An Ontology-based Possibilistic Framework for Extracting Relevant

Terms from Job Advertisements

Albeiro Espinal

1,2 a

, Yannnis Haralambous

1 b

, Dominique Bedart

and John Puentes

IMT Atlantique, Lab-STICC, CNRS UMR 6285, Brest, France

DSI Global Services, Le Plessis Robinson, France

Keywords:

Recruitment Process, Relevant Term Extraction, Recruiter’s Behavior Modeling, Textual Relevance Marker,

Ontology of Job Offer, Cognitive Uncertainty Measure, Ontology-based Belief-Desire-Intention Architecture.

Abstract:

In a traditional recruitment process, large amounts of resumes and job postings are often handled manually,

which is very time-consuming. Existing machine learning techniques for automatic resume ranking lack accu-

racy in accessing relevant information in job offers, which is crucially needed in order to ensure the pertinence

of resumes. We present a context-driven possibilistic framework for extracting such information from job

postings, in the form of relevant terms. In our process, after considering the recruiters’ speciﬁc organiza-

tional context, we analyze their term relevance evaluation strategies in job advertisements. By interviewing

a group of recruiters and analyzing their behavior, we have derived a ﬁrst set of textual relevance markers.

Existing term-extraction methods from the literature were also applied to extract such textual relevance mark-

ers. We have evaluated all markers using cognitive uncertainty measures and we have integrated them into

an ontology-based Belief-Desire-Intention architecture. Doing this, we have improved the F1 score and recall

measures of existing state-of-the-art term extraction approaches by 20% and 29% respectively. Besides, our

framework is open-ended: it is possible to add new textual markers at any time as nodes of a fuzzy decision

tree, the calculation of which depends on the context and domain of job offers.

1 INTRODUCTION

In a recruitment process, recruiters interact with two

essential documents: Curricula Vitae (CV) and Job

Advertisements (JA). The former are documents sent

by applicants to communicate their qualiﬁcations,

while the latter are often written by recruiters to de-

ﬁne the particular requirements of a new job opening

(Cabrera-Diego et al., 2019). In general, recruiters

are the main actors involved in managing these docu-

ments (Breaugh, 2013). They consider the pertinence

of each applicant’s CV and the requirements of the

JA. Based on this comparison, they select the most

relevant applicants to be contacted initially. This pro-

cedure is known as the “screening phase” (Cabrera-

Diego et al., 2019).

A screening phase becomes signiﬁcantly complex

and time-consuming when recruiters need to evalu-

ate a large number of CVs and JAs (Cabrera-Diego

et al., 2019). This phenomenon is indeed globally ex-

perienced by multiple recruitment ofﬁces (Zhao et al.,

2021). To reduce such workload, machine learning

https://orcid.org/0000-0002-1907-3424

https://orcid.org/0000-0003-1443-6115

methods have been proposed for automatically rank-

ing applicants’ CVs depending on the content of a JA.

However, these methods do not achieve the expected

identiﬁcation of relevant information expressed in the

JA. As a result, CVs are often ranked according to

background information, unrelated to the given JA.

To understand and overcome this limitation, we

studied the way recruiters evaluate information rel-

evance in JAs. The study of recruiters’ expertise is

essential, since they develop strategies to identify the

most relevant requirements of job positions.

As a ﬁrst step in the modeling of the recruiter’s

cognitive process we chose to start with a very basic

linguistic stratum, namely the one of terms (Frantzi

et al., 2002; Cram and Daille, 2016). Therefore we

identify “relevant terms” in the JA, which the recruiter

expects to ﬁnd in the CV. Of course, chances are the

terms used in the CV may be different, but they inter-

act with the relevant ones through a semantic network,

therefore an ontology is necessary to access a network

of similarities between terms.

To identify relevant terms we deﬁne the notion of

textual markers, that is: text features on various lin-

guistic strata that can be used for relevance evalua-

tion. The main contribution of our study is to provide

Espinal, A., Haralambous, Y., Bedart, D. and Puentes, J.

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements.

DOI: 10.5220/0011521700003332

In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 163-174

ISBN: 978-989-758-611-8; ISSN: 2184-3236

163

an ontology-based possibilistic framework to deﬁne

and evaluate textual markers associated to informa-

tion relevance in JAs. This framework is sufﬁciently

large to encompass the various recruiters’ strategies

and viewpoints. We also show how the organizational

context, as represented by the recruiters, can be ex-

ploited for automating and improving the extraction

of relevant information from JAs.

This article is organized as follows. Section 2

describes the state of the art. We summarize some

key principles of the proposed framework in Sec-

tion 3. The set of JA textual markers derived from

the recruiters’ perceptions and strategies is presented

in Section 4. Experimental results and discussion,

along with conclusions and perspectives are presented

in Sections 5, 6 and 7, respectively.

2 STATE OF THE ART

In general, the automated ranking of resumes in re-

lation to the content of a JA consists of three main

steps, namely, pre-processing, representation, and

ranking. Pre-processing includes text extraction from

documents in digital format, like .pdf, .doc, or .txt

ﬁles, followed by segmentation and semantic an-

notation of sections (Cabrera-Diego et al., 2019).

Moreover, noise reduction is necessary to identify

and remove stop words, as well as special sym-

bols (Roy et al., 2020). Thereafter, by means of

ontology-based (C¸ elik, 2016), document-embedding-

based (Zhu et al., 2018), or word-count-based tech-

niques (Cabrera-Diego et al., 2019), representations

of CVs and JAs are deﬁned, their contents are com-

pared applying similarity metrics (Guo et al., 2016),

and a ranking of resumes is obtained.

Various approaches have been proposed to deter-

mine ranking. For instance, recruiters’ feedback has

been combined with the evaluation of a relevance al-

gorithm (Kessler et al., 2012) (Cabrera-Diego et al.,

2019). Other techniques, centered on word embed-

ding, used ranking methods based on neural net-

works such as CNNs, LSTMs, RNNs, and DNNs

(Deng et al., 2018). Attention-based models were

also deﬁned to rank resumes with respect to JAs

(Zhao et al., 2021) (Wang et al., 2021). Furthermore,

joint embeddings were speciﬁed, in order to deter-

mine a common semantic space and compute cosine

similarities of CVs and JAs (Zhu et al., 2018).

The process of ranking applicants’ resumes en-

compasses a fundamental step, rarely studied in the

literature, namely the task of explicitly identifying the

most relevant information expressed in a JA, in order

to rank the corresponding resumes. This question of

explicit information extraction from JA has been ap-

proached under the perspective of document indexing,

at the document representation stage. Making use of

an n-gram and a graph representation of each docu-

ment, the RAKE (Rapid Automatic Keyword Extrac-

tion) (Rose et al., 2010) algorithm computes terms’

frequency and degree, to identify the most relevant

document terms. The YAKE! (Yet Another Keyword

Extraction) algorithm (Campos et al., 2018) uses tex-

tual and statistical features to detect the most pertinent

information expressed in documents. A parallel com-

bination of textual and graph-based statistical features

is used by the FRAKE (Fusional Real-Time Keyword

Extraction) algorithm (Zehtab-Salmasi et al., 2021)

for the same purpose. Finally, (Dagli et al., 2021)

studied the viability of topical and BERT (Bidirec-

tional Encoder Representations from Transformers)

models for identifying the most relevant information

in JAs.

Identifying the most relevant information in JAs is

essential to ensure the pertinence of an automatic re-

sumes ranking process. Yet, from the recruiter’s point

of view, relevant information in a JA is not equiva-

lent to extracted information for document indexing.

Additionally, existing studies in the ﬁeld of document

indexing do not consider the organizational context of

JA processing. This point is critical since the under-

representation of the organizational context limits sig-

niﬁcantly over time the application of machine learn-

ing methods (Martin Jr. et al., 2020).

In this study, we aim to improve the task of iden-

tifying automatically the most relevant information in

JAs, by giving a central role to the representation of

the organizational context. This context is a strong

component of the recruiters expertise, since their be-

havior is considerably inﬂuenced by the dynamics

and needs of each enterprise (Breaugh, 2013). The

ontology-based possibilistic framework that we pro-

pose intends to leverage organizational context inte-

gration, through the analysis of recruiters’ strategies

and viewpoints, for deriving context-driven textual

markers of information relevance in JAs.

3 FRAMEWORK DESCRIPTION

Our framework is structured in three axes: The ﬁrst

axis analyses what is relevant for recruiters in JAs.

We specify the organizational context of the docu-

ments and derive textual markers of information rel-

evance. These markers are represented by an ontol-

ogy. The second axis evaluates the pertinence of the

markers from a recruiters’ perspective. To this end,

a fuzzy decision tree is constructed to identify unde-

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

164

tected relationships between markers. These relation-

ships become new markers associated to information

relevance in a JA. As for the third axis, it determines

the best textual markers in a possibilistic agent archi-

tecture.

3.1 Deﬁnitions

Recruiters and job applicants write JAs and CVs re-

spectively, with a single goal: to express the set of

professional skills that they need or offer (Zhao et al.,

2021). A professional skill can be deﬁned as the mo-

bilization of knowledge, know-how, and interpersonal

abilities to achieve the goals of a professional activity

(Le Vrang et al., 2014). A professional activity can be

deﬁned as a series of related tasks, requiring special

knowledge and skills (Kiselyova et al., 2021).

A recruitment process is a sequence of steps un-

dertaken to identify, engage, and retain the best appli-

cant for a speciﬁc job position. With the help of ﬁeld

experts, recruiters identify the set of skills for a new

job opening. The main needs are then expressed in a

JA. Applicants respond to the JA by sending their re-

sumes, which are manually evaluated by recruiters in

order to determine whether or not the content of the

resumes meets the JA requirements (Cabrera-Diego

et al., 2019).

The screening phase demands the recruiter to gain

knowledge of the organizational context in order to

achieve successful results (Roy et al., 2020). Follow-

ing this phase, the most suitable proﬁles are contacted

on a personal basis in order to gain insight into their

experience and skills. Finally, recruiters conduct face-

to-face interviews with the top-ranked candidates.

3.2 Organizational Context and

Recruiters’ Strategies

In this section, we provide further details about the

protocol that we propose for extracting and construct-

ing textual features from recruiters’ strategies and

viewpoints. In general, a recruitment process takes

place in the organizational context of an enterprise.

We represent the organizational context of our do-

main, adapting an existing deﬁnition of societal con-

text (Martin Jr. et al., 2020). Accordingly, the orga-

nizational context surrounding JAs is represented by

identifying the main social actors, artifacts, actions

on artifacts, and enterprise processes, associated with

this type of document. Those aspects of the organi-

zational context can signiﬁcantly inﬂuence the life-

cycle of CVs and JAs in the hiring processes. As a

consequence, the representation of the organizational

context is a fundamental preliminary step. Our ap-

proach makes use of the UNC-method, which was

deﬁned for representing organizational contexts (Za-

pata Jaramillo and Arango Isaza, 2009), as:

1. The organizational context representation is con-

structed together with recruiters from preliminary

interviews (Zapata Jaramillo and Arango Isaza,

2009). Pre-conceptual schema and domain mod-

els are applied to identify and model the main

contextual concepts and their relationships. From

these representations, a mother-ontology is di-

rectly derived. This ontology covers the main con-

cepts associated with JAs and the particular orga-

nizational context where these documents are pro-

cessed.

2. Based on these diagrams, recruiters goals related

to the life-cycle of JAs in the recruitment pro-

cesses, are identiﬁed and represented in a hierar-

chical structure.

3. Process diagrams (Zapata Jaramillo and

Arango Isaza, 2009) represent models of en-

terprise processes related to JAs.

4. Using a ﬁshbone chart, the relationships between

organizational problems related to JAs and their

causes are also represented.

5. As suggested by (Zapata Jaramillo and

Arango Isaza, 2009), all the previous dia-

grams are combined in a Process Explanatory

Table in order to unify the representation of the

organizational context.

6. From this representation and the derived ontology,

the analysis of recruiters perceptions and strate-

gies about information relevance in JAs is carried

out.

7. Recruiters proceed to annotate relevant informa-

tion in JA documents, handled in different recruit-

ment processes. As recruiters read and annotate

the most relevant information in JAs, each an-

notation is observed and described in detail, ap-

plying the controlled language proposed by (Zap-

ata Jaramillo and Arango Isaza, 2009).

8. Our approach identiﬁes and classiﬁes common

and transverse recruiters’ behaviors when anno-

tating documents into two categories: explicit

ones (such as selecting a term) and implicit ones

(such as refraining from selecting of a term).

Transverse annotation behaviors are described by

semantic rules. The mother-ontology conceived

in step (1) is used to model the concepts and rela-

tions required by each rule. These semantic rules

are textual markers of information relevance in

JAs.

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements

165

9. Cognitive uncertainty associated to derived tex-

tual markers is estimated by computing an am-

biguity measure, as deﬁned in (Yuan and Shaw,

1995). In Section 4, we formalize the process of

evaluating textual markers.

10. Relevance markers with an ambiguity level less

than or equal to an ambiguity threshold are se-

lected and implemented in an ontology-based pos-

sibilistic agent architecture to simulate recruiters’

behaviors. Relations between markers are anal-

ysed by calculating a fuzzy decision tree (Yuan

and Shaw, 1995). In this tree, each marker is con-

sidered as a decision node. The least ambiguous

relations derived from this tree are identiﬁed as

new relevance markers in the agent’s architecture,

in order to improve its deliberation process of in-

formation relevance in JAs.

3.3 Ontological Resources

As a fundamental part of our framework, a mother-

ontology contains the main concepts and relations

inherent to the speciﬁc context of JAs. This on-

tology is complemented by integrating existing on-

tologies related to the speciﬁc organizational context.

In our case, we integrated the following sources of

knowledge. The internal ontology of DSI Group, a

French consulting ﬁrm. It contains over 36,000 pro-

fessional skills, lexically represented in French lan-

guage. By complementing it with international on-

tologies like ESCO (Le Vrang et al., 2014), we also

reconstructed ontologies that underpin professional

skills frameworks such as O*NET

, CIGREF

, and

ROME

, based on text-to-RDF-triple transformations

(Somodevilla Garc

ıa et al., 2018).

Additionally, we built an ontology based on

14,000 anonymous CVs and 2,000s JAs, in a semi-

automatic manner, by applying ontology learning

techniques (Alfonso-Hermelo et al., 2019). Specif-

ically, we extracted and processed the professional

skills’ sections of these documents. Then, we inte-

grated all ontologies to the mother-ontology by us-

ing a hybrid approach based on transformer models

(BERT), a terminological variation analysis (Cram

and Daille, 2016), and ontology quality metrics (Mc

Gurk et al., 2017). In Figure. 1, we present an upper-

view of the resulting ontology.

https://www.onetonline.org/

https://www.cigref.fr/

https://www.pole-emploi.fr/employeur/vos-

recrutements/le-rome-et-les-ﬁches-metiers.html

3.4 Beliefs, Desires and Intentions

Architecture of the Agent

According to multiple empirical studies, ”...hu-

mans’ semantic inferences are uncertain and context-

sensitive...” (Pavlick and Kwiatkowski, 2019). The

same principle can be applied to the process of ex-

tracting relevant terms from JAs. In fact, recruiters

semantically infer relevant terms by reading the JA’s

text, ”...[assuming] common human understanding of

language [and] common background knowledge...”

(Pavlick and Kwiatkowski, 2019). Consequently, we

propose a possibilistic architecture to incorporate hu-

man semantic reasoning and its associated uncertainty

for enhancing the automatic extraction of relevant

terms.

Speciﬁcally, this architecture (Da Costa Pereira

and Tettamanzi, 2010) is based on what has been de-

ﬁned as the agent’s beliefs, desires, and intentions

paradigm, linked to modeling of variables applying

possibility (Figure. 2).

The module of beliefs is composed by two sub-

modules. The ﬁrst one consists of a distribution of

trust degrees τ, which represents a trust level that

the agent assigns to its sources φ. In our context,

there are three categories of sources: JAs, ontologi-

cal resources, and textual markers of information rel-

evance. We deﬁne trust degrees through quality met-

rics, in order to detect possible inconsistencies, as

malformed JA texts, ontology conformity issues (Mc

Gurk et al., 2017), or ambiguous textual relevance

markers. Each trust degree is associated to a toler-

ance threshold β

. A source with a trust degree τ

below a given tolerance threshold β

will be rejected

by the agent.

The second sub-module makes use of a possibil-

ity distribution π, which represents the agent’s be-

liefs B. Each term is characterized by a possibility

distribution formed by the levels of relevance (possi-

bility degrees) provided by each marker. This possi-

bility distribution induces a measure of possibility Π

(Da Costa Pereira and Tettamanzi, 2010). It shows to

what extent a term is pertinent. Then, the possibility

measure is associated to its dual necessity measure N

(Da Costa Pereira and Tettamanzi, 2010). The latter

can be interpreted as to what extent it is impossible

not to conclude that a term is relevant. Note that pos-

sibility degrees of beliefs are limited to the trust de-

gree of their respective information about knowledge

sources. A belief modiﬁcation operator ∗ is deﬁned

for this purpose (Da Costa Pereira and Tettamanzi,

2010).

The second agent’s component is the module of

desires, represented by an utility distribution (for-

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

166

Figure 1: Upper-view of the mother-ontology created from the context representation.

Figure 2: Basic architecture of the dynamic BDI agent

(Da Costa Pereira and Tettamanzi, 2010).

mally a possibility distribution), which indicates

whether or not the use of a textual relevance marker

is convenient according to the current agent’s be-

liefs, desires, and desire-generation rules. Desire-

generation rules R

are logical conditions expressing

dependence relationships between beliefs and desires

(Da Costa Pereira and Tettamanzi, 2010). The eval-

uation of desire-generation rules allows the agent to

estimate to what extent ∆ it is justiﬁed to make use of

each marker. Textual markers associated to positive

degrees become justiﬁed desires ϒ.

The third agent’s component, namely intentions,

consists of actions that the agent decides to execute

based on its current beliefs, justiﬁed desires, and

goals. A plan outlines the sequence of actions to be

taken. In our context, these actions refer to natural

language processing methods that are necessary to au-

tomatically extract the relevant terms from JAs.

4 TEXTUAL MARKERS

The algorithm YAKE! (Campos et al., 2018) already

uses textual markers, but deﬁned in an informal way.

In this section we will formally deﬁne the notion of

textual marker in a way that it is compatible with

YAKE! textual markers, as well as with textual mark-

ers actually used by recruiters.

4.1 Preliminary Deﬁnitions

According to (Cram and Daille, 2016), a term is “a

functional class of lexical units used in discourse”. In

our context, terms are identiﬁed based on their ter-

mhood (Frantzi et al., 2002), measured by using the

weirdness ratio (Cram and Daille, 2016).

Terms are extracted by applying the most frequent

morpho-syntactic patterns to multiple specialized cor-

pora (Cram and Daille, 2016). Nearly all of them cor-

respond to nominal phrases.

Let d

be a JA belonging to a corpus C and T

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements

167

,.. .,t

} the set of terms of d

Let R

⊆ T

be the set of most relevant terms in

. Each term t

∈ R

is considered as relevant under

a possibility degree α

Let A

= {a

,.. .,a

} be the set of sections of

(job description, proﬁle details, etc). Each section

can be represented by a subset of terms from T

. A

term can belong to multiple sections.

Let E

= {e

,. .., e

} be a set of qualifying ad-

jectives and nouns that are linked to a subset of terms

in T

by syntax dependencies.

Let O = {o

,. .., o

} be a set of ontologies

(as the one presented in Section 3). Let c

s,1

s,2

,. .., c

s,k

} be the set of concepts of ontology

and T

= {t

j,1

j,2

,. ..,t

j,l

} the set of terms lexi-

cally representing concept c

in a given language.

The mother-ontology O contains, among other

concepts, the structure of a textual document: sec-

tions, paragraphs, sentences, syntagms, terms, words,

morphemes, etc.

4.2 Deﬁnitions of Markers

In this section, textual markers #1 to #10 have been

obtained by monitoring the behavior of expert re-

cruiters and by interacting with them. Textual mark-

ers #11 to #16 correspond to those of the YAKE! term

extraction approach.

4.2.1 Textual Marker #1: Presence of

Professional Skills or Job Types in Title

Sections

“In the event that a term in the title matches one of

the terms used to represent professional skills or job

types, then it can potentially be considered as rele-

vant.”

Let a

∈ A

be the title section of d

. Let t

,. ..,t

} be the set of terms contained in a

. T

is the set of terms lexically representing a professional

skill or job type concept c

in the ontology o

. We

request that:

∀t

∃c

∈ o

∧t

∈ T

∧t

∈ t

] → t

∈ R

, (1)

with a possibility degree α

k,1

∈ [0, 1].

4.2.2 Textual Marker #2: Terms Representing

Professional Skills in a Job Description

Section or Proﬁle Description Section

In general, a term used to represent a professional

skill in a job description section or in a proﬁle de-

scription section is more likely to be chosen as a rele-

vant term. Let s

and s

be the sets of terms in the job

description section and the proﬁle description section,

respectively. Set t

∈ T

. Let T

be the set of terms

representing a professional skill concept c

in the on-

tology o

. We request that:

∀t

∃c

((t

∈ s

∨t

∈ s

) ∧t

∈ T

) → t

∈ R

, (2)

with a possibility degree α

k,2

∈ [0, 1].

4.2.3 Textual Marker #3: Relevance of Job

Posting Sections

“Recruiters will choose with a higher possibility de-

gree terms used in the title, job description, and pro-

ﬁle description sections, rather than terms belonging

to other sections (company description, contract de-

tails, etc).”

This marker is not redundant with markers #1 and

#2, since we do not request terms to be professional

skills. Let S = s

∪ s

⊆ T

, where: s

is the set

of terms of the title section; s

is the set of terms of

the job description section; and s

is the set of terms

of the proﬁle description section.

Let t

∈ T

∩ S. Then, we request that:

∀t

∈ T

∧t

/∈ S) → (P(t

∈ R

) > P(t

∈ R

)),

(3)

with a possibility degree α

k,3

∈ [0,1]. P(t

∗

∈ R

) de-

notes the possibility of t

∗

being chosen as a relevant

term.

4.2.4 Textual Marker #4: Terms Dependent on

Pertinence Expressions

“A term that bears a syntax dependency with a syn-

tagm of the JA is more likely to be chosen as a rele-

vant term.”

• Let t

∈ T

∩ T

for some c

• We deﬁne a “pertinent expression” e

as a syn-

tagm used by the recruiter who wrote the JA (i.e.,

you master C#, good knowledge of cloud com-

puting). Let us suppose that e

is syntactically

dependent with t

. More speciﬁcally, let t

be a

qualifying adjective or a noun modiﬁer directly

dependent with e

. Then, we request that:

∀t

∃e

∈ T

∧e

∈ E

∧is

dependent(t

))

→ t

∈ R

, (4)

with a possibility degree α

k,4

∈ [0, 1].

4.2.5 Textual Marker #5: Terms Used in Traces

of Professional Activities Descriptions

“It will be more likely for a term representing a pro-

fessional concept to be considered as relevant when

a job description explicitly describes an interaction

with this concept.”

We deﬁne a trace of a professional activity de-

scription as a sentence in a job posting, describing a

worker’s action on an object. Be b

∈ d

a trace of

a professional activity description represented by the

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

168

set of terms T

. We request that b

contains at least

one verb and one dependent object. We assert that the

terms t

representing these objects will have a greater

possibility degree of being chosen as relevant. There-

fore:

∀t

∈ T

∧ is object(t

)) → t

∈ R

, (5)

with a possibility degree α

k,5

∈ [0, 1].

4.2.6 Textual Marker #6: Terms Representing

High Risk Professional Skills/Activities

This marker intends to provide more relevance to

terms representing professional skills or activities for

which an employee’s mistake can signiﬁcantly de-

grade the company’s economic activity. Value 0 in-

dicates that a potential mistake will not have a sub-

stantial impact on the economic activity, while value 1

indicates that an error on this skill or activity will have

a strong impact.

Depending upon the unique context of each enter-

prise, this textual marker allows the agent to adapt its

own behavior to meet the speciﬁc needs of each job

opening.

Let M be an ontology that contains the set of pro-

fessional skills and activities of a given company. M

contains a set of concepts c

= {c

M,1

M,2

,. .., c

M,k

The recruiter manually assigns a risk level ε

M,k

∈

[0,1] to each professional skill or activity.

Let s

be a term in a job posting d

representing

a professional skill or activity in M. Among the con-

cepts associated to s

, let c

M,l

be the concept with the

maximal risk level. If this risk level is greater than

a threshold β

M,l

, then s

is chosen as a relevant term

and:

∀s

∃c

M,l

∈ T

∧ c

M,l

∈ M ∧ s

∈ T

M,l

∧ is greater than(ε

M,l

,β

M,l

) → s

∈ R

, (6)

with possibility degree α

j,6

∈ [0, 1]

4.2.7 Textual Marker #7: Actions Expressed in

Management JAs

It is also important for recruiters to identify the type

of actions required by management JAs. For instance,

some jobs focus on team management, while others

include accountability activities or even development

tasks. This is because management jobs can be quite

diverse.

Be d

a management job posting. The agent de-

tects management JAs based on a Latent Dirichlet Al-

location model, trained on 14,000 curricula vitae. Let

be a verbal term of d

. If t

belongs to the trace of a

professional activity f

and corresponds to the head of

its syntactic tree, then it potentially is a relevant term.

We deﬁne this marker as:

∀t

∃ f

( f

∈ d

∧t

∈ f

∧ is management(d

)∧

is verb(t

) ∧ is head of(t

, f

)) → t

∈ R

, (7)

with a possibility degree α

k,7

∈ [0, 1].

4.2.8 Textual Marker #8: BERT Semantic

Similarity of Professional Skills

“Speciﬁc terms used to represent professional skills

that are semantically close (in the sense of BERT) to

already discovered relevant terms, will be considered

relevant.”

Let t

∈ R

and t

∈ T

. We deﬁne the speciﬁcity

function (Cram and Daille, 2016) of a term f (t) as its

relative frequency in a speciﬁc corpus C

, divided by

its relative frequency in a multi-language corpus C

Furthermore, we deﬁne g(t

) as the BERT se-

mantic similarity between two terms. We used a

SBERT model pre-trained on the Wikipedia corpus

(Reimers and Gurevych, 2019) to derive semantically

meaning from complex terms. This model was ﬁne-

tuned on the professional skill standards CIGREF, e-

CF, c2i, and ROME, being formalized as follows:

∀t

∈ R

∧ g(t

) > 0) → t

∈ R

, (8)

with a possibility degree deﬁned by the normalized

equation :

= ∥(1 − α

) ∗ g(t

) ∗ f (t

))∥. (9)

4.2.9 Textual Marker #9: Relevance of the

Economic Activity Sector

“Terms referring to an economic activity sector re-

quired by a job posting (e.g., ﬁnance, banks, aeronau-

tics, etc.) will be chosen as potentially relevant.”

This implies that:

∀t

∈ T

∧ is sector requirement(t

)) → t

∈ R

(10)

with a possibility degree α

k,9

∈ [0, 1].

Economic activity sectors are identiﬁed by align-

ing job posting terms and economic activity concept

labels, provided by ESCO, O*NET, ROME, and CI-

GREF standards.

4.2.10 Textual Marker #10: Professional Skill

Prerequisites

Let there be a prerequisite relation between two pro-

fessional skills c

and c

as given in an ontology o

Relations of this type can be obtained from ontolo-

gies such as ESCO. If c

is a prerequisite of c

and c

is considered as relevant (under a certain possibility

degree) then c

inherits the possibility degree of c

∀t

∃c

∈ o

∧c

∈ o

∧t

∈ T

∧t

∈ T

∧

is prerequisite(c

) ∧t

∈ R

) → t

∈ R

, (11)

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements

169

with a possibility degree α

k,10

∈ [0,1]. This possibil-

ity degree is equal to the possibility degree of t

∈ R

4.2.11 Textual Marker #11: YAKE! Casing

Upper-cased terms tend to be more relevant. In

our context, this YAKE! marker is related to the re-

cruiters’ behavior, as they tend to capitalize terms per-

taining to professional skills:

∀t

∈ T

∧ is upper cased(t

)) → t

∈ R

(12)

We deﬁne the possibility degree of this rule based

on the normalized YAKE! equation:

k,11

) =



max(TF(U(t

),TF(A(t

))

ln(TF(t

))



, (13)

where TF(U (t

)) is the number of times that t

ap-

pears uppercased, TF(A(t

)) is the number of occur-

rences of t

as an acronym (for details see (Campos

et al., 2018)) and TF(t

) is the term frequency.

4.2.12 Textual Marker #12: YAKE! Term

Position

The hypothesis of this marker is that terms appearing

at the beginning of the document tend to be more rel-

evant.

∀t

∈ T

∧ is position marker activated(t

))

→ t

∈ R

, (14)

with a possibility degree given by the following nor-

malized YAKE! equation:

) = ∥ ln(ln(3 + Median(Sent(t

))))∥, (15)

where Sent(t

) is the set of positions of the sentences

containing t

4.2.13 Textual Marker #13: YAKE! Term

Frequency Normalization

The most commonly used terms are more relevant.

Hence:

∀t

∈ T

∧ is frequency marker activated(t

))

→ t

∈ R

, (16)

with a possibility degree given by the following nor-

malized equation proposed by YAKE!:

k,13

) =



TF(t

)

MeanTF + σ



, (17)

where TF(t

) is the number of occurrences of t

which is balanced by the mean and standard deviation

of frequency.

4.2.14 Textual Marker #14: YAKE! Term

Relatedness to Context

This YAKE! marker is based on the following hypoth-

esis: “The higher the number of different terms that

co-occur with a candidate term t on both sides, the

least signiﬁcant term t will be”:

∀t

∈ T

∧ is relatednes activated(t

)) → t

∈ R

(18)

with a possibility degree obtained from the normal-

ized YAKE! equation:

k,14



1 + (DL + DR···) ∗

TF(t

)

maxTF



, (19)

where

DL[DR] =

t,w

∑

k∈A

t,w

CoOccur

t,k

. (20)

t,w

| corresponds to the number of different terms in

a window of size w and TF is the term frequency.

4.2.15 Textual Marker #15: YAKE! Different

Sentences

“A term will be more relevant depending on how often

it is used within different sentences,” represented as:

∀t

∈ T

∧ is sentences marker activated(t

))

→ t

∈ R

, (21)

with a possibility degree obtained from the normal-

ized equation:

k,15



SF(t

)

#Sentences



, (22)

where SF(t

) is the number of sentences containing t

and #Sentences is the total number of sentences of d

4.2.16 Textual Marker #16: YAKE! Overall

Score

We include the global relevance score proposed by

YAKE! which is based on markers #11, #12, #13, #14,

and #15 (Campos et al., 2018). Let t

∈ d

. A term is

considered as “partially relevant” if it’s predicted as

such by the overall score:

∀t

∈ T

∧ is predicted by yake(t

)) → t

∈ R

(23)

with a possibility degree α

k,16

∈ [0, 1].

4.3 Evaluation of Relevance Markers

According to (Yuan and Shaw, 1995), “a classiﬁcation

task can be viewed as a rational action that a decision

maker will take in consistency with his or her knowl-

edge”. In our case, the recruiter’s task of annotating

relevant terms is associated with uncertainties, which

have been measured in the literature through ambigu-

ity measures (Yuan and Shaw, 1995). Based on ambi-

guity estimation, we propose a protocol for evaluating

the pertinence of relevance markers.

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

170

4.3.1 Protocol Formalization

Let U = {t} be the set of terms of a given JA. Let

C be a fuzzy set representing the recruiter perceived

levels of term relevance in a JA. C is represented by

the membership function µ

, which maps recruiters’

annotations into C. C is constituted by two fuzzy sub-

sets: the subset C

of relevant term levels and the sub-

set C

of non relevant term levels. Both fuzzy cat-

egories are represented by triangular functions. We

will call R (resp. R

) the subset of C (resp. C

)

obtained from the fuzziﬁcation of recruiters annota-

tions on the JA.

Each JA term t can be described by a set of rele-

vance markers I

, obtained from recruiters strategies

and viewpoints. We denote A(t) = {I

,. .., I

Each relevance marker provides a degree of pos-

sibility for a term to be selected as relevant. The next

step involves fuzzifying these possibility degrees by

applying a membership function µ

. Although µ

was built in an equivalent manner as µ

, the speciﬁc

codomain of each marker I

was taken into account.

We interpret the result of this fuzziﬁcation process as

an evidence E

After fuzzifying I

and the recruiters’ annotations,

we evaluate the possibility of describing the fuzzi-

ﬁed recruiters’ annotations R based on the evidence

obtained from I

. More speciﬁcally, we evaluate

the ambiguity level of the following rule: If E

Then

R. To do it, a subsethood measure between the ev-

idence E

and the experts’ classiﬁcation R has been

proposed (Yuan and Shaw, 1995). In our context, as

a substitute for subsethood, we use the Hamming dis-

tance between fuzzy sets to reﬂect more explicitly to

what extent recruiter’s knowledge R approximates E

S(E

,R) =

∑

t∈U



(t) − µ

(t)



. (24)

The Hamming distance measures the degree to

which a term is relevant or not, based on the avail-

able evidence E

. Applying this distance, we deﬁne

the possibility π of classifying a term as relevant (R

)

or non relevant (R

), in relation to recruiter’s strate-

gies and viewpoints as:

π(R

| E

) =

S(E

)

max(S(E

),S(E

))

. (25)

Possibility is intrinsically related to the concept of

ambiguity (Yuan and Shaw, 1995). Based on E

, the

more likely we are to deduce that a term is relevant or

not, the less ambiguity there is. From π(R | E

), we

estimate the ambiguity level associated to marker I

or equivalently, to the evidence E

as:

G(E

) = g(π(R|E

)) =

∑

i=1

(π

∗

− π

∗

i+1

)ln(i), (26)

where π

∗

= {π

∗

,π

∗

,. .., π

∗

} is the possibility distribu-

tion π(R | E

) permuted and sorted so that π

∗

≥ π

∗

i+1

for i ∈ {1, .. ., n} and π

∗

n+1

= 0. In our case, n = 2 as

we evaluate ambiguity when deciding whether a term

is relevant (R

) or not (R

) based on I

This ambiguity function G indicates to what de-

gree it can be inferred that a term is pertinent or not,

depending on the relevance marker I

. A value of

0 means that there is no ambiguity and a value of

ln(n) represents the maximal ambiguity level (Yuan

and Shaw, 1995).

Since A = {I

,, ... ,, I

} is the set of relevance

markers, based on the least ambiguous markers with

an ambiguity level below or equal to a threshold σ,

we train a fuzzy decision tree (Yuan and Shaw, 1995).

We replace the classical information entropy mea-

sure with the previously presented ambiguity metric.

As proposed by (Yuan and Shaw, 1995), the path of

branches from the roots to the leaves is converted to

simpliﬁed logical rules. These new rules represent

logical relations between markers associated to in-

formation relevance. The least ambiguous ones are

added as new sources of information in the agent ar-

chitecture. This allows us to enrich the deliberation of

the agent. The process of building the fuzzy decision

tree is an important step in evaluating the pertinence

of markers. A fuzzy decision tree can be used for

discovering unseen relationships between them, and

also for automating the classiﬁcation task of relevant

terms’ prediction.

5 EXPERIMENTAL RESULTS

We evaluated the proposed framework in the setting

of the recruitment ofﬁce of DSI Group. We inter-

viewed four recruiters, which we refer to in this study

as A, B, C, and D. These recruiters spent multiple

days or even weeks leading recruitment processes and

managing their associated JAs, exchanging on a daily

basis with management and technical experts. In

other words, they gained an in-depth understanding of

the most fundamental needs of their JA by acquiring

high levels of contextual knowledge.

Based on preliminary interviews with these re-

cruiters, the organizational context surrounding the

JAs was represented. Then, we interviewed an ex-

pert recruiter (recruiter A) in order to identify which

job requirements were most relevant to the managed

JAs. During this interview, we asked the recruiter the

following question about ﬁve JAs: what are the re-

quirements expressed in this JA that you would not

accept a candidate without?

According to personal viewpoints and knowledge,

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements

171

Table 1: Experimental results. Precision, recall, and F1-score levels of each method on 20 job advertisements applying RAKE

(Rose et al., 2010), FRAKE (Zehtab-Salmasi et al., 2021), BERT topics approach (Dagli et al., 2021), YAKE! (Campos et al.,

2018), and our Agent.

Metric/Method RAKE FRAKE BERT YAKE! Our Agent

Recall@N, Precision@N

F1-Score@N

0.02 0.09 0.17 0.10 0.38

Recall@2N 0.08 0.17 0.20 0.27 0.56

Precision@2N 0.04 0.08 0.10 0.13 0.28

F1-Score@2N 0.05 0.11 0.14 0.18 0.38

this recruiter marked a set of terms that represented

the most essential requirements of each JA. The ma-

jority of annotated terms corresponded to professional

skills. A small number of terms related to locations or

key aspects of previous responsibilities. The recruiter

annotated a minimum of 4 terms and a maximum of

10 terms per JA. In this speciﬁc application case, JAs

had, in average, 100 terms.

By applying our framework, we derived textual

markers of information relevance associated to the

recruiter strategies and viewpoints. These markers

were formalized, implemented, and became the cor-

nerstone of a BDI agent. We evaluated the validity

and reproducibility of this agent on 20 recruitment

processes managed by recruiters B, C and D. This

process was executed as follows:

• We asked recruiters B, C and D to annotate the

most relevant information expressed in the set of

JAs associated to their recruitment processes.

• We compared their annotations in relation to the

most relevant terms predicted by the agent derived

from recruiter A’s strategies and viewpoints.

• We compared the performance of the proposed

agent to state of the art term extraction meth-

ods. We used precision@K, recall@K and F1-

score@K metrics to asses the suitability of each

model, in predicting the top N and top 2N most

relevant terms of each annotated JA (where N rep-

resents the number of terms annotated by the re-

cruiter on each document).

Even if theoretically our approach should be com-

pared to other approaches speciﬁcally designed for

JAs, we discarded that aim for two reasons. To the

best of our knowledge, currently there are not auto-

mated and open source methods for extracting specif-

ically relevant information from JA. Also, there are no

public JA corpora annotated for the extraction of rele-

vant information. For these reasons, we compared our

approach to the performance of domain-independent

state of the art term extraction methods. In their re-

spective papers, these approaches were evaluated on

annotated academic documents, news texts, and sci-

entiﬁc articles (Campos et al., 2018). Therefore, this

is the ﬁrst time that they are evaluated on JAs.

Table 1 presents the results of our experiments.

As indicated by the metrics, our algorithm’s results

are signiﬁcantly better than the other four algorithms.

Particularly, compared to the best results of those al-

gorithms (most of them from YAKE!), our approach

improvements vary from 15% to 29%, being 56% for

Recall@2N the highest performance.

Regarding the performance of the BDI agent,

three aspects of its behavior were essential for achiev-

ing and preserving a superior F1-Score at @N and

@2N. Firstly, context-driven markers associated with

lower levels of ambiguity, signiﬁcantly inﬂuence the

extraction process of the agent. For instance, markers

#1, #4 and #5 have usually been sufﬁcient per se for

determining whether a term (under the conditions of

each marker) is relevant or not. Secondly, context-

driven markers with medium ambiguity levels tend

to reduce their associated cognitive uncertainty, when

used together with other markers. Thirdly, we ob-

served that the inclusion of cognitive uncertainty mea-

sures to determine trust degrees, allowed the agent

to control the relative ambiguity associated to some

markers as #12.

Finally, our results reﬂect the complexity of ex-

tracting relevant terms from short documents such as

JAs. Work on YAKE! had already evidenced this is-

sue on corpora made up of short length documents.

Various corpora were involved in their work, e.g.,

WWW, KDD and pak2018 (Campos et al., 2018). In

our case, YAKE! has a similar behavior (18%) when

compared to its reported maximal of 17.2% for these

datasets of short documents. On the other hand, a

maximal F1-Score@2N of 5% and 14% for RAKE

(n-gram) and BERT (embeddings) respectively, could

be attributed to the inadequacy of document repre-

sentations to reproduce the concepts of an organiza-

tional context. In fact, the n-gram approach tends to

underrepresent terms, while the embedding approach

may not be aligned with the inherent knowledge of a

given organizational context, even after a ﬁne-tuning

process. It also appears that the FRAKE algorithm

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

172

is limited by the short length of JAs, given that it is

centered on statistical graph-based markers. Accord-

ing to these elements, with a maximal F1-Score of

38%, the performance of our approach illustrates that

context-driven markers can complement relevance in-

formation extraction in JAs, by integrating knowl-

edge of the speciﬁc organizational context, from a re-

cruiter’s perspective.

6 DISCUSSION

The automatic extraction of relevant terms from JAs

is a complex task that has been rarely studied, despite

its key role to determine the ranking of CVs during

the screening phase of recruitment. We propose an

approach that relies on a multisource ontology and a

possibilistic framework, utilized by a dynamic BDI

agent model, to analyze simple rules that deﬁne 16

textual markers. Given the current lack of speciﬁc

approaches for this task, we compared our approach

to the closest related techniques, i.e., automatic term

extraction approaches for document indexing. Out

of 16 textual markers, we deﬁned 10 markers upon

an expert recruiter’s strategies and viewpoints, while

6 were integrated from the YAKE! algorithm. Re-

sults show that the deﬁned BDI agent performs better

than automatic term extraction algorithms, in extract-

ing relevant information from JAs. Obtained perfor-

mances suggest that representing the organizational

context of JAs, in terms of recruiter’s strategies and

viewpoints, is likely to improve the identiﬁcation of

information relevance, beyond relevant terms.

Also, we observed that our implementation is

likely to detect more coherent sets of relevant terms

than other approaches. Given that the agent extracts

terms following a terminological analysis (Frantzi

et al., 2002; Cram and Daille, 2016), it detects com-

plex terms often chosen as relevant by recruiters. Be-

sides, we observed that some textual representations,

such as the one proposed by RAKE (Rose et al.,

2010), underrepresent complex terms, reducing pre-

dictability considerably.

An overview of results reveals that performance

measures of all algorithms, except Recall@2N of our

algorithm, are below the expected baseline perfor-

mance. This can be explained by the fact that statisti-

cal markers on which are based the works of (Campos

et al., 2018) (Rose et al., 2010) (Zehtab-Salmasi et al.,

2021) are insufﬁcient to determine appropriately term

relevance in JAs. Our results show that context-based

qualitative relevance markers are essential for achiev-

ing a better F1-score, given that they are independent

of the JA’s size. E.g., it is not unusual to ﬁnd JAs

where the extent of the enterprise description is large

compared to the job description. As a consequence,

statistical markers tend to give a higher score to terms

from the former section that aren’t necessary to de-

scribe essential job requirements.

After performing combinatorial analysis of tex-

tual markers, we found that the agent gets a bet-

ter F1-score by executing textual markers as follows.

The ﬁrst step is to generate a population of beliefs

about terms relevance, using independent markers

with lower ambiguity levels. For instance, marker #1

based on titles contents or marker #6 focused on ﬁ-

nancial impact of professional skills/activities. Then,

dependent markers with low ambiguity levels (such as

#10) can be applied in order to reinforce the current

population’s beliefs about terms relevance.

7 CONCLUSION AND

PERSPECTIVES

Multiple analyses and resources are necessary to au-

tomatize a recruitment process. While a rich ontology

is essential to cover the wide spectrum of professional

qualiﬁcations, skills, and experience, the study of JAs

permits to identify relevant terms intended to facilitate

CV ranking. Additionally, the ambiguity of employed

language must also be taken into account. Since sta-

tistical parameters do not provide optimal representa-

tions of relevant document terms in a JA, we propose

to correlate human decisions with identiﬁed textual

markers. This approach relies strongly on termino-

logical analysis and hypotheses formulated after ob-

serving how recruiters work. Our results indicate that

the terminological analysis is a fundamental step for

improving the linguistic quality and contextual rele-

vance of information extraction from JAs. Our frame-

work has therefore the potential to assess the rele-

vance of text markers within speciﬁc contexts, and to

be adapted to the evolving organizational context of

each company.

Further work is still required to conﬁrm the var-

ious insights that have been identiﬁed, in particular

concerning recruiters’ way of working when match-

ing JAs and CVs, as well as the relevance of other

models to represent the semantic space of relevant

terms. A larger corpus and more recruiters will be

necessary to evaluate whether the applicability of tex-

tual relevance markers can be affected by variations

in the organizational context. Compatible automatic

approaches to identify relevant terms in CVs to match

JAs are foreseen, integrating the recruiter’s expertise.

Other machine learning models will be deﬁned look-

ing for a more precise representation of the intrin-

An Ontology-based Possibilistic Framework for Extracting Relevant Terms from Job Advertisements

173

sic and extrinsic relationships between the concepts

constituting a speciﬁc organizational context in which

JAs are processed.

REFERENCES

Alfonso-Hermelo, D., Langlais, P., and Bourg, L. (2019).

Automatically Learning a Human-Resource Ontology

from Professional Social-Network Data. In Canadian

AI 2019: Advances in Artiﬁcial Intelligence, volume

11489 of LNAI, pages 132–145. Springer.

Breaugh, J. A. (2013). Employee Recruitment. Annual Re-

view of Psychology, 64:389–416.

Cabrera-Diego, L. A., El-B

eze, M., Torres-Moreno, J. M.,

and Durette, B. (2019). Ranking r

esum

es automati-

cally using only r

esum

es: A method free of job offers.

Expert Systems with Applications, 123:91–107.

Campos, R., Mangaravite, V., Pasquali, A., Jorge, A. M.,

Nunes, C., and Jatowt, A. (2018). YAKE! Collection-

Independent Automatic Keyword Extractor. In Ad-

vances in Information Retrieval, pages 806–810.

Springer.

C¸ elik, D. (2016). Towards a semantic-based information

extraction system for matching r

esum

es to job open-

ings. Turkish Journal of Electrical Engineering and

Computer Sciences, 24(1):141–159.

Cram, D. and Daille, B. (2016). Terminology extraction

with term variant detection. In Proceedings of ACL-

2016 system demonstrations, pages 13–18.

Da Costa Pereira, C. and Tettamanzi, A. G. B. (2010). An

integrated possibilistic framework for goal generation

in cognitive agents. In Proceedings of the 9th Interna-

tional Conference on Autonomous Agents and Multia-

gent Systems: Volume 1, pages 1239––1246.

Dagli, R., Shaikh, A. M., Mahdi, H., and Nanivadekar, S.

(2021). Job Descriptions Keyword Extraction using

Attention based Deep Learning Models with BERT.

In 3rd International Congress on Human-Computer

Interaction, Optimization and Robotic Applications

(HORA), pages 1–6.

Deng, Y., Lei, H., Li, X., and Lin, Y. (2018). An im-

proved deep neural network model for job matching.

In 2018 International Conference on Artiﬁcial Intelli-

gence and Big Data (ICAIBD), pages 106–112.

Frantzi, K. T., Ananiadou, S., and Tsujii, J. (2002). The

C-value/NC-value Method of Automatic Recognition

for Multi-word Terms. Research and Advanced Tech-

nology for Digital Libraries, 1513:585 – 604.

Guo, S., Alamudun, F., and Hammond, T. (2016).

esuMatcher: A personalized r

esum

e-job matching

system. Expert Systems with Applications, 60:169–

182.

Kessler, R., B

echet, N., Roche, M., Torres-Moreno, J. M.,

and El-B

eze, M. (2012). A hybrid approach to manag-

ing job offers and candidates. Information Processing

and Management, 48(6):1124–1135.

Kiselyova, E. I., Koroshchenko, K. R., and Robson, G.

(2021). Legal Horizons Content of the Job Descrip-

tion: Features and Areas of Concern. Legal Horizons,

14(2):63–69.

Le Vrang, M., Papantoniou, A., Pauwels, E., Fannes, P.,

Vandensteen, D., and De Smedt, J. (2014). ESCO:

Boosting job matching in Europe with semantic inter-

operability. Computer, 47(10):57–64.

Martin Jr., D., Prabhakaran, V., Kuhlberg, J., Smart, A.,

and Isaac, W. S. (2020). Extending the Machine

Learning Abstraction Boundary: A Complex Sys-

tems Approach to Incorporate Societal Context. arXiv

2006.09663.

Mc Gurk, S., Abela, C., and Debattista, J. (2017). To-

wards Ontology Quality Assessment. http://ceur-ws.

org/Vol-1824/ldq paper 2.pdf.

Pavlick, E. and Kwiatkowski, T. (2019). Inherent disagree-

ments in human textual inferences. Transactions of

the Association for Computational Linguistics, 7:677–

694.

Reimers, N. and Gurevych, I. (2019). Sentence-BERT: Sen-

tence Embeddings using Siamese BERT-Networks.

In Proceedings of the 2019 Conference on Empiri-

cal Methods in Natural Language Processing, pages

3982–3992. Association for Computational Linguis-

tics.

Rose, S., Engel, D., Cramer, N., and Cowley, W. (2010).

Automatic Keyword Extraction from Individual Docu-

ments, chapter 1, pages 1–20. John Wiley and Sons.

Roy, P. K., Chowdhary, S. S., and Bhatia, R. (2020). A Ma-

chine Learning approach for automation of Resume

Recommendation system. Procedia Computer Sci-

ence, 167:2318–2327.

Somodevilla Garc

ıa, M., Vilari

no Ayala, D., Pineda, I., So-

modevilla Garc

ıa, M., Vilari

no Ayala, D., and Pineda,

I. (2018). An Overview of Ontology Learning Tasks.

Computaci

on y Sistemas, 22(1):137–146.

Wang, X., Jiang, Z., and Peng, L. (2021). A Deep-Learning-

Inspired Person-Job Matching Model Based on Sen-

tence Vectors and Subject-Term Graphs. Complexity,

2021:1–11.

Yuan, Y. and Shaw, M. J. (1995). Induction of fuzzy deci-

sion trees. Fuzzy Sets and Systems, 69(2):125–139.

Zapata Jaramillo, C. M. and Arango Isaza, F. (2009). The

UNC-method: a problem-based software develop-

ment method. Ingenier

ıa e Investigaci

on, 29:69–75.

Zehtab-Salmasi, A., Feizi-Derakhshi, M.-R., and Balafar,

M.-A. (2021). FRAKE: Fusional Real-time Auto-

matic Keyword Extraction. arXiv 2104.04830.

Zhao, J., Wang, J., Sigdel, M., Zhang, B., Hoang, P., Liu,

M., and Korayem, M. (2021). Embedding-based Rec-

ommender System for Job to Candidate Matching on

Scale. arXiv 2107.00221v1.

Zhu, C., Zhu, H., Xie, F., Ding, P., Xiong, H., Ma, C., and

Li, P. (2018). Person-Job Fit: Adapting the Right Tal-

ent for the Right Job with Joint Representation Learn-

ing. ACM Transactions on Management Information

Systems, 9:1–17.

FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications

174