Conversations’ Polarization on Social Media:
A Sentiment Analysis of Conversations about Goods and Services in Facebook and Twitter
Alice Brunelli- 11103515 Master’s Thesis
Graduate School of Communication
Master’s programme Communication Science Persuasive Communications
Supervisor: Mw. Dr. Guda van Noort February 3rd, 2017
Abstract
Social media have become the easiest touch-point between users and brands. Users can leave reviews and express their opinions about products and brands become aware of consumers’ tastes, feedbacks and judgements. Sometimes it can happen that the conversations about a certain brand or product become polarized. This means that users tend to express increasingly extreme opinions only because they are influenced by what other people say. Because both consumers and brands rely on social media, it is important to identify which conditions contribute to the creation of polarized conversations which provide biased information. The aim of this research is to investigate whether conversations are more polarized on Facebook or Twitter and whether the type of product (services or goods) increases polarization. 3886 conversations have been analysed using automated sentiment analysis. The results show that, as expected, conversations on Twitter are more polarized than on Facebook. However, contrary to the expectations, conversations about goods are more polarized than services ones. Theoretical and practical implications are discussed.
Keywords: polarization, conversations, Facebook, Twitter, goods brands, service brands.
Conversations’ Polarization on Social Media:
A Sentiment Analysis of Conversations about Goods and Services in Facebook and Twitter
Social networking sites (SNS) are known as spaces where users exchange different kinds of information and opinions about a great variety of topics (Drury, 2008). For instance, brands with related products and services are often the main protagonists of users’ comments (Fournier & Avery, 2011) through which, users evaluate them and provide information about their personal experience (Muntinga, Moorman, & Smit, 2011; Hornikx & Hendriks, 2015). Users’ comments, namely reviews, represent a very powerful tool for both brands and consumers. On the one side, companies can monitor comments and opinions to be able to respond to complaints or negative reviews and avoid potential reputational harm (Hornikx & Hendriks, 2015). On the other side, consumers can use these reviews as a source of
information about products and services to reduce the risks related to purchasing (Hornikx & Hendriks, 2015). Moreover, consumers also use such reviews in their decision-making process and to create an opinion (Guo & Zhou, 2016).
Users’ comments and opinions may not necessarily be unbiased judgements. Rather they can be modelled according to what other people say (Hong & Kim, 2016). It can happen, for example, that the initial dislike for a product will become stronger after users become aware that the majority of people holds a negative attitude towards it. In this scenario, where the individual position or changes or becomes more extreme after being exposed to
likeminded points of view, we can speak of opinion polarization (Lee, 2007). As polarized reviews are biased, the opinions that consumers form relying on these reviews can be biased to the same extent. Since this situation can be problematic, the first aim of this study is to examine polarization in relation to opinions about brands.
to the political sphere (e.g. Baldassarri & Geldman, 2008; Meffert, Chung, Joiner, Waks, & Garst, 2006), to homosexuality (Munro & Ditto, 1997), same-sex marriage and minority rights (Taber & Lodge, 2006) and abortion (Mouw & Sobel, 2001). Less attention has been devoted to polarization in social media settings and in relation to “everyday” topics such as the quality of a certain product, that is of particular importance because it can be used by consumers when they form their opinion (Guo & Zhou, 2016). Social media could serve as an ideal ground for investigation of this phenomenon because it allows users to choose the information they are exposed to (Sunstein, 2001). As a consequence, opinion polarization on social media would lead people to express more extreme opinions because they mainly expose themselves to information in line with their pre-existing point of view (Hong & Kim, 2016). Moreover, social media allow an “easy” collection of data because these are publicly shared by users and available for research purposes. Data can be then analysed by, for
instance, considering the words used in the comments. This way of analysis is named content analysis and allows to “quantify content in terms of predetermined categories and in a
systematic and replicable manner.” (Bryman, 2015, p. 289). Of particular interest for this study is that type of content analysis which focuses on the sentiment of the content.
Previous research indeed suggests that opinion polarization frequently occurs on social media. These conclusions are drawn on SNS in general (Lee, Choi, Kim, & Kim, 2014) or on a specific platform, such as Twitter (Hong & Kim, 2016). But social media is not to be considered as umbrella concept (Voorveld, van Noort, Muntinga, & Bronner, 2016) because different platforms aim at satisfying different psychological and social needs (Dunne, Lawlor, & Rowley, 2010) and also differ from one another because they lead consumers to a different media experience (Voorveld et. al, 2016). In particular, the most common ones Facebook and Twitter (Statista, 2016) fulfil basically the same needs, namely social connections, entertainment, relaxation, information seeking and sharing and opinion
expression (Dunne et al., 2010; Whiting & Williams, 2013). But while Facebook is mainly used to fulfil the social needs, Twitter is more prominently used to fulfil the need for information and opinion expression (Whiting & Williams, 2013; Quan-Haase, Martin, & McCay-Peet, 2015). As a consequence, Facebook is considered as a place to build online relationships and express oneself, whereas Twitter serves as a room where users discuss and express their opinions (Hughes, Rowe, Batey, & Lee, 2012). Since polarization is strictly related to opinions and opinions play a different role in the two platforms, I believe that it is not possible to draw conclusions about polarization in social media without differentiating between the two platforms. Moreover, they represent a suitable place for studying
polarization because, differently from other very popular SNS such as Snapchat and Instagram, they are mainly based on words rather than on pictures, allowing the textual analysis of users’ comments. Therefore, the second aim of this research is to examine the difference between the two social media platforms.
In examining polarization regarding opinions about brands, differences among types of brands should also to be considered. More specifically it is important to examine the differences between service brands and goods brands (Hill, 1977) because, depending on the brand type, reviews have a different relevance for consumers (Hornikx & Hendriks, 2015).
The main difference between goods and services is that goods and their quality are tangible, while services and their quality undergo a more subjective evaluation (Parasuraman, Zeithaml, & Berry, 1985). As a consequence, the risk associated with the purchase of
services is higher than that associated with goods, and even the relevance of pre and post-purchase information and reviews is influenced by the aforesaid risk level (Hornikx & Hendriks, 2015). Therefore, the third aim of this study is to examine the differences between goods and service brands in relation to polarization.
In conclusion, this study aims to investigate to what extent conversations about brands are more polarized in different social network platforms and whether polarization differs between goods and service brands. Thereby, this research contributes to the existing literature by determining which (virtual) environment and which type of content facilitate or inhibit the creation of polarized opinions. This knowledge will enrich the existing literature by testing whether previous findings about polarization hold in a different context, namely
conversations about brands. Moreover, interesting practical implications for brands and consumers can emerge. Brands can have a deeper knowledge which will help them to allocate their media budget and to take the most responsive actions to manage the possible positive or negative effects of polarization. Consumers can get a better awareness that sometimes
reviews may be biased and non-reflective of the real quality of a product. Theoretical Background
Opinion Polarization: The Concept and Origins
Opinion polarization implies a change in opinion from positive to negative or vice versa. Alternatively, opinions can become more extreme in the direction that one already holds, thus becoming more positive or more negative (Myers & Lamm, 1975). This
phenomenon has been defined in different ways. On the one side, polarization is defined as a phenomenon according to which the members of a social group split because they have opposing views. This process becomes more evident over time since more and more
individuals inside the group tend to agree with one side or the other, the so-called poles. The pendulum effect is a primary consequence of polarization. This means that poles have the tendency to make increasingly contradictory statements about a given topic worsening the psychological distance between them (Myers & Lamm, 1975). This first definition portrays polarization as an attribute of the group. On the other side, polarization is also seen as an attribute of the single individual who changes his/her initial position in line with that shared
by the majority of people in his/her in-group and will strengthen it after being exposed to likeminded points of view (Lee, 2007). This implies that a person’s opinion will become more extreme after noticing that others hold the same beliefs (Yardi & Boyd, 2010). This is expressed in social media by a “discussion thread”. This term is defined as “chain of written ideas or opinions (exchanged among two or more participants in an online discussion) linked to the sequence in which they were espoused by the participants” (discussion thread, n.d.) in Facebook or Twitter.
In social media context, when examining brand conversations, considering
polarization as an attribute of a person or a group is problematic because, differently from offline interactions, the number of people that participate a discussion can be potentially very high (Yardi & Boyd, 2010). In addition, these people do not actually have to be part of the same group. Moreover, users join a discussion with one comment only, and this makes it difficult to track how the opinions of every single user become more extreme over time (Yardi & Boyd, 2010). In addition, because the users observe the whole discussion, it could be that it is the conversation that shapes his/her opinion rather than a single comment which is part of it. These reasons increase the importance of examining the whole discussion thread and not the single extreme opinions. Furthermore, the analysis of the discussion thread allows a more precise measurement since it considers how the general sentiment of the whole
discussion thread changes. Therefore, this study examines polarization as an attribute of the discussion thread.
The origins of the polarization phenomenon can be explained by the cognitive dissonance theory (Festinger, 1957). This theory assumes that whenever individuals are confronted with information that is not in line with their own attitudes or beliefs, they feel dissonance and find themselves in an uncomfortable psychological state. To reduce or eliminate this state they make use of heuristics or mental shortcuts. Selective exposure is one
of these and implies that individuals will actively look for likeminded information and avoid counterarguments to reinforce their own beliefs and reduce dissonance (Kim, Hsu, & de
Zúñiga, 2013). For instance, after a purchase, a consumer can be unsure that he/she made the right choice. The consumer can, therefore, decide to use social media to find support for the choice. To do so he/she is likely only to look for information which confirms his/her decision and to avoid counterarguments. This will solve the internal conflict the consumer feels and reassure him/her about the purchase decision (Telci, Maden, & Kantur, 2011)
Online polarization: Differences between Facebook and Twitter
In the online context, opinion polarization is most likely to occur on social media (Hong & Kim, 2016). There, users can both gather information and produce content. For instance, they can learn about the existence of a new product but also use reviews to form a personal opinion about a product (Belch & Belch, 2015; Guo & Zhou, 2016). As mentioned, social media is not to be considered as umbrella concept because every platform is different from the others for its technical characteristics (e.g. layout), for its unique content and for the way users experience it (Voorveld et al., 2016). Before deepening into the differences
between the platforms and how these influence polarization, it is important to give a general overview of what social media are.
Online social networks can be defined as online platforms through which users share content (namely information, ideas, pictures, videos) and communicate with others (Drury, 2008). They allow users to have an active role in building their own network by deciding which other users are part of it (Hughes et al., 2012). Because of this, it is likely that selective exposure is even stronger in social media. In fact, users can decide which kind of information must appear on their homepage and thus determine how heterogeneous the information they will be exposed to in their social network will be (Sunstein, 2001). If selective exposure is very strong, it means that a user will visualize very homogeneous information, that is
information consistent with each other. The less heterogeneous the information and arguments that appear on one’s home page are, the more selective exposure has occurred. However, every social media differs from the others in terms of structure and use people make of it. For instance, the two most common ones Facebook and Twitter (Statista, 2016) are similar to each other because they are places where users express themselves (Drury, 2008). However, they also have other characteristics which make them very different from one another that lead people to use them for various purposes (Dunne et al., 2010; Voorveld et al.,2016).
Twitter is a microblogging website that was founded in 2006. It allows users to tweet (publish content), reply and retweet (forward) posts that do not exceed 140 characters. Users build their network by following those who post the information they want to read. Tweets mainly aim at sharing information, opinions, news and complaints (Smith, Fischer,
&Yongjian, 2012). Twitter seems to be more focused on the exchange of information and opinions (Kwak, Lee, Park, & Moon, 2010). Although users have conversations on Twitter, these are more unidirectional instead of reciprocal. The perception is increased even more by the fact that other contacts are named “followers” and by the fact that users are not allowed to decide who is going to follow them on the social media (Davenport, Bergman,Bergman, & Fearrington, 2014). Facebook is a social networking site born in 2004 that allows users to create personal profiles that include personal information, interests, tastes and pictures. Moreover, they can also share posts without characters’ restrictions and interact with their friends by writing on their profile page and commenting their posts and pictures. Through Facebook, users can keep up with their friends’ lives, with rumours and gossip. They can also communicate and maintain relationships (Smith et al., 2012). Facebook is a place for
socialization and reciprocal conversation, and it can be considered as a more heterogeneous social network (Hughes et al.,2012). Users do not build their network according to what other
people have to say, rather they send friendship requests mainly to friends or acquaintances with whom they want to share personal information and have online relationships (Hughes et al.,2012).
As these brief overviews show, the two platforms differ for many aspects. Some of these are of particular interest for this research because they suggest that polarization can occur differently on Facebook and Twitter. First, the limited number of characters that Twitter, unlike Facebook, has could lead users to express less nuanced thoughts and comments. Second, while Facebook encourages users to build a personal profile with as much personal information as possible, Twitter does not give the same possibility. Therefore, Twitter allows users to regain part of the anonymity lost with social media (Hughes et
al.,2012) reducing people’s inhibition to express extreme attitudes (Lee, 2007). The last and maybe more relevant difference is that while Facebook serves as a way to cultivate
friendships, Twitter is mainly used as a source of information and to express one’s opinion. According to the aforementioned studies (Sunstein, 2001; Lee et al., 2014), people have the tendency to look for information which supports their own opinion or attitude. Consequently, it is likely that Twitter users, more than Facebook ones, build a network of likeminded points of view. For this reason, Twitter can be considered as a sort of echo chamber that can
potentially and easily contribute to form polarized opinions (Hong & Kim, 2016). Twitter can, therefore, be considered a less heterogeneous SNS.
As mentioned in the previous part of this paper, the focus of this research is on opinions that users have about brands. While examining the relationship between Twitter or Facebook and polarization, conversations about brands will be taken into consideration. From these considerations about the differences between the two social media, the first hypothesis of this study is the following:
Product Type: Goods and Services and Polarization
So far, this study has provided the rationale for studying polarization on different social media platforms. However, the motivations why it is important to study the
phenomenon in relation to brands are still missing. Before distinguishing between the types of brands, it is important to understand why social media are so important for brands.
According to economic sociologists, the market is not a predefined space, rather is it a process built on exchanging of ideas and opinions from both brands and consumers (Song, Cheon, Lee, Lim, Chung, & Rim, 2014). Brands, on the one side, provide information, support consumers during and after the purchase and acknowledge consumers’ opinions and evaluations (Muntinga et al., 2011). Consumers, on the other side, consider possible
purchases, search for information provided both by other users and brands, make decisions and then evaluate a product (Belch & Belch, 2015; Guo & Zhou, 2016). This idea of market as result of the intertwining of brands and consumers is also a key point of the social
consumer decision journey (Belch & Belch, 2015). According to this view, the decision-making process that leads to the purchase is not linear, rather is it a circular process in which feedbacks play a crucial role. In this perspective, social media represent a valuable resource because they allow consumers to actively search for information and marketers to be actively involved with their consumers’ journey. However, the journey consumers go through when considering a purchase does not always have the same length. Rather it depends on several factors like, for instance, the risk consumers associate with the product. Riskier products have longer cycles which imply a higher importance of feedbacks.
Considering the risk factor, it is possible to distinguish products into two different categories, namely services and goods (Hill, 1977). As mentioned before, goods are
essentially objects or manufactured products which are tangible and have constant quality as shoes, food and electronic devices. Services are intangible, and their evaluation varies from
consumer to consumer, for instance, an airline, a bank or an online retailer (Parasuraman et al., 1985). Moreover, the risk associated with purchases also differs, so that services are perceived as riskier than goods (Hornikx & Hendriks, 2015). This view is also supported by Suwelack, Hogreve and Hoyer (2011) who, taking the definition of Nelson (1974),
distinguish between experience goods and search goods.
Experience goods are products that need to be personally experienced before purchase and evaluation. They correspond to what was defined as services in this paper. Differently, search goods evaluation can be assessed considering information that is already available and they correspond to what was defined as goods in this paper. The evaluation of the first ones is more subjective and less diagnostic (Hoch & Deighton, 1989) and the risk associated with them is higher. The second ones can be evaluated before the purchase thus presenting a lower risk.
Hornikx and Henriks (2015) found that information on social media is more likely to concern services than goods. The authors also share the idea that pre-purchase information and post-purchase evaluation are more important for services than for goods. Because these studies suggest that users tend to provide a greater amount of opinions and evaluations for services and that these are more diagnostic, it is likely that opinions will be more polarized in relations to services than for goods.
H2: Conversations about service brands are more polarized than those about goods brands
Platform Type Facebook vs Twitter Product Type Goods vs Services Conversation Polarization H2 H1
Figure 1. Model of the study
Method Sample
The research question on this paper focused on whether conversations about brands are more polarized according the social network platform in which they take place and whether these become more intense depending on the type of product that the brand sells (services vs goods). The unit of analysis in this paper is therefore discussion thread, also referred to as conversation in this paper, on Facebook and Twitter. These are defined here as conversations built of brand-related posts and their comments posted by users or by the brands themselves on the official Facebook profile or Twitter account of the brands. Moreover, discussion threads were defined as such if they were composed of the first post and at least of two comments.
The two platforms have proper terminology to call the elements of the discussion threads (i.e. tweet, replies, post and comments). However, for convenience the words “first post” will be used in this paper to refer both to Facebook initial posts and tweets which begin a conversation, the word “comment” both for Facebook comments and twitter replies and the word “post” as general terms for both first posts and comments.
Six brands were selected to build the sample. These are a fashion brand (New
Balance), a technology brand (Samsung), a food brand (McDonald’s), an online retailer brand (Amazon), a financial services brand (Deutsche Bank) and an airline (EasyJet). According to the definition used previously in this paper (Hill, 1977) the first 3 representative of the goods category and the last 3 of the service one. These brands were chosen because they are 1) internationally known, 2) they have an official Facebook page and Twitter account 3) which is daily used both by themselves and by users. However, since the Facebook data for the fashion and the food brands chosen were not available for the download, these two were replaced with other two brands representative of the same category and that also have similar
price range and target of the original ones. These brands are Converse and Subway. In total 30540 posts and 126792 tweets were collected. After cleaning the data, the final dataset includes 2422 Facebook conversations and 1464 twitter conversations. Of the Facebook conversations the majority (75.6%) is made of one original post and less than 6 comments, while the majority of conversations from Twitter (93.7%) is composed of the original post and less than 3 comments.
Table 1.
Brands and Data
Facebook Twitter Brands Number of Facebook posts Number of discussions Number of Tweets Number of discussions Goods: Fashion brand 3254 157 344 27 Technology brand 4825 430 16377 140 Food brand 4621 267 14035 378 Services: Online retailer brand 5912 436 91104 819 Financial service brand 4557 443 174 5 Airline brand 7371 689 4758 95 Procedure
A particular type of automated content analysis, namely sentiment analysis, was used to answer the research question and hypotheses. Content analysis is a systematic approach of analysis of texts or documents. It aims at quantifying in a replicable way the media content through the use of predetermined categories. In the case of sentiment analysis, the aim is that of determining how positive or negative the content of the text is (Bryman, 2015)). Moreover, this method is particularly suitable for this study because it is unobtrusive and does not
directly involve users. This allowed the collection of unbiased data to detect the phenomenon of polarization which usually takes place outside the individuals’ awareness (Lee et al. 2014). In addition, the automated feature of content analysis enhances the accuracy of the analysis and allows examining a larger amount of data (Macnamara, 2005).
The data were collected between November, 24th and December, 1st 2016 with the use
of Facebook and Twitter Application Programming Interface (API) that allowed the
extraction of the data from the two platforms. In the Facebook’s case, Jupyter Notebook was used. It is a server-client application, that with the use of a script allowed the download of the last 10.000 users’ posts published on the official pages of the chosen brands. The result was an Excel document that allowed to understand which comments belonged to the same conversations, thus identifying the discussion threads. Twitter does not allow collecting data as Facebook does, for this reason a server was set up during the data collection time. This server allowed collecting 1% of the data available that contained previously determined
keywords. Queries, that could be meaningful to download the data about conversations on
Twitter, were built looking at the accounts of the brands to see whether common patterns occur. Moreover, the literature about conversation and discourse analysis (e.g. Wetherell, Taylor, & Yates, 2001; Belkaroui, Faiz, & Elkhlifi, 2014) gave an idea of which elements are usually present in discussions. In particular, there are usually questions, replies to these and other comments where users express their opinion (see Appendix for an overview of the queries).
The data were then cleaned excluding posts which were not in English and conversations
composed by less that 3 posts. Then they were analysed with the use of a software that
determines the sentiment of the text, namely Sentistrength. The functioning of the software will be better explained in the following paragraph.
Sentistrenght: A tool for sentiment analysis. To analyse the data, I carried out the automated content analysis focusing on the sentiment of the posts and tweets. To do so, I used “Sentistrength”, a software that produces an estimate of the sentiment content of strings of texts. This software is based on a list of sentiment terms and word stems (Thelwall, 2013) but also considers other elements as repeated punctuation, exclamation marks, use of capital letters and also emoticons. This software works in a way that it attributes to each text both a positive value from +1 (no positive sentiment) to +5(very strong positive sentiment) and a negative one from -1 (no negative sentiment) to -5 (very strong negative sentiment). For example:
“Your service is absolutely terrible!!!”
This post posted from a user on the Facebook page of Deutsche Bank was analysed by Sentistrength in the following way:
Your[0] service[0] is[0] absolutely[0] terrible[-3][-1 LastWordBoosterStrength][-0.6 EmphasisInPunctuation] [[Sentence=-6,1=word max, 1-5]][[[1,-5 max of sentences]]] Positive value= +1; negative value= -5
This post scored +1 on the positive scale, meaning that it contains no positive sentiment and -5 (very strong negative sentiment) on the negative scale.
Variables
Platform type. This variable was coded 0 when the conversation took place on Facebook or 1 on Twitter. 37,7% of the total discussions were retrieved from Twitter and 62,3% from Facebook.
Product type. This variable was coded 0 when the discussions concerned a brand representative of the goods category or 1when it concerned a service brand. 36% of the conversations are about goods, while 64% about services.
are. The sentiment was coded for each post and a negative and a positive score (value
between 1 and 5 and between -1 and -5) was assigned to each post by Sentistrength. Then, in order to measure polarization, I considered these two values and I calculated the sum between the positive and the negative value in order to have a unique value expressing the sentiment of the texts (M=.03, SD=1.13). For instance, “I absolutely love having Kenmore! I never have to use my warranty, but I know that their warranty is a good one!” was attributed a positive value of +4 and a negative value of -1, therefore the final score of this post was +3.
Finally, to assess whether the polarization took place, it was considered how extreme every post was rated. Then each post was compared with the others belonging to the same thread. Since the polarization implies that opinions become more extreme while the
discussion evolves, last comments should be rated as more extreme than the ones posted at the beginning of the discussion. In line with this, I calculate the polarization value for every conversation by calculating the difference between the sentiment of the first post and that of the last comment of the related conversation (M=-.11, SD= 1.52). Among the conversations the majority (39,2%) were polarized in a negative direction, 31,7% or the conversations were positively polarized, while 31% were not polarized (Table 2). Polarization did not happen with the same intensity in all the cases, rather it varied according to how much extreme was the last post of the conversation compared to the first one. Positively polarized conversations scored between +1 (slightly positively polarized) and +5 (extremely positively polarized), while negatively polarized conversations scored between -1 (slightly negatively polarized) and -5 (extremely negatively polarized).
Table 2.
Frequency and Meaning of Polarized Conversations
Value Meaning Percentage
Negative polarization (31.7%) -5 Extremely polarized 2% -4 Very polarized 1.4% -3 Polarized 4.3% -2 A little polarized 10.4% -1 Slightly polarized 21% Positive polarization (39.2%) +1 Slightly polarized 18.7% +2 A little polarized 8.7% +3 Polarized 3% +4 Very polarized 1% +5 Extremely polarized .3% Results
Hypothesis 1 expect the there would be a relationship between the platform and polarization. In particular, the effect was expected to be stronger for Twitter. Hypothesis 2 stated that the product type will moderate the relationship between the platform and
polarization. Both hypotheses were tested with a regression using platform (0= Facebook, 1= Twitter) as independent variable, product type (0= goods brands, 1= service brands) as moderator and polarization as dependent variable. Table 3 shows how the variables are correlated.
Table 3.
Correlations of the Variables included in the Study
Platform Product Platform x Product
Platform -
Product -.02 -
Platform x Product .72* .42* - Note. N= 3886, *p <.001
The findings show that the model as a whole is significant F (3, 3882) = 11.94, p < .001. The independent variables predicted the 0.9% of the variance in integration (adj. R2 = .008). The average VIF of the independent variables is 2.56. This value means that
multicollinearity is not a problem for the analysis. The platform type b = .19, b* = .06, t = 2.28, p = .02, 95% CI [.03, .35] has a small positive effect on polarization. Conversations on Twitter are more polarized than on Facebook. The product type b = -.13, b* = -.04, t = -2.06, p = .04, 95% CI [-.26, -.01] also has a small positive direct effect on polarization. This means that conversations about service brands are less polarized that the ones about goods brands. The interaction effect between the platform and the product type is not significant b = .14, b* = .04, t = 1.38, p = .17, 95% CI [-.06, .35]. Therefore, hypothesis 1 stating that conversations are more polarized on Twitter than on Facebook is accepted. However, no support was found for hypothesis 2 because the product type does not moderate the relationship between the platform type and polarization. However, it is important to notice that a direct effect of product type on polarization b = -.13, b* = -.04, t = -2.06, p = .04, 95% CI [-.26, -.01] was found, suggesting that polarization is stronger in discussions about brands producing goods.
Conclusions and Discussion
The study reported in this paper focused on a psychological phenomenon that happens continuously in everyday life both in the offline and online setting: opinion polarization. The rationale behind this research was the lack of empirical studies that investigated this
phenomenon in the online world, in particular in social media that have become more and more fundamental for sharing information and communication (Drury, 2008). Among the ones available, Facebook and Twitter were chosen because they are the most used by users to leave reviews about their favourite or disliked products and brands (Statista, 2016; Muntinga et al., 2011). The choice of not considering social media as a whole but distinguishing between them was made on the basis of the Use and Gratification theory applied to media (Dunne et al., 2010; Whiting & Williams, 2013) and other studies (e.g. Muntinga et al., 2011; Voorveld et al., 2016). These have underlined that every platform differs from the other because it fulfils different users’ needs and gives them a different media experience. Among products, a basic distinction between goods and services was made (Hill, 1977) starting from the assumption that depending on the purchase risk associated with a product, reviews have a different importance (Hornikx & Hendriks, 2015). From these premises, two hypotheses were formulated and tested using automated sentiment analysis. The first hypothesis stated that conversations about brands are more polarized on Twitter than on Facebook, while the
Platform Type Facebook vs Twitter Product Type Goods vs Services Conversation Polarization b=.19 b= -.13 b=.14
Figure 2. Model and results: bold lines are significant results (p<.05), dotted lines are non significant results (p>.05).
second hypothesized that conversations were even more polarized when they concern services brands instead of goods brands. These hypotheses were tested by automatically analysing 3886 conversations related to goods brands (i.e. fashion, food, and technology brands) and service brands (i.e. online retailer, financial services, and airline brands). Automated sentiment analysis was used to analyse conversation polarization by first determining the sentiment (from -5 to +5) of each post of the conversation and than considering whether and how the sentiment changed from the first to the last post.
The results confirmed the first hypothesis. Conversations about brands on Twitter were more and more extreme throughout their development. On the other hand, this
phenomenon was less likely to occur for conversations on Facebook. Support for this finding could be found in the literature that investigates the difference between the two social
networks. When users want to express an opinion about a brand or product and build a conversation around it, they prefer Twitter rather than Facebook because Twitter is the platform for informing and sharing information (Kwak et al., 2010). This finding provided further support for the echo chamber view (Hong & Kim, 2015) according to which Twitter is particularly susceptible to selective exposure. Therefore, users on Twitter surround
themselves with likeminded points of views and this increases the probability of polarizations (Lee, et al., 2014)
The results were also in line with studies that examined differences between social media, though not focusing on polarization. For instance, Voorveld et al. (2016) studied the different feelings that users experience when using a particular platform. Although their study considered a number of platforms wider than that in this study, the results were similar. In particular, people used different social media for different purposes and each platform gave users a unique experience. The findings confirmed that Facebook was used to interact with others and out of boredom, while Twitter as an informative tool.
The second hypothesis integrated a new variable in the model, namely product type. Based on the findings of Honikx and Hendriks (2015) and of Suwelack, Hogreve, and Hoyer (2011), it was expected that this would moderate the relationship between platform type and conversation polarization. More specifically, it was supposed that conversations would be even more polarized when services, rather than goods, were the main topic. Thus it was hypothesized that conversations about services on Twitter should have been the most polarized. However, contrary to the expectations, the moderation was not significant. Therefore, the second hypothesis was not supported. However, the results showed an unexpected finding. In fact, product type influenced conversation polarization directly, thus being not a moderator but a second independent variable. In particular, conversations about goods were more polarized than those concerning services. Therefore, both platform type and product type contributed in explaining the polarization phenomenon but rather than being related with one another they are two separate causes.
The results about how the platform type influenced conversation polarization are inconsistent with previous findings (e.g. Hornikx & Hendriks, 2015). Although Hornikx and Hendriks (2015) did not focus on polarization but rather on users’ opinions and reviews, their results suggested that users expressed more opinions about services than goods. Therefore, it was hypothesised that polarization, which is based on opinions, would also happen
accordingly. However, the results of this research showed that conversations about goods are more polarized than the ones about services, thus being in contrast with expectations. An explanation for this finding may be that because goods have tangible characteristics and consistent quality they are easier to evaluate than services (Hoch & Deighton, 1989). As a consequence, opinions about goods could be more extreme just because they are easier to formulate. Another possible explanation may be that the results depended on the sample
which may not have been fully representative. The brands chosen to build the sample were selected because they matched the definition given for goods and services. However, the distinction between the two categories is not clear, but it is rather blurred. In fact, many goods also have a service component and vice versa. For instance, a pair of shoes matches the goods category since it is tangible and consumers could express an opinion before actually trying it. However, the same consumer could order the shoes online and have them delivered at home. In this case, the online conversation about goods may actually be related to the service that the goods brand delivers.
Anyway, the possible explanations provided above are just speculation. Especially because these findings are not in line with the existing literature, further research is needed to
replicate the study. For instance, it could be interesting to use a different sample or carry out content analysis regarding topics and issues discussed in the conversations rather than selecting the sample according to the brand category.
Practical Implications
This study can have interesting implications from a marketing communication perspective. First, it could provide further knowledge about social networks that represent a critical area of interest for marketers. Social networks have become the main place where companies meet potential consumers and vice versa (Fisher, 2009). Because of that, brands are allocating in this area a big portion of their budget for marketing. However, the return on investment is often not as conspicuous as expected. By gaining more knowledge about social networks, marketers could invest their money in a more conscious way. Second, brands should be aware of the phenomenon since it is likely to affect the opinions of a large group of consumers (Lee, 2007). If the majority of a group likes a product, this can become a sort of social norm to which other people could adhere to. Social norms are defined as properties of a certain group that characterize where the group positions itself along a certain dimension
(Prentice, 2008), in this case, the liking or disliking for a specific product or brand. A user who is part of a group or wishes to be part of it can conform to the opinions of that group. If the group is polarized, that individual and potentially many others could adopt the same polarized opinion, and if it is negative, this opinion change can damage companies.
According to this study, brands should regularly monitor social media to on prevent negative polarization to happen or promote the positive one. If the polarization is positive, it might be important to engage in actions that further foster the positive social media
discussion. On the contrary, if the polarization is taking a negative turn, the brand should engage in responsive actions of webcare to minimize the possibility of reputational harm.
The focus should be mainly on negative polarization because, as shown in many studies (e.g. Skowronski & Carlston, 1889; Ahluwalia, Burnkrant, & Unnava, 2000), negative information is more diagnostic and is weighted more than positive information. Therefore, it is easier for a brand to be damaged by negative polarization than advantaged by the positive one. In relation to this, the present study shows that companies should devote more attention and webcare efforts on Twitter than on Facebook and if they provide goods rather than services. Furthermore, consumers could also take advantage from the results of this study. They could become aware that sometimes opinions and evaluations of a certain product may be biased and thus not reflective of its real quality.
Limitations and Future Research
The study presents some limitations that should be taken into consideration when reading its results. The main limitations are due to the difficulties of retrieving data from Twitter. Because of its policies, only a small amount of the tweets published are publicly available for the download. To download the data a limited number of queries, which were thought to be meaningful for the purposes of this study, were created. However, other useful data may have been excluded, threatening the validity of the results due to the missing data.
Future research could use different queries. Moreover, during download, data collection exceeded the limits of the server several times. This implies that randomly some tweets related to the queries were not downloaded. In addition, the timeframe for data collection on Twitter is quite small (1 week) and therefore the results may not be completely representative and generalizable. The amount of data retrieved from Twitter is less than the one downloaded from Facebook. Because the time periods for data collection are different and small, the external validity might have been threatened by external events which influenced the conversations. For instance, these events might have made the conversations more extreme then they would normally be. Another limitation of the study could be related to the
methodology used to analyse the data, namely the automated sentiment analysis
(Sentistrength). For developing Sentistrength, 2600 social media comments were tested several times by different coders in order to guarantee the reliability and accuracy of the sentiment interpretation (Thelwall, 2013). Although the software used was created and tested to be as similar to the manual content analysis as possible, there is always the doubt whether the interpretation of a machine can be as reliable as the one of a real person. For instance, the software “does not attempt to use grammatical parsing (e.g., part of speech tagging) to disambiguate between different word senses” (Thelwall, 2013, p. 4). Therefore, if a word has more than one meaning, the software may not be able to interpret it. Future research could use manual coding for sentiment analysis, maybe on a subsample to check the reliability of the automated sentiment analysis.
Future research could devote its attention to replicate this study on a different sample to see whether the same unexpected result, namely conversations about goods are more polarized, occurs. Moreover, it would also be interesting to differentiate between negative and positive polarization and test the effects of the platform and product type on these two sides of the phenomenon. The results also showed that the conversations were positively or
negatively polarized. This means that conversations either become more positive or more negative throughout their development. However, in this study, no distinction was made between positive or negative polarization. By examining which platform “promotes” which kind of polarization, companies could act even more specifically to minimize or maximize its impact.
References
Ahluwalia, R., Burnkrant, R. E., & Unnava, H. R. (2000). Consumer response to negative publicity: The moderating role of commitment. Journal of marketing research, 37(2), 203-214.
Baldassarri, D., & Gelman, A. (2008). Partisans without constraint: Political polarization and trends in American public opinion. AJS; American journal of sociology, 114(2), 408. Belch, G. E., & Belch, M. A. (2015). Advertising and promotion: An integrated marketing
communications perspective. Boston: McGraw-Hill.
Belkaroui, R., Faiz, R., & Elkhlifi, A. (2014, November). Conversation analysis on social networking sites. In Signal-Image Technology and Internet-Based Systems (SITIS), 2014 Tenth International Conference on (pp. 172-178). IEEE.
Bryman, A. (2015). Social research methods. Oxford university press.
Davenport, S. W., Bergman, S. M., Bergman, J. Z., & Fearrington, M. E. (2014). Twitter versus Facebook: Exploring the role of narcissism in the motives and usage of different social media platforms. Computers in Human Behavior, 32, 212-220.
Discussion thread. (n.d.). retrieved from:
http://www.businessdictionary.com/definition/discussion-thread.html
Drury, G. (2008). Opinion piece: Social media: Should marketers engage and how can it be done effectively?. Journal of Direct, Data and Digital Marketing Practice, 9(3), 274-277.
Dunne, Á., Lawlor, M. A., & Rowley, J. (2010). Young people's use of online social
networking sites-a uses and gratifications perspective. Journal of Research in Interactive Marketing, 4(1), 46-58.
Festinger, L. (1957). A theory of cognitive dissonance. Stanford, CA: Stanford University press.
Fisher, T. (2009). ROI in social media: A look at the arguments. Database marketing & costumer and strategy management, 16(3), 189-195.
Fournier, S., & Avery, J. (2011). The uninvited brand. Business horizons, 54(3), 193-207.
Guo, B., & Zhou, S. (2016). Understanding the impact of prior reviews on subsequent reviews: The role of rating volume, variance and reviewer characteristics. Electronic
Commerce Research and Applications, 20, 147-158.
Hill, T. P. (1977). On goods and services. Review of income and wealth, 23(4), 315-338. Hoch, S. J., & Deighton, J. (1989). Managing what consumers learn from experience. The
Journal of Marketing, 1-20.
Hong, S., & Kim, S. H. (2016). Political polarization on twitter: Implications for the use of social media in digital governments. Government Information Quarterly.
Hornikx, J., & Hendriks, B. (2015). Consumer tweets about brands: A content analysis of sentiment tweets about goods and services. Journal of Creative Communications, 10(2), 176-185.
Hughes, D. J., Rowe, M., Batey, M., & Lee, A. (2012). A tale of two sites: Twitter vs. Facebook and the personality predictors of social media usage. Computers in Human Behavior, 28(2), 561-569.
Kim, Y., Hsu, S. H., & de Zúñiga, H. G. (2013). Influence of social media use on discussion network heterogeneity and civic engagement: The moderating role of personality traits. Journal of Communication, 63(3), 498-516.
Kwak, H., Lee, C., Park, H., & Moon, S. (2010). What is Twitter, a social network or a news media? Paper presented at the Proceedings of the 19th international conference on World Wide Web, Raleigh, NC. 10.1145/1772690.1772751
Lee, E. J. (2007). Deindividuation effects on group polarization in computer-mediated communication: The role of group identification, public-self-awareness, and perceived argument quality. Journal of Communication, 57(2), 385-403.
Lee, J. K., Choi, J., Kim, C., & Kim, Y. (2014). Social media, network heterogeneity, and opinion polarization. Journal of communication, 64(4), 702-722.
Macnamara, J. R. (2005). Media content analysis: Its uses, benefits and best practice methodology. Asia-Pacific Public Relations Journal, 6(1), 1.
Meffert, M. F., Chung, S., Joiner, A. J., Waks, L., & Garst, J. (2006). The effects of
negativity and motivated information processing during a political campaign. Journal of Communication, 56(1), 27-51.
Most famous social network sites worldwide as of September 2016, ranked by number of active users (in millions). (2016, September). Retrieved from:
https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-of-users/
Mouw, T., & Sobel, M. E. (2001). Culture wars and opinion polarization: The case of abortion1. American Journal of Sociology, 106(4), 913-943.
Munro, G. D., & Ditto, P. H. (1997). Biased assimilation, attitude polarization, and affect in reactions to stereotype-relevant scientific information. Personality and Social
Psychology Bulletin, 23(6), 636-653.
Muntinga, D. G., Moorman, M., & Smit, E. G. (2011). Introducing COBRAs: Exploring motivations for brand-related social media use. International Journal of Advertising, 30(1), 13-46.
Myers, D. G., & Lamm, H. (1975). The polarizing effect of group discussion: The discovery that discussion tends to enhance the average prediscussion tendency has stimulated new insights about the nature of group influence. American Scientist, 63(3), 297-303.
Nelson, P. (1974). Advertising as information. Journal of political economy, 82(4), 729-754. Parasuraman, A., Zeithaml, V., & Berry, L.L. (1985). A conceptual model of service quality
and its implications for future research. Journal of Marketing, 49(4), 41–50. Prentice, D. A. (2008). Mobilizing and weakening peer influence as mechanisms for
changing behavior. Understanding peer influence in children and adolescents, 161-180. Quan-Haase, A., Martin, K., & McCay-Peet, L. (2015). Networks of digital humanities
scholars: The informational and social uses and gratifications of Twitter. Big Data & Society, 2(1), 2053951715589417.
Skowronski, J. J., & Carlston, D. E. (1989). Negativity and extremity biases in impression formation: A review of explanations. Psychological bulletin, 105(1), 131.
Smith, A. N., Fischer, E., & Yongjian, C. (2012). How does brand-related user-generated content differ across YouTube, Facebook, and Twitter?. Journal of Interactive
Marketing, 26(2), 102-113.
Song, G. Y., Cheon, Y., Lee, K., Lim, H., Chung, K. Y., & Rim, H. C. (2014). Multiple categorizations of products: Cognitive modeling of customers through social media data mining. Personal and Ubiquitous Computing, 18(6), 1387-1403.
Sunstein, C. R. (2001). republic.com. Princeton, NJ: Princeton University Press
Suwelack, T., Hogreve, J., & Hoyer, W. D. (2011). Understanding money-back guarantees: Cognitive, affective, and behavioral outcomes. Journal of Retailing, 87(4), 462-478. Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs.
American Journal of Political Science, 50(3), 755–769. doi:10.1111/j.1540-5907.2006. 00214.x.
Telci, E. E., Maden, C., & Kantur, D. (2011). The theory of cognitive dissonance: A
marketing and management perspective. Procedia-Social and Behavioral Sciences, 24, 378-386.
Thelwall, M. (2013). Heart and soul: Sentiment strength detection in the social web with sentistrength. Proceedings of the CyberEmotions, 1-14.
Voorveld, H. A. M., van Noort, G., Muntinga, D. G., & Bronner, F. E. (2016, July). Social media experiences and appreciations for brand communication: The differentiating role of type of platform. Paper presented at ICORIA 2016, Ljubljana, Slovenia. Wetherell, M., Taylor, S., & Yates, S. J. (Eds.). (2001). Discourse as data: A guide for
analysis. Sage.
Whiting, A., & Williams, D. (2013). Why people use social media: A uses and gratifications approach. Qualitative market research,
Yardi, S., & Boyd, D. (2010). Dynamic debates: An analysis of group polarization over time on twitter. Bulletin of Science, Technology & Society, 30(5), 316-327.
Appendix
List of Queries used to Download the Data from Twitter
Brand name Queries
Food brand mcdonalds is, believe mcdonalds, think mcdonalds, mcdonalds opinion, mcdonalds experience, what mcdonalds, how mcdonalds, who mcdonalds, where mcdonalds, when mcdonalds, agree mcdonalds, disagree mcdonald, why mcdonalds, mcdonalds true, mcdonalds feel, it seems mcdonalds
Fashion brand
newbalance is, believe newbalance, think newbalance, newbalance opinion, newbalance experience, what newbalance, how newbalance, who
newbalance, where newbalance, when newbalance, agree newbalance, disagree newbalance, why newbalance, newbalance true, newbalance feel, it seems newbalance
Technology brand
Samsung is, believe Samsung, think Samsung, Samsung opinion, Samsung experience, what Samsung, how Samsung, who Samsung, where Samsung, when Samsung, agree Samsung, disagree Samsung, why Samsung,
Samsung true, Samsung feel, it seems Samsung
Airline easyjet is, believe easyjet, think easyjet, easyjet opinion, easyjet experience, what easyjet, how easyjet, who easyjet, where easyjet, when easyjet, agree easyjet, disagree easyjet, why easyjet, easyjet true, easyjet feel, it seems easyjet
Financial service brand
deutschebank is, believe deutschebank, think deutschebank, deutschebank opinion, deutschebank experience, what deutschebank, how deutschebank, who deutschebank, where deutschebank, when deutschebank, agree deutschebank, disagree deutschebank, why deutschebank, deutschebank true, deutschebank feel, it seems deutschebank.
Online retailer brand
amazon is, believe amazon, think amazon, amazon opinion, amazon
experience, what amazon, how amazon, who amazon, where amazon, when amazon, agree amazon, disagree amazon, why amazon, amazon true, amazon feel, it seems amazon.
