How to Boost Customer Relationship Management via Web Mining

Beneﬁting from the Glass Customer’s Openness

Frederik Simon B

aumer and Bianca Buff

Semantic Information Processing Group, Paderborn University, Warburger Str. 100, 33098 Paderborn, Germany

Keywords:

Customer Relationship Management, Web Mining, Local Grammars.

Abstract:

Customer Relationship Management refers to the consistent orientation of a company towards its customers.

Since this requires customer-speciﬁc data sets, techniques such as web mining are used to acquire information

about customers and their behavior. In this case study, we show how web mining can be used to automatically

collect information from clients’ websites for Customer Relationship Management systems in Business-to-

Business environments. Here, we use tailored local grammars to extract relevant information in order to

build up a data set that meets the required high quality standards. The evaluation shows that local grammars

produce substantial high-quality results, but turn out to be too rigid in some cases. In summary, our case study

demonstrates that web mining in combination with local grammars is suitable for Business-to-Business CRM

as long as the information demand can be deﬁned precisely and the requested information is available online.

1 INTRODUCTION

The goal of Customer Relationship Management

(CRM) is to establish, maintain and make use of

successful customer relationships (Link, 2001; Chen

and Popovich, 2003; Zeng et al., 2003). In order

to achieve this, CRM combines and generalizes cus-

tomer information to provide helpful answers to a

great number of business-relevant questions (Buttle

and Maklan, 2015). Although CRM has long been

perceived as business strategy in the Business-to-

Consumer (B2C) sector, it can be applied in both

Business-to-Business (B2B) and B2C environments

(Kurt et al., 2018; Buttle and Maklan, 2015). Its im-

plementation in companies can range from a single

marketing tool to an enterprise-wide strategy.

Regardless of the conditions of use, the acqui-

sition and maintenance of the underlying customer

database is very important for a successful CRM. Ac-

cording to Kurt et al. (2018), the care of necessary

customer data is challenging because customer data

is not often stored in CRM systems but in separate

ﬁles. This is particularly the case in B2B environ-

ments, where key account managers or consultants

manage a smaller number of clients and where per-

sonal loyalty has a high priority (Kurt et al., 2018).

This results in incomplete, outdated and conﬂicting

data sets, which are no operation basis for an efﬁ-

ciently working CRM. Modern CRM systems are not

only seeded with information provided by customer

service and sales staff, but also gather data from e-

commerce systems, access logs or behavior analysis

tools. Information is automatically collected, pro-

cessed and enriched, both in B2C and B2B environ-

ments. A common example are personal usage pro-

ﬁles on web shops. In this case study, we focus on

clients’ company websites as data source for CRM

systems in the B2B context. In the following, we ex-

plain how Web (Content) Mining (WM) can be used

in B2B CRM and how local grammars (LGs) can be

applied to create high quality data sets (Gross, 1993).

Sec. 2 presents related work. Based on this, the show

case (Sec. 3) and a system design for a WM system

will be presented (Sec. 4). The system will be eval-

uated (Sec. 4.3.3) and its results are then discussed

(Sec. 5). Finally, a research outlook is given in Sec. 6.

2 RELATED WORK

In the following, we describe existing work with fo-

cus on CRM in B2B companies as well as on WM

and LGs, which can be used in CRM to create and en-

rich underlying databases (Gross, 1997). Besides very

different deﬁnitions of CRM, numerous CRM goals

exist (Kurt et al., 2018). The primary goal of CRM

can be described as an increase in corporate value and

corporate success, whereby the three secondary goals

Bäumer, F. and Buff, B.

How to Boost Customer Relationship Management via Web Mining Beneﬁting from the Glass Customer’s Openness.

DOI: 10.5220/0007828301290136

In Proceedings of the 8th International Conference on Data Science, Technology and Applications (DATA 2019), pages 129-136

ISBN: 978-989-758-377-3

129

(1) improvement of customer loyalty, (2) improve-

ment of customer proﬁtability and (3) improvement

of customer acquisition are named (Kurt et al., 2018).

In this case study, we focus on improving customer

proﬁtability through the acquisition of CRM-relevant

information from the World Wide Web (WWW). This

information extraction (IE) and monitoring task can

be done manually or with computational assistance.

Since it is important to identify proﬁtable customers

and perceive business-relevant changes as quickly as

possible and to react adequately to them, extensive

automation is desirable. At this point, WM reveals its

value for CRM.

The idea of WM is to use the wide range of in-

formation offered by the WWW for a multitude of

questions. However, WM is deﬁned ambiguously

in the literature (Zhao and Bhowmick, 2003; Kosala

and Blockeel, 2000; Cooley et al., 1997). Within the

scope of this work, WM can be deﬁned as IE us-

ing the WWW as data source with the goal to col-

lect predeﬁned information about B2B clients. How-

ever, WM can be used and implemented in many dif-

ferent ways. Some existing approaches use pattern-

based WM to build databases for question answering

or to build web IE systems (Zhang and Lee, 2002;

Chang et al., 2001). A follow-up of this approach is

the use of (linguistic) grammars to extract informa-

tion, relations, etc. from natural language (NL) texts.

As ﬁnite state machines, such grammars are used in

many ﬁelds of computational linguistics. “From the

linguistics point of view, ﬁnite state machines are ade-

quate for describing relevant local phenomena in lan-

guage research and for modeling some parts of NL,

such as its phonology, morphology, or syntax” (Paji

et al., 2013). Therefore, LGs describe sequences of

words forming semantic units and syntactic struc-

tures. In addition, they provide information about

the morpho-syntactic properties of the elements de-

scribed therein, which can be syntactically or seman-

tically shaped (Geierhos, 2010). Therefore, they are

well suited for the analysis of the structure of proper

names, can deal with domain-speciﬁc knowledge and

sub-languages and capture the interdependencies of

the lexemes among themselves.

In general, LGs are visualized in the form of

graphs. The combination of parameterized graphs

with a dictionary can be extremely effective for the

syntactic analysis of simple sentences (Lee and Geier-

hos, 2011). Graphs are very user-friendly and intu-

itive representations for local grammars, which are

far superior to equivalent formalisms such as regu-

lar expressions. With the help of various drawing

programs, local grammars can easily be created, ex-

tended and edited.

While Iftene and Balahur-Dobrescu (2008) apply

grammar induced patterns on Wikipedia to extract re-

lations between named entities, Paji

c et al. (2013)

use ﬁnite state machines for large Web Monitoring.

Therefore, the potential main beneﬁt of automatic

data acquisition is the following: It saves time and

reduces CRM administration costs. Furthermore, it is

not practicable for e.g. the sales staff to acquire and

maintain information for each single customer manu-

ally – even in B2B context with a smaller number of

customers – since they are not CRM experts or knowl-

edge engineers. This refers especially to information

that does not occur in everyday business processes.

An example is meta data such as the number of em-

ployees of a client’s company to assess the sales po-

tential. However, existing work is mostly dedicated

to WM as an effective technique to extract business

value from companies’ own websites, web shops and

social network channels – mostly with focus on B2C

companies (Mouthami et al., 2013; Kietzmann et al.,

2011; Culnan et al., 2010). But for an overall un-

derstanding of the business market, some companies

also need to monitor activities and analyze informa-

tion on their competitors social media channels and

business websites in order to increase competitive ad-

vantage (Sheng et al., 2005; Croll and Power, 2009).

For example, He et al. (2013) use WM to analyze

user-generated texts on social media sites of the three

largest U.S. pizza chains in order to learn more about

the communication behavior of competing compa-

nies. The better a company is informed about its com-

petitive activities, the better it can manage its own ac-

tivities. Therefore, there exist business software so-

lutions that already offer WM and also implement in-

terfaces to CRM systems in the B2B area. This usu-

ally includes the extraction of non-sector-speciﬁc in-

formation such as address data, the names of the man-

aging director or other executives. Depending on the

information need of companies, however these gener-

alized approaches can be too short-sighted. Basically,

the tools, which are necessary to conduct WM in the

context of B2B CRM, are not new: Machine Learn-

ing (ML) and WM techniques have already been ap-

plied on IE tasks on user-generated content (Tsytsarau

and Palpanas, 2012; Freitag, 2000, 1998) and used

in CRM (Cioca et al., 2013). LGs as IE technique

have also proven themselves (Gross, 1993; Geierhos,

2010). Even though the individual techniques taken

by themselves are well established in various contexts

and use cases, we are not aware of a work that collec-

tively applies them as a combined B2B WM system

to a real business case.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

130

3 SHOW CASE

In the following, we will take a closer look at the use

of WM in B2B CRM systems to discuss the appli-

cability in this context. As show case, we present

the experiences we have gained in cooperation with

a medium-sized special tool production and distribu-

tion company for craft businesses.

The company is active in the B2B sector and

supplies craft businesses with focus on e-commerce

but still relying on traditional sales. The sales staff

receive client-related information from a CRM sys-

tem, which contains both online sales and information

about direct sales, client communication, meta infor-

mation such as number of employees, business areas

in which the client’s craft businesses operates and fur-

ther notes. As this is a highly competitive market, it is

necessary to react quickly to relevant changes among

clients. If there are indications that, for example, a

client is entering a new business area (which is cov-

ered by the company) or that the client’s staff is grow-

ing (additional tools and equipment might be needed),

the sales department is automatically informed.

However, since the number of clients is constantly

growing, manual maintenance of the CRM database

is no longer possible. While the sales staff is used

to update information about the clients manually, this

will be increasingly automated in the future. In the

scope of this paper, we focus on the following meta

information, which is to be stored in the CRM system:

1. Foundation of the client’s business (e.g. 1995)

2. Number of employees (e.g. 20)

3. Business areas (e.g. “painters and varnishers”)

This is a selection of information that the company

uses to analyze the potential of clients and to make

the sales department as effective as possible. We have

chosen these three types of information because they

can be expressed very differently in NL and therefore

have to be collected in different manner. While, for

example, the company foundation is given almost en-

tirely as (numerical) date (e.g. “1998”, “1 February

1998”), the number of employees varies in format (“A

large team of 10 employees”, “we are more than four-

teen hard-working men”).

This circumstance must be taken into account dur-

ing WM. The overall assumption is that customer’s

potential can be read off at least in part from this infor-

mation. For instance, a growing number of employ-

ees awaits selling more equipment and tools. In addi-

tion, an expansion or change in operations, for exam-

ple from painting to painting and refurbishment, may

also indicate that new equipment can be sold. So far,

this information has been entered into the CRM sys-

tem by the sales staff. For this purpose, some rough

estimates have been made: The number of employ-

ees, for example, can be estimated from the number

of company vehicles. The year of foundation often

results from publicly accessible documents and the

classiﬁcation into business areas can be derived from

the sold products. However, this is a time-consuming

task.

For this reason, alternatives were searched for and

it turned out that more than 60% of the clients have a

company website (already available in CRM) and that

often the searched information is available on these

websites. In order to automatically acquire the infor-

mation needed to maintain the CRM system, it is nec-

essary to develop a WM software that automatically

searches and extracts the information. We give insight

into our work on a WM system in the following.

4 SYSTEM DESIGN

The system to be developed should be able to extract

a predeﬁned information from a given website of a

client. All relevant information bits should be col-

lected and transmitted to sales employees, who can

ﬁnally check it and transfer it into the existing CRM

system or reject it. Since the data in the CRM system

must be of high quality, manual veriﬁcation of the in-

formation is an absolute requirement by any company.

However, in order to keep the maintenance time as

short as possible, a very high extraction quality (pre-

cision) should be guaranteed.

4.1 Data Set

As a starting point for our work we have received

2,000 website URLs of existing clients. Of these ad-

dresses, only 1,680 (84%) were still available online.

This does not shed a good light on the current CRM

database. In order to learn more about the way in

which the required information is presented online,

the websites were analyzed manually and correspond-

ing texts were stored.

In addition, we searched the WWW for further

websites of craft businesses and gathered the rele-

vant text. This way, a text corpus of 6,571 sentences

(118,805 tokens, 19,056 types) was created. Exam-

ples for such sentences are: “When the company was

founded on 1st December 1992, only three men were

employed.” or “Today, an average of 100 employees

work on major projects throughout Germany” (Trans-

lated from German).

How to Boost Customer Relationship Management via Web Mining Beneﬁting from the Glass Customer’s Openness

131

Table 1: Number of annotations in the training corpus per class.

Class Annotations Sample sentences

No. Employees 1,778

Unser Team besteht zur Zeit aus 23 Mitarbeitern

Our staff currently consists of 23 employees

Founding year 2,516

Seit 1996 sind wir Ihr Malerbetrieb in Berlin

Since 1996 we are your painting company in Berlin

Business area 4,736

Zum Beruf des Malers geh

ort auch Wissen

uber Brandschutz

The profession of a painter also includes knowledge about ﬁre protection

4.2 Data Acquisition via Web Crawling

In general, a crawler is a software application that

systematically processes a deﬁned, but not necessar-

ily limited, dataset ad inﬁnitum with regard to de-

ﬁned goals and based on conﬁgured instructions for

action. The term (World Wide) Web Crawler refers

to a tool whose targeted data set results from the con-

tent of the WWW (Castillo, 2004). The goals behind

web crawling are diverse: While global search en-

gines aim at the complete indexing of existing web

resources, so-called focused crawlers limit the results

with regard to given criteria (Menczer et al., 2004).

This can be a Top Level Domain (e.g. “.de”), a sub-

ject area (e.g. painting tools) or the relevance to an

information need. To extract information from web-

sites and to monitor websites for changes, a focused

web crawler is required that iteratively accesses given

URLs from the CRM database, downloads the source

codes (HTML) of all relevant pages and stores the

collected data with a timestamp for further process-

ing (cf. Fig. 1). However, the relevance criterion is

difﬁcult to deﬁne, basically all sub-pages of a website

are relevant, which can potentially contain relevant in-

formation. During the initial analysis of clients’ web

pages during text corpus creation, we were able to

create a list of URL patterns, which served as a black-

list to speed up the crawling process. For example, it

is not necessary to access support pages (“/faq/”) or

privacy pages (“/datenschutz/”) because they mostly

do not contain relevant information. Furthermore, we

limited the crawling depth to a maximum of four sub-

pages and the maximum number of pages per client to

150. Even with these limitations, the average acquisi-

tion time is 3.5 minutes per website, which is mainly

due to weak web servers

We have tried classiﬁcation techniques to detect

and ignore irrelevant sub-pages, but the relevant infor-

mation is often represented on sub-pages with com-

pletely different content, so we had too many false

positives and there was a considerable loss of infor-

mation. Therefore, all pages that are not blacklisted

are currently passed to the IE component.

It is necessary to reduce the crawling speed in order not

to attract negative attention, cause damage or be banned.

CRM DB

WM DB

Get (initial) URL

Start

Download HTML

Extract n hyperlinks

Save HTML of i/n

i +=1,

i < n

Terminate

i = n

Figure 1: Basic functionality of the web crawler.

4.3 Information Extraction

The automatic transformation of HTML-source-code

into a semantically structured document (template) is

challenging and requires text preprocessing such as

HTML-stripping and sentence splitting in order to

achieve good results (cf. Sec. 4.3.1). In the follow-

ing, IE is conducted by means of LGs. While ML

methods have proven to be robust in the context of IE

in the past (Chang et al., 2006; Dollmann and Geier-

hos, 2016), LGs are convincing due to their high pre-

cision (Bsiri et al., 2008), since the linguistic con-

text of searched information can be deﬁned very pre-

cisely (Gross, 1993; Geierhos, 2010). As Bsiri et al.

(2008) show, LGs can also achieve a high recall on

user-generated content from the WWW, but require

extensive preprocessing. By choosing LGs, we hope

to meet the high demands on data quality made by our

practice partner.

4.3.1 Data Preprocessing & Data Annotation

We applied data preprocessing to each accessed page

of a client’s website which is structured into the fol-

lowing steps: (1) HTML-removal in order to get

the plain text including removing boilerplate con-

tent, such as navigation, headers and footers

, (2)

Paragraph cleaning via regular expressions to remove

We used: https://pypi.org/project/jusText/.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

132

paragraphs that are too short (less than four words),

(3) Sentence splitting

as IE has to be done on a sen-

tence basis in order to make the annotation of the

relevant information for the practice partner as fast

and simple as possible. The initial annotation of the

sentences with regard to the requested information

(cf. Sec. 3) was carried out by the research team. We

have token-based annotations in three classes (“Em-

ployees”, “Foundation” and “Keyword”). It is very

important for the resulting WM system that it is con-

tinuously fed with new annotated data to improve. For

this reason, we have developed a convenient graphi-

cal annotation interface which enables company em-

ployees to annotate single tokens from sentences re-

garding the matching class by simply drag’n’drop the

tokens. Table 1 shows the number of annotations for

each class.

The annotation of tokens for the company founda-

tion and the annotation of keywords for later alloca-

tion to business areas is not challenging due to the low

number of structural variants. Years are usually given

in two-digit (“99”), four-digit (“1999”) notation or as

a date (“09.02.1988”). Keywords are diverse in their

lexical form, but consist mainly of one or two words.

The number of employees is more difﬁcult to anno-

tate: On the one hand, the numbers are often written

out (“ten skilled workers”), which increases the num-

ber of variants. On the other hand, teams are often

subdivided with regard to their function (“The team

consists of 3 painters and 2 electricians”) or given as

a rough estimation “more than 20”. In these cases, all

relevant tokens are annotated.

4.3.2 Local Grammars

We employed LGs

to transform the preprocessed text

into a semantically structured form. This means that

the original text is automatically enriched with an-

notations that allow the information to be extracted.

This task is conducted by a multitude of LGs that al-

low the modeling of a high degree of syntactic vari-

ation. This ﬂexibility could not be reached by stan-

dard string search, where a list of collected phrases

would be matched with the document, since minor

morpho-syntactic variation prevents the match. Here,

bootstrapping with LGs (Gross, 1999) is a much more

promising method, as illustrated by the following ex-

ample (translated from German).

• “The story of our company, which has 350 quali-

ﬁed employees, began in 1927.”

• “I have a team of twenty men.”

We used: https://spacy.io/.

We used the multilingual corpus processing suite Uni-

tex/GramLab, https://unitexgramlab.org.

• “The company employs more than hundred peo-

ple worldwide in 18 national companies.”

Though these three phrases are different on the

morpho-syntactic level, they contain the same type

of information. By using a LG, it is possible to ef-

ﬁciently represent all three phrases (cf. Fig 2). In the

following, we describe the process of LG modeling

by using the example “number of employees”.

Altogether, the sentences related to employ-

ees in the corpus are composed of 25,406 tokens

(4,306 types). In order to be able to derive pat-

terns, we have examined the context in which the

relevant information is embedded in more detail.

We have found a total of 34 general identiﬁers

such as “Mitarbeiter” (employees), “Beschftigte” (co-

workers), “Fachkrfte” (skilled workers), “Mnner”

(men), “Leute” (people) and “Personen” (persons) as

well as 270 speciﬁc job identiﬁers such as “Maler”

(painter), “Lackierer” (varnisher), “Zimmermann”

(carpenter), “Dachdecker” (roofer) and “Heizungs-

und Sanitrtechniker” (heating and sanitary techni-

cian). The context is further deﬁned by location or

company speciﬁcations such as “Firma” (ﬁrm), “Un-

ternehmen” (company) and “Branche” (sector). Ex-

amples for frequent syntactical constructions are:

• “{who} employ(s) {adv.} {number} {identiﬁer}”

• “{identiﬁer}: {number}”

• “{number} {identiﬁer} are employed by {who}”

Particularly frequent adverbs are “zurzeit”

(presently) and “inzwischen” (meanwhile). Another

specialty are words, which relativize the concrete

number of employees. Examples are “circa”, “rund”

(around) oder “durchschnittlich” (average) and

“insgesamt” (in total). Because LGs support lists,

lexica and sub-graphs, the lexical variants can easily

be considered. However, only what is known can

be modeled. So there will always be constructions

that have not yet been taken into account. With our

current LG (cf. Fig. 2), we are able to cover 88% of

the annotated sentences in the corpus (cf. Tab. 2).

Table 2: Covered sentences in the training corpus [%].

Class Coverage

No. Employees 0.88

Founding year 0.91

Business area 0.85

Some cases cannot be covered because no addi-

tional context is given. This is, for example, the case

if the number of employees is not embedded in a sen-

tence but given as a single number with no context.

How to Boost Customer Relationship Management via Web Mining Beneﬁting from the Glass Customer’s Openness

133

Figure 2: LG for extracting numbers of employees (simpliﬁed).

4.3.3 Evaluation

To evaluate the quality of our WM system regarding

the LGs, we designed a manually annotated test cor-

pus composed of 1,000 sentences. This test corpus

allows us to provide values for recall and precision of

the automatically retrieved results.

Table 3: Evaluation of the LGs performances [%].

Class Precision Recall F-score

No. Employees 0.89 0.72 0.80

Founding year 0.91 0.75 0.82

Business area 0.82 0.47 0.60

Table 3 shows promising results of precision (87%

on average) and recall (65% on average). The ex-

pected low recall value can be explained by uncov-

ered syntactic constructions and missing entries in the

word lists, which have to be constantly maintained

and extended. However, the precision suffers above

all from lexical ambiguities that can arise when the

LGs are weakened in the sense of a higher recall as

can be seen in the following example: “[...] Lukas

32 angestellter Maler [...]” (“[...] Lukas 32 em-

ployed painter [...]”). Here, 32 employees are de-

tected, since the adjective “angestellter” (employed)

is equated with the noun “Angestellter” (employee)

in the LG. The argument that the initial sentence is

grammatically questionable does not apply, because

this is a basic assumption for user-generated content

that LGs have to face. In our case, since the recog-

nized information is checked again, such an error can

be intercepted. However, it is necessary to consider

such cases in future work.

Moreover, knowing that IE is nowadays essen-

tially inﬂuenced by ML methods, we have also ap-

plied an established method to our topic as a compar-

ative value. For short text like the sentences used in

this paper, the MIT Information Extraction (MITIE)

library

was already used in the past (Geyer et al.,

We used: https://github.com/mit-nlp/MITIE/.

2016; King, 2009). MITIE is an open source NLP

library focused on Named Entity Extraction and is

known for state-of-the-art statistical ML. Applied to

our test corpus, MITIE

could convince with a good

average recall value (0.82), but had a comparatively

low precision (0.67). This result may be due to the

small amount of data in the corpus and is not suitable

for a ﬁnal judgment. In fact, MITIE shows very good

results in the extraction of keywords. Its exclusive us-

age can be promising and sensible in an area where

precision is less important due to the large number of

information bits per website and the creation of LGs

is too time-consuming due to the number of variants.

4.4 Monitoring & Data Integration

The monitoring currently includes an automatic re-

visit of the websites once a week. The system stores

extracted information in a database to detect changes

when pages are revisited. If no change is detected

during revisit, no data is transferred to the CRM sys-

tem. However, if a change is detected (this also in-

cludes if the website is not accessible), a change no-

tiﬁcation is sent for review. A notiﬁcation contains

information about the source (URL), the paragraph,

the sentence, and the extracted information. In this

way, the employee can disambiguate and review in-

formation (e.g. conﬂicting foundation data), merge in-

formation bits (e.g. several entries on the number of

employees), and detect overseen information that oc-

curs in the same paragraph. As exchange format, we

chose JSON, which seemed to be suitable because of

the lower overhead compared to XML. A client ap-

plication was developed on the part of the practice

partner, which receives the JSON and presents it to

an employee in a user-friendly way. The employee

makes the ﬁnal decision to transfer the extracted in-

formation to the CRM system. If the extraction result

is erroneous, the employee can reject it. In this case,

an error is reported to the IE system and stored in the

Dictionary: 200,000 words, No. Features: 271.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

134

database. On the one hand, this is valuable informa-

tion for the revision of the LGs and on the other hand,

it is very important training data for further work with

ML techniques. This way, we have fulﬁlled both the

initial requirement for strict data reviewing and cre-

ated further training data by the help of an expert-in-

the-loop approach.

5 DISCUSSION

The approach described in this case study is work in

progress but very promising: information stored or to

be stored in CRM can be automatically updated or

added. The use of LGs proves to be effective because

the achieved quality of information can be regarded

as very high. However, challenges remain that re-

quire further work: The IE system reliably extracts

the type of information we are looking for. However,

this type of information occurs in different contexts

on the clients’ websites. For example, there are web-

sites that present the clients’ company’s history and

contain several (sometimes contradictory) data on the

foundation and expansion of the company. Further-

more, sentences such as “1997: The company cel-

ebrates its 120th anniversary with seven employees,

including ofﬁce workers and trainees” include an in-

dication of when the company was founded (1877),

but the automated identiﬁcation is challenging. The

situation is similar for statements such as “For more

than 60 years”, which do not allow any temporal map-

ping. Complex syntactic and semantic constructions

such as “In the meantime our team includes 5 jour-

neymen, two trainees as well as my wife Nicole and

my daughter Stefanie in the ofﬁce” are also very difﬁ-

cult to grasp, both for LGs and ML techniques. Here

it could turn out to be an advantage that experts re-

view the sentences and can interpret them before the

data is transferred into the CRM. In the future, such

records are to be extracted and transferred as ﬂagged

candidates.

From the point of view of B2B CRM, we can re-

port that 83% of the processed websites could be as-

signed to a business area on the basis of the extracted

keywords. In addition, we were able to ﬁnd and ex-

tract information on the foundation of the respective

company in 31% of all processed websites. In com-

parison, the information on the size of the team could

be extracted much less often automatically, only in

12% of the cases such information was extracted in

live operation.

6 CONCLUSION

As could be shown in this case study, WM com-

bined with suitable IE can lead to noteworthy results

in B2B CRM contexts. In order for such an imple-

mentation to be successful, some challenges and as-

sumptions have to be taken into account. The chal-

lenge for LGs is of course the poor quality of the NL

texts and the lack of context. While the poor text

quality can be partly compensated by a weakening

of the LGs (which probably damages the precision),

a missing context cannot be compensated because a

LG cannot apply on the one hand and on the other

hand cannot disambiguate important information. As-

sumptions that have to be made are that the relevant

information can be deﬁned as a kind of pattern and is

available online on the company’s website. For exam-

ple, addresses and tax numbers can easily be acquired

from public websites, especially thanks to their struc-

tured character, whereas individual product names or

general product characteristics are more difﬁcult to

extract. In future work, we will create more LGs to

acquire further information for the B2B CRM. More-

over, we will improve the reliability of the current

LGs by improving the coverage of the used linguis-

tic resources (e.g. domain-speciﬁc dictionaries).

REFERENCES

Bsiri, S., Geierhos, M., and Ringlstetter, C. (2008). Struc-

turing Job Search via Local Grammars. Advances in

Natural Language Processing and Applications. Re-

search in Computing Science, 33:201–212.

Buttle, F. and Maklan, S. (2015). Customer Relationship

Management: Concepts and Technologies. Routledge,

3 edition.

Castillo, C. (2004). Effective Web Crawling. PhD thesis,

University of Chile.

Chang, C.-H., Kayed, M., Girgis, M. R., and Shaalan, K. F.

(2006). A Survey of Web Information Extraction Sys-

tems. IEEE Transactions on Knowledge and Data En-

gineering, 18(10):1411–1428.

Chang, C.-H., Lui, S.-C., and Wu, Y.-C. (2001). Apply-

ing Pattern Mining to Web Information Extraction.

In Cheung, D., Williams, G. J., and Li, Q., editors,

Advances in Knowledge Discovery and Data Mining,

pages 4–15, Berlin, Heidelberg, Germany. Springer.

Chen, I. J. and Popovich, K. (2003). Understanding cus-

tomer relationship management (CRM): People, pro-

cess and technology. Business Process Management

J., 9(5):672–688.

Cioca, M., Ghete, A. I., Cioca, L. I., and G

ıfu, D. (2013).

Machine Learning and Creative Methods Used to

Classify Customers in a CRM Systems. In Innovative

Manufacturing Engineering, volume 371 of Applied

Mechanics and Materials, pages 769–773.

How to Boost Customer Relationship Management via Web Mining Beneﬁting from the Glass Customer’s Openness

135

Cooley, R., Mobasher, B., and Srivastava, J. (1997). Web

Mining: Information and Pattern Discovery on the

World Wide Web. In Proc. of the 9th IEEE Int. Conf.

on Tools with Artiﬁcial Intelligence, pages 558–567.

Croll, A. and Power, S. (2009). Complete Web Monitor-

ing: Watching your visitors, performance, communi-

ties, and competitors. O’Reilly Media.

Culnan, M., McHugh, P., and I. Zubillaga, J. (2010). How

Large U.S. Companies Can Use Twitter and Other So-

cial Media to Gain Business Value. MIS Quarterly

Executive, 9:243–259.

Dollmann, M. and Geierhos, M. (2016). On- and Off-

Topic Classiﬁcation and Semantic Annotation of

User-Generated Software Requirements. In Proc.

of the 2016 Conf. on Empirical Methods in Natural

Language Processing, pages 1807–1816, Austin, TX,

USA.

Freitag, D. (1998). Information Extraction from HTML:

Application of a General Machine Learning Ap-

proach. In Proc. of the 15th National/10th Conf.

on AAAI/IAAI, AAAI ’98/IAAI ’98, pages 517–523,

Menlo Park, CA, USA. AAAI.

Freitag, D. (2000). Machine Learning for Information Ex-

traction in Informal Domains. Machine Learning,

39(2):169–202.

Geierhos, M. (2010). BiographIE - Klassiﬁkation und Ex-

traktion karrierespeziﬁscher Informationen. Linguis-

tic Resources for Natural Language Processing. Lin-

com.

Geyer, K., Greenﬁeld, K., Mensch, A., and Simek, O.

(2016). Named Entity Recognition in 140 Characters

or Less. In 6th Workshop on Making Sense of Micro-

posts (#Microposts2016), pages 78–79.

Gross, M. (1993). Local grammars and their representation

by ﬁnite automata. Data, Description, Discourse, Pa-

pers on the English Language in honour of John McH

Sinclair, pages 26–38. Harper-Collins.

Gross, M. (1997). The Construction of Local Grammars.

Finite-state language processing. The MIT Press.

Gross, M. (1999). A Bootstrap Method for Constructing

Local Grammars. In Proc. of the Symposium on Con-

temporary Mathematics, pages 229–250. University

of Belgrad.

He, W., Zha, S., and Li, L. (2013). Social media competitive

analysis and text mining: A case study in the pizza

industry. Int. J. of Info. Management, 33(3):464 – 472.

Iftene, A. and Balahur-Dobrescu, A. (2008). Named Entity

Relation Mining Using Wikipedia. In In Proc. of the

6th Int. Language Resources and Evaluation Conf.

Kietzmann, J. H., Hermkens, K., McCarthy, I. P., and Sil-

vestre, B. S. (2011). Social media? Get serious! Un-

derstanding the functional building blocks of social

media. Business Horizons, 54(3):241 – 251. SI: Social

Media.

King, D. E. (2009). Dlib-ml: A Machine Learning Toolkit.

J. Mach. Learn. Res., 10:1755–1758.

Kosala, R. and Blockeel, H. (2000). Web Mining Research:

A Survey. SIGKDD Explor. Newsl., 2(1):1–15.

Kurt, A., Glauser, O. J., and Weiss, E. (2018). Opti-

mierung des Customer Relationship Managements in

B2B-M

arkten, pages 207–224. Springer Fachmedien

Wiesbaden, Wiesbaden, Germany.

Lee, Y. S. and Geierhos, M. (2011). Buy, Sell, or Hold?

Information Extraction from Stock Analyst Reports.

In Beigl, M., Christiansen, H., Roth-Berghofer, T. R.,

Kofod-Petersen, A., Coventry, K. R., and Schmidtke,

H. R., editors, Modeling and Using Context, pages

173–184, Berlin, Heidelberg. Springer Berlin Heidel-

berg.

Link, J. (2001). Grundlagen und Perspektiven des Customer

Relationship Management, pages 1–34. Springer,

Berlin, Heidelberg, Germany.

Menczer, F., Pant, G., and Srinivasan, P. (2004). Topi-

cal Web Crawlers: Evaluating Adaptive Algorithms.

ACM Trans. Internet Technol., 4(4):378–419.

Mouthami, K., Devi, K. N., and Bhaskaran, V. M. (2013).

Sentiment analysis and classiﬁcation based on textual

reviews. In Proc. of the 2013 Int. Conf. on Info. Com-

munication and Embedded Systems, pages 271–276.

Paji

c, V., Vitas, D., La

zeti

c, G. P., and Paji

c, M. (2013).

WebMonitoring Software System: Finite State Ma-

chines for Monitoring the Web. Computer Science and

Info. Systems, 10(1).

Sheng, Y. P., Mykytyn, Jr., P. P., and Litecky, C. R.

(2005). Competitor Analysis and Its Defenses in the

e-Marketplace. Commun. ACM, 48(8):107–112.

Tsytsarau, M. and Palpanas, T. (2012). Survey on mining

subjective data on the web. Data Mining and Knowl-

edge Discovery, 24(3):478–514.

Zeng, Y. E., Wen, H. J., and Yen, D. C. (2003). Cus-

tomer relationship management (CRM) in business-

to-business (B2B) e-commerce. Info. Management &

Computer Security, 11(1):39–44.

Zhang, D. and Lee, W. S. (2002). Web Based Pattern Min-

ing and Matching Approach to Question Answering.

In Proc. of Text REtrieval Conf. 2002.

Zhao, Q. and Bhowmick, S. S. (2003). Sequential Pattern

Mining : A Survey. ITechnical Report CAIS Nayang

Technological University Singapore, 1:26.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

136