Conference PaperPDF Available

Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements

Automated Analysis of Job Requirements for Computer Scientists in
Online Job Advertisements
Joscha Gr¨
uger1,2 a and Georg J. Schneider1 b
1Computer Science Department, Trier University of Applied Sciences, Main Campus, Trier, Germany
2University of Trier, Department of Business Information Systems II, 54286 Trier, Germany
Data Analysis, Web Mining, Natural Language Processing, Information Retrieval, Machine Learning, Job Ads,
The paper presents a concept and a system for the automatic identification of skills in German-language job
advertisements. The identification process is divided into Data Acquisition, Language Detection, Section
Classification and Skill Recognition. Online job exchanges served as the data source. For identification of the
part of a job advertisement containing the requirements, different machine-learning approaches were compared.
Skills were extracted based on a POS-template. For classification of the found skills into predefined skill classes,
different similarity measures were compared. The identification of the part of a job advertisement containing
the requirements works with the pre-trained LinearSVC model for 100% of the tested job advertisements.
Extracting skills is difficult because skills can be written in different ways in the German language – especially
since the language allows ad-hoc creation of compound. For extraction of skills, POS templates were used.
This approach worked for 87.33% of the skills. The combination of a fasttext model and Levenshtein distance
achieved a correct assignment of skills to skill classes for 75.33% of the recognized skills. The results show
that extracting required skills from German-language job ads is complex.
The labor market for IT specialists is very complicated
due to the always evolving and changing IT environ-
ment. New programming languages, frameworks and
development concepts appear on a yearly basis. Some
of these technologies and concepts are very trendy
only for a certain period, some manifest and play an
important role for a long time. For a company it is
vital to find IT specialists that are competent with the
strategic technologies used in their enterprise. In a
study by Bitkom in the year 2015, seven out of ten ICT
companies (70 percent) state that there is currently a
shortage of IT specialists. In 2014, only 60 percent of
those surveyed said that there was a shortage (Bitkom,
2016). 38.7 percent of the IT vacancies forecasted for
the top 1000 companies in Germany in 2015 are diffi-
cult to fill due to a lack of a fitting qualification, and
6.1 percent of the vacancies will remain vacant due
to a lack of suitable candidates (Weitzel et al., 2015).
Hence it is vital for a national economy to educate
students that have not only an excellent qualification
in the theoretical foundations of Computer Science but
also up-to-date skills in current technologies for this
urgent need. Likewise, being competent with technolo-
gies in demand, it helps the alumni to easily find a job.
Universities of Applied Sciences in Germany usu-
ally target praxis-oriented topics and they have tradi-
tionally a close link to the professional field. There-
fore lab courses, student projects or applied research
projects usually reflect the current trends in the IT
to complete the theoretical foundations with practi-
cal experiences using modern IT systems. Hence the
students can easily find a job with the competencies
acquired during their education. Universities try to
solve the problem of staying up-to-date with their ed-
ucation by regularly reviewing their curricula. They
observe developments in the industry to offer either
specialized courses that complement the sound theo-
retical foundation of their study program or to change
programming languages, frameworks or systems for
the lab courses or projects. Students usually embrace
these changes and also gladly choose new course of-
ferings to get the best possible chances on the labor
market. For both students and universities, the ques-
tion is, what should be taught on the practical side and
Grüger, J. and Schneider, G.
Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements.
DOI: 10.5220/0008068202260233
In Proceedings of the 15th International Conference on Web Information Systems and Technologies (WEBIST 2019), pages 226-233
ISBN: 978-989-758-386-5
Copyright c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
what should students learn to best meet the require-
ments of the industry? One possibility would be to
manually review the job advertisements in different
newspapers and online platforms. However, this is a
tedious and time-consuming task. Our research hy-
pothesis is that this task can be automated in a way
that the findings are at least nearly as good as a human.
German, however, is a challenge compared to English,
where similar approaches already exist, as German
has a rich morphology, umlauts, four cases and much
fewer corpora for training.
This paper describes an approach for an automated
survey of current job requirements for computer sci-
entists on the German labor market. Obviously, this
survey has to be carried through regularly to identify
trends and manifesting technologies.
The suggested process to realize the automated
survey is based on technologies of Natural Language
Processing and Machine Learning. The results of the
analysis can be used, for example, to check the content
of university curricula or to identify new technological
trends at an early stage. In addition, the procedure can
be used to implement a skill-based job search. The
next section discusses the current state of research and
various papers regarding the extraction of skills from
job ads. Section 3 introduces our approach to job ad
analysis. Both the processing steps and the evaluation
of each step are described in detail. Finally, Section 4
concludes the results and discusses future work.
Some studies conclude that education does not comply
with the requirements of employers in the IT sector.
Kwon Lee and Han (2008) for example concluded
for the US labor market that most universities attach
great importance to hardware and operating systems,
although the employers surveyed are rarely interested
in these skills. They see for example deficits in the
teaching of skills in the economic and social cate-
gory. Yongbeom et al. (2006) also speak of a skill
gap between employers’ requirements and universi-
ties’ IT curricula. Among other things, the lack of
project management, Enterprise-Resource-Planning
(ERP) and information security modules in the cur-
ricula is criticized. Scott et al. (2002) criticize poor
database knowledge and the lack of skills in CASE/-
modeling tools and Business Process Reengineering
(BPR) techniques among graduates. Students also
lacked skills in XML and iterative development.
There are a number of reasons for the skill gap:
one is the rapid technological change, another is the
discordance between the content of the curricula of
the universities and the required competencies of the
industry (Scott et al., 2002; Milton, 2000). In addition,
too long revision cycles of curricula relative to the
speed of technology change and a lack of knowledge
at universities about new and upcoming technology
are cited as reasons for the gaps (Lee et al., 2002).
In order to keep curricula up-to-date, universities
need to know which competencies are currently and
in the long term required by the industry. For iden-
tifying the skills required various approaches exist.
Prabhakar et al. (2005) researched online job adver-
tisements for computer scientists in the US in 2005
with regard to the changing demand for skills over the
year. For this purpose, they examined the job advertise-
ment to see whether it contained one of 59 keywords
or not. The approach for identifying requested skills
in the IT sector of Gallagher et al. (2010) is interview-
based. His team interviewed 104 senior IT managers.
The questions were very general, e.g. it was asked
whether programming skills were required, and not for
concrete programming languages like Java. Sibarani
et al. (2017) developed an ontology-guided job market
demand analysis process. Their method is based on
the self-defined SARO ontology and a defined set of
skills. Using the predefined skills and ontology, they
perform a named-entity tagging. The identified skills
are linked by a co-word analysis. Litecky et al. (2010)
used web and text mining techniques for retrieving
and analysing a data set of 244.460 job ads. They
scraped the data from online job exchanges to extract
titles and requirements based on predefined keywords.
Wowczko (2015) took a different approach. They anal-
ysed descriptions of vacancies and reduced the words
used in the descriptions until only significant words
remained. Custom word lists, stemming, removing
stopwords, removing numbers, stripping whitespaces,
etc. were used to clean up the data.
The problem with all these approaches except
Wowczko (2015) and Gallagher et al. (2010) is that
they are based on fixed keyword lists. Thus, only abil-
ities contained in the lists are recognized. New tech-
nologies or skills described in any other way cannot be
recognized. Abbreviations like Active Directory and
AD are assigned to different classes or remain unrecog-
nized. Additionally, some processes were performed
manually and are hence rather time-consuming and are
only carried out periodically. Wowczko (2015) also
finds false positives like strong,excellent and can. Con-
sequently, an automated procedure that monitors job
advertisements permanently would simplify this proce-
dure enormously. Thus the approaches also search in
areas of the job advertisement where no requirements
are described (e.g. in the company description).
Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements
This paper describes a concept and a resulting auto-
mated procedure to search for competencies in all
job advertisements. The approach also recognizes
unknown competencies and maps them to the same
class if they are semantically similar or equal. For
the extraction of the skills, only the area in which the
requirements for the applicant are formulated is used.
The focus of the approach is on job advertisements
in German. The target group of the analyzed job ad-
vertisements are computer scientists. Online job ex-
changes such as and were used
as data sources. The process for identifying skills is
divided into four steps (see figure 1): Data Acquisition,
Language Detection, Section Classification and Skill
3.1 Data Acquisition
A web crawler based on the Scrapy framework was
developed for data retrieval. The web crawler searches
German online job exchanges for jobs in the IT sector.
The online job exchanges, and were used as data sources. The
crawler extracts the HTML job ads found, as well as
metadata such as company name, job title, and work
location. The defined process works on the HTML
files of the job advertisements. The metadata could be
used for extensions such as geographic analysis or job
title analysis.
3.2 Language Detection
The result of the data acquisition are all job ads, di-
rected at computer scientists of the online job ex-
changes mentioned. Seven percent of the extracted
job advertisements are written in English, 93% of the
advertisements are written in German. As described,
the approach is aimed at ads in German, the English-
language job advertisements must be filtered out.
In order to be able to filter out non-German adver-
tisements a language recognition according to Shuyo
was implemented (Shuyo, 2010). The basic assump-
tion of the approach is that certain n-grams occur
in different frequency in different languages (see ta-
ble 1). Based on a quantitative corpus analysis using
Wikipedia pages, all mono-, bi- and trigrams of the
languages to be detected were counted and the proba-
bility with which each n-gram occurs in the respective
language was calculated. Shuyo provides the results
of the quantitative analysis in so-called profile files.
Table 1: Comparison of the probability with which the n-
grams occur in the languages German and English (based on
Shuyos profile files.
n-gram German English
D 0.00648 0.00253
ie 0.01158 0.00259
hum 0.00013 0.00029
To recognize the language, the text to be classi-
fied is fragmented into mono-, bi- and trigrams. The
algorithm randomly selects individual n-grams and
calculates the probability with which these occur in
the languages tested (Shuyo, 2010).
For evaluation, 741 job advertisements were classi-
fied by language. 8.77% of the job ads tested were in
English, 91.23% in German. The algorithm achieved
an accuracy of 100% and examined a maximum of 15
n-grams per text.
3.3 Section Classification
After all English-language job adverts have been fil-
tered out, the position of the requirements within the
job advertisement must be identified. Position means
in which HTML list element the requirements are
listed. For this purpose classification algorithms of
machine learning were trained and compared.
For training 150 job advertisements were seg-
mented. To this end, all HTML list elements with
more than 20 characters contained in the job adver-
tisements were assigned to a category. The categories
are Requirements,Offer,Tasks and uncategorized. Re-
quirements describes the employer’s requirements for
the applicant. Elements classified as tasks contain the
tasks to be performed on the position. Offer-sections
contain information about the offers and benefits of
the company. The models were trained based on the
segmented data of 50 job advertisements. The test
corpus includes 100 segmented job advertisements.
The scikit-learn
implementations of Random-
ForestClassifier, LinearSVC (Linear Support Vector
Classification), MultinomialNB (Naive Bayes classi-
fier for multinomial models), LogisticRegression, De-
cisionTreeClassifier and BernouliNB (Naive Bayes
classifier for multivariate Bernoulli models) were
trained and compared in the scikit-learn standard con-
figuration. For tokenization, a word tokenizer was
used. Features were represented by a count matrix.
The result of the count matrix was transformed into
normalized tf (term frequency) representation, contain-
ing 3349 features. To compare the algorithms, each
algorithm was trained 5 times on 50 segmented job ads
1Version: 0.20.3 (
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
Item Pipelines
Data Aquisition
filter out non German ads
using n-Grams
Language Detection
Support Vector Maschine
using nlkt
Categories: Tasks,
Requirements, Offer and
Section Classification
Res. detached compounds
Skill Cleaning
Semantic Matching
Skill Recognition
Figure 1: System architecture.
Figure 2: Box plot of 5 tests per algorithm. LinearSVC
reached an accuracy of 0.950291.
Table 2: Comparing accuracy, precision and recall. The fea-
tures were tokenized by words and the frequency measured
by TF.
algorithm accuracy precision recall
RandomForestCl. 0.7496 0.7215 0.5933
LinearSVC 0.9502 0.9416 0.9434
MultinomialNB 0.8902 0.8804 0.8103
LogisticRegr. 0.8874 0.8754 0.8043
DecisionTreeCl. 0.8100 0.8115 0.8249
BernoulliNB 0.8690 0.8185 0.7664
and tested with 100 job ads. Rated by accuracy Lin-
earSVC and MultinomialNB were the best algorithms.
LinearSVC reached an average accuracy of 0.950291
and MultinomialNB 0.890297 (see figure 2 and table
The results are achieved because skill/requirements
sections contain words like
(eng. knowl-
(eng. ideally) and
good) in high frequency and almost exclusively (fig-
ure 3). For optimization, the classification algorithms
were also tested with stemmed features, on a feature
set with and without stop words, using n-grams, us-
ing TF-IDF (inverse document frequency) and with
a pruned feature set. Table 3 shows that LinearSVC
using TF-IDF, without further optimization, gives the
best results in terms of skill section recognition.
The confusion matrix in table 4 shows that all re-
quirement sections were correctly recognized. In ad-
dition to requirement sections, other areas were also
classified as requirement sections. This leads to a pre-
oder gute
80 Requirements
Figure 3: Frequency with which the features occur in require-
ment sections and in other areas (Tasks, Offer, uncategorized,
based on 50 job ads).
Table 3: Comparing algorithms by accuracy (acc), precision
(pre) and recall (rec) for different settings. For stemming
NLTK (Natural Language Toolkit) German Snowball is used.
The stop word list contains 621 German stop words. Pruning
means removing tokens that are included in most job ads or
in almost no job ads (<0.5% or >99.5%).
LinSVC Mult.NB LogReg
0.9502 0.8902 0.8874
0.9416 0.8804 0.8754
0.9434 0.8103 0.8043
TF-IDF, prune
>99.5 %
0.9368 0.8899 0.9421
0.9321 0.9543 0.9247
0.9308 0.9274 0.8163
0.9372 0.9032 0.8821
0.9471 0.8875 0.8799
0.8960 0.8249 0.7963
0.9449 0.9137 0.9006
0.9512 0.9426 0.8895
0.9116 0.8424 0.8193
stem, prune
0.9393 0.9000 0.9052
0.9245 0.8724 0.9248
0.9283 0.8205 0.8312
0.9163 0.8902 0.8240
0.9438 0.8820 0.8439
0.8436 0.8103 0.7227
cision of 0.826 for the Requirements class. For the
further recognition of the skills, this would mean that
skills would also be searched in e.g. offer sections.
Recall and precision of the requirements class also
Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements
Table 4: Confusion matrix of prediction with LinearSVC and
TF-IDF. The results show that texts of the classes Unclassi-
fied, Task and Offer were also classified as Requirements.
n =
Unc. Offer Req.
Task recall
14 2 1 0 0.823
2 57 16 11 0.662
0 0
0 1
1 11 4 47 0.746
prec. 0.823 0.814
0.826 0.81
Table 5: Confusion matrix of prediction with LinearSVC and
TF-IDF under the premise that there is only one section con-
taining requirements per job ad. The Requirement Section
is the area for which the highest probability of containing
requirements has been calculated.
n =
Unc. Offer Req.
Task recall
14 3 0 0 0.823
2 76 0 8 0.883
0 0
0 1
1 9 0 53 0.841
prec. 0.823 0.864
1 0.869
show that in some job advertisements several areas
were classified as requirement sections. Therefore the
premise was defined that in every online job advertise-
ment exactly one requirement section exists. Based on
the trained model,
calculates the probability with
which text
is a requirement section. A requirement
of a job advertisement
consisting of a set
of section texts Siis defined as:
Table 5 shows the result including the premise. No
false positives are assigned to the requirements class
and the recognition of the other classes also improves.
3.4 Skill Detection
After the requirement sections have been identified,
the skills they contain must be extracted. A POS
tagging was performed for this purpose. For the
tagging the nltk ClassifierBasedTagger was trained
with the Tiger Corpus of the Institute for Machine
Language Processing of the University of Stuttgart.
The Tiger Corpus consists of 90,000 tagged tokens
and 50,000 sentences (IMS, 2018). The analy-
sis of POS tagged skills shows that they can be
included in requirement descriptions in a very hetero-
geneous form. Skills can include nouns, foreign words,
compound words, detached compounds, combi-
nations with cardinals, etc. Examples for skills
in job ads are
ISO 27001
gute Deutsch- und Englischkenntnisse
Erfahrung in der Programmierung mit Java.
A multi-stage process was defined to identify the
skills. First, detached compounds like German and
English skills are resolved. In the next step, the skills
get chunked by POS templates. False positives are
then filtered out via a blacklist. In order to be able to
count semantically similar skills in a class, they are
assigned to a class via a semantic matching.
3.4.1 Resolve Detached Compounds
In German words can be concatenated to mean the
same as the sum of two words (e.g. eng: admin-
istrator password – de:
By a conjunction, several compositions with the
same end or beginning can be shortened without
losing their meaning (e.g.
und Serveradministrator
can be written as
Client- und Serveradministrator
). In order to
understand the connections between the conjunctions,
they must be resolved.
For resolving detached compounds like shown in
the listing a compound resolver module was developed.
The compoundResolver-module searches for tokens
that are classified as TRUNC (Truncate) by the POS
tagger and end with a hyphen punctuation mark. If
the following token is classified as KON (Conjunction)
and the following as NN (regular nominal) than the reg-
ular nominal is split into syllables using the pyphen
library. Examples of resolving detached compounds
and the POS-tagging are:
Deutsch- und Englischkenntnisse
-> Deutschkenntnisse und Englischkenntnisse
Hochschul- oder Universit¨
-> Hochschulabschluss oder Universit¨
By merging the syllables, all possible combinations
were generated while keeping the order of the syl-
lables. Based on a dictionary of German nominals,
each syllable is tested. The shortest combination con-
tained in the Dictionary is identified as the first part
of the compound. The rest of the word is merged
with the truncated part of the compound. The new
String replaces the truncated token. An example of the
generated syllable combinations and the result after
resolving is:
[’Eng’, ’Englisch’, ’Englischkennt’,
’Englischkenntnis’, ’Englischkenntnisse’]
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
[(’Deutschkenntnisse’, ’NN’), (’und’, ’KON’),
(’Englischkenntnisse’, ’NN’)]
For regular nouns that are not compounds (like
Wirtschafts- (Informatik)
) the algorithm has
been extended. By checking if just the whole word
matches with an entry of the dictionary the token can
be joined with the truncated token. An example of
resolving Wirtschafts- (Informatik) is:
[(’Wirtschafts-’, ’TRUNC’), (’)’, ’$(’),
(’Informatik’, ’NN’)]
[(’Wirtschaftsinformatik’, ’NN’), (’)’, ’$(’),
(’Informatik’, ’NN’)]
3.4.2 Skill Chunking
Job requirements are specified in job advertisements
in various forms. To extract the requirements, the text
is chunked using POS templates and regular expres-
sions. Four structures were identified for matching
skills. Each structure is shown with examples and the
appropriate expression for the chunking with nltk Reg-
Sequence of nouns, cardinal numbers, proper
names or/and foreign words like:
HTML Kenntnisse
Windows 10
. The code shown
indicates that the search is for single or consecutive
occurrences of foreign words, normal nouns, proper
names, or cardinal numbers.
These sequences can be separated
by hyphens, such as
IT-System-Kaufmann,SAP IS-A.
Within a skill description, articles, preposi-
tions, or prepositions with articles can be included:
Kenntnisse [der] Informatik, Erfahrung [
im] Projektmanagement, Erfahrung [in der]
Arbeit [mit] Eclipse.
In addition, the description of a skill can con-
tain attributive adjectives, adverbial or pred-
icative adjectives:
Kenntnisse [in der]
technischen Informatik, Erfahrung [in
der] objektorientierten Programmierung.
These regular expressions were testet on a corpus
of 60 job ads. 87.33% of the skills mentioned in the
job ads could be recognized correctly. 12.67% of the
skills could not be extracted. In addition, 21.57% of
the extracted tokens are false positives.
3.4.3 Skill Cleaning
After the Skill Chunking the result contains 21.57%
false positives like
. These are not job
requirements and it is necessary to remove them from
the results. For this purpose, the results are cleaned via
a filter. The filter uses a blacklist of 117 tokens that are
classified as non-requirements. The list contains to-
kens that are often part of the requirement description
but are not requirements in themselves. These include
words like
(eng. knowledge) and
(eng. experience). The tokens were collected based on
50 job ads and on the token frequency.
The filter compares the stemmed blacklist tokens
with the stemmed extracted skills. The stemming is
used to filter out flexions of the blacklist words, too.
If these match, the extracted token is discarded. For
stemming the German Snowball Stemmer is used.
3.4.4 Semantic Matching
After the skill cleaning the set of extracted tokens
contains only skills. The extracted skills can be syntac-
tically different, but semantically similar or identical.
For example:
IT-Sicherheit [ IT-Security ,
S ic h e r he i t s a sp e k t e de r I T ]
GW T [ K e nn tn i ss e i n G WT , G o og le
We b T oo l ki t ]
In order to get an overview of the most frequently re-
quired skills, exact matching is not relevant. Therefore,
semantically similar requirements must be assigned
to the same class. To this end, different text distance
algorithms were compared. The evaluation of the algo-
rithms is based on a test data set that contains 200 skills
and relation to the correct target class. Two fast text
models (Bojanowski et al., 2017; Grave et al., 2018)
one word2vec model
and four syntactic text distance
algorithms were evaluated.
Table 6 shows the result of the comparison. With
an accuracy of 44.00% the fasttext Model of Grave
et al. (2018) achieves the best results, followed by the
Levenshtein distance with 39.33%. The fasttext model
recognizes the semantic relation between token like
. But especially abbreviations and
inflections are mostly not recognized. Levenshtein pro-
vides good results when large character strings match.
If only small parts of the string match, the Levenshtein
Bojanowski et al. (2017):
4Grave et al. (2018):
Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements
Table 6: Comparing accuracy for text distance algorithms.
algorithm accuracy
Fasttext (Grave et al. (2018)) 0.4400
Levenshtein 0.3933
editex 0.3800
Hamming 0.2866
Fasttext (Bojanowski et al. (2017)) 0.2133
Burrows–Wheeler transform Run-
length encoding (BWT RLE)
word2vec 0.1333
distance is also large. This leads to a wrong match-
ing of extracted tokens, which contain a very short
keyword like C++ and other meaningless words like
In order to take advantage of both, the fasttext
model and the Levenshtein distance, both distances
were combined and supplemented by preprocessing
and stemming (see Algorithm 1). In preprocessing,
the set of 511 skills
is extended by the respective
stemmed skills. If the token
is contained in the set
it is assumed that this is the best possible result. If not,
a set
W={(w0,p0), ..., (wn,pn)}
of the 500 seman-
tically most similar words
and the corresponding
is calculated by the model
. It is as-
sumed that stemmed(
) also has the similarity
). From the set of all
the intersection
is formed (
). The similarity of each element
of the intersection is weighted with the normalized
Levenshtein distance and if
is a substring of
is a substring of
is weighted with
wc =1.3
The algorithm returns the word
with the greatest
similarity pi. This results in an accuracy of 75.33%.
3.5 Required Competencies
The algorithm was tested on 491 job ads. Figure
4 shows that professional experience, programming
experience and a university degree are the most de-
manded competencies. The most important soft skills
are the ability to work in a team, communication skills
and a sense of responsibility. The most demanded pro-
gramming languages are Java, C and Python. Linux,
SAP and VMware lead the list with experience in han-
dling concrete products.
Knowing the demands on employees is important
for universities and students. Due to the lack of
instruments to identify requirements for employees on
the German labor market, a procedure was developed
to automatically extract them from job advertisements.
The procedure is subdivided into Data Acquisition,
Language Detection, Section Classification and a
multi-stage Skill Detection procedure. For data acqui-
sition the online job portals,
and were used as data sources. The
language detection method according to Shuyo, based
on n-grams, was used for speech recognition and
reached an accuracy of 100% for the tested 300 job ads.
Algorithm 1:Semantic Matching.
Input: token to classify t, list of skills S, model mwith m(x,n)returning
sequence of n tuples of the most similar words to xand the similarity.
Output: semantically most similar word wito given token t
wc 1.3
SS∪ {stemmed(s)|sS}
if tSthen
return t
Wm(t,500)// Result: {(w0,p0), ...(wn,pn)}
W0W∪ {(stem(wi),pi)|(wi,pi)W}
W← {(wi,pi)|(wi,pi)W0wiS}
for all (w,p)Wdo
if wsubstring of tor tsubstring of wthen
end if
if p>resultpthen
end if
end for
return resultw
end if
To classify the sections, different methods of ma-
chine learning were compared. The best results with
an accuracy of 0.9502 were achieved by the linear im-
plementation of a Support Vector Machine using the
one-vs-rest approach for multi-class problems. Taking
into account the premise that per job advertisement
only one HTML list element defines the employee re-
quirements, an accuracy of 100% could be achieved.
The extraction of skills is based on natural lan-
guage processing techniques. Using part of speech
templates, tokens are extracted that correspond to the
pattern of skills. False positives are filtered out via a
blacklist. A synonym dictionary was created to cor-
rectly assign semantically identical or similar skills to
a common class. This dictionary contains descriptions
for skills in relation to semantically similar skills. Us-
ing a text distance algorithm based on fasttext word
embedding and the Levenshtein distance the tokens
are assigned to the skill classes. With this method,
75.33% of the known skills can be assigned correctly.
In the future, the semantic similarity recognition
process could be improved. The inclusion of exter-
nal data sources and training of word embeddings on
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
data bases
IT security
process thinking
sense of responsibility
IT specialists
Education and training
project management
computer science degree
communication skills
team ability
university degree
professional experience
Figure 4: Most wanted competences based on 491 job ads (March 28, 2019).
job ads could improve the assignment. The procedure
could also be extended to other occupational groups
and, for example, provide up-to-date statistics online
in real-time. In combination with other data in the
job ads, maps showing the required skills in certain
regions could be generated. In addition, the evaluation
of the transferability of the approach to the English
language and a comparison of the results would be
Bitkom (2016). 51.000 offene stellen
ur it-spezialisten. Retrieved from
Bojanowski, P., Grave, E., Joulin, A., and Mikolov, T. (2017).
Enriching word vectors with subword information.
Transactions of the Association for Computational Lin-
guistics, 5:135–146.
Gallagher, K. P., Kaiser, K. M., Simon, J. C., Beath, C. M.,
and Goles, T. (2010). The requisite variety of skills for
it professionals. Commun. ACM, 53(6):144–148.
Grave, E., Bojanowski, P., Gupta, P., Joulin, A., and Mikolov,
T. (2018). Learning word vectors for 157 languages. In
Proceedings of the International Conference on Lan-
guage Resources and Evaluation (LREC 2018).
IMS (2018). Tiger corpus. Retrieved from ims.uni-
Kwon Lee, C. and Han, H. (2008). Analysis of skills require-
ment for entry-level programmer/analysts in fortune
500 corporations. Journal of Information Systems Edu-
cation, 19.
Lee, S., Koh, S., Yen, D., and Tang, H.-L. (2002). Percep-
tion gaps between is academics and is practitioners:
an exploratory study. Information & Management,
Litecky, C., Aken, A., Ahmad, A., and Nelson, H. J. (2010).
Mining for computing jobs. IEEE Software, 27(1):78–
Milton, T. (2000). Cross training the answer to e-commerce
staff shortages. Computer Weekly.
Prabhakar, B., Litecky, C. R., and Arnett, K. (2005). It skills
in a tough job market. Communications of the ACM,
Scott, E., Alger, R., Pequeno, S., and Sessions, N. (2002).
The skills gap as observed between is graduates and
the systems development industry–a south african ex-
perience. Informing Science.
Shuyo, N. (2010). Language detection library for java. Re-
trieved from
Sibarani, E. M., Scerri, S., Morales, C., Auer, S., and Col-
larana, D. (2017). Ontology-guided job market demand
analysis. In Hoekstra, R., Faron-Zucker, C., Pellegrini,
T., and de Boer, V., editors, Proceedings of the 13th In-
ternational Conference on Semantic Systems - Seman-
tics2017, pages 25–32, New York, New York, USA.
ACM Press.
Weitzel, T., Eckhardt, A., Laumer, S., Maier, C., and Stet-
ten, A. v. (2015). Recruiting trends 2015: Eine em-
pirische untersuchung mit den top-1.000-unternehmen
aus deutschland, sowie den top-300-unternehmen aus
den branchen finanzdienstleistung, health care und it.
Retrieved from
Wowczko, I. (2015). Skills and vacancy analysis with data
mining techniques. Informatics, 2(4):31–49.
Yongbeom, K., Jeffrey, H., and Mel, S. (2006). An update
on the is/it skills gap. Journal of Information Systems
Education, 17(4):395–402.
Automated Analysis of Job Requirements for Computer Scientists in Online Job Advertisements
... In paper [19], the authors created a system for automated identification of skills for German-language job advertisements. They used and compared several machine learning approaches for identification of the job position requirements. ...
Full-text available
There are currently many job portals offering job positions in the form of job advertisements. In this article, we are proposing an approach to mine data from job advertisements on job portals. Mainly, it would concern job requirements mining from individual job advertisements. Our proposed system consists of a data mining module, a machine learning module, and a postprocessing module. The machine learning module is based on the SDCA logistic regression. The postprocessing module includes several approaches to increase the success rate of the job requirements identification. The proposed system was verified on 20 most searched IT job positions from the selected job portal. In total, 9971 job advertisements were analyzed. Our system’s verification is finding all job requirements in 80% of analyzed advertisements. The detected job requirements were also compared with the Open Skills database. Based on this database and the extension of IT job positions with other typical job skills, we created a list of the most frequent job skills in selected IT job positions. The main contribution is the development of a universal system to detect job requirements in job advertisements. The proposed approach can be used not only for IT positions, but also for various job positions. The presented data mining module can also be used for various job portals.
... 2) The method based on statistics mainly used the statistical attributes of the distribution of skill words to identify keywords. One type was to sort keywords by feature quantitative indicators; for example, Grüger & Schneider [20] used TF-IDF (term frequency-inverse document frequency) values to filter skill keywords. Statistical-based methods do not require syntactic and semantic information and do not rely on labeling data, but the accuracy rate is relatively low. ...
Full-text available
The occupational profiling system driven by the traditional survey method has some shortcomings such as lag in updating, time consumption and laborious revision. It is necessary to refine and improve the traditional occupational portrait system through dynamic occupational information. Under the circumstances of big data, this paper showed the feasibility of vocational portraits driven by job advertisements with data analysis and processing engineering technicians (DAPET) as an example. First, according to the description of occupation in the Chinese Occupation Classification Grand Dictionary, a text similarity algorithm was used to preliminarily choose recruitment data with high similarity. Second, Convolutional Neural Networks for Sentence Classification (TextCNN) was used to further classify the preliminary corpus to obtain a precise occupational dataset. Third, the specialty and skill were taken as named entities that were automatically extracted by the named entity recognition technology. Finally, putting the extracted entities into the occupational dataset, the occupation characteristics of multiple dimensions were depicted to form a profile of the vocation.
... Several text-mining and natural language processing techniques have been applied to resumes and job postings. Elimination of the stop words, bag of words model, part of speech tagging (POS), removal of the noisy data, normalization (stemming and lemmatization, n-gram model, TF-IDF technique, and named entity recognition are some of the used text-mining and natural language processing tasks in literature (Singh et al., 2010;Ayishathahira et al., 2018;Duan et al., 2019;Grüger and Schneider, 2019). Chuang et al. (2008) introduced the study on the processing of semi-structured Chinese resumes. ...
Conference Paper
Full-text available
Designing secure, resilient, and scalable IoT networks is challenging because of heterogeneity, resource-constrained devices, and no guarantees of reliable network connectivity, which requires the implementation of new secure communication protocols with assurance compliance with security levels requirements of applications. In this context, one major question is how to map the multiple adaptive chaos-based security protocols to Class of Fog Applications (CoFA) on Fog integrated Cloud architecture. The approach to identifying the levels of chaos-based security protocols and discriminate the applications arriving at the Fog into CoFA for secure communication by keeping the important constraints like how much public channel is secure, how much time it takes to communicate from one node to another, and the rate at which synchronization error reaches close to zero. This paper thus introduces a set of CoFA for Fog applications and the associated level of chaos-based security protocols for secure communication. We have introduced dynamical adaptive chaos masking methodology for secure transmission wherein, we have divided into four sets of adaptive, secured communication protocol which starts with level one: security consists of non-identical chaotic (one master and one slave), level two: consist of non-identical hyper-chaotic (one master, one slave), level three: with non-identical complex hyper-chaotic system and, level four with hyper-chaotic complex and real hyper-chaotic (two masters, two slaves). We compared existing literature and state of the art; we found that our technique best results in synchronization error recovery and time duration. To the best of our knowledge, such mapping of CoFA with level-based secure communication protocols for Fog integrated Cloud architecture is introduced the first time. In the analysis phase, hyper-chaotic is more complex than a chaotic, similarly hyper-chaotic complex, with more complexity than hyper-chaotic and chaotic, ensuring more robust and secure transmission on a public channel.In this study, a chaotic system with a hyperchaotic system is used to increase the security of the chaotic communication system, which has not been previously implemented for this purpose in the literature for Fog integrated Cloud architecture. The used chaotic system includes chuas system, modified jerk system and genesio-tesi system. Also, to map Fog request to the respective levels of security, a CoFA was introduced. These properties increase the complexity of the chaotic system and the security level. The sliding adaptive mode control method is preferred for the synchronization part of the secure communication.
... In some studies, it has been observed that it is made dependent on the field. For example, in the study in 2019, "skill recognition" was carried out on the German language and only in Computer Science with machine learning algorithms such as supervised learning algorithms like Random Trees [3]. ...
Conference Paper
Full-text available
This paper discusses the results of an investigation into the skills gap between Information Systems (IS) graduates at the University of Cape Town (UCT) and the South African Systems Development Industry. Three objectives were defined for this study. Firstly to measure the alignment between the level of skills possessed by students and the level of skills demanded by development companies. Secondly to identify and compare the most prominent specific skills that industry requires with the skills of students and thirdly to determine whether the students obtained the skills directly through UCT. The study revealed that there was alignment between the importance rating of companies and the skills of students in some areas, but not in others. Although correlation exists between the specific skills and technologies that industry requires and those which students possess, knowledge of certain technologies is lacking from the formal IS curriculum.
Full-text available
Through recognizing the importance of a qualified workforce, skills research has become one of the focal points in economics, sociology, and education. Great effort is dedicated to analyzing labor demand and supply, and actions are taken at many levels to match one with the other. In this work we concentrate on skills needs, a dynamic variable dependent on many aspects such as geography, time, or the type of industry. Historically, skills in demand were easy to evaluate since transitions in that area were fairly slow, gradual, and easy to adjust to. In contrast, current changes are occurring rapidly and might take an unexpected turn. Therefore, we introduce a relatively simple yet effective method of monitoring skills needs straight from the source—as expressed by potential employers in their job advertisements. We employ open source tools such as RapidMiner and R as well as easily accessible online vacancy data. We demonstrate selected techniques, namely classification with k-NN and information extraction from a textual dataset, to determine effective ways of discovering knowledge from a given collection of vacancies.
Full-text available
This paper presents the most up-to-date skill requirements for programmer/analyst, one of the most demanded entry-level job titles in the Information Systems (IS) field. In the past, several researchers studied job skills for IS professionals, but few have focused especially on "programmer/analyst." The authors conducted an extensive empirical study on programmer/analyst skill requirements by collecting and analyzing 837 job ads posted on Fortune 500 corporate websites for three years. The results indicate that the programmer/analysts in the Fortune 500 are expected to fulfill a combination of roles from computer program writers to technical experts as well as businessmen. The results of this study are discussed in terms of their implications for the IS 2002 curriculum model.
Full-text available
A Web content mining approach identified 20 job categories and the associated skills needs prevalent in the computing professions. Using a Web content data mining application, we extracted almost a quarter million unique IT job descriptions from various job search engines and distilled each to its required skill sets. We statistically examined these, revealing 20 clusters of similar skill sets that map to specific job definitions. The results allow software engineering professionals to tune their skills portfolio to match those in demand from real computing jobs across the US to attain more lucrative salaries and more mobility in a chaotic environment.
Full-text available
Introduction IT professionals are beset by ongoing changes in technology and business practices. Some commentators have suggested that, in order to stay competitive, IT professionals should retool themselves to gain competency in specific in-demand technical skills. This article argues that thriving in such a dynamic environment requires competency in a broad range of skills, including not only technical skills, but non-technical skills as well. Our research shows that IT departments in non-IT companies report that while both technical and non-technical skills are important, the skills most critical to retain in-house and most sought in new mid-level employees are non-technical skills such as project management, business domain knowledge and relationship skills. These skills are critical because they enable IT departments to work effectively with other departments, internal users, and external customers and suppliers. Non-technical skills leverage technical skills to augment the organization's overall effectiveness in designing and delivering solutions to meet an organization's challenges and opportunities. These findings depart from previous articles emphasizing technical skills as a basis for valuing IT workers and other research recommending business-oriented skills only for those managing IT workers, not for IT professionals themselves. Our findings lead us to the realization that in today's environment of continuous and fast-paced change, a mix of skills is essential for IT professionals. We believe that the Law of Requisite Variety can help explain the need for greater breadth of knowledge and skills among IT professionals. From cybernetics, the Law of Requisite Variety states that adapting to change requires a varied enough solution set to match the complexity of an environment. In this case, IT workers need a broad enough range of knowledge and skills to meet the demands of their increasingly dynamic and complex profession. Based on our research, we offer a framework outlining six skill categories. We believe that all six skill categories are critically important for the career development of IT professionals.
Continuous word representations, trained on large unlabeled corpora are useful for many natural language processing tasks. Many popular models to learn such representations ignore the morphology of words, by assigning a distinct vector to each word. This is a limitation, especially for morphologically rich languages with large vocabularies and many rare words. In this paper, we propose a new approach based on the skip-gram model, where each word is represented as a bag of character n-grams. A vector representation is associated to each character n-gram, words being represented as the sum of these representations. Our method is fast, allowing to train models on large corpus quickly. We evaluate the obtained word representations on five different languages, on word similarity and analogy tasks.
In den „Recruiting Trends 2014“ untersucht das Centre of Human Resources Information Systems (CHRIS) der Universitäten Bamberg und Frankfurt am Main mit Unterstützung und im Auftrag von Monster Worldwide Deutschland bereits im zwölften Jahr in Folge die Personalbeschaffung in den Top-1.000-Unternehmen aus Deutschland. Dabei wurden seit Beginn der jährlichen Studien im Jahr 2003 auf Basis der Antworten von insgesamt rund 1.800 teilnehmenden Unternehmen zahlreiche Entwicklungen und Trends im Kontext der Rekrutierung identifiziert und über die Jahre hinweg begleitet, was neben der Untersuchung topaktueller Themen stets auch hochinteressante Längsschnittauswertungen und robuste Einsichten in „good practises“, Fakten und Folklore im Recruiting ermöglicht. Die komplementäre jährliche Kandidaten-Studie „Bewerbungspraxis“ zeichnet mit den Antworten von bislang über 110.000 Teilnehmern das komplementäre Bild aus Bewerbersicht. Dieses Jahr haben sich 128 der 1.000 größten deutschen Unternehmen an den „Recruiting Trends 2014“ beteiligt (Rücklaufquote 12,8 Prozent). Die Verteilung der Stichprobe der 128 Studienteilnehmer ist dabei gemäß dem aktuellen Datenbankregister von Bisnode (vormals Hoppenstedt) hinsichtlich der Merkmale Umsatz, Mitarbeiterzahl und Branchenzugehörigkeit repräsentativ für die Grundgesamtheit der Top-1.000-Unternehmen aus Deutschland (vgl. Kapitel 2.15). Die „Recruiting Trends 2014“ umfassen 10 Themenschwerpunkte der modernen Personalbeschaffung in den 1.000 größten deutschen Unternehmen, ausgehend von aktuellen Trends und Entwicklungen über den Personalbedarf und die Stellenausschreibung bis hin zur Kandidatenauswahl und Weiterbildung von Recruitern (vgl. Abbildung 1). Des Weiteren beinhaltet die Studie die Resultate ausgewählter Befragungen zur Personalbeschaffung in den jeweils 300 größten deutschen Unternehmen aus den Branchen Health Care, IT und Maschinenbau, sowie drei Fallstudien bei der Otto GmbH & Co KG, der Krones AG und der VOITH GmbH. Die Studie schließt mit einer Beschreibung der Methodik, einer Analyse der Zusammensetzung der Studienteilnehmer und verschiedenen Tests auf die Repräsentativität der erhobenen Stichproben. Neben den Branchenanalysen und den Einblicken in die Praxis werden die Ergebnisse aus der Befragung der Top-1.000-Unternehmen aus Deutschland an einigen Stellen auch durch Ergebnisse aus der aktuellen Bewerberstudie „Bewerbungspraxis 2014“ des CHRIS ergänzt, um Unternehmen und Bewerbern einen umfassenden Überblick über die Personalbeschaffung in Deutschland zu ermöglichen.
It has long been recognized that there are significant perception gaps between information systems (IS) academics and IS practitioners with regards to the required IS knowledge/skills to perform their professional jobs successfully. Unfortunately, there is no consensus about which knowledge/skills are more important in the IS profession. With the rapidly changing IS technology and IS industry, a study was urgently needed to find and understand the gaps between these two groups. We report here on a study to measure the gap.
IT professionals need to be aware of the specific skills that are in demand to survive in the competitive IT job market. Professionals should also pay attention to the skills for which demand may be growing in order to better position themselves in the tough job market. Web programming, Unix, C++, Java, SQL programming, and Oracle database are emerging to be the top six skills in the current job market. Java is now demanded in more than one-fifth of all jobs, which shows that Java has become a mainstream programming language. Among database skills, Oracle is still the most demanded database skill, with over 22% of all positions requiring Oracle, and has maintained a commanding lead over other database skills. Industry certification now is found to be of immense importance to give a candidate an edge in a tight job market.
Learning word vectors for 157 languages
  • E Grave
  • P Bojanowski
  • P Gupta
  • A Joulin
  • T Mikolov
Grave, E., Bojanowski, P., Gupta, P., Joulin, A., and Mikolov, T. (2018). Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2018).