Flow of information from media to twitter : effects on corporate reputation and outcomes

!

!

!

!

!

!

!

Flow of Information from Media to Twitter: Effects on

Corporate Reputation and Outcomes

By Katia Meggiorin Student nr. 10230718

!

!

!

Research Master in Communication Sciences Master Thesis

!

!

!

!

!

!

!

!

!

This study investigates the flow of influence from media to Twitter users in the creation of a company’s reputation. Two-step flow of communication and agenda-setting theories from media effects literature are used to show the influence of media on Twitter discussions as mediated by opinion leaders. To test the influence of media on belief formation, we investigate corporate reputation as formed in online environments and its effect on a target company’s productive outcomes. In particular, third-level agenda-setting is used to show the influence of media on the prominence construct of corporate reputation. The sentiment towards the target company is then compared across groups to investigate the existence of a similar flow. Finally, a time series analysis is performed to test direct and indirect effects of the identified groups on the company’s outcomes. Results suggest the influence of media in defining predominant topics, but not necessarily in defining the semantic network construction. Based on these findings, we stress the need for further investigation on the processes of sentiment formation as it relates to topic generation in social media use.

!

Introduction

Recent developments in communication technologies have given rise to a society where meaning is constantly produced and consumed (Castells, 2011). The new possibilities and dynamics of web 2.0 (O’Reilly, 2007) paved the way for media convergence: a cultural shift which “alters the relationship between existing technologies, industries, markets, genres and audiences” (Jenkins, 2004, p. 33). This technological shift has called for a renegotiation of the

relationship between information producers and consumers (Jenkins, 2004), especially in the context of social media, characterized by the predominance of user-generated content (Bruns, 2007). In precisely this user-generated domain, the role of institutionalized media as a dominant information source begs to be questioned.

Both empirical research and recent current events underscore the importance of social media in the production and dissemination of influential information. In the emblematic Arab Spring, for example, the uprising against the established Government was largely organized through Twitter. In the concurrent wake of this political turmoil, international newspapers, TV channels and other traditional media depended on Twitter as a vital information source for their news reports (Lotan, Graeff, Ananny, Gaffney, Pearce & Boyd, 2011). Similarly, research on a devastating flood in Pakistan points to Twitter as an influential wellspring of information, identifying the tweets of Pakistani citizens as primary informants of news media (Murthy & Longwell, 2013).

In these two massive crises, Twitter’s impactful role may have been a function of the difficulties that traditional media faced in accessing information. While challenging social conditions posed an obstacle for journalists seeking to report on the events, the extensive

diffusion of user-generated information through social media led journalists to turn to Twitter as a primary information source. Given Twitter’s role in supplementing and even dominating the mediated information flow in crisis situations, we wondered whether similar patterns of information production might be present in social circumstances of a less urgent nature.

In order to investigate the interaction between traditional media reports and user-generated content on Twitter, we analyzed the dynamics of news consumption that helped form the social reputation of a major Italian bank.

RQ: To what extent and how do traditional media interact with Twitter users-generated

content to construct the reputation of a company in the online environment?

!

Theoretical Background

This study aims to better understand the processes by which Twitter users consume media news. In particular, we examine the process of belief formation on Twitter, insofar as it

contributes to the corporate reputation of a major Italian bank. To pursue this inquiry, we draw from two distinct bodies of literature within communication science: first, corporate reputation (Rindova, Pollock & Hayward, 2006) and second, media effects, with its distinct theories of two-step flow of communication (Katz, 1957) and agenda-setting (Carroll & McCombs, 2003).

In the literature, corporate reputation is defined as an intangible asset of a firm founded upon two crucial characteristics (Rindova, Pollock & Hayward, 2006, p. 50): the ability of the firm to “attract large scale public attention” (prominence) and the elicitation of “positive emotional responses from the public” (sentiment). Numerous studies acknowledge and support the importance of a positive reputation for a firm’s welfare (Bontis, Booker & Serenko 2007; Lange, Lee & Dai, 2011, Rindova, Williamson, Petkova & Sever, 2005).

This valuable intangible asset is assumed to be at least partially manageable by each given corporation (Lange et al., 2011; Walker, 2010). Hence, there exist numerous corporate

strategies designed to develop both prominence and positive public sentiment, including those that intervene with the production of reputation-pertinent news, such as press releases and press conferences.

For while corporations are commonly considered responsible for developing their reputations, the media are also presumed to play a substantial role in shaping public opinion, explaining the breadth of literature on media effects in communication science (e.g. Neuman & Guggenheim, 2011).

But within the media effects literature, researchers disagree on the underlying processes that drive media influence. The two-step flow model of communication (Katz, 1957) and agenda-setting perspective (McCombs & Shaw, 1972) represent different bodies of theory within the broader media effects literature. The former attends to the social context within which media messages circulate, attributing significance to the specific ways in which information is received, while the latter focuses on the salience, interpretation, and cognitive organization of the

presented message as a central predictor of the occurrence of media effects (Neuman & Guggenheim, 2011). Arguing that the two theories combined can best explain the extent of traditional media’s influence on Twitter conversations, we will draw from both in this paper.

Prior research examining the influence of traditional media on online users has yielded mixed results. Indeed, recent studies have found only partial effects of traditional media on the opinion formation of online populations (Johnson & Kaye, 2004; Murthy & Longwell, 2013). A common explanation for these findings is that the ease of production and consumption of information in social media might decrease user's dependence on traditional media (Anderson,

2006; Carpentier & De Cleen, 2007; Jenkins, 2004). An example is the aforementioned study of Twitter use in response to a natural catastrophe in Pakistan. The study identified Twitter users as representing a more impactful source of primary information as compared to traditional media (Murthy & Longwell, 2013). Similarly, a study on source credibility revealed that public opinion rated bloggers as more credible than traditional media (Johnson & Kaye, 2004).

On the other hand, there are also studies reporting significant effects for both traditional media and social media. In particular, it has been found that both sources of information

influence people’s opinion, however they appear to influence different attributes of the same construct (Bruhn, Schoenmueller & Schafer, 2012; Pfeiffer & Zinnbauer, 2010).

This paper confronts the discrepancies among previous findings on this topic, probing the interactive dynamics between traditional media and user-generated media. In particular, we investigate whether there is a significant flow of influence from traditional sources to Twitter users. To pursue this aim, we examine how traditional media news are disseminated, enriched, and elaborated upon by Twitter users, as well as how the underlying corporate reputation impacts the productive outcomes of our target company.

Two-Step Flow of Communication Theory

The first step towards understanding Twitter’s dynamics of information dissemination is to recognize its inherent social component. Like any social network, Twitter is characterized precisely by its interpersonal influence among users. In contrast to what some critics argue, existing studies demonstrate that affective relationships between people may not be

emotional links and reciprocal trust are enforced rather than diminished in online environments (Bruhn et al., 2012; Johnson & Kaye, 2004; Walther, 1996). Thus, there is reason to hypothesize that interpersonal relations play a role in the spread of influence in the Twitter context as well (Katz, 1957; Weimann, 1991).

The two-step flow of communication (Katz, 1957) identifies a narrow group of people that are perceived by their community as reliable diffusers of information and opinions. Other community members depend on these individuals to identify credible sources and select information deemed relevant for the community. These people are usually regarded as opinion leaders: they are part of the community they influence; they are influential only for specific topics; and they can be found at any social level (Weimann, 1991). In other words, according to two-step flow theory, in every community that holds a shared topic of interest, there exist identifiable opinion leaders.

In order to identify opinion leaders, one first has to define the community of interest in which the opinion leader operates. However, in the social media this may be a challenging task. While off-line communities are usually clearly structured and rather stable across time, in Twitter and the on-line environment in general, communities are transient and diffuse, with people gathering only temporarily around a topic of common interest (Arvidsson, 2013). Hence, it is not a user’s history or demographic features that give rise to their influential role, but rather the provisional pertinence of their tweets’ content. In other words, opinion leaders are central to the information flow precisely because of the information they share, and not because of any pre-existing credentials.

More broadly speaking, a traditional hierarchical structure diffuses information in the two-step flow from media to opinion leaders and from opinion leaders to the general public. However, this formulation is challenged by the social network structure, which spreads

information loosely and not through predefined hierarchies (Arvidsson, 2013; Colleoni 2012). In order to account for the interpersonal influence exercised in a community that is temporary, unstable, and network-structured, we propose an alternative means of identifying opinion leaders in Twitter. In particular, we define the community as time- and topic- specific, including all users participating in the discussion at any time point during the studied period. We examine the social structure among users to identify the primary sources of information,

pinpointing the actors most actively involved in selecting relevant information sources and spreading news along the network. In line with the two-step flow of communication theory, we expect to find traditional media among the primary sources of information.

H1: In Twitter, traditional media are mainly found among the primary sources of

information.

According to the two-steps flow of communication theory, by analyzing the network we expect to also find users not involved in media news production, but still central to the production and dissemination of information. We regard these influential users as Opinion leaders in the network.

Agenda-Setting Theory

To research influence flow, we need to investigate how content is spread and consumed among Twitter users. Given the opportunity to produce, consume and modify information on

Twitter (Anderson, 2006; Carpentier & De Cleen, 2007; Jenkins, 2004), users are able to revolutionize media content. Hence, we first pursue our theoretical construction by addressing the issue of whether and to what extent Twitter users recall such content. Subsequently, we shift our focus from actors to the content itself, drawing from the agenda-setting theory of media effects.

Agenda-setting theory interprets the influence of media on peoples in terms of saliency of discussed topics and their cognitive organization (Neuman & Guggenheim, 2011). Priming certain topics by discarding others is presumed to affect public perception of the relative importance of topics (Carroll & McCombs, Guo, 2013; 2003; McCombs & Shaw, 1972; Vargo, Guo, McCombs & Shaw, 2014). In other words, “[mass media] may not be successful much of the time in telling people what to think, but it is stunningly successful in telling its [audience] what to think about" (McCombs & Shaw, 1972, p. 177).

According to the theory, the perceived importance of topics is measured by first-level agenda-setting, which uses a topic’s prominence to compare agenda-setting of different groups. Second-level agenda setting deepens the analysis by focusing on the attribute of the topic combined with its perceived importance (Carroll & McCombs, 2003).

Recently, a third-level has been proposed to capture the centrality of topic in peoples’ discussions, as well as the cognitive links defining their semantic structure (Guo, 2013). Third-level agenda-setting questions the reliability of focusing exclusively on the prominence of topics, since this method is unable to account for contextual meaning. Indeed, the way topics are

Hence, including interlinks among topics may be necessary for a stronger and all-embracing explanation of the agenda-setting effect (Guo, 2013; Vargo et al., 2014). Implied in this argument is the idea that the agendas of two distinct groups can be considered equivalent only if they are identical in their topics’ network.

Flow of influence in terms of topics’ prominence.

In our first hypotheses we implicitly identified three groups of users with, presumably, three different agendas: media, opinion leaders and regular users. The former are represented by national newspapers, TV channels, and other institutionalized media. The seconds comprise actors whose main activity appears to be the identification of relevant sources of information and the consequent spread of relevant content across the network. The latter consist of Twitter users who are neither central in creating information nor influential in spreading it, as compared to other group members.

According to the two-step flow of communication theory, opinion leaders are thought to pass on information from media to regular users by selecting relevant news and/or by

commenting on topics and events important to them. In terms of agenda-setting, this should produce a similar agenda between media and opinion leaders. Supporting this assumed similarity, various studies in the literature reveal re-tweeting as one of the most popular activities among Twitter users (Bruns & Stieglitz, 2013; Jones, 2014; Kwak, Lee, Park & Moon, 2010; Murthy & Longwell, 2013). In particular, re-tweeting is most likely to take place when the tweet carries a link (Lindgren & Lundstrom, 2011; Jones, 2004), which is typically the case in the dissemination of a news report. On the other hand, the same studies recognize that re-tweet rates vary largely

depending on the topic (Kwak et al., 2010). Despite these preliminary findings, existing research has not examined which characteristics of content are associated with a tweet’s likelihood of being re-tweeted (Kwak et al., 2010; Wu, Hofman, Mason & Watts, 2011). This calls into question the implication that media content may be re-tweeted indiscriminately. On the basis of the reported studies, we cannot rule out the possibility that opinion leaders operate a heavy filter on media content. Indeed, keeping a long tail perspective in mind, we could assume that less popular news off-line may be gathered and made viral once brought online (Anderson, 2006). In other words, news that are primed by traditional media agenda may be discarded by Twitter users in favor of marginal news deemed more important to them. Increased accessibility to information introduced by web 2.0 has therefore augmented Twitter users’ autonomy from the agenda of traditional media, since they are free to actively select and comment on information they care about. Hence, in a third level agenda-setting comparison between media and opinion leaders, we expect the two groups to show similarities, but not identical agendas.

H2: Opinion leaders’ agenda is weakly correlated to media’s agenda.

Based on the two-step flow of communication theory, we also expect to observe the influence of opinion leaders’ agenda on regular users’ agenda due to the existing interpersonal influence linking the two groups. Indeed, opinion leaders are actively spreading information through the community, and their centrality in the information flow suggests that their tweets are largely shared by other users. Moreover, as discussed above, since online communities are defined in our study as people provisionally gathering around a given topic of common interest, tweets’ content, rather than users’ history and demographics, is assumed to be at the base of their influential role.

Thus, regular users tend to turn to opinion leaders, rather than to media, in identifying topics to discuss. Relatedly, according to agenda-setting theory, opinion leaders are central to the

information flow mainly because of the information they share and its role in influencing the topics discussed by regular users. Hence, in comparing the third-level agenda-setting of these two groups, we expect to observe a high level of overlap in topic, representing very similar agendas, which differ from that of traditional media. In sum, it follows that regular users’ third-level agenda will be weakly correlated to media’s third-third-level agenda.

H3: Regular users’ agenda is strongly correlated to opinion leaders’ agenda and weakly

correlated to media’s agenda.

Flow of influence in terms of topics’ sentiment.

With hypotheses two and three of this study, we address the influence of media in determining which news and topics about the target company attract large scale of public attention

(prominence). Another fundamental component of corporate reputation, however, is the public’s emotional response towards the company, i.e. sentiment (Rindova et al., 2006). Indeed, public sentiment appears to have a direct effect on the productive outcomes of the company (Aula, 2010; Bontis et al., 2007; Byrd, 2012; Pollock, Rindova & Maggitti, 2008).

Consequently, we examine the sentiment expressed towards the target company among the three distinct groups of actors. According to the two step-flow of communication, when opinion leaders take news from media and pass it on to other members of the community, they often enrich it with personal opinions and sentiments (Katz, 1957). The emotional content of tweets has been demonstrated to affect their popularity of tweets, as it increases attention and enlarges

influence (Wang, Wang & Zhu, 2013). Hence, we are led to hypothesize a mediator role of opinion leaders, for their function in spreading sentiment from media to regular users, and driving users’ sentiment in terms of direction (positive vs. negative) in regards to the topic discussed. On the other hand, there is limited evidence among existing research to support this hypothesis. Hence, we formulate a second research question addressing the mediation role of opinion leaders in the formation of sentiment towards the organization.

RQ2: To what extent do opinion leaders mediate sentiment towards the company as expressed

by media and regular Twitter users?

Corporate Reputation Effects on Bank’s Hard Outcomes

After investigating the influential flow from media to Twitter users, we now address the effectiveness of media in affecting a company’s hard outcomes both directly and indirectly through Twitter discussions. To this end, we examine to what extent corporate reputation, composed of prominence and sentiment, as constructed online by the three groups, (Rindova & Martins, 2012; Rindova et al., 2005) affects the company productive outcomes.

In accordance with two-step flow of communication theory, we expect traditional media to indirectly exercise their influence on the company’s outcomes through Twitter users. In line with previous hypotheses, we expect media to significantly impact opinion leaders’ expressed

sentiment, but only weakly influence their agenda. Similarly, we expect opinion leaders to strongly affect regular users’ agenda, but only partially affect their sentiment. In sum, traditional media are expected to have an indirect effect on company’s hard outcomes through opinion leaders’ expressed sentiment, which in turn has a weak influence on regular users’ sentiment. At

the same time, media are anticipated to have an indirect effect on the company’s hard outcomes through a weak influence on opinion leaders discussed topics, which in turns strongly affects regular users’ ones.

H4: Media have an indirect effect on company’s hard outcomes by weakly influencing

opinion leaders’ agenda, which in turn strongly influences regular users’ agenda.

H5: Media have an indirect effect on company’s hard outcomes by strongly influencing

opinion leaders’ sentiment, which in turn weakly influences regular users’ agenda.

!

Methods

This section introduces the data employed in this study, and describes how they were organized and prepared for the hypothesis-testing phase.

Sample Selection

In order to research the effect of media news on Twitter discussions, we compare news titles as asserted by media to tweets published by Twitter users. Using media titles rather than entire articles prevents any bias due to differing content lengths among the two sources, as there is a limit of 140 characters in a given text in Twitter technology. This constraint has the effect of condensing articulated thoughts into concise sentences, whereas media articles can express opinions in greater detail. We believe that this difference leads to distinct uses of the language, and consequently difference in keyword usage. Thus, news titles and tweets, which are similar in length, will be used as units of analysis: Mtitles= 98.5, SD = 29.5; Mtweets127.8, SD = 41.5. This

With respect to the data collection, both media news and tweets were included in the sample only when their content concerned the target bank directly. Moreover, no distinction was made based on whether the bank was the subject of the title/tweet or simply marginally

mentioned in it. Both titles and tweets were collected on a daily basis to cover a period of time of 230 days: from 16th May 2013 to 31st December 2013. The collection of media titles was

performed through LexisNexis database. News published by national Italian newspapers that mentioned the target bank’s name anywhere in its content was retrieved in the Italian language. Meanwhile, all tweets in the Italian language that mentioned the target bank anywhere in their body, and contained one or more hashtags, were collected in tandem. The tweet collection process was enabled by adapting the API provided by Twitter, and was performed by a business partner of the target bank. Data were available to the research team because the target bank served as sponsor of this project. Together with the tweets retrieved, the sponsor company also provided the data of bank accounts opened and closed by customers during the time period of analysis. Open/closure data of bank accounts covered the entire Italian territory for the time period of 16th May 2013 to 31st December 2013.

Definition of Social Media’s Experimental Groups

In order to test our hypotheses, we defined three groups to compare: Media, Opinion leaders and Regular users. Media titles, used to represent the media group, included all article titles retrieved through LexisNexis according to the process described in the previous section. As for the data from Twitter, we split the users, and consequently their tweets, into two groups: opinion leaders and regular users.

With the network analysis software Pajek, we constructed a network of Twitter users involved in conversations about the bank during the entire period of analysis. Links among users were defined on the basis of their reciprocal mentions (Twitter @ mentions). We used this directed network to perform the Hubs and Authorities analysis (Kleinberg, 1999). This analysis gives two scores to each actor- the authority score and the hub score- in order to measure one’s importance in the network. In order to be deemed an authority, a node needs to be pointed at by many good hubs; good hubs, in turn, are composed of actors pointing to many good authorities (Kleinberg, 1999). That is, actors are evaluated for the number of mentions they receive and for the quality of the mentioning source. The algorithm is based on the assumption that if user A mentions user B, he/she has, to some extent, conferred authority on B (Kleinberg, 1999, p. 2). We prefer this authority measure to the users’ in-degree value because the former takes into account the quality of the in-link (i.e. received mention). Unlike the in-degree value, Hub and Authority analysis does not consider all sources of mention to be equal, such that those of actors who themselves espouse high quality mentions (i.e. mentions from authoritative sources) are weighted more. Moreover, Hubs and Authority analysis gives two scores to each actor: one to measure his absolute authority on the topic, and the other to measure his ability to link to highly authoritative actors.

On the basis of these two measures, Pajek offers a classification of users into four distinct categories. The first category lists actors in the network who are recognized authorities in the investigated topic (i.e. the target bank). The second category comprises authorities who also function as important bridges linking other authoritative actors (i.e. they score high in both Hubs

and Authority measures). The third category encompasses actors who have not authority as primary information sources, but are hubs in the network that bridge important authoritative users. Finally, the fourth category groups together users who are not recognized as authoritative in the network, nor are they regarded as linking important authorities. Notice that this

categorization is made on two continuous variables and to obtain a categorization in four groups a threshold must be set. We decided to identify the top 10% of authoritative users, so we

instructed Pajek to code users below this threshold as neither authoritative nor important hubs. This boundary was selected on the basis of the Pareto rule as applied to social media (Bruns & Stieglitz, 2013), which affirms that the top 10% of most active and most visible users produce about 90% of the content (Bruns & Stieglitz, 2013).

The analysis shows a clear split among Twitter users into authoritative and not authoritative users. Several users in the dataset presented Twitter accounts of media, from national newspapers, TV channels, local newspapers, etc. As this study differentially represents the media group using news titles retrieved from LexisNexis, we opted to remove tweets of media accounts from the dataset. Preventing a biased report, this methodological decision assures that remaining Twitter users are not officially representing any established media. As for the remaining users, those who were categorized as authoritative actors by the Hubs and Authority algorithm (i.e. Authority, Hub&Authority or Hub) were coded in the Opinion leaders group. Lastly, users who did not belong to any authority category were coded as Regular users. Notably, Regular users may include Twitter accounts of politicians, companies and other established institutions, revealing that categorization is topic specific and heavily depends upon the specific

structure of the network investigated. One particular case involves Twitter accounts from the target company, mainly classifiable be authorities in the network. Media effects theories would relegate these as primary information sources forward the Media group (e.g. Westphal & Deephouse, 2011). However, testing the influence of the company on media news production is not our aim, so we have included company Twitter accounts in the Opinion leaders group according to their Hub and Authority score.

Topics Definition and Keywords Extraction

In order to identify the main discussion topics and respective keywords, we used the text analysis software Alceste, which groups textual data on the basis of similarity in language (Illia, Sonpar & Bauer, 2012). The software does not have a pre-existing list of words for the

categorization. Instead it garners categories from the present data and classifies titles into groups on the basis of co-occurring words (Illia et al., 2012). Hence, the resulting classes are similar for the words used and the topic addressed. In addition to the identification of main topics discussed, Alceste also returns a list of words most correlated to each identified class and the corresponding correlation score.

For better accuracy in topic identification, we split the title dataset into four groups of the same eight week time-span, excluding the last group, which covered nine weeks. This choice allowed Alceste to distinguish on the basis of topics discussed, rather than on their sources, which may differ in terms of language or tone (e.g. across left and right wing newspapers), and therefore bias our keyword extraction. The analysis lead to accuracy levels between 86% and

97%. Note that these values are far above the acceptance level of 70% recognized for Alceste (Illia et al., 2012). Across the four time groups, Alceste identified three major topics:

● Finance, which gathers all news related to the target company’s stock market shares, as

well as their fluctuation.

● Society, which includes all topics related to Italian society that pertain to the target bank.

● Trades, which refers to money loans furnished by the bank to support Italian companies.

For each identified topic, we extracted the top 30 words most correlated to it. To more accurately identify the topics, we reported only nouns, therefore excluding grammatical particles from the list, such as articles, adjectives and verbs. We found nouns to be more topic-specific, offering a clear distinction between categories. In the rare cases where an overlap was found, we assigned the contended keyword to the topic, according to its correlation with the Alceste score. A special case was made for the name of the target company, which was strongly correlated to each of the topics identified. For this reason we decided to exclude it from our analysis, with the exception of the third-level agenda setting test, in which the company name was essential to the construction of the semantic network. Meanwhile, the other analyses in the study would be biased by inclusion of this keyword, given the frequent presence of the bank’s name in the dataset, as compared to the other keywords. As we will explain in more detail shortly, other analyses are based on the absolute frequency count of keywords mentioned (i.e. prominence) or on the average sentiment expressed (i.e. sentiment). Both measures are sensitive to the over-representation bias associated with the bank’s name high presence.

In order to construct the semantic network, we combined keywords with each title and tweet that mentioned them, since keywords rather than titles/tweets were central to our analysis. If a keyword was mentioned in the title/tweet, then it was coded under the keyword itself and its representative category. If more than one keyword was found in the same title/tweet, then it was duplicated so that each row in the dataset contained only one keyword. We also used media news’ keywords for the categorization of twitter groups. This helps to make the three groups comparable, as it favors the identification of a flow of information from media to twitter users. In other words, the use of the same keywords and topics across groups allows us to investigate the extent to which Twitter users recall the media’s agendas in their discussions.

After coding each title and tweet with keywords and their related categories, we constructed a semantic network for each group. We first constructed a two-mode network, using both keywords and titles/tweets as nodes, and then used Pajek software to create a one-mode network, in which keywords served as the only nodes. This procedure enabled us to link keywords depending on whether they were simultaneously used in the same title/tweet. Additionally, the weight of each link indicates the frequency (i.e. number of titles/tweets) with which two keywords are used together. These constructed semantic networks will enable a comparison of third-level agenda-setting between the three groups (Vargo et al., 2014).

Sentiment Measure

Titles- and tweets- sentiment was coded on a three-levels scale ranging from negative to positive (-1 to 1). A priori, we coded sentiment according to both the tone applied by the user and the coverage of the news itself (Pang & Lee, 2008; Pak & Paroubek, 2010). For example, a tweet

was coded negative if the user displayed a negative opinion towards the bank (e.g. an insult to the bank customer service), or if the tweet contained negative news about the bank (e.g. a loss of several million euros by the bank in the first half of the year). This operationalization of

sentiment captures both the expressed tone and the decision to spread negative- rather than positive- news (Rindova et al., 2006).

Due to the high number of data to categorize, we created a machine learning software to aid in codification of tweets’ sentiment. The applied machine learning is based on the Passive-Aggressive algorithm (PA) (Crammer, Dekel, Keshet, Shalev-Shwartz, & Singer, 2006), and implemented with a Pairwise Coupling with majority voting method (Hastie & Tibshirani, 1998), in order to account for multi-class categorization. Three individuals coded the same 1459 tweets, yielding a good level of intercoder reliability (Kalpha= .81, p= .026). On the basis of the three coders’ training, the machine learning software coded the total 14179 tweets, producing a satisfactory reliability level: recall 0.62, precision 0.71, f-measure 0.69 and balanced-accuracy 0.8. Unfortunately, the same machine learning process could not be used for the sentiment of titles because of the different grammatical constructions and symbolisms used across the two sources (e.g. absence of mentions and hashtags in news titles). Consequently, the titles from the media group were manually coded by a single coder using the Coding Manual that was applied to the tweets (Appendix A).

Average sentiment expressed was found to differ among groups: Media (M= .04, SD= . 08), Opinion leaders (M= -.08, SD= .19), and Regular users (M= -.17, SD= .30). In preparation

for the proceeding analysis, we coded the sentiment of each experimental group on a daily basis, enabling the performance of a Time Series analysis.

Prominence Measure

To measure Prominence of discussed topic across the three groups, we counted the frequency with which each keyword was mentioned by each experimental group. The frequency count was calculated on a daily basis and corresponded to the absolute number of keywords, representing the number of references made daily to the target bank. This measure of prominence does not have a delimited scale. Expressed prominence in Media group was the highest on average (M= 27.11, SD= 23.52), followed by Regular users (M= 5.43, SD= 7.86) and Opinion leaders (M= 4.54, SD= 5.99).

Open/closure index

In order to create a daily index of the rates by which bank accounts opened and closed, we used the daily data provided by the bank. Starting from the daily absolute number of

openings and closures, we subtracted the closures from the openings in order to obtain the index of interest. The index does not have a delimited scale and includes both positive and negative values (M= - 0.47, SD= 7.86). In particular, negative values indicate a higher number of closures compared to openings processed by the bank in a specific day.

!

This section reports the results of the analyses performed to investigate the study’s hypotheses. For each hypothesis testing we will describe the analysis performed, as well as the obtained results.

Primary Sources of Information: Hub and Authority Analysis

In order to test the first hypothesis of the study, we constructed an actor’s network consisting of all users tweeting about the target bank in the period of analysis: 3632 users in 230 days. We constructed a directed one-mode network, based on the mentions among users (i.e. Twitter’s @ mentions) and ran a Hub and Authority analysis (Murthy & Longwell, 2013), with a threshold set at the top 10% (i.e. top 363 users) (Bruns & Stieglitz, 2013). The analysis returned a categorization of all users into four categories: Authority, Hub&Authority, Hub, and no

Authority level. The resulting crosstab (Table 1) shows that media are most present in the authority category, and represent less than 1% of users who do not have a central role in the spreading of information (i.e. “none” Authority category in Table 1). Finally, the chi-square test shows a significant correlation between media membership and authority category affiliation: X2_{= 140.57, p< .001. According to the analysis, hypothesis one of the study is supported.}

Table 1 - Crosstable between Media membership and Authority category

  Authority category (Pajek)

None Authority Hub&Authority Hub

Media membership No media 2959 257 71 265 99,10% 90,80% 88,80% 93,60% Media 27 26 9 18 0,90% 9,20% 11,30% 6,40% Chi-square 140.57 p< .001 Percentages in columns

!

Third Level Agenda-Setting: QAP Test

In order to test hypotheses two and three of the study, we compared the semantic networks developed by each group over the course of the 230 days analyzed. In particular, we created a semantic network for each group and then compared them with the QAP test (Vargo et al., 2014). The test determines the correlation between two networks on the basis of two

proprieties of links: the absence or presence of a link between two nodes, and its weight (Vargo et al., 2014). In order for this analysis to be reliable, all compared networks have to include the same list of keywords, whereas in table B2, some keywords that appear in the Media group are missing from both the Opinion leaders group (11 words) and the Regular users group (19 words). For the purpose of the current analysis, we added the missing words without links to the other keywords in the group, enabling an appropriate comparison among groups without biasing the results regarding the existence of links between keywords.

The visual representations of semantic networks show a substantial difference between the Media group and the two Twitter groups (Figures B1, B2 and B3), which is verified by the QAP test. Here, Media and Opinion leaders show a weak and significant correlation (r= 0.22, p= .004), supporting our second hypothesis, while Opinion leaders and Regular users show a strong and significant correlation: r= 0.78, p< .001. Finally, we use the QAP test to calculate the correlation between Media and Regular Users: r= 0.13, p= .003, demonstrating a correlation weaker than that between Media and Opinion leaders. This finding suggests that the Opinion Leaders’ group mediates the semantic networks of media and regular users, supporting our third hypothesis.

Sentiment comparison: One-Way Analysis of Variance (ANOVA)

In order to investigate the mediating role of Opinion leaders between Media and Regular users in terms of sentiment, we compared levels of this construct across the three groups with a one-way analysis of variance (ANOVA) (Kiousis, Bantimaroudis & Ban, 1999). The ANOVA test treated group membership as an independent variable, and sentiment as a dependent variable. Results of the test were significant (F(2,234)= 21.00, p< .001) with the Media group scoring the highest (M= .04, SD= .08), followed by Opinion Leaders (M= -.08, SD= .19), and then Regular users (M= -.17, SD= .30). The post hoc Bonferroni test shows a significant difference across the three groups in terms of sentiment: Media and Opinion leaders (Mdif= .12, p< .001), Opinion

leaders and Regular users (Mdif= .08, p= .041), and Regular users and Media (Mdif= .20, p< .

001). Additionally, we tested the sentiment towards the three topics separately, as suggested in the existing literature (Kiousis et al., 1999).

The test on the Financial topic yielded significant results: F(2, 81)= 13.79; p<.001. In particular, Media group scored the highest (M= 0.04; SD= .05), followed by Opinion leaders group (M= -0.08; SD= .14), and Regular users (M= -.24; SD= .31). The post hoc Bonferroni test showed no significant difference between Media and Opinion leaders’ groups (Mdif= .11; p= .

066). On the contrary, Regular users show a sentiment mean that is significantly different from both Opinion leaders (Mdif= .16; p= .009) and Media (Mdif= .27; p< .001).

As for the Trades, the one-way analysis of variance (ANOVA) test is significant overall F(2, 65)= 4.52; p= .015. Media group expressed the most neutral sentiment (M= 0.01; SD= .04), followed by opinion leaders (M= -.16; SD= .26), and regular users (M= -.18; SD= .39). The post

hoc Bonferroni test shows that Opinion leaders do not significantly differ from either Media group (Mdif= .17; p= .067) or Regular users (Mdif= .03; p= 1). On the other hand, regular users

differ significantly from Media group in this domain (Mdif= .19; p= .031).

Finally, an one-way analysis of variance (ANOVA) between the three groups on Society topics revealed significant results: F(2, 82)= 7.92; p= .001. The statistics reported show that Media group scores the higher (M= .07; SD= .11), followed by Opinion leaders (M= -.04; SD= . 17), and regular users (M= -.09; SD= .20). The post hoc Bonferroni test shows that Media differ significantly from Opinion leaders (Mdif= .11; p= .035) and Regular users (Mdif= .11; p= .001),

whereas Opinion leaders and Regular users group show no difference (Mdif= .06; p= .574).

Overall, these results suggest that influence on sentiment among the groups depends on the topic. Media and Opinion leaders’ groups are not distinct in the expressed sentiment, apart from the Society topic, whereas Opinion leaders and Regular users show a significant difference for the Financial topic. Finally, Media and Regular users express significantly different sentiment in all the three topics analyzed. Hence, these results generally support the mediation role of Opinion leaders between Media and Regular users’ groups. In particular, Opinion leaders seem to direct the sentiment of Regular users, which in turn make it more extreme. However, we also noticed that the inter-groups behavior heavily changes depending on the topic. Future research on sentiment should investigate the three groups’ behavior, as separated per topic.

Effect on Company’s Hard Outcomes: Time Series Analysis

In order to investigate the effect of corporate reputation (prominence and sentiment) as formed on Twitter in regards to the productive outcomes of the target company, we perform a

time series analysis. We test two separate VAR models to investigate the two constructs of corporate reputation: prominence and sentiment. Each model will include four variables: one per group plus the open/closure index.

VAR model of prominence.

Before running the VAR model, we check the time series of the four variables to exclude any problems with the data that may interfere with the reliability of the analysis. Figure C4 reports the time series distribution of Media prominence (tm_prom), Opinion Leaders’

prominence (ol_prom), and Regular users’ prominence (ru_prom), plotted together. The figure does not show evidence of any concerns. We also checked the time series for the Open/closure index (Figure C5) in a separate graph because the divergence in scales among the variables would have prevented a clear interpretation of the time series. As for the variables in the model, the Open/closure index time series does not appear to present any concerns in its distribution. However, we decide to perform the Augmented Dickey-Fuller test to rule out the null hypothesis that the variables present a unit root. Test statistics calculated by the Augmented Dickey-Fuller are all beyond the critical value of -2.88, which allows us to reject the null hypothesis for all variables in the model: ADFMediaP= -11.340, p< .001, ADFOp.leadersP= -8.033, p< .001,

ADFReg.usersP= -9.976, p< .001, ADFo/c.indexP= -12.144, p< .001.

In order to select the appropriate lag length for the VAR model, we inspect the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) for VAR models with lag from 1 to 8 (Table C3). The two indices describe the improvement in the log-likelihood, which is penalized by each additional lag. Due to the fact that both indices are based on the

negative of the log-likelihood, lower values of these fit statistics are preferred (Brandt &

Williams, 2007). As a consequence, we chose to perform a 4-lags VAR model, since the low AIC and BIC statistics infer a better fit to the data at this time lag. The model selected includes four endogenous variables: Media prominence, Opinion leaders’ prominence, Regular users’

prominence, and Open/closure index.

Results from the Granger Causality test show that Media prominence significantly affects Opinion leaders’ prominence (X2_{= 13.61, p= .009), but not that of Regular users (X}2_{= 9.12, p= .}

058). In addition, Opinion leaders’ prominence significantly affects the prominence of the Regular users’ discussions (X2_{= 23.08, p< .001). Finally, the index of open/closure of bank}

accounts is significantly influenced by both Opinion leaders (X2_{= 10.85, p= .028) and Regular}

users (X2_{= 11.24, p= .024), whereas it is not significantly influenced by Media prominence (X}2₌

6.84, p= .144). Interestingly, the Open/closure index is also shown to have a significant effect on the Twitter discussions of both Opinion leaders (X2_{= 38.79, p< .001) and Regular users (X}2₌

45.15, p< .001).

By inspecting the graphs from the Impulse Response Function (Figure C6), we note that a shock in Media prominence appears to cause a weak increase in corporate prominence among Opinion leaders on the first day after the shock. Much stronger and durable is the effect of Opinion leaders on corporate prominence among Regular users. Indeed, the IRF graph shows that a shock in corporate prominence among Opinion leaders causes an increase of prominence among Regular users for the first four days after the shock. Moreover, both Opinion leaders’ and Regular users’ discussions are shown to have an effect of Open/closure index on the second day

after a shock on both groups. However, while a shock in Opinion leaders prominence has a negative effect on Open/closure index, a shock among Regular users has a weak but positive effect on the index. Finally, a shock of the Open/closure index causes a decrease on corporate prominence on discussions of both Opinion leaders and Regular users for the first two days after the shock.

Subsequently, we investigate the residuals’ correlation between variables in the model through the decomposition of forecast error variance. In particular, the analysis reports the variance of the forecast error as decomposed in percentages, representing the forecast variance attributable to other variables in the model. Table C5 reveals that the size of effects among variables in the model of prominence is sizeable. For example, at lag-8 about 12% of the variance in Opinion leader group is attributable to Media group and more than 15% to Open/ closure index. As for the variance of Regular users, at lag-4 nearly 35% of variance is

represented by Opinion leaders and an additional 8% by the Open/closure index. Moreover, from lag-4 the relative size effect of both variables on Regular users variance continues to increase steadily, with the variance represented in the Open/closure index increasing more than that of Opinion leaders. However, at lag-8 the former is still about 20 percentage points below the variance explained by the latter. Finally, we note that the variance of the Open/closure index at lag-8 is significantly accounted for by the three groups, at a level of about 3% each. Much higher is the percentage of variance among both Opinion leaders and Regular users: at lag-8 the

variance explained reaches about 15% in both groups, according to the index. Furthermore, while the variance of both Media group prominence and the Open/closure index is mainly determined

independently, the effect of other variables among both Opinion leaders’ and Regular users’ prominence is clearly much stronger. This is particularly true for the Regular users’ group, whose variance at leg-8 is only 42% affected by itself.

These results indicate that there is an effect of media on Opinion leaders’ prominence, and also an effect of Opinion leaders on Regular users, while there is no significant effect of Media on regular users. Moreover, both Opinion leaders and Regular users have an effect on the Open/closure index, while Media group does not. This finding offers clear support for hypothesis four of this study.

VAR model of sentiment.

In order to prepare our variables for the Time Series analysis, we draw the time series of each along the period of analysis (Figures C5 and C7) (already discussed for the Open/closure index). In Figure 7, we see the variables Media sentiment (tm_sent), Opinion leaders’ sentiment (ol_sent) and Regular users’ sentiment (ru_sent) plotted together. There seems to be no evidence of series moving in the same direction, evidencing no apparent problem with the data. The Augmented Dickey-Fuller test is performed to further disconfirm the null hypothesis of a presence of unit root in the data (Brandt & Williams, 2007). The test shows that no variables in the model appear be affected by unit root. Indeed, their test values all fail beyond the critical value of -2.88: ADFMediaS= -14.256, p< .001, ADFOp.leadersS= -14.348, p< .001, ADFReg.usersS=

-14.155, p< .001, ADFo/c.indexS= -12.144, p< .001.

Next, we select the lag length of the VAR model. In Table C6, we report the values for the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) for VARs

with 1 to 8 lags. The two statistics describe the improvement in the model at the increase of lag. In particular, the smaller the values from both statistics, the better the fit of the model (Brandt and Williams, 2007). Lag 1 seems to be a good choice for the model since both AIC and BIC are at their lowest values. Accordingly, we performed a 1-lag VAR model with Media sentiment, Opinion leaders’ sentiment, Regular users’ sentiment, and Open/closure index as endogenous variables.

The Granger Causality test reported in Table C7 shows only a significant effect of Opinion leaders’ sentiment on Regular users sentiment (X2_{= 8.73, p= .003). Consequently, we}

inspect the graphs from the Impulse Response Function (Figure C8) to observe the effect of Opinion leaders’ sentiment on Regular users’ sentiment along the time. The graph indicates that Regular Users respond positively to a shock in Opinion leaders’ sentiment on the first day after the shock. However, the response appears to be very weak, and decays the second day. In other words, an increase in bank-sentiment among Opinion leaders discussions causes an increase in bank-sentiment within Regular users discussions. However, this effect appears to be weak and short in duration.

Finally, we investigate the decomposition of the Forecast Error Variance. This analysis uses the divergence between the estimated model and the actual values of the vector of

endogenous variables, to determine the level of interdependence of variables’ errors. From Table C8, we see that variables’ variances are mainly self-determined, and not strongly affected by other variables in the model. An exception exists within the Regular users group, whose variance

at lag-2 is explained by Opinion leaders at a total percentage of 13.50%. However, this value stays stable after this point in time.

On the basis of these results, we see no effect of Media on Opinion leaders’ or Regular users’ sentiment. Similarly, since there is no evidence of an impact of the three groups on the company’s outcomes, we reject hypothesis five of the study.

!

Discussion

This paper successfully addresses the need for a deeper comprehension of influential dynamics between offline media and Twitter discussions. The split of communicative actors into three distinct groups, as supported by the two-step flow of communication theory, has shown to be useful in describing the flow of information from Media to Twitter Regular users, as mediated by Opinion leaders. On the basis of agenda-setting theory, the influential flow was investigated in terms of topics discussed and then enriched with a comparison of related sentiment, so as to capture the construct of corporate reputation (Rindova et al., 2006). The effects of corporate reputation, described as the joint effect of prominence and sentiment, were tested on corporate hard outcomes. The time series analysis revealed evidence of an effect of Twitter discussions on company outcomes, as well as vice versa. Next, we report the main contributions of the study to the existing literature in the field.

In accordance with the two-step flow of communication (Katz, 1957), this study supports the existence of an information flow from media to Twitter users as mediated by opinion leaders. In particular, media are shown to affect the perceived importance of topics discussed (McCombs

& Shaw, 1972), while the increased independence of social media users in the creation and spread of information (Bruns, 2007) appears to manifest in the construction of semantic network of topics. Hence, we have found support for the theory of agenda-setting, insofar as it concerns the centrality of topics as addressed by the first level of agenda-setting (McCombs & Shaw, 1972). At the same time, we find evidence of a renegotiated relationship between information producers and consumers (Jenkins, 2004) in comparing the third-level agendas of the

experimental groups. In discussing the very same topics, Twitter users organize themselves in a distinct semantic network as compared to media.

The second major contribution of the study is the adoption of the recently proposed third-level agenda-setting (Guo, 2013) in comparing semantic networks as constructed by media and Twitter users. This approach is successful in describing the correlation between groups’ semantic networks. We thus encourage the embrace of third-level agenda setting as a means of integrating first and second levels (Carroll & McCombs, 2003), given its ability to account for contextual meaning that leads to a stronger and more all-embracing explanation of the agenda-setting effect (Guo, 2013; Vargo et al., 2014).

The final contribution of the study is the support offered to the literature on corporate reputation and its effects on company outcomes. By applying the definition of corporate reputation as articulated in prominence and sentiment towards the company (Rindova et al., 2006), this study describes the effects of reputation as created online on a company’s productive outcomes. In particular, we successfully demonstrated the effect of prominence on company outcomes, but failed to find support for a similar effect of the sentiment construct. In addition,

our findings suggest that the sentiment construct heavily depends on the discussed topic. Further research should investigate its effect on company outcomes by accounting for the specific topic discussed.

In addition to these contributions, the study also presents some limitations and opportunities for future research. Together with an effect of Twitter discussion on company outcomes, our results also reveal a converse influence of company performance on Twitter discussions. In particular, company outcomes appear to decrease company prominence among Twitter users’ discussions. Moreover, the effect of company outcomes on Twitter discussions appears to be stronger than the inverse effect. This interesting finding may be due to word-of-mouth dynamics among the company’s customers. We might hypothesize that changes in customers’ satisfaction towards the bank prompt word-of-mouth activity, which in turn might spread online through Twitter social media. However, the study we have performed was not appropriately constructed to investigate this phenomenon, meriting further study on this topic.

!

References

Anderson, C. (2006). The long tail: Why the future of business is selling less for more. New York: Hyperion

Arvidsson, A. (2013). The potential of consumer publics. Ephemera: Theory \& Politics in

Organization, 13(2), 367–391.

Aula, P. (2010). Social media, reputation risk and ambient publicity management. Strategy

Bontis, N., Booker, L. D. & Serenko, A. (2007). The Mediating Effect of Organizational Reputation on Customer Loyalty and Service Recommendation in the Banking Industry. Management Decision 45, 1426–1445.

Brandt, P., & Williams, J. T. (2007). Multiple Time Series Models. Thousand Oaks: Sage Publications.

Bruhn, M., Schoenmueller, V., & Schäfer, D. B. (2012). Are Social Media Replacing Traditional Media in Terms of Brand Equity Creation? Management Research

Review 35(9): 770–790.

Bruns, A. (2007). Produsage. Paper presented at the 6th ACM SIGCHI conference on Creativity \& cognition (C\&C '07), Washington, DC.

Bruns, A., & Stieglitz, S. (2013). Towards more systematic Twitter analysis: Metrics for tweeting activities. International Journal of Social Research Methodology, 16(2), 91–108.

Byrd, S. (2012). Hi fans! Tell us your story!: Incorporating a stewardship-based social media strategy to maintain brand reputation during a crisis. Corporate Communications: An

International Journal, 17(3), 241 - 254

Carpentier, N., & De Cleen, B. (2007). Bringing Discourse Theory into Media Studies.

Journal of Language and Politics, 6(2), 265–293.

Carroll, C. E. & McCombs, M. (2003). Agenda-setting effects of business news on the public’s images and opinions about major corporations. Corporate Reputation

Review, 6(1), 36–46.

Colleoni, E. (2012). New forms of Digital Marketing Research. In Belk, R. & Liama, R. (Eds.). Companion to Digital Consumption (pp. 124-134). London: Routledge. Crammer, K., Dekel, O., Keshet, J., Shalev-Shwartz, S. & Singer, Y. (2006). Online

passive-aggressive algorithms. Journal of Machine Learning Research, 7, 551-585.

Guo, L. (2013). Toward the third level of agenda setting theory: A network agenda setting model. In Johnson, T. (Eds.). Agenda setting in a 2.0 world: New agendas in

communication (pp. 112–133). New York, NY: Routledge.

Hastie, T. & Tibshirani, R. (1998). Classification by pairwise coupling. The Annals of

Statistics, 26(2), 451-471.

Illia, L., Sonpar, K., & Bauer, M.W. (2012). Applying Co-occurrence Text Analysis with ALCESTE to Studies of Impression Management. British Journal of Management,

25(2), 352-372.

Jenkins, H. (2004). The cultural logic of media convergence. International Journal of

Cultural Studies, 7(1), 33-43

Johnson, T. J., & Kaye, B. K. (2004). Wag the blog: How reliance on traditional media and the Internet influence credibility perceptions of Weblogs among blog users.

Journalism & Mass Communication Quarterly, 81(3), 622-642.

Jones, J. (2014). Switching in Twitter's Hashtagged Exchanges. Journal of Business and

Technical Communication, 0, 1-26.

Katz, E. (1957). The two-step flow of communication: un up-to-date report on an hypothesis. Public Opinion Quarterly, 21, 61-78.

Kiousis, S., Bantimaroudis, P. & Ban, H. (1999). Candidate Image Attributes: Experiments on the Substantive Dimension of Second Level Agenda Setting. Communication

Research, 26(4), 414-428.

Kleinberg, J.M. (1999). Authoritative Sources in a Hyperlinked Environment. Journal of the

Association for Computing Machinery, 46(5), 604-632.

Kwak, H., Lee, C., Park, H., & Moon, S. (2010). What is twitter, a social network or a news

media? Paper presented at the tenth World Wide Web Conference. Raleigh, NC.

ACM Press, 2010.

Lange, D., Lee, P. M. & Dai, Y. (2011). Organizational reputation: a review. Journal of

Management, 37(1), 153–184.

Lindgren, S., & Lundström, R. (2011). Pirate culture and hacktivist mobilization: The cultural and social protocols of #wikiLeaks on Twitter. New Media & Society 13(6), 999-1018.

Lotan, G., Graeff, E., Ananny, M., Gaffney, D., Pearce, I., & Boyd, D. (2011). The

revolutions were tweeted: Information flows during the 2011 Tunisian & Egyptian revolutions. International Journal of Communication, 5, 1375–1405.

McCombs, M. E., & Shaw, D. L. (1972). The agenda-setting function of mass media. Public

Opinion Quarterly, 36(2), 176–187.

Meraz, S. (2011). Using time series analysis to measure intermedia agenda-setting influence in traditional media and political blog networks. Journalism & Mass Communication

Quarterly, 88(1), 176-194.

Murthy, D., & Longwell, S. A. (2013). Twitter and disasters: The uses of Twitter during the 2010 Pakistan floods. Information, Communication & Society, 16, 837–855.

Neuman, W. R., & Guggenheim, L. (2011). The evolution of media effects theory: A six-stage model of cumulative research. Communication Theory, 21, 169–196.

O'Really, T. (2007). What is Web 2.0: Design Patterns and Business Models for the Next Generation of Software. Communications & Strategies, 1, 17-37

Pak, A. & Paroubek, P. (2010). Twitter as a Corpus for Sentiment Analysis and Opinion

Mining. Paper presented at the seventh meeting of the European Language

Resources Association, Valletta, Malta.

Pang, B. & Lee, L. (2008). Opinion mining and sentiment analysis. Foundations and

Trends® in Information Retrieval, 2(1–2), 1-135.

Pfeiffer, M. & Zinnbauer, M. (2010). Can old media enhance new media? How traditional advertising pays off for an online social network. Journal of Advertising Research,

50(1), 42-49.

Pollock, T. G., Rindova, V. P. and Maggitti, P. G. (2008). Market watch: Information and availability cascades among the media and investors in the US IPO market. Academy

of Management Journal, 51(2), 335–358.

Rindova, V. P. & Martins, L. L. (2012). Show me the money: A multidimensional

perspective on reputation as an intangible asset. In Barnett, M. L. & Pollock, T. G. (Eds.). The Oxford handbook of corporate reputation (pp. 16-33). Oxford: University Press.

Rindova, V. P., Williamson, I. O., Petkova, A. P., & Sever, J. M. (2005). Being good or being known: An empirical examination of the dimensions, antecedents, and consequences of organizational reputation. Academy of Management Journal, 48, 1033-1049.

!

!

Rindova, V. P., Pollock, T. G. & Hayward, M. L. (2006). Celebrity firms: The social construction of market popularity. Academy of Management Review, 31(1), 50–71. Vargo, J.C., Guo, L., McCombs, M. & Shaw, D.L. (2014). Network Issue Agendas on

Twitter During the 2012 U.S. Presidential Election. Journal of Communication,

64(2), 296-316.

Walker, K. (2010). A Systematic Review of the Corporate Reputation Literature: Definition, Measurement, and Theory. Corporate Reputation Review, 12(4), 357-387.

Walther, J.B. (1996). Computer-mediated communication: impersonal, interpersonal and hyperpersonal interaction. Human Communication Research, 23, 3-43

Wang, C.J., Wang, P.P., Zhu, J.J.H. (2013). Discussing occupy wall street on Twitter:

longitudinal network analysis of equality, emotion, and stability of public discussion.

Cyberpsychology, behavior and social networking, 16(9), 679-85.

Weimann, G. (1991), The Influentials: Back to the Concept of Opinion Leaders. Public

Opinion Quarterly, 55, 267–79.

Westphal, J. D., & Deephouse, D. L. (2011). Avoiding bad press: Interpersonal influence in relations between GEOs and journalists and the consequences for press reporting about firms and their leadership. Organization Science, 22, 1061-1086.

Wu, S., Hofman, J. M., Mason, W. A. & Watts, D. J. (2011). Who says what to whom on

!

Appendix A – Codebook for sentiment analysis

!

Scale range: from -1 to 1 -1: negative sentiment

0: neutral sentiment +1: positive sentiment

!

Each title or tweet coded has to show an unique sentiment score ranging from -1 to +1. Moreover, the content of the analysis is the target company. It implies that sentiment is coded on

the subject of the title/tweet, when this is the target bank. Similarly, sentiment is coded on the object of the title/tweet when the target bank is the object of the sentence. In the cases where the

target bank neither subject nor object of the title/tweet, the sentiment is coded neutral due to the marginality of the target bank to the discourse.

Negative sentiment coding:

A title/tweet is coded with negative sentiment if is clearly expresses a criticism towards the target company. The criticism may assume the forms of negative news about the company (e.g. revenue loss) and/or negative tone towards the bank. In other words, both negative coverage and negative

opinions are coded as negative sentiment.

Hints for coding negative sentiment are: use of insults towards the bank, negative appellations to refer to the bank (e.g. robbers), opinionated interpretation of an event involving the bank which

expresses a negative feeling towards the bank. Examples:

• RT @Cantiere: Soldi alle persone non alle banche Sanzionata #(target bank’s name) che devasta i territori #sollevazione #fuckausterity #30N http://t.c

• Adisu e (target bank’s name), rendetemi ci che mi spetta!!!! #borsadistudio

• RT @_MagliaNera_: #(target bank’s name) 6.000 esuberi. Anche i furti ai nostri danni non bastano pi per mantenere un numero eccessivo di lavoratori.

Neutral sentiment coding:

A title/tweet is coded as neutral if the information reported is neither positive nor negative towards the bank. It may assume two forms. First, the news is neutral in itself when it reports

events lacking any positive or negative inference (e.g. opening of a new branch of the bank). Second, the news is neutral when the bank is nor the subject or the object of the title/tweet. In

other words, the sentiment is not expressed towards the bank, so it cannot be coded. Hints for coding neutral sentiment are: the bank is not an active actor in the title/tweet (e.g. bank’s branch used as geographical reference), the news is objective and reports information without clear negative or positive impact on the bank (e.g. transfer of a bank’s office to a new

location). Examples:

• #(target bank’s name) #basket premio Retina doro a novembre siete tutti invitati #lasquadrachenoncera #roma. Ci sar anche @bancoroma

• Da #(target bank’s name) sette #certificati bonus cap legati a indici azionari... http://t.co/ EC82wakksA

• Ma lo spot radio della #cna quasi uguale a quello di. (target bank’s name) di qualche tempo fa

Positive sentiment coding:

A title/tweet is coded with positive sentiment if clearly reports good news about the company. This may assume two forms. First, the news is objectively positive for the bank (e.g. it won a trial). Second, the personal interpretation of the news is positive towards the bank (e.g. overture of a new bank’s branch is commented with enthusiasm). The latter may also include promotional

material spread by the bank itself.

Hints for coding positive sentiment are: positive enthusiasm and exhortations to act in favor of the bank (e.g. exhortation to visit the last floor of the new bank’s skyscraper), promotional messages about bank’s products and services (e.g. new bank account with low interests), use of

smiles to convey approval (e.g. “☺” and “;)”). Examples:

• #(target bank’s name) assume 500 giovani. Il sindacato: Finalmente una notizia positiva #lavoro http://t.co/dTePyzZBGF

• Domenica attiva!Si parte w/ @citroenitalia alla volta del 25imo piano della @(target bank’s name)_PR Tower!! #DS3CabrioTour http://t.co/J3ABxpndeR

• @Milanosecrets: #milano #(target bank’s name) http://t.co/wJxzqterAgmitica Manu! Perfetta per la mia collezione :)

!

!

!

Appendix B – QAP test

!

Table 2 - Keywords list per category and keywords frequency count per experimental group

Topic category Keywords Media group Opinion leaders

group Regular users group

Finance attes 192 23 18 autog 114 4 missing banc 1814 418 715 bors 2021 94 85 bpm 208 7 1 calo 255 24 4 carige 67 7 6 corr 343 42 74 debol 354 12 missing europ 795 54 52 exor 55 3 missing ferragamo 64 1 missing fiat 406 14 11 finmeccanica 225 8 2 ftse 145 32 12 mediaset 494 15 14 milano 1741 106 255 mps 367 94 170 peggio 106 9 13 popolare 124 9 8 positiv 229 13 12 rialz 283 27 3 saipem 153 14 12 scivol 144 6 missing stm 100 9 2

street 128 3 missing telecom 551 34 50 tod 79 6 15 usa 263 18 44 wall 130 3 3 aziend 51 38 37 Society bond 191 34 11 calcio 78 84 74 credit 247 114 92 crescere 31 9 9 crisi 209 33 36 econom 66 182 74 euro 1037 145 145 famigl 48 32 24 fond 182 58 56 garan 56 13 20 ghizzoni 433 119 98 giovan 30 37 29 governo 119 20 30 grupp 105 21 28 imprendi 44 4 10 imprese 96 83 72 intervista 30 3 7 invest 128 63 122 italia 405 189 308 milion 187 82 123

minozzi 15 missing missing

nicastro 44 21 14

prestit 59 87 29