Challenges and Potentials for Keyword Extraction from Company

Websites for the Development of Regional Knowledge Maps

Christian Wartena

and Montserrat Garcia Alsina

Hochschule Hannover, Expo Plaza 12, 30539 Hannover, Germany

Universitat Oberta de Catalunya, Rambla del Poblenou, 256, 08018 Barcelona, Spain

Keywords:

Text Mining, Keyword Extraction, Territorial Intelligence, Regional Innovation Systems.

Abstract:

Regional Innovation Systems describe the relations between actors, structures and infrastructures in a region in

order to stimulate innovation and regional development. For these systems the collection and organization of

information is crucial. In the present paper we investigate the possibilities to extract information from websites

of companies. First we describe regional innovation systems and the information types that are necessary to

create them. Then we discuss the possibilities of text mining and keyword extraction techniques to extract this

information from company websites. Finally, we describe a small scale experiment in which keywords related

to economic sectors and commodities are extracted from the websites of over 200 companies. This experiment

shows what the main challenges are for information extraction from websites for regional innovation systems.

1 INTRODUCTION

A basic prerequisite for regional development, for

stimulating regional innovation and for improved co-

operation between companies in a region, is a high

quality overview of all companies in that region, their

activities and their strengths. However, hardly any

such overviews exist. Information about companies

is often incomplete, outdated or focuses only on one

speciﬁc branch of industry. Ofﬁcial lists, if available

at all, also suffer from the problem that ofﬁcially reg-

istered companies might be inactive or just be admin-

istrative constructs being part of a larger company,

that has to be considered as a unit when activities, cus-

tomers, products, services, etc. have to be described.

Building a regional overviewwe face several chal-

lenges. In the ﬁrst place, we have to identify which

companies in a region are involved in innovation and

regional development. The second challenge is the se-

lection of information sources. In the third place, we

should determine which information from each com-

pany is relevant to identify its strengths, and to pro-

mote innovation and regional development. Finally,

we have to describe the information about a company,

its activities, products and customers in a uniform

way. As with regard to second challenge, it seems nat-

ural to extract information from the company’s web-

site to obtain basic information (formal data, products

and activities) about a company, since virtually every

company has a website. There are other sources of

up to date information, like patent databases or com-

mercial directories. However, in the present paper we

will investigate the possibilities to use the websites

of companies, leaving other sources for future work.

We identify what the main challenges are and we ex-

plore the possibilities to use them as an information

source to describe the knowledge existing in a region.

In order to approach the third and fourth challenge we

combine expertise from competitive and territorial in-

telligence and data mining and more speciﬁcally key-

word extraction (KWE).

While we sketch the overall approach for the con-

struction of comprehensive regional knowledge maps

in (Garcia-Alsina et al., 2013), in this paper we fo-

cus on the possibilities of KWE. In a small empir-

ical study we investigate the possibilities to extract

knowledge from company websites for the construc-

tion of regional knowledge maps of innovation sys-

tems. We use simple thesaurus based KWE tech-

niques, but show how these techniques can provide

very rich information about companies when used in

combination with a highly structured thesaurus. In

contrast to previous work on KWE we do not obtain

a ﬂat list of keywords but we get a keyword based de-

scription of a company for different aspects. In the

current study we have evaluated the extracted key-

words for two aspects: the economic sector a com-

pany belongs to and products produced by the com-

241

Wartena C. and Garcia Alsina M..

Challenges and Potentials for Keyword Extraction from Company Websites for the Development of Regional Knowledge Maps.

DOI: 10.5220/0004660002410248

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval and the International Conference on Knowledge

Management and Information Sharing (SSTM-2013), pages 241-248

ISBN: 978-989-8565-75-4

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

pany. In the experiment we also see the limits of the

approach. In the ﬁrst place, the extraction of key-

words from complete websites consisting of pages on

completely different topics, differs substantially from

classical KWE of uniform texts. Furthermore, we do

not try to ﬁnd keywords for the texts, but keywords

that apply to the company behind these texts.

The remainder of the paper is organized as fol-

lows. In section 2 we discuss the theoretical back-

ground and related work in territorial intelligence and

text mining. In section 3 we describe an experiment to

access the feasibility of KWE as a tool to support the

construction of regional knowledge maps. The results

of the experiment are given and discussed in section 4.

We ﬁnish the paper with a conclusion and an outlook

to future work.

2 RELATED WORK

2.1 Territorial Intelligence

Studies about the regional economics are done with

methods coming from different disciplines: Econ-

omy, Geography, Sociology, Science of Education,

Information and Communication Science (Mollo,

2009). More speciﬁcally, in the last 15 years a line

of research has been developed, that identiﬁes inno-

vation processes as a key to the regional development

(Girardot and Brunau, 2010; OECD and EUROSTAT,

2005).

Territorial Intelligence (TI) is a collective posture

that explores the territorial possibilities by the collec-

tion of information and its treatment in order to an-

ticipate risks and threats (Herbaux, 2008). To (Gi-

rardot, 2008) the territorial intelligence is the science

whose object is the sustainable development of ter-

ritories and whose subject is the territorial commu-

nity. TI focuses not only on the economic efﬁcacy

and efﬁciency of the development model, but also in-

cludes all the dimensions that affect sustainable devel-

opment (social, political, cultural, and environmen-

tal). It involves information about resources available

in the territory, products, services (individual or col-

lective, private or public) people needs, which activi-

ties are taking part in a territory and the territorial dy-

namics (Girardot and Brunau, 2010; Girardot, 2008;

Herbaux, 2008). Hence, the collection of information

about the territory plays an important role to planning

strategic actions to put in touch different actors.

According to (Jimenez et al., 2011; Asheim and

Gertler, 2005; Doloreux and Parto, 2005; Lundvall

and Johnson, 1994) a Regional Innovation System

(RIS) or National Innovation System (NIS) consists

of the relation between actors, structures and infras-

tructures involved in a region, as well as the knowl-

edge ﬂow between them, that serve the goal of bring-

ing innovative performance of companies in the re-

gion. Besides, an efﬁcient system of distribution

and access to knowledge contributes to increase the

amount of innovative opportunities (Lundvall, 1998;

David and Foray, 1995). More speciﬁcally, RIS and

NIS as research area have developed a framework to

study the factors that enable innovation and regional

economic development. Some of the features of this

framework are: a) institutions (rules, norms, or orga-

nizations) (Lundvall and Christensen, 2003; Nelson,

1993), b) innovation processes, c) knowledge’s ﬂows

which take place within the regional learning process

(Lundvall,1992); d) the social capital and the regional

context in which innovation happens, and that pro-

vides a set of rules, conventions and norms that pre-

scribe behavioral roles and shape expectations (Do-

loreux and Parto, 2005; Nahapiet and Ghoshal, 1998);

e) inﬂuence of national or local or social idiosyn-

crasies, which inﬂuence the social process (Salavisa

and Vali, 2012; Lundvall, 1998), and f) the inter-

sectorial differences to explain innovation activities

(Doloreux et al., 2008).

Considering these antecedents, we underline the

relevance of the role of: a) knowledge in the inno-

vation process and consequently, knowledge manage-

ment in ﬁrms and regions, where knowledge maps

play an important role (Barinani et al., 2011; Driessen

et al., 2007), b) the role of companies as actors in

a regional innovation system, and c) the informa-

tion about the environment that the companies should

achieve and d) controlled vocabularies, like thesauri,

with a common understanding of central terms to

classify and retrieveknowledge (F¨arber and Rettinger,

2013; Garcia-Alsina and Ortoll, 2012; Canongia,

2007; Escorsa et al., 2000). Taking into account the

amount of documents, it is difﬁcult extract relevant

information, and classify this information, so seman-

tic technologies and data mining are needed (F¨arber

and Rettinger, 2013; Eckert et al., 2007).

2.2 Keyword Extraction

Our current goal is to classify companies and other or-

ganizations. A number of information sources can be

used to do this. In the present study we focus however

exclusively on the web sites as a source of informa-

tion. Thus we can consider our task as a text classi-

ﬁcation problem. Text classiﬁcation is a well studied

ﬁeld. A good overview of techniques and approaches

is given by (Sebastiani, 2002).If there is a large num-

ber of categories, which is usual the case if each term

KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

242

of a thesaurus is considered as a potential document

class, the standard classiﬁcation approach cannot be

applied, as not enough training data are available. In

these cases a KWE approach is usually better suited.

In its basic form, KWE is nothing else than select-

ing the most salient terms of a document. Determin-

ing such terms has been studied since the mid of the

previous century (Salton and Buckley, 1987). How-

ever, salience turns out not to be the only criterion

for keywords. More features can be found, indicat-

ing whether a term is suited as a keyword or not.

These features, together with the relevance weight of

the term, can be used in a supervised machine learn-

ing setting to learn how to distinguish keywords from

non-keywords. This approach to KWE was proposed

by (Frank et al., 1999) and (Turney, 2000).

Alternatively, the set of possible keywords can be

restricted by the terms of a thesaurus or some other

restricted vocabulary. This approach is followed by

(De Campos et al., 2007) who use information from

the thesaurus in combination with Bayesian statistics

to suggest keywords. (Wang et al., 2007) uses PageR-

ank to determine the most central words in the graphs

which is constructed with the WordNet relations be-

tween the potential keywords. The community struc-

ture formed by the potential keywords and their rela-

tions is used by (Grineva et al., 2009). In (Malais´e

et al., 2007a) and (Gazendam et al., 2010) words with

a large number of relations get higher weights in order

to promote central concepts.

3 SYSTEM DESCRIPTION

In order to investigate the potentials of KWE we crawl

the pages from the websites of over 200 companies

and store them in a local repository. We annotate the

data manually for several aspect in order to be able to

evaluate KWE algorithms. In the following we will

describe the thesaurus, the data sources and the KWE

in more detail.

3.1 Thesaurus

We use the STW Thesaurus for Economics

as a

source for potential keywords. The STW is organized

in subthesauri. In order to describe the company we

extract only words that are used as descriptors in the

subthesauri Commodities and Economic Sectors. For

classiﬁcation of the economic sectors it might seem

more natural to use the NACE (Nomenclature statis-

tique des Activit´es ´economiques dans la Communaut´e

http://zbw.eu/stw/versions/latest/about.en.html

Europ´eenne.) classiﬁcation, that is an ofﬁcial stan-

dard in the European Union for the classiﬁcation of

economic activities. However, the STW Economic

Thesaurus has the advantages that

1. it is available in SKOS format (Isaac and Sum-

mers, 2009), and thus can be read easily by text

analysis software;

2. it has a lot of different descriptors for each cate-

gory, that might be found in texts;

3. it contains a classiﬁcation of products and related

areas as well.

Especially the second property is essential for our ap-

proach: since the descriptors are usually short terms,

that are likely to be found in texts, we can use these

descriptors as a list of potential keywords. In order

to use the approach for the construction of a regional

knowledge map, we either have to deﬁne a mapping

from STW sectors to NACE categories, or we have to

transform the NACE thesaurus into SKOS and popu-

late the categories with appropriate descriptors.

The main disadvantage of the STW is, that it is

designed to index documents about economics, not

to classify companies and their products. We are not

aware of any usage of the STW for this purpose. The

difference between the two goals might mainly be one

of focus and coverage: the sectors that is most written

about, and thus are expected to have a good coverage

in the STW, need not to be the same sectors to which

many companies belong.

As already said above, the STW Thesaurus for

Economics consists of 7 subthesauri. Each thesaurus

has several subject categories, that can contain sub-

categories. Subject categories represent either a tech-

nical intermediate level of description (e.g. Branches

of Industry) or classes of products, economic sectors,

etc. Each category has a denotation and a number of

descriptors. The descriptors are subclasses or exam-

ples of the category they belong to. E.g. fashion is a

descriptor of the category Textile and Clothing Indus-

try (W.06.01.11). Also the technical categories have

descriptors, that usually refer to instances that cannot

be classiﬁed into one of the subclasses.

3.2 Data Sources

In order to investigate the possibilities of describing

companies by extracting keywords from their web-

sites we have compiled a set of 229 companies from

10 economic sectors. These companies are not situ-

ated in one region, as we do not yet aim at the con-

struction of a regional knowledge map in this phase

of the development. All companies are German and

ChallengesandPotentialsforKeywordExtractionfromCompanyWebsitesfortheDevelopmentofRegionalKnowledge

Maps

243

have German websites. We have classiﬁed the com-

panies and their products according to the main cate-

gories of the STW. The companies were selected by

students, that had the task of selecting 20 websites of

companies in a speciﬁc branch or sector as a source

for keywords in the related domain.

We have used the crawler4j

to crawl the websites.

For 21 companies the crawling was not successful and

no pages could be retrieved. Thus 208 companies re-

main. Since a few companies have a site with a high

number of pages, we limited the number of pages to

be retrieved to 120. The limitation services the prac-

tical goal of keeping the size of the corpus moder-

ate, but also has more fundamental reasons: A few

companies have very large websites, which makes

the amount of information per company very unbal-

anced. Moreover, we expect that even a large com-

pany should be described rather well on the ﬁrst two

levels of a web site. If we crawl in a breadth ﬁrst

way, as we do, we might expect that at some point we

have seen the core information of a company. If more

pages follow, we might get more and more speciﬁc in-

formation on detailed topics, that even could obscure

the more important and central information. The limit

of 120 is rather arbitrary and turned out to be a size

that allows us for almost all companies in our list to

crawl the complete site. In total 14673 pages were

retrieved, which averages to 70.5 pages per company.

We did not do any boiler plate removal since it

turned out that in many cases essential information

is removed. E.g. a list of products or departments is

often given as a menu, that might be removed. For

companies with very limited websites, or in cases

where the products or departments mentioned in the

list point to other domains (we crawl only pages from

one domain for each company) the most important in-

formation then would be removed.

3.3 Establishment of a Ground Truth

In order to create a base for evaluation of the auto-

matic KWE, we classiﬁed all companies manually

with the subject categories of the subthesauri Com-

modities and Economic Sectors. These categories are

quite broad and can be assigned with a low rate of

error and subjectivity and do not require very deep

analysis of the information available about a com-

pany. In many cases several economic sectors and

products had to be assigned. In the ﬁrst place this

was necessary for a number of large companies or

business groups that are active in several (related sec-

tors), like machinery construction and electrical in-

dustry. The second reason for multiple assignments

http://code.google.com/p/crawler4j/

are businesses that can be viewed from different per-

spectives. E.g. there is a number of goat farms in

our data set. These farms can be classiﬁed as Animal

Husbandry (W.01.02), but since they usually produce

goat cheese, they also can be classiﬁed as Food and

Tobacco Industry (W.06.01.12), but as they usually

have a shop where they sell the cheese, Retail Trade

(W.10.05) is also not completely wrong. Moreover,

such farms often offer possibilities to view the ani-

mals to the public and they have a small restaurant or

even offer cottages for rent. Thus a number of further

subject categories apply.

3.4 Keyword Extraction

As a ﬁrst step for KWE all texts are analyzed using

a GATE pipeline (Cunningham et al., 2002) that con-

sists of a language guesser, a tokenizer, a sentence

splitter, a part of speech tagger and lemmatizer an

ontology lookup component and several JAPE gram-

mars. The language guesser is used to ensure that

only German texts are analyzed. The ontology lookup

component Apolda (Wartena et al., 2007) ﬁnds all la-

bels of thesaurus terms in the texts, including multi-

word terms and inﬂected forms. The JAPE grammar

formalism is part of the GATE software, and enables

the deﬁnition of patterns over words and previous an-

notations. The Jape Grammars are used to extract ad-

dresses, phone numbers and names of companies.

Since advanced algorithms for KWE usually have

been reported to give only limited beneﬁt over a sim-

ple tf.idf weighting scheme, we decided to extract the

words with the highest tf.idf score. In a later phase

other algorithms can be used, but using a simple al-

gorithm makes identiﬁcation of problems easier. As

well for the computation of the document frequen-

cies as for the computation of the term frequencies

we consider the whole of all web pages of a company

as one document. The intuition behind this is, that a

term like Gesch

aftsf

uhrer (Manager) that is found on

348 pages of 111 companies in our data set is much

less indicative for the business of a company than the

word Schiff (ship) that occurs on approximately the

same number of pages (330) but only for 29 compa-

nies (a number of shipyards is in our data set). The

tf.idf value is computed as usual:

tf.idf(w, d) = n

(w) · log



df(w)



(1)

where n

(w) is the number of occurrences of w in

document d, N is the total number of documents and

df(w) is the number of documents in which w occurs.

KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

244

Table 1: Precision of extracted keywords for economic sec-

tors and commodities.

Sectors Commodities

prec@1 0.076 0.27

prec@2 0.088 0.27

prec@5 0.063 0.23

4 RESULTS AND DISCUSSION

All extracted keywords are descriptors in the subthe-

sauri Commodities or Economic Sectors in the STW.

As mentioned above, we classiﬁed the companies ac-

cording to the main categories in these subthesauri.

We consider a keyword correct if it is a descriptor

of a category assigned to the company. Thus we can

compute the precision, i.e. the fraction of correct key-

words, for each top n elements of the ranked keyword

list. The precision of the top n keywords is referred

to as prec@n. Since we have no exhaustive list of

keywords that should be assigned to a company, we

cannot compute a recall value. The precision for eco-

nomic sectors and commodities is given in Table 1.

The results show immediately that the extraction

of product categories is much easier than the extrac-

tion of economic sectors. This is not very surprising

since companies will tell about their products on their

website, not about the sector they operate in. Even

worse, they might write about the use of their prod-

ucts in the sectors their customers are from.

Though we did not use any advanced technique

for KWE, but rather establish a baseline, the results

are in the same order of magnitude as reported in

literature. E.g. (Medelyan and Witten, 2005) ﬁnd

a prec@5 of 0.21 for assignment of terms from the

Agrovoc thesaurus to documents, (Gazendam et al.,

2009) ﬁnd a prec@5 of 0.23 for assignment of terms

from the Dutch GTAA thesaurus for audio-visual

archives. In this context it should be noted, that the

comparison is useful to get a feeling for the order of

magnitude we have to think about. In classical work

on KWE, keywords are used as general descriptors of

a text. Here, we use keywords to describe a speciﬁc

aspect of a text. Furthermore, we look for keywords,

that are not necessarily good descriptors for the texts

they are extracted from, but we look for keywords de-

scribing the company behind the website.

When looking at the results for the economic sec-

tors, the approach seems not to be very successful: the

majority of the keywords found where judged to be ir-

relevant in our evaluation. In the following we will,

take a closer look, at the irrelevant keywords that were

found and propose ways to improve the results.

4.1 Error Analysis

For most cases, in which no keywords from the right

class have been assigned the underlying property is

data sparseness. Sparseness might have led either

to unprecise manual classiﬁcation, or to lack of the-

saurus terms that could be used by the analysis of the

internet sites. A typical case of the former is a set

of over 10 companies that are specialized in library

(software) systems. The closest branch in the STW

is Information Services (W.19.05). The KWE ﬁnds

mainly words related to libraries. Since a software

company writing software for libraries is not a library,

these keywords are evaluated as incorrect. A related

problem arises from our decision to use the subject

categories to classify companies and to evaluate the

descriptors found as keywords. In most cases this

works quite well. E.g. the category Fishery (W.03)

has descriptors like Aquaculture, Fishery, Fishermen,

Fishery ﬂeet, etc. These terms both are likely to occur

in texts and are good descriptors for companies in this

sector. However, a category like Branches of Indus-

try (W.06.01) has subcategories, representing speciﬁc

branches, that have descriptors describing them, but

it has also descriptors itself. These descriptors rep-

resent branches, like Toy industry that do not have

other descriptors themselves. Consequently we have

to annotate a toy manufacturer as Branches of Indus-

try (W.06.01), which is not very adequate.

If we analyze the wrong keywords, we ﬁnd ﬁve

brought classes of errors:

Minor Aspects. In a number of cases, the keyword

found is in fact correct, but its category was not

used to classify the company, usually because the

category does not reﬂect the main business of that

company. This is especially the case for the prod-

ucts category, where we usually assigned only the

main category, while many other product cate-

gories might apply.

Random Classes. In some cases the extracted key-

words seem completely random. This is the result

of the idf-weighting scheme. If some seldom term

is mentioned two or three times on a small web-

site it will get a very high score. This effect is

reinforced by the fact that the correct terms usu-

ally do not have a very high tf.idf value, since

we started collecting several companies from the

same sector. This problem is observed frequently

in thesaurus based KWE. A method to exclude

completely irrelevant terms, is to analyze the (the-

saurus) relations between the keywords initially

found: if a keyword has no relation (or only few

relations) to other keywords, it is likely that it is

not related to the main topic of the document. Var-

ChallengesandPotentialsforKeywordExtractionfromCompanyWebsitesfortheDevelopmentofRegionalKnowledge

Maps

245

ious proposals have been made to operationalize

this idea (Malais´e et al., 2007a), (Grineva et al.,

2009), (Gazendam et al., 2010).

Customer Classes. Web pages are usually designed

for the customers of a company. This might result

in texts that contain much more terms describing

the branch of the customers of a company than

the company itself. E.g. for most pharmaceutical

companies descriptors for the branch Health care

systems (W.25), which is reserved for hospitals

and physician practices, are found. Another typi-

cal example are the shipyards, classiﬁed as Vehicle

Construction (W.06.01.03), but for which many

terms are found from the categories Port Manage-

ment (W.12.01.03.03) and Shipping (W.12.01.03).

Ambiguous Terms. Some terms are in some way

ambiguous and might refer to two different

branches of industry, but are only used for one of

both in the STW. E.g. the word radio is used a de-

scriptor for Broadcasting Industry (W.19.03), but

is in our data set found for a company producing

radio sets. In a number of cases, the structure of

the thesaurus is also problematic. E.g. the term

sail, sometimes found on the pages of shipyards,

is an alternative term for off-the-peg textiles that

belongs to the category Clothing (P.19).

Broader Terms. A typical error is the selection of

a term that is too broad, but in fact not incor-

rect. This is a common problem for thesaurus

based KWE and alternative evaluation methods

have been proposed to deal with this (Medelyan

and Witten, 2005) and (Gazendam et al., 2009).

Finally, keywords that do not match our self de-

ﬁned gold standard do not have to be wrong. Ex-

actly these keywords might be very interesting and

informative. The keywords that do not match with

the obvious economic sector, might give hints to hid-

den expertise. E.g., a company from the sector Vehi-

cle construction (W.06.01.03), which main product is

ships construction (Watercrafts; P.09.03), but with an

extracted keyword Textiles (P.18), might indeed have

expertise on textiles that could also be useful for other

companies. Nevertheless, ﬁrst the main classiﬁcation

for each company has to be established.

4.2 Thesaurus Enhancement

The majority of the problems arises from data spar-

sity in the thesaurus. It makes of course no sense to

add ad-hoc categories for products or sectors that are

completely missing, like medical equipment industry,

since we use a thesaurus because it is standardized

and stable. However, we can add more descriptors,

non-preferred labels, spelling variants and synonyms

to the thesaurus. This would let the original thesaurus

intact. The extended version of the thesaurus needs

only to be used by the KWE algorithm, while the re-

sults are still descriptors or categories from the ofﬁ-

cial thesaurus. When more labels are available, espe-

cially synonyms and alternative labels, the ﬁnal result

is based on much more data, and the occurrence of an

irrelevant word will have less impact. Now, in many

cases only two or three thesaurus terms are found on

a page, and it is more or less by chance whether these

are relevant or irrelevant terms. Enrichment of a the-

saurus to enhance thesaurus based KWE is a common

method used e.g. by (Tiun et al., 2001), (Malais´e

et al., 2007b).

There are various ways to extend the thesaurus in

the proposed way. In the ﬁrst place, the STW has

many links to equivalent terms in other thesauri, like

the German Subject Headings Authority File (SWD),

that has many alternative labels for each term. Using

the link between the SWD and the STW these alter-

native labels could be added to the STW as well. Fur-

thermore, general dictionaries could be used to ﬁnd

synonyms and ﬁnally there exist very good methods

to ﬁnd synonymsstatistically in large corpora(see e.g.

(Weeds et al., 2004) for an overview of methods of

distributional similarity).

4.3 Future Work

For future work, we have to extract more types of in-

formation. Obviously, there are other sources of infor-

mation and a challenge for the future will be to inte-

grate the knowledge about the regional economy that

comes from different sources, uses different vocab-

ularies, is written for different target audiences and

might even be contradictory.

The other line of research we have to follow is the

improvement of the information extraction methods.

Here we can start with more advanced methods, but

we also have to take into account the peculiarities of

websites, the large size of web sites, (consisting of

70 pages on average, while a lot of KWE research is

done on short abstracts) and the fact that the keywords

should not describe the website, but the company be-

hind the website.

5 CONCLUSIONS

We have argued that automatic methods to ﬁnd and

aggregate information on the internet are deserved for

the construction of concise regional knowledge maps

KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

246

in which economic activities and potentials of compa-

nies and other organizations in a region are described.

The abundance of information rich websites of com-

panies seems to offer good possibilities to do so.

In an experimental study we have shown that web-

sites of companies indeed can be used as a source of

information. In the ﬁrst place, factual information like

addresses and phone numbers can be extracted rela-

tively easily, which we did not discuss in more detail

in the paper. Secondly, the description of various as-

pects of a company can be described separately by au-

tomatically extracted keywords when using an appro-

priately structured thesaurus. Thus we go a step fur-

ther than is usually done in KWE, where only one ﬂat

list of keywords is produced. We have evaluated the

keywords for two aspects. For the commodities this

works quite well, yielding a quality of results com-

parable to KWE results reported in literature. For the

aspect economic sectors, the precision of the results is

not very high. Here we are confronted with the prob-

lem that websites do not describe companies but are

in the ﬁrst place description and advertisement for a

wide audience and potential customers. Thus the ter-

minology on a good website will tell us at least as

much about the audience as about the company. The

information about products is much more factual and

less dependent on a point of view than the economic

sector, which explains the difference in performance

between the two aspects. For other aspects we will

haveto developtechniques that are better able to sepa-

rate the information about the writer and the audience

of a website.

Overall we have shown that analysis of web in-

formation and KWE constitute an interesting source

of information for territorial intelligence and the con-

struction of concise knowledge maps, that is worth to

be explored further.

ACKNOWLEDGEMENTS

The research presented in this paper was partially

funded by the Spanish Ministry of Education, Culture

and Sport (Ref. CAS 12/00155).

REFERENCES

Asheim, B. and Gertler, M. (2005). The geography of in-

novation: regional innovation systems. In Fagerberg,

J., Mowery, D., and Nelson, R., editors, The Oxford

Handbook of Innovation., pages 291 – 317. Oxford

University Press, Oxford.

Barinani, A., Agard, B., and Beaudry, C. (2011). Com-

petence maps using agglomerative hierarchical clus-

tering. Journal of Intelligence Manufacturing, pages

1–12.

Canongia, C. (2007). Synergy between competitive in-

telligence (CI), knowledge management (KM) and

technological foresight (TF) as a strategic model of

prospecting — the use of biotechnology in the devel-

opment of drugs against breast cancer. Biotechnology

Advances, 25(1):57–74.

Cunningham, H., Maynard, D., Bontcheva, K., and Tablan,

V. (2002). A framework and graphical development

environment for robust nlp tools and applications. In

Proceedings of the 40th Annual Meeting of the Associ-

ation for Computational Linguistics, July 6-12, 2002,

Philadelphia, PA, USA, pages 168–175. ACL.

David, P. and Foray, D. (1995). Assessing and expanding

the science and technology knowledge base. STI Re-

view, 14:13–68.

De Campos, L. M., Fernandez-Luna, J. M., Huete, J. F.,

and Romero, A. E. (2007). Automatic indexing from

a thesaurus using bayesian networks. In Mellouli, K.,

editor, Symbolic and Quantitative Approaches to Rea-

soning with Uncertainty, pages 865–877. LNCS 4724,

Springer.

Doloreux, D., Nabil, A., and Landry, R. (2008). Mapping

regional and sectoral characteristics of knowledge-

intensive business services: Evidence from the

province of Quebec (Canada). Growth and Change,

39(3):464–496.

Doloreux, D. and Parto, S. (2005). Regional innovation sys-

tems: Current discourse and unresolved issues. Tech-

nology in Society, 27:133–153.

Driessen, S., Huijsen, W., and Grootveld, M. (2007). A

framework for evaluating knowledge-mapping tools.

Journal of Knowledge Management, 11(2):109 – 117.

Eckert, K., Stuckenschmidt, H., and Pfeffer, M. (2007).

Interactive thesaurus assessment for automatic docu-

ment annotation. In Proceedings of the 4th interna-

tional conference on Knowledge capture., pages 103–

110. ACM.

Escorsa, P., Rodriguez, M., and Maspons, R. (2000). Tech-

nology mapping, business strategy and market oppor-

tunities. Competitive Intelligence Review, 11(1):46–

57.

F¨arber, M. and Rettinger, A. (2013). A semantic wiki

for novelty search on documents. In Proceedings of

the 13th Dutch-Belgian Workshop on Information Re-

trieval, pages 60–61, Delft.

Frank, E., Paynter, G. W., Witten, I. H., Gutwin, C.,

and Nevill-Manning, C. G. (1999). Domain-speciﬁc

keyphrase extraction. In Proceedings of the Sixteenth

International Joint Conference on Artiﬁcial Intelli-

gence, IJCAI 99, Stockholm, Sweden, July 31 - August

6, 1999., pages 668–673.

Garcia-Alsina, M. and Ortoll, E. (2012). La Inteligen-

cia Competitiva: evoluci´on hist´orica y fundamentos

te´oricos. Trea, Gij´on.

Garcia-Alsina, M., Wartena, C., and Lieberam-Schmidt,

S. (2013). Regional knowledge maps: potentials

and challenges. In Fifth International Conference

on Knowledge Management and Information Sharing

(KMIS 2013).

ChallengesandPotentialsforKeywordExtractionfromCompanyWebsitesfortheDevelopmentofRegionalKnowledge

Maps

247

Gazendam, L., Wartena, C., and Brussee, R. (2010). The-

saurus based term ranking for keyword extraction. In

Tjoa, A. M. and Wagner, R., editors, Database and

Expert Systems Applications, DEXA, 10th Interna-

tional Workshop on Text-based Information Retrieval,

TIR, pages 49–53. IEEE.

Gazendam, L., Wartena, C., Malais´e, V., Schreiber, G.,

De Jong, A., and Brugman, H. (2009). Automatic an-

notation suggestions for audiovisual archives: Evalu-

ation aspects. Interdisciplinary Science Reviews, 34,

2(3):172–188.

Girardot, J.-J. (2008). Evolution of the concept of territo-

rial intelligence within the coordination action of the

european network of territorial intelligence. Ricerca e

Sviluppo per le politiche sociali, 1(1-2):11–29.

Girardot, J.-J. and Brunau,

E. (2010). Territorial intelli-

gence and innovation for the socio-ecological transi-

tion. In 9th. International conference of Territorial

Intelligence, ENTI, Strasbourg.

Grineva, M. P., Grinev, M. N., and Lizorkin, D. (2009). Ex-

tracting key terms from noisy and multitheme docu-

ments. In Proceedings of the 18th International Con-

ference on World Wide Web, WWW 2009, Madrid,

Spain, April 20-24, 2009, pages 661–670.

Herbaux, P. (2008). Tools for territorial intelligence and

generic scientiﬁc methods. In Internationa Annual

Conference on Territorial Intelligence. Besanc¸on:

16.16 October.

Isaac, A. and Summers, E. (2009). Skos simple knowledge

organization system primer. W3C Working Group

Note. http://www.w3.org/TR/skos-primer/.

Jimenez, F., Fern´andez, I., and Men´endez, A. (2011). Los

sistemas regionales de innovaci´on: revisi´on concep-

tual e implicaciones en am´erica latina. In Los Sistemas

Regionales de Innovaci´on en Am´erica Latina. Banco

Interamericano de Desarrollo, Washington.

Lundvall, B., editor (1992). National Systems of Innova-

tion: Towards a Theory of Innovation and Interactive

Learning. Pinter, London.

Lundvall, B.-A. (1998). Why study national systems and

national styles of innovations? Technology Analysis

& Strategic Management, 10(4):407 – 421.

Lundvall, B. A. and Christensen, J. L. (2003). Broaden-

ing the analysis of innovation systems–competition,

organisational change and employment dynamics in

the danish system. In P., C., Heitor, M., and BA,

L., editors, Innovation, Competence Building and So-

cial Cohesion in Europe. Towards a Learning Society,

chapter Broadening the analysis of innovation sys-

tems–competition, organisational change and employ-

ment dynamics in the Danish system., pages 144–179.

Cheltenham UK: Edward Elgar.

Lundvall, B.-A. and Johnson, B. (1994). The learning econ-

omy. Journal of Industry Studies, 1(2):23–42.

Malais´e, V., Gazendam, L., and Brugman, H. (2007a). Dis-

ambiguating automatic semantic annotation based on

a thesaurus structure. In Hathout, N. and Muller,

P., editors, Actes de la 14e conf´erence sur le Traite-

ment Automatique des Langues Naturelles (commu-

nications orales), pages 197–206, Toulouse. Associ-

ation pour le Traitement Automatique des Langues.

Malais´e, V., Isaac, A., Gazendam, L., and Brugman, H.

(2007b). Anchoring dutch cultural heritage thesauri

to wordnet: two case studies. In ACL 2007, pages 57–

63.

Medelyan, O. and Witten, I. H. (2005). Thesaurus-based

index term extraction for agricultural documents. In

Proc. of the 6th Agricultural Ontology Service work-

shop.

Mollo, M. (2009). The survey on territory research in eu-

rope,. In International Conference of Territorial In-

telligence, Papers on Tools and methods of Territorial

Intelligence (MSHE), Besanc¸on.

Nahapiet, J. and Ghoshal, S. (1998). Social capital, intel-

lectual capital, and the organizational advantage. The

Academy of Management Review, 23(2):242–266.

Nelson, R. R., editor (1993). National Innovation Systems:

A Comparative Study. Oxford University Press, Ox-

ford.

OECD and EUROSTAT (2005). Oslo Manual: Guide-

lines for collecting and interpreting innovation data.

OECD Publising and European Commission. 3rd. edi-

tion.

Salavisa, I. and Vali, M. (2012). Social Networks, Innova-

tion and the Knowledge Economy. Routledge.

Salton, G. and Buckley, C. (1987). Term weighting ap-

proaches in automatic text retrieval. Technical report,

Cornell University.

Sebastiani, F. (2002). Machine learning in automated

text categorization. ACM computing surveys (CSUR),

34(1):1–47.

Tiun, S., Abdullah, R., and Kong, T. E. (2001). Auto-

matic topic identiﬁcation using ontology hierarchy.

In Gelbukh, A. F., editor, Computational Linguis-

tics and Intelligent Text Processing, Second Interna-

tional Conference, CICLing 2001, Mexico-City, Mex-

ico, February 18-24, 2001, Proceedings, volume 2004

of Lecture Notes in Computer Science, pages 444–

453. Springer.

Turney, P. D. (2000). Learning algorithms for keyphrase

extraction. Inf. Retr., 2(4):303–336.

Wang, J., Liu, J., and Wang, C. (2007). Keyword extraction

based on pagerank. Advances in Knowledge Discov-

ery and Data Mining, 4426:857–864.

Wartena, C., Brussee, R., Gazendam, L., and Huijsen, W.

(2007). Apolda: A practical tool for semantic anno-

tation. In Database and Expert Systems Applications,

DEXA, 7th International Workshop on Text-based In-

formation Retrieval, TIR, pages 288–292. IEEE.

Weeds, J., Weir, D. J., and McCarthy, D. (2004). Character-

ising measures of lexical distributional similarity. In

COLING 2004, Proceedings of the 20th International

Conference on Computational Linguistics.

KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

248