Towards Semi-Automatic Approach of Building an Ontology: A Case

Study on Material Handling Data

Sepideh Sadat Sobhgol

1 a

, Mario Thron

1 b

and Giuliano Persico

2 c

ifak e.V. Magdeburg, Germany

Demag Cranes & Components GmbH, Germany

Keywords:

Ontology, Semantic Search, Keyword Extraction, Term Relation, Similarity Measurements.

Abstract:

InnoSale project aims to improve sales processes for complex industrial equipment and services using AI

technologies. The project addresses the challenges of time-consuming back-ofﬁce support and interpreting

customer requests using different vocabularies. As partners involved in the project, we are developing a semi-

automated approach to the creation of an ontology for the material handling domain by merging existing

terminology from leading companies in the industry. This ontology will serve as the basis for a semantic

search engine to improve the generation of quotations and the matching of customer requirements. Through

the use of historical data and advanced machine learning techniques, the search engine streamlines the sales

process, reducing manual effort and improving response times. The results showcases how the utilization of

machine learning and NLP techniques can aid in constructing an ontology in a semi-automatic fashion. The

study demonstrates the effectiveness of extracting terms, identifying synonyms, and uncovering various re-

lationships, contributing to the development of an ontology. These approaches offer potential for improving

the ontology construction process and enhancing semantic search capabilities, leading to more effective in-

formation retrieval. This position paper, being concise in nature, presents our initial ﬁndings and progress in

this endeavor. It’s important to note that, based on new sources of information and ongoing research in the

future, the results and conclusions may evolve or differ.

1 INTRODUCTION

The production of industrial goods that involves cre-

ating a variety of products by adding multiple options

to base products is often referred to as modular pro-

duction or modular manufacturing. It offers two ap-

proaches to product customisation: pre-existing prod-

uct options sourced from a catalogue, or custom op-

tions created for a unique case. Off-the-shelf options

allow for quick and efﬁcient customisation, while cus-

tom options offer the opportunity to create a truly

unique product tailored to a speciﬁc customer’s needs,

but require more time and resources to develop and

test.

Finding similar custom options in a manufac-

turer’s project history can be an efﬁcient approach

to creating new custom options. However, ﬁnding

these similarities in a project history spanning several

https://orcid.org/0000-0002-9746-3612

https://orcid.org/0009-0002-0648-9903

https://orcid.org/0009-0008-5050-473X

decades can be challenging and time consuming. One

of the goals of the InnoSale project was to establish

efﬁcient methods for identifying and exploiting pre-

existing custom options that are related to a new cus-

tomer request.

Our semantic search approach involves roughly

the following steps:

• Generate ontology (semi-automatically)

• Map custom option projects to ontology concepts

• Search those projects based on concept mappings

This article describes our advances with regard to

semi-automatically generating an ontology for a mod-

ular production domain, especially for the material

transportation domain. An example of such a custom

option is a heat shield for bearings or gears of a trans-

portation system when it is to be installed in a steel

plant and the heat is coming from a direction speciﬁc

to that customer.

In project documents, the technical terms com-

monly used by experts can differ signiﬁcantly from

the layman’s terms used in customer requests. This

248

Sobhgol, S., Thron, M. and Persico, G.

Towards Semi-Automatic Approach of Building an Ontology: A Case Study on Material Handling Data.

DOI: 10.5220/0012210000003598

In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 2: KEOD, pages 248-254

ISBN: 978-989-758-671-2; ISSN: 2184-3228

disparity necessitates the use of an ontology that

establishes synonym relationships between these

terms. By incorporating an ontology, semantic search

techniques can yield more accurate and relevant re-

sults than a simplistic word indexing approach.

The discipline of ontology is introduced by

philosophers such as Aristotle, but has been deﬁned

differently for use in computer science by several

authors. A summary is given in (Guarino et al.,

2009). The authors distinguish between the original

deﬁnition of an ontology in computer science as an

”explicit speciﬁcation of a conceptualization” (Gru-

ber, 1995) and an extended deﬁnition by later pub-

lications as a ”formal speciﬁcation of a shared con-

ceptualization” (Borst, 1997). For our purposes, the

term ”shared” is not important, since the terms used

in project documents and by customers of the mod-

ular product manufacturer are usually non-shareable

intellectual property of that company. The term ”for-

mal” implies the use of a formal exchange format

for ontologies such as OWL, the Web Ontology Lan-

guage (Baader et al., 2005; Motik et al., 2012), which

is not required for our non-shared ontology. The us-

ability of our semantic search approach depends a lot

on its performance and thus, we focus more on efﬁ-

cient custom data structures than on exchange of the

ontology. Therefore, we use the term ”ontology” in

its original, more general sense, as deﬁned in (Gru-

ber, 1995) and not in its more modern meaning.

There is no automated process for ontology mod-

eling. Terms, such as nouns or important/frequent

terms in the domain, are used to create the concept

space. However, manual effort from domain experts

is still necessary to form concepts and relationships

in the ontology. To reduce this manual effort, ma-

chine learning (ML) and natural language process-

ing (NLP) can be utilized, especially when Descrip-

tion Logic (DL)-based languages are not used. Differ-

ent layers in ontology learning and building are shown

in Fig 1. The upper layers are based on terminology

and terms with similar meanings, represented in the

lower layers. Generating terms and their synonyms

semi-automatically can help reduce the workload for

domain experts.

In recent years, semantic search engines have

gained attention due to their ability to provide more

comprehensive and accurate search results compared

to traditional word index-based search engines. There

are several approaches to semantic search, which are

based e.g. on word embeddings or on ontologies. In

InnoSale project, we explore an ontology-based se-

mantic search approach that can retrieve results when

synonyms or generalisations of the search terms are

present in the target documents. In this article, we

Figure 1: Ontology Learning Layer Cake (Cimiano et al.,

2009).

discuss an efﬁcient method for creating an ontology

from existing sources, such as controlled vocabular-

ies, that can be used for semantic search, which then

could reduce the gap between searching the speciﬁc

words in documents and keywords in query (Ramku-

mar and Poorna, 2014).

At ﬁrst, our ontology will be a data structure

which uniﬁes the terminologies of project partners in

material handling domain by bringing synonyms to-

gether to create a concept or ﬁnding a concept hier-

archy. Further-on, the uniﬁed ontology also needs to

be updated by new terms, which are extracted from

incoming inquiries and which cannot be found in

the original version of the uniﬁed ontology. There-

fore, Named Entity Recognition (NER) (Al-Moslmi

et al., 2020) can be used to extract the underlying key-

words from customer inquiries in terms of analyzing

text at the word or subword level. Discovering the en-

tities unveils the underlying structure of the data and

thus better serves as a step towards semi-automatic

approach of creating an ontology. For manual editing

the ontology, an Ontology Editor will be developed

which provides a graphical user interface for this task.

2 BACKGROUND

In this section, we brieﬂy establish key terms and con-

text to understand our study. We start with some ex-

planations of ontology-based semantic search and ba-

sic listing on different methods of synonym detection.

Semantic search is a document retrieval process

that goes beyond simply relying on word occur-

rences in documents. Instead, it leverages domain

knowledge, which can be represented through an

ontology—a formal speciﬁcation of concepts and

their relationships (Hotho et al., 2006). Ontology-

based semantic search uses ontologies to understand

the meaning of user queries and the content being

searched. This enhances the accuracy and relevance

of search results. In this approach, the search engine

Towards Semi-Automatic Approach of Building an Ontology: A Case Study on Material Handling Data

249

analyzes both the query and the data against the ontol-

ogy, enabling it to interpret the query’s intent and un-

cover semantic relationships between concepts. This

understanding of semantics allows the search engine

to provide more precise and contextually relevant

search results. Ontologies in semantic search pro-

vide more advanced capabilities than keyword-based

search. They consider related concepts, synonyms, hi-

erarchical relationships, and other semantic connec-

tions to deliver more accurate and comprehensive re-

sults. Ontology-based semantic search is particularly

advantageous in domains with intricate and special-

ized terminology, where comprehending the semantic

context plays a crucial role in retrieving pertinent in-

formation (Ding et al., 2004; Mangold, 2007). The

connection between documents and ontologies plays

a crucial role in semantic search approaches. There

are two main approaches: tight coupling and loose

coupling. In tight coupling, documents explicitly refer

to concepts in the ontology, making it easier to resolve

homonymies. However, it requires signiﬁcant effort in

annotating documents with semantic information. On

the other hand, in loose coupling, documents are not

bound to a speciﬁc ontology, which presents the chal-

lenge of selecting the appropriate ontology. While

loose coupling provides ﬂexibility, it has limitations

in terms of semantic resolution, especially in scenar-

ios like the World Wide Web. Ontology-based se-

mantic search engines utilize ontologies, comprising

concepts, properties, constraints and axioms. Stan-

dard properties including synonym-of, hypernym-

of, meronym-of, instance-of, negation-of are used to

capture relationships in semantic search, enhancing

capabilities but introducing dependencies on ontology

structure (Hotho et al., 2006).

In order to enhance the ontology’s richness

by incorporating synonym relationships between

terms, one common technique is based on linguistic

resources, such as dictionaries and thesauri, which

provide explicit synonyms for a given word. An-

other approach involves utilizing corpus-based meth-

ods, where large collections of text are analyzed to

identify co-occurring words or patterns that indi-

cate synonymy. Additionally, distributional similarity

methods leverage word embeddings or vector repre-

sentations to measure the semantic similarity between

words and identify synonyms. Machine learning tech-

niques, including supervised and unsupervised algo-

rithms, have also been employed for synonym and

relation extraction by training models on annotated

datasets or using clustering algorithms to group sim-

ilar words (Zelenko et al., 2003; Nguyen and Grish-

man, 2015; Han et al., 2020; Mohammed, 2020). Fi-

nally, hybrid approaches combining multiple methods

have shown promise in achieving more accurate syn-

onym detection results (Blondel and Senellart, 2002;

Wang and Hirst, 2009; Yıldız et al., 2014).

3 EXISTING TERMINOLOGIES

The considered project partners already maintained

and still maintain different terminologies, which de-

ﬁne the vocabulary to be used for naming of products

and parts in their projects. The terminologies shall

be uniﬁed into a single terminology. It will cover a

broader set of terms used by the technical experts than

one of the existing terminologies. To ensure data pri-

vacy rights, we are unable to upload our terminol-

ogy. However, we will describe the structure of those

ﬁles here instead. Table 1 and Table 2 present an

Excel ﬁle that illustrates the structure of one of the

current terminologies. Fig 2 illustrates the structure

of the Acrolinx Database. These terminologies in-

clude translations in various languages. The Acrolinx

database may also include synonyms for terms. In

Acrolinx, terms that are synonyms in different lan-

guages are assigned the same entry ID.

Table 1: Standard Terms.a.

DE EN ES

Abdeckband Masking tape Banda protectora

Abdeckblech Cover plate Placa protectora

Abdeckblech Ger

ateseite Cover plate equipm.side Chapa prot. lad aparell.

Table 2: Standard Terms.b.

FR IT CS

Bande de protection Nastro coprente Zakr

yvac

ı p

aska

ole de protection Lamiera copertura Kryc

ı plech

ole protect. c. appareil Lamiera copert.apparecchi Kryc

ı plech strany stroje

4 STEPS TOWARDS

SEMI-AUTOMATIC CREATION

OF AN ONTOLOGY

This section outlines the steps involved in creating an

ontology. The creation of the ontology involves the

following steps:

• Importing existing terminologies: The vocabulary

used by the manufacturers is deﬁned in terminolo-

gies which is in a format of an Excel ﬁle and

Acrolinx database for this purpose. These ter-

minologies include translations in different lan-

guages, and efforts should be made to unify them.

• Creating an uniﬁed terminology: The uniﬁed ter-

minology incorporates term abstractions to enable

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

250

Figure 2: Acrolinx.

the identiﬁcation of product variants, thus forming

an ontology.

• Regular updates based on incoming in-

quiries: Manufacturers receive inquiry emails

from customers who may use different vocabulary

than what was used in previous projects. The on-

tology should be updated with new words, which

may need to be manually linked as synonym

terms or term abstractions.

The following section provides further details on

this approach.

4.1 Unifying Terminologies

As mentioned in Section 1 and depicted in Fig 1, in

the lower layer of the ontology cake, synonyms can be

connected to form a concept. Subsequently, the con-

cept hierarchy can be explored. Fig 3 provides us with

an understanding of how we can unify the terminolo-

gies to create our new term structure. Terms and their

Figure 3: Terms, synonyms, concepts, abstract-speciﬁcation

relations.

synonyms can be stored as complex character string-

based tables. However, in terms of space efﬁciency, it

is more advantageous to store terms in a database ta-

ble, with one column for the term (as a string) and the

other for the termID (as an integer). This approach

allows for easier establishment of relations between

terms based on the integer IDs. Subsequently, if there

is a relation between terms, they can be connected

to each other using their respective ”IDs”. This can

be achieved through the use of relational or NoSQL

databases or by simply storing them in ﬁles. The ter-

mID can serve as a unique identiﬁer, which may al-

ready exist in the Acrolinx term database. However, if

the term is sourced from an Excel ﬁle or is entirely

new, a new termID must be generated for it. Syn-

onyms are assigned the same conceptID, and concepts

can have various types of relations with each other, in-

cluding abstract-speciﬁcation relations. Fig 4 shows

the data model of the ontology. We have deﬁned three

entities. The term entity represents a structure of how

terms are being stored in database. The concept en-

tity shows that synonyms are assigned the same con-

ceptID and concept concept entity are representative

of the relations between concepts. The data struc-

ture is stored in Sqlite, which functions as a rela-

tional database. A Python script is implemented to

build the database and its corresponding tables. Fur-

thermore, another Python script is utilized to import

terminologies into the database and merge any exist-

ing synonyms. Totally, we have 78,603 terms in term

table in different languages.

Figure 4: Ontology Data Model.

4.2 Relation Identiﬁcation

In our study, we are primarily interested in two types

of relations: synonym and abstract-speciﬁcation re-

lations. We examine synonym relations to address the

variation in vocabulary used by customers and experts

in inquiry emails. Utilizing synonyms helps identify

Towards Semi-Automatic Approach of Building an Ontology: A Case Study on Material Handling Data

251

the most similar previous projects, aiding in the cur-

rent inquiry process. Additionally, the uniﬁed termi-

nology should integrate term abstract-speciﬁcation re-

lation to facilitate the discovery of product variants as

well. In following sections, we will discuss our ap-

proach to ﬁnding those relations.

4.2.1 Synonym Detection

As mentioned earlier, the Acrolinx database contains

synonyms for certain terms. However, we are also in-

terested in exploring techniques to expand the num-

ber of synonyms, near synonyms, or similar terms.

Firstly, we calculated the word vector for each term

using SentenceTransformer in Python. Next, we uti-

lized a distance function provided by SciPy in Python

to calculate the similarity between the word vectors

using cosine similarity. We set the threshold to be

greater than 0.6 in order to avoid obtaining a large

number of irrelevant terms. In Table 3 and Table 4, we

present similar terms to the term ’cover’ as well as

family terms related to ’cover’ in both English and

German. Here, we notice that when a target term and

a similar term share a common part, it is more likely

for the model to perceive these two terms as highly

similar to each other. We have also observed that al-

though there is a degree of semantic similarity be-

tween the target and returned terms, however, these

terms cannot be considered exact synonyms and do

not fall into the category of synonyms. Therefore, we

classify them as similar terms for simplicity.

Table 3: Similar Terms in German.

Query Similar Terms

abdeckung Abst

utzung, Ablage

abdeckung komplett Abtragbock komplett

abdeckung links Gewindering links

abdeckung l

ufter Aufh

angeumklammerung

abdeckung l

ufter Gitterabdeckung

abdeckung rechts Gewindering rechts

Table 4: Similar Terms in English.

Query Similar Terms

cover support, shelf

cover complete support stand complete

cover links cover left, threaded ring left

cover fan hanging bracket, grid cover

cover right threaded ring right

4.2.2 Abstract-Speciﬁcation Relation

We are also interested in another relationship, which

involves integrating term abstractions and speciﬁca-

tions. This integration aims to enhance the discov-

ery of product variants as well. Hierarchical rela-

tions between terms in a text are essential for or-

ganizing concepts and understanding the semantic

structure. As it is shown in Fig 5, for instance, con-

sider the term ’cover’. It serves as an abstract term

representing a general concept. Within this hierar-

chy, we can identify speciﬁc types of covers, such as

’cover plate’ and ’cover surface’, which can be con-

sidered as subcategories or instances of the abstract

term. We implemented a tree-like algorithm to de-

tect such relation (parent-child) between terms in a

given domain. This structure is stored in a table within

a SQLite database containing three columns: ’con-

ceptID1’, ’conceptID2’ and ’relationType’. In this

context, ’conceptID1’ represents the child, ’concep-

tID2’ represents the parent, and ’relationType’ de-

notes the nature of their connection. In summary, the

identiﬁcation and storage of hierarchical relations be-

tween terms in a text provide a foundation for or-

ganizing concepts and understanding semantic struc-

tures. At present, this process requires approximately

20 minutes to handle 15,000 English terms, which is

longer than our initial expectations. Therefore, one of

our ongoing tasks is to optimize this algorithm, aim-

ing to achieve faster execution time.

Figure 5: Hierarchical Relation between Terms.

4.2.3 Relation Identiﬁcation Through Model

Creation

The analytical study which is carried out in this sec-

tion is mainly inspired by the work presented in (Mo-

hammed, 2020) which addressed the problem of syn-

onyms identiﬁcation from text corpus using super-

vised neural network. After analyzing the previous

section, it becomes evident that cosine similarity is

not always indicative of synonymy. It is uncommon

for the most similar word to be an actual synonym

of the target word. This brings up an important ques-

tion: Can we develop a system that classiﬁes whether

two words are synonyms based on their vector rep-

resentations? If so, how can we acquire labeled data

to train such a system. To tackle this challenge, we

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

252

approached the synonymy identiﬁcation problem as

a classiﬁcation task. As a learning algorithm, we uti-

lized a deep learning model provided by keras library

in python. The deep learning model built with Keras

consists of a 128 embedded layer followed by a Dense

layer, with the activation function set to softmax. With

a simple model we were able to get around 86.1%

accuracy on the test set. To acquire labeled training

data, we retrieved pairs of synonyms from our sqlite

database. These synonyms were previously popu-

lated with data from the Acrolinx database. Addition-

ally, we included more pairs that exhibit an abstract-

speciﬁcation relationship, which were obtained from

Section 4.2.2. The training dataset, in total, comprises

795 synonym pairs and 794 pairs demonstrating an

abstract-speciﬁcation relationship. For every classi-

ﬁcation run, the data was divided into training and

testing sets in a 75:25 ratio, respectively. Despite the

small size of the training dataset, the outcomes of

the classiﬁcation experiments are promising. In this

binary classiﬁcation scenario, the accuracy surpasses

our initial expectations, indicating the viability of em-

ploying supervised learning for the task of identify-

ing synonyms using word embeddings as features. We

have reevaluated our model by testing it on pairs of

terms that were not present in the training set. The

results are presented in Table 5, where 0 and 1 corre-

spond to the labels ’synonym’ and ’abstract-speciﬁc’

respectively. In future research, we aim to investigate

whether we can minimize misclassiﬁcation errors by

incorporating the deﬁnition of each term as an addi-

tional parameter into this model.

Table 5: Executing Model on Test Data.

term1 term2 Actual Label Predicted Label

kran kran pcc-250 1 1

kettenzug Kettenzug PKVUN 1 1

Kran Kranb

uhne 1 1

Kettenumlenkrad Kettenumlenkung 0 1

Benachrichtigung Mitteilung 0 0

Seilscheibe Umlenkrad Umlenkrolle 0 1

eingangswelle welle 1 1

5 REGULAR UPDATE OF

ONTOLOGY BASED ON

INCOMING INQUIRIES

The manufacturer gets inquiry emails from cus-

tomers, who use possibly a different vocabulary than

used by experts in previous projects. The ontol-

ogy should be updated accordingly by possible new

terms, which need to be manually related as syn-

onym terms or term abstractions. In order to ex-

tract keywords from incoming email, we applied

Named Entity Recognition (NER) (Al-Moslmi et al.,

2020) on some sample data from our project part-

ners. We utilized a pre-trained model from the spaCy

library to perform NER tasks. A named entity rep-

resents a tangible object in the real world and is as-

signed a label, such as ’person’, ’date’, ’country’ and

so on (Srinivasa-Desikan, 2018). To update our on-

tology, we have created a user interface using An-

gular web technology. Once the keywords are ex-

tracted, sales engineers can review and identify the

relevant terms. The selected terms, along with their

associated relationships, are then processed and in-

serted into the database. Fig 6 illustrates the various

stages involved in the evaluation of a new inquiry and

the subsequent update of the ontology by the ontology

editor.

Figure 6: Evaluation of incoming customer inquiries.

6 CONCLUSIONS

This research paper presents a data-driven method for

constructing an ontology. The purpose of this ontol-

ogy is to facilitate semantic search for past offers or

projects in relation to incoming inquiry texts. The ap-

proach employed is not entirely automatic, as a fully

automatic process would result in low-quality term

deﬁnitions and relationships. Our intention is not to

create an ontology from scratch but to build it by

utilizing existing resources, such as our terminolo-

gies. In our study, we tackled the task of detecting

synonyms and abstract-speciﬁcation relations using

word embeddings, while also identifying hierarchical

structures between terms. Sentence transformers were

employed to construct word embeddings, and we con-

ducted a qualitative evaluation of the most similar

words to speciﬁc targets. Our investigation revealed

that distributional similarity does not always imply

synonymy; instead, similarities may be attributed to

other functional factors, such as domain similarity

resulting from the presence of common words be-

Towards Semi-Automatic Approach of Building an Ontology: A Case Study on Material Handling Data

253

tween terms. Additionally, In our study, we utilized

word embeddings as features to train the deep learn-

ing model. Moving forward, we aim to enhance these

features by incorporating additional information in-

ferred from the context itself, allowing the features

to capture the linear contexts in which the words

typically appear. This will aid in distinguishing syn-

onymy from other sense relations. Furthermore, our

current models generate a single vector representa-

tion or word embedding for each term. However, con-

textual models can generate word representations that

are inﬂuenced by the surrounding words in the sen-

tence (Deb and Chanda, 2022). To achieve this, it

is crucial to consider the deﬁnitions of individual

terms, rather than solely focusing on the terms them-

selves. Therefore, this paper is being presented as

an ongoing project, and our next objective is to en-

hance the ontology by integrating term deﬁnitions

sourced from additional information channels in fu-

ture. Furthermore, we demonstrated that embeddings

can serve as effective features for training deep learn-

ing model in classiﬁcation tasks. In order to capture

new terms, we utilize NER and employ an Ontology

Editor to streamline the process of updating the on-

tology. In conclusion, this research paper illustrates

that employing machine learning and NLP techniques

enables the development of an ontology in a semi-

automatic manner by extracting terms and detecting

relations between terms. The study highlights the po-

tential of these approaches in enhancing the ontology

construction process.

ACKNOWLEDGEMENTS

This work is supported by ITEA3 under the supervi-

sion of the German Federal Ministry of Education and

Research (FKZ: 01IS21084 (InnoSale)).

REFERENCES

Al-Moslmi, T., Oca

na, M. G., Opdahl, A. L., and Veres,

C. (2020). Named entity extraction for knowledge

graphs: A literature overview. IEEE Access, 8:32862–

32881.

Baader, F., Horrocks, I., and Sattler, U. (2005). Description

logics as ontology languages for the semantic web.

In Mechanizing mathematical reasoning, pages 228–

248. Springer.

Blondel, V. D. and Senellart, P. P. (2002). Automatic ex-

traction of synonyms in a dictionary. vertex, 1:x1.

Borst, W. N. (1997). Construction of Engineering Ontolo-

gies for Knowledge Sharing and Reuse. Phd thesis,

Universiteit Twente, Enschede.

Cimiano, P., M

adche, A., Staab, S., and V

olker, J. (2009).

Ontology learning. In Handbook on ontologies, pages

245–267. Springer.

Deb, S. and Chanda, A. K. (2022). Comparative analysis

of contextual and context-free embeddings in disaster

prediction from twitter data. Machine Learning with

Applications, page 100253.

Ding, L., Finin, T., Joshi, A., Pan, R., Cost, R. S., Peng,

Y., Reddivari, P., Doshi, V., and Sachs, J. (2004).

Swoogle: a search and metadata engine for the se-

mantic web. In Proceedings of the thirteenth ACM

international conference on Information and knowl-

edge management, pages 652–659.

Gruber, T. R. (1995). Toward principles for the design of

ontologies used for knowledge sharing. Int. J. Hum.

Comput. Stud., 43:907–928.

Guarino, N., Oberle, D., and Staab, S. (2009). What Is an

Ontology?, pages 1–17. Springer Berlin Heidelberg,

Berlin, Heidelberg.

Han, X., Gao, T., Lin, Y., Peng, H., Yang, Y., Xiao, C.,

Liu, Z., Li, P., Sun, M., and Zhou, J. (2020). More

data, more relations, more context and more openness:

A review and outlook for relation extraction. arXiv

preprint arXiv:2004.03186.

Hotho, A., J

aschke, R., Schmitz, C., and Stumme, G.

(2006). Bibsonomy: A social bookmark and publi-

cation sharing system. International Journal.

Mangold, C. (2007). A survey and classiﬁcation of seman-

tic search approaches. International Journal of Meta-

data, Semantics and Ontologies, 2(1):23–34.

Mohammed, N. (2020). Extracting word synonyms from

text using neural approaches. Int. Arab J. Inf. Technol.,

17(1):45–51.

Motik, B., Patel-Schneider, P. F., Parsia, B., Bock, C., Fok-

oue, A., Haase, P., Hoekstra, R., Horrocks, I., Rutten-

berg, A., Sattler, U., and Smith, M. (2012). OWL 2

web ontology language: Structural speciﬁcation and

functional-style syntax. W3c recommendation, W3C.

Nguyen, T. H. and Grishman, R. (2015). Relation extrac-

tion: Perspective from convolutional neural networks.

In Proceedings of the 1st workshop on vector space

modeling for natural language processing, pages 39–

48.

Ramkumar, A. S. and Poorna, B. (2014). Ontology based

semantic search: An introduction and a survey of cur-

rent approaches. In 2014 International Conference on

Intelligent Computing Applications, pages 372–376.

Srinivasa-Desikan, B. (2018). Natural Language Process-

ing and Computational Linguistics: A practical guide

to text analysis with Python, Gensim, spaCy, and

Keras. Packt Publishing Ltd.

Wang, T. and Hirst, G. (2009). Extracting synonyms from

dictionary deﬁnitions. In Proceedings of the Interna-

tional Conference RANLP-2009, pages 471–477.

Yıldız, T., Yıldırım, S., and Diri, B. (2014). An integrated

approach to automatic synonym detection in turkish

corpus. In Advances in Natural Language Processing:

9th International Conference on NLP, PolTAL 2014,

Warsaw, Poland, September 17-19, 2014. Proceedings

9, pages 116–127. Springer.

Zelenko, D., Aone, C., and Richardella, A. (2003). Kernel

methods for relation extraction. Journal of machine

learning research, 3(Feb):1083–1106.

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

254