Exploring Corporate Reputation based on Sentiment Polarities That

Are Related to Opinions in Dutch Online Reviews

R. E. Loke

and J. Vergeer

Centre for Market Insights, Amsterdam University of Applied Sciences, Amsterdam, The Netherlands

Keywords: Aspect Based Sentiment Analysis (ABSA), Opinion Mining, English Texts, Dutch Texts.

Abstract: This research demonstrates the power and robustness of the vocabulary method by Hernández-Rubio et al.

(2019) for aspect extraction from online review data. We showcase that this algorithm not only works on the

English language based on the CoreNLP toolkit, but also extend it on the Dutch language, specifically with

aid of the Frog toolkit. Results on sampled datasets for three different retailers show that it can be used to

extract fine-grained aspects that are relevant to acquire corporate reputation insights.

1 INTRODUCTION

As evidenced by GfK (2019), the number of online

purchases in the Netherlands has continuously risen

in recent years. Along with this rise in purchases, the

importance of online reviews is also evident as over

90% of online shoppers research into an online

retailer through customer reviews (Statista, 2019). In

addition, it is well known that the number of online

reviews that a retailer has, affects its sales (Chevalier

& Mayzlin, 2006). Following these trends, the

increasing importance of online reviews might

suggest that online reviews form a relevant company

asset that, at least to some extent, give an indication

of an organization’s reputation. At the moment little

research has been done on this subject; the literature

on corporate reputation that does exist mainly

employs surveys which is not always the most readily

available data. If it were possible to partly base

corporate reputation on online reviews it could

logically provide meaningful insights to stakeholders

like online retailers to boost their performance. Thus,

corporate reputation retained from online reviews

could prove to be an intangible yet important resource

for an organization. Intangible resources are in

general known to be important due to the business

value they can potentially create and the difficulty of

replication by competitors (Roberts & Dowling,

2002).

https://orcid.org/0000-0002-7168-090X

With regard to processing capabilities of online

reviews that are related to corporate reputation, a

comprehensive analysis using unsupervised

techniques and the automatic extraction of aspects

could be useful and might be worth researching into

as there are often no explicit ground thruths available

for online review data. Our research will therefore

focus on this angle. Hastie, Tibshirani and Friedman

(2009) describe unsupervised learning as attempting

to infer properties from data without having a ground

truth available: unsupervised learning techniques lack

a clear cut measure of success and are generally used

when there is no ground truth available in the data. It

is, at least, an interesting alternative to expensive

supervised solutions that require example output

labels along with online review data to work properly.

In addition, the performance of supervised systems is

often limited due to a certain practical imbalance that

is commonly observed in data in terms of the number

of positive reviews available in comparison to the

number of negative reviews. The vocabulary method

by Hernández-Rubio et al. (2019) is a candidate

unsupervised technique that allows to extract fine-

grained aspects from texts and facilitates automatic

labeling of their sentiment polarities. It depends on

dependency parsing and, with the method, already

excellent results have been obtained in the English

language on general texts. In this paper, we research

into applying and extending this precise method to

online reviews in Dutch.

Loke, R. and Vergeer, J.

Exploring Corporate Reputation based on Sentiment Polarities That Are Related to Opinions in Dutch Online Reviews.

DOI: 10.5220/0011285500003269

In Proceedings of the 11th Inter national Conference on Data Science, Technology and Applications (DATA 2022), pages 423-431

ISBN: 978-989-758-583-8; ISSN: 2184-285X

423

Table 1: Definitions on the three dimensions that comprise

corporate reputation according to Lange et al. (2011, p.

155).

Dimension Definition

Being known

“Generalized awareness or

visibility of the firm; prominence

of the firm in the collective

erce

tion”

Being known

for something

“Perceived predictability of

organizational outcomes and

behavior relevant to specific

audience interests”

Generalized

favorability

“Perceptions or judgements

of the overall organization as

good, attractive and

riate”

In the rest of the paper, we review relevant

literature in Section 2, describe our processing

framework as well as the online review data that we

obtained in Section 3, give results in Section 4, and

discuss and conclude the work in Section 5.

2 LITERATURE REVIEW

In this section, an overview of recent scientific

literature on corporate reputation and online reviews

is given. Thereafter, we also focus on two relevant

generic text mining methods, being topic extaction

and sentiment analysis.

2.1 Corporate Reputation

Definitions of corporate reputation can differ when

looking at multiple sources. Barnett, Jermier &

Lafferty (2006, p. 34) define corporate reputation,

after reviewing the definitions used in a multitude of

other sources, as the “Observers’ collective

judgments of a corporation based on assessments of

the financial, social, and environmental impacts

attributed to the corporation over time”. This is, at

least in part, in line with Hall’s (1993) findings where

an organization’s reputation is usually built over

multiple years, yet also fragile. Gotsi and Wilson

(2001) present two schools of thought, one where

corporate reputation is synonymous with corporate

image and one where the two concepts are seen as

different, but generally are also seen as interrelated.

Barnett et al. (2006) argue that we should move away

from seeing corporate reputation and image as being

one and the same, which is in line with the second

school of thought of Gotsi and Wilson (2001). This

definition, where corporate reputation and image are

interrelated yet not synonymous, would be logical to

adopt for this research with online reviews as the

other vocal point of interest. Online reviews

encompass a customer’s experience with a company

and not necessarily how that company is seen by said

customer.

Table 2: The seven dimensions that make up corporate

reputation according to Fombrun et al. (2015).

Dimension Definition

Products &

Services

Whether the company’s

offerings are considered of high

quality, value, and service, and if

customer needs are met.

Innovation

Whether or not a company is

innovative and ada

tive.

Workplace

Perceptions of the way a

company shows concern for its

employees along with rewarding

and treatin

them fairl

Governance

Perceptions of how a

company is held regarding ethics,

fairness and transparency.

Citizenship

How a company is perceived

regarding the environment,

whether it supports good causes

and its contributions to societ

Leadership

Whether leaders within a

company are perceived as

visionary and if they endorse their

com

anies.

Performance

Perceived financial

performance, profitability, and

growth.

Lange, Lee & Dai (2011) found that corporate

reputation can be categorized along three different

dimensions; see Table 1. Given the context of this

research, the being known for something dimension

seems to be most relevant as specific aspects like

delivery and product quality would intuitively be the

most prevalent in reviews. Generalized favorability

might be difficult to assess on individual review level

but could be seen along the lines of overall sentiment

on a company. Lange et al. (2011) do underline that

corporate reputation should be viewed along more

than one dimension. While useful in conceptualizing

corporate reputation, Lange et al. (2011) really focus

on defining corporate reputation and its underlying

dimension as opposed to quantifying it. They do not

mention any specific variables that could be used to

model the dimensions. This seems to occur often as

other research into corporate reputation, like Lange et

al. (2011), generally put a heavy focus on defining

corporate reputation and the role it plays in an

organization’s success instead of attempting to, for

example, quantify it. One of the articles that does lean

more in this direction is the article by Fombrun, Ponzi

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

424

and Newburry (2015). This article attempts to

validate a system that assesses an organization’s

reputation from the perspectives of different

stakeholders. They do this by validating the

dimensions that this system uses, namely: “Products

& Services”, “Innovation”, “Workplace”,

“Governance”, “Citizenship”, “Leadership” and

“Performance”. For definitions on these dimensions,

see Table 2. For our research the possibility of

extracting all seven dimensions from online reviews

does not seem likely. A dimension like, for example,

“Leadership” is not likely to be relevant in an online

customer review. A dimension like “Products &

Services” and perhaps “Innovation” could be

relevant, however.

Unlike Lange et al. (2011), Fombrun et al. (2015)

do mention specific variables that make up their

found dimensions, see Table 3. These variables show

that this specific model has been made to be based on

the perspectives of multiple stakeholders. The “stands

behind” variable of the Products & Services

dimension focuses on the company while the other

variables of that dimension focus more on the

products and services themselves. This is relevant

since in this research, the online reviews from

Trustpilot generally encompass the views of the

customer stakeholder group.

Table 3: Variables that make up the seven corporate

reputation dimensions of Fombrun et al. (2015).

Dimension Variables

Products &

Services

High Quality; Good Value;

Stands Behind; Meets Customer

Needs

Innovation

Innovative; First to Market;

Adapts to Change

Workplace

Rewards Employees Fairly;

Concern for Employees’ Well-

Being; Equal Opportunities in

Workplace

Governance

Open and Transparent;

Behaves Ethically; Fair in Doing

Business

Citizenship

Protects Environment;

Supports Good Causes; Positive

Influence on Societ

Leadership

Strong and Appealing Leader;

Clear Vision of Future; Well

anized; Excellent Mana

ers

Performance

Profitable; Good Financial

Results; Stron

Growth Pros

ects

In conclusion, corporate reputation is defined as a

collective view on a corporation. The reputation is

built over time and can be fragile. In the context of

online reviews, it likely can only be measured in the

“being known for something” and “generalized

favorability” dimensions of Lange et al. (2011).

These dimensions are quite broad however and were

not designed to be used directly as quantification

dimensions. The model by Fombrun et al. (2015)

could be used, but likely only in part as it concerns

more stakeholder groups than that single-faceted

research would logically focus on. The amount of

work done in attempting to quantify corporate

reputation remains quite limited and the models that

do exist are likely not completely applicable.

2.2 Online Reviews

Online reviews can be a valuable resource for insights

in an organization’s performance. Wang, Lu and Zhai

(2010) state that with Web 2.0 it is possible for large

groups of people to express their opinions on many

things like products and services. They continue in

saying that these opinions contribute to both other

users and retailers as they enable both types of parties

to gather information and make educated decisions.

Xing and Zhan (2015) do note some flaws in

this, the first being that with every single person being

able to post their opinions, the quality of these

opinions cannot be guaranteed. They note some

examples like online spammers and fake reviews. The

second flaw Xing and Zhan (2015) note is that a

ground truth in the context of, for example, online

reviews is a “tag of a certain opinion” and not

necessarily an established truth. It thus can be useful

for gauging the opinions of a customer base, but they

do not necessarily display an accurate description of

the quality of a product/service. This is exemplified

by Korfiatis, García-Bariocanal and Sánchez-Alonso

(2012) who state that the experience that multiple

consumers have with the same product can be

different due to, for example, differences in

expectations. It can however be argued that the

aggregated opinions of a customer base are a strong

indicator of quality, making the true truthfulness of

online reviews somewhat context reliant.

2.3 Text Mining

The analysis of textual data, also known as text

mining, can have a different definition depending on

the area in which it is being applied (Hotho,

Nürnberger and Paaß, 2005). The latter authors focus

on the area which is also most relevant for this

research and as such define text mining as “the

application of algorithms and methods from the fields

machine learning and statistics to texts with the goal

of finding useful patterns” (Hotho et al., 2005, p. 4).

Exploring Corporate Reputation based on Sentiment Polarities That Are Related to Opinions in Dutch Online Reviews

425

The paper by Hotho et al. (2005) mentions multiple

methods that fall under text mining, of which some

are relevant to this research and need to be discussed

and explored further.

To analyze text data it first is in general in step (1)

necessary to find structure in them, this can be

achieved with techniques like topic modelling (step

1A) or through the classification of syntactic

dependencies (step 1B). After that, in step (2), the

data can be analyzed on aspects like the prevalent

sentiments in the text. As the granularity of the data

increases, however, the difficulty of the analysis will

also increase.

The success of working with textual data can be

reliant on the language used in the data and on data

preparation (pre) processing that is commonly done

in step (0). Relatively simple operations like

stemming (reducing words to their root (Lovins,

1968)) and removing stop words (words that carry a

grammatical function but do not carry a meaning of

the document’s content (Wilbur & Sirotkin, 1992))

might largely still be successful but more refined

operations like lemmatization (changing words to

their basic forms (Korenius, Laurikkala, Järvelin, &

Juhola, 2004)) can be more difficult. The Stanford

CoreNLP toolkit (Manning, et al., 2014) carries a

lemmatize option but does not support the Dutch

language. An alternative is the Frog toolkit (Van den

Bosch et al., 2007). Frog was specifically made for

the Dutch language and can, among others, lemmatize

words, add part of speech tags (POS) and add

syntactic dependency tags.

Below we describe topic extraction (step 1A) and

sentiment analysis (step 2). Modeling of syntactic

relationships (step 0) and classification of syntactic

dependencies (step 1B) will be described in Section

3.2 and Section 3.3.

2.3.1 Topic Extraction

The first step besides data preparation often is to find

the most important groups of aspects that can be

found in review data. By applying LDA and BTM,

underlying topics in the available reviews can be

extracted to gain insight into the aspects these reviews

focus on. These aspects could in our application that

is directed to social listening from Trustpilot online

reviews for example be subjects like product quality

or delivery time.

According to Hu, Boyd-Graber, Satinoff and

Smith (2014), the problem of dealing with large

volumes of unstructured textual data remains a

persistent problem. To gain insight in underlying

themes that can be found in texts, topic modeling can

be a potential solution. A well-known and often used

topic modeling technique is “Latent Dirichlet

Allocation” (LDA) by Blei, Ng and Jordan (2003).

The idea behind this technique is that a text document

is composed of a finite number of underlying “topics”

or themes. In the article where they originally

presented their findings Blei et al. (2003) test the

effectiveness of the topics found by LDA by using

them as features in a classification model. They report

that not only did LDA reduce the dimensionality of

the model (it was tested against a model using word

occurrence), but it also achieved high performance.

As such it can be seen as a method that can bring

structure to text data which could solve a problem

with the analysis of text data mentioned by Hotho et

al. (2005), namely that text data is unstructured and

as such cannot easily be processed by a computer.

It must be noted however that topic models are not

perfect. According to Hu et al. (2014), it can very well

occur that a found topic is nonsensical, practically a

duplicate of another topic or a combination of

multiple topics that logically speaking should be

separate.

Another method for extracting topics is the

“biterm topic model” (BTM) by Yan, Guo, Lan, and

Cheng (2013) whose authors state that conventional

topic models like LDA may not work well on short

texts due to data sparsity in these documents. Weng,

Lim, Jiang, and He (2010), for example, attempt to

remedy this sparsity by aggregating the texts that they

analyze based on author. Instead of assuming that a

text is comprised of a mixture of topics, BTM

assumes that the entire corpus is a mixture of topics

where the cooccurrence of two words (a biterm) is

independently related to a topic. With BTM it could

be interesting to extract topics from the reviews as

these are generally shorter than, say, a chapter of a

book.

Since LDA and BTM are being unsupervised

learning techniques (there is no predefined ground

truth available) a potential outcome is a collection of

nonsensical topics. To assess topic models and find

the optimal number of topics, the perplexity metric is

commonly used as was done in Blei et al. (2003).

It is important to mention that topic modeling

approaches (step 1A) proved to be unsuccessful for

our application goals. Therefore, we adopted step 1B

instead.

2.3.2 Sentiment Analysis

After that text data has been structured it can be more

readily analyzed, for example through sentiment

analysis. Sentiment analysis is analyzing the

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

426

sentiment that a group of people hold towards specific

entities like products, services, or organizations (Xing

& Zhan, 2015; Liu, 2012). According to Liu (2012)

there are, broadly speaking, three levels of sentiment

analysis: document level, sentence level and entity

level. On the document level, sentiment is assessed

over the entire text, on sentence level, sentiment is

assessed per sentence. Entity level is the most fine-

grained level where sentiment is really specified on

the opinions themselves. Liu does note that as the

analysis level gets more specific, the level of

difficulty of the analysis also rises. This is in line with

Nguyen, Shirai and Velcin (2015) who state that

sentiment analysis on social media texts is difficult

due to problems like short text lengths, spelling errors

and uncommon grammar.

Given the focus of our research, sentiment

analysis is likely to play an important role. There are

multiple ways to approach it, Farhadloo and Rolland

(2013), for example, use a machine learning

approach, whereas Vashishtha and Susan (2019) use

a rule-based approach. The common factor of both

approaches is that they use sentiment scoring in some

shape or form. These scores represent the sentiment

of the analyzed aspect(s) be they individual words,

sentences, or entire texts.

3 DATA AND METHODOLOGY

In this section, we describe the online review datasets

that we obtained, the modeling of syntactic

relationships in the review texts as well as the aspect

extraction and sentiment labeling that we applied in

our processing framework.

3.1 Online Review Datasets

The primary data that this research employs concerns

online reviews collected from trustpilot.com as this

website by far contains the largest number of online

reviews. Trustpilot.com is an often-used website for

writing and sharing online reviews, it is in the top 1%

most popular websites and boasts over 97 million

reviews on over 420,000 companies (Trustpilot,

2020).

Our specific data concern reviews on 3 different

online retailers operating in the Netherlands:

Bol.com, Agradi, and ABOUT YOU. The number of

reviews, being exactly 100 for each company, is

relatively low and is only meant for the purpose of

demonstrating the use of any techniques found to be

performing well. In future work, once techniques

have been completely finetuned properly, we

obviously plan to process larger datasets for diverse

stakeholders and our partners.

The data was collected using the web scraping

framework Scrapy in the Python programming

language. The ratings on the Trustpilot website that

come along with the reviews were scraped as well to

test several validity assumptions. In general, ratings

can give a sense of the overall sentiment of a review

but do not say much on the more fine-grained topics

that corporate reputation is assessed on.

3.2 Modeling of Syntactic

Relationships

An interesting way that turns out to be quite useful of

finding structure in unstructured textual data is to

look at the syntactic relations in texts. According to

Qiu, Liu, Bu and Chen (2011) the syntactic relation

between two words (A and B) can be defined as: A

depends on B, or B depends on A. These

dependencies can differ in nature, an adjective might

be dependent on a noun but that same noun might be

dependent on a verb. One relation might specify the

subject of a sentence, others might specify a

modification of the meaning of that sentence (adding

a sentiment or modifying the intensity of that

sentiment). Applying this definition to a full sentence,

a syntactic tree can be made, visualizing relationships

between words. For an example of such a

visualization, see Figure 1. Annotations of these

relations between words are usually not readily

available (especially in user generated content like

online reviews) and as such an often-used solution is

a so-called dependency parser like the one available

in the Stanford CoreNLP toolkit (Manning, et al.,

2014; Chen & Manning, 2014).

Figure 1: Syntactic tree of the sentence: “The big house and

the nice garden”, classified by the Stanford CoreNLP

toolkit.

Syntactic dependencies can be extracted in more

languages than just English, the Stanford CoreNLP

toolkit can work with five different languages

(Manning, et al., 2014). The Dutch language is not

supported, however; an alternative is the Frog toolkit

Exploring Corporate Reputation based on Sentiment Polarities That Are Related to Opinions in Dutch Online Reviews

427

by Van den Bosch, Busser, Canisius and Daelemans

(2007) which was designed specifically for the Dutch

language. Frog is capable of many of the tasks

CoreNLP can perform, one of which being able to

extract syntactic dependency tags, as evidenced in

Figure 2.

Figure 2: Syntactic tree of the sentence: “Het grote huis en

de mooie tuin”, classified by the Frog toolkit.

In terms of syntactic dependency tags, CoreNLP

and Frog largely convey the same information.

Looking at the syntactic trees in Figures 1 and 2 there

is at least one structural difference, however. The way

conjunctions like “and” in Figure 1, and “en” in

Figure 2, are noted makes for different types of

relations between words, between English and Dutch.

Individually this might be of minimal impact, but it is

something to take note of when using an algorithm

that takes syntactic dependency tags as input.

3.3 Aspect Extraction

There are multiple approaches that can be taken to

extract aspects or entities from a text. The

(unsupervised) vocabulary-based method described

by Hernández-Rubio, Cantador, & Bellogín (2019, p.

404) is based on the syntactic dependency and POS

tags generated by the Stanford CoreNLP toolkit and

extracts sentiment polarities that are related to aspects

on entity level in English texts. This algorithm

generates outputs in the form of:

(noun, adjective, modifier, isAffirmative) (1)

where noun is the extracted aspect with adjective

specifying a positive, neutral, or negative opinion

(sentiment polarity whereto the adjective adheres to,

see next subsection how the sentiment polarity is

being determined) which is potentially modified by

modifier. The variable isAffirmative specifies whether

the sentiment polarity must be inverted due to the

word “not” being used.

The vocabulary method of Hernández-Rubio et al.

(2019, p. 404) is in their paper available as pseudo-

code. This code has been implemented from scratch

in Python, using the results gained from applying the

Stanford CoreNLP toolkit for English texts and the

Frog toolkit for Dutch texts. As has been mentioned

before, in comparison to CoreNLP, Frog uses

different syntactic dependency tags, but largely

conveys the same information.

3.4 Sentiment Classification

Different approaches can be adopted to map

adjectives to sentiment polarities.

A labor-intensive possibility for finding ground

truths would be to manually tag adjectives in reviews

according to their sentiment.

Fang and Zhang (2015) take the approach of

calculating the overall sentiment on sentence level

using predefined sentiment scores by looking at the

used words. With this approach it may be possible to

assess the correctness of the sentiment polarities by

aggregating the polarities on review level to be

compared with the review ratings. However, to the

best knowledge of the authors, there is no general

applicable list of Dutch words available wherein

every word has a sentiment score attached.

There are lists of positive and negative words

available, positive, and negative Dutch lexicons,

however (Chen & Skiena, 2014). Using these lists,

adjectives can be roughly classified as positive or

negative. A neutral sentiment only gets assigned if an

adjective has conflicting polarities and on average is

neither positive nor negative. This is the approach that

we have adopted in our processing framework. Note

that we only make use of a binary sentiment

classification based on word use, and that there is not

any scoring involved. This makes that the

modification of sentiments, e.g., the use of the word

“very”, is not accounted for in the current

implementation.

4 RESULTS

For initial testing of the combination of Frog along

with the vocabulary method, the 30 testing sentences

used by Hernández-Rubio et al. (2019) were

translated into Dutch and processed by the algorithm.

The workings of the method in combination with

CoreNLP had been tested in an earlier stage and were

comparable to the original results of Hernández-

Rubio et al. (2019).

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

428

Table 4: Aspects extracted from simple sentences.

The hotel staff was

friendly.

('staff-3', 'friendly-5', '-',

True)

Het hotelpersoneel was

vriendelijk.

('hotelpersoneel',

'vriendelijk', '-', True)

Table 5: Aspects extracted from slightly more complicated

sentences.

The hotel had friendly

and efficient staff.

('staff-7', 'efficient-6', '-',

True),

('staff-7', 'friendly-4', '-',

True

)

Het hotel had

vriendelijk en efficiënt

ersoneel.

('efficient', 'vriendelijk', '-',

True), ('hotel', 'vriendelijk',

'-', True

)

Table 6: Aspects extracted from simple sentences (inverted

sentiments).

The hotel staff was not

friendly.

(‘staff-7’, ‘friendly-6’, ‘-‘,

False)

Het hotelpersoneel was

niet vriendelijk.

(‘hotelpersoneel’,

‘vriendelijk’, ‘-‘, False)

Table 7: Aspects extracted from slightly more complicated

sentences (inverted sentiments).

The hotel staff was

not friendly and

efficient.

('staff-3', 'efficient-8', '-',

True),

('staff-3', 'friendly-6', '-',

False)

Het hotelpersoneel

was niet vriendelijk

en efficient.

('efficient', 'vriendelijk', '-

', False), ('hotelpersoneel',

'vriendeli

k', '-', False)

Tables 4, 5, 6 and 7 show the results of the aspect

extraction. The algorithm for the Dutch language

seems competent on simple, straight to the point

sentences; see Table 4. Slightly more complicated

sentences, namely those that involve conjunctions,

work suboptimal; see Table 5. This is likely due to

Frog’s labeling of conjunctions structurally differing

from Stanford’s CoreNLP. The algorithm can pick up

inverted sentiments (“not friendly”) in simple

phrases; see Table 6. The Dutch word “efficiënt” had

to be transformed to “efficient” due to Frog

registering “ë” as a special character which heavily

influences how it processes the rest of the sentence;

see Table 7. This version is also classified as a noun

while it should be an adjective.

Figures 3, 4 and 5 show results of the aspect

extraction and sentiment classification for the online

review datasets for the retailers Bol.com, Agradi, and

ABOUT YOU that were scraped from Trustpilot.

Note that use of personal pronouns such as, e.g.,

“ik” seems to be prevalent in all figures; these aspects

give little insight without their contexts and can in

daily practice easily be filtered out so that they do not

occlude more interesting and more relevant aspects.

This observed personal pronoun problem is not

exclusive to the Dutch language however as

Hernández-Rubio et al. (2019) mention in their paper

that their implementation suffered the same issue.

Bol.com customers seem to be divided on whether

they are positive or negative on Bol’s customer

service (“service”, “klantenservice”); see Figure 3.

Delivery seems to be the aspect that is most often

used in Agradi reviews, for the majority being

positive; see Figure 4. Agradi customer service is

quite often neutral, signifying that diverging

sentiments are prevalent on this aspect in individual

reviews. When price is mentioned in this sample of

Agradi reviews it is only done so in a positive context.

Relatively speaking, delivery is mentioned quite

often in ABOUT YOU reviews; this occurs in

positive contexts, see Figure 5. The extracted aspects

seem, for the vast majority, to be used in a positive

context. This could be due to the sample, but it is clear

that ABOUT YOU, on average, is regarded to be the

most positive on the Trustpilot samples out of the

three companies. This could, perhaps in part, explain

the overwhelming general positivity of customers for

ABOUT YOU, but more data should be analyzed.

5 DISCUSSION & CONCLUSION

Although the vocabulary method described by

Hernández-Rubio et al. (2019) was originally

designed for the English language with the CoreNLP

toolkit, it works relatively well with the Frog toolkit

on individual Dutch sentences for extracting aspects.

The overall results of the vocabulary method on a

sample of online Dutch reviews gives some

interesting insights on the broad sentiments prevalent

in different aspects. These results demonstrate the

potential of this algorithm and show that it can extract

fine-grained aspects.

The outputs generated by the vocabulary method

are fine-grained aspects; according to Liu (2012, p.

81), the next step would be to group these aspects in

Exploring Corporate Reputation based on Sentiment Polarities That Are Related to Opinions in Dutch Online Reviews

429

“synonymous aspect categories”. The need for this is

evidenced in Figures 3 and 4 where two different

Dutch words are used to denote the category customer

service. It is recommended that further research is

done on this as Liu (2012, p. 82) calls it critical for

opinion analysis.

For practical reasons like data availability the

focus of this research and any demonstrations of

found solutions concern reviews on online retailers.

There is also the possibility of using alternative

sources of information instead of online reviews.

Dijkmans, Kerkhof, & Beukeboom (2015) for

example report that social media engagement of

airlines positively affects perceived corporate

reputation. Although more engaged customers are

also exposed to posts with a negative sentiment,

Dijkmans et al. (2015) find that the net effect remains

positive. Social media messages might very well be

an interesting addition or alternative to online

reviews. The solutions developed during this research

are likely to be applicable to online reviews

concerning other stakeholders and other textual data

like social media messages. Though the scope of this

specific research concerns the e-commerce

stakeholder its applications by no means must

necessarily be limited in the same manner.

Assessing corporate reputation on aspects

generated by this method, using for example the

model described by Fombrun et al. (2015), might very

well be possible but would require more research as

the current results are too fine-grained to adequately

base estimations on.

Overall, the vocabulary-based method seems to be

a very interesting technique that can extract aspects

from reviews: the fine-grained aspects give some

insights into what topics are prevalent in these

reviews and the sentiment can be assessed. More

research will have to be done in the future to draw

concrete conclusions in relation to corporate

reputation, but the demonstrations done in this

research do give a positive signal.

Figure 3: Top 10 most used aspects for Bol.com.

Figure 4: Top 10 most used aspects for Agradi.

Figure 5: Top 10 most used aspects for ABOUT YOU.

ACKNOWLEDGEMENTS

This paper has been inspired on the MSc master

project of Joël Vergeer who was involved via the

master Digital Driven Business at HvA. Thanks go to

Selmar Meents as well as several anonymous

reviewers for providing some useful suggestions to an

initial version of this manuscript. Rob Loke is

assistant professor data science at CMIHvA.

REFERENCES

Barnett, M. L., Jermier, J. M., & Lafferty, B. A. (2006).

Corporate reputation: The definitional landscape.

Corporate reputation review, 26-38.

Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent

dirichlet allocation. Journal of machine Learning

research, 993-1022.

Chen, D., & Manning, C. D. (2014). A Fast and Accurate

Dependency Parser Using Neural Networks.

Proceedings of the 2014 conference on empirical

methods in natural language processing (EMNLP),

740-750.

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

430

Chen, Y., & Skiena, S. (2014). Building sentiment lexicons

for all major languages. Proceedings of the 52nd

Annual Meeting of the Association for Computational

Linguistics, 383-389.

Chevalier, J. A., & Mayzlin, D. (2006). The effect of word

of mouth on sales: Online book reviews. Journal of

marketing research, 345-354.

Dijkmans, C., Kerkhof, P., & Beukeboom, C. J. (2015). A

stage to engage: Social media use and corporate

reputation. Tourism management, 58-67.

Fang, X., & Zhan, J. (2015). Sentiment analysis using

product review data. Journal of Big Data.

Farhadloo, M., & Rolland, E. (2013). Multi-class sentiment

analysis with clustering and score representation. 2013

IEEE 13th international conference on data mining

workshops (pp. 904-912).

Fombrun, C. J., Ponzi, L. J., & Newburry, W. (2015).

Stakeholder tracking and analysis: The RepTrak®

system for measuring corporate reputation. Corporate

Reputation Review, 3-24.

GfK. (2019, September 19). Quarterly number of online

purchases in the Netherlands from 1st quarter 2015 to

2nd quarter 2019 (in million units) [Graph]. Retrieved

April 8, 2020, from Statista: https://www.statista.com/

statistics/556353/quarterly-number-of-online-purchase

s-in-the-netherlands/

Gotsi, M., & Wilson, A. M. (2001). Corporate reputation:

seeking a definition. Corporate communications: An

international journal, 24-30.

Hall, R. (1993). A framework linking intangible resources

and capabiliites to sustainable competitive advantage.

Strategic management journal, 607-618.

Hastie, T., Tibshirani, R., & Friedman, J. (2009). The

elements of statistical learning. New York: Springer.

Hernández-Rubio, M., Cantador, I., & Bellogín, A. (2019).

A comparative analysis of recommender systems based

on item aspect opinions extracted from user reviews.

User Modeling and User-Adapted Interaction, 381-441.

Hotho, A., Nürnberger, A., & Paaß, G. (2005, May). A brief

survey of text mining. Ldv Forum, 19-62.

Hu, Y., Boyd-Graber, J., Satinoff, B., & Smith, A. (2014).

Interactive topic modeling. Machine learning, 423-469.

Korenius, T., Laurikkala, J., Järvelin, K., & Juhola, M.

(2004). Stemming and lemmatization in the clustering

of finnish text documents. Proceedings of the thirteenth

ACM international conference on Information and

knowledge management, 625-633.

Korfiatis, N., García-Bariocanal, E., & Sánchez-Alonso, S.

(2012). Evaluating content quality and helpfulness of

online product reviews: The interplay of review

helpfulness vs. review content. Electronic Commerce

Research and Applications, 205-217.

Lange, D., Lee, P. M., & Dai, Y. (2011). Organizational

reputation: A review. Journal of management, 153-184.

Liu, B. (2012). Sentiment analysis and opinion mining.

Morgan & Claypool Publishers.

Liu, Y., Liu, Z., Chua, T. S., & Sun, M. (2015). Topical

word embeddings. Twenty-Ninth AAAI Conference on

Artificial Intelligence.

Lovins, J. B. (1968). Development of a stemming

algorithm. Mechanical Translation and Computational

Linguistics, 22-31.

Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R.,

Bethard, S., & McClosky, D. (2014). The Stanford

CoreNLP natural language processing toolkit.

Proceedings of 52nd annual meeting of the association

for computational linguistics: system demonstrations,

55-60.

Nguyen, T. H., Shirai, K., & Velcin, J. (2015). Sentiment

analysis on social media for stock movement

prediction. Expert Systems with Applications, 9603-

9611.

Qiu, G., Liu, B., Bu, J., & Chen, C. (2011). Opinion word

expansion and target extraction through double

propagation. Computational linguistics, 9-27.

Roberts, P. W., & Dowling, G. R. (2002). Corporate

reputation and sustained superior financial

performance. Strategic management journal, 1077-

1093.

Statista. (2019). Online Reviews. Retrieved April 8, 2020,

from Statista: https://www.statista.com/study/50566/

online-reviews/

Trustpilot. (2020, August 8). Over Trustpilot. Retrieved

from Trustpilot: https://nl.trustpilot.com/about

Van den Bosch, A., Busser, B., Canisius, S., & Daelemans,

W. (2007). An efficient memory-based

morphosyntactic tagger and parser for Dutch. LOT

Occasional Series, 191-206.

Vashishtha, S., & Susan, S. (2019). Fuzzy rule based

unsupervised sentiment analysis from social media

posts. Expert Systems with Applications, 112834.

Wang, H., Lu, Y., & Zhai, C. (2010). Latent aspect rating

analysis on review text data: a rating regression

approach. Proceedings of the 16th ACM SIGKDD

international conference on Knowledge discovery and

data mining, (pp. 783-792).

Weng, J., Lim, E. P., Jiang, J., & He, Q. (2010).

Twitterrank: finding topic-sensitive influential

twitterers. Proceedings of the third ACM international

conference on Web search and data mining, (pp. 261-

270).

Wilbur, W. J., & Sirotkin, K. (1992). The automatic

identification of stop words. Journal of information

science, 45-55.

Xing, F., & Zhan, J. (2015). Sentiment analysis using

product review data. Journal of Big Data.

Yan, X., Guo, J., Lan, Y., & Cheng, X. (2013). A biterm

topic model for short texts. Proceedings of the 22nd

international conference on World Wide Web, (pp.

1445-1456).

Exploring Corporate Reputation based on Sentiment Polarities That Are Related to Opinions in Dutch Online Reviews

431