Brand Feedback and eWOM Sentiment on Twitter

(1)

FACULTY OF EDUCATION AND BUSINESS STUDIES

Department of Business and Economics Studies

Brand Feedback and eWOM Sentiment on Twitter

Evaluating the effect of company response strategies in online public conversations

Linnea Rudenius

2020

Student thesis, Bachelor degree, 15 HE Business Administration

Study Programme in Business Administration Supervisor: Jonas Molin and Lars-Johan Åge

Examiner: Maria Fregidou-Malama

(2)

FACULTY OF EDUCATION AND BUSINESS STUDIES

Department of Business and Economics Studies

(3)

i

Acknowledgements

I would like to thank my supervisors, Jonas Molin and Lars-Johan Åge for their

valuable feedback during the entire process of conducting this study. I am also grateful for the helpful comments provided by my examiner Maria Fregidou-Malama. Finally, I would like to thank Charlotta Einarsson for helping me distribute the link to a web page built for collecting Twitter labels, resulting in over 500 manually labelled tweets.

(4)

ii

Abstract

Title: Brand Feedback and eWOM Sentiment on Twitter: Evaluating the effect of company response strategies in online public conversations

Level: Student thesis, final assignment for Bachelor Degree in Business Administration Author: Linnea Rudenius

Supervisor: Jonas Molin and Lars-Johan Åge Examiner: Maria Fregidou-Malama

Date: 2020-06-08

Aim: The purpose of this study is to investigate how brand feedback response strategies impact eWOM sentiment in public conversations between consumers and businesses.

Method: The study has been conducted by collecting over one million tweets from Twitter and performing automated sentiment analysis of these tweets. After processing the tweets, 39 773 conversations were identified containing brand feedback. Two hypotheses were formulated regarding brand feedback and eWOM sentiment, and the sentiment scores of the 39 773 conversations used to test the hypotheses.

Result & Conclusions: There was, in general, a slight positive change in sentiment for conversations containing brand feedback. However, the results also showed that providing brand feedback does not always lead to a positive change in sentiment. The results showed no correlation between response speed and sentiment change in the conversations.

Contribution of the thesis: Previous studies within brand feedback and eWOM have to a large extent been experimental. This study contributes to the field by investigating brand feedback and eWOM from real data, using over 100 000 tweets in the analysis. The results of this study show that eWOM sentiment is not affected by response speed, which

contradicts one previous study. This study also contributes by showing that brand feedback only has a slight positive effect on eWOM sentiment, which means that an improved eWOM alone may not be reason enough for companies to invest in brand feedback.

Suggestions for future research: Further research could address limitations in the methodology, such as more sophisticated sentiment analysis or evaluating other social

(5)

iii media platforms. Research could also address the lack of correlation between response speed and sentiment change and investigate intervening variables or moderating effects.

Key words: eWOM, brand feedback, sentiment analysis, service recovery, social media, Twitter

(6)

iv

Sammanfattning

Titel: Brand Feedback and eWOM Sentiment on Twitter: Evaluating the effect of company response strategies in online public conversations

Nivå: Examensarbete på Grundnivå (kandidatexamen) i ämnet företagsekonomi Författare: Linnea Rudenius

Handledare: Jonas Molin och Lars-Johan Åge Examinator: Maria Fregidou-Malama

Datum: 2020-06-08

Syfte: Syftet med studien är att undersöka hur svarsstrategier vid brand feedback påverkar sentiment av eWOM i publika konversationer mellan konsumenter och företag.

Metod: Över en miljon tweets samlades in från Twitter och sedan applicerades en automatisk analys av sentiment på dessa tweets. Genom insamlade tweets identifierades 39 773 konversationer som innehöll brand feedback. Två hypoteser togs fram relaterade till brand feedback och sentiment av eWOM, och de värden som beräknats genom den

automatiska analysen användes för att testa hypoteserna på de 39 773 konversationerna.

Resultat & slutsats: Resultaten visade att det, generellt sett, fanns en viss skiftning av sentiment i positiv riktning för konversationer som innehöll brand feedback. Dock visade resultaten också att brand feedback inte alltid leder till en positiv förändring. Resultaten visade ingen korrelation mellan svarstid och förändring av sentiment i konversationerna.

Examensarbetets bidrag: Tidigare studier inom brand feedback och eWOM har till stor del varit experimentella. Denna studie bidrar till området genom att undersöka brand feedback och eWOM i verklig data, med en analys av över 100 000 tweets. Resultaten av studien visar att eWOM sentiment inte påverkas av svarstid, vilket står i kontrast mot vad en tidigare studie visat. Studien bidrar också genom att visa att brand feedback bara har en viss positiv effekt på eWOM sentiment, vilket innebär att en förbättrad eWOM i sig inte nödvändigtvis är en tillräcklig anledning för företag att investera i brand feedback.

Förslag till fortsatt forskning: Fortsatt forskning kan adressera begräsningar i

metodologin använd i studien; genom att använda mer sofistikerad teknik för att analysera sentiment eller genom att utvärdera andra sociala nätverk. Forskning kan också kunna

(7)

v adressera bristen på korrelation mellan svarstid och förändring av sentiment, och

undersöka mellanliggande variabler eller modererande effekter.

Nyckelord: eWOM, brand feedback, sentiment analysis, service recovery, social media, Twitter

(8)

vi

List of Abbreviations and Definitions

API: Application Programming Interface, a collection of definitions and protocols for interacting with a system or software.

AUC: Area under the curve, the area under the ROC curve. Used as an aggregate measure of performance for all possible classifier thresholds.

Brand Feedback: A company’s written response to consumer feedback online.

Compound Score: A normalized, weighted composite score of text sentiment produced by the VADER sentiment analysis tool.

eWOM: Electronic Word-Of-Mouth, any brand-related content posted publicly by potential, actual or former customers on the internet.

Response Speed: The difference in time from when the initial consumer feedback was posted until the company provided a response.

ROC curve: Receiver Operating Characteristic curve. A graphical plot that illustrates the performance of a binary classifier

Sensitivity: The ability of a binary classifier to correctly identify positive results.

Specificity: The ability of a binary classifier to correctly identify non-positive results.

VADER: Valence Aware Dictionary and sEntiment Reasoner, the sentiment analysis tool used in this study.

(11)

1

1. Introduction

The way people are communicating with businesses has changed drastically since the rise of internet and social media. Through social media, individuals can share information about brands with a vast amount of people by writing reviews, sharing user stories or by

mentioning the brand in a post. This kind of communication on the internet constitute what is commonly referred to as electronic word-of-mouth (eWOM), which is defined by Henning- Thurau, Gwinner, Walsh and Gremler (2004) as “any positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of people and institutions via the Internet” , and can thus be used to describe any brand-related content posted publicly on the internet. As consumer communication shifts from the traditional passive role to a more active one, eWOM becomes central to the

development of marketing strategies (King, Racherla & Bush, 2014).

The impacts of eWOM on important brand aspects such as brand image, trust and loyalty have proved to be significant in several different studies. Farzin and Fattahi (2018) investigated the effects of eWOM in Iran and found that an increase in positive eWOM regarding products and brands in social media networks enhanced consumer perceived brand image and purchase intentions. Seifert and Kwon (2019) investigated the impacts of eWOM on aspects such as perceived brand value and trust, and the results of their study

demonstrated that eWOM had a significant effect on brand trust. Similarly, Ladhari and Michaud (2015) found that comments generated on the Facebook network influenced the trust and attitude toward hotels. In their literature review, King et al. (2014) also state that several studies have confirmed the effects of eWOM on aspects such as brand trust and loyalty.

When evaluating the impacts of brand-related eWOM, many studies rely on analysis of text sentiment; that is analysis of whether the content expresses a positive, negative or neutral view about the brand. Cheung and Thadani (2012) explain in their literature review that many researchers have considered the persuasive effects of eWOM sentiment. They state that positively framed eWOM encourages people to adopt a service or product and that negatively framed eWOM discourages them by emphasizing weaknesses and problems. Lee, Rodgers and Kim (2009) evaluated the influence of sentiment in consumer product reviews and found

(12)

2 that positive reviews increased the attitude towards the brand, but that even a moderate

amount of negativity in a review would cancel out this positive effect. Seifert and Kwon (2019) specifically propose that sentiment influences the level of consumers’ brand value and find that consumers are more likely to communicate favorably about a brand if they

encounter other positive posts on social media networks.

1.1. Problem Description

Considering the impact of eWOM on consumers’ attitudes toward a brand, an interesting question for both researchers and practitioners is how eWOM should be managed. According to Bhandari and Rodgers (2018), brands and advertisers often participate in the eWOM communication by responding to negative product reviews. The authors refer to a company’s written response to consumer feedback as brand feedback, which is a term that will also be used in this study. The purpose of brand feedback is for the company to strengthen or uphold the promise they have made to satisfy the needs and desires of the consumers. Bhandari and Rodgers (2018) claim that there is a lack of understanding of the role that brands play in the eWOM process, and that a thorough knowledge of the effects of brand participation on aspects such as brand trust is missing. Esmark Jones, Stevens, Breazeale and Spaid (2018) similarly claim that empirical investigation into the impact of responses to eWOM is missing, and that this gap is highly relevant to academics and practitioners since the wrong kind of response can have a significant impact on attitude toward the company.

To address the research gap of brand feedback, Bhandari and Rodgers (2018) conducted an experiment using online reviews collected from various shopping sites and concluded that brand feedback had a significant positive effect on purchase intention via brand trust.

However, they also showed that brand feedback can potentially have negative effects.

Esmark Jones et al. (2018) conducted three different experiments that also considered product reviews and concluded that any response is better than no response in terms of resulting consumer attitudes toward the company. The role of brand feedback is closely related to online service recovery, which has been studied by Hogreve, Bilstein and Hoerner (2019) and Schaefers and Schamari (2016). The researchers argue that company responses do not just impact the complaining customer, but potentially also hundreds of virtual observers. Thus, online complaints and recovery attempts may have a significantly higher impact on consumer brand perceptions than traditional complaints.

(13)

3 Although the described studies address the brand feedback research gap, there are still several areas that could be considered in more depth. For instance, most studies within the area are conducted using experiments and few studies use large amounts of real data. Some studies (e.g. Bhandari & Rodgers, 2018) suggest brand feedback has an impact on purchase

intentions, and other studies suggest online recovery attempts have an impact on the attitude of observers (e.g. Hogreve et al., 2019). However, there are still several areas that can be investigated related to the performance of different brand feedback or recovery strategies. For instance, Fan and Niu (2016) evaluated the effect of recovery speed on customer satisfaction but conducted no in-depth analysis of different company response times.

To shed further light on the impact of brand feedback, this study aims to expand previous research by utilizing big data mining techniques on a dataset of over one million tweets from Twitter. I propose that automatic sentiment analysis of conversations containing brand feedback can be used to determine the success of the feedback, and that different response strategies can be evaluated through this approach. Drawing on insights from previous

research, eWOM sentiment is an important aspect when considering consumer perceptions of brand trust and brand value, which makes it an interesting aspect to consider in the context of brand feedback. The study will expand on existing research by considering aspects that previous research has investigated and mentioned to be important, but not evaluated in depth using large amounts of real data. The aspects that will be considered are; providing versus not providing brand feedback and the role of response speed in brand feedback. These aspects have previously been investigated by Esmark Jones et al. (2018), Bhandari and Rodgers (2018), Fan and Niu (2016) and Istanbulluoglu (2017), but all of the studies have been limited to experimental settings or small amounts of data and neither have considered the sentiment of eWOM produced at the end of the recovery attempt.

1.2. Purpose

The purpose of this study is to investigate how brand feedback response strategies impact eWOM sentiment in public conversations between consumers and businesses.

(14)

4

1.3. Research Questions

1. How does the presence of brand feedback influence eWOM sentiment in public conversations between consumers and businesses on Twitter?

2. How does the response speed of brand feedback influence eWOM sentiment in public conversations between consumers and businesses on Twitter?

1.4. Limitations

As previously stated, eWOM sentiment can be described as any brand-related content posted publicly on the internet. It may thus originate from any public social media network or webpage. However, analyzing sentiment from several different social media networks or webpages causes some complexity, both in terms of how data can be collected and how data should be analyzed when originating from different contexts. Because of this, this study will focus on the Twitter social media network only. The reason for choosing Twitter as a data source is that Twitter provides easy access to a large amount of data, and the microblogging format used on Twitter is suitable for automated sentiment analysis. Furthermore, researchers suggest that the open system and the broadcast nature of Twitter communication makes Twitter a useful source for exploring social media (Abney, Pelletier, Ford and Horky, 2017).

When conducting automated sentiment analysis, there are many different tools one may use to perform the task and depending on which tool one uses, the results may vary. One

implication of this is that the results of this study will largely depend on what tool is used to perform the automated analysis. To reason about the validity of the approach, one may ideally want to evaluate different sentiment analysis tools and compare the results they produce. However, using multiple sentiment analysis tools and comparing the results are considered out of scope for this study. Instead, only one tool will be selected and evaluated.

Although Twitter to a large extent consists of posts submitted by real users, there is also a vast number of bots posting content with malicious intent (Jamison, Broniatowski & Crouse Quinn, 2019). Such content may distort the collected data. However, finding potential bot posts and removing them has been considered out of scope for this study as it requires a substantial amount of additional work.

(15)

5 There are a few other, methodological limitations that are relevant for the results of this study. These will be discussed further in Chapter 3.

1.5. Disposition

The chapters in this study are organized as follows; first, a literature review of existing literature within eWOM and brand feedback will be presented. At the end of this chapter, I will present a summary and propose two hypotheses that will be used to answer the research questions. In Chapter 3, the methodology used in the study is presented, along with a

discussion about the quality of the study and the limitations of the methodology used.

Chapter 4 shows the results produced and Chapter 5 presents a discussion of the results in relation to the literature presented in Chapter 2. Finally, Chapter 6 presents the conclusions drawn in the study, along with contributions and suggestions for further research.

(16)

6

2. A Review of the Literature

There are several different areas of research that are relevant for the purpose of this study.

The first area of interest concerns the effects of eWOM on consumer brand perceptions.

Papers that investigate the effects of eWOM on consumer brand perceptions appear to have focused mainly on three aspects; brand image, attitude and trust. In a literature review by King et al. (2014), the authors suggest that eWOM significantly affects several individual- level outcomes such as consumers’ willingness to pay in a product category, consumer

engagement and levels of trust and loyalty towards a brand. Looking at individual studies that have been conducted within the area, most focus on some particular variable and its influence on brand image, attitude or trust. This literature review will begin by presenting each one of these variables and the related research separately.

Many of the articles presented in this literature review make use of sentiment analysis in their studies. Text sentiment can be defined as whether the content expresses a positive, negative or neutral view about a brand. Although positive and negative eWOM will be mentioned in several places in this section, a separate headline will also be presented considering eWOM sentiment in the context of brand evaluations.

Although brands are affected by eWOM, the interactions between consumers and brands online are not unidirectional. Just as consumers may post about companies online, companies can respond and interact with consumers. Brands are thus not only affected by eWOM but can also play a key part in generating it. As described by Bhandari and Rodgers (2018), brands and advertisers today often participate in the eWOM communication process by responding to negative product reviews. The authors call this process brand feedback, which is also one of the main topics in this study. In the literature review, literature regarding both brand feedback and the closely related topic of online service recovery will be presented.

Results from previous studies within all topics will then be used to form hypotheses that can help answer the proposed research questions.

2.1. eWOM and Brand Image

Brand image is described by Reza Jalilvand and Samiei (2012) as the comprised attributes and benefits associated with a brand that make the brand distinctive, and so it distinguishes

(17)

7 the firm’s offer from competition. Similarly, Aaker (1996, p. 71) describes brand image as

“how the brand is perceived”. Looking at papers that investigate the relationship between eWOM and brand image generally seem to agree that eWOM may have a significant effect on brand image.

In a study by Charo et al. (2015), the authors evaluated the impact of eWOM on brand image and purchase intention by sending a questionnaire to members of a specific group on

Facebook. The results of the questionnaire showed that 47.5% of respondents thought online reviews had a moderate impact on brand image, and 23% thought it had a major impact. Only 8% of the respondents thought the reviews had no impact on brand image. From analysis of the results, the authors concluded that the impact of eWOM on brand image is significant. On the same topic, Krishnaumurthy and Kumar (2018) conducted an experiment to evaluate the impact of eWOM on brand expectations and image. The authors argued that understanding customer expectations of a brand is a starting point to understanding their concept of the brand image. In the study, respondents were presented with a scenario of an impending purchase and then got to read an optional amount of eWOM before expressing their expectations of the brand and product. The results of the study showed that customers that spend more time taking part of eWOM formed a higher expectation and a better image of the brand in their minds.

Farzin and Fattahi (2018) argued that eWOM communications might have a strong influence on brand image and conducted a study where a questionnaire was distributed to 400 business administration students in Iran. One of their hypotheses was that the eWOM behavior of consumers that are members of social media network sites has a direct, positive effect on brand image. The hypothesis was based on previous studies suggesting that customers build a positive image about a brand when they notice positive eWOM that is associated with the brand. The results of the study supported the hypothesis and indicated that consumer eWOM behavior had a positive and significant impact on brand image. Similarly, in a study by Reza Jalilvand and Samiei (2012) the authors distributed a questionnaire to evaluate the impact of eWOM on brand image in the automotive industry in Iran. The conclusion of this study also indicated that eWOM has a considerable effect on brand image.

(18)

8

2.2. eWOM and Brand Attitude

Brand attitude is described by Wu and Wang (2011) as the continuous preference or loathing tendency of the consumer towards a certain brand. Furthermore, they describe that brand attitude consists of the overall evaluation the consumer has towards the brand, and that this evaluation in turn forms the base of the brand image. Just as with brand image, studies conducted within the area of eWOM and brand attitude seem to indicate that eWOM can have a significant influence on brand attitude.

In a study by Kudeshia and Kumar (2017) the authors examined the effect of user-generated positive eWOM on brand attitude and purchase intentions. The study was conducted through a questionnaire sent to individuals who were users of Facebook, and a total of 311 valid responses were collected. The results of the study indicated a direct positive relationship between positive eWOM, brand attitude and purchase intention. Wu and Wang (2011) also considered brand attitude but focused on the effects of eWOM source credibility on

consumers’ attitudes toward a brand. The authors concluded that high credibility towards the eWOM message source increases perceived quality and improves brand attitude. The study was conducted through questionnaires to a sample of 60 people and considered two types of products; electronic foods and fast-moving consumer goods.

In the studies by Kudeshia and Kumar (2017) and Wu and Wang (2011), the results indicated that eWOM would have an effect on brand attitude. However, both of these studies were conducted using questionnaires for a small sample size. Thus, the generalizability of the results may be questioned. Also, both of these studies consider positive eWOM only and does not take into consideration what effects negative eWOM may have on brand attitude.

Continuing on the topic of brand attitude, Park and Jeon (2018) conducted a study to evaluate the effects of eWOM on brand attitude change with a focus on differences across cultures.

Their study examined the effect of eWOM sequence (presentation order of positive/negative posts) on changes in brand attitude from a cross-cultural perspective. South Korea and USA were selected for a comparative analysis, and a controlled experiment was conducted on segregated experimental groups. The result indicated that no significant different existed in brand attitude changes depending on the presentation order for the South Korea group, but that the American group showed a significant attitude change when first seeing negative posts

(19)

9 compared to when first seeing positive posts. The order of positive and negative reviews was also examined by Coker (2012), who conducted an experiment to evaluate the effect of presentation order on consumer attitudes and judgements. In the experiment, online hotel reviews ordered from positive to negative resulted in a more positive evaluation than if the same reviews were ordered from negative to positive. The author argued that these results indicate that positive affect holds a stronger effect on brand attitude than negative affect. That is, participants were less successful at adjusting their attitudes from a positive to negative valence than they were adjusting their attitudes from a negative to positive valence.

When looking at the order of positive and negative eWOM, it appears as though consumers more easily changes their attitude from negative to positive than positive to negative.

However, According to Park and Jeon (2018), this would not be true for eastern cultures.

Although the presentation order may be important when consumers form their attitude toward a brand, none of the studies investigating presentation order mention the aspect of time. That is, whether the changes in attitude from reading positive and negative reviews depends on at what time they were written. It is not mentioned whether the presentation of reviews

represented a chronological order, or whether respondents believed reviews in the top to be newer or older than reviews at the bottom.

2.3. eWOM and Brand Trust

Ladhari and Michaud (2015) argued that trust is one of the most influential factors in online sales, and that trust refers to a willingness to rely on an exchange partner. Several studies seem to confirm the hypothesis that eWOM has a significant impact on brand trust. More specifically, Seifert and Kwon (2019), Awad and Ragowsky (2008), Baker and Kim (2019) and Ladhari and Michaud (2015) contribute to the discussion, and they all seem to indicate that eWOM significantly affects brand trust.

Seifert and Kwon (2019) evaluated how sentiment in eWOM affects the level of consumers’

brand engagement and brand trust. The authors argued that the positive or negative sentiment expressed in brand-related eWOM may play a critical role in altering consumers’ trust for the brand. One of the hypotheses of the study was thus that consumers’ brand trust change is more positive when exposed to a positive social network-based eWOM compared to negative eWOM. The results of the study supported this hypothesis, and Seifert and Kwon (2019)

(20)

10 drew the conclusion that positive eWOM leads to a more positive trust change, and that negative eWOM leads to a negative trust change. The authors also stated that consumers are more likely to communicate with and about the brand in a more favorable light when they encounter positive brand eWOM on social media. Negative eWOM, however, might lead to avoidance behavior.

The effects of eWOM on brand trust has also been studied by Awad and Ragowsky (2008), but with the approach of comparing differences between genders. The authors hypothesized that the effect of eWOM quality on trust of an online vendor will be significantly different in magnitude for men and women. The results of a survey with over 1500 responses indicated that the hypothesis was supported, and that there was a significant difference in establishing trust online for men versus women. From evaluating the effects of an online WOM system on a retailer site, the results showed that the presence of such a system contributes to more perceived trust for men than for women.

Continuing on the concept of brand trust, Baker and Kim (2019) took a different approach and examined the effect of exaggerated eWOM. Their study was conducted in the context of online tourism reviews, and the authors defined exaggerated eWOM as “intentionally

distorted communications by customers that misrepresent their consumption experiences”.

The study was conducted by a mixed methods approach, where the authors first conducted two qualitative studies using critical incident technique, and then a quantitative study with an experimental design. The results of the study revealed that positive comments about a firms’

products and services can bring positive attitudinal changes in consumers’ perceptions and that negative comments may cause negative changes. Ladhari and Michaud (2015) also conducted a study within the hospitality industry, but with the purpose of examining the effect of comments generated on Facebook on the choice of a hotel. The authors specifically looked at the influence of comments on booking intentions, the trust and attitude toward the hotel as well as the perception of its website. By conducting an experiment using positive and negative Facebook comments, 800 university students answering a survey confirmed the hypotheses that comments generated on the Facebook network influence both attitude and trust toward a hotel. The more positive the comments about the hotel were, the more positive would the attitude be and the higher the trust.

(21)

11

2.4. eWOM Sentiment and Brand Evaluations

While some studies focus on outcomes, others focus on different eWOM characteristics and their effect on consumer evaluations. For instance, many researchers base their studies on investigations of eWOM sentiment. Cheung and Thadani (2012) state in their literature review that research in the area of consumer behavior has shown that consumers pay more attention to negative information than positive information, and that people tend to weigh negative information higher than positive information during evaluation. However, Wang, Cunningham and Eastin (2015) argue that current research is conflicted and that some research claim that consumers perceive positive messages to be more persuasive than

negative ones, while other studies show that negative information is more attention grabbing in general.

In a study by Wang et al. (2015), the role of different message content characteristics on attitudinal and behavioral responses to online consumer reviews was examined. The results from the study suggested that message sentiment of online consumer reviews has a

persuasive effect on both purchase intentions and consumers’ attitudes towards the review and the product. The results also indicated that positive reviews have a greater impact than negative reviews. Wang et al. (2015) argued that these findings contradict previous research, as previous research suggests that negative WOM messages have a stronger impact on brand evaluation than positive ones.

Related to the concept of eWOM sentiment is the existence of emotions in eWOM. Through an experiment conducted by Kim and Gupta (2012), the authors investigated how consumers interpret emotional expressions in online user reviews and the subsequent impact on product evaluations. The authors defined emotions as high intensity, valanced feeling states that are associated with the product of interest. The findings of the study suggested that negative emotional expressions in a negative review tend to decrease the perceived information value of the review and make consumers’ product evaluations less negative. The authors argued that the reason for this might be that consumers attribute the negative emotions to the reviewer’s irrational dispositions. However, the authors also stated that positive emotional expressions in a positive review do not influence consumers’ product evaluations. They suggested that consumers are likely to judge emotions in a review as inappropriate when negative and acceptable when positive.

(22)

12 Although many studies investigate eWOM sentiment, few of them consider neutral eWOM.

According to Tang, Fang and Wang (2014), studies have not examined the effects of neutral user generated content systematically but instead assume that it has no effect on product sales. The authors examined how the existence of neutral content influences consumers to interpret and perceive positive and negative content. In the study, Tang et al. (2014) showed that the existence of neutral content may, depending on its characteristics, either amplify or attenuate both positive and negative effects of other content.

When it comes to the effect of eWOM sentiment on product- and brand evaluations, researchers seem to agree that negative eWOM contributes to negative evaluations and positive eWOM contributes to positive evaluations. Wang et al. (2015) also suggest that positive eWOM has a stronger effect on product and brand evaluations than negative eWOM, and that this contradicts previous research. However, to the extent of this literature review, no research has been found where the results indicate that negative eWOM should have a greater impact on brand evaluations than positive eWOM. Kim and Gupta (2012) focused on

emotions rather than sentiment, although their definition of emotions suggest emotional expressions are eWOM with a strong negative or positive sentiment. Tang et al. (2014) gave additional insight into the sentiment discussion by stating that neutral eWOM may also have an impact on evaluations.

2.5. Brand feedback and Service Recovery on Social Media

The behavior of responding to negative product reviews online is referred to by Bhandari and Rodgers (2018) as brand feedback and conceptualized as “a brands attempt to reinforce the validity of the brand promise and reinstate potential lost trust resulting from negative eWOM”. Bhandari and Rodgers (2018) claimed that the trend for businesses to provide feedback to customer complaints online has increased substantially in recent years, but that scholarship has not adequately kept pace to understand the role of the brand in the eWOM process. Similarly, Esmark Jones et al. (2018) stated that empirical investigation into the impact of responses to eWOM seems to be missing. Closely related to what Bhandari and Rodgers (2018) refers to as brand feedback is service recovery, which is defined by Wilson, Zeithaml, Bitner and Gremler (2016) as the actions taken by an organization in response to a

(23)

13 service failure. Literature concerning brand feedback relates heavily to literature concerning service recovery, and thus they will be presented together here.

2.5.1. Effects of Brand Feedback on Individual Consumer Outcomes

In 2014, Sparks and Bradley (2014) conducted a study on managerial responses to online reviews of hotels. The authors drew on service marketing and justice literature to develop a conceptual scheme to capture and organize likely recovery strategies. They called this the

“Triple A” typology, which comprises three main components; acknowledgement, accounts and actions. The typology was tested by interviewing managers about the response strategies they used and by conducting a qualitative analysis of 150 online management responses.

Although the authors defined a framework of response strategies for online reviews, they did not test whether these strategies had any effect on customers. In the article by Sparks and Bradley (2014), the authors mention that future research should investigate whether customer perceptions of hotel brands will be more positive if companies respond to complaining reviews. This question is addressed by Esmark Jones et al. (2018) and Bhandari and Rodgers (2018), although not within the hospitality industry.

To address the gap of knowledge on the effects of brand feedback on variables such as brand trust and purchase intentions, Bhandari and Rodgers (2018) conducted an experiment using online reviews gathered from shopping sites. In the experiment, 453 undergraduate students were randomly assigned to one of eight scenarios where they got to read online product reviews and then responded to questions about the investigated variables. The results of the study showed that the presence of brand feedback to a negative review had an indirect positive effect through an increased brand trust, and a direct negative effect on consumers’

purchase intention. The authors suggested that this result indicates that the effects of brand feedback may not be as straightforward as prior research suggests.

Esmark Jones et al. (2018) investigated the impact of responses to negative eWOM and how those responses affected readers’ evaluations of the company. However, in contrast to Bhandari and Rodgers (2018), Esmark Jones et al. (2018) investigated different types of responses; both responses from other consumers and from the company itself. With their study, the authors tried to answer the question of what effect the source of a company- positive response to negative eWOM have on consumer outcomes as well as how situational

(24)

14 factors such as magnitude and sequence of reviews influence the findings. The authors also addressed the question of how an attempt to attribute a failure to the original poster impact consumer outcomes. To answer the questions, Esmark Jones et al. (2018) conducted three experiments where participants got to read the description and reviews posted for a product.

The results of the experiment showed that consumer responses to negative eWOM leads to a higher product satisfaction, and that any response is better than no response from the

company.

The studies by Bhandari and Rodgers (2018) and Esmark Jones et al. (2018) are similar in the way that they were both conducted using experiments to evaluate the effect of management responses on different consumer outcomes. Although they focus on somewhat different variables, their findings both highlight the complexity of management responses. As

Bhandari and Rodgers (2018) stated, brand feedback may have both a positive and a negative effect on purchase intentions which makes it hard to predict the results of providing brand feedback. Esmark Jones et al. (2018) on the other hand considered feedback from other consumers as well and showed that although brand feedback may have a positive effect, positive responses from other consumers are the most effective in terms of customer satisfaction. The authors stated that any response from the company may be better than no response, but negative responses from other consumers may lead to a lower customer satisfaction. Since consumer responses cannot be controlled by the company, this is another aspect that needs to be considered when forming a strategy for handling online complaints.

To investigate the influence of management responses on customer satisfaction, Gu and Ye (2014) conducted a quantitative analysis from data retrieved by scraping an online Chinese travel agency for customer reviews and management responses. When performing the analysis, the authors looked for customers who had posted a review for the same hotel more than once and compared how their level of satisfaction changed depending on whether they had received a response to their previous review or not. The results indicated that the influence of online management responses varied greatly across customers, and that

providing management responses can increase the future satisfactions of customers who are unsatisfied but has limited influence on the satisfaction of other customers. Litvin and Hoffman (2012) also investigated management responses to online hotel reviews but did not express that the influence had such a great variation. Instead, they found that that managing

(25)

15 negative reviews can significantly affect consumer attitudes. Similarly, Liu, Jayawardhena, Dibb and Ranaweera (2019) found that a hotel responding to negative online reviews will lead to consumers having a higher future eWOM continuance intentions than when offering no response.

2.5.2. Virtual observers

Gu and Ye (2014) addressed the fact that the public nature of an online recovery effort requires the service providers to consider not only how their responses influence the complaining customer, but also how they influence consumers who observe the complaint and response. This is a topic also addressed by Hogreve, Bilstein and Hoerner (2019), who claimed that virtual observers of an online recovery attempt can be significant sources of WOM. The authors stated that on average, 835 observers read each complaint posted by a dissatisfied customer. By conducting four different experiments the authors found that a transparent service recovery in which virtual observers can follow the full dialogue between the complainant and the service provider enhances WOM and purchase intentions. However, the authors acknowledge that a transparent service may have negative effects if the recovery is unsuccessful.

Just as Hogreve et al. (2019), Schaefers and Schamari (2016) found that the presence of virtual observers enhances the positive effects of recovery success in social media compared to traditional recovery processes. However, Schaefers and Schamari (2016) does not

acknowledge that an unsuccessful service recovery would lead to additional negative effects because of the virtual observers. Instead, the authors argue that the negative consequences of unsuccessful recovery remain unaffected. Due to these findings, Schaefers and Schamari (2016) claim that there is more to gain from providing a transparent service recovery via social media than there is to lose, as the positive outcomes are enhanced while the negative outcomes remain unaffected.

2.5.3. Response Strategies

While some studies, such as the ones by Bhandari and Rodgers (2018), Gu and Ye (2014) and Litvin and Hoffman (2012), investigate the presence of brand feedback and its effects on customer outcomes, other studies focus on specific types of feedback. For instance, Johnen and Schnittka (2019) evaluated whether apologies are better than defensive strategies. The

(26)

16 authors posed that observers’ processing of complaints and the effectiveness of a response strategy depends on the benefits those observers seek when browsing. The results of their study indicated that defensive strategies may be preferred in some settings, while

accommodative responses are preferred in others. These results are somewhat contradictory to the results of a study by Weitzl and Hutzinger (2017), who found that accommodative responses yield significantly more favorable brand-related outcomes within bystanders than defensive responses.

Feedback types were also investigated by Abney et al. (2017), who considered the difference between adaptive and generic brand feedback in the context of service recovery. The authors hypothesized that adaptive responses would lead to a higher satisfaction, repurchase intention and tolerance for failure. To test their hypotheses, the authors conducted an experiment to capture customers’ evaluations of various responses provided to customer complaints on Twitter. The results of the experiment showed that consumers perceive a positive difference in highly directed responses versus other less personalized messages.

Fan and Niu (2016) also evaluated the effectiveness of different service recovery strategies using social media networks. By conducting a manual analysis of 347 conversations on Twitter, the authors investigated several different aspects such as customer emotions,

satisfaction and recovery speed. In approximately half of the conversations, the authors found that the customer felt better after the exchange with the company agents, while the other half still felt frustrated or irritated. However, in 75 percent of the cases the customer expressed satisfaction at the end of the conversation. As for the role of recovery speed, which was explained as the time period from the initial complaining tweet to the end of the exchange, the authors found that this had no direct effect on customer satisfaction. Furthermore, an analysis between responses and customer satisfaction showed that providing further

directions resulted in less satisfied customers, and that responses not asking the customer to take further initiatives resulted in higher customer satisfaction.

Similar to Fan and Niu (2016), Istanbulluoglu (2017) investigated the impact of recovery speed on customer satisfaction for consumers who had complained on social media. The author collected data through a survey sent to consumers who had complained and received a response from a company on Twitter or Facebook. In total, 422 responses were used for

(27)

17 analysis and the author concluded, in contradiction to the conclusions of Fan and Niu, that consumers’ satisfaction with complaint handling on social media is related to the response time. On Twitter, Istanbulluoglu (2017) claim that consumers expect companies to reply to their complaints within 1-3 hours.

2.6. Summary and Hypotheses

Several researchers state that eWOM has a significant impact on brand image (e.g. Charo et al., 2015; Farzin & Fattahi, 2018), brand attitude (e.g. Kudeshia & Kumar, 2017; Wu &

Wang, 2011) and brand trust (e.g. Seifert & Kwon, 2019; Awad & Ragowsky, 2008). Many of these studies rely on eWOM sentiment, where findings indicate that positive eWOM relate to increased measures of the variables and negative eWOM relate to decreased measures.

Furthermore, eWOM sentiment has an impact on brand evaluations where positive eWOM leads to more positive evaluations and negative eWOM leads to more negative evaluations (e.g. Wang et al., 2015; Liu, Hu & Xu, 2018). According to Tang et al. (2014) neutral eWOM may also attenuate or amplify positive and negative effects of other content.

Looking at studies conducted within the area of brand feedback, research indicates that brand feedback can have an effect on individual level outcomes such as customer satisfaction and attitude (e. g. Bhandari & Rodgers, 2018; Esmark Jones et al, 2018; Litvin & Hoffman, 2012). However, brand feedback and online recovery attempts also influence the perceptions of virtual observers (e. g. Hogreve et al., 2019; Schaefers & Schamari, 2016). Thus,

successful brand feedback may contribute to a positive consumer attitude not only for the complaining customer, but also for the observers. Different response strategies have been examined by, for instance, Abney et al. (2017) and Weitzl and Hutzinger (2017). Generally, researchers seem to agree that accommodative responses are more successful than defensive or generic ones. However, the studies on response strategies are scarce and some studies contain contradictory results, which means drawing conclusions based on these are hard.

(28)

18 Table 1: A summary of the literature and takeaways most relevant for this study.

Reference Research Design Main takeaway

Charo et al. (2015) Survey The impact of eWOM on brand image is significant.

Krishnaumurthy and

Kumar (2018) Experimental Customers that spend more time taking part of eWOM formed a better brand image.

Kudeshia and Kumar (2017)

Survey There is a direct positive relationship between positive eWOM and brand attitude.

Coker (2012) Experimental Participants were less successful at adjusting their attitudes from a positive to negative valence.

Seifert and Kwon (2019)

Survey Positive/negative eWOM leads to a more positive/negative trust change

Cheung and Thadani (2012)

Literature Review Consumers pay more attention to negative information than positive information Wang, Cunningham

and Eastin (2015) Experimental Positive reviews have a greater impact than negative reviews.

Bhandari and Rodgers (2018)

Experimental The presence of brand feedback to a negative review had an indirect positive effect through an increased brand trust, and a direct negative effect on consumers’

purchase intention.

Esmark Jones et al.

(2018)

Experimental Any response is better than no response from the company. Negative responses from other consumers may lead to a lower customer satisfaction.

Litvin and Hoffman

(2012) Experimental Managing negative reviews can

significantly affect consumer attitudes.

Liu, Jayawardhena, Dibb and Ranaweera (2019)

Experimental Responding to reviews lead to consumers having a higher future eWOM continuance intentions than when offering no response.

Hogreve, Bilstein and

Hoerner (2019) Experimental A transparent service recovery in which virtual enhances WOM and purchase intentions. A transparent service may also have negative effects.

Schaefers and Schamari (2016)

Experimental There is more to gain from providing a transparent service recovery via social media than there is to lose.

Abney, Pelletier, Ford and Horky (2017)

Experimental Consumers perceive a positive difference in highly directed responses versus other less personalized messages.

Fan and Niu (2016) Qualitative analysis of data from Twitter

Recovery speed had no direct effect on customer satisfaction.

Istanbulluoglu (2017) Survey Consumers’ satisfaction with complaint handling on social media is related to the response time.

(29)

19 A summary of the literature and takeaways most relevant for this study can be found in Table 1. A more extensive summary is included in Appendix A. There are a few gaps that can be identified from this literature review. Looking at the methodology used in studies relating to eWOM, a majority of them (e.g. Park & Jeon, 2018; Kim & Gupta, 2012; Bhandari &

Rodgers, 2018; Schaefers & Schamari, 2016) have been conducted as experiments. A few studies have been conducted using real social media data (Tang et al., 2014; Gu & Ye, 2014;

Johnen & Schnittka, 2019), but no study has been identified that conducts a quantitative analysis using a large amount of data from Twitter. This is the first gap that this study aims to fill. Looking at the topics of the different studies, they concern the impact of eWOM on different brand aspects as well as the effect of brand feedback and online service recovery on consumer level outcomes. Studies have investigated the effects of brand feedback and online service recovery on aspects such as customer satisfaction (e.g. Esmark Jones et al., 2018), purchase intentions (e.g. Bhandari & Rodgers, 2018) and the attitude of observers (e.g.

Hogreve et al., 2019). However, they have not investigated whether brand feedback has any effect on subsequently generated eWOM, nor in depth investigated the response speed. This is the second research gap that this study aims to fill.

The research questions asked in this study both relate to eWOM sentiment in public conversations between consumers and businesses on Twitter, and how it is influenced by brand feedback. To answer these questions, the influence of brand feedback on eWOM sentiment needs to be measured somehow. Given that this study will be conducted using large amounts of data, the measure needs to be quantifiable and it should be possible to derive without manual labor. I argue that if brand feedback has any influence on eWOM sentiment, the sentiment should differ from the beginning of the conversation, before brand feedback, to the end of the conversation, after brand feedback. Thus, the influence may be measured by considering the change in sentiment of the eWOM produced at the end of the conversation compared to the beginning of the conversation.

Considering research question 1, several researchers such as Bhandari and Rodgers (2018), Esmark Jones et al. (2018) and Litvin and Hoffman (2012) have investigated the presence of brand feedback and seem to agree that providing brand feedback is better than not providing any feedback. To investigate the influence of the presence brand feedback on eWOM

(30)

20 sentiment and answer research question 1, I draw from these conclusions and form the

following hypotheses using the previously described definition of influence.

H10: When brand feedback is present in a conversation, there will be no change in eWOM sentiment from the beginning to the end of the conversation

H1A: When brand feedback is present in a conversation, there will be a significant, positive, change of eWOM sentiment from the beginning to the end of the conversation

Research question 2 regards response speed; the difference in time from when an initial consumer feedback tweet was posted until the company provided a response. Although both Fan and Niu (2016) and Istanbulluoglu (2017) evaluated the effect of recovery speed on customer satisfaction, they only considered specific ranges of recovery speed and the conclusions drawn were contradictory. No in-depth analysis was made in either of these studies to investigate the correlation between recovery speed and customer satisfaction. To form the hypotheses that should answer research question 2, I draw on the conclusions of Istanbulluoglu (2017), who stated that a quick response time leads to a higher satisfaction, and theory of procedural justice, which states that customer expect fairness in terms of timeliness of the complaint process and that the complaint is handled quickly (Wilson et al., 2016). Thus, it is expected that the change in sentiment from the beginning to the end of the conversation will correlate with the time from the initial consumer post until the first

company response.

H20: The company response speed is not correlated with change of eWOM sentiment from the beginning to the end of the conversation.

H2A: There is a significant, negative, correlation between the time until the first company response and change of eWOM sentiment from the beginning to the end of the conversation.

(31)

21

3. Methodology

The purpose of this study is to investigate how brand feedback strategies impact eWOM sentiment in conversations between consumers and businesses, and I initially posted two research questions related to this purpose. In Chapter 2, these where then formulated as two different hypotheses. To guide choices and reasoning about the methodological approach taken in this study, the research philosophy of positivism has mainly been used. Positivism is, as described by Saunders, Lewis and Thornhill (2009, p. 135), related to the philosophical stance of the natural scientist and entails working with an observable social reality to produce law-like generalizations. In this study, the observable social reality can be considered the conversations of tweets fetched from Twitter, and the generalizations produced should lead to answers to the two research questions.

As by Saunders et al. (2009, p. 135) describe it, positivism focuses on strictly scientific empiricist methods designed to yield pure data and facts uninfluenced by human interpretation or bias. Furthermore, Aliyu, Bello, Kasim and Martin (2014) argue that

positivism can be regarded as a research strategy and approach that is rooted on the principle that truth and reality is free and independent of the viewer and observer. Sanders et al. (2009, p. 135) do, however, state that even a researcher adopting a positive stance exercises choice in issues related to the study. Thus, the research will in some way always be influenced by the researcher. Following the ideals of positivism, one focus for this study has been to produce results that can be reproduced and are based on methods that are not dependent on human intervention or interpretation. Through the use of statistical methods and

programmatic analysis of texts, my intention is to produce results that are completely independent of my own interpretation. However, for analyzing sentiment the human

interpretation is key to evaluating the performance of an automated tool. Although the results of the hypothesis testing may be independent from my own interpretation, I cannot reason about these results without also comparing the tool used to human interpretation. Thus, excluding human interpretation entirely from the research methodology has not been considered desirable.

Looking at typical research methods used within positivism, Saunders et al. (2009, p. 136) describe these as highly structured methods using large samples and quantitative analysis.

(32)

22 This too is similar to the approach that has been used in this study. The methodology used for testing the hypotheses in this study can be divided into three stages; data collection, sentiment analysis and hypothesis testing. In this chapter, these three stages will first be described separately. Then, the quality and limitations of the study will be discussed in terms of validity and reliability. Finally, a section on the ethical implications of the methodology will be presented as well as a brief discussion about replicability.

3. 1. Data Collection

The data used in this study has been collected using Twitter’s API (Application Programming Interface). An API can be described as a collection of definitions and protocols for interacting with a system or software. Thus, an API may serve as a gateway between two different services or products, allowing them to interact without knowledge of each other’s internal implementation. In the case of Twitter, the company offers a set of different APIs for different purposes and types of consumers (Twitter, 2020a). In this study, the API endpoint for fetching individual tweets and the Twitter Standard Search API has been used. The Twitter Standard Search API allows consumers to search and collect tweets posted within the last seven days. The APIs used are part of the public set of APIs from Twitter meaning that they can be used by anyone with a Twitter developer account. To comply with the terms and conditions for using the APIs, only aggregated statistics about the tweets retrieved will be presented in this study, and not the data itself.

To achieve the purpose of this study, data was needed that represented conversations between consumers and businesses on Twitter. The initial collection of this data was conducted using the Twitter Standard Search API, which is the most convenient API for fetching multiple tweets that match a specific criterion. The API does, however, come with a couple of limitations. First of all, it is only possible to retrieve tweets posted within the last 7 days.

Furthermore, there is no way of explicitly retrieving lists of conversations or specifying that the tweets should be posted on accounts dealing with customer support. Instead, it is possible to search for specific tweet identifiers, or tweets posted to or from a specific account. In order to retrieve conversations between consumers and businesses, a list of Twitter accounts

dealing with customer support thus had to be produced.

(33)

23 To find customer support Twitter accounts, I first went to Twitter and searched for terms that are commonly associated with customer support accounts; such as “support”, “help”, “ask”

and “cares”. For the purpose of making the process of data collection as easy as possible, I tried to find Twitter accounts that contained a large number of consumer tweets and replies.

With this in mind, I conducted an additional search with terms including the names of the world’s largest companies by revenue. In total, a list of 120 different accounts were identified that deal with customer support on Twitter. Figure 1 shows the distribution of industries the identified accounts belonged to, and Figure 2 shows the size of the companies by revenue.

Figure 1: A bar chart showing industry on the x-axis and number of accounts on the y-axis. (Own construction: The industry was selected based on the main business of the company)

Figure 2: A Pie chart showing the size of companies based on revenue in USD. (Own construction:

Revenue was found in reports from the companies. Revenue could not be found for all companies).

0 2 4 6 8 10 12 14 16

Computer Software Conglomerate

Courier

Consumer Electronics Social Media

Transportation Streaming Media

Video Games Cloud Computing

Airline Hospitality

Telecommunications Retail

Publishing Financial Services

Recruitment Automotive

Restaurant Logistics

Web Hosting

Distribution of industries for Twitter accounts

$1 Million-$1 Billion

$1 Billion-$5 Billion

$5 Billion-$20 Bilion 25%

$20 Billion-$100 Billion 28%

$100 Billion-$500 Billion 13%

Size of companies based on revenue in USD

$1 Million-$1 Billion $1 Billion-$5 Billion $5 Billion-$20 Bilion

$20 Billion-$100 Billion $100 Billion-$500 Billion

(34)

24 Once the list of accounts was acquired, I conducted two searches for each account using the Twitter Search API; one fetching tweets posted by the account, and one fetching tweets posted to the account. The query also specified that all tweets should be in English. In order to get a large dataset, the process was repeated four weeks in a row. Once the initial list of tweets had been produced, a complementary search was conducted to make sure complete conversations had been retrieved. This complementary search made use of the endpoint from the Twitter API that allows lookup of any tweet by its id, and recursively called this endpoint using the in_reply_to_status_id field (described in Table 2) from the initially collected tweets. Thus, in this complementary search it was possible to retrieve tweets from

conversations that were older than 7 days. A complete description of the searches and the code used can be found in Appendix B. In total, 1 000 103 tweets were initially collected.

When fetching tweets using the Search API, there are several fields one may extract and store. Table 2 shows the fields that were stored for each tweet in this study.

Table 2: A table of fields stored for each tweet used in the study.

Field Descritpion

id A unique ID of the tweet

user_id A unique ID of the user posting the tweet (does not contain any personal data)

created_at The time the tweet was originally posted

in_reply_to_status_id If this was a reply, the ID of the tweet it replied to in_reply_to_user_id If this was a reply, the ID of the user it replied to

pos The score of positive sentiment, calculated as described in Section 3.2

neg The score of negative sentiment, computed by the VADER sentiment analysis tool described in Section 3.2

neu The score of neutral sentiment, computed by the VADER sentiment analysis tool described in Section 3.2

compound The compound sentiment score, computed by the VADER sentiment analysis tool described in Section 3.2

(35)

25 One thing that is noticeable in Table 2 is that the actual text of the tweets has not been stored.

The reason for this is that the Twitter developer policy (Twitter, 2020b) states that if Twitter content is stored offline, it must be kept up to date with the current state of that content on Twitter. More specifically, one must delete or modify any content if it is deleted or modified on Twitter. For the purpose of this thesis, only the sentiment of a tweet is of interest and not the content of the tweet itself. Thus, for the sake of simplicity, the text in the tweets were never stored.

3. 2. Sentiment Analysis

Sentiment is described by Seifert and Kwon (2019) as the contextual polarity of a text, or whether a piece of writing expresses a positive, negative or neutral opinion about a subject.

Furthermore, the authors describe sentiment of a social media network-based brand-related eWOM as whether the users’ narrative on a brand experience expresses a positive, negative or neutral view about the brand. Sentiment analysis can thus be described as the process of identifying positive, negative or neutral expressions in a text. According to Seifert and Kwon (2019), sentiment analysis has been one of the most prominent ways of mining social media text in both academic and industry research due to its predictive power.

With the definition of sentiment analysis as the process of identifying positive, negative or neutral expressions in a text, it can be treated as a problem of classification. Rogers and Girolami (2017, p. 169) describe that in classification problems, one is typically presented with a set of N training objects {𝑥_#, … , 𝑥_&}, each with a corresponding label 𝑡₎ that describes which class the object belongs to. For the problem of sentiment analysis, there are three classes, 𝑡₎ = {−1, 0,1}, representing negative, positive and neutral sentiment respectively. In a typical classification problem, the N training objects are used to train a classification algorithm which can then be used to predict the class 𝑡_)./ for an unseen object 𝑥_)./. However, such an approach requires access to an extensive set of training data with

corresponding labels. In this study, labelled data is scarce and thus cannot be used to train an algorithm. Instead, a classification algorithm that does not require any training data needed to be used.

There are several available tools for conducting sentiment analysis on texts. In this study, the VADER (Valence Aware Dictionary and sEntiment Reasoner) sentiment analysis tool as

Brand Feedback and eWOM Sentiment on Twitter

FACULTY OF EDUCATION AND BUSINESS STUDIES