CIRA: A Competitive Intelligence Reference Architecture for

Dynamic Solutions

Marco Spruit and Alex Cepoi

Department of Information and Computer Sciences, Utrecht University, Princetonplein 5, Utrecht, The Netherlands

Keywords: Competitive Intelligence, Web Intelligence, Reference Architecture, Text Analytics.

Abstract: Competitive Intelligence (CI) solutions are the key to enabling companies to stay on top of the changes in

today’s competitive environment. It may come as a surprise, then, that although Competitive Intelligence

solutions already exist for a few decades, there is still little knowledge available regarding the implementation

of an automated Competitive Intelligence solution. This research focuses on designing a Competitive

Intelligence reference Architecture (CIRA) for dynamic systems. We start by collecting Key Intelligence

Topics (KITs) and functional requirements based on an extensive literature review and expert interviews with

companies and Competitive Intelligence professionals. Next, we design the architecture itself based on an

attribute-driven design method. Then, a prototype is implemented for a company active in the maritime &

offshore industry. Finally, the architecture is evaluated by industry experts and their suggestions are

incorporated back in the artefact.

1 INTRODUCTION

The business environment has changed a lot during

the last decades—constantly increasing competition

due to globalisation, shorter product lifecycle,

increased popularity of outsourcing as a means of cost

reduction... These are just a few of the many reasons

why companies nowadays need to exploit a lot more

information than before about the competitive

environment on which to base their strategic

decisions (Zanasi, 2001). They need to perform

thorough analyses before committing company

resources towards a product which may not last long

on the market, due to fierce competition or any other

reason.

It is no wonder, then, that a lot of academic

activity has been carried out over the years in the

domain of Competitive Intelligence (CI), a field

which focuses on monitoring the competitive

environment with the aim of providing actionable

intelligence that will provide a competitive edge to an

organisation (Safarnia et al., 2011).

The large growth in the size of unstructured

content created by consumers as well as the

increasing availability of traditional media on the

Internet have brought the necessity of quantitative

analysis of this information via Competitive

Intelligence Capturing Systems (Ziegler, 2012),

tasked with collecting and performing analyses in an

automated manner. A study performed back in 1998

shows that 90% of all information needed by a

company to make informed critical decisions was

already available on the Internet (Teo and Choo,

2001), but Oder (2001) noted that the software

industry was “a long way from delivering a satisfying

business or competitive intelligence solution”.

Some implementation suggestions have been

proposed over the years (e.g. Zhao and Jin, 2010), but

they have not been validated in practice by a

prototype or cross-analysed for similarity. To the

authors’ knowledge there is no in-depth analysis of

what technological solutions should be used to meet

the different requirements of the CI solution

components, or how one would architect a CI solution

based on a set of requirements. With such a big gap

in literature, it is of little wonder that the development

of competitive intelligence systems is still an ongoing

endeavour in most enterprises (Zhao and Jin, 2011).

Therefore, our main research objective is to

develop a technical reference architecture for a

dynamic competitive intelligence system, which can

be easily customisable depending on a specific set of

requirements.

2 RESEARCH METHOD

Our research method starts with a structured literature

Spruit, M. and Cepoi, A..

CIRA: A Competitive Intelligence Reference Architecture for Dynamic Solutions.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 1: KDIR, pages 249-258

ISBN: 978-989-758-158-8

249

review adopting a snowball method. We start from

two searches for competitive intelligence, one

including all papers and one selecting only on papers

newer than 2009. We reviewed the first 10 results of

each query, and for each of them looking up relevant

papers who cite or are cited by them (for a total of 58

papers).

By using the coding method in grounded theory,

we extract the concepts describing KITs, solution

requirements or techniques to be used in

implementation and label them. This process is

iterative, as labels are continually refined and

duplicates are merged.

In order to make sure these are still aligned with

recent developments and in order to better explore the

state of the art, we performed expert interviews with

two companies involved in implementing such a

system and CI professionals. A total of 6 interviews

were conducted with 2 companies and 3 CI

professionals.

We construct a Competitive Intelligence

Reference Architecture (CIRA) by following the

attribute-driven design method proposed by Bass et

al. (2012). This involves an iterative process which

decomposes the architecture into parts. Each part is

designed and tested individually based on functional

requirements and quality attribute requirements.

Evaluation is performed using a mixed method: a

single use-case analysis and expert evaluations which

commented on the design and gave feedback for

further improvements. The use case involves

implementing a prototype for a CI solution for a

company active in the maritime & offshore industry

and assessing how well it performs. The architecture

model is then assessed by an expert panel of 3

professionals with experience in system design and

architecture. Their suggestions for improvement were

incorporated into the final artefacts.

3 LITERATURE REVIEW

There is a large consensus among publications

regarding the three major phases that should make up

any competitive intelligence solution: (i) collection

and storage, (ii) analysis and (iii) dissemination. Data

collection will be the phase responsible for extracting

data from internal or external sources into a collection

of entities we refer to as posts. A post is an abstract

article or unitary text snippet (e.g. Spruit and Vlug,

2015) together with all its metadata (e.g. date, author,

source, user comments). The analysis phase will be

running various algorithms to create CI insights,

which will then be disseminated to the end-users.

Note that these are also compatible with CRISP-DM

based processes (e.g. Otten and Spruit, 2011).

Table 1: Competitive Intelligence solution requirements identified in literature.

CI Phase # Functional Requirement # Sources (e.g.)

Extraction 1 Track changes in posts 4 (Chen et al., 2002)

2 Collect both structured and unstructed data 2 (Rao, 2003)

Analysis 3 Credibility analysis 7 (Zhao and Jin, 2011)

4 Topic exploration 5 (Ziegler, 2012)

5 Document summarisation 2 (Bose, 2008)

6 Scenario simulation 2 (Bose, 2008)

7 Event deduplication 2 (Chen et al., 2002)

8 Integrate structured and unstructured data 2 (Fan et al., 2006b)

9 Product comparison 1 (Xu et al., 2011)

10 Predict competitor data 1 (Cobb, 2003)

11 Company, people and markets profiles 1 (Bose, 2008)

12 Multilingual support 1 (Chen et al., 2002)

Dissemination 12 Ad-hoc querying 8 (Mikroyannidis et al., 2006)

14 Monitoring & alerts 5 (Fan et al., 2006b)

15 Personalised information routing 5 (Fan et al., 2006a)

16 Information visualisation 5 (Rao, 2003)

17 Newsletter summaries 1 (Prescott, 1995)

18 Security policies 1 (Bose, 2008)

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

250

Figure 1: The Key Intelligence Topics (KITs) model.

CIRA: A Competitive Intelligence Reference Architecture for Dynamic Solutions

251

3.1 Key Intelligence Topics and

Solution Requirements

Our literature review resulted in identifying the

following Key Intelligence Topics (KITs) as depicted

in Figure 1. We identify three main types of entities

we would like to extract intelligence about:

companies (including competitors, companies in the

supply chain, and B2B clients), people (including

B2C customers, key employees, and stakeholders)

and markets (inclusing research activity & trends,

regulators, and policy makers).

The next step involves identifying functional

requirements for a CI solution. While the list is not

meant to be a complete reference, we found that we

did not uncover any new requirements during the

interviews process. We classified the functional

requirements into the three main phases for a better

overview in Table 1.

3.2 Architectural Fragments

In the three phases proposed by Bernhardt (1994), we

consider two intermediary steps between data

collection and analysis: pre-processing (the

elimination of clutter from data sources, such as ads

and navigational elements) and post-processing

(including annotating content with part of speech, and

named entities).

We have identified the architectural fragments for

CI implementations presented throughout the

literature. Zhao and Jin (2011) proposes an

architecture reliant on the extraction of entities and

relationships using Information Extraction techniques

and modelling them onto an ontology. Using a rule-

based technique, intelligence insights would be

generated and be subject to a credibility analysis

process. While not going into detail on how this

process would work, he proposes a social-network-

based evaluation model, where the credibility of each

insight would depend on the credibility of the facts

and the sources from where these facts have been

extracted.

Ziegler (2012) proposes a similar architecture, but

proposes using a plain storage for posts. Indexes

would be build by preprocessing the content and

weighing the resulting terms while unsynchronized

annotators would generate annotation data stored in a

separate storage system.

Dai (2013) identifies two main types of analysis

which can be performed: opinion mining (which

extracts general perception from users comments)

and event detection (which clusters posts reporting

the same events). Using the systems mentioned above

as building blocks, Dai (2013) proposes a high-level

CI solution which she calls Data Analysis and

Visualisation AId for Decision Making (DAVID).

Liu et al. (2009) proposes a solution to be used for

trend analysis, called event change detection. The

system analyses differences in the conditional and

consequent parts of association rules in order to detect

emerging patterns, unexpected changes patterns and

added/perished rules.

To analyse the similarities and differences between

these fragments, we compare the 5 fragments. First

we extract the activities for presented in each of the

fragments and then we develop a comparison table by

specifying all activities and comparing them across

all the other fragments. We mark the presence of each

activity with either 3 (if they are present), 7 (if they

are absent), n/a if they are not applicable to the current

fragment (Wei and Lee (2004) and Hu and Liu (2004)

are but subfragments of a complete CI solution) or a

custom string (which means they are present under a

different name or are of a specific type). The

comparison table is present in Table 2.

Table 2: Comparative analysis of Competitive Intelligence fragments identified in literature.

Action Fragment Zhao & Jin

(2011)

Ziegler (2012) Wei & Lee

(2004)

Hu & Liu

(2004)

Dai (2013)

Preprocessing Detect Content



DOM-based

learning

n/a n/a n/a

Post Storage

 

n/a

 

Postprocessing POS Tagger

    

NER + ERD

    

Entity Storage ontology annotations n/a

 

Analysis Event Deduplication (implied)

   

Credibility Evaluation SN Model



n/a

 

Opinion Mining



polarity n/a



(postprocessing)

Trend Detection



co-occurrences n/a

 

Results Storage



annotations n/a n/a Knowledge Base

Dissemination Adhoc Querying

 

n/a n/a



KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

252

4 REFERENCE ARCHITECTURE

Experts deemed the list of solution requirements to be

comprehensive during our interviews, as we did not

uncover any new ones. They advised designing a

modular customizable systems using an annotation

based storage. They also suggest storing metadata and

summary instead of fulltext and using a gazetteer

approach to Named Entity Recognition (NER) since

it generally yields more accurate results.

By combining the identified solution require-

ments, and design decisions identified in literature

and interviews according to the Attribute-Driven

Design method by Bass et al. (2012), we here propose

the high-level Competitive Intelligence Reference

Architecture (CIRA) presented in Figure 2.

4.1 Collection and Storage

We distinguish two types of components which

handle the collection part of a CI solution: either a

crawler (in case the input is an unstructured data

source) or an Extract, Transform, Load (ETL) tool

(for ingesting structured data). The crawlers and ETL

tools would run independently of one another.

We define a crawler as a component which can

execute actions and traverse any unstructured data

source in a predefined way. A web crawler is

probably the most well-know example, although a

generic crawler should be able to extract data from

other types of data sources (all sorts of data APIs, rdf,

etc.).

A web crawler would be responsible with

authenticating itself on the web source (if needed),

following a predefined traversal logic and extracting

from each post the metadata and content. There are,

to the authors’ knowledge, three ways of

implementing the extraction part of the process:

manually-provided selectors (e.g. XPath), template-

based extraction (Zhai and Liu, 2005) or fully

automated extraction (Ziegler, 2012).

The post items then enter through a pipeline of

preprocessors which have the task of cleansing the

data. Examples include stripping html tags, normalize

dates, etc. Optionally, we can at this point filter out

duplicate (or rather unchanged) post items by looking

up a hash value in the database. The post items now

enter a pipeline of postprocessors and filters before

being finally passed to the storage middleware.

The role of postprocessors is to implement post-level

analysis processes (NER could be a candidate here)

during the collection phase. The role of the filters is

to discard the posts which are not relevant for the

user. One usage example is a NER postprocessor and

a filter which discards all posts in which no named

entities have been detected.

Figure 2: The Competitive Intelligence Reference Architecture (CIRA).

CIRA: A Competitive Intelligence Reference Architecture for Dynamic Solutions

253

While many types of storage engines may do the job,

especially for smaller volumes, a column- or

document-oriented database would work best, due to

the flexibility in storage schema (personal communi-

cation, Company C1, 8 jan 2014).

In order to properly update existing data, an

unique identifier needs to exist for each post. This

unique key would be a combination of the source and

an unique site_id (which can be chosen at discretion).

In order to avoid data consistency issues when data

importers and analysis modules run concurrently, we

recommend storing newly crawled posts in separate

snapshots and import them at the end of the collection

process.

4.2 Analysis

The functional requirements of the analysis phase

have been extracted in Table 1. Except for

multilingual support, which involves adapting the

analysis algorithms for other languages, each of the

requirements refer to a distinct analysis module.

Any analysis module which operates at the post

level (without information about the rest of the posts

in the storage) can be implemented as a postprocessor

module. Generally that is recommended as

postprocessors greatly reduce data consistency issues.

It is common for analysis modules to have

dependencies between themselves, so we need to

figure out a way to schedule them in the correct order.

For our usecase running the modules sequentially in

a topological order (the most basic algorithm for task

scheduling) should be a more than adequate solution,

but other more sophisticated algorithms can be used

as well.

One issue that was raised during the expert

validation was preserving data consistency between

analysis modules, since new posts can be imported

while a dependency module is running and the next

module would then find new posts for which the

dependent module has not run. In order to cater for

this the scheduler needs to index the items present

before the chain of analysis modules is scheduled and

use that subset for all modules it schedules. This can

be achieved in an number of different ways, among

which we mention creating a snapshot of post item

identifiers and storing a timestamp and filtering based

on it in each module.

4.3 Information Dissemination

We spent little time discussing information

dissemination, partly because the topic is generic, not

specific to a CI solution. During our literature review

we identified the main methods of delivery to be ad-

hoc queries, monitoring, alerts and digests.

The main methods of enhancing the results are

personalised information routing and visualisation

techniques. Fan et al. (2006a) performs a thorough

comparison of the possible algorithms which can be

used for information routing and evaluates their

performance. Whatever the choice, the architecture

remains the same: a personalised information router

component which gets results either from the

database or from the analysis modules and decides to

which of the users to send them.

5 USE CASE ANALYSIS

In order to test the architecture created, we developed

a prototype for a company active in the maritime &

offshore industry. Their main requirements for a CI

solution were to be able to quickly explore articles

based on competitors, markets and areas as well as

identify patterns and trends over the years in the large

amount of available data.

When we started the use-case design, a

commercial product for information collection was

already being developed, so, in order to bootstrap the

process and due to time constraints, we decided to use

it as a part of our CI solution prototype process. While

it does not perfectly match the architecture presented

in the previous section, it serves as an example of how

existing systems can be used as a “collateral” in

developing a CI solution starting from a reference

architecture (Bass et al., 2012).

Figure 3: High-level Architecture of the prototype.

The technical architecture of this prototype is

depicted in Figure 3. The main analysis components

used were NER as a postprocessor and a Topic

exploration module. Using NER, we were able to

identify for each of the extracted posts:

• Organisations

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

254

Table 3: Example post item in our data set.

{

“lang”:

“english”,

“title”:

“Brent falls under $106”,

“url”:

“http://www.upstreamonline.com/live/article1355570.ece”,

“text”:

“Oil fell under $106 a barrel on Wednesday to a six-week low as

concerns eased about [...]”,

“publish_date”:

“2014/03/19”,

“annotations”:

{

“areas”:

[“Asia > Russia”, “Commonwealth of Independent States > Ukraine”],

“markets”:

[“Offshore Oil & Gas”],

“orgs”:

[“US Federal Reserve”, “American Petroleum Institute”,

“Energy Information Administration”]

“people”:

[“Christopher Bellew”, “President Vladimir Putin”, “Brent”],

} }

(using both a gazetteer and a pre-trained NER)

• People (using a pre-trained NER)

• Areas (using a gazetteer NER)

• Markets (using a gazetteer NER)

For the gazetteer NER, we made use of two

taxonomies which were provided by the client, one

for areas (area > country, e.g. “Asia >

Korea, South”) and another for markets (domain

> market, e.g. “Seagoing Transport > LNG

Tankers”).

5.1 Topic Exploration

We adopt and modify the event change detection

algorithm proposed by Liu et al. (2009) in order to

exploit how association rules between tags evolve

over time (identifying here trends as well as

unexpected new/perished/changed rules).

The example post in Table 3 contains 9 tags

([“areas”, “Asia > Russia”], [“areas”,

“Commonwealth of Independent States >

Ukraine”], [“markets”, “Offshore Oil &

Gas”], …) grouped into 4 features (“areas”,

“markets”, “orgs”, “people”).

Table 4: The top organisations which are co-mentioned

with Norway in the data set.

Arti-

cles

Quarter

occurrences

Feature Co-occurring

organisation

|++++++++|

[“orgs”, “Hyundai”]

|++++++++|

[“orgs”, “Daewoo”]

|++++++++|

[“orgs”, “Samsung”]

|++++++++|

[“orgs”, “Rolls-Royce”]

|++++++++|

[“orgs”, “Bergen Group”]

The next step is what we call tag promotions, a

process which consolidates the tags with the same

meaning into one. Example promotions include

acronyms (OSE becoming Oslo Stock Exchange) or

named entities identified by the classifier being

“promoted” to the named entity specified in the

taxonomy (Hyundai Heavy Industries becoming

Hyundai).

As part of topic exploration we first identify the

top tags for each feature. Table 4 shows an example.

For example, Table 4 lists the top organisations which

are co-mentioned with Norway:

At the beginning of each line in Table 4 is the total

number of articles tagged with both Norway and that

specific organisation, as well as a list of occurrences,

e.g. |+++ ++++|. In this example, an occurrence

was found in each of the quarters except Q4 2012;

there are 8 quarters for 2012-2013 period, and each +

signifies that there was at least one occurrence for that

quarter.

Table 5: Two examples of association rules in the data set.

Confi-

dence

Occurrences Rule

0.84 |++++++++| [“markets”, “Seagoing

Transport > LNG Tanker”] 

[“areas”, “Asia”]

0.73 |++++++++| [“markets”, “Seagoing

Transport > LNG Tanker”]

[“areas”, “Asia”]  [“areas”,

“Asia > Korea, South”]

The second step is to divide all articles into

quarters (3 calendar months periods) and generate

rules using the association rule learning technique for

each of the quarters. To this extent we use a standard

Apriori algorithm for sparse datasets (Agrawal and

Srikant, 1994), which despite its simplicity is one of

CIRA: A Competitive Intelligence Reference Architecture for Dynamic Solutions

255

the best algorithms for association rule learning (Hipp

et al., 2000). The output of the algorithm are rules in

the form as shown in Table 5.

The first rule means that 84% (the confidence) of

all articles tagged with LNG Tanker (LNG stands for

liquified natural gas) will also be tagged with Asia

and that 73% of all articles tagged with LNG Tanker

and Asia are also tagged with South Korea. These two

rules seems to suggest that South Korea is a leading

authority in the LNG market, a fact which was later

confirmed to us by the client.

However, the association rule learning is going to

generate a lot of such rules (in our experiments a

couple hundred per quarter), making it complicated to

analyse them all and spot trends or changes in the

marketplace. The final step is the event change

detection algorithm which compares the rules mined

from two successive quarters in order to identify new,

perished or growing trends as well as unexpected

changes. For each of these changes, Liu et al. (2009)

propose a way to calculate the degree of change (also

marked as deg) Below we present each type of

changes with examples.

1. Emerging: Emerging rules, as shown in Table 6,

are the ones that have been found also in the previous

quarter and are seeing an increase in support in

current quarter (we use a threshold of 20%, i.e. a ratio

between the two values of 1.2 or larger).

Table 6: Rules for query “emerging 2013 Q3 => 2013 Q4”.

Change Rule

+1.05 [“areas”, “Asia > Russia”] 

[“areas”, “Arctic Region”]

+1.05 [“areas”, “Arctic Region”] 

[“areas”, “Asia > Russia”]

+0.78 [“areas”, “Asia > Russia”] 

[“markets”, “Offshore Oil & Gas”]

2. Added: Added rules, as shown in Table 7, are rules

that have been found in this quarter but were not

found in the previous one.

Table 7: Rule for query “added 2012 Q4 => 2013 Q1

(query: LNG Tanker)”.

Change Rule

+0.20 [“markets”, “Seagoing Transport > LNG

Tanker”] [“areas”, “Asia > Japan”] 

[“areas”, “Asia > Korea, South”]

Table 8: Perished rule for query “perished 2012 Q4 => 2013

Q1 (query: Shell)”.

Change Rule

+0.21 [“markets”, “Offshore Oil & Gas”] [“orgs”,

“Shell”]  [“areas”, “Asia > Korea , South”]

3. Perished: Perished rules, as shown in Table 8, are

the ones that have been found in previous quarters,

but not in the current one.

4. Consequent Change: Consequent change rules, as

shown in Table 9, are unexpected changes in the

consequent part of a rule. The rule found in the

previous quarter is not found in the current one, but

instead there is a rule with a very similar conditional

part and a different consequent part.

Table 9: Consequent change rule for query “consequent

change 2013 Q1 => 2013 Q2”.

Change Rule

+1.05 [“areas”, “Europe > Norway”]  [“areas”,

“Asia > China”]  [“areas”, “Europe > Nor-

way”]  [“areas”, “Asia > Korea , South”]

5. Condition Change: Condition change rules, as

shown in Table 10, are unexpected changes in the

condition part of a rule. They operate similarly to the

consequent changes.

Table 10: Condition change rule for query “condition

change 2012 Q4 => 2013 Q1”.

Change Rule

+1.01 [“markets”, “Ship Repair”]  [“areas”, “Asia

> China”]  [“markets”, “Seagoing

Transport > Container Vessel”]  [“areas”,

“Asia > China”]

Finally, we classify the most popular rules by

averaging their confidence across quarters in Table

11. This shows the most important trends which have

been happening for the entire period of two years. For

example the most prominent trends identified with the

entire dataset suggest Hong Kong’s interest in search

and rescue vessel markets as well as Daewoo and

Hyundai doing business together in South Korea.

5.2 Client Feedback

Due to KDIR page constraints we refer to Cepoi

(2014) for the in-depth evaluation of our prototype.

6 DISCUSSION

During the research process we discovered many of

its modules to be research topics in their own right.

To cater for this variability we limited ourselves to a

higher level architecture which focuses more on the

interaction and requirements of the analysis modules

rather than their implementation details.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

256

Table 11: The Top-3 association rules indicate the most important trends during the two year-period.

Confidence Occurrences Rule

1.0 |++++++++| [“areas”, “Southeast Asia > Hong Kong ( SAR)”] 

[“markets”, “Environmental Safety & Control > Search and Rescue Vessel”]

0.994 |++++++++| [“areas”, “Southeast Asia > Hong Kong ( SAR)”] 

[“markets”, “Defence & Security > Search and Rescue Vessel”]

0.946 |++++++++| [“orgs”, “Daewoo”] [“orgs”, “Hyundai”]  [“areas”, “Asia > Korea , South”]

We argue that our reference architecture is more

flexible than the framework proposed by Zhao and Jin

(2011), which assumes a CI solution only focusing on

basic rule-based intelligence generation, while

catering equally well for that scenario. It is also more

descriptive than the architectures from Ziegler (2012)

or Dai (2013), which specify little more than the

possible analysis modules.

The architecture has been validated both by industry

experts and by being instantiated in a prototype,

which delivered results evaluated by the customer to

be accurate. We believe this research will prove

useful for future implementation projects.

REFERENCES

Agrawal, R and R Srikant (1994). Fast algorithms for

mining association rules. VLDB, 487-499, Santiago.

Bass, L. et al. (2012). Software architecture in practice.

Addison-Wesley.

Bernhardt, D. (1994). ‘I want it fast, factual, actionable’—

tailoring competitive intelligence to executives’ needs.

Long Range Planning, 27(1), 12–24.

Bose, R. (2008). Competitive intelligence process and tools

for intelligence analysis. Industrial Management &

Data Systems 108(4), 510–528.

Cepoi, A. (2014). A Reference Architecture for a Dynamic

Competitive Intelligence Solution. MSc thesis, Utrecht

Univ. http://dspace.library.uu.nl/handle/1874/298561.

Chen, H. et al. (2002). CI Spider: a tool for competitive

intelligence on the Web. Decision Support Systems

34(1), 1–17.

Cobb, P (2003). Competitive Intelligence through Data

Mining. Journal of competitive intelligence and

management 1(3), 80–89.

Dai, Y. (2013). Designing Text Mining-Based Competitive

Intelligence Systems. Doctoral dissertation, UEF.

Fan, W. et al. (2006a). An integrated two-stage model for

intelligent information routing. Decision Support

Systems 42(1), 362–374.

Fan, W. et al. (2006b). Tapping the power of text mining.

Communications of the ACM 49(9), 76–82.

Hipp, J. et al. (2000). Algorithms for association rule

mining — a general survey and comparison. ACM

SIGKDD Explorations Newsletter 2(1), 58–64.

Hu, M. and Liu, B. (2004). Mining and summarizing

customer reviews. 10

ACM SIGKDD international

conference on Knowledge discovery and data mining.

Liu, D. et al. (2009). Mining the change of event trends for

decision support in environmental scanning. Expert

Systems with Applications, 36(2), 972–984.

Mikroyannidis, A. et al. (2006). PARMENIDES: towards

business intelligence discovery from web data, Int.

Conference on Web Intelligence, 1057–1060.

Nadeau, D. and Sekine, S. (2007). A survey of named entity

recognition and classification. Lingvisticae

Investigationes 30(1), 3–26.

Oder, N. (2001). The Competitive Intelligence Opportunity,

1–3.

Otten, S. and Spruit, M. (2011). Linguistic engineering and

its applicability to business intelligence: towards an

integrated framework. Int. Conf. on Knowledge

Discovery and Information Retrieval, Paris, 460–464.

Prescott, J. (1995). The evolution of competitive

intelligence. International Review of Strategic

Management 6, 71–90.

Rao, R (2003). From unstructured data to actionable

intelligence. IT Professional 5(6), 29–35.

Safarnia, H. et al. (2011). Review of Competitive

Intelligence & Competitive Advantage in the Industrial

Estates Companies in the Kerman City. International

Business and Management 2(2), 47–61.

Spruit, M., and Vlug, B. (2015). Effective and Efficient

Classification of Topically-Enriched Domain-Specific

Text Snippets. International Journal of Strategic

Decision Sciences 6(3), 1–17.

Teo, T. S. and W. Y. Choo (2001). Assessing the impact of

using the Internet for competitive intelligence.

Information & management 39(1), 67–83.

Wei, C.-P. and Y.-H. Lee (2004). Event detection from

online news documents for supporting environmental

scanning. Decision Support Systems 36(4), 385–401.

Xu, K et al. (2011). Mining comparative opinions from

customer reviews for Competitive Intelligence.

Decision Support Systems 50(4), 743–754.

Zanasi, A. (2001). Competitive intelligence through data

mining public sources. Competitive Intelligence Review

9(1), 44–54.

Zhai, Y. and B. Liu (2005). Extracting Web Data Using

Instance-Based Learning. Web Information Systems

Engineering–WISE 2005, 318–331.

Zhao, J. and P. Jin (2010). Conceptual Modeling for

Competitive Intelligence Hiding in the Internet. Journal

of Software 5(4).

CIRA: A Competitive Intelligence Reference Architecture for Dynamic Solutions

257

Zhao, J. and P. Jin (2011). Extraction and Credibility

Evaluation of Web-based Competitive Intelligence.

Journal of Software 6(8), 1513–1520.

Ziegler, C.-N. (2012). Competitive Intelligence Capturing

Systems. Mining for Strategic Competitive Intelligence,

51–62.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

258