Comparing Competences on Academia and Occupational Contexts based

on Similarity Measures

Alexandra Gonz

alez Eras

, Pablo Quezada S.

, Patricia Lude

na Gonz

alez

and Carolina Gallardo

Departament of Computer Science, Universidad T

ecnica Particular de Loja, UTPL, 1101608, Loja, Ecuador

Departamento Organizaci

on y Estructura de la Informaci

on, Universidad Polit

ecnica de Madrid, 28031, Madrid, Spain

Keywords:

Body of Knowledge, Machine Learning, Similarity Measures, Professional Proﬁle, DISCO II, VSM, NLP.

Abstract:

This paper is a ﬁrst contribution of a schema to match professional and work competences through similarity

measures. In this contribution we focus on the determination of connections between university proﬁles based

in standards (body of knowledge and thesauri) similarity measures and Natural Language Processing (NLP)

techniques. Our ﬁrst experiments proved that this hybrid schema got a promise results in the recognition of

competency patterns in order to apply in the laboral context.

1 INTRODUCTION

One of the main concerns of the software industry is

to develop the talent of its human resources, since; the

quality and innovation of its products and services de-

pend on a great extent of knowledge, and the ability

and the talent of software engineers. Therefore, the

relationship university-employment becomes a cycle

comparison of skills, college proﬁles which meet the

competencies of future professional while labor pro-

ﬁles have the skills required to ﬁll a job position.

However, in reality it is almost impossible to com-

pare competencies, mainly due a problems as: in-

compatible proﬁles (Fazel-Zarandi and Fox, 2013)

(Stevens, 2013) and unstructured, ambiguous, and

sometimes incomplete data (Fazel-Zarandi, 2013).

Looking for a solution, models and platforms have

been proposed in order to proﬁle standardization

(Draganidis and Mentzas, 2006) and comparison

through competency frameworks (e-CF

, SIOC

O*NET

). Nevertheless, the actors rarely use these

tools, or they have only been proposed for one lan-

guage context without a real application in others. As

a result, is difﬁcult to obtain a standardization of pro-

ﬁles that permit the comparison of competence.

e-Qualiﬁcations Framework, available online at http://

www.ecompetences.eu

Semantically-Interlinked Online Communities, avail-

able online at http://sioc-project.org

Occupational Information Network, available online at

https://www.onetonline.org

This paper is a ﬁrst approximation to develop a

strategy in order to compare professional and work

competences. In order to develop the ﬁrst experiment,

we used the standard DISCO II to compared with col-

lege proﬁles and thereby obtain a middle ground that

allows us to meet the reporting inconsistencies. In the

same context, we focus on the combination of simi-

larity measures (Harispe et al., 2013), (Turney et al.,

2010), (Turney, 2006) and NLP techniques based on

n-grams to ﬁnd common patterns between university

proﬁles and DISCO II.

2 CONTEXT

The Fig. 1 shows a picture about the context of this

research where university proﬁles are showed as cir-

cles, job offers as triangles and competencies are rep-

resented by the colors of the ﬁgures.

We realize that competencies have a different de-

gree of presence in the proﬁles, as represented by the

different sizes of the ﬁgures. Also, there are groups of

proﬁles (academic or curriculums), covering a greater

or lesser extent job offers; also they share competen-

cies in each context, besides new skills required in

the job offers. Otherwise, standards offer a compe-

tence computational representation, that sometimes it

is insufﬁcient to show competency meanings.

Therefore, we comprise the main issues of this re-

search in the following main points:

1. Job market always requires competencies that

540

González Eras A., Quezada S. P., Ludeña González P. and Gallardo C..

Comparing Competences on Academia and Occupational Contexts based on Similarity Measures.

DOI: 10.5220/0005491405400546

In Proceedings of the 11th International Conference on Web Information Systems and Technologies (WEBIST-2015), pages 540-546

ISBN: 978-989-758-106-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Context problematic.

Universities have problems to cover (new/missing

competencies). Sometimes the proﬁles have dif-

ferent structure and lack of information. Besides,

the proﬁles clearly not describe the competencies

or the competency elements, so we cannot make a

comparison between them (Paquette et al., 2012).

The job proﬁles are more close to describe job

features like activities or roles and not like com-

petencies (Malzahn et al., 2013). Moreover, the

curriculum for a university study normally gives

only dependencies between courses and a generic

description de capacidades en un area especﬁca

(Dorn and Pichlmair, 2007). These descriptions

do not have a clear relation with competencies and

por ende es difcil identiﬁcar nuevas competencias

as como a aquellas competencias que fueron omi-

tidas o no cubiertas por el perﬁl (Fazel-Zarandi,

2013).

2. Job offers and university proﬁles share compe-

tencies, but with different level of meaning ( dif-

ferent conceptualization). Mainly due to differ-

ent interpretations that each actor has about the

deﬁnition of competency (Fazel-Zarandi and Fox,

2009). For instance, the term competency ap-

pears to be used at times to refer to actions and

their consequences and at others to refer to cogni-

tive skills and personality characteristics (Stevens,

2013) , or the competencies of individuals may

be expressed in terms of qualiﬁcations and cer-

tiﬁcations, such as academic degrees (Malzahn

et al., 2013) or as learning outcomes within edu-

cational processes (Paquette, 2007). Other aspect

to consider is the degree of performance of com-

petences in which the labor market related activ-

ities and functions (Bizer et al., 2005), whereas

in academia is related to generic descriptions of

occupational areas (Dorn and Pichlmair, 2007).

3. Standards have few ﬂexible and complex

schemas. Standards have the function of facilitate

the exchange of competence descriptions. Stan-

dards as : IMS RCDEO 2002, IEEE RCD 2004,

HR-XML are the core of many Applications

of Manager of Competences (Draganidis and

Mentzas, 2006). In the same context each indi-

vidual prefers to use a free vocabulary to convey

their competencies (Fazel-Zarandi and Fox,

2009). Similarly thesauri and taxonomies have

been proposed to deﬁne skills and competencies

(SIOC, O*NET, DISCO II) but, stakeholders do

not use these schemas to create their documents

(Malzahn et al., 2013), because they often

lack motivation to add them into their proﬁles

(Hansen et al., 2011) because the speciﬁcations

are difﬁcult to apply (Malinowski et al., 2006),

therefore over time the proﬁles become out-dated

(Fazel-Zarandi and Fox, 2013).

3 RELATED WORKS

3.1 NLP Techniques

NLP techniques contribute to the recognition of the

text portions used for comparison. In this sense, many

of works in the literature focused on the connection of

ontologies With NLP tools, to extract competencies

as Entities and relations. In (Malzahn et al., 2013)

The NLP techniques are used to extract patterns Com-

petency. In (Janev and Vranes, 2011) a framework

that extracts patterns of personal and enterprise skills

since different text sources used. (Yahiaoui et al.,

2006) uses a semantic annotation scheme based on a

multi ontology framework for labeling CVs and job

offers by STI concepts instances.

3.2 Similarity Measures

When we think about similarity, we associate the per-

spective of two entities sharing in some degree a set

of characteristics, and a similarity measure capturing

the strength of this semantic interaction in connec-

tion with their meaning (Turney et al., 2010). Mea-

sures could estimate the similarity/dissimilarity be-

tween speciﬁc kind of semantic representation on

which is based the comparison, for instance units of

language as words, sentences, paragraphs and docu-

ments (Harispe et al., 2013).

This idea represent the semantic between Vector

Space Model (VSM) in which Vector Space Model

ComparingCompetencesonAcademiaandOccupationalContextsbasedonSimilarityMeasures

541

(VSM), is consider how as a point in a space (a vector

in a vector space). Points that are close together in this

space are semantically similar and points that are far

apart are semantically far (Turney et al., 2010). For

instance, to measure the similarity between a query

and a document (Manning et al., 2008), or within al-

gorithms that measures the similarity of semantic re-

lations (Turney, 2006).

VSMs have several attractive features. VSMs ex-

tract knowledge automatically from a given corpus,

thus they require much less Labour than other ap-

proaches to semantics, such as lexicons, thesauri and

ontologies (Manning et al., 2008).

In the same context, VSM is used in job re-

cruitment recommender systems para rankear proﬁles

based in candidate-offer matrixs (Linden et al., 2003),

and the hybrid schema where with the combination

of NLP techniques and similarity measure for ex-

ample Coseno, to cluster the competences (Malzahn

et al., 2013). In (Buitelaar and Eigner, 2008) linguis-

tic patterns of knowledge are combinated with dis-

tance measures like Euclidean L1norm to build se-

mantic networks based on term frequency. In (Re-

ichhold et al., ) a measure of role relevance added to

Cosine distance match job offer vectors.

4 METHODOLOGY

In order to develop this paper we proposed the next

methodology based in 3 stages. (Figure 2).

Figure 2: First proposal of a hybrid schema to get a baseline.

4.1 Preprocessing

The sources for our corpus was taken from the follow-

ing sources:

• Career proﬁles of Software Engineering and re-

lated, which were taken from university websites

of Latin American universities, selected from the

database of Webometrics

Ranking Web of Universities, available at http://

www.webometrics.info/

In order to choose the candidates phrases in uni-

versity proﬁles we select paragrahps of the sec-

tions: description, occupational ﬁeld, skills, com-

petences and knowledge areas, in which show

competences sentences. In the Figure 3 we can

show these competences.

Figure 3: Example of academic competencies.

To resolve the problem of complexity of sentences

we divided the paragraphs following the proposed

scheme: abilities represent in verbal structure and

some topics related with the knowledge area that

represent the nominal structure. In the case of

complex sentences (more than one verb) we per-

form sentence repetition based on the following

rule: for each verb in a sentence, a new sentence

will be added which is the same sentence but with

one verb.

• Standards: we are considering the standard

DISCO II which offers competence phrases that

have been chosen based on a consensual process.

4.2 NLP Process

The process of NLP is based on a superﬁcial repre-

sentation of texts (Metzler et al., 2007), in which we

submit the corpus to the following tasks:

1. Tokenization, stop words: words are separated

according to the spaces between them based on

the Snowball list of stop words for the Spanish

language, then a manual review of the words, to

eliminate cases that were not considered (articles,

prepositions, conjunctions and numbers).

2. Lemmatization: to regularize surface variations of

words by converting them to the same form with

reducer el sparsity problem. The types of normal-

ization applied were: case folding (converting all

words to lower case) and lemmatization (changing

the words by its canonical form).

3. N-grams: we obtained unigrams and bigrams

since university proﬁles (P) and standard (D), with

the propose of develop a ﬁrst identiﬁcation of

common pattern between corpus.

4.3 Similarity Measures Process (SM

Process)

In the SM process we focused on the combination of

similarity measures as example Cosine 1, which al-

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

542

lows comparing two frequency vectors whether they

be raw o weighted or they have different lengths (Tur-

ney et al., 2010). Cosine captures the idea that the

length of the vectors is irrelevant; the important thing

is the angle between the vectors (Deerwester et al.,

1990).

cos(x, y) =

∑

i=1

∗ y

)

∑

i=1

∑

i=1

(1)

As a complement, we used a lexical character 2,

that calculated the relation between unigrams and bi-

grams, giving high weight to similar bigrams. Then,

we built a matrix of similarities between both cor-

puses (D x P matrix) counting the number of simi-

lar unigrams (nu) and the number of similar bigrams

(nb). The value of similarity gave more importance to

bigrams. In the same context we develop a distance

analysis where D x P matrix was submitted to VSM,

which included Singular Value Decomposition tech-

nique to reduce the Sparsity. Additional we develop a

n-grams in order to search patters compose by verbal

and nominal structures.

SM = nu ∗ 0.5 + nb (2)

5 EXPERIMENTATION

With the goal of matching university proﬁles (P) with

the standard DISCO II (D), we conducted the follow-

ing experiments:

• Distance Analysis: to compare both corpus and

get the distances between them. We also use

the entropy to determine whether it is possible to

compare P and D with no other additional infor-

mation.

• Perfect alignment of P and D: to establish the sim-

ilarity between P and D on the basis of different

types of alignment.

• Bigrams analysis: to detect the occurrence of pat-

terns.

5.1 Distance Analysis

For the analysis of distances, D x P matrix was then

submitted to SVD under the framework of a distance

matrix and cosine as a distance measure. The Table1

shows preliminary results:

We get that the maximal distance is 1.33, the av-

erage distance is around 0.82, and the standard devi-

ation is around 0.21. With these numbers, we clearly

Table 1: Results of Distance Analysis between P and D.

Criteria Values

Maximal distance 1.3271

Average distance 0.8212

Standard deviation 0.2082

(a) SVD of D versus P (SVD = U * S * V

: where U=

proﬁles and V= disco)

(b) Disco versus Proﬁles (DxP)

Figure 4: Plot 2D of SVD over D and P.

know that most of the sentences are dissimilar. The

Figure 4 shows the results of the SVD in 2D.

In the second plot we can conﬁrm the dissimilarity

between D and P (the white space), but also we notice

that there is a lot of redundancy in P since many sen-

tences differ from others by only one token. This is

due to the fact that P contains many repetitions To the

diagonal analysis we start with a following hypoth-

esis: the diagonal of the distance matrix is distorted

when information is missing. The Figure 4 conﬁrms

that D and P have a distorted diagonal, which means

that D and P share few similar sentences. We can see

these sentences in strong red. We can see peaks in

the graph, which means that some sentences in P are

similar to many sentences in D. Besides; many sen-

tences of P (the white area over the phrase 100) are

not covered at all in D (in x).

5.2 Alignment of Corpus

To achieve a measure of similarity, the purpose is to

ComparingCompetencesonAcademiaandOccupationalContextsbasedonSimilarityMeasures

543

ﬁnd a perfect alignment between D and P. Then, How

does it looks when the alignment is perfect that is if

the proﬁles are just like the standard?. Likewise, we

made some tests with the purpose of compare D and

5.2.1 Comparing P with P

The Figure 5 shows plots of SVD over P with P, which

we can see that the similarity matrix is just perfect.

Also, the matrix of cosine distance gives a good diag-

onal.

(a) SVD of P versus P (SVD = U * S * V

: where U

and V= proﬁles)

(b) Proﬁles versus Proﬁles (PxP)

Figure 5: Plots 2D to compare of P with P.

Although, some sentences in P are redundant (big

rectangles in the diagonal). So, some sentences in P

are similar with many others (up to 6 or 7).

5.2.2 Comparing D with D

The Figure 6 shows the plots of SVD between D with

D, we can see that the similarity matrix of D with D.

The matrix of cosine distance gives reveal of

course a diagonal, but not so clean. It means the D

is not so clean as we thought. Many sentences are

similar to others; this is due to the fact that the en-

tire hierarchy of DISCO II is used. Besides, the plot

shows many sentences of D are similar with others in

D (as shown by the peaks).

Figure 6: Plots 2D to compare of D with D.

Figure 7: First proposal of a hybrid schema to get a baseline.

In 7, we can show in the remarked bigrams which

involve a verb like ”gestionar” or a nominalization

like ”diseo” followed by a noun, these kind of bi-

grams very often express competence. Therefore, we

suppose that removing the verb we will remove many

competences.

Besides we found bigrams like base dato and sis-

tema which gives us a idea about common patterns

and the posibles patterns for new experiments, as the

case of mapping of domain areas.

6 ANALYSIS AND DISCUSSION

The comparison scheme based on unigrams and bi-

grams and similarity measures is promising to ﬁnd

similar items. The results obtained in the bigrams

analysis conﬁrm the need to explore other possibili-

ties of combinations between noun phrases and verb

phrases. The use of patterns can provide new possi-

bilities for the interpretation of isolated values.

There is much redundancy in P and D and it is

necessary to perform a pre-cleaning over the corpus

before applying any comparison scheme. The inter-

pretation of negative values (isolated) is not clear in

the case of P x P matches. We cannot be certain that

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

544

these values are equivalent to missing information or

a derivative problem of redundancy in P and D. In

future experiments should address the analysis of the

speciﬁc statements that cause these values.

The comparison of P and D could be performed

at different levels of the standards hierarchy, in order

to reduce the number of outliers. To achieve greater

entropy in the matching we have to consider the use of

parts of DISCO II rather than all the entire hierarchy

and thus reach a lower level of redundancy in P x D.

7 CONCLUSIONS AND FUTURE

WORKS

We propose a model to compare university and job

offer proﬁles based on similarity measures.

The use and combination of different similarity

measures to get a high performance will be develop-

ing. Also, using standards to framework construction

and validation permit give a solution of vocabulary

mismatch problem.

As a result, we expect to get similarity indicators

between university and job offer proﬁles.

We propose a comparison with job proﬁles based

on competencies and guides the referents in engineer-

ing context.

ACKNOWLEDGEMENTS

We thank our colleagues from Politechnical Univer-

sity of Madrid and Universidad T

ecnica Particular de

Loja who provided insight and expertise that greatly

assisted the research of this paper.

REFERENCES

Bizer, C., Heese, R., Mochol, M., Oldakowski, R., Tolks-

dorf, R., and Eckstein, R. (2005). The impact of

semantic web technologies on job recruitment pro-

cesses. In Wirtschaftsinformatik 2005, pages 1367–

1381. Springer.

Buitelaar, P. and Eigner, T. (2008). Topic extraction from

scientiﬁc literature for competency management. In

The 7th International Semantic Web Conference.

Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas,

G. W., and Harshman, R. A. (1990). Indexing by latent

semantic analysis. JAsIs, 41(6):391–407.

Dorn, J. and Pichlmair, M. (2007). A competence manage-

ment system for universities. In ECIS, pages 759–770.

Draganidis, F. and Mentzas, G. (2006). Competency

based management: a review of systems and ap-

proaches. Information Management & Computer Se-

curity, 14(1):51–64.

Fazel-Zarandi, M. (2013). Representing and Reasoning

about Skills and Competencies over Time. PhD the-

sis, University of Toronto.

Fazel-Zarandi, M. and Fox, M. S. (2009). Semantic match-

making for job recruitment: an ontology-based hybrid

approach. In Proceedings of the 8th International Se-

mantic Web Conference.

Fazel-Zarandi, M. and Fox, M. S. (2013). Inferring and

validating skills and competencies over time. Applied

Ontology, 8(3):131–177.

Hansen, D. L., KHOPLAR, H., and Zhang, J. (2011).

Recommender systems and expert locators. Under-

standing Information Retrieval Systems: Manage-

ment, Types, and Standards, pages 435–447.

Harispe, S., Ranwez, S., Janaqi, S., and Montmain, J.

(2013). Semantic measures for the comparison

of units of language, concepts or instances from

text and knowledge base analysis. arXiv preprint

arXiv:1310.1285.

Janev, V. and Vranes, S. (2011). Ontology-based com-

petency management: the case study of the mihajlo

pupin institute. J. UCS, 17(7):1089–1108.

Linden, G., Smith, B., and York, J. (2003). Amazon. com

recommendations: Item-to-item collaborative ﬁlter-

ing. Internet Computing, IEEE, 7(1):76–80.

Malinowski, J., Keim, T., Wendt, O., and Weitzel, T. (2006).

Matching people and jobs: A bilateral recommenda-

tion approach. In System Sciences, 2006. HICSS’06.

Proceedings of the 39th Annual Hawaii International

Conference on, volume 6, pages 137c–137c. IEEE.

Malzahn, N., Ziebarth, S., and Hoppe, H. U. (2013). Semi-

automatic creation and exploitation of competence on-

tologies for trend aware proﬁling, matching and plan-

ning. Knowledge Management & E-Learning: An In-

ternational Journal (KM&EL), 5(1):84–103.

Manning, C., Raghavan, P., and Sch

utze, H. (2008). Intro-

duction to information retrieval. cambridge university

press, cambridge, uk.

Metzler, D., Dumais, S., and Meek, C. (2007). Similarity

measures for short segments of text. Springer.

Paquette, G. (2007). An ontology and a software framework

for competency modeling and management. Educa-

tional Technology & Society, 10(3):1–21.

Paquette, G., Rogozan, D., and Marino, O. (2012). Com-

petency comparison relations for recommendation in

technology enhanced learning scenarios. In Pro-

ceedings of the 2nd Workshop on Recommender Sys-

tems for Technology Enhanced Learning (RecSysTEL

2012), volume 896, pages 23–34. Citeseer.

Reichhold, M., Kerschbaumer, J., Fliedl, G., and Winkler,

C. Automatic generation of user role proﬁles for opti-

mizing enterprise search.

Stevens, G. W. (2013). A critical review of the science and

practice of competency modeling. Human Resource

Development Review, 12(1):86–107.

ComparingCompetencesonAcademiaandOccupationalContextsbasedonSimilarityMeasures

545

Turney, P. D. (2006). Similarity of semantic relations. Com-

putational Linguistics, 32(3):379–416.

Turney, P. D., Pantel, P., et al. (2010). From frequency to

meaning: Vector space models of semantics. Journal

of artiﬁcial intelligence research, 37(1):141–188.

Yahiaoui, L., Boufa

ıda, Z., and Pri

e, Y. (2006). Semantic

annotation of documents applied to e-recruitment. In

SWAP.

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

546