A Semantic Approach for Generating Graphical Representation from
Aircraft Maintenance Text
Thi-Bich-Ngoc Hoang
a
, Ba-Huy Tran
b
and Marzieh Mozafari
c
Capgemini Engineering, France
Keywords:
Maintenance Task, Process Ontology, Graphical Illustration, Technical Language Processing, Information
Extraction, Aircraft Maintenance Manual.
Abstract:
Industrial maintenance is a strategic business function. Over the past twenty years, the role of maintenance in
companies has become increasingly important both technologically and economically. However, maintenance
service has not taken into account the frequent change in maintenance knowledge, the users’ perspective
(the training, the origins, or the cultures), and the users’ support documents preferences. In this article, we
propose preliminary results of an approach to make industrial maintenance universal. We use natural language
processing techniques to extract core information from maintenance text and then construct a knowledge
base to store all relevant information about maintenance processes, domain information, and corresponding
graphics. Finally, we generate a graphical representation of input text to help better understand the procedure,
thus increasing the user experience and the performance of maintenance operations in terms of reducing time
and cost. This approach is first applied to aircraft maintenance and can be applied to maintenance in other
industry domains as well.
1 INTRODUCTION
Industrial maintenance is a strategic business func-
tion. It can be defined as all the troubleshooting and
repair actions, adjustment, overhaul, control and ver-
ification of material or even immaterial equipment.
Over the past twenty years, the role of maintenance in
companies has become increasingly important, both
technologically and economically. Whether it is in-
dustrial maintenance expenditure or dedicated staff,
the maintenance sector shows a significant increase
on all points. In France, annual expenditure on main-
tenance is around 18 billion euros and requires 70,000
jobs
1
.
In addition, the needs of user maintenance play-
ers evolve over time and cannot be satisfied by the
services currently provided by computer maintenance
support systems on the market. Indeed, these services
are based on the knowledge initially formalized but
which is not systematically updated. Thus the ser-
vices offered after a few years are no longer in line
a
https://orcid.org/0000-0002-8693-9195
b
https://orcid.org/0000-0002-2578-9138
c
https://orcid.org/0000-0002-1384-7548
1
https://metgroupe.fr/les-chiffres-cles-maintenance/
with current knowledge. We must take into account
the dynamic aspect of knowledge, to meet the needs
of users and improve the performance of help soft-
ware offering these services.
Different activities in industrial maintenance gen-
erate a vast volume of written data in the form of re-
ports, historic records, plans, and schedules. Many
text processing tasks out of these textual data can
be effectively automated using Natural Language
Processing (NLP). Extracting practical information
from maintenance documentations presents a unique
challenge in the domain of information extraction
(IE), because these instructional texts include multi-
ple steps with specific objects which should be per-
formed sequentially. In addition, this process highly
depends on the quality of the raw data and the way
it is processed with NLP such as pre-processing, To-
kenization, Part-of-Speech tagging (POS), Name En-
tity Recognition (NER), etc.
The aim of this study is to leverage NLP along
with knowledge bases to improve the performance
of maintenance documents analysis, simplify aircraft
maintenance processes, and insure semantic interop-
erability. To that end, we extract core information
in aircraft maintenance documents including descrip-
tion, warnings, cautions, notes, actions, related com-
Hoang, T., Tran, B. and Mozafari, M.
A Semantic Approach for Generating Graphical Representation from Aircraft Maintenance Text.
DOI: 10.5220/0011543000003335
In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 3: KMIS, pages 169-176
ISBN: 978-989-758-614-9; ISSN: 2184-3228
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
169

ponents, and information related to each component
using NLP techniques. Then we develop an ontology
to describe the maintenance procedure and domain,
and also generate illustrations using the graphic rep-
resentation of resources or contexts in our knowledge
base to better guide the user. To the best of our knowl-
edge, this is the first attempt in the aeronautic main-
tenance domain, and this approach can be applied to
other industrial domains as well.
The rest of the paper is organized as follows. Sec-
tion 2 presents a literature review of the informa-
tion extraction from text and semantic representation
of maintenance procedure. The NLP techniques for
extracting the information from maintenance docu-
ments and the methodology for generating ontology
and graphical representation of maintenance proce-
dures are described in Section 3. Finally, Section 4
draws some conclusions and offers a view of possible
future work.
2 RELATED WORK
2.1 Information Extraction from Text
Information extraction from text has attracted a num-
ber of studies in recent years. However, most of the
information extraction systems have been developed
for domains such as medical (Xu et al., 2017a; Ge
et al., 2020; Lopes et al., 2019), biomedical (Nayel
et al., 2019; Gao et al., 2021; Yang et al., 2022), and
others (Beltagy et al., 2019; Wadden et al., 2020; Aly
et al., 2021). Extracting domain specific information
from maintenance documents has received limited at-
tention (Dixit et al., 2021; Sharp et al., 2017).
NLP is an approach to extract information from
text written by humans. Approaches applied in recent
NLP systems are grouped into two categories: Rule-
Based Methods (RBM) and Learning-Based Methods
(LBM). RBM refers to the modeling where the rela-
tionships and patterns in data are defined by the hu-
man while in LBM, these relationships and patterns
are figured out and trended out by the machine. RBM
is interpretable and suitable for rapid development
and domain transfer, and requires pre-defined vocabu-
laries (Adnan and Akbar, 2019; Valenzuela-Esc
´
arcega
et al., 2015; Patel and Tanwani, 2019). In general, the
performance of LBM is better in terms of precision
and recall but appropriate feature selection is impor-
tant. In addition, generating training data is a time-
consuming task.
Xu et al. (Xu et al., 2017b) recognized medi-
cal concepts and terminology such as diseases, drugs,
treatments, or procedures from unstructured medical
text. They used the bidirectional Long-Short Term
Memory (biLSTM) and conditional random fields to
identify medical named entity relies on character-
based word representations learned from the super-
vised corpus. Using another approach, Beltagy et al.
(Beltagy et al., 2019) introduced SciBERT which is
based on BERT (Devlin et al., 2018) but retrained on
a large scientific paper corpus. The authors showed
that their model is effective on NLP tasks such as
sequence tagging, sentence classification and depen-
dency parsing.
In the domain of maintenance, Dixit et al. (Dixit
et al., 2021) proposed a method to extract entities of
interest from maintenance records based on their ex-
tends on an existing domain dictionary. This dictio-
nary includes lists of components, positions, observa-
tions, and actions and is used to identify correspond-
ing elements in an input sentence. Their approach got
preliminary result, which mostly depends on the qual-
ity of the domain dictionary. Sharp et al. (Sharp et al.,
2017) introduced a proof-of-concept pipeline combin-
ing machine learning and natural language processing
techniques to cluster and tag maintenance data. They
achieved the accuracy around 70% when categorizing
and labelling a free form maintenance log entry from
a set of known labels.
The above approaches neither target to our objec-
tives nor are available online for public use. Thus, in
our work, we only consider NLP tools that are effi-
cient and available for use as follows.
Stanford CoreNLP (Singh et al., 2013): allows
users to perform a variety of NLP tasks, such as
part-of-speech tagging, tokenization, or named en-
tity recognition. The advantages of this tool are the
scalability and optimization for speed, making this
tool relevant for processing large amounts of data,
and performing complex operations. Spacy (Honni-
bal and Montani, 2017) : Spacy offers components for
uses an object-oriented approach to NLP handling and
supports pre-trained statistical neural network mod-
els and word vectors. Comparing to NLTK and other
libraries, Spacy well interfaces with all major deep
learning frameworks and performs faster; however,
it lacks flexibility and does not support many lan-
guages. Natural Language Toolkit - NLTK (Honnibal
and Montani, 2017) supports common tasks in NLP
by offering a model trained on a wide range of cor-
pora and lexical resources. It processes and presents
all data in form of strings and does not support object-
oriented. One of the disadvantages of NLTK is that it
requires significant resources and time when perform-
ing on massive amount of data (Al Omran and Treude,
2017).
As showed in previous research (Honnibal and
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
170

Montani, 2017; Al Omran and Treude, 2017), when
evaluating these NLP tools in various data collections
such as Java API documents, Stack Overflow, and
Github Readme files, Spacy makes the best perfor-
mance for NLP tasks, especially for POS tagging. In
addition, Spacy supports transfer learning, which can
be used to import knowledge from annotated exam-
ples into the pipeline to improve its efficiency. Thus,
we choose Spacy POS tagging to identify actions, re-
lated objects, warnings, guides, and other information
in aircraft maintenance text.
2.2 Semantic Representation of
Maintenance Procedures
To develop our generic ontology helping capture all
necessary knowledge from which illustrations can be
generated, we examine three types of ontologies.
Ontologies for Procedure: A maintenance text
can be viewed as a set of maintenance procedures,
that are our main objects of study. We consider a
maintenance procedure as a particular procedure for
doing something involving one or more steps or op-
erations. In the literature, there exist many ontolo-
gies describing such a procedure. For business pro-
cess modeling, the Business Process Model and No-
tation (BPMN) is a widely used standard. The spec-
ifications gave rise to the construction of ontologies,
such as the BPMN Ontology (Rospocher et al., 2014)
or BPMN 2.0 Ontology (Natschl
¨
ager, 2011). For in-
dustrial procedure modeling, Karray et al. (Karray
et al., 2012) has introduced an ontology for industrial
maintenance, and Chungoora et al. (Chungoora et al.,
2013) introduced an ontology for the manufacturing
process. Besides, there are several works inspired by
the Process Specification Language (PSL), a frame-
work to describe the structure of process executions
such as that of (Gr
¨
uninger, 2009). Ontologies for pro-
cedures can be built based on ISO specifications, as
presented in (Fraga et al., 2018). Except for the last
one, these models are complex and only the fragment
of them that deals with process description is related
to our study. Furthermore, the proposed models (and
their submodels) aren’t adapted to our needs. One so-
lution is to reuse and extend the models as introduced
in (Annane et al., 2019) or (Tarbouriech et al., 2021).
Domain Ontologies: We next examine the do-
main ontologies representing the knowledge of the
domain where the maintenance is applied on. In
the aeronautical field, to our best knowledge, there’s
only an ontology, called Aircraft Ontology (Ast et al.,
2014), which is reused in (Stefanidis et al., 2020).
However, the ontology is not well-structured and does
not cover all aircraft components we need.
Ontologies for Graphics: To represent knowl-
edge about graphics, an ontology is required. As one
of the earlier works, Niknam et al. (Niknam and
Kemke, 2011) presented an ontology for basic com-
puter graphics and geometric shapes. There are af-
terward several studies on iconic ontologies as intro-
duced in (Kuicheu et al., 2012; Ma and Cahier, 2014;
Lamy and Soualmia, 2017). The work of (Lamy and
Soualmia, 2017) is the most relevant to our context
as the authors propose an iconic ontology that plays
the main role and is linked with the domain ontology
through a mapping ontology.
2.3 Generating Graphics from
Knowledge Base
In the literature, a common method of generating
graphics from information, specifically from a knowl-
edge base, is to rely on SVG graphics which is an
XML-based markup language for describing vector
graphics based on two dimensions. This method has
been used in several works such as the automatic gen-
eration of maps (Ipfelkofer et al., 2006), simulation
of 2D models (Lehtonen and Karhela, 2006), gener-
ation of medical icons (Lamy et al., 2008) or gener-
ation of traditional medicine recipes (Kouame et al.,
2020). Although several works have been proposed in
this area, no application or source code has been pub-
licly available. Very few studies have been invested
on graph generation methodology. (Kouame et al.,
2020) drew pictograms using Inkscape, and automat-
ically generated icons and recipes as SVG image files
from ontology knowledge, using Python scripts.
3 METHODOLOGY
In this section, we present our proposed semantic-
centric approach to simplify maintenance processes
by graphical representation. We first use NLP tech-
niques to extract core information from maintenance
documents, and then use these pieces of information
to populate our ontology. We also use some external
resources to improve our knowledge base and inte-
grate images corresponding to each aircraft compo-
nent. Finally, we illustrate a graphical representation
of maintenance processes using information from our
built knowledge base.
Our approach can be divided into three steps pre-
sented in Figure 1.
A Semantic Approach for Generating Graphical Representation from Aircraft Maintenance Text
171

Figure 1: The workflow of our approach to generate graphical representation from maintenance text.
3.1 Extracting Core Information from
Aircraft Maintenance Text
We conducted experiments and evaluated our model
on the Aircraft Maintenance Manual (AMM) for WT9
Dynamic LSA issued by Aerospool on 22 May 2017
2
,
and it is freely available and usable for all applica-
tions.
3.1.1 Tools and Techniques
In this work, we target to an NLP tool which is flex-
ible enough to be tailored to suit our demands and
requirements for analyzing the aircraft maintenance
text. In addition, the tool should be high perfor-
mance and effective. Therefore, we choose Spacy
(Honnibal and Montani, 2017) which is showed as the
most effective NLP tools when evaluated on several
data collections compared to the other current open-
source tools (Al Omran and Treude, 2017; Bird et al.,
2009). When applying Spacy to analyse aircraft main-
tenance, we tailor the model to fulfil our requirements
such as retraining the model and applying rules to re-
tokenize text complement to the Spacy tokenization.
3.1.2 Extracting Core Information from Text
Our objective is to extract core pieces of information
from aircraft maintenance texts. Each maintenance,
that we call process, is assumed to be structured in a
format as illustrated in Figure 2.
Each process includes several pieces of informa-
tion such as identification number, name, description,
warnings, cautions, notes, and steps of instruction.
We identify and extract these pieces of information
as follows:
• The process identification and name are included
in the first line of each process. The identification
2
https://www.aerospool.sk/downloads/RTC/AS-AMM-
01-000 I1 R1 20180202.pdf
Figure 2: The process of ‘Removing transponder antenna’.
is the first part of the line which includes numbers
connected to each other by ’-’ while the name is
the rest part of the first line.
• The warnings, cautions, and notes of each process
are recognized by the paragraph right after corre-
sponding key words such as Warnings, Cautions,
and Notes. If there is no keyword specified, we
consider paragraphs which contain words in a list
(’must’, ’never’, ’do not allow’, ’ensure’, ‘guar-
antee’, ’should’, ’have to’, ’need’) as the warning
of the process.
• The description of the process is the paragraph
which appears after warnings, cautions, notes (if
any) and contains words to guide how to perform
process name.
• The steps of instruction are the rest of the pro-
cess text. Each step starts by an ordinary number
and contain at least one sentence. Each sentence
is considered as one task. Each task can have
one or several subtasks (clauses) which include at
least one action and one related component. Each
task/subtask points to its previous task/subtask.
Our main objective is extracting core information
from steps. For each step in a maintenance process,
we identify actions, related components, and informa-
tion related to each component such as the status, di-
rection, and position. In addition, we also recognize
corresponding warnings and cautions of each instruc-
tion step.
All these elements are recognized by Spacy with
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
172

our own customization and adaptation. We base on
the POS, head text, and children text of each token in
an instruction step to determine the classification (ac-
tion, component...) of each token or group of token.
Actions: An action is the main verb of the sen-
tence/clause which is recognized by the NLP tool. We
hypothesize that the subtask A is the ’previous sub-
task’ of the subtask B if the main verb in the sub-
task A is the head text of the main verb in the sub-
task B. As shown in the analysis of the step 4 of the
process 34-50-11 illustrated in Figure 3, the action of
the subtask 1 is ’unscrew’ while the action of sub-
task 2 is ’remove’. The head text of ’remove’ is ’un-
screw’ thus the subtask 1 is the ’previous subtask’ of
the subtask 2.
Components: Components related to a certain ac-
tion in a subtask is the nouns of which their head text
is that action. These nouns are recognized by Spacy
with our customization and adaptation. In Figure 3,
the component related to the action ’unscrew’ is ’the
nut’ while ’the washer’ and ’labels’ are components
related to ’remove’.
Information Related to a Certain Component:
We identify information related to a certain compo-
nent such as the status, the position, the direction,
and other quantity. These elements are recognized as
adverb or adverb positions by Spacy and then they
are checked with our predefined lists to determine
the corresponding information related to components.
Our predefined list of words describing the position
of a component is (’among’, ’around’, ’behind’, ’be-
neath’, ’between’, ’by’, ’in’, ’into’, ’inside’, ’near’,
’next to’, ’on’, ’over’, ’across’, ’below’, ’above’,
’against’, ’under’, ’beside’, ’in front of’, ’through’,
’underneath’) while the list of words describing the
status of a component is (’ON’, ’OFF’) and the list
of words describing the direction of a component is
(’to’, ’from’). The number of each object is stored in
the ’quantity’. In the step 4 of the process 34-50-11,
the ’transponder’ is pulled out ’from’ (the direction)
’the fuselage bottom part’.
Warning and Guide: We identify warnings in a
subtask by checking if the subtask contains gerund
verb in our predefined list. The warning phrase will
start from the gerund verb to the end of the sen-
tence/clause. Guides is specified in a similar way to
warnings considering gerund verbs that are not in the
gerund warning verb list. In the step 4 of the process
34-50-11 (Figure 2, the phrase ’considering the sur-
face’ is specified as a general ’guide’ for this step.
The analysis of the step 4 of the process 34-50-11
(’Unscrew the nut; remove the washer, label; and pull
out the transponder antenna from the fuselage bottom
part.’) is presented in the Figure 3.
Figure 3: The analysis of the task 4 in the ‘34-50-11 Re-
moving transponder antenna’ process.
While using Spacy to identify actions (verbs),
components (nouns), and other information, we find
that this tool works well on technical documents,
specifically on aircraft maintenance manual. How-
ever, there still exists a number of false positive (FP)
and false negative (FN). We deal with this issue by
customizing the Spacy model. We first create a list
of aircraft components which is constructed and ver-
ified by experts in this domain. This list is then used
to retrain the Spacy to recognize components that the
original Spacy can not. This list could also be used
to verified components recognized by Spacy to elim-
inate wrong recognition. In addition, we retokenize
and apply some matched rules to improve the Spacy
performance.
3.1.3 Storing the Core Information Extracted
from Maintenance Text
{
” i d p r o c e s s ” : ” wt9 −34 − 50 −11” ,
” i d s t e p ” : wt9 −4−S4 ,
” t a s k s ” : [
{
” id ” : ” wt9 −T4 ” ,
” s u b t a s k s ” : [
. . . .
{
” id ” : ” wt9 −T4 −3” ,
” a c t i o n ” : [ ” p u l l o u t ” ] ,
” o b j e c t s ” : [
{
” id ” : ” 5 ” ,
” name ” : ” t r a n s p o n d e r a n t e n n a ” ,
” d i r e c t i o n ” : ” fr om ” ,
” p o s i t i o n ” : ” ” ,
” s t a t u s ” : ” ” ,
” q u a n t i t y ” : ” ”
} ,
{
” id ” : ” 1 0 ” ,
” name ” : ” f u s e l a g e b o t t o m p a r t ” ,
” d i r e c t i o n ” : ” ” ,
” p o s i t i o n ” : ” ” ,
” s t a t u s ” : ” ” ,
” q u a n t i t y ” : ” ”
}
] ,
” w a r n i n g ” : ” ” ,
” g u i d e ” : ” ” ,
” p r e v i o u s s u b t a s k ” : ” wt9 −T4 −2”
}
] ,
” p r e v i o u s t a s k ” : ” 3 ”
}
]
}
Listing 1: An excerpt of the output in json given by the
analysis of task 4 of the process 35-50-11.
After analyzing the maintenance text, we store the
output in JSON structure as shown in the Code 1.
Each step has an ID, a name, a list of tasks, and point
A Semantic Approach for Generating Graphical Representation from Aircraft Maintenance Text
173

to a previous task. The previous task of a task can be
itself if the current task is the first one of the process.
For each task, we store an Id, a name, a list of sub-
tasks, and a previous task ID. In the substask, pieces
of information included are the identification, the list
of actions, the related objects, the warning, and the
guide. The ID, name, direction, position, status, and
quantity are identified for all objects related to a cer-
tain action. The JSON output of the analysis for the
step 4, process 34-50-11 is illustrated by the Code 1.
3.2 Building Knowledge Base
3.2.1 Ontology Development
We try to propose a more generic ontology as possible
to apply our approach to the aeronautic maintenance
domain. As maintenance documents of this domain
aren’t always publicly available, we developed our
ontology based on an XA41/XA42 AMM, a WT9 Dy-
namic LSA AMM, and some internal Airbus AMMs.
Figure 4 depicts our ontology composed of two parts:
1. Maintenance Procedure Description: To de-
scribe maintenance procedures, we consider the
following concepts:
• Process, task, and subtask: The Process class
represents the general maintenance procedure
that is composed of successive tasks. A Task
can contain in turn subtasks (Subtask) and can
refer to another process.
• Resources: Resource is held by an Enterprise.
It’s categorized as a Device or a Tool used to
perform a manual operation (an Act) on a Com-
ponent or a part of a component (Component
Part).
• Context: This abstract class represents addi-
tional information about a process or task. It
could be a warning, a preliminary condition for
action, a state to switch to, or the precise posi-
tion of the component.
2. Graphical Information: To represent informa-
tion about graphical objects, we use the Graphic
class that is specialized by Resource Graphic and
Context Graphic. hasGraphic is introduced as
an annotation property so that individual of Con-
text and concepts representing a resource can be
linked a Graphic.
3.2.2 Knowledge Base Construction
We integrate several resources to enrich our knowl-
edge base as follows:
Aircraft Maintenance Procedures: These pieces
of information were extracted using the NLP tech-
niques presented in Section 3.1. Extracted concepts
could be enriched by many ways: using dictionary,
such as WordNet
3
(for common nouns) or VerbNet
4
(for actions); using open data, such as DBPedia
5
or
Wikidata
6
(for named entity); or by domain experts.
This part is under investigation. So far we applied
WordNet and VerbNet while the other resources will
be left for future work.
Aircraft Components: Aircraft components are
presented by concepts belonging to the domain ontol-
ogy and are validated by experts to enrich the vocab-
ulary.
Tools and Devices: We have populated a prelim-
inary dataset of popular tools and devices used for
maintenance tasks.
Graphics: Aircraft components are drawn by spe-
cialists or given by enterprises. Information of inter-
val objects inside the graphics and the corresponding
aircraft components are imported. In addition, we im-
ported icons for the context (warnings, cautions or
notes) and tools.
Figure 5 represents an excerpt of the knowledge
graph describing the task 4 of the process, using the
developed ontology. A number of classes have been
populated beforehand, for example, action verbs like
Install and Pull, that have Act as parent class; or air-
craft components like WT9 Antenna and WT9 Fuse-
lage, that specialize the Component class.
3.3 Generating Illustration
The generating process is inspired from (Kouame
et al., 2020). As presented, a resource or a con-
text can have a graphic representation based on that
we can generate the corresponding illustration for a
maintenance task. Regarding Figure 6, the first im-
age (Init) shows our origin SVG image (wt9 uc.svg),
inside each component has a proper ID correspond-
ing to the one of the knowledge graph. Thanks to the
identifier and ordering, components can be removed
(or hidden) or even animated. Furthermore, an ar-
row can be also added to represent the direction of the
maintenance action (currently only up and down), as
demonstrated by the next image. In this manner, the
whole maintenance process can be demonstrated by a
sequence of illustrations, each of which corresponds
to a particular task.
3
https://wordnet.princeton.edu/
4
https://verbs.colorado.edu/verbnet/
5
https://www.dbpedia.org/
6
https://www.wikidata.org/
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
174

Figure 4: An ontology representing maintenance procedures.
Figure 5: An excerpt of the knowledge graph describing a
task of the maintenance process.
Figure 6: Graphical representation of Subtask 4-1 in the
process 34-50-11.
4 CONCLUSION
In this paper, we introduce a method to simplify air-
craft maintenance processes combining NLP tech-
niques, knowledge base, and graphical representation.
We first extract actions, components and related in-
formation in processes and then use these results as
an input to populate our ontology. We also use exter-
nal resources to improve our knowledge base. In ad-
dition, we integrate images corresponding to aircraft
components into the ontology and use them to gen-
erate a graphical representation for each task in the
processes.
In the future, we have a plan to build a ground
truth on a big dataset and use it to evaluate our NLP
method. In addition, we will improve our knowledge
base by apply more external resources and by having
validation from domain experts.We would also want
to construct a collection of images corresponding to
all aircraft components. This will help us effectively
generate graphical representation of tasks in mainte-
nance process.
We suppose that the approach of simplifying
maintenance processes we built have a broad range of
applications in several industrial domains such as car
maintenance, ship maintenance or mechanic machine
maintenance.
REFERENCES
Adnan, K. and Akbar, R. (2019). An analytical study of
information extraction from unstructured and multidi-
mensional big data. Journal of Big Data, 6(1).
Al Omran, F. N. A. and Treude, C. (2017). Choosing an nlp
library for analyzing software documentation: a sys-
tematic literature review and a series of experiments.
In 2017 IEEE/ACM 14th international conference on
mining software repositories (MSR). IEEE.
Aly, R., Guo, Z., Schlichtkrull, M., Thorne, J., Vlachos,
A., Christodoulopoulos, C., Cocarascu, O., and Mit-
tal, A. (2021). Feverous: Fact extraction and verifi-
cation over unstructured and structured information.
arXiv preprint arXiv:2106.05707.
Annane, A., Aussenac-Gilles, N., and Kamel, M. (2019).
Bbo: Bpmn 2.0 based ontology for business pro-
cess representation. In 20th European Conference on
Knowledge Management (ECKM 2019).
Ast, M., Glas, M., Roehm, T., and Luftfahrt, V.
(2014). Creating an ontology for aircraft de-
sign. Deutsche Gesellschaft f
¨
ur Luft-und Raumfahrt-
Lilienthal-Oberth eV.
Beltagy, I., Lo, K., and Cohan, A. (2019). Scibert: A
pretrained language model for scientific text. arXiv
preprint arXiv:1903.10676.
Bird, S., Klein, E., and Loper, E. (2009). Natural language
processing with Python: analyzing text with the natu-
ral language toolkit. ” O’Reilly Media, Inc.”.
Chungoora, N., Young, R. I., Gunendran, G., Palmer, C.,
Usman, Z., Anjum, N. A., Cutting-Decelle, A.-F.,
A Semantic Approach for Generating Graphical Representation from Aircraft Maintenance Text
175

Harding, J. A., and Case, K. (2013). A model-driven
ontology approach for manufacturing system interop-
erability and knowledge sharing. Computers in indus-
try.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2018). Bert: Pre-training of deep bidirectional trans-
formers for language understanding. arXiv preprint
arXiv:1810.04805.
Dixit, S., Mulwad, V., and Saxena, A. (2021). Extracting
semantics from maintenance records. arXiv preprint
arXiv:2108.05454.
Fraga, A. L., Vegetti, M., and Leone, H. P. (2018). Se-
mantic interoperability among industrial product data
standards using an ontology network. In ICEIS (2).
Gao, S., Kotevska, O., Sorokine, A., and Christian, J. B.
(2021). A pre-training and self-training approach for
biomedical named entity recognition. PloS one.
Ge, S., Wu, F., Wu, C., Qi, T., Huang, Y., and Xie, X.
(2020). Fedner: Privacy-preserving medical named
entity recognition with federated learning. arXiv
preprint arXiv:2003.09288.
Gr
¨
uninger, M. (2009). Using the psl ontology. In Handbook
on Ontologies. Springer.
Honnibal, M. and Montani, I. (2017). Natural language
understanding with bloom embeddings, convolutional
neural networks and incremental parsing. Unpub-
lished software application. https://spacy. io.
Ipfelkofer, F., Lorenz, B., and Ohlbach, H. J. (2006).
Ontology driven visualisation of maps with svg-an
example for semantic programming. In Tenth In-
ternational Conference on Information Visualisation
(IV’06). IEEE.
Karray, M. H., Chebel-Morello, B., and Zerhouni, N.
(2012). A formal ontology for industrial maintenance.
Applied ontology.
Kouame, A., Brou, K. M., Lo, M., and Lamy, J.-B. (2020).
Visual representation of african traditional medicine
recipes using icons and a formal ontology, ontomed-
trad. In MIE.
Kuicheu, N. C., Wang, N., Tchuissang, G. N. F., Siewe,
F., and Xu, D. (2012). Description logic based icons
semantics: An ontology for icons. In 2012 IEEE 11th
International Conference on Signal Processing. IEEE.
Lamy, J.-B., Duclos, C., Bar-Hen, A., Ouvrard, P., and
Venot, A. (2008). An iconic language for the graphi-
cal representation of medical concepts. BMC medical
informatics and decision making.
Lamy, J.-B. and Soualmia, L. F. (2017). Formalization
of the semantics of iconic languages: An ontology-
based method and four semantic-powered applica-
tions. Knowledge-Based Systems.
Lehtonen, T. and Karhela, T. (2006). Ontology approach
for building and visualising process simulation mod-
els using 2d vector graphics. In SIMS Proceedings
of the 47th Conference on Simulation and Modeling.
Finnish Society of Automation, SIMS-Scandinavian
Simulation Society.
Lopes, F., Teixeira, C., and Oliveira, H. G. (2019). Con-
tributions to clinical named entity recognition in por-
tuguese. In Proceedings of the 18th BioNLP Workshop
and Shared Task.
Ma, X. and Cahier, J.-P. (2014). Graphically structured
icons for knowledge tagging. Journal of information
science.
Natschl
¨
ager, C. (2011). Towards a bpmn 2.0 ontology. In
International Workshop on Business Process Model-
ing Notation. Springer.
Nayel, H. A., Shashirekha, H., Shindo, H., and Mat-
sumoto, Y. (2019). Improving multi-word entity
recognition for biomedical texts. arXiv preprint
arXiv:1908.05691.
Niknam, M. and Kemke, C. (2011). Modeling shapes and
graphics concepts in an ontology. In SHAPES.
Patel, R. and Tanwani, S. (2019). Application of machine
learning techniques in clinical information extraction.
In Smart Techniques for a Smarter Planet. Springer.
Rospocher, M., Ghidini, C., and Serafini, L. (2014). An
ontology for the business process modelling notation.
In FOIS.
Sharp, M., Sexton, T., and Brundage, M. P. (2017). Toward
semi-autonomous information. In IFIP International
Conference on Advances in Production Management
Systems. Springer.
Singh, J., Joshi, N., and Mathur, I. (2013). Development
of marathi part of speech tagger using statistical ap-
proach. In 2013 International Conference on Ad-
vances in Computing, Communications and Informat-
ics (ICACCI). IEEE.
Stefanidis, D., Christodoulou, C., Symeonidis, M., Pallis,
G., Dikaiakos, M., Pouis, L., Orphanou, K., Lam-
pathaki, F., and Alexandrou, D. (2020). The icarus
ontology: A general aviation ontology developed us-
ing a multi-layer approach. In Proceedings of the 10th
International Conference on Web Intelligence, Mining
and Semantics.
Tarbouriech, C., Bernard, D., Vieu, L., Barton, A., and
´
Ethier, J.-F. (2021). Flight procedures description us-
ing semantic roles. In CEUR Workshop Proceedings.
Valenzuela-Esc
´
arcega, M. A., Hahn-Powell, G., Surdeanu,
M., and Hicks, T. (2015). A domain-independent rule-
based framework for event extraction. In Proceedings
of ACL-IJCNLP 2015 System Demonstrations.
Wadden, D., Lin, S., Lo, K., Wang, L. L., van Zuylen,
M., Cohan, A., and Hajishirzi, H. (2020). Fact or
fiction: Verifying scientific claims. arXiv preprint
arXiv:2004.14974.
Xu, K., Zhou, Z., Hao, T., and Liu, W. (2017a). A bidi-
rectional lstm and conditional random fields approach
to medical named entity recognition. In International
Conference on Advanced Intelligent Systems and In-
formatics. Springer.
Xu, K., Zhou, Z., Hao, T., and Liu, W. (2017b). A bidi-
rectional lstm and conditional random fields approach
to medical named entity recognition. In International
Conference on Advanced Intelligent Systems and In-
formatics. Springer.
Yang, T., He, Y., and Yang, N. (2022). Named entity recog-
nition of medical text based on the deep neural net-
work. Journal of Healthcare Engineering, 2022.
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
176