Nuclear energy: Twitter data mining for social listening analysis

Vol.:(0123456789) 1 3 Social Network Analysis and Mining (2023) 13:29 https://doi.org/10.1007/s13278-023-01033-8 ORIGINAL ARTICLE Nuclear energy: Twitter data mining forsocial listening analysis EnaraZarrabeitia‑Bilbao1· MaiteJaca‑Madariaga1· RosaMaríaRio‑Belver2· IzaskunÁlvarez‑Meaza1 Received: 8 December 2022 / Revised: 18 January 2023 / Accepted: 20 January 2023 © The Author(s) 2023 Abstract Knowing the presence, attitude and sentiment of society is important to promote policies and actions that influence the development of different energy sources and even more so in the case of an energy source such as nuclear, which has not been without controversy in recent years. The purpose of this paper was to conduct a social listening analysis of nuclear energy using Twitter data mining. A total of 3,709,417 global tweets were analyzed through the interactions and emotions of Twitter users throughout a crucial year: 6months before and 6months after the beginning of Russian invasion of Ukraine and the first attack on the Zaporizhzhia NPP. The research uses a novel approach to combine social network analysis methods with the application of artificial neural network models. The results reveal the digital conversation is influenced by the Russian invasion of Ukraine. However, tweets containing personal opinions of influential people also manage to enter the digital conversation, defining the magnitude and direction of the debate. The digital conversation is not constructed as a public argument. Generally, it is a conversation with non-polarized communities (politics, business, science and media); neither armed conflict or military threats against Zaporizhzhia NPP succeed in rousing anti-nuclear voices, even though these events do modify the orientation of the sentiment in the language used, making it more negative. Keywords Nuclear energy· Twitter· Social network analysis· Artificial neural networks· Russia–Ukraine conflict 1 Introduction It is an undeniable fact that in recent years there has been an intense worldwide public debate on nuclear energy (DiazMaurin 2014; Diaz-Maurin and Kovacic 2015). The urgency of mitigating climate change, as well as dependence on fossil fuels, has reopened the debate on this energy source (Friederich and Boudry 2022). Thus, while its detractors call for its abolition due to its dangerousness and the problems derived from waste generation (Greenpeace 2022; NRDC 2022), its defenders fundamentally see it as a way toward decarbonization (IAEA 2020; NEI 2022). The United States (historical supporter of this energy source and country with most nuclear power plants in operation (IAEA-PRIS 2022) is cautiously embracing nuclear power, despite certain environmentalists’ persistent concerns, to help achieve its goal of a net-zero carbon economy for the nation by 2050 (USDS-USEOP 2021). Japan (country operating the third most nuclear power plants), where nuclear power has been anathema since the Fukushima Daiichi nuclear power plant disaster in 2011 (Schneider and Froggatt 2012), is advocating the operation of reactors that have remained inactive for ten years since the accident (Reuters 2021). There is no doubt that nuclear energy-related developments have also been compulsive in Europe. In November 2021, France (country operating the second most nuclear power plants) announced that it will relaunch the construction of nuclear reactors in the country (LeFigaro 2021), thus reversing the discussion on an energy source that seemed obsolete after the catastrophe at the Fukushima plant. In addition, on New Year’s Eve 2021, nuclear power experienced a partial shutdown in Berlin and a resurgence in Brussels: Germany shut down three of its six operating nuclear power plants as part of the approved plan to shut down all atomic production by the end of 2022 (Joly 2021). However, at the same stroke of midnight, the European Commission announced its pioneering proposal to change the classification of green energy to include * Enara Zarrabeitia-Bilbao enara.zar[email protected] 1 Faculty ofEngineering, University oftheBasque Country (UPV/EHU, Bilbao, Spain 2 Faculty ofEngineering, University oftheBasque Country (UPV/EHU, Vitoria-Gasteiz, Spain Social Network Analysis and Mining (2023) 13:29 1 3 29 Page 2 of 17 nuclear power and natural gas as green energy (European Commission 2022). In this scenario, however, an event that has changed the worldwide debate about nuclear energy has been the Russian invasion of Ukraine on February 24, 2022. The armed conflict between Russia and Ukraine has undoubtedly influenced energy policies around the world and has led countries to push for a rapid transition to greener energy sources. Nuclear power has been in decline since the Fukushima disaster more than a decade ago, but the energy and environmental crisis is prompting reappraisal of the industry. Hence, the Russian invasion of Ukraine; rising natural gas prices; and Russian energy dependence have reopened the debate on whether nuclear power can help solve the challenges of energy security and climate change. Nonetheless, the March 4, 2022, attack of the Zaporizhzhia nuclear power plant (NPP) in Ukraine (the largest nuclear power plant in Europe) has caused enormous social upheaval. The attack on the Zaporizhzhia NPP is the first attack, with pernicious intent, on a nuclear power plant (without a radiation accident). There have been previous nuclear disasters such as Kyshtym (USSR, 1957), Windscale Piles (UK, 1957), Three Mile Island (USA, 1979), Chernobyl (USSR, 1986) and Fukushima (Japan, 2011); however, none were intentionally provoked. Knowing the presence, attitude and sentiment of society is important to promote policies and actions that influence the development of different energy sources (Ibar-Alonso etal. 2022). Hence, being aware that in the post-Fukushima era, nuclear energy became a delicate situation and public opinion became pessimistic (Kim etal. 2016), so analyzing how recent events have affected society can be considered important and of great interest to the scientific community. That is to say, what the social response on this occasion has been and whether it has led to any change in strategy or positioning with respect to this energy source. To this end, the interactions and discussions generated on Twitter will be analyzed. Twitter is considered a social network that represents an ideal scenario for diverse interactive audiences and a contemporary digital town square (Fernández-Gómez etal. 2018; Orbegozo-Terradillos etal. 2022). Based on this scenario, the purpose of this paper was to conduct a social listening analysis of nuclear energy. With this purpose, on the one hand, the relationships established (discussion communities generated) on the social network in relation to nuclear energy will be analyzed throughout a whole year and; on the other hand, the overall feelings and attitudes toward nuclear energy will be determined through sentiment analysis of tweets. All this, with special emphasis on the Russian invasion of Ukraine and the Zaporizhzhia NPP attack. 2 Twitter data mining andnuclear energy The information age or information society is defined by the social changes that have taken place since the final decades of the twentieth century, derived from the development of Information and Communication Technologies together with the development of networked social structures, which have impacted all areas of human activity (Mansell 2009; Castells 2010; Del-Fresno-García 2014). Today, a dense social network of interactions links people, information, events, places, etc., facilitating or limiting the flows of information, ideas and perceptions, among others, in an instantaneous and massive networked communication systems (Del-Fresno-García 2014). In this context, Twitter is a discursive space that has emerged in the heat of the digital agora, allowing people to make themselves visible: being seen and heard under the protection of the most diverse causes, slogans or situations of conformity or nonconformity (Baer 2016). Several studies have shown how Twitter plays a key role in times of crisis and conflict (Orbegozo-Terradillos etal. 2019). In addition, it makes it possible to capture a wide variety of data in real time, and retrospectively, providing access to data records of human activity in the digital sphere over time (Del-Fresno-García 2014). Nonetheless, the value of the data alone is not high; however, data are the raw material from which more value can be generated. The specific process used to harness data as a raw material and generate is known as data mining, and the operation essentially consists of capturing a series of information records and interpreting them (Morales-iGras 2020a). In this regard, numerous researchers have used Twitter data to achieve different objectives related to nuclear energy through the analysis of such data and using various data mining techniques. In this scenario, it is worth noting that most of the scientific works related to Twitter data mining and nuclear energy refer to the disaster at the Fukushima nuclear power plant. As a consequence of the Fukushima nuclear catastrophe, large amounts of radioactive materials leaked out, causing radioactive pollution of water. Hence, shortly after the catastrophe, public opinion was formed through various platforms, including digital social network services such as Twitter (Seung-Hoi etal. 2016). The exchange of information on digital social networks has far-reaching positive effects, such as real time and high broadcastability properties. Moreover, Twitter users are simultaneously consumers and contributors of information (Veil etal. 2011). However, in the Fukushima power plant disaster, the information that spread rapidly included misleading reports, such as the claim that iodine is useful for treating radioactivity as a substitute for stable iodine, Social Network Analysis and Mining (2023) 13:29 1 3 Page 3 of 17 29 and that became a problem (Aoki etal. 2018). Risk information transmitted by digital social networks had a strong influence on both the perception of risk and public reactions to an unusual catastrophe (Chung 2018). All in all, after the Fukushima nuclear power plant accident, risk and environmental communication has become a widely studied topic through the social network Twitter (Li etal. 2016). In the months following the accident, science-based tweets decreased; however, tweets with more emotional expressions began to spread and Twitter worked as a new public sphere especially in terms of anti-nuclear movements (Kim 2014). Likewise, it was observed that the role of influencers was crucial for spreading information (Tsubokura etal. 2018; Sano etal. 2021). In this context, with the emergence of social networks, new ways of expressing opinions have emerged, making the analysis of emotions and feelings displayed on digital platforms key toward better understanding users' opinions on a given topic (Arumugam etal. 2021). Hence, Twitter users’ emotional response after the Fukushima nuclear disaster, mainly negative and opposed to nuclear energy, has been studied in detail (Miura etal. 2015; Su etal. 2016; Kim etal. 2016; Hasegawa etal. 2020). Identifying the process by which people emotionally respond could be useful in risk communication when similar disasters occur in future (Hasegawa etal. 2020). However, despite the fact that the Fukushima incident has been widely analyzed through Twitter data mining, other fields of study as well as other geographical locations have also been covered, such as political debates about nuclear withdrawal initiatives or narrative policy frameworks (Arlt etal. 2018; Gupta etal. 2018). In all these studies, various data mining tools, techniques and methodologies are used. One of the techniques widely used is social network analysis (SNA) (Rantasila etal. 2018; Yagahara etal. 2018), which allows knowledge to be generated according to the relational structure of virtual interactions (Morales-i-Gras 2020b). In addition, increasingly sophisticated tools are being used to provide more precise answers to more complex realities. In this sense, within deep learning algorithms, artificial neural networks (ANNs) are used, especially for deciphering public opinion or analyzing emotions and sentiments toward nuclear energy (Liu and Na 2018; Khatua etal. 2020; Arumugam etal. 2021). However, these techniques are usually used separately, i.e., SNA techniques are not combined with ANN. 3 Research questions andmethodology The study focuses on the interactions and emotions of Twitter users regarding nuclear energy throughout a crucial year: 6months before and 6months after the beginning of Russian invasion of Ukraine and the first attack on the Zaporizhzhia NPP, i.e., from September 1, 2021, to August 31, 2022. Three specific periods have been studied: First period, before the beginning of the conflict; second period, the first three weeks of the conflict, when the first attack on the nuclear power plant took place; and third period, after the beginning of the conflict, i.e., after the first three weeks have elapsed. For that purpose, the research uses a novel approach to combine social network analysis (SNA) methods with the application of artificial neural network (ANN) models. In other words, a set of methods for the analysis of social interactions that specifically investigate relational structures and their representation as networks (SNA) are combined with algorithms modeled as elementary units or neurons connected in such a way to form a network capable of solving complex nonlinear problems (ANN) (see Analysis methods section). 3.1 Research questions In order to address the objective of social listening analysis on Twitter about nuclear energy using Twitter data mining, the following questions have been used as a guide: RQ-1: What have been the main characteristics of the overall conversation about nuclear energy that has taken place throughout a year? Are there morphological differences between the conversations generated before, during and after the beginning of the Russian invasion of Ukraine and the attack on the Zaporizhzhia NPP? RQ-2: Which have been the main communities generated in the overall conversation about nuclear energy throughout a year? How have these communities changed before, during and after the beginning of the Russian invasion of Ukraine and the attack on the Zaporizhzhia NPP? RQ-3: Who have been the leaders in the overall conversation about nuclear energy throughout a year? How have these influential players changed before, during and after the beginning of the Russian invasion of Ukraine and the attack on the Zaporizhzhia NPP? RQ-4: What has been the sentiment about nuclear power throughout a year? How has the sentiment expressed in nuclear energy-related tweets changed before, during and after the beginning of the Russian invasion of Ukraine and the attack on the Zaporizhzhia NPP? 3.2 Research methodology 3.2.1 Collecting text data foranalysis The first step, in order to address the main objective of the study, was data extraction for the Twitter social network and data preparation. Social Network Analysis and Mining (2023) 13:29 1 3 29 Page 4 of 17 In this regard, first, the search query considered to select the tweets that make up the sample of this study was: “nuclear energy” OR nuclearenergy OR “nuclear power” OR nuclearpower. Subsequently, the tweets (from September 1, 2021, to August 31, 2022) were downloaded using the Twitter API for Academic Research (Twitter 2022) and Twarc (a command line tool and Python library for collecting and archiving Twitter JSON data via the Twitter API (Twarc 2022)). The data were collected retroactively in two batches. On August 25, 2022, tweets sent from September 1, 2021, through July 31, 2022, were downloaded, and on September 1, tweets sent throughout August 2022 were downloaded. Once the tweets were obtained, data preprocessing was necessary for subsequent data analysis. Thus, OpenRefine open-source software was used for data cleanup and transformation (OpenRefine 2022). This cleaning and transformation consisted, mainly, in extracting mentions in order to synthesize different networks based on which users mention which other users in the conversation itself. 3.2.2 Analysis methods The second step was the empirical approach of the research, developed in two stages. In the first stage, networks metrics and main communities were analyzed (answering RQ-1, RQ-2 and RQ-3), and in the second stage, the emotional stage of the digital discussions was studied (answering RQ-4). To obtain the different discussion networks generated about nuclear energy (a network that takes into account the discussion for the whole year, another network for the first period, another for the second period and the last one for the third period), all the mentions in the digital conversations were extracted, synthesizing networks based on which users mentioned other users in the conversation itself. The resulting networks were exported to the social network analysis open-source software Pajek (Batagelj and Mrvar 2002), and in order to analyze the established relationships between users of the digital conversation sphere, different metrics were obtained, at both the global and individual level (at network level and at nodal level). Hence, through the global metrics, the morphology of the networks was analyzed, and through the metrics at the nodal level, the key actors of the conversations were identified. The Louvain multi-level algorithm was applied to identify the communities that have participated in the digital conversations. This algorithm allows densely interconnected groups of nodes (Twitter user) to be generated, as well as obtaining the best network partitions or the highest modularity figure. Thus, the groups or clusters representative of the digital conversation for the different periods under analysis were identified. The synthesized networks and communities detected in the previous steps were further processed with Gephi software, an open-source tool of proven validity for improving the visualization and analysis of large network graphs (Bastian etal. 2009; Orbegozo-Terradillos etal. 2022). Gephi’s Force-Atlas 2 algorithm (Jacomy etal. 2014) was employed because of its usefulness in bringing nodes that form part of the same communities closer together and away from those with which they are less algorithmically related. In the second stage of the empirical approach, artificial neural networks were used to process the tweets and achieve the goal of analyzing the sentiments generated in those tweets. Sentiment analysis, also known as opinion mining or emotion classification, is a combination of natural language processing (NLP) and text mining, with the aim of analyzing the text data of social media, among others, and having that information, mine the user’s emotions (Salloum etal. 2017; Jain and Kaushal 2018). It has been detected that artificial neural networks totally surpass more traditional machine learning models (Jain and Kaushal 2018). Moreover, the key to achieve a high accuracy model seems to be in creating an architecture that combines different deep neural networks (Sosa 2017; Kamiş and Goularas 2019; Umer etal. 2021). Therefore, this study used a model adapted from Periwal (2021) that produces a sentiment analysis of tweets using deep learning algorithms. Thus, the architecture of the neural network model is led by the Bi-LSTM and attention as detailed below. To do that, first of all, unprocessed text data from the sentiment140 dataset were gathered and preprocessed (Go etal. 2009; Kaggle 2022; Sentiment140 2022). This dataset contains 1,600,000 tweets extracted using the Twitter API, which have already had their sentiment classified as a positive, labeled as 1, or negative, labeled as 0. There are 800,000 tweets of each classification type respectively, which means that the dataset is not skewed. After preprocessing, data were split into a training dataset and a test dataset. Then, text was tokenized and padded. The output achieved is the input for our sentiment classification neural network model, which is a deep learning sequence model. The model has an architecture composed of the following layers: an embedding layer created using the word2vec model to convert tweets into word vectors (Mikolov etal. 2013); a Bi-LSTM neural network layer to capture the semantic meaning of the text (Elfaik and Nfaoui 2021; Chandra etal. 2021); an attention mechanism layer to extract the most relevant words; a dense layer which adds a fully connected layer in the model and that the argument passed specifies the amount of nodes in that layer; and finally, a last dense layer with a sigmoid activation function to achieve the emotion classification (Fig.1). After training and testing the described model, it was applied to unlabeled tweets related to nuclear energy, returning the emotion of each of them. Social Network Analysis and Mining (2023) 13:29 1 3 Page 5 of 17 29 Figure1 summarizes the used sentiment analysis neural network model (Kaggle 2022). Figure2 summarizes the methodological process applied in the study. 4 Results anddiscussion The research collected a total sample of 3,858,024 tweets. Since English terms were used for the search, almost all the tweets were in English (96.15%), so in order to maintain homogeneity throughout the study, it was decided to work with the tweets in English, i.e., with a sample of 3,709,417 tweets. Figure3 shows that, as expected, the distribution of tweets has not been homogeneous over time. Specifically, unusual activity is observed with 606,055 tweets (16.3% of total tweets) on March 4, 2022. The attack on Zaporizhzhia, Europe’s largest nuclear power plant (CNN 2022), generated global social alarm reflected in the social network activity. Regarding the three periods under study, 18.5% of the tweets collected correspond to period I, 35.7% of the tweets correspond to period II and 45.8% of the tweets correspond to period III (Fig.3). Thus, the impact that the conflict has had on the intensification of the digital conversation about nuclear energy can be observed. By isolating the three periods for a clearer view of what happens in each of the periods, Figs.4, 5 and 6 show that the distribution of tweets has not been homogeneous over time. Different protagonists, topics or events have led to increased user activity on the social network. The first period (Fig.4) shows odd activity on December 8 and 9, due to a tweet posted by Elon Musk supporting nuclear power plants: “Unless susceptible to extreme natural disasters, nuclear power plants should not be shut down.”1 Likewise, on February 13, an American alt-right political activist, Jack Posobiec, posted a tweet related to American politics, which had a great response: “Biden’s National Fig. 1 Sentiment analysis neural network model applied in the study Fig. 2 Methodological process applied in the study Identify main query terminology (“nuclear energy” or nuclearenergy or “nuclear power” or nuclearpower) and timespan (09/01/2021-08/31/2022) Twitter data extraction & preparation Collect data (Twitter API for Academic Research and Twarc) Data cleaning and preparation (OpenRefine) Data analysis & visualization with big data techniques Networks metrics and main community analysis (Pajek, Gephi, PowerQuery) Sentiment analysis (Deep learning algorithms: model using word2vec, BiLSTM and attention mechanism) Achievement of objective / conclusions Social Network Analysis Artificial Neuronal Networks 1 https:// twitt er. com/ elonm usk/ status/ 14686 89628 16105 2674. Social Network Analysis and Mining (2023) 13:29 1 3 29 Page 6 of 17 Security Advisor was involved in a illegal spying operation on the previous President. He is now escalating tensions with a nuclear power in Eurasia bc Biden’s poll numbers are down.”2 In addition, on February 22, the proximity of the outbreak of the Russian invasion is noted on the network. The second period (Fig.5), as previously emphasized, shows the impact of the Russian invasion of Ukraine on network activity related to nuclear energy and the great social upheaval generated by the attack on the Zaporizhzhia NPP. In the last period (Fig.6), the unusual network activity on July 6 is due to the European Parliament's acceptance of the inclusion of nuclear energy and gas as environmentally sustainable economic activities (European Parliament 2022). Finally, the upturn in the number of tweets in August is, above all, due to the upsurge of continuous Russian attacks of the Zaporizhzhia NPP and tweets in favor of nuclear energy posted by influential actors such as Elon Musk: 0 100000 200000 300000 400000 500000 600000 1-Sep 5-Sep 9-Sep 13-Sep 17-Sep 21-Sep 25-Sep 29-Sep 3-Oct 7-Oct 11-Oct 15-Oct 19-Oct 23-Oct 27-Oct 31-Oct 4-Nov 8-Nov 12-Nov 16-Nov 20-Nov 24-Nov 28-Nov 2-Dec 6-Dec 10-Dec 14-Dec 18-Dec 22-Dec 26-Dec 30-Dec 3-Jan 7-Jan 11-Jan 15-Jan 19-Jan 23-Jan 27-Jan 31-Jan 4-Feb 8-Feb 12-Feb 16-Feb 20-Feb 24-Feb 28-Feb 4-Mar 8-Mar 12-Mar 16-Mar 20-Mar 24-Mar 28-Mar 1-Apr 5-Apr 9-Apr 13-Apr 17-Apr 21-Apr 25-Apr 29-Apr 3-May 7-May 11-May 15-May 19-May 23-May 27-May 31-May 4-Jun 8-Jun 12-Jun 16-Jun 20-Jun 24-Jun 28-Jun 2-Jul 6-Jul 10-Jul 14-Jul 18-Jul 22-Jul 26-Jul 30-Jul 3-Aug 7-Aug 11-Aug 15-Aug 19-Aug 23-Aug 27-Aug 31-Aug Nº of tweets Date (years 2021-2022) Period I Period II Period III Fig. 3 Daily trend of tweets throughout the year (from September 1, 2021, to August 31, 2022) 0 2000 4000 6000 8000 10000 12000 14000 Nº of tweets Date (period I) Tweet posted by Elon Musk supporting nuclear power plants Tweet posted by Jack Posobiec related to American politics Days prior to the Russian invasion Fig. 4 Daily trend of tweets over the first period (from September 1, 2021, to February 23, 2022) Fig. 5 Daily trend of tweets over the second period (from February 24, 2022, to March 16, 2022) 0 100000 200000 300000 400000 500000 600000 700000 Nº of tweets Date (period II) Attack on the Zaporizhzhia NPP 2 https:// twitt er. com/ jackp osobi ec/ status/ 14928 74434 72574 4643. Social Network Analysis and Mining (2023) 13:29 1 3 Page 7 of 17 29 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 Nº of tweets Date (period III) European Parliament's acceptance of nuclear energy as environmentally sustainable economic activity Continuous attacks of the Zaporizhzhia NPP and Tweet posted by Elon Musk in favor of nuclear energy Fig. 6 Daily trend of tweets over the third period (from March 17, 2022, to August 31, 2022) 3 https:// twitt er. com/ elonm usk/ status/ 15632 92201 04343 1424. Table 1 Metrics extracted from the analyzed conversations (Zarrabeitia-Bilbao etal. 2022b) a Number of users who have participated in the conversation, i.e., only users who have interpellated or have been interpellated. Metrics analyzed Overall networks for different periods Global Period I Period II Period III Total impacts (tweets or retweets) 3,709,417 684,961 1,325,015 1,699,441 Users (nodes)a1,243,554 350,377 574,937 668,543 Average impacts (per user) 2.98 1.95 2.3 2.54 Arcs (interactions) 3,767,699 758,625 1,436,134 1,697,983 Density 0.00000244 0.00000618 0.00000434 0.00000380 Average degree 6.05956637 4.33033561 4.99579606 5.07965232 Maximum distance 28 24 28 32 Average distance 7.27924 6.64913 10.04885 7.61276 Input degree centralization 0.07459838 0.05662457 0.07531547 0.06023185 Output degree centralization 0.00236256 0.00425212 0.00143234 0.00362949 Betweenness centralization 0.01229997 0.00685957 0.00209164 0.00790441 Number of clusters 20,985 12,186 9,326 13,388 Modularity (Louvain multi-level algorithm) 0.688192 0.760213 0.665745 0.717920 “Countries should be increasing nuclear power generation! It is insane from a national security standpoint & bad for the environment to shut them down.”3 All in all, it is observed that the role of influencers such as Elon Musk is crucial for spreading messages. 4.1 Networks metrics andmain community analysis (RQ‑1, RQ‑2, RQ‑3) 4.1.1 Network metrics Metrics extracted from the analyzed conversations show the main characteristics of the conversations about nuclear energy in the different periods and morphological differences between them (Table1). As mentioned in the previous section, the number of impacts in the second period (1,325,015 impacts in 3weeks) was relatively large (35.7% of the impacts for the whole year). This was somewhat foreseeable due to the media and social commotion generated by the Russian invasion of Ukraine and the attack on the Zaporizhzhia NPP. It can be seen that the most cohesive network is that of the first period (density 0.000618%); however, the low-density figures indicate that there are still many strategic interpellations to be explored by the nuclear energy ecosystem. Moreover, input degree centralization is higher in the case of the second period (7.53%) than for the other periods. The group of nodes that receive many mentions is larger than in the rest of the cases, placing us in a scenario of a less horizontal dialogic. However, it is observed that all centrality values are low. Thus, it is inferred, both for the different periods and for the overall set, that small groups of actors Social Network Analysis and Mining (2023) 13:29 1 3 29 Page 8 of 17 did not capitalize the reception of mentions in the digital discussions (input degree centralization); that there was no single group that posted most of the mentions (output degree centralization); and that the digital network is distributed horizontally rather than being monopolized by a few users (betweenness centralization). Finally, in all cases, the modularity is greater than 0.3, which indicates, for the whole time span studied, a community structure of great mathematical significance (Newman and Girvan 2004). All in all, except for the high relative activity in the second period, there are no significant morphological differences in the networks of the different time spans analyzed. 4.1.2 Main community analysis Regarding the central characters of the digital discussion, Figs.7, 8, 9 and 10 illustrate, per different time span studied, which eight main communities (representing in all cases more than 3.4% of the actors in the network) have participated in the digital conversation and the position that each community holds.4 Tables2, 3, 4 and 5 show who the key players of the conversation and what the main topics discussed in each of the communities have been. On the one hand, the analysis of the protagonists of the conversations has been conducted by analyzing both the unweighted and weighted input degree. The first of these metrics indicates the number of “neighbors” a user has without counting the number of times the user talks to the “neighbors”, i.e., which users have the largest number of unique audiences. However, this metric is combined with the weighted input degree, since for long study periods, it indicates that a user is important not only in a conjunctural way, having only once been retweeted a lot, but also in a structural way. And, on the other hand, the content that has been important in each of the communities has been determined by carefully reading the most “viralized” tweets in each community. First of all, it can be seen that there is alignment between the agents with the highest input degree and the highest weighted input degree. In addition, no account displays, a priori, suspicious behavior. No user has a suspiciously high weighted input degree level compared with the unweighted input degree level, so that, a priori, no account is artificially promoted by other automated accounts. From the community and leader analysis carried out, it can be inferred that, in general, actors related to politics, business, science and media lead the digital discussions about nuclear energy on Twitter. Therefore, this social network is a dissemination tool that works, in this case, as a means of expression for specific communities: Namely technical or collegiate communities, with a knowledge or interest in the topic, and that support nuclear energy as a national, international and environmental strategy. Fig. 7 Most important communities’ network throughout the year (from September 1, 2021, to August 31, 2022) 4 The order of the colors will be maintained for all the networks and will be in accordance with the size of the community (from largest to smallest). Therefore, the fact that a color is repeated from one network to another does not mean that it is the same community, in fact, from one network to another there will always be differences in the communities. Social Network Analysis and Mining (2023) 13:29 1 3 Page 9 of 17 29 Fig. 8 Most important communities’ network over the first period (from September 1, 2021, to February 23, 2022) Fig. 9 Most important communities’ network over the second period (from February 24, 2022, to March 16, 2022) Social Network Analysis and Mining (2023) 13:29 1 3 29 Page 16 of 17 Batagelj V, Mrvar A (2002) Pajek—analysis and visualization of large networks. Springer, Berlin, Heidelberg, pp 477–478 Castells M (2010) The Rise of the Network Society, 2nd edn. Chandra S, Kumar Gourisaria M, Gm H, etal (2021) Semantic analysis of sentiments through web-mined twitter corpus. In: International semantic intelligence conference,. Delhi, pp 25–27 Chung IJ (2018) Interactive dynamics of media communication, risk perception and public response to Fukushima nuclear disaster. Korea Assoc Policy Stud 27:281–303 CNN (2022) March 4, 2022 Russia-Ukraine news European Commission (2022) EU Taxonomy: Commission begins expert consultations Del-Fresno-García M (2014) Haciendo visible lo invisible: visualización de la estructura de las relaciones en red en Twitter por medio del análisis de redes sociales. Prof La Inf 23:246–252. https:// doi. org/ 10. 3145/ epi. 2014. may. 04 Diaz-Maurin F (2014) Going beyond the nuclear controversy. Environ Sci Technol 48:25–26. https:// doi. org/ 10. 1021/ ES405 282Z/ ASSET/ IMAGES/ LARGE/ ES201305282Z_ 0001. JPEG Diaz-Maurin F, Kovacic Z (2015) The unresolved controversy over nuclear power: a new approach from complexity theory. Glob Environ Chang 31:207–216. https:// doi. or g/ 10. 1016/j. gloen vcha. 2015. 01. 014 Elfaik H, Nfaoui E-H (2021) Deep attentional bidirectional LSTM for Arabic sentiment analysis in Twitter. In: 2021 1st International conference on emerging smart technologies and applications (eSmarTA), pp 1–8 Fernández-Gómez JD, Hernández-Santaolalla V, Sanz-Marcos P (2018) Influencers, personal branding and political ideology on Twitter. Cuadernos.info. https:// doi. org/ 10. 7764/ cdi. 42. 1348 Friederich S, Boudry M (2022) Ethics of nuclear energy in times of climate change: escaping the collective action problem. Philos Technol 35:1–27. https:// doi. org/ 10. 1007/ s1334702200527-1 Go A, Bhayani R, Huang L (2009) Twitter sentiment classification using distant supervision. CS224N Project Rep 1:2009 González V (2011) Educación para la ciudadanía democrática en la cultura digital. Comumicar 18:131–138. https:// doi. org/ 10. 3916/ C36201103-0 Greenpeace (2022) #Nuclear. https:// www. green peace. org/ inter natio nal/ tag/ nucle ar/. Accessed 6 Oct 2022 Gupta K, Ripberger J, Wehde W (2018) Advocacy group messaging on social media: using the narrative policy framework to study Twitter messages about nuclear energy policy in the United States. Policy Stud J 46:119–136. https:// doi. org/ 10. 1111/ psj. 12176 Hasegawa S, Suzuki T, Yagahara A etal (2020) Changing emotions about fukushima related to the fukushima nuclear power station accident—how rumors determined people’s attitudes: social media sentiment analysis. J Med Internet Res 22:18662. https:// doi. org/ 10. 2196/ 18662 IAEA (2020) Driving deeper decarbonization with nuclear energy. https:// www. iaea. org/ bulle tin/ drivi ngdeeperdecar boniz ationwithnucle arenergy. Accessed 6 Oct 2022 IAEA-PRIS (2022) The Database on Nuclear Power Reactors. https:// pris. iaea. org/ pris/ home. aspx. Accessed 8 Oct 2022 Ibar-Alonso R, Quiroga-García R, Arenas-Parra M (2022) Opinion mining of green energy sentiment: a Russia–Ukraine conflict analysis. Mathematics 10:1–22. https:// doi. org/ 10. 3390/ math1 01425 32 Jacomy M, Venturini T, Heymann S, Bastian M (2014) ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLoS One 9:e98679. https:// doi. org/ 10. 1371/ journ al. pone. 00986 79 Jain K, Kaushal S (2018) A comparative study of machine learning and deep learning techniques for sentiment analysis. In: 2018 7th International conference on reliability, infocom technologies and optimization (Trends and Future Directions) (ICRITO), pp 493–487 Joly J (2021) Germany begins nuclear phase-out, shuts down three of six nuclear power plants . euronews Kaggle (2022) Sentiment140 dataset with 1.6 million tweets. https:// www. kaggle. com/ datas ets/ kazan ova/ senti ment1 40. Accessed 4 Oct 2022 Kamiş S, Goularas D (2019) Evaluation of deep learning techniques in sentiment analysis from Twitter data. In: Proc—2019 Int Conf Deep Learn Mach Learn Emerg Appl Deep 2019 12–17. https:// doi. org/ 10. 1109/ DEEPML. 2019. 00011 Khatua A, Cambria E, Ho S, Na JC (2020) Deciphering public opinion of nuclear energy on Twitter. In: Proceedings of the international joint conference on neural networks. institute of electrical and electronics engineers Inc. Kim J, Brossard D, Scheufele DA, Xenos M (2016) “Shared” information in the age of big data: exploring sentiment expression related to nuclear energy on twitter. Journal Mass Commun Q 93:430–445. https:// doi. org/ 10. 1177/ 10776 99016 640715 Kim H (2014) 3.11 East Japan great earthquake and twitter: focusing on the relationship between SNS and anti-nuclear movements. 日 本硏究 273–295 LeFigaro (2021) Emmanuel Macron annonce la construction de nouveaux réacteurs nucléaires de type EPR Li N, Akin H, Yi-Fan LS etal (2016) Tweeting disaster: An analysis of online discourse about nuclear power in the wake of the Fukushima Daiichi nuclear accident. J Sci Commun 15:1–20. https:// doi. org/ 10. 22323/2. 15050 202 Liu Z, Na JC (2018) Aspect-based sentiment analysis of nuclear energy tweets with attentive deep neural network. In: Lecture Notes in Computer Science. Springer, pp 99–111 Mansell R (2009) The information society Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation of word representations in vector space. In: 1st International conference on learning representations. pp 1–12 Miura A, Komori M, Matsumura N, Maeda K (2015) Expression of negative emotional responses to the 2011 Great East Japan Earthquake: analysis of big data from social media. Shinrigaku Kenkyu 86:102–111. https:// doi. org/ 10. 4992/ JJPSY. 86. 13076 Morales-i-Gras J (2020a) Datos masivos y minería de datos sociales: conceptos y herramientas básicas. 32 Morales-i-Gras J (2020b) Minería de datos de los medios sociales, técnicas para el análisis de datos masivos. 33 NEI (2022) Voices for Nuclear Energy. https:// www. nei. or g/ voicesfornucle arenergy. Accessed 7 Oct 2022 Newman M, Girvan M (2004) Finding and evaluating community structure in networks. Phys Rev E 69:26113 NRDC (2022) Nuclear Power 101. https:// www. nrdc. org/ stori es/ nucle arpower101. Accessed 8 Oct 2022 OpenRefine (2022) A free, open source, powerful tool for working with messy data. https:// openr efine. org/. Accessed 28 Apr 2021 Orbegozo-Terradillos J, Morales-i-Gras J, Larrondo-Ureta A (2019) Feminisms outraged at justice: the online conversation on “La Manada” case. IC—Rev. Científica Inf. y Comun 2019:249–283 Orbegozo-Terradillos J, Morales-i-Gras J, Larrondo-Ureta A (2022) Twitter y la (de)construcción del mito: Maradona y el activismo digital feminista. Cuadernos.info 52:181–203. https:// doi. org/ 10. 7764/ cdi. 52. 34147 European Parliament (2022) Taxonomy: MEPs do not object to inclusion of gas and nuclear activities . https:// www. europ arl. europa. eu/ news/ en/ pressroom/ 20220 701IP R34365/ taxon omymepsdonotobjecttoinclu sionofg asan dnucle aractiv ities. Accessed 23 Oct 2022 Periwal N (2021) Twitter Sentiment Analysis using Word2Vec, BiLSTM | Kaggle. https:// www. kaggle. com/ code/ stoic static/ twitt ersenti mentanal y sisusingword2 v ecbilstm. Accessed 5 Nov 2022 Social Network Analysis and Mining (2023) 13:29 1 3 Page 17 of 17 29 Rantasila A, Sirola A, Kekkonen A etal (2018) #fukushima Five years on: a multimethod analysis of twitter on the anniversary of the nuclear disaster. Int J Commun 12:928–949 Reuters (2021) Japan’s new PM defends pro-nuclear stance in parliamentary debut. https:// www. reute rs. com/ world/ asiapacifi c/ japansnewpmkishi dafaceoppos itionquest ionsfirsttime20211011/. Accessed 9 Oct 2022 Rustam F, Ashraf I, Mehmood A etal (2019) Tweets classification on the base of sentiments for US airline companies. Entropy 21:1078. https:// doi. org/ 10. 3390/ E2111 1078 Salloum SA, Al-Emran M, Monem AA, Shaalan K (2017) A survey of text mining in social media: Facebook and Twitter perspectives. Adv Sci Technol Eng Syst J 2:127–133 Sano Y, Torii HA, Onoue Y, Uno K (2021) Simulation of information spreading on twitter concerning radiation after the Fukushima nuclear power plant accident. Front Phys 9:1–9. https:// doi. org/ 10. 3389/ fphy. 2021. 640733 Schneider M, Froggatt A (2012) 2011–2012 World nuclear industry status report. Bull at Sci 68:8–22. https:// doi. org/ 10. 1177/ 00963 40212 459126 Sentiment140 (2022) Sentiment140. http:// help. senti ment1 40. com/ home. Accessed 4 Oct 2022 Seung-Hoi K, Yu-I H, Meeyoung C, etal (2016) Public Discourse on Environmental Pollution and Health in Korea: tweets following the Fukushima nuclear accident. In: Social web for environmental and ecological monitoring. pp 187–190 Sosa PM (2017) Twitter Sentiment Analysis using combined LSTMCNN Models. Eprint Arxiv 1–9 Su LYF, Cacciatore MA, Liang X etal (2016) Analyzing public sentiments online: combining humanand computer-based content analysis. Information Commun Soc 20:406–427. https:// doi. org/ 10. 1080/ 13691 18X. 2016. 11821 97 Tsubokura M, Onoue Y, Torii HA etal (2018) Twitter use in scientific communication revealed by visualization of information spreading by influencers within half a year after the Fukushima Daiichi nuclear power plant accident. PLoS ONE 13:1–14. https:// doi. or g/ 10. 1371/ journ al. pone. 02035 94 Twarc (2022) twarc. https:// twarcproje ct. readt hedocs. io/ en/ latest/. Accessed 20 Feb 2022 Twitter (2022) Twitter API for Academic Research. https:// devel oper. twitt er. com/ en/ pr odu cts/ twitt erapi/ acade micresea rch. Accessed 20 Feb 2022 Umer M, Ashraf I, Mehmood A etal (2021) Sentiment analysis of tweets using a unified convolutional neural network-long shortterm memory network model. Comput Intell 37:409–434. https:// doi. org/ 10. 1111/ COIN. 12415 USDS-USEOP (2021) The Long-Term Strategy of the United States, Pathways to Net-Zero Greenhouse Gas Emissions by 2050. Washington Veil SR, Buehner T, Palenchar MJ (2011) A work-in-process literature review: incorporating social media in risk and crisis communication. J Contingencies Cris Manag 19:110–122. https:// doi. org/ 10. 1111/J. 14685973. 2011. 00639.X Yagahara A, Hanai K, Hasegawa S, Ogasawara K (2018) Relationships among tweets related to radiation: Visualization using cooccurring networks. JMIR Public Heal Surveill 4:1–14. https:// doi. org/ 10. 2196/ publi cheal th. 7598 Zarrabeitia-Bilbao E, Morales-I-Gras J, Rio-Belver RM, GarechanaAnacabe G (2022a) Green energy: Identifying development trends in society using Twitter data mining to make strategic decisions. Prof La Inf 31:1–16. https:// doi. org/ 10. 3145/ epi. 2022. ene. 14 Zarrabeitia-Bilbao E, Rio-Belver R-M, Alvarez-Meaza I, Martínez de Alegría-Mancisidor I (2022b) World environment day: understanding environmental programs impact on society using twitter data mining. Soc Indic Res 164:263–284. https:// doi. or g/ 10. 1007/ s1120502202957-y Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.