Introduction

Public perceptions of science have been studied extensively since the mid-twentieth century by means of public opinion polls like the Eurobarometers in Europe, the National Science Foundation (NSF) surveys in the United States, or the reports of the Spanish Foundation for Science and Technology (FECYT) in Spain, among others. Ever since they were introduced, these surveys have sought to measure public interest in, knowledge of, and attitudes towards science (Davis, 1958), although they have not been immune to criticism (Bauer et al., 2007; Pardo, 2001). Science communication in particular has often been presented as an essential strategy for fostering public engagement with science (European Commission, 2008).

Given that the current trend in science communication is to involve the public in scientific dialogue (Nisbet & Scheufele, 2009), the performance of science communicators on social networks, which have massive numbers of user accounts and high levels of participation, has become a question of particular interest. Academics have highlighted the need for a better understanding of how these new virtual environments affect science communication (Brossard & Scheufele, 2013). At the same time, social networks themselves offer the opportunity for close-range investigation of public debates about scientific issues, with special attention to new voices and different contexts (Kapoor et al., 2018; Shan et al., 2014), on media platforms where content is often accessed without mediators. It is for this reason that studies of social networks can constitute a useful complement to traditional surveys of public perceptions of science (Li et al., 2019). With respect to public participation, researchers have highlighted the need to assess public involvement in open discussions about science in these environments, considering aspects like ease of access to content, the type of information disseminated and even the type of audience, among other factors (López-Pérez & Olvera-Lobo, 2019). It has been suggested, however, that audiences that follow science accounts do not generally interact with them, as in most cases they use them merely to keep updated (Álvarez-Bornstein & Montesi, 2019). It is worth noting that the content generated every day on Twitter is intimately linked to current scientific developments (Veltri, 2013; Wilkinson & Thelwall, 2012; Zhao et al., 2011), a quality that makes it especially attractive for research on science communication (Büchi, 2016), and that constitutes one of the most powerful reasons for its selection for this study, along with the open access it offers to bulk data. On the other hand, it has been suggested that the best strategy for communication on Twitter is actively focusing on increasing followers rather than relying on keyword searches to make scientific content visible (Mohammadi et al., 2018). With this in mind, to test the methodology proposed in this article I have chosen to use the Twitter account of the popular science communicator Neil de Grasse Tyson, who has gathered a very large number of followers (more than 13 million in 2019).

The objective here is to present the possibilities offered by a methodology designed to analyse sets of tweets and assess the impact of communications on Twitter about scientific issues, with the aim of revealing unexplored dimensions of the public interest in science and attempting to answer a question that has been inspired by the results of traditional surveys: What triggers an interest in science? Scientific advances and discoveries or aspects associated with everyday human life, like cultural, political, or even emotional factors? Attempting to answer this question will help reveal trends that indicate the most effective ways of communicating science. My starting hypothesis is that the social network Twitter, understood as a space for public participation, could prove useful for learning about these issues. To address the questions outlined above, this study includes the following stages: (1) development of an algorithm in Swift programming language that can analyse large sets of data and measure the degree of interest and impact of information released on Twitter based on conversations about science; (2) application of the algorithm to a set of publicly available data extracted from the account of the astrophysicist Neil de Grasse Tyson; (3) graphic representation and interpretation of results; and (4) assessment of the scope of the study.

It should be noted first of all that there is very little consensus about what methods are reliable for research on Twitter and what information it can reveal to us (Veltri & Atasanova, 2015), since rigorous methodologies that would permit reliable systematic analysis have yet to be developed (Kahle et al., 2016). However, there are various exploratory studies of science communication on Twitter in specific areas; for example, the studies that have aroused the most academic attention are those related to public perceptions of risk, like research on the climate debate (Pearce et al., 2014) or studies of health issues aimed at understanding the emotional stance of the public (Becker et al., 2016). To study the content of tweets, computational techniques are used in order to systematically analyse large volumes of data, using text mining, natural language processes, network analysis, etc., and qualitative approaches with the participation of a human encoder who is aware of the conceptual context of the communication being researched and is therefore able to draw subtle information out of the tweets (Uren & Dadzie, 2015). This study therefore combines a mixed methodology to leverage the potential of both types of analysis.

It is worth noting that due to the format of Twitter, users express themselves in very brief terms, carefully choosing relevant words to reflect their ideas, which in principle makes it easier to explore key elements based on frequently used words. In this sense, the content of tweets allows for a semantic representation (Narr et al., 2011) that facilitates the analysis both of elements central to the communication, and of peripheral elements in simple terms, assigning them levels of relevance. Another common approach is what is known as “sentiment analysis” applied to tweet content, and in this study I also attempt to measure the emotional charge of the communications, given that emotional messages on Twitter are more likely to be retweeted (Stieglitz & Dang-Xuan, 2013; Veltri & Atanasova, 2015) and reflect the emotional perceptions of users expressed in natural language (Dehkharghani et al., 2014), generally identified based on predetermined lists of words.

Finally, it is also important to acknowledge the studies that suggest users are more inclined to tweet about their personal daily activities than about informational publications (80% compared to 20%) (Dann, 2010; Naaman et al., 2010). Twitter is, after all, both an information network and a social network (Myers et al., 2014), and its use as a research tool can take a wide range of approaches, from studying the potential to engage audiences by retweets (Kwak et al., 2010) to the activity of scientific journalists on the platform (Arrabal & De-Aguilera, 2016), or even the role of teachers in motivating and engaging their students (Santoveña & Bernal, 2019). There is also a widely accepted idea that personal profiles work better than institutional accounts in terms of interactions with the public (Pérez-Rodríguez et al., 2018).

Material and methods

Methodological proposal

For this study I propose a methodology aimed at identifying trends among audiences of science communication through an analysis of communications that are openly available on Twitter. The set of tweets to be studied can be collected easily through the application programming interface (API) offered by the platform itself (Twitter, 2019), for example, using a simple algorithm programmed in R language. In this case, it is recommended to clean the text of the messages with the tidytext R package (Silge & Robinson, 2016), which processes the text and prepares it for analysis, eliminating words with no semantic value like connectors or stopwords. Then, to analyse the content of representative sets of tweets, an algorithm in Swift language has been designed. The functions of this algorithm are described in this section. Quantitative techniques for performing systematic analysis are combined with a qualitative approach for classification of terms into categories.

First of all, to determine what the content in the sample is about and how it can be quantified in order to estimate its impact, the basic, indivisible unit of analysis used in this methodology is the word or term (Blei et al., 2003). The measurements to estimate impact are based on its frequency of use and on the number of retweets and likes associated with it, appropriately standardised based on the number of times the word appears in the sample. In addition, to reveal the extent to which a particular type of information is more interesting to the receiving audience than another, two other coefficients are proposed: “popularity” and “polemicity”. The idea underlying the popularity indicator is that the more a word is retweeted (i.e., the more retweets it accumulates) the more popular it is considered, also taking into account how often it appears in the sample. For example, an unusual word with a lot of retweets would be considered especially popular. To define this, the variable retweetRate (retweet ratio) has been used, providing a measurement of the interest aroused by the term, i.e., the extent to which it is shared. It is important to note that popularity may be either positive or negative, as content that a given user disapproves of may be retweeted and thus gain visibility. On the other hand, polemicity, understood as the degree of controversy triggered by the information, is defined as the ratio of the retweetRate to the favoriteRate (like ratio), which will be higher if the content of the tweet is retweeted widely but receives fewer likes. The variables described above and the popularity and polemicity indicators for words and categories are presented in Table 1, being comparable by definition between different datasets.

On the other hand, to determine the level of interest triggered by scientific content either on its own or correlated with other types of information in the tweet, a classification of words by category is proposed. This method has been used before in other studies investigating the relative interest in science compared to other topics on Twitter (Zhao et al., 2011). For this study, the categories proposed (which are described in Table 2) are the following: “science”, “culture”, “political-social”, “beliefs”, “media”, and “emotional”. These categories have been defined using my own criteria inspired by previous studies. Specifically, the category “culture” is drawn from proposals that explore cultural factors in studies of public perceptions of science (Bauer et al., 2012; Pardo, 2001); “political-social” is based on Twitter studies that highlight concerns of this type in relation to scientific controversies, usually in the area of climate change (Pearce et al., 2014); “beliefs” refers to the frequent interactions between science and religion and to the growing interest in studies of pseudoscience (Moreno-Castro et al., 2019); “media” is included due to its relevance to communication studies; and “emotional” is derived from studies that apply sentiment analysis to tweets, along with the observation that “an emotional connection [...] can be a powerful ‘way in’ to a science experience for non-experts, capturing initial attention and increasing feelings of bonding with the communicator or other participants” (Kaiser, 2014: 28).

The classification of words into categories is done manually based on their meaning. Computer-assisted research methods are not suited to this type of content analysis as such methods assume that the terms have the same meaning in any context (Matthes & Kohring, 2008), while the use of a human encoder can ensure a better interpretation of the context of the discussion. Thanks to the algorithm design, this categorisation is relatively simple when the words are organised in order of relevance in a list based on frequency of use and cumulative likes and retweets. At the same time, the algorithm also generates two files with nodes and edges organised in a visual representation in the form of a semantic network (for example, using the popular software Gephi) that shows the relationships between the metrics proposed and the content of the tweets collected, such as depicting the weight of each category in comparison with the others, or showing whether the scientific concepts present in the data sample are central to the communication or merely peripheral. The tasks required for the analysis are detailed in Table 3.

It is important to note that the list of categorised words is cumulative, which means that it will always grow in size and thus favour subsequent analysis applied to new sets of tweets, for which much of the work will already be done, apart from the classification of the most relevant words in the new communication by executing the algorithm. This will provide increasingly refined analyses of science communications, easily reproducible by other researchers and at a low cost in terms of intellectual and financial resources compared to public opinion polls, which require extensive work and substantial funding, and which have static results. Of course, this algorithm design is not restricted to the area of science, as its variables are compatible with other conceptual spaces. This tool is available to any researchers who may need it, simply by sending me a request via email.

Analysis and results: The impact of Neil deGrasse Tyson’s public communications

Analysis

After identifying the most relevant words with the algorithm, the top 1,250 terms were classified manually. These offer an idea of the thematic preferences of the user in question. Among the most frequently used words are scientific concepts like “Mars”, “space”, and “physics”. This result is to be expected given that the account holder is a science communicator who works in the field of astrophysics. Also appearing frequently are concepts unrelated to science itself, such as the word “film”, making reference to the cultural industry of cinema, and the word “happy”, referring to an emotional state.

Figure 1 contains a series of pie charts graphically depicting the relative importance of each category in Tyson’s communications, based on: 1) his own thematic preferences; 2) cumulative likes by category; and 3) cumulative retweets by category.

To estimate the impact of his communications, the other two pie charts show how these thematic preferences trigger different quantities of likes and retweets for each category. Here we find that the impact of words containing an emotional charge (e.g. “joy”, “shit”, “hostile” or “cry”) is markedly bigger than their representation in Tyson’s communications would suggest. Words referring to social or political issues also have a much bigger impact, while the impact of scientific terms is smaller. On the other hand, words containing cultural information and those referring to beliefs have an impact in proportion with their relative presence in his communications. The percentages for cumulative likes are: “science”: 28%; “emotion”: 35%’ “culture”: 9%; “political-social”: 23%; “media”: 1%; “beliefs”: 4%; while the percentages for cumulative retweets are: “science”: 29%; “emotion”: 36%’ “culture”: 9%; “political-social”: 20%; “media”: 1%; “beliefs”: 4%.

Finally, in order to gauge how popular or controversial the content of Tyson’s communications is, the popularity and polemicity coefficients were calculated for the different categories. These are also represented in Figure 1 with bar graphs (using comparative measurements whose values are not meaningful in themselves). The most popular communications are those related to beliefs, political or social issues and emotions, while the most unpopular are those related to media. Similarly, the level of polemical content in the communications is clearly higher in the category of beliefs than in any other category.

The final step was to find out what happens when scientific information is combined with other types of information, which was explored by representing the semantic network of Tyson’s communications, also classified into categories, showing the relationships and connections between words. This depiction can be found in Figure 2, where the size of the nodes represents: 1) the communication preferences of the communicator (a node is bigger when a particular term appears more often in the sample); 2) the terms that received the most retweets; and 3) the terms that received the most likes. As can be seen, Tyson constructs his communication through central scientific concepts (network on the left), but it is evident in the other two conceptual network depictions that the impact of information on users is bigger for terms that are peripheral to the communication, especially for the categories of “emotion” and “political-social”.

Discussion and conclusions

The results discussed in the previous section have two dimensions: the strictly methodological dimension and the dimension related to the case chosen for this study. Based on the analysis of the case study with the tool proposed, the most striking result is that scientific content gives way to other types of information in the collected set of tweets by the science communicator Neil deGrasse Tyson. Although Tyson’s communications are dominated by words associated with scientific content, the tweets with an emotional charge are the ones that receive the most attention from users, in terms both of likes and of retweets, followed by words referring to social and political content. Conversely, the media category is clearly the one with the weakest result in terms of the interest it arouses, and it also has the lowest popularity rate, despite the fact that its presence in Tyson’s communications is quite high. This phenomenon is left open to interpretation.

In light of these results, this study constitutes a data-based approach that suggests that science on its own is not as interesting to the public as other subjects. Although it only studies one specific case, it is worth highlighting that Tyson is a science communicator with a huge influence and a large number of followers, and the sample used for this study contains half of his tweets. It should also be noted that messages posted by Tyson go first to his followers, an audience supposedly interested in science (although they are also subsequently disseminated via retweets).

The values generated for the two proposed indicators of popularity and polemicity reveal some clear differences between the two in the individual categories that may constitute points of interest. For example, the “science” category, despite not being particularly popular with receiving users, exhibits a certain level of polemical content, perhaps because of Tyson’s social activism on issues like the climate crisis.

While the idea of polemicity is based on identifying which topics generate the most debate or controversy, popularity is a measurement of the attention they receive, regardless of whether that attention is positive or negative. The aspects revealed by these calculations might otherwise go unnoticed if there is a low incidence of words used in the sample in a particular category; this is the case, for example, of “beliefs”, which proves an extremely popular and controversial category, despite its minimal presence in the set studied. Presumably due to their sensitive nature, these are issues that do not predominate in the communications of the user studied but that have a big impact in the Twitter ecosystem. It should also be noted that while the results based on likes and retweets are quite somewhat similar, an interpretation for this that seems plausible is that the action of retweeting gives the message greater visibility, thereby increasing the potential number of likes.

With respect to word matches in the semantic network, represented to evaluate the levels of centrality of terms with differing degrees of appeal, it is clear for the sample examined that concepts peripheral to scientific discussion, i.e., referring to adjacent issues, are of more interest than central concepts, and are for the most part non-scientific. One possible interpretation for this is that there are particular subjects that suddenly attract attention, but they are not subjects that normally appear in Tyson’s communications. This question requires more in-depth analysis to identify the causes. On the other hand, the similarity between the two semantic networks depicting the cumulative number of likes and retweets in nodes may again be related to the fact that the spreading of a tweet on Twitter favours the accumulation of likes by giving greater visibility to it and exposing it to other users.

This case study thus hints at specific strategies for strengthening science communication: linking scientific information to socio-political issues and/or expressing it in emotional terms. Of course, as this study is limited to the particular case of Tyson’s tweets, the results obtained cannot be extrapolated to the whole study universe of science communication, and my intention is certainly not to make such a generalisation, but merely to point to a trend that should be researched in greater depth.

Given that the results of applying the tool proposed here to the particular case of a famous science communicator are reasonably consistent, it is worth asking: Would studies of the communications of other science communicators offer similar results? What differences would be identified in the case of institutional accounts, or if current scientific issues are the focus instead of users? It would thus be worthwhile to apply this tool to other specific Twitter profiles and to general discussions that receive significant media attention. Another incentive for further research is the potential of the tool to provide comparable periodic assessments and to support governments in the preparation of specific scientific communication plans or similar initiatives (to offer one example).

In short, due to the limitations of surveys to gauge public perceptions of science (Bauer et al., 2012; Pardo, 2001), this study has effectively considered a number of unexplored areas, confirming the potential of research on social networks to complement such surveys (Li et al., 2013). In particular, the confirmation with empirical data of the effect of emotional content in scientific communication is especially noteworthy (Kaiser, 2014).1