Interactive Domain-Specific Knowledge Graphs

from Text: A Covid-19 Implementation

Vinı́cius Melquı́ades de Sousa1[0000−0002−1282−5857] and

Vinı́cius Medina Kern2[0000−0001−9240−304X]
1
Universidade Federal de Santa Catarina, Florianópolis, BRA
[email protected]
2
Universidade Federal de Santa Catarina, Florianópolis, BRA - Funded by CNPq,
Research Productivity Grant 314140/2018-2
[email protected]

Abstract. Information creation runs at a higher rate than informa-

tion assimilation, creating an information gap for domain specialists
that usual information frameworks such as search engines are unable to
bridge. Knowledge graphs have been used to summarize large amounts
of textual data, therefore facilitating information retrieval, but they re-
quire programming and machine learning skills not usually available to
domains specialists. To bridge this gap, this work proposes a framework,
KG4All (Knowledge Graphs for All), to allow for domain specialists to
build and interact with a knowledge graph created from their own chosen
corpus. In order to build the knowledge graph, a transition-based system
model is used to extract and link medical entities, with tokens repre-
sented as embeddings from the prefix, suffix, shape and lemmatized fea-
tures of individual words. We used abstracts from the COVID-19 Open
Research Dataset Challenge (CORD-19) as corpus to test the framework.
The results include an online prototype and correspondent source code.
Preliminary results show that it is possible to automate the extraction
of entity relations from medical text and to build an interactive user
knowledge graph without programming background.

Keywords: Knowledge graphs · COVID-19 · Information retrieval soft-

ware · Natural language processing · Personalized analytics.

1 Introduction
Shannon’s Mathematical Theory of Communication [19] is understood as the
Information Science debut [4]. Ever since Shannon’s work the field has evolved
into a number of sub-fields, following the advances in society. One of such fields
is Information Retrieval, which was considered to be the Information Science
main core [17]. It started in the 1970’s and its focus was on the creation of
retrieval indexes and the physical allocation of information. As the technological
development took place the focus shifted towards information processing and
efficient information retrieval, digitally speaking [5]. The use of knowledge graphs
to represent human knowledge, and therefore as a way into information retrieval,
2 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

has been receiving attention both from academia and industry. A knowledge
graph can be defined as a structured representations of facts, in the form of
entities and relations and its semantic description [10]. A knowledge graph is
composed by triplets in the form (head entity, relation, tail entity), Figure 1
depicts an example of a knowledge graph where on the left-side it is presented
the triplets and on the right-side the representation in a graph form.

Fig. 1. Example of Knowledge Graph. Extracted from [10].

The construction of knowledge graphs can be classified into two main groups:
(i) manually/curated or (ii) automatic/semi-automatic. The first group consists
of allocating domain specialists to annotate, in accordance with a set of rules, the
entities, relations and descriptions [22]. Manually constructed knowledge graphs
are time consuming and tends to advance at a slower pace than information
development. On the other hand, automatic/semi-automatic knowledge graphs
are built upon a workflow, usually starting from a text corpus, from which en-
tities and relations are inferred. Automatic/semi-automatic constructed knowl-
edge graphs are able to keep up with the information creation, at the cost of (i)
quality, that is, the entities and relations are not as accurate as when the knowl-
edge graph is manually annotated [9] and (ii) having to deal with engineering
challenges, such as data acquisition and storage, text parsing, information ex-
traction, etc. While companies such as Google and Microsoft have the necessary
resources to solve these challenges, smaller organizations and independent re-
searchers are required to have programming skills in order to be able to use
the advances of research in the information retrieval through knowledge graphs
[18]. In other terms, the use of machine learning in information retrieval through
knowledge graphs results in an increase on the complexity demanded to make
use of such advances. The higher the complexity, the more limited is the number
of people capable of making use of the gains allowed by those advances [14] [8].
The information accessibility and availability for possible users is one of the
tasks that Information Science is responsible for [15], as the general view of
the information process, from creation to utilization, is a core activity of the
area [2]. Domain specialists is a particular group of users, with real needs, that
could benefit from using knowledge graphs. They are not usually proficient in
 Interactive Domain-Specific Knowledge Graphs 3

programming/machine learning optimization skills and, at the same time, their

information needs are not fulfilled by regular knowledge frameworks, such as
google [11]. Therefore, if: (i) domain specialists cannot assimilate, through hu-
man cognition, the information in the same pace that the information is created
[9]; (ii) regular knowledge frameworks are not sufficient to fulfill the domain
specialists information needs; and (iii) domain specialists do not have the tech-
nical skills in order to make use of algorithms that would allow them to process
and interact with a large amount of information. Then, it can be stated that
a framework that allowed domain specialists to create and interact with their
own knowledge graphs without requiring programming skills would be a step
towards narrowing the information creation and assimilation gap. The present
work depicts the preliminary results towards a framework that aims to assert
the previously stated problem. In other words, it is presented the preliminary
work of a framework that aims to allow domain specialist to make use of the
advantages of Knowledge Graphs research by creating its own knowledge graph.
The work is organized as follows. Section 2 presents the used methods in
order to achieve the results shown in section 3. A discussion about the results is
found in section 4 and, finally, section 5 concludes the present work.

2 Methods
This section presents the methods that were used in order to create the presented
results. The sub-section 2.1 depicts the search result for similar works, followed
by the sub-section 2.2 that presents the general overview of the proposed frame-
work. Sub-section 2.3 explains the NLP technique that was used to build the
knowledge graph. And finally, sub-section 2.4 is responsible for justifying the
use of network visualization.

2.1 Similar Works

In order to execute a search for similar works at least three search parameters
have to be defined: (i) Scientific Bases; (ii) Keywords; and (iii) inclusion and ex-
clusion criteria. Such definitions are as follows. (i) Searched Scientific Bases are:
Web of Science, Scopus, IEEE Xplore and Association for Computing Machin-
ery Digital Library (ACM). (ii) Chosen keywords: Knowledge Graph, text OR
corpus and Graphical Interface OR Web Application. (iii) The inclusion criteria
is listed:

1. Present a framework to build a knowledge graph from text corpus;

2. Present a form of interacting with the knowledge graph;
3. Make the source code or the framework available for use.

And, finally, the exclusion criteria:

1. Not Present a framework to build a knowledge graph from text corpus;
2. Not Present a form of interacting with the knowledge graph;
4 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

3. Not Make the source code or the framework available for use;
4. Not being an scientific paper;
5. Not being in English or Portuguese

The search resulted in seventy-two (72) retrieved papers, after removing du-
plicate papers a total of sixty-nine (69) paper abstracts were read by the authors.
For each abstract it was attributed the inclusion and exclusion criteria. Figure 2
depicts the distribution count of paper for each criteria combination. The papers
placed within the black rectangle refers to the papers to which were attributed
at least one inclusion criteria and none exclusion criteria. That is, these are the
works considered to be similar to the present one. The work myDIG: Personal-
ized illicit domain-specific knowledge discovery with no programming [11] was
the only one that was classified as a similar work by the criteria defined.

Fig. 2. Inclusion and Exclusion Criteria Count

The work was developed at the Information Sciences Institute of University of

Southern California, and presents a framework that allows investigative domain
specialists to build and interact with their own knowledge graphs from the web
 Interactive Domain-Specific Knowledge Graphs 5

pages. As one would expect, there are similarities and differences between myDIG
and the present work.
The main similarity is found in the problem to be solved. Both works acknowl-
edges that domain specialists struggle to keep up with the information creation.
At the same time, the advances of data processing with Machine Learning, that
would allow a way to narrow the gap between information creation and assimi-
lation, requires programming and machine learning skills, that is not commonly
found in domain specialists, restricting the number of domain specialists that
can make use of such advances.
On the other hand, the main difference is found in the user profile. Both
works have in mind domain specialists. However, while myDIG is focused in
a case where the user has a well defined idea of what she is looking for, the
present work focuses on the step where the domain specialist needs to have an
overview of the knowledge relation in his corpus, that is, an easy to assimilate
and interactive content summarization. Another difference is found in the input
data, myDIG uses web pages while the present paper works is build upon natural
language text. One final difference worth mentioning is related to the availability
of the framework. The myDIG paper indicated a GitHub repository with the
framework code, and therefore is was not attributed to it the third exclusion
criteria. However, when the authors of the present work read the full myDIG
paper it was explained that the engine that transform web pages into a knowledge
graph is maintained by a private company and its not available and therefore it
was not explained how it worked. This work on the other hand was built upon
open source technologies and is also completely available3 .
Next sub-section presents the proposed framework that aims at allowing
domain specialists with no programming skills to benefit from machine learning
advances.

2.2 Proposed Framework - KG4All

Figure 3 presents the proposed framework, named as KG4All, that stands for
Knowledge Graphs for All. The image can be read starting from the left side
user icon and following the lines direction. The user uploads a corpus to a web
application. This web application then sends the text from the corpus to a back-
end. This stage is where the machine learning algorithms are used in order to
build a knowledge graph from the texts. Once the Knowledge Graph is created
the web application makes use of interactive tools, allowing the user to interact
with the knowledge contained in the corpus that was uploaded.
This work, as mentioned in the last paragraph of 1, presents preliminary
results of the process of building the KG4All. Specifically, it presents the first
results of the elements inside the black rectangle. That is, the corpus element,
highlighted with red in Figure 3, has not been implemented yet.
3
Web interface code:https://github.com/viniciusmsousa/kg4all. Data Processing
workflow: https://github.com/viniciusmsousa/KG4All-data-processing-explained.
At the current stage these components are not connected in the application, as
explained in section 3.
6 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

Fig. 3. Proposed framework

A few practical considerations should be made. The choice to build a web

application was made as the result of the following reasoning: The utilization
of digital a tool is necessary mainly due to the fact that large amount of data
processing is only possible through computers. Therefore, the real decision to be
made is whether to build an web application or a smartphones app. The authors
have chosen to build the web application for the following reasons. First, given
the authors background build an web application presented less technical chal-
lenges. And secondly, people tend to be more productive on personal computers
when compared to smartphones[1]. In order to build the presented framework
the authors used the Shiny R package [7], which is a framework to build web
application using the statistical programming language R [16]. Examples of oth-
ers apps built with the framework can be found in the maintainer official gallery
web page4 . The main advantage of the framework is that it allows the creation
of fully functional web application with in a relatively simple structure. The
en core sci sm model from the SciSpacy [13] python package was the choice to
build the NLP tasks, that are explained in the sub section 2.3. Finally, as ex-
plained in sub section 3.1, the implementation was made using the metadata file
from the COVID-19 Open Research Dataset Challenge (CORD-19) [3] and the
raw data used for the results presented can be found in the link5 .
With that in mind the rest of this section presents the steps taken in order
to achieve the preliminary result, shown in section 3.

4
https://shiny.rstudio.com/gallery/
5
https://drive.google.com/drive/folders/1YAHpv4-93rqMy94CyP830fRzN81Cwk9? usp =
sharing
 Interactive Domain-Specific Knowledge Graphs 7

2.3 NLP
The objective is to allow the user to upload it’s own corpus into KG4All, then
from this corpus a knowledge graph is built. This section presents the text pro-
cessing tasks that are responsible to create the triplets set from the texts. In
other words, the Machine Learning block in the Figure 3. The general task, i.e.,
extract triplets from natural language text, can be splited into two sub-taks: (i)
Name Entity Recognition and (ii) Entity Linking.
Name Entity Recognition (NER) labels sequences of words in a text which
are the names of things, such as person, company, etc. [21]. For example take
the following natural language statement:
Armstrong landed on the moon.
After a NER processing this statement could be annotated as follows:
Armstrongperson landed on the moonlocation .
Since KG4All is implemented in the medical domain it is needed a source to
get medical entities definitions. The Unified Language Medical System (UMLS)
[6] provides just that. A few examples are shown in Table 1 and the full database
with the definitions and relations from UMLS used in this work can be found in
this link6 .

Table 1. Examples of Medical Entities fro the UMLS.

Entity Type Entity Name

Intellectual Product Clinical Trial Objective
Virus Avipoxvirus
Cell Component Azurophilic granules
Temporal Concept Priority
Bird Aves
Intellectual Product Report (document)
Population Group Donor person

Therefore, an example of a NER annotated medical text could look like:

The reportIntellectual Product on the AvipoxvirusVirus is the current
priorityTemporal Concept .
The second sub-task is called Entity Linking which aims at finding a relation
between two entities [21]. For example, by reading the statement Armstrong
landed on the moon the human cognition interprets the semantic meaning
and concludes that there is a link between the entities Armstrong and moon.
And this link is landed on. Finally, this knowledge can be represented in a
triplet form as:
6
https://drive.google.com/drive/folders/1kEw1r JA7pI5V ycmaXBV wbN 0XM W U M sST ?usp =
sharing
8 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

(Armstrong,landed on,moon)

Entity linking aims at using algorithms to detect these relations. The algo-
rithms usually integrate three steps to link entities [21]:

1. Entity mention spotting: Detects mentions in the text of multiple entities;

2. Entity mention mapping: Lists the possibles entities from a formal knowledge
base;
3. Candidate Selection: Selects, based on a criteria, which candidates are indeed
linked with the mentioned entity.

Therefore, by completing this two sub-task it is possible to build a knowledge

graph from text. An example of the data processing workflow developed by
authors to create a KG from Medical text can be found on this github page
(prepared by the authors)7 . It presents the use of the SciSpacy [13] which is a
open source python [20] framework dedicated to dealing with scientific texts from
medical domain. The framework allows a large range of tasks, but the purpose
of this research it was focused on the Named Entity Recognition and Entity
Linking.
Once the data processing workflow is completed it is possible to create tools
to allow the user to interact with the knowledge graph without having to program
anything, this is the topic of the sub section 2.4.

2.4 Knowledge Graph Visualization

Having the extracted triplets from the corpus the next step towards the proposed
framework is to allow a user to interact with the knowlege graph. As shown in
section 1, a knowledge graph has a network structure, i.e., nodes (the entities)
connected by edges (the relations).
Producing and examining a network plot is often one of the first step in a
network analysis, since its overall purpose is to allow a better understanding of
the underling structure in the data [12]. Figure 1 is an example of how a network
can be visualized in order to reveal the underlying structure of the data. The use
of aesthetics can enhance certain feature from the data in a better visual form.
Color the nodes to indicate different types of nodes and change the edge size to
depict the relation strength or count are two ways to do so. Therefore, the first
interaction element implemented in the KG4All is the tool that allows the user
to view a network graph from the knowledge graph extracted form the corpus,
by selecting a document of interest.
In summary, this section started demonstrating, in sub section 2.1, the re-
search gap in allowing domain specialists to benefit from the advances in the
machine learning and natural language processing research in order to inter-
act with a large number of documents. Second, sub section 2.2, presented and
explained in a high level a possible way towards fulfilling the gap previously
7
https://github.com/viniciusmsousa/KG4All-data-processing-
explained/blob/main/01DataProcessingExplained.ipynb
 Interactive Domain-Specific Knowledge Graphs 9

mentioned with the KG4All framework. Thirdly, sub section 2.3 explained the
tasks of entity recognition and entity linking, which are the tasks that the pre-
sented work relies upon. And finally, the current sub section justified the choice
of using network graphs to create an interactive knowledge graph as kick start
to the web application. The next session presents the results obtained as this
research evolves.

3 Results
This section presents the functional prototype of the KG4All framework. As
stated before the KG4All source code is open source, currently it is present
in two github repositories8 due to the fact that the Web Application is not
integrated with the Machine Learning back end yet. And the application can be
accessed throught the link viniciusmsousa.shinyapps.io/KG4All9 . The prototype
current main features are: (i) detects the relations within the abstracts from the
COVID-19 Open Research Dataset Challenge (CORD-19) [3] and (ii) connect
these relations to the UMLS relations mapping. The following of this presents
the domain implementation and test corpus in sub section 3.1. Next, the web
interface components in sub section 3.2. The triplets display component in sub
section 3.3 and finally the interactive graph in sub section 3.4.

3.1 Domain Implementation

The Covid-19 pandemic breakout in early 2020 and changed most people’s life.
The subject became an important topic in the international organizations agen-
das. Due to the fact that this research was taking place during the pandemic peak
in Brazil the authors decided to implement the proposed framework in the Med-
ical Domain. Specifically, it was choosen to develop a data processing workflow
that works with medical texts, based in the Unified Medical Language System
[6] and tested it to generate the results using the abstracts of papers related
to the coronavirus extracted from the COVID-19 Open Research Dataset Chal-
lenge (CORD-19) [3], which is the result of a response coordinated by the White
House to make available the scientific publications related to the coronavirus.

3.2 Web Interface Components

Figure 4 presents the KG4All Web Interface, i.e., the interface that the domain
specialist interacts with. It is composed mainly by three components marked in
the figure. Component 1 is the search bar that allows the user to search for a
desired document using words. Component 2 presents the triplets extracted from
the abstract of the document selected in component 1. And finally, component
3 presents the interactive graph visualization of the relations.
The sub section 3.3 explains the KG4All’s triplets component.
8
Web application: https://github.com/viniciusmsousa/kg4all. Data Processing:
https://github.com/viniciusmsousa/KG4All-data-processing-explained.
9
viniciusmsousa.shinyapps.io/KG4All
10 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

Fig. 4. KG4All Web Interface (’Influenza Virus: A Brief Overview’)

3.3 Triplets Component

The triplets are shown in component 2 of the KG4All. Figure 5 is a zoom in

component 2. Each row each represents on relation found in the text. The relation
in itself is in the columns head name, relation type and tail name. The other
columns of the table presents additional information about the relation. We
highlight that the column entitled relation count depicts the number of times
that the relation occurred in the whole corpus.

Fig. 5. Triplets Component (’Influenza Virus: A Brief Overview’)

For example, from the first line of the figure it can be seen that the entity
Influenza (head entity) is a process of (relation) the Influenza Virus (tail
entity). Next step in to connect this entities with other entities found in the
corpus, through the UMLS. This is presented in the sub section 3.4.
 Interactive Domain-Specific Knowledge Graphs 11

3.4 Interactive Graph

The last implemented component is the interactive knowledge graph, presented

in Figure 6. One might note that there more relations in the graph than in
the triplets table. This is due to the fact that the graph shows the relations
found in the abstract with the relations found in the whole corpus that involves
the entities from the selected text. This allows the domain specialist to have
a general view of how the selected text is related with the whole corpus. It
is worth mentioning that even though it is not implemented, the authors are
studying ways to explicit the document from which the additional relations are
from.

Fig. 6. Interactive Knowledge Graph (’Influenza Virus: A Brief Overview’)

As is expected in network visualizations KG4All use some aesthetics to add

more information to the relations. The nodes color demonstrates to which group
each node belong to. For example, Influenza is classified by the UMLS as a
Disease or Syndrome. On the other hand influenza A virus is classified as
a virus. Another aesthetics used in the interactive graph is the edge (or link)
size. It proportional to the number of times that the relation was present in the
corpus.
Section 4 presents considerations about the results, improving directions that
are in the authors workflow.

4 Discussion

A few considerations about KG4All itself. First, the current implementation

uses a selected dataset to create the interactive knowledge graph, however this
is a temporary implementation. Once the upload interface and the integration
between the web interface and the back end model is done, KG4All will have
an upload interface where the users will be allowed to upload their own medical
corpus. Second, there are two factors that impacts KG4All computing cost: (i)
12 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

The model used to detect the medical entities and (ii) the size of the corpus that
is submitted to the data processing pipeline. The model that is current being used
is the en core sci sm [13] and once it is loaded it uses 132MiB of memory. And the
dataset used to create the prototype, with 81.354 medical abstracts, used around
10GB while running on win10 with intel i7. It is worth noting that in practical
use the authors expect smaller corpus, for instance, the result os a search in
scientific articles database. Third, the use of machine learning algorithms to
extracted the triplets cannot guarantee that all the entities relations present
in the text will be extracted. How ever, as shown in the SciSpacy paper [13]
the amount of relations detected are not insignificant, providing a reasonable
summarizing of the knowledge present in the corpus. And, finally, there are both
some implementations as well as corrections to be made on the current state.
For example, a way to explicit from each document the entity was extracted,
when it is a relations that is not in the selected document is a implementation
to be made. In some cases there are overlapping of the edges name, which is a
correction in the back log.
Besides the practical differences from the myDIG [11] work, explained in
sub section 2.1, the authors believes that KG4All complements the myDIG,
in the sense that the same issue, gap between domain specialists information
assimilation and creation, is being addressed. And contributing for a different
group of information users by focusing in an open source tool for knowledge
graph creation and interaction.

5 Conclusion

The present work has argued that there is a gap between information creation
and assimilation. This gap impacts the domain specialists, a group that the tra-
ditional information tools does not satisfies their information necessities. It has
also been argued that the research advances in the field of information retrieval
through knowledge graph using machine learning algorithms is evolving and pro-
vides ways to narrow the information gap. However, such advances are relied on a
high level of computational and mathematical complexity. This high complexity
results in the need of programming and machine learning skills in order to make
use of the advances. Such skills are not commonly found in domain specialists.
It was presented the KG4All prototype, which is a framework that will allow
users to upload their own corpus and interact with a knowledge graph created
from the corpus without the need of programming skills. The domain that the
KG4All is implemented is the medical one, due to the fact that the Covid-19
pandemic took place while this research was taking place. There are other works
proposing solutions to the same problem however with a differences in the target
domain specialists user profile, and therefore, the present work contributes to
the research on how to make the advances in knowledge graph through machine
learning usage.
 Interactive Domain-Specific Knowledge Graphs 13

References
1. Adepu, S., Adler, R.F.: A comparison of performance and preference on mobile
devices vs. desktop computers. In: 2016 IEEE 7th Annual Ubiquitous Comput-
ing, Electronics Mobile Communication Conference (UEMCON). pp. 1 – 7 (2016),
https://ieeexplore.ieee.org/document/7777808
2. Agarwal, R., Dhar, V.: Editorial—Big Data, Data Science, and Analytics: The
Opportunity and Challenge for IS Research. Information Systems Research 25(3),
443 – 448 (2014), https://doi.org/10.1287/isre.2014.0546
3. AI, A.I.F.: https://www.kaggle.com/allen-institute-for-ai/
CORD-19-research-challenge
4. Araújo, C.A.A.: Correntes teóricas da ciência da informação. Ciência da in-
formação 38, 192 – 204 (12 2009), http://www.scielo.br/scielo.php?script=
sci_arttext&pid=S0100-19652009000300013&nrm=iso
5. Araújo, C.A.A.: Fundamentos da Ciência da Informação: correntes teóricas e o con-
ceito de informação. Perspectivas em Gestão & Conhecimento 4(1), 57 – 79 (2014),
https://periodicos.ufpb.br/ojs2/index.php/pgc/article/view/19120
6. Bodenreider, O.: The unified medical language system (umls): integrating biomed-
ical terminology. Nucleic acids research 32(suppl 1), D267–D270 (2004)
7. Chang, W., Cheng, J., Allaire, J., Xie, Y., McPherson, J.: shiny: Web Application
Framework for R (2020), https://CRAN.R-project.org/package=shiny, r pack-
age version 1.5.0
8. Elbashir, M., Collier, P., Davern, M.: Measuring the effects of business intel-
ligence systems: The relationship between business process and organizational
performance. International Journal of Accounting Information Systems 9(3),
135 – 153 (2008), https://www.scopus.com/inward/record.uri?eid=2-s2.
0-51249116446&doi=10.1016%2fj.accinf.2008.03.001&partnerID=40&md5=
f6748444fd6918d43aa33b5de2c118d3, cited By 254
9. Hoyt, C.T., Domingo-Fernández, D., Aldisi, R., Xu, L., Kolpeja, K., Spalek,
S., Wollert, E., Bachman, J., Gyori, B.M., Greene, P., Hofmann-Apitius, M.:
Re-curation and rational enrichment of knowledge graphs in Biological Expres-
sion Language. Database 2019 (jan 2019), https://academic.oup.com/database/
article/doi/10.1093/database/baz068/5521414
10. Ji, S., Pan, S., Cambria, E., Marttinen, P., Yu, P.S.: A Survey on Knowledge
Graphs: Representation, Acquisition and Applications (2020), https://arxiv.
org/abs/2002.00388
11. Kejriwal, M., Szekely, P.: myDIG: Personalized illicit domain-specific knowledge
discovery with no programming. Future Internet 11(3) (2019), https://www.mdpi.
com/1999-5903/11/3/59
12. Luke, D.A.: A user’s guide to network analysis in R. Springer (2015)
13. Neumann, M., King, D., Beltagy, I., Ammar, W.: ScispaCy: Fast and Robust
Models for Biomedical Natural Language Processing. In: Proceedings of the 18th
BioNLP Workshop and Shared Task. pp. 319 – 327. Association for Computa-
tional Linguistics, Florence, Italy (2019), https://www.aclweb.org/anthology/
W19-5034
14. Olszak, C., Ziemba, E.: Approach to building and implementing Business Intelli-
gence systems. Interdisciplinary Journal of Information, Knowledge, and Manage-
ment 2, 135 – 148 (2007), https://www.scopus.com/inward/record.uri?eid=
2-s2.0-77749242597&partnerID=40&md5=fd70fbb98a2ddee0b6daf68f28050db5,
cited By 81
14 Vinı́cius Melquı́ades de Sousa and Vinı́cius Medina Kern

15. Pinto, A.L., Silva, A.M., Sena, P.M.B.: Ontologias baseadas na visualização da
informação das redes sociais. Prisma.com (Portugual) 24(13), 5 – 24 (2010), https:
//www.brapci.inf.br/index.php/res/v/68060
16. R Core Team: R: A Language and Environment for Statistical Computing. R
Foundation for Statistical Computing, Vienna, Austria (2020), https://www.
R-project.org/
17. Saracevic, T.: Ciência da informação: origem, evolução e relações. Perspectivas em
Ciência da Informação 1(1) (1996), http://portaldeperiodicos.eci.ufmg.br/
index.php/pci/article/view/235
18. Sen, S., Li, T.J., Team, W., Hecht, B.: WikiBrain: Democratizing Computation on
Wikipedia (2014), https://doi.org/10.1145/2641580.2641615
19. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical
Journal 27(3), 379 – 423 (1948), https://onlinelibrary.wiley.com/doi/abs/
10.1002/j.1538-7305.1948.tb01338.x
20. Van Rossum, G., Drake, F.L.: Python 3 Reference Manual. CreateSpace, Scotts
Valley, CA (2009)
21. Waitelonis, J.: Linked Data Supported Information Retrieval. Ph.D. thesis, Karl-
sruher Institut für Technologie (2018)
22. Yuan, J., Jin, Z., Guo, H., Jin, H., Zhang, X., Smith, T., Luo, J.: Constructing
biomedical domain-specific knowledge graph with minimum supervision. Knowl-
edge and Information Systems 62(1), 317 – 336 (2020), https://link.springer.
com/article/10.1007/s10115-019-01351-4