Issue Supremacy: The influencer role of organizations in online social networks Katerin Kunfermann
Master Thesis
Graduate School of Communication Universiteit van Amsterdam
Author Note
Student Number: 12225223
Master’s Programme: Corporate Communications Supervision: Anne Kroon
Abstract
The concept of issue arenas has mostly been applied within the context of crisis communications research. The present contribution examines online discussions about the issue of Artificial Intelligence (AI) by drawing parallels to the concept of innovation processes. Innovation processes are evolving dynamically over time. This proposes a new theoretical approach that is different from the one of crisis communication. Analysing the Twitter network, a total of 90’666 Tweets were scraped. Organizational Twitter users within the set of Tweets were examined with regards to their influencer position. The influencer position consists of explicit (e.g. retweets, number of followers) and implicit influence (e.g. sentiment) measures. Examining the
relationship between explicit and implicit measures, it turned out that the sentiment was
inversely correlated to the influencer position of an organization within the network of Artificial Intelligence. Moreover, sentiment polarity has shown to not be correlated to influencer positions of organizational Twitter users. The negative relationship found between influencer position and sentiment suggests that organizations can take on an influential position within a network by not recoiling from sharing negative sentiment. Future research should explore the pros and cons of an organization’s aspiration to strive or not to strive for an influential position.
Issue Supremacy: The influencer role of organizations in online social networks
Artificial Intelligence (AI) is on the rise. AI technology has been predicted to enhance the productivity of businesses by up to 40% (Accenture, 2019). It will impact workplaces (Jarrahi, 2018), bring new means of transportation such as autonomous driving technologies (Dai, Zeng, Yu, & Wang, 2019; Fraedrich, Heinrichs, Bahamonde-Birke, & Cyganski, 2019; Xia, Jin, Kong, Xu, & Zeng, 2019), change educational systems (Bajaj & Sharma, 2018; Chassignol,
Khoroshavin, Klimova, & Bilyatdinova, 2018), disrupt communication technologies (Galloway & Swiatek, 2018; Shank, Graves, Gott, Gamez, & Rodriguez, 2019; Sodhro et al., 2019), influence healthcare, for example in the area of risk assessment of diseases or in patient care (Becker, 2019; Hamet & Tremblay, 2017; Holmes, Sacchi, Bellazzi, & Peek, 2017); AI will affect agricultural issues such as grain production and help reduce the impact of agrochemicals on the environment (Elahi, Weijun, Zhang, & Nazeer, 2019; Patrício & Rieder, 2018); and it will also change the entertainment industry (El Beheiry et al., 2019; Sitterding, Raab, Saupe, & Israel, 2019; Zhang & Dahu, 2019). Using AI, it is possible to transfer mental skills, such as
understanding, reasoning, planning, communication and perception to software programs (Al Aani, Bonny, Hasan, & Hilal, 2019; Nagy & Hajrizi, 2018; Ponce & Gutiérrez, 2019).
For organizations, AI offers the opportunity to increase efficiency of processes effectively at low costs (Kelnar & Kostadinov, 2019). New algorithms, the availability of training data, specialized hardware, cloud services, open source resources, greater investments and increased interest enable an inflection point in AI capabilities (Kelnar & Kostadinov, 2019). As AI
algorithms are based on historical data they require a lot of time to be developed. AI will impact organizations in many ways: business terms and regulations need to be adjusted (Wall, 2018; Wright & Schultz, 2018), AI will entail fundamental changes in work force requirements
(Abubakar, Behravesh, Rezapouraghdam, & Yildiz, 2019) and from a broader perspective, profound changes in the global economy are expected (Dirican, 2015; Farrow, 2019; Montes & Goertzel, 2019). In the United States alone 10 billion USD in venture capital was invested into AI technologies in 2018 (Djordjevic, 2019). That being so, organizations should not await a revolution but invest in research and engage in discussions that are being held about AI in order to co-determine and shape the developments (Mahidhar & Davenport, 2018). Organizational actors are seen as major contributors to the progresses made in AI (Clarke, 2019) as they have the power to initiate AI projects and co-determine whether or not AI is to be seen as threat or opportunity. Ergo, it is crucial to understand the discussions that evolve around AI.
In the current climate of continuous change, social media permit organizations to stay on top of current discussions revolving around AI (Mention, Barlatier, & Josserand, 2019).
Influencers play an important role when it comes to those online discussions. While a vast amount of literature already focuses on the interactions of organizations with influencers on social media (Alp & Öğüdücü, 2019; Arora, Bansal, Kandpal, Aswani, & Dwivedi, 2019; Bank, Yazar, & Sivri, 2019; Bigné, Oltra, & Andreu, 2019), the discussions about AI on social media remain to a large extent unexplored. Organizations striving for an influencer position within the discussions gain credibility (Druckman, 2001), legitimacy (Coombs, 2002) and power (Lukes, 1974) and therewith a position of supremacy. Explicit influence measures within the context of Twitter describe easy accessible information about a certain actor such as number of followers and retweets whereas implicit influence measures are not as easy to access as the explicit ones. An example for an implicit influence measure is the average sentiment of an actor. The
relationship between the explicit and implicit influence measures explains not only the current attitude within the issue arena of actors towards a specific issue, but moreover is expected to
shed light on the diverse and complex variety of approaches on the identification of influencers on Twitter.
The leading research question targets the issue of AI and is formulated as follows: “How can the influencer position of an organization in an online social network be detected and how does this position relate to the expressed sentiment of that organization within the network?” This approach is insofar novel, as it does not view the organization as a starting point of analysis but the topic, namely AI. Building a network around the topic of AI and identifying influential organizational actors within the network allows an explorative approach rather than analysing a given set of organizational actors and their twitter communication. Moreover, the issue arena theory is reconsidered from an unconventional perspective, eliminating possible biases that are presumed to have existed in previous research due to a strong focus on crisis communication. Thus, on a theoretical level, the issue arena theory is connected with innovation processes, viewing issues as arising and evolving similar to the way innovations are. On an empirical level, explicit influence measures are used and connected to sentiment providing insights into implicit influence measures and their relationship to explicit influence measures of an actor.
In order to examine the leading research question, firstly, the underlying theoretical framework is introduced. From a theoretical perspective, the issue and the arena lay the necessary theoretical foundation for this study to be conducted. These two components of the issue arena are to be analysed theoretically, based on literature. Secondly, the actors and communication within the issue arena are to be analysed operationally. How data is retrieved, influencer positions and sentiment are computed, is explained. Results are shown and the discussion evaluates the findings. Lastly, the conclusion will recapitulate the findings.
Theoretical Background
When examining the structure and content of online discussion about AI, the main underlying theoretical concept of issue arenas needs to be explained. Issue arenas set the contextual and structural boundaries for organizations and their space of action. Therein every actor influences the others by engaging in discussions about a certain issue, whereas those boundaries are not set but evolve in a dynamic manner, embedded in a situational context and intertwined with other issue arenas. The concept of issue arenas is rooted in the ideas that, firstly, organizational roles in networks matter more than organizational structures and, secondly, that arenas are places of interaction between organizations and their audiences (Luoma-aho & Vos, 2009; Wessells, 2007). As stages or platforms (Craib, 1978), issue arenas are situated in virtual or physical environments, develop dynamically and allow their actors to interact (Luoma‐Aho & Vos, 2010), thereby building a network of organizations and individuals. When analysing an issue arena four aspects are relevant (Vos, Schoemaker, & Liisa Luoma-aho, 2014). Vos et al. (2014) propose using an analytical framework to approach issue arenas with four aspects: Issue-related aspects (The Issue), places of interaction (The Arena), actors involved (The Actors) and course of debate (The Communication). At this point, they argue that both issue and arena should be seen as a given set of thematic and structural (fluid) boundaries and therefore be deduced theoretically, while the actors and communication are constantly shaping the issue and the arena and should be analysed in practice for each idiosyncratic issue arena.
The Issue
The fact that a common definition of the ‘issue’ in an issue arena is inexistent underlines the observation that its specific characteristics are highly dependent on issue-related aspects, namely issue context and issue characteristics (Vos et al., 2014). The ‘issue’ having a negative
connotation has been researched extensively (Hellsten, Jacobs, & Wonneberger, 2019; W. Liu, Lai, & Xu, 2018; Sommerfeldt & Yang, 2017; Stølsvik, 2019; Xiong, Cho, & Boatwright, 2019) with the underlying axiom that organizations have to participate in arenas in order to ensure organizational survival (Luoma‐Aho & Vos, 2010). The negative connotation of issues, however, is not explicitly mentioned in the definition of issues that are defined as: “places where
stakeholders and organizations discuss societal issues” (Hellsten et al., 2019, p. 36).
In the absence of a structured analysis of the issue, it seems legitimate to use an approach based on the four moments of innovation transformation (Aka, 2019). As the issue is not to be considered entirely negative, it is justifiable to adopt the procedural perspective of the four moments of innovation that allow an analysis of the evolving issue over time while it is adjusting dynamically to its environment. The idea of moments being intertwined and not merely gradual accurately grasps the evolution of an issue (Aka, 2019; Akrich, Callon, Latour, & Monagahan, 2002).
With AI having emerged very early, long before technology made self-learning devices possible (Mayor, 2018), it is not the innovation process itself that is relevant in this context but the four moments of innovation processes of AI. AI can be defined as: “The theory and
development of (computer) systems able to perform tasks that normally require human
intelligence, such as visual perception, speech recognition, decision-making, and translation of languages (Oxford Dictionaries, 2019).”
The first moment, Problematization, is characterized by the fundamental question ‘what is intelligence?’ and an intimidated attitude towards AI including the fear of it threatening human existence and causing destruction and omnipresent chaos (Mayor, 2018). The contextual features of AI as an issue touch upon different disciplines such as philosophy, psychology, linguistics,
machine vision, cognitive science, mathematics, logic and ethics (O’Regan, 2016). Rooted in ancient imagination, the theories of René Descartes sum up the fundamental idea of AI, ‘cogito ergo sum’ (I think, therefore I am). The main goal of AI technologies is to develop a machine that has awareness of itself and its environment, is able to learn and evolve over time, has a free will and is ethical (Descartes, 1999).
The second moment, Interessement, describes processes in which actors manifest
themselves by answering fundamental questions or by stopping to ask them while experimenting with possible problem sets and their solutions (Aka, 2019). It is characterized by trial-and-error and experiments that are limited by the size and speed of the memory and processors of
computational devices (Buchanan, 2005). As Minsky (1968) summarized: “The most central idea […] was that of finding heuristic devices to control the breadth of a trial-and-error search. A close second preoccupation was with finding effective techniques for learning. In the post-1962 era the concern became less with ‘learning’ and more with the problem of representation of knowledge (however acquired) and with the related problem of breaking through the formality and narrowness of the older systems. The problem of heuristic search efficiency remains as an underlying constraint, but it is no longer the problem one thinks about, for we are now immersed in more sophisticated sub problems, e.g., the representation and modification of plans (Minsky, 1968, p. 9)”.
The third moment, Enrolment, defines and coordinates the roles assumed by different actors (Aka, 2019) once AI has been introduced to different fields of research and practice: AI was introduced to the field of medicine by Ted Shortliffe in 2012 (Shortliffe, 2012). Furthermore, Deep Blue was the first supercomputer able to identify and analyse up to 60 billion moves in chess, which led to the defeat of then World Chess Champion (Campbell, Hoane Jr., & Hsu,
2002). The Enrolment moment is characterized by a certain apathy. As global economies were highly volatile in the early 2000’s, AI research was strongly tied to tangible results and was expected to bring capitalist value (Farrow, 2019). At the same time, AI investments and
inventions started to rise in Asian countries (G. Li, Hou, & Wu, 2017) and the industry started to professionalize (Hallinen, 2015).
The focal moment of this contribution, Mobilization, is the one in which networks are extended and new relationships are initiated and established (Aka, 2019). With the advent of continuously improving self-learning abilities, the first self-driving cars have been introduced and are being tested on public roads under the supervision of company staff (Teoh & Kidd, 2017). AI applications have become available to the average person as they have become less expensive (Farrow, 2019). The Mobilization moment is characterized by “the widespread acceptance of a paradigm that can be called ‘data-driven’ or ‘statistical AI’ (Farrow, 2019, p. 65)”. Globalization and capitalism are driving the agenda and data starts to have economic value (Brynjolfsson & Mcafee, 2017; Farrow, 2019; Keane, 2003; Kurzweil, 2005). Beyond that, AI is inspiring optimistic as well as pessimistic worldviews, it raises ethical questions, and data protection and privacy are becoming more and more important (Combs, 2017; Crevier, 1995; Farrow, 2019).
The Actor
The Actor-Network Theory (ANT) describes actors as part of a network: “emphasizing the adoption of the principle that human and non-human actors maintain a symbiotic
relationship, forming a social network of (material and non-material) elements. This socio-technical network defends the horizontality of the relationship between technology and society (Joia & Soares, 2018, p. 2203).” Consequently, it is assumed that actors in real life and actors in
an online social network, for example, are symmetrical (Joia & Soares, 2018). Depending on the issue and the arena, many or only a few actors may be involved (Vos et al., 2014). Three different qualities describe the actors (Vos et al., 2014): credibility (Druckman, 2001), legitimacy
(Coombs, 1992) and power (Lukes, 1974).
Organizations have an interest in avoiding their stakeholders to be harmed, which might make them loose trust in AI technologies and oppose their use (Clarke, 2019). The delicate trust relationship between organizations and their stakeholders has been especially studied in the field of CSR communications (Joo, Miller, & Fink, 2019; C.-K. Lee, Song, Lee, Lee, & Bernhard, 2013; Phillips, Thai, & Halim, 2019). The concept of public trust influences the way an
organization is perceived to the extent that higher levels of salience, uncertainty, and outrage lead to a lack of trust in an organization (Coombs, 2014). To that end, it has been found that higher levels of trust result in positive outcomes for an organization, such as greater credibility (Felix, Gaynor, Pevzner, & Williams, 2017; Pevzner, Xie, & Xin, 2015; Wei & Zhang, 2018). Being an implicit influence measure, trust is directly and strongly linked to the degree of importance of an actor (Asim, Malik, Raza, & Shahid, 2019). As influential actors have great potential to
accelerate the dissemination of information, the influencer position of an actor is heavily
dependent on the trust of a given node on a network and community level (J. Zhang, R. Zhang, J. Sun, Y. Zhang, & C. Zhang, 2016) and the ability to drive action and receive people’s
engagement (Arora et al., 2019; Freberg, Graham, McGaughey, & Freberg, 2011).
In short: The influencer position of an organizational actor in an online social network allows organizations to gain the trust needed in order to gain credibility as a legitimate actor within a specific issue. Organizations should not only interact with influencers but also strive for
an influencer position themselves in order to gain a status that allows them to be perceived as credible, legitimate and powerful.
In social network analysis, user influence is one of the most studied concepts. Social networks can be represented as graphs where nodes represent users and edges show interpersonal ties such as retweets among them (Riquelme & González-Cantergiani, 2016). On Twitter, four types of relationships exist: user-to-user, user-to-tweet, tweet-to-tweet, and tweet-to-user. These four types of relationships allow users to follow, mention, reply, retweet and “like” each other’s content (Riquelme & González-Cantergiani, 2016). Thus, Twitter is specifically predestined to be researched with regards to such influence measures (Borge-Holthoefer & Moreno, 2012; R. Li, Lei, Khadiwala, & Chang, 2012; Maruyama, Robertson, Douglas, Semaan, & Faucett, 2014).
Most commonly, variations of the PageRank algorithm are used to measure user influence in terms of his or her ability to spread information and to be retweeted by other influential users (Riquelme & González-Cantergiani, 2016). The original PageRank algorithm uses link information in a set of pages to determine which pages are most pointed to and thus, most important relatively to other pages in the set (Page, Brin, Motwani, & Winograd, 1999; Safronov & Parashar, 2003). Consequently, users are influential when “his or her actions in the network are capable of affecting the actions of many other users in the network” (Riquelme & González-Cantergiani, 2016, p. 960). In particular, an influencer score can be assigned to an actor within a network using retweet edges in respect to weights (Jabeur, Tamine, & Boughanem, 2012).
The Arena
Issues are being discussed in physical as well as in virtual spaces. Each place of
Within the realm of issue arenas, social media provide the possibility for many to express
themselves (Vos et al., 2014). Furthermore, the fact that social media are enabled by the Internet, has been suggested to accelerate interaction (Self, 2010) and cause issue contagion (Coombs, 2002). Debates on social media regarding a specific issue can reinforce traditional media
attention and vice versa (Meriläinen & Vos, 2011). These arguments of awareness and interaction reinforcement legitimize social media as an important issue. In social science, the term ‘arena’ is referred to as “spaces where issues of concern are debated and negotiated (Raupp, 2019, p. 2)”. Organizations enter the issue arena with a strategic interest to influence the public agenda, opinion or policy (Crable & Vibbert, 1985).
Moreover, there is a number of economic and non-economic reasons that organizations are motivated to adopt social media as a communication channel (Lei, Li, & Luo, 2019):
influence on negative messages (L. F. Lee, Hutton, & Shu, 2015), technological competence and impact on internal operations (Schaupp & Bélanger, 2013), media presence (Jung, Naughton, Tahoun, & Wang, 2017), management styles (Bamber, 2010). As shown in the previous section, organizations have their strategic goals in mind when they try to position themselves so they will be perceived as powerful and knowledgeable actors. Especially in healthcare communication this has been researched extensively (Nayak & Linkov, 2019; Ramage & Moorley, 2019; Wilson, Ranse, Cashin, & McNamara, 2014). Viewing issues from a perspective without a negative connotation as an innovation process, organizations should strive for interactions with their customers in order to capture knowledge and gain reputation (Prahalad & Ramaswamy, 2004; Torres de Oliveira, Indulska, Steen, & Verreynne, 2019). Social media interaction improves both radical and incremental innovation and strengthens relationships between an organization and its
stakeholders (Martini, Massa, & Testa, 2012; Piller, Vossen, & Ihl, 2012; Torres de Oliveira et al., 2019).
Twitter allows issues to evolve dynamically (Das, Dutta, Medina, Minjares-Kyle, & Elgart, 2019). It has gained importance not only for political actors (Kruikemeier, 2014; Pal & Gonawela, 2017; Yaqub, Chun, Atluri, & Vaidya, 2017) but also for economic actors (Majumdar & Bose, 2019; Nisar & Prabhakar, 2018; Reboredo & Ugolini, 2018; Shen, Urquhart, & Wang, 2019). Especially, Twitter has been found to have strong predictive power for companies’ operational outcomes (Lei et al., 2019). Accordingly, the aggregate opinion on Twitter
successfully predicts an organization’s earnings (Bartov, Faurel, & Mohanram, 2017), and the number of messages is positively related to trade volumes (Sprenger, Tumasjan, Sandner, & Welpe, 2014). In addition, companies tweet less when existing information suggests neutral outcomes (Crowley, Huang, & Lu, 2018) while they tend to enhance their performance by tweeting more (Huang, Lu, & Su, 2016). The interaction with non-organizational users in this context is insofar important as it influences investor’s valuation of an organization (Kadous, Mercer, & Zhou, 2017). By enabling anyone with access to the Internet to engage in discussions and expanding conversations outside of the academic sphere, the enabling function of Twitter reveals major themes, stances on the impact and public perspectives on issues such as AI (Goldberg & Rosenkrantz, 2019).
The Communication
The communication in issue arenas deserves to be analysed specifically for each issue. Based on the perspective of issues as innovation processes, the current literature sees online social networks in general as enablers and drivers of innovation (Bhimani, Mention, & Barlatier, 2018). Organizations can capture and create new knowledge by involving internal and external
stakeholders in different phases of the innovation process (Piller et al., 2012). Thus, social media serve as a tool to engage stakeholders, build ideas, set up forums for communication, broadcast information, spot trends, collaborate and motivate participation (Bhimani et al., 2018; Turban, Bolloju, & Liang, 2011). In this context, it has been found that influencers on Twitter use numerous hashtags and mentions. Moreover, they have a larger number of people they follow (Lahuerta-Otero & Cordero-Gutiérrez, 2016). These findings have a relatively low importance, as influencer positions in these studies are set into relation with the amount of interaction (hashtags, retweets and mentions) and followership (number of followers): Explicit influence measures. The validity of such an approach is insofar limited, as the influencer position itself is to be defined with those explicit measures. It is needed, to take a measure that is not explicit, in order to truly be able to make a statement about the characteristics of an influencer.
An opinion holder is the owner of an opinion. Opinion holders can be organizations, groups, individuals, etc. There are two distinct types of opinion holders: direct and indirect (Liu, 2015; Paltoglou & Giachanou, 2014). “The direct opinion is expressed directly on an entity, whereas the indirect opinion is expressed on an entity based on some positive or negative effects related to some other entities. In a social media setting, on Twitter in particular, opinion holders are the authors of the tweets (Alharbi & de Doncker, 2019, p. 51).” Opinion holders can
consequently publish Tweets containing positive, negative or neutral sentiment. This has often been researched within the context of stock market prices (R. Chen, Yu, Jin, & Bao, 2019; Gupta & Banerjee, 2019; Jiang, Lee, Martin, & Zhou, 2019; Qadan & Aharon, 2019; Shi, Tang, & Long, 2019). In that context, it has been found that influencers take on an important role on Twitter with regards to their sentiment and their influence on financial markets (Groß-Klußmann, König, & Ebner, 2019).
AI being a very polarizing topic in current research and popular media, it raises many ethical questions and a need for critical attention is expressed (Galloway & Swiatek, 2018). It has been shown that people react with emotions such as surprise, amazement and disappointment or unease when encountering AI (Shank et al., 2019). Thus, the first hypothesis assumes that influencer positions are enhanced by a sentiment polarity in:
H1 The level of polarity in Tweets sent by organizations positively predicts the influencer position of organizations compared to organizations that publish Tweets containing neutral sentiment.
Furthermore, in social psychology, the negativity bias theory states that humans are more likely to focus on bad news than on good news (Rozin & Royzman, 2001). In an online social network, this theory has been supported inasmuch as negative sentiment in posts received a higher number of comments and shares (N. Kumar, Nagalla, Marwah, & Singh, 2018). Generally speaking, negative Tweets are published more often. When it comes to scientific discussions on Twitter regarding the fields of mathematics and computers, citizens and representatives from
organizations outside of universities engage in more negative discussions about topics within these fields (Didegah, Mejlgaard, & Sørensen, 2018). Regarding other technical and ethically critical issues such as the oil and gas industry, Gupta and Banerjee (2019) find a predominance of negative sentiment in news reports about the Organization of Petroleum Exporting Countries. Furthermore, negative sentiment is higher in periods of higher reserves, exports and production (Gupta & Banerjee, 2019). These findings hint at a dominance of negative sentiment in ethically questionable environments in a way that is similar to the ethical questions arising with the
implementation of AI technologies (Clarke, 2019). The second hypothesis can therefore be formulated as follows:
H2 Stronger positive sentiments expressed in Tweets published by organizations on Twitter are negatively related to the influencer position of that organization.
Methods
In order to test the hypotheses explained above, data analysis requires four
methodological steps: First, Twitter data needs to be scraped and sentiment scores have to be given to each Tweet within the dataset. Second, one needs to show how organizations have been identified within the dataset in order to conduct further analysis on this data subset. Third, one needs to present descriptive statistics regarding the dataset and show the sample distribution. In a fourth step, one has to show how the influencer rank was computed using the PageRank
algorithm as a starting point. Lastly, hypotheses need to be tested with a linear regression analysis.
Scraping Twitter Data The scraper was built using the programming language
TypeScript and is running continuously on a nodeJS server. The scraping is implemented by connecting to the Twitter streaming API and monitors the keywords ‘#AI’ and
‘#ArtificialIntelligence’. In order to use the Twitter API, credentials were requested from and then issued by Twitter to authenticate the connection. Using a standard API, rate limitations apply and restrict data retrieval (Twitter, 2019b). When the rate limitation threshold was reached, the scraper would wait for a set amount of time before reconnecting to the API. The reconnecting strategy was used as proposed in the Twitter API documentation (Twitter, 2019a). Tweets sent
during absent time were therefore not included in the dataset. Additionally, software bugs and other technical issues impeded the collection of data. The data was saved in a Mongo DB
database, which is a not only structured query language (NoSQL) database. Data in Mongo DB is stored in JSON documents (Mongo DB, 2019).
Sentiment Analysis The sentiment analysis was integrated into the Twitter Scraper using
AFINN, a simple word matching tool (Nielsen, 2011). “Sentiment analysis, also known as opinion mining, is an important type of text analysis that addresses the problem of detecting, extracting and analysing opinion oriented text, identifying positive and negative opinions, and measuring how positively or negatively an entity is regarded” (Alharbi & de Doncker, 2019, p. 51). The AFINN word list contains words from the public domain Original Balanced Affective Word List, slang words by Urban Dictionary, Wiktionary and The Compass DeRose Guide to Emotion words. This lexicon-based method utilizes affective dictionaries to estimate the
affective content of text segments (Alharbi & de Doncker, 2019). To exclude ambiguities, words such as ‘patient, ‘firm’, ‘mean’, ‘power’ and ‘frank’ were not included (Nielsen, 2011). The AFINN sentiment analysis resulted in a numerical output ranging from -5 (negative) to 5
(positive). Each Twitter user was given a sentiment score by calculating the mean of each Tweet sent by the user.
Identification of Organizations In order to test the above hypotheses, organizational
actors in the dataset need to be manually identified using a binary annotation (0 and 1) for organization and non-organization. Based on the majority of perspectives on organizational actors in the literature (Robichaud, Giroux, & Taylor, 2004), actor-hood is considered a
prerequisite for an organization to be identified as such: “Organizations are externally defined as actors by other actors in society. Organizations have sovereignty to act independent of the wishes
of their members. Organizations are a type of social tool designed by individuals in a manner that is both unattainable by any given individual” (King et al., 2010, p. 298). To not discriminate between different types of organizations, be it due to different types of corporate constitutions, legal forms, stages of organizational development or operations in different industries, all types of organization were included into the dataset and no distinction was made between, e.g., start-ups and big corporations or governmental, non-profit and for-profit organizations. The specific criteria used to identify whether or not a Twitter user acts as an organization or not is shown in Table 1.
Table 1
Criteria binary identification: organization (1) / non-organization (0)
Type of Account Annotation Description Logic Online
communities Fan accounts Accounts of technical tools
0 Twitter users who call themselves an online community, sharing e.g. Machine Learning-News,
Developer Communities for Questions and Answers. Twitter users that share information about a certain person or topic of interest (e.g. fan community of Elon Musk). Twitter users who share news about a technical tool such as a software
No formal control mechanisms (Ivaturi & Chua, 2019) Volunteer efforts (Metzler,
Günnemann, & Miettinen, 2019)
News media accounts
0 Traditional news media that act as Twitter users and share articles (e.g. The Guardian, sharing its own articles)
Main goal is news dissemination (all topics) (Orellana-Rodriguez & Keane, 2018)
Role accounts Freelancers
0 Twitter users who hold a specific role within an organization (e.g. CTO of Microsoft, founders, inventors, investors, researchers, marketers)
Twitter users who act as freelancers and earn money by providing services (e.g. software developer, speakers, artists etc.)
Mainly act as individuals rather than organizations (Aladwani, 2015)
Offer goods and services but not as an organization but as an individual (Grothe-Hammer, 2019)
News distribution accounts, spam and bots, robot profiles
0 Twitter users who distribute news media articles from other news media, distribute fake news or
Dissemination of commercial URL’s, fake news, abusive content, automated generation of large
Note: Organizational Twitter users are annotated with 0 and non-organizational Twitter users are coded with 1
Descriptives The dataset was collected from the 15th of March 2019, 11:36:03, until the 15th of May 2019, 18:04:06. A total of 90’666 Tweets has been collected. Figure 1 shows the distribution of Tweets over time.
automated content
Twitter users who are not actual people, but e.g. chatbots
volume content (Inuwa-Dutse, Liptrott, & Korkontzelos, 2018)
Non-available accounts
0 Twitter users who were blocked or are not available anymore
Not active anymore
No access to account information Educational
Platforms
1 Twitter users that are linked to online platforms that offer courses (e.g. tutorials)
Belongs to the group of online firms (like Netflix, Spotify) providing AI knowledge (Kaplan & Haenlein, 2019)
AI Projects 1 Twitter users that act in the name of a project group conducting research or projects, events and conferences regarding AI
Research community with interest in promoting themselves as experts Is a project-based organization or
project-oriented organization (Turner, 2018)
Topic-specific News media Apps and Software Podcasts
1 News media that acts solely within the field of AI Learning Twitter users that act as an ‘App’ or ‘Software’ with regards to AI Twitter users that act as
‘Podcasts’ with regards to AI
Use of social media as business practice
AI focus – goal is to position themselves as AI experts
Figure 1 Overview of amount of Tweets scraped per day. Scraper downtimes have been excluded from
this overview
The dataset included 10’129 unique Twitter users. 6’999 Tweets in total included the #AI (case insensitive) and 90’562 Tweets the #ArtificialIntelligence (case insensitive). Figure 2 shows how the number of Tweets per account (M = 6, SD = 33.38) was distributed.
Figure 2 Number of Tweets per account. Only the first 58 accounts are shown, who published ≤ 70
Tweets 0 1000 2000 3000 4000 5000 6000 7000 8000 15/03/2019 20/03/2019 25/03/2019 08/04/2019 28/04/2019 07/05/2019 15/05/2019 Number of Tweets 0 500 1000 1500 2000 2500 Number
The most common language scraped was English. Figure 3 shows which other languages were represented in the dataset. Very clearly, English is the predominant language (88.63 %), followed by the category Undefined (4 %) that includes characters such as Emojis that are not applicable or not identifiable as a language. Other languages represented in the dataset are Spanish (1.8%), followed by Japanese (1.7%), French (1.17%), German (0.6%), Portuguese (0.43%), Italian (0.3%) and Dutch (0.14%). Other (1.2%) includes languages such as Danish, Catalan, Romanian, Finnish, Polish, Indonesian, Swedish, Estonian, Tagalog, Russian and others.
Figure 3 Language distribution Tweets
The Tweet dataset was first filtered for English language Tweets resulting in 8’579 Twitter users. 5287 accounts that published two or less Tweets in the time span of data collection were
removed from the sample. Furthermore, 34 accounts were removed because they were blocked or no longer available. The final sample size resulted in N = 615 organizational Twitter users. The sentiment distribution (M = .022, SD = .753, min = -1.00, max = 2.00) of those Tweets is shown in Figure 4. Most Tweets had a sentiment score of 0.0 (36.13 %). More Tweets exhibited a positive sentiment, with 23.08% containing a sentiment of 0.1 and 31.10% a sentiment of 0.2.
0 10000 20000 30000 40000 50000 60000 70000 80000 Number
Figure 4 Sentiment Distribution (language = en), rounded sentiments (0.1)
Influencer Analysis The PageRank algorithm served as a basis for the identification of
influencers (Page et al., 1999). The InfRank algorithm is a variation of the PageRank algorithm and measures user influence in terms of the ability to spread information and to be retweeted by other influential users (Jabeur et al., 2012; Riquelme & González-Cantergiani, 2016). Each Twitter user was attributed an initial influence score based on popularity. The influence of a Twitter user was confirmed if other influencers were involved in retweets. InfRank considers the social network of Twitter users and gives influence scores using retweet edges with respect to weights (Jabeur et al., 2012). The following variables in Table 2 are relevant.
Table 2
Network Topology Variables (Jabeur et al., 2012, p. 113)
Variable Description
𝑈 Set of Twitter user nodes
𝐸 = 𝑈 × 𝑈 Set of edges denoting relationships between Twitter users
Σ𝐸= {𝑓, 𝑟, 𝑚} Alphabet of edge labels with 𝑓, 𝑟, 𝑚 corresponds, respectively, to
following, retweeting and mentioning associations ℓ𝐸: 𝐸 →Σ𝐸 Associates a label to each edge
𝒪: 𝑈 × Σ𝐸 → 𝑈 × 𝑈 ⋯
× 𝑈
Associates to each Twitter user 𝑢𝑖 ∈ 𝑈 the set of successor nodes with connecting edges are labeled by �� ∈Σ𝐸
0 2000 4000 6000 8000 10000 12000 14000 16000 -1 ,0 -0 ,9 -0 ,8 -0 ,7 -0 ,6 -0 ,5 -0 ,4 -0 ,3 -0 ,2 -0 ,1 0 ,0 0 ,1 0 ,2 0 ,3 0 ,4 0 ,5 0 ,6 0 ,7 0 ,8 0 ,9 1 ,0 1 ,1 1 ,2 1 ,3 1 ,4 1 ,5 1 ,6 1 ,7 1 ,8 1 ,9 2 ,0 Number
Variable Description 𝐼: 𝑈 × Σ𝐸 → 𝑈 × 𝑈 ⋯
× 𝑈
Associates to each Twitter user 𝑢𝑖 ∈ 𝑈 the set of predecessor nodes with connecting edges are labeled by 𝜄 ∈Σ𝐸
Furthermore, the Relationship Weight needs to be specified. While 𝑇 (𝑢𝑖) represents the set of tweets of Twitter user 𝑢𝑖, 𝑅 (𝑢𝑖) represents the set of Tweets Twitter user 𝑢𝑖 has retweeted. The retweeting relationship used in the InfRank algorithm is shown in Table 3.
Table 3
Relationship Weight
Retweeting Relationship Description
𝑤𝑟(𝑢𝑖, 𝑢𝑗) = |𝑇 (𝑢𝑗) ⋂ 𝑅 (𝑢𝑖)| |𝑇(𝑢𝑖)|
Retweeting relationship: The retweet edge is defined as follows: a retweet from Twitter user 𝑢𝑖 to another Twitter user 𝑢𝑗 if Twitter user 𝑢𝑗 retweeted at least one Tweet from 𝑢𝑖. A retweeting association is considered reliable to the extent that Twitter user 𝑢𝑖 publishes retweets that belong to Twitter user 𝑢𝑗.
Note: from (Jabeur et al., 2012, p. 113)
A popularity score was attributed to each influencer. The popularity score was considered confirmed if the actor involved other influential actors in its retweets. Figure 5 shows the InfRank algorithm.
Error! Reference source not found.Influence was defined using retweet edges. The last step of
calculating InfRank consisted in choosing a damping factor d which determines the convergence rate (Bressan & Peserico, 2010). Most commonly it is set to 0.85 (Alp & Öğüdücü, 2019;
Buzzanca, Carchiolo, Longheu, Malgeri, & Mangioni, 2018; Chakhmakhchyan & Shepelyansky, 2013; Frahm & Shepelyansky, 2019; Kandiah & Shepelyansky, 2012; Zengin Alp & Gündüz Öğüdücü, 2018).
Analysis To run a regression analysis, six assumptions were tested (Neter, Kutner,
residuals was carried out on the dataset in order to detect outliers. As it is usual with long-tailed data, the dataset contained some outliers (std. residual min = -.387, std. residual max = 21.089). Regardless of these outliers, the analysis was continued. The second test – which was run in order to see if the data met the assumption of collinearity – indicated that multicollinearity was nothing to be concerned about (IQ Scores, Tolerance = 1.000, VIF = 1.000). With regards to the assumption of independent errors, the data did not meet the requirements (Durbin-Watson value = .282). Furthermore, the histogram of standardized residuals did not indicate an approximate normal distribution of the errors, which was also due to the long-tailed data, nor did the normal P-P plot of standardized residuals. The scatterplot of standardized residuals showed, furthermore, that they did not meet homogeneity of variance nor linearity. Lastly, the data did meet the
assumption of variances greater than zero (InfRank: Variance = 15.054, Sentiment: Variance = .006).
Results
First, a visual representation of the Twitter network including all users was created using the network analysis and visualization software Gephi (Figure 6).
Figure 5 Visualization Twitter Network, each node is represented by a dot and each line represents a
retweet connection Nodes = Twitter Users, Edges = retweets, Layout algorithms = ForceAtlas2 and Yifan
Hu
Gephi is an open-source software that uses a 3D engine to display large networks in real-time (Bastian, Heymann, & Jacomy, 2009). To do so, the original PageRank algorithm was used to determine the size of each node (Page et al., 1999), whereas retweets represented the edges in the network. In order to make the graph readable, the ForceAtlas2 algorithm was used.
repulse each other, and like springs, edges attract their nodes. The movement they create converges towards a balanced state (Jacomy, Venturini, Heymann, & Bastian, 2014). For two reasons, the balanced state did not completely converge in this example. Firstly, readability was not improved in a fully converged graph and secondly, there is a speed and precision trade-off. While speed accelerates the convergence and thus the balanced state, precision tends to prevent it. A more precise simulation requires more computing time. Additionally, the Yifan Hu layout algorithm was applied. The difference to ForceAtlas2 is that the pair of nodes taken into
consideration is only the one of adjacent nodes, which reduced the complexity; hence a the Yifan Hu layout algorithm can be computed faster (Khokhar, 2015).
The first hypothesis suggests that organizations that publish Tweets containing a polarity in sentiment (M = .108, SD = .071) score higher on the influencer scale (M = .627, SD = 3.88) compared to organizations that publish Tweets containing neutral sentiment. In order test the first hypothesis absolute sentiment values were used. The regression analysis resulted in 0.2% not significantly explained variance of the influencer rank (R2 = .002, F (1, 613) = 1.143, p < .286). It was found that sentiment did not significantly predict influencer rank (β = - 2.357, p < .286).
The second hypothesis suggests that sentiments (M = .105, SD = .074) expressed in Tweets published by organizations on Twitter are negatively correlated to the organization’s influencer position (M = .627, SD = 3.88). In order to test the second hypothesis, a single linear regression was conducted. The results of the regression indicated that sentiment significantly explained 0.8% of the variance of the influencer rank (R2 = .008, F (1, 613) = 5.036, p < .025). It was found that sentiment significantly predicted influencer rank (β = - 4.705, p < .025). These results show that, as assumed in the second hypothesis, there is a negative relationship between sentiment and influencer rank.
Discussion
This contribution examines the relationship between the influencer position of organizational actors within the Twitter network of AI and the implicit influence measure of sentiment.
Firstly, it was shown how an influencer position within a Twitter network can be detected using the PageRank algorithm as a starting point. This approach is among the most undisputed ones in the current research literature and has been implemented in research settings such as keyword detection (M. Kumar & Rehan, 2018), phishing tweet and rumour detection (Liew, Sani, Abdullah, Yaakob, & Sharum, 2019; Sicilia, Lo Giudice, Pei, Pechenizkiy, & Soda, 2018), detection of citizen engagement (Bonsón, Perea, & Bednárová, 2019) and communities of practice in the media industry (Komorowski, Huu, & Deligiannis, 2018). The interaction
enabling function of social media has attracted a vast amount of researchers trying to understand human behaviour with regards to how a small number of users can create chain reactions of influence based on word-of-mouth (Alp & Öğüdücü, 2019). Since there is no objective fundamental truth, influencer identification is a difficult task.
Explicit influence measures are the ones that are immediately accessible when one is looking at a network and they do not require further analysis (e.g. likes, number of followers, retweets, mentions, etc.). Implicit influence measures are more difficult to measure, as they are, by their very nature, not explicit, but carry implicit information that needs further analysis in order to be detected. Examples for such implicit influence measures are expertise of the user, interest of the user, legitimacy of the user, user sentiment, and user authenticity. Additionally, a blurry distinction between explicit and implicit influence measures makes it difficult to find one overarching influence measure that is superior to others. Rather, it makes sense to proceed from
the same base algorithm and adjust the algorithm according to specific research interests, in the manner it has been done in this contribution. Implicit influence measures must either be
completely excluded and considered measurements from outside the network or be included within the influence measure itself. The latter, although, does not contribute to a clear distinction between explicit and implicit measurements but rather blurs the distinction even more.
Moreover, this approach makes it difficult to find general algorithms that work in different settings. The success of the PageRank algorithm can be accredited to its versatility and usability on Web Pages and in online social networks such as Twitter. To summarize and answer the first part of the research leading question: It is suggested to use the same algorithm for influence measures and adjust the algorithm bottom-up or include measures according to specific interests of research rather than adopting a top-down approach with the idea of finding one all-including (explicit and implicit) algorithm.
Secondly, it was found that organizations that publish Tweets on Twitter containing a negative sentiment were able to significantly predict their influencer position. The analysis of sentiments on Twitter has been subject to a vast amount of studies (Abid, Alam, Yasir, & Li, 2019; Behrendt & Schmidt, 2018; Reboredo & Ugolini, 2018). The same holds true for the influencer position of Twitter users (Alp & Öğüdücü, 2019; Arora et al., 2019; Francalanci & Hussain, 2017). The novelty of this contribution lies in the connection between explicit influence measures and a implicit influence measure, i.e. sentiment. The negative relationship found shows that influencer positions increase with overall negativity of the Twitter user. As the study focused only on organizational Twitter users, it follows that organizations are well advised to incorporate a rather negative sentiment into the Tweets they publish when striving for an influential position in a certain issue. While traditional public relations theories see negative sentiment as something
harmful to an organization (Chen, Hung-Baesecke, & Chen, 2019; Gesualdi, 2019; Sommerfeldt, Yang, & Taylor, 2019; Waymer & Heath, 2019), social media follows a different logic.
Consequently, negative sentiment on social media should not always be considered harmful, as it can enhance the awareness about an organization as an important influencer within a field
provided that the negative sentiment is not directly addressed towards the corporation but towards the issue.
The negative connotation of issues was successfully disconnected from the issue. As shown visually (see Figure 6), an issue and with it, the network around it can evolve without the negative connotation assumed in other studies (Neu, Saxton, Rahaman, & Everett, 2019;
Pourebrahim, Sultana, Edwards, Gochanour, & Mohanty, 2019; Xiong et al., 2019).
Furthermore, organizations that published Tweets containing stronger sentiment did not rank higher as influencers (InfRank score) than organizations that published Tweets containing more neutral sentiment. Interestingly, the first hypothesis was not confirmed. The sentiment polarity thesis as a predictor of a higher influencer position was thus not confirmed. This finding is insofar interesting as influencers are not polarizing in general but rather only slightly negative. As influencers by their very definition speak to a broader audience, a more homogenous
sentiment enables them to address more people and thus provoke discussions between them with a stronger sentiment. In a similar way, Daniel, Neves and Horta (2017) found that Microsoft as an organizational Twitter user announced the purchase of Nokia rather neutrally with regards to sentiment but still provoked much agitation among its users. In line with this contribution, the sentiment polarity was not found at the epicentre of information dissemination (influencer) but within the epicentre’s target group (Daniel, Neves, & Horta, 2017). Other than suggested, the
sentiment concentration is thus assumed to be high in the periphery of the network (see Figure 6).
To summarize and answer the second part of the research leading question: A negative relationship had been detected between the influencer position of an organization on Twitter participating in the issue arena of AI and their average user sentiment.
The theoretical implications of this contribution are to be seen in the context of the issue arena theory. A structured analysis of an issue arena has been proposed with regards to four aspects of the issue arena: the issue, the arena, the communication and the actors. Furthermore, the issue arena of AI has been located on Twitter and a visualization of the Twitter network helps to understand the current situation of few, well interconnected and active users, while most users are dispersed and not very well connected. With regards to active and passive stakeholders (Hellsten et al., 2019), the visualization of the issue arena is also beneficial in so far as it shows that a majority of users are not connected to the major contributors (see Figure 6). Furthermore, although a majority of influential users has been shown to be negative rather than positive, the negative bias was excluded by examining a topic that is actually controversial at times but which, if one excludes crisis communication, cannot be judged as easily as purely negative.
Some limitations apply to this contribution. First, the choice of PageRank as a base algorithm was based on best practice. A reflection on the use of the PageRank algorithm as a foundation is crucial in order to determine whether the PageRank algorithm can be transferred to into social media environments directly. In this context, one needs to consider that celebrities and people of interest might also play an important role. This suggests the idea that their behaviour might influence the behaviour of organizations and their communication strategies and that the interest of organizations in being recognized as influencers can vary (Van Norel, Kommers, Van
Hoof, & Verhoeven, 2014). Thinking one step further, the importance of different influencers around an organization would be an interesting topic to study.
Second, a more sophisticated way of topic detection is suggested in order to capture a legitimate ‘snapshot’ of reality (Wang, Chi, Liu, & Wang, 2019). Focusing on hashtags to detect topical actors may be critical insofar as Twitter users may have used hashtags unintentionally or with an intention other than to engage in discussions about AI. Clustering algorithms, for example, may be an option to exclude noise from the sample (Makkonen, Ahonen-Myka, & Salmenkivi, 2004; Shuoying & Jin, 2016).
Third, no distinction was made between different types of organization. This is insofar interesting as different types of organizations may have different types of interest to market themselves as AI experts. Not all organizations may share an interest to be recognized as an influencer, which leads us to a fundamental question that should be explored: “Should organizations strive for an influencer position in Twitter networks and, if so, what different strategies can be applied to different types of organizations?” When conducting a future large-scale analysis, this question could be analysed by implementing an AI algorithm that
automatically detects organizations and clusters them into different types.
Conclusion
This contribution analysed the topical influencer role of organizations within the issue of AI with regards to the sentiments expressed within the online social network of Twitter. The research question was formulated as follows: “How can we detect the influencer position of an organization in an online social network of experts and to what extent does this position relate to the sentiment towards them within the network?” Building on the theoretical framework of issue arenas, issues and arenas were analysed on a theoretical basis, whereas actors and
communication within the issue arena were analysed using statistics. Influencer positions were analysed based on the PageRank algorithm, which is the most common algorithm used to compute influence in an online social network. Regarding sentiment it was found that sentiment did in fact negatively influence the influencer position of an organization. If organizational Twitter users strategically aim for an influencer position within a certain issue on Twitter, they should consider the sentiments they express on Twitter. Future research should focus on the motivation why organizations should or should not strive for an influencer position within a network.
References
Abid, F., Alam, M., Yasir, M., & Li, C. (2019). Sentiment analysis through recurrent variants latterly on convolutional neural network of Twitter. Future Generation Computer Systems, 95, 292–308. https://doi.org/10.1016/j.future.2018.12.018
Abubakar, A. M., Behravesh, E., Rezapouraghdam, H., & Yildiz, S. B. (2019). Applying artificial intelligence technique to predict knowledge hiding behavior. International Journal of Information Management, 49, 45–57. https://doi.org/10.1016/j.ijinfomgt.2019.02.006 Accenture. (2019). Future of Artificial Intelligence Economic Growth | Accenture. Retrieved 29
April 2019, from https://www.accenture.com/sk-en/insight-artificial-intelligence-future-growth
Aka, K. G. (2019). Actor-network theory to understand, track and succeed in a sustainable innovation development process. Journal of Cleaner Production, 225, 524–540. https://doi.org/10.1016/j.jclepro.2019.03.351
Akrich, M., Callon, M., Latour, B., & Monagahan, A. (2002). The key to success in innovation part i: the art of interessement. International Journal of Innovation Management, 06(02), 187–206. https://doi.org/10.1142/S1363919602000550
Al Aani, S., Bonny, T., Hasan, S. W., & Hilal, N. (2019). Can machine language and artificial intelligence revolutionize process automation for water treatment and desalination? Desalination, 458, 84–96. https://doi.org/10.1016/j.desal.2019.02.005
Aladwani, A. M. (2015). Facilitators, characteristics, and impacts of Twitter use: Theoretical analysis and empirical illustration. International Journal of Information Management, 35(1), 15–25. https://doi.org/10.1016/j.ijinfomgt.2014.09.003
Alharbi, A. S. M., & de Doncker, E. (2019). Twitter sentiment analysis with a deep neural network: An enhanced approach using user behavioral information. Cognitive Systems Research, 54, 50–61. https://doi.org/10.1016/j.cogsys.2018.10.001
Alp, Z. Z., & Öğüdücü, Ş. G. (2019). Influence Factorization for identifying authorities in Twitter. Knowledge-Based Systems, 163, 944–954.
https://doi.org/10.1016/j.knosys.2018.10.020
Arora, A., Bansal, S., Kandpal, C., Aswani, R., & Dwivedi, Y. (2019). Measuring social media influencer index- insights from facebook, Twitter and Instagram. Journal of Retailing and Consumer Services, 49, 86–101. https://doi.org/10.1016/j.jretconser.2019.03.012
Asim, Y., Malik, A. K., Raza, B., & Shahid, A. R. (2019). A trust model for analysis of trust, influence and their relationship in social network communities. Telematics and Informatics, 36, 94–116. https://doi.org/10.1016/j.tele.2018.11.008
Bajaj, R., & Sharma, V. (2018). Smart Education with artificial intelligence based determination of learning styles. International Conference on Computational Intelligence and Data Science, 132, 834–842. https://doi.org/10.1016/j.procs.2018.05.095
Bamber, J. (2010). Wang 2010,’What‟ s my style? The influence of top managers on voluntary corporate financial disclosure’. The Accounting Review, 85, 1131.
Bank, S., Yazar, E. E., & Sivri, U. (2019). Can social media marketing lead to abnormal portfolio returns? European Research on Management and Business Economics, 25(2), 54–62. https://doi.org/10.1016/j.iedeen.2019.04.006
Bartov, E., Faurel, L., & Mohanram, P. S. (2017). Can Twitter Help Predict Firm-Level Earnings and Stock Returns? The Accounting Review, 93(3), 25–57. https://doi.org/10.2308/accr-51865
Bastian, M., Heymann, S., & Jacomy, M. (2009). Gephi: an open source software for exploring and manipulating networks. Presented at the Third international AAAI conference on weblogs and social media.
Becker, A. (2019). Artificial Intelligence in Medicine: What is It Doing for Us Today? Health Policy and Technology. https://doi.org/10.1016/j.hlpt.2019.03.004
Behrendt, S., & Schmidt, A. (2018). The Twitter myth revisited: Intraday investor sentiment, Twitter activity and individual-level stock return volatility. Journal of Banking & Finance, 96, 355–367. https://doi.org/10.1016/j.jbankfin.2018.09.016
Bhimani, H., Mention, A.-L., & Barlatier, P.-J. (2018). Social media and innovation: A
systematic literature review and future research directions. Technological Forecasting and Social Change. https://doi.org/10.1016/j.techfore.2018.10.007
Bigné, E., Oltra, E., & Andreu, L. (2019). Harnessing stakeholder input on Twitter: A case study of short breaks in Spanish tourist cities. Tourism Management, 71, 490–503.
https://doi.org/10.1016/j.tourman.2018.10.013
Bonsón, E., Perea, D., & Bednárová, M. (2019). Twitter as a tool for citizen engagement: An empirical study of the Andalusian municipalities. Government Information Quarterly. https://doi.org/10.1016/j.giq.2019.03.001
Borge-Holthoefer, J., & Moreno, Y. (2012). Absence of influential spreaders in rumor dynamics. Physical Review E, 85(2), 026116.
Bressan, M., & Peserico, E. (2010). Choose the damping, choose the ranking? Selected Papers from the 3rd Algorithms and Complexity in Durham Workshop ACiD 2007, 8(2), 199– 213. https://doi.org/10.1016/j.jda.2009.11.001
Brynjolfsson, E., & Mcafee, A. (2017). The business of artificial intelligence. Harvard Business Review.
Buchanan, B. G. (2005). A (very) brief history of artificial intelligence. Ai Magazine, 26(4), 53– 53.
Buzzanca, M., Carchiolo, V., Longheu, A., Malgeri, M., & Mangioni, G. (2018). Black hole metric: Overcoming the pagerank normalization problem. Information Sciences, 438, 58– 72. https://doi.org/10.1016/j.ins.2018.01.033
Campbell, M., Hoane Jr., A. J., & Hsu, F.-H. (2002). Deep Blue. Artificial Intelligence, 134(1–2), 57–83. https://doi.org/10.1016/S0004-3702(01)00129-1
Chakhmakhchyan, L., & Shepelyansky, D. (2013). PageRank model of opinion formation on Ulam networks. Physics Letters A, 377(43), 3119–3123.
https://doi.org/10.1016/j.physleta.2013.10.003
Chassignol, M., Khoroshavin, A., Klimova, A., & Bilyatdinova, A. (2018). Artificial Intelligence trends in education: a narrative overview. 7th International Young Scientists Conference on Computational Science, YSC2018, 02-06 July2018, Heraklion, Greece, 136, 16–24. https://doi.org/10.1016/j.procs.2018.08.233
Chen, R., Yu, J., Jin, C., & Bao, W. (2019). Internet finance investor sentiment and return comovement. Pacific-Basin Finance Journal.
https://doi.org/10.1016/j.pacfin.2019.05.010
Chen, X., Hung-Baesecke, C.-J. F., & Chen, Y.-R. R. (2019). Constructing positive public relations in China: Integrating public relations dimensions, dialogic theory of public relations and the Chinese philosophical thinking of Yin and Yang. Public Relations Review. https://doi.org/10.1016/j.pubrev.2019.04.004
Clarke, R. (2019). Principles and business processes for responsible AI. Computer Law & Security Review. https://doi.org/10.1016/j.clsr.2019.04.007
Combs, C. C. (2017). Terrorism in the twenty-first century. Routledge.
Coombs, W. T. (1992). The Failure of the Task Force on Food Assistance: A Case Study of the Role of Legitimacy in Issue Management. Journal of Public Relations Research, 4(2), 101–122. https://doi.org/10.1207/s1532754xjprr0402_03
Coombs, W. T. (2002). Assessing online issue threats: Issue contagions and their effect on issue prioritisation. Journal of Public Affairs, 2(4), 215–229.
Coombs, W. T. (2014). Ongoing crisis communication: Planning, managing, and responding. Sage Publications.
Crable, R. E., & Vibbert, S. L. (1985). Managing issues and influencing public policy. Public Relations Review, 11(2), 3–16. https://doi.org/10.1016/S0363-8111(82)80114-8
Craib, I. (1978). Erving Goffman: Frame Analysis. Philosophy of the Social Sciences, 8(1), 79– 86. https://doi.org/10.1177/004839317800800108
Crevier, D. (1995). AI: The tumultuous history of the search for artificial intelligence. Journal of the History of the Behavioral Sciences, 31(3), 273–278. https://doi.org/10.1002/1520-6696(199507)31:3<273::AID-JHBS2300310314>3.0.CO;2-1
Crowley, R., Huang, W., & Lu, H. (2018). Discretionary disclosure on Twitter.
Dai, H., Zeng, X., Yu, Z., & Wang, T. (2019). A scheduling algorithm for autonomous driving tasks on mobile edge computing servers. Journal of Systems Architecture, 94, 14–23. https://doi.org/10.1016/j.sysarc.2019.02.004
Daniel, M., Neves, R. F., & Horta, N. (2017). Company event popularity for financial markets using Twitter and sentiment analysis. Expert Systems with Applications, 71, 111–124. https://doi.org/10.1016/j.eswa.2016.11.022
Das, S., Dutta, A., Medina, G., Minjares-Kyle, L., & Elgart, Z. (2019). Extracting patterns from Twitter to promote biking. IATSS Research, 43(1), 51–59.
https://doi.org/10.1016/j.iatssr.2018.09.002
Descartes, R. (1999). Discourse on method and meditations on first philosophy. Hackett Publishing.
Didegah, F., Mejlgaard, N., & Sørensen, M. P. (2018). Investigating the quality of interactions and public engagement around scientific papers on Twitter. Journal of Informetrics, 12(3), 960–971. https://doi.org/10.1016/j.joi.2018.08.002
DiPaola, S., Gabora, L., & McCaig, G. (2018). Informing artificial intelligence generative techniques using cognitive theories of human creativity. Postproceedings of the 9th Annual International Conference on Biologically Inspired Cognitive Architectures, BICA 2018 (Ninth Annual Meeting of the BICA Society), Held August 22-24, 2018 in Prague, Czech Republic, 145, 158–168. https://doi.org/10.1016/j.procs.2018.11.024
Dirican, C. (2015). The Impacts of Robotics, Artificial Intelligence On Business and Economics. World Conference on Technology, Innovation and Entrepreneurship, 195, 564–573. https://doi.org/10.1016/j.sbspro.2015.06.134
Djordjevic, N. (2019, January 28). AI Statistics About the Latest Trends and AI Developments in 2019. Retrieved 29 April 2019, from Tech Jury website: https://techjury.net/stats-about/ai/ Druckman, J. N. (2001). On the Limits of Framing Effects: Who Can Frame? The Journal of
El Beheiry, M., Doutreligne, S., Caporal, C., Ostertag, C., Dahan, M., & Masson, J.-B. (2019). Virtual Reality: Beyond Visualization. Journal of Molecular Biology, 431(7), 1315–1321. https://doi.org/10.1016/j.jmb.2019.01.033
Elahi, E., Weijun, C., Zhang, H., & Nazeer, M. (2019). Agricultural intensification and damages to human health in relation to agrochemicals: Application of artificial intelligence. Land Use Policy, 83, 461–474. https://doi.org/10.1016/j.landusepol.2019.02.023
Farrow, E. (2019). To augment human capacity—Artificial intelligence evolution through causal layered analysis. Futures, 108, 61–71. https://doi.org/10.1016/j.futures.2019.02.022 Fassin, Y. (2009). The Stakeholder Model Refined. Journal of Business Ethics, 84(1), 113–135.
https://doi.org/10.1007/s10551-008-9677-4
Felix, R., Gaynor, G., Pevzner, M., & Williams, J. L. (2017). Societal trust and the economic behavior of nonprofit organizations. Advances in Accounting, 39, 21–31.
https://doi.org/10.1016/j.adiac.2017.09.003
Fraedrich, E., Heinrichs, D., Bahamonde-Birke, F. J., & Cyganski, R. (2019). Autonomous driving, the built environment and policy implications. Transportation Research Part A: Policy and Practice, 122, 162–172. https://doi.org/10.1016/j.tra.2018.02.018
Frahm, K. M., & Shepelyansky, D. L. (2019). Ising-PageRank model of opinion formation on social networks. Physica A: Statistical Mechanics and Its Applications, 526, 121069. https://doi.org/10.1016/j.physa.2019.121069
Francalanci, C., & Hussain, A. (2017). Influence-based Twitter browsing with NavigTweet. Information Systems, 64, 119–131. https://doi.org/10.1016/j.is.2016.07.012
Freberg, K., Graham, K., McGaughey, K., & Freberg, L. A. (2011). Who are the social media influencers? A study of public perceptions of personality. Public Relations Review, 37(1), 90–92.
Galloway, C., & Swiatek, L. (2018). Public relations and artificial intelligence: It’s not (just) about robots. Public Relations Review, 44(5), 734–740.
https://doi.org/10.1016/j.pubrev.2018.10.008
Gesualdi, M. (2019). Revisiting the relationship between public relations and marketing: Encroachment and social media. Public Relations Review, 45(2), 372–382. https://doi.org/10.1016/j.pubrev.2018.12.002
Goldberg, J. E., & Rosenkrantz, A. B. (2019). Artificial Intelligence and Radiology: A Social Media Perspective. Current Problems in Diagnostic Radiology, 48(4), 308–311. https://doi.org/10.1067/j.cpradiol.2018.07.005
Groß-Klußmann, A., König, S., & Ebner, M. (2019). Buzzwords build Momentum: Global Financial Twitter Sentiment and the Aggregate Stock Market. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2019.06.027
Grothe-Hammer, M. (2019). Organization without actorhood: Exploring a neglected phenomenon. European Management Journal, 37(3), 325–338.
https://doi.org/10.1016/j.emj.2018.07.009
Gupta, K., & Banerjee, R. (2019). Does OPEC news sentiment influence stock returns of energy firms in the United States? Energy Finance 2017, 77, 34–45.
https://doi.org/10.1016/j.eneco.2018.03.017
Hamet, P., & Tremblay, J. (2017). Artificial intelligence in medicine. Insights Into the Future of Medicine: Technologies, Concepts, and Integration, 69, S36–S40.
https://doi.org/10.1016/j.metabol.2017.01.011
Hellsten, I., Jacobs, S., & Wonneberger, A. (2019). Active and passive stakeholders in issue arenas: A communication network approach to the bird flu debate on Twitter. Public Relations Review, 45(1), 35–48. https://doi.org/10.1016/j.pubrev.2018.12.009
Holmes, J. H., Sacchi, L., Bellazzi, R., & Peek, N. (2017). Artificial Intelligence in Medicine AIME 2015. Artificial Intelligence in Medicine AIME 2015, 81, 1–2.
https://doi.org/10.1016/j.artmed.2017.06.011
Huang, W., Lu, H., & Su, B. (2016). Green disclosure? Social media and prosocial behavior. Working paper. University of Toronto.
Inuwa-Dutse, I., Liptrott, M., & Korkontzelos, I. (2018). Detection of spam-posting accounts on Twitter. Neurocomputing, 315, 496–511. https://doi.org/10.1016/j.neucom.2018.07.044 Ivaturi, K., & Chua, C. (2019). Framing norms in online communities. Information &
Management, 56(1), 15–27. https://doi.org/10.1016/j.im.2018.05.015
J. Zhang, R. Zhang, J. Sun, Y. Zhang, & C. Zhang. (2016). TrueTop: A Sybil-Resilient System for User Influence Measurement on Twitter. IEEE/ACM Transactions on Networking, 24(5), 2834–2846. https://doi.org/10.1109/TNET.2015.2494059
Jabeur, L. B., Tamine, L., & Boughanem, M. (2012). Active Microbloggers: Identifying Influencers, Leaders and Discussers in Microblogging Networks. In L. Calderón-Benavides, C. González-Caro, E. Chávez, & N. Ziviani (Eds.), String Processing and Information Retrieval (pp. 111–117). Springer Berlin Heidelberg.
