Ontology Learning from Clinical Practice Guidelines

Samia Sbissi

1 a

, Mariem Mahfoudh

2,3 b

and Said Gattouﬁ

1 c

SMART Laboratory, Tunis University, Tunis, Tunisia

MIRACL Laboratory, University of Sfax, Sfax, Tunisia

ISIGK, University of Kairouan, Kairouan, Tunisia

Keywords:

Ontology Learning, Ontology Enrichment, SWRL, Word2Vec.

Abstract:

In order to assist professionals and doctors to make decisions about appropriate health care for patients who are

at risk of cardiovascular disease, we propose a decision support system based on OWL (Ontology Language

Web) ontology with SWRL (semantic web rule language) rules. The idea consists to parse clinical practice

guidelines (i.e. documents that contain recommendations and medical knowledges) to enrich and exploit

existing cardiovascular domain ontology. The enrichment process is conducted by ontology learning task. We

ﬁrst pre-process the text and extract the relevant concepts. Then, we enrich the ontology not only by OWL DL

axioms, but also SWRL rules. To identify the similarity between terms texts and ontology concepts, we have

used a combination of methods as levenshtein similarity and Word2Vec.

1 INTRODUCTION

The clinical guidelines (CG) contain a set of recom-

mendations and knowledge used to guide health pro-

fessionals in making appropriate decisions and im-

proving the quality of care. Although health profes-

sionals are familiar with these guides, text-based ver-

sions have several limitations (Cabana et al., 1999;

Francke et al., 2008). The text of the recommenda-

tions may contain undeﬁned terms in well-accepted

terminology, and unclear sentences which may lead

to an ambiguous interpretation of the content. The

size/complexity of the recommendations may be an

obstacle also in the sense that it may hide relevants

informations or discourage specialists from reading

all the document (Bonacin et al., 2013). In order to

deal with these problems, some tools were proposed

to code the clinical guidelines and to create computer-

interpretable guidelines (CIG) and there is an increas-

ing demand to convert this unstructured information

into structured information.

Ontology plays a key role in representing the

knowledge hidden in texts and makes it human and

computer understandable. An ontology is a formal

and structural way of representing the concepts and

https://orcid.org/0000-0002-5301-5156

https://orcid.org/0000-0001-7860-8604

https://orcid.org/0000-0001-7914-6165

relations of a shared conceptualization. More pre-

cisely, it can be deﬁned as concepts, relations, at-

tributes and hierarchies present in the domain. On-

tologies can be created by extracting relevant in-

stances of information from text using a process

called ontology population. However, handcrafting

such big ontologies is a difﬁcult task, and it is im-

possible to build ontologies for all available domains

(Asim et al., 2018). Therefore, instead of handcraft-

ing ontologies, research trend is now shifting toward

automatic extract ontology from the text that is de-

ﬁned as an ontology learning process (Maedche and

Staab, 2001).

The process of ontology learning begins with the

extraction of terms and their synonyms from the

text. The corresponding terms and synonyms are con-

verted to the form of concepts. Then, taxonomic and

non-taxonomic relations between these concepts are

found. Finally, axiom schemata are instantiated and

general axioms are extracted from unstructured text.

This whole process is known as ontology learning

layer cake (Gardent and Mahfoudh, 2016).

In our work, we address the problem of ontology

learning that could be applied to the analyzed medical

text, in order to enrich an existing ontology. The

enriched ontology aims to infer and produce a recom-

mendation task. To this end, we collaborate with the

hospital of the "Rabta" (Tunis) to make an assistance

system that helps doctors to make decisions about

312

Sbissi, S., Mahfoudh, M. and Gattouﬁ, S.

Ontology Learning from Clinical Practice Guidelines.

DOI: 10.5220/0008169903120319

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 312-319

ISBN: 978-989-758-382-7

patients who are at risk of cardiovascular disease,

especially aortic dissection. The aortic dissection

is a partial disruption of the wall of the aorta that

may at any time evolve towards a complete rupture,

with consequent death (Criado, 2011). One of the

ontologies we found close to our domain is the

CVDO ontology

. It is an owl ontology, designed

to describe the entities related to cardiovascular

diseases. This ontology will be learnt and enriched

from a text called clinical practice guidelines (Erbel

et al., 2014). It is an evolving reference document

that contains a set of recommendations which

aim to assist professionals to master a medical

domain. Recommendation example: In patients

with an abdominal aortic diameter of

25-29 mm, new ultrasound imaging should

be considered 4 years later). Our goal is to

transform these recommendations into logical forms,

more precisely into Semantic Web Rule Language

(SWRL) rules. SWRL (Horrocks et al., 2004) is a

semantic web language, which is integrated directly

within OWL (Ontology web language) ontologies.

It allows deﬁning rules in the form of logical impli-

cations between conditions and conclusions. This

transformation will be conducted by the ontology

learning process. We think that the ontology learning

process will help increase the number of transformed

rules from text to ontology (Sbissi et al., 2019) due

to the lack of concepts and relations in the initial

ontology.

The remainder of this paper is organized as fol-

lows. Section two gives a summary of related work.

Section three presents our approach. Section four

presents the implementation and results. Finally, we

conclude and we give some future work.

2 RELATED WORK

Studies done by (Séroussi et al., 2010) highlight

physicians low adherence to the text-based guide-

lines (27.2%), and high adherence to the evaluated

electronic version of the guidelines (86.1%). They

proposed a system called ASTI-GM that has been

designed to be computer-based thinking support on

how to decide, is a demand guideline-based CDSS

where the user interactively characterizes her patient

by browsing the system knowledge base to obtain

the recommended treatment. The translation from

text-based guidelines to computer-interpret-able one

requires a well-deﬁned description language. Sev-

eral formalization techniques and methodologies have

http: //purl.bioontology.org/ontology/CVDO

been proposed in the literature. One of the most used

is the ontology. An important beneﬁt of ontology is

the ability to specify axioms for reasoning.

The mapping process from text is called ontol-

ogy learning and it can be of three types : man-

ual, semi-automatic and automatic. In the manual

method, an ontology is constructed from the scratch

by domain experts and knowledge engineers using

the most painstaking procedures (Maedche, 2013). In

the semi-automatic method, the domain experts and

trained users use semi-automatic prototype. For ex-

ample, (Dramé et al., 2014) build a semi automatic-

multilingual domain ontology using UMLS Metathe-

saurus and parallel corpus. However, these meth-

ods are time-consuming and require domain experts.

Consequently, the automatic method of ontology

learning is becoming a major trending. Several sys-

tems are proposed : Text-to-Onto (Cimiano et al.,

2009), OntoGain (Drymonas et al., 2010), etc.

There are three major approaches for ontology

learning that are often used: statistical methods (e.g.

C/CN value, T F − IDF, word2vec etc.), machine

learning methods, and linguistic approaches (e.g.

POS patterns, parsing, WordNet, discourage anal-

ysis, etc.). Authors in (Wohlgenannt, 2015) have

built an ontological learning system by collecting ev-

idence from heterogeneous sources in a statistical ap-

proach. The candidate concepts were extracted and

the "is a" type of relations was constructed by us-

ing chi − square" co-occurrence signiﬁcance score.

(Doing-Harris et al., 2015) made use of cosine sim-

ilarity, T F −IDF, also called C −value statistic, and

POS to extract the candidate collocates for construct-

ing an ontology.

In (Jiang and Tan, 2010; Wong et al., 2012),

authors used statistical methods to extract concepts

by computing the relevance of document words

based on term frequency-inverse document frequency

(TD/IDF) and similar measures. Often these methods

are combined.

In the medical ﬁeld, there are a lot of non-

taxonomic relationships, such as symptoms and eti-

ologies of diseases, indications of medicines, aliases

of diseases or medicines, etc. Among them, ontol-

ogy learning from this type of text plays a big role

in using statistical and linguistic methods. (Mikolov

et al., 2013) propose a skip gram model implemented

in the word2vec system. The key idea is that words

with similar contexts should have similar meanings.

For example, if we see the two sentences "the pa-

tient complained of aortic dissection symptoms" and

"the patient reported aortic dissection symptoms", we

might infer that "complained" means the same thing

as "reported". As a result, these two words should be

Ontology Learning from Clinical Practice Guidelines

313

close in the representation space. (Minarro-Giménez

et al., 2014) learn to embed from unstructured med-

ical corpora crawled from PubMed, Merck Manuals,

Medscape and Wikipedia.

In this paper, we take this line of work further

by showing how to learn medical concept and rela-

tion from medical recommendation text. Speciﬁcally,

we show how to use a claims ontology consisting of

cardiovascular disease and recommendation text. We

show that with simple algorithmic adjustments, it is

possible to use the word2vec algorithm to learn em-

beddings on this type of longitudinal. In addition of

linguistic approach, mostly the extraction of concepts

was done together with the concept hierarchy extrac-

tion by looking for speciﬁc patterns in the texts. These

patterns included the Hearst patterns and other lexico-

syntactic patterns conducted by Pos tagging techni-

cal from NLP (Liu et al., 2011; Biemann, 2005).

We learn in this process new concepts and relations

that will be added to the ontology. The process of

adding new elements in ontology is called ontology

enrichment. All the previous work are interested to

extract new concepts and relations (taxonomic or se-

mantic) in order to enrich the ontology. However, in

our knowledge, very little works take into account to

enrich ontology by SWRL rules constructed after the

results obtained by ontology learning.

3 PROPOSED APPROACH

Our approach is described in ﬁg.1 and described be-

low.

3.1 Ontology Learning

Different sub-tasks are included in the ontology learn-

ing: relevant terminology extraction, synonym terms

identiﬁcation, concepts construction, concepts hierar-

chy organization, learning relations, relations hierar-

chy organization and axioms extraction (Asim et al.,

2018; Gyawali et al., 2017). To accomplish these sub-

tasks, we begin by the indispensable pre-processing

step.

Pre-processing. Our corpus is composed of a set of

medical recommendations deﬁned by the European

Society of Cardiology ESC. An example of a recom-

mendation is presented below.

"If the anatomy is favourable and the expertise

available, endovascular repair (TEVAR) should be

preferred over open surgery."

Figure 1: Approche overview.

To pre-process our corpus, we have used linguis-

tic technical as part of speech tagging, sentence pars-

ing and lemmatization which are linguistic-based pre-

processing techniques used in almost every ontology

learning methodology. More details can be ﬁnding in

our previous work presented in (Sbissi et al., 2019).

Term/Concept Extraction. Several approaches use

linguistic method to extract terms and concepts (Is-

mail et al., 2015; Panchenko et al., 2016; Atapattu

et al., 2017). The text is tagged with parts of speech

to extract syntactic structures in a sentence such as

noun phrases and verb phrases.

In (Sbissi et al., 2019), we have used Levenshtein

measure and WordNet ontology to search similarity

between text and the ontology. In order to amelio-

rate our results, we propose to integrate other linguis-

tic and syntactic methods. We also use a statistical

method, Word2vec. Word2vec (Mikolov et al., 2013)

computes continuous vector representations for large

text data-sets. It provides high performance for mea-

suring syntactic and semantic similarities.

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

314

Taxonomic Relations. Subsumption relations (also

known as "is-a" or inclusion relations) provide a tree

view of the ontology and determine inheritance be-

tween concepts.

3.2 Ontology Enrichment

Ontology enrichment consists of adding or modify-

ing the existing ontology by performing one or sev-

eral ontology learning tasks (Mahfoudh et al., 2013).

We are not interested in only simple concepts and tax-

onomic relations, but also in SWRL rules extracted

automatically from medical guidelines. The enrich-

ment process attempts to facilitate text understand-

ing and automatic processing of textual resources,

moving from words to concepts and relationships. It

starts by extracting concepts/relationships from plain

text using linguistic processing such as part-of-speech

(POS) tagging and phrase chunking. The extracted

concepts and relationships are then arranged in the

initial ontology, using syntactic and semantic analy-

sis techniques. The text contains a set of recommen-

dations. The following example presents one recom-

mendation and how it is treated.

"In all patients with AD, medical therapy including

pain relief and blood pressure control is

recommended."

⇓

[(’In’, ’IN’), (’all’, ’DT’), (’patients’, ’NNS’),

(’with’, ’IN’), (’AD’, ’NNP’), (’,’, ’,’), (’medical’,

’JJ’), (’therapy’, ’NN’), (’including’, ’VBG’),

(’pain’, ’NN’), (’relief’, ’NN’), (’and’, ’CC’),

(’blood’, ’NN’), (’pressure’, ’NN’), (’control’,

’NN’), (’is’, ’VBZ’), (’recommended’, ’VBN’)]

Let (NN) be a noun, (VB) be a verb, etc. The

evolution of our ontology CVDO is conducted not

only by some changes of the addition of elements like

(concepts, object properties and data properties) re-

sulting from the ontology learning process but also

by adding the swrl rules extracted from the text of

the recommendations. The following example is an

SWRL extracted from a recommendation.

In patients with abdominal aortic diameter of 25

to 29 mm, new ultrasound imaging should be

considered.

⇓

Patient(?p) ∧hasAbdominalDiametre(?p, ?d) ∧

swrlb : greaterT han(?d, 25) ∧swrlb :

lessT han(?d, 29)− >

recommendedDiagnosis(?p, ”ultrasounImaging”)

In our case, two big changes are replicated to the

ontology as it is illustrated in ﬁgure 2.

• OWL Changes: AddClass, AddSub-

Class, AddDataProperty, AddObjectProp-

erty,AddDataPropertyAssertion, AddObjectProp-

ertyAssertion, AddIndividual, etc.

• SWRL Changes: AddAtom, AddClassAtom,

SWRLBuiltIn, SWRLExpression, etc.

Figure 2: Changes replicated to ontology.

We can beneﬁt from the syntactic relations of the

terms extracted from the text to determine the type of

change that will be added to the ontology. Example

a concept should be a noun, object property and data

property are a verb. After parsing the text, the extrac-

tion of syntactic relations between the terms as well

as the part of speech tag is used. We focus on whether

each extracted term will be a candidate to enrich our

initial ontology or not. We use a similarity measure to

compare each extracted term to the content of the ini-

tial ontology. The most populate change of ontology

enrichment is the aid of a new concept in the ontology.

The following example explains the different steps of

this process.

Ontology Learning from Clinical Practice Guidelines

315

AddClass(Cc,CVDO):

Input:

• Cc: candidate concept to be added (Cc is a

noun(NN)).

• CVDO: our existing ontology.

• Recommandation.txt: the text of recommenda-

tion analysed and parsed.

Case 1:

If Cc does not exist in CVDO Then:

Search similarity between Cc and concepts of

CVDO: Levenshtein measure is used.

• If we obtain a similarity: sim(Cc]=Ci Then

: add semantic relation IsSimilar(Cc)=Ci or

AddLabel.

Case 2:

If Cc no exist in CVDO and no similarity obtained

with concepts of the ontology Then:

AddClass(Cc,CVDO).

An other changes conducted by constructing

SWRL is add a class of atom as represented by the

following steps:

AddClassAtom(C,var):

C should be a class existing in CVDO

• If var is a variable representing an OWL indi-

vidual Then

write C(?var).

• If var is a name of individual Then

write C(var).

3.3 Evaluation of Ontology Evolution

During the process of ontology enrichment guided by

ontology learning, we search to maintain the consis-

tency of the ontology. The managing and the evalua-

tion of the evolution of ontologies can be at different

layers (Petasis et al., 2011):

• Lexical, vocabulary or data layer. We focus on

which concepts and instances have been included

in the ontology and the vocabulary used to iden-

tify them.

• Relational layer: the relations between the con-

cepts of the ontology:

– Hierarchy, taxonomy: an ontology almost al-

ways includes hierarchical inclusion relations

between its concepts.

– Semantic relations: it concerns other relations

besides inclusion and can be evaluated sepa-

rately.

We propose in our approach to preserving consis-

tency: each transformation is deﬁned by a set of

negative application conditions (NAC) and derived

changes (DCH) (Mahfoudh et al., 2015). The ontol-

ogy inconsistencies treated by our work are:

• Data redundancy that can be generated following

and add or rename operation. This type of incon-

sistency is corrected by the NACs.

• Axioms contradiction, the addition of a new ax-

iom should not be accepted if it contradicts an ax-

iom already deﬁned in the ontology. Many cases

are considered: (1) two classes cannot be disjoint

and equivalent at the same time, (2) two classes

that share a subsumption relation cannot be dis-

joints, etc.

In some cases, if we apply a change to one ontology

entity, it can depend on other ontology elements. Re-

ferring to previous work in (Mahfoudh et al., 2015),

the Table 1 presents some changes and the ontol-

ogy concepts which are related. We used the follow-

ing vocabulary: (Class (C), Property (P), ObjectProp-

erty(OP), DataProperty(DP), Individual (I), DataType

(DT)).

Table 1: Dependancy between changes and ontology enti-

ties.

C I OP DP

AddConcept

√

AddIndividual

√ √

AddDataproperty

√

AddObjectProperty

√ √

AddSubClasses

√ √

AddObjectPropertyAssertion

√ √

AddDataPropertyAssertion

√ √

A number of changes could affect the ontology

when it is requested to be reﬂected in the existing on-

tology. Add concept is the most common change in

any ontology. New concepts emerge and have to be

accommodated in the concept hierarchy (subclass).

The addition of subclasses requires certain conditions

and generates changes at the ontology level.

3.3.1 Add Subclass

The AddSubClass (C1;C2) is deﬁned as follow (Mah-

foudh et al., 2015):

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

316

• Precondition: C1;C2, the classes should exist in

the ontology.

• Negative Changes(NCH):

1. C1 vC2: condition to avoid redundancy. If C1

already is a subclass of C2, we will not add it;

2. C2 v C1: the subsumption relation cannot be

symmetric;

3. C1 v ¬C2:classes which share a subsumption

relation cannot be disjoint;

4. ∃Ci ∈ C(O).(C1 v Ci) ∧(Ci v C2): if there is

a class Ci which is the subClassOf the class C2

and the superClass of C1, then, C1 is already a

subClass of C2;

5. ∃(Ci,C j) ∈ C(O).(Ci v C1) ∧ (C j v C2) ∧

(Ci v ¬C j): classes which share a subsump-

tion relation cannot have subClasses that are

disjoint;

• Results: C1 vC2 , the axiom will be added to the

ontology.

4 IMPLEMENTATION AND

RESULTS

We developed a Java-based implementation to test

our approach. Stanford CoreNLP is used for pre-

processing of text to determine Pos tagging and chun-

ked tree. In the process of ontology learning and en-

richment, we need to read from ontology and to write

or add elements in the ontology. For this task, we

used the Jena API. Our ontology CVDO contains ini-

tially 514 concepts and the recommendations text is

composed of 614 words.

4.1 Search Links

Search links between CVDO ontology and a pre-

processed text of clinical practice guideline refer to

the semantic annotation process. We used the con-

cepts names to produce an expanded list of equivalent

or related terms. Each term of the input text may be

associated with one or more entities from the ontol-

ogy. To ﬁnd the similarities, we have used (Sbissi

et al., 2019) :

1. exact matching: identiﬁes the identical entities

(String) in the text and in the domain ontology ;

2. morphological matching: identiﬁes the entities

with a morphological correspondence;

3. syntactical similarities: using Levenshtein mea-

sure (Levenshtein, 1966);

4. semantic matching: identiﬁes the synonyms rela-

tions with WordNet ontology.

We present in table 2 the result of links process.

Table 2: Search links.

number of links Links(%)

Initially 28 4.38%

With similarity 190 30%

Only 30% of similarities was extracted. The only

relation between text and ontology is "is-similar-to".

This type of relationship is sufﬁcient to an enrichment

task. We pass to the step 2 that’s ontology learning.

4.2 Ontology Learning

Term Extraction. Linguistic and syntactic analysis is

employed head-modiﬁer principal to identify and ex-

tract complex terms in which the head of the complex

term takes the role of hypernym. X is a hyponym of

Y if Y is a type of X. Example a dissection aortic is a

hyponym of dissection aortic type B.

we used linguistic features (POS, etc.) and word em-

bedding features (word2vec).

Table 3: Word2vec process.

Word2vec Word2vec

(unigram) (bigram)

Iteration1 71.8% 76.2%

IterationN 76.6% 81.8%

We keep the process for word2vec as simple as

possible. After word2vec model generation, we ﬁx

and apply the built-in word2vec similarity function to

get terms related to the seed terms.

In table 3, we iterate the algorithm, In the ﬁrst itera-

tion, the algorithm needs ﬁrst user intervention to re-

move from the result ﬁle all words that are far from

the domain. After the third iteration, the algorithm

offers automatically correspondence.

On the plus side, the word2vec implementation

is extremely simple and provides a high-percentage

of relevant concept candidates. On the minus side,

candidates suggested by word2vec are (as expected)

sometimes even too strongly related to the seed terms,

for example, syntactic variations such as plural forms

or near-synonyms.

Word2vec with bigram produce a better result. To ex-

plain it, let the same example of the rule:

"In all patients with AD, medical therapy in-

cluding pain relief and blood pressure control is

recommended."

*bigram[0,10]:(’in,’all’), (’all’,patients’),

..(”pain’,’relief’),..,(’blood’,’pressure’).

Ontology Learning from Clinical Practice Guidelines

317

In our case bigram could have good result because

the majority of concepts are composed noun. Also,

the same case to object property.

4.3 Ontology Enrichment

Thanks to the ontology learning technique, we were

able to extract concepts and relations between con-

cepts that were missing in the ontology. The Figure

3 represents some concepts, object property and dat

property that we managed to extract.

Figure 3: New elements conducting the ontology enrich-

ment.

The ontology enriched with the rules is shown in

Figure 4. We illustrate in this ﬁgure some enriched

concepts with relations.

Figure 4: Enriched ontology.

In the Figure 5, we can remark that we have two

recommendations in case of aortic dissection "AD".

Actually, these two rules belong to one recommenda-

tion. Our word2vec extraction process has considered

that surgery and urgent

urgey are distinct. So to im-

prove the results we want to use T F −IDF to measure

cosine similarity, T F −IDF, also called C −value.

Figure 5: Conlict rules.

5 CONCLUSION

The paper presented a method to automatically learn-

ing ontology from unstructured text, a clinical prac-

tice guidelines (CG) in order to enrich an existing on-

tology. CG presents a set of recommendations and

knowledge which aims to assist doctors to make de-

cisions about appropriate health care for patients who

are at risk of cardiovascular disease.

The ontology learning process starts with

analysing the text by the pre-processing technics

using the Stanford core NLP. Then, it passes to

relevant terminology extraction, synonym of terms

identiﬁcation, concepts construction, concept hi-

erarchy organization, learning relations, relations

hierarchy organization and axioms extraction. To

extract term/concept, we used Levenshtein mea-

sure, WordNet ontology and the statistical method

word2vec. For other relations we used the chunking

tree and hearst pattern to search hierarchic relations.

Once these elements are extracted, we have updated

the ontology by adding them. The ontological

enrichment process do not treat only OWL concepts

and axioms but also integrates SWRL rules which

will be used to reasoning tasks.

As a future work, we aim to use enriched ontol-

ogy and the SWRL rules to build a medical decision

support system for the cardiologists.

REFERENCES

Asim, M. N., Wasim, M., Khan, M. U. G., Mahmood, W.,

and Abbasi, H. M. (2018). A survey of ontology learn-

ing techniques and applications. Database, 2018.

Atapattu, T., Falkner, K., and Falkner, N. (2017). A com-

prehensive text analysis of lecture slides to generate

concept maps. Computers & Education, 115:96–113.

Biemann, C. (2005). Ontology learning from text: A survey

of methods. In LDV forum, volume 20, pages 75–93.

Bonacin, R., Pruski, C., and Da Silveira, M. (2013). Ar-

chitecture and services for formalising and evaluating

care actions from computer-interpretable guidelines.

International Journal of Medical Engineering and In-

formatics, 5(3):253–268.

Cabana, M. D., Rand, C. S., Powe, N. R., Wu, A. W.,

Wilson, M. H., Abboud, P.-A. C., and Rubin, H. R.

(1999). Why don’t physicians follow clinical practice

guidelines?: A framework for improvement. Jama,

282(15):1458–1465.

Cimiano, P., Mädche, A., Staab, S., and Völker, J. (2009).

Ontology learning. pages 245–267.

Criado, F. J. (2011). Aortic dissection: a 250-year perspec-

tive. Texas Heart Institute Journal, 38(6):694.

Doing-Harris, K., Livnat, Y., and Meystre, S. (2015). Au-

tomated concept and relationship extraction for the

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

318

semi-automated ontology management (seam) sys-

tem. Journal of biomedical semantics, 6(1):15.

Dramé, K., Diallo, G., Delva, F., Dartigues, J. F., Mouil-

let, E., Salamon, R., and Mougin, F. (2014). Reuse

of termino-ontological resources and text corpora for

building a multilingual domain ontology: an applica-

tion to alzheimer’s disease. Journal of biomedical in-

formatics, 48:171–182.

Drymonas, E., Zervanou, K., and Petrakis, E. G. (2010).

Unsupervised ontology acquisition from plain texts:

the ontogain system. pages 277–287.

Erbel, R., Aboyans, V., Boileau, C., Bossone, E., Bar-

tolomeo, R. D., Eggebrecht, H., Evangelista, A., Falk,

V., Frank, H., et al. (2014). 2014 esc guidelines on the

diagnosis and treatment of aortic diseases. European

heart journal, 35(41):2873–2926.

Francke, A. L., Smit, M. C., de Veer, A. J., and Mistiaen,

P. (2008). Factors inﬂuencing the implementation of

clinical guidelines for health care professionals: a sys-

tematic meta-review. BMC medical informatics and

decision making, 8(1):38.

Gardent, C. and Mahfoudh, M. (2016). Overview and com-

parison of existing deep semantic parsers. Technical

report, ModelWriter Project, LORIA-CNRS.

Gyawali, B., Shimorina, A., Gardent, C., Cruz-Lara, S., and

Mahfoudh, M. (2017). Mapping natural language to

description logic. In European Semantic Web Confer-

ence, pages 273–288. Springer.

Horrocks, I., Patel-Schneider, P. F., Boley, H., Tabet, S.,

Grosof, B., Dean, M., et al. (2004). Swrl: A semantic

web rule language combining owl and ruleml. W3C

Member submission, 21:79.

Ismail, R., Bakar, Z. A., and Rahman, N. A. (2015). Extract-

ing knowledge from english translated quran using nlp

pattern. Jurnal Teknologi, 77(19).

Jiang, X. and Tan, A.-H. (2010). Crctol: A semantic-based

domain ontology learning system. Journal of the

American Society for Information Science and Tech-

nology, 61(1):150–168.

Levenshtein, V. I. (1966). Binary codes capable of cor-

recting deletions, insertions and reversals. In Soviet

physics doklady, volume 10, pages 707–710.

Liu, K., Hogan, W. R., and Crowley, R. S. (2011). Nat-

ural language processing methods and systems for

biomedical ontology learning. Journal of biomedical

informatics, 44(1):163–179.

Maedche, A. and Staab, S. (2001). Ontology learning for

the semantic web. IEEE Intelligent systems, 16(2):72–

79.

Mahfoudh, M., Forestier, G., Thiry, L., and Hassenforder,

M. (2013). Consistent ontologies evolution using

graph grammars. In International Conference on

Knowledge Science, Engineering and Management,

pages 64–75. Springer.

Mahfoudh, M., Forestier, G., Thiry, L., and Hassenforder,

M. (2015). Algebraic graph transformations for for-

malizing ontology changes and evolving ontologies.

Knowl.-Based Syst., 73:212–226.

Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).

Efﬁcient estimation of word representations in vector

space. arXiv preprint arXiv:1301.3781.

Minarro-Giménez, J. A., Marin-Alonso, O., and Samwald,

M. (2014). Exploring the application of deep learning

techniques on medical text corpora. Studies in health

technology and informatics, 205:584–588.

Panchenko, A., Faralli, S., Ruppert, E., Remus, S., Naets,

H., Fairon, C., Ponzetto, S. P., and Biemann, C.

(2016). Taxi at semeval-2016 task 13: a taxonomy

induction method based on lexico-syntactic patterns,

substrings and focused crawling. In Proceedings of

the 10th International Workshop on Semantic Evalua-

tion (SemEval-2016), pages 1320–1327.

Petasis, G., Karkaletsis, V., Paliouras, G., Krithara, A., and

Zavitsanos, E. (2011). Ontology population and en-

richment: State of the art. pages 134–166.

Sbissi, S., Mahfoudh, M., and Gattouﬁ, S. (2019). Map-

ping clinical practice guidelines to SWRL rules. pages

283–292.

Séroussi, B., Bouaud, J., Sauquet, D., Giral, P., Cornet,

P., Falcoff, H., and Julien, J. (2010). Why gps do

not follow computerized guidelines: an attempt of ex-

planation involving usability with asti guiding mode.

Studies in health technology and informatics, 160(Pt

2):1236–1240.

Wohlgenannt, G. (2015). Leveraging and balancing het-

erogeneous sources of evidence in ontology learning.

pages 54–68.

Wong, W., Liu, W., and Bennamoun, M. (2012). Ontology

learning from text: A look back and into the future.

ACM Computing Surveys (CSUR), 44(4):20.

Ontology Learning from Clinical Practice Guidelines

319