. Nagarajan, I. V. Ramakrishnan, (2003) “OntoMiner: Bootstrapping and
populating ontologies from domain-specific Web sites”. IEEE Intelligent Systems, Vol. 18, No. 5,
pp.24-33.
[28] W. S. Wu, A. H. Doan, C. Yu, W. Y. Meng, (2006) “Bootstrapping domain ontology for semantic
web services from source web sites, Technologies for E-Services”. LNCS, Vol. 3811, pp.11-22.
[29] A. Segev, Q. Z.Sheng, (2012) “Bootstrapping Ontologies for Web Services”. IEEE Transactions on
Services Computing, Vol. 5, No. 1, pp.33-44.
[30] C. Y. Zhang, W. R. Xu, Z. Y. Ma, S. Gao, Q. Li, J. Guo, (2015) “Construction of semantic
bootstrapping models for relation extraction”. Knowledge-Based Systems, Vol .83.
[31] OWL Reference, (2003) “OWL Web Ontology Language Reference”. http://www.w3.org/ TR/owl-
ref/.
[32] OWL 2 Overview, (2012) “OWL 2 Web Ontology Language Document Overview (Second Edition)”.
http://www.w3.org/TR/owl2-overview/.

A BOOTSTRAPPING METHOD FOR AUTOMATIC CONSTRUCTING OF THE WEB ONTOLOGY INSTANCES AND PROPERTIES

More Related Content

What's hot (12)

Similar to A BOOTSTRAPPING METHOD FOR AUTOMATIC CONSTRUCTING OF THE WEB ONTOLOGY INSTANCES AND PROPERTIES (20)

Recently uploaded (20)

A BOOTSTRAPPING METHOD FOR AUTOMATIC CONSTRUCTING OF THE WEB ONTOLOGY INSTANCES AND PROPERTIES