Introduction

One of the most relevant aspects in the rapid and innovative evolution of the digital environment in the last decade has been the implementation of Internet tools to extend the communication networks, widening the media’s target. Therefore, it follows that mass media from around the world have confidently opted for the use of interactive tools and social media to publish, inform and interact with their readers and audiences (^{Caballero, 2001}; ^{García de Torres et al., 2008}; ^{Lasorsa, Lewis & Holton, 2012}; ^{Said-Hung et al., 2013}).

This paper presents the results of a large-scale quantitative study that addressed issues such as the use of the web in journalistic production (^{Micó, Canavilhas, Masip & Ruiz, 2008}) and the efficiency and speed with which the mass media channels generate knowledge (^{Rogers, 2003}). The main objective is to characterize the modeling and diffusion process of news topics in social media, using Big Data as a way of understanding the world in the age of information, where in order to produce more knowledge, more data is necessary (^{Mukherjee & Shaw, 2016}; ^{Ularu et al., 2012}). Thus, this philosophy, that has permeated different areas of study, provides us with an adequate structure of thought to approach research questions that seemed unattainable before, although the metrics limit to some extent the scope of the analysis by the very nature of the data; that is, because of its great size and complexity.

For such reasons, this study used Topic Modeling (among other techniques of automated data analysis) and an entire year of publications (54 878 tweets) from the Twitter account of the Colombian newspaper El Tiempo (@ElTiempo).

The studies that have addressed the diffusion of news have been centered on the understanding of important news or world-class events (^{Greenberg, 1964}; ^{Henningham, 2000}; ^{Rogers & Seidel, 2002}). In the 1960’s, research pointed to the fact that a news story could take between one or two days to complete the dissemination process, even when having a great spread in the traditional media (^{Deutschmann & Danielson, 1960}). This, without a doubt, has drastically changed with time, and especially with the advent of the Internet and social media.

The concept of news diffusion has been notably influenced by the work of ^{Rogers (2003)}, for whom the news can be examined from the point of view of prominence, a concept that expresses the degree of importance of a news event as considered by individuals. For this reason, it is understood that out of all communication channels that are spread in the media spectrum, the public chooses and structures the news they consume. With the surge of social media, the study of news diffusion is becoming a field of study that has been developing dramatically as new forms of informing and communicating emerge. This level of complexity is due, as well, to the unpredictable nature of the occurrence of a news topic, combined with its rapid dissemination (^{Rogers, 2000}).

The variety of the news disseminated through the social media makes it so that the study of the diffusion of news topics through long periods of time allows for the modeling of these specific topics within more general topics. In fact, most of the longer-term topics have been found to be labelled as “politics”, “business” or “news”, and that similar news topics tend to temporally organize themselves into topic chains (^{Kim & Oh, 2011}). ^{Kim & Oh (2011)} found that a few “unique topics” or short-duration topics that appear when using the “topic modeling” tool are incoherent (although some of them could represent relevant events such as the death of someone famous or the increase of aviation safety). This is due to the fact that topic modeling is a method that learns topic structures from large collections of documents without human supervision (^{Arora et al., 2013}), a necessary development taking into account that, with all the information available online, we have reached the point that it is humanly impossible to process it all (^{Blei, 2012}).

^{Zhao et. al. (2011)} addressed the issue of Twitter being just another news feed that was faster than traditional news media, tapping into the unsupervised topic modeling, a way of extracting topics (underlyingsemantic topics), using only the words that were found in a set of documents (^{Blei & McAuliffe, 2007}). In this way, then, each tweet can be associated to a topic, and each topic to a specific category (^{Zhao et al., 2011}). The previous discussions as well as the growing dynamics of news diffusion from traditional media using social media have led us to examine this process in the Twitter account of a media outlet of national scale and ask ourselves:

RQ1a: What are the news topics that emerge from the tweets published by the information channel in Twitter from the Colombian newspaper El Tiempo during a one-year period?

On the other hand, diverse research works have contributed to establishing the characteristics of the journalistic information transmitted through the Twitter channels and what properties of the messages are responsible for making the users be willing to follow the informational accounts of this social network (^{Argüelles & Muñoz, 2012}; ^{Lotan et al., 2011}; ^{Schultz & Sheffer, 2012}; ^{Stubbs, 2001}; ^{Ure & Parselis, 2013}; ^{Wasike, 2013}). Besides the formal characteristics of the content found in social media (length, links, etc.), the tone or the sentiment expressed in the text has been one of the categories that can better characterize the content, as they allow us to elucidate automatically if the messages contain positive, negative or neutral sentiments in their structure (^{Leetaru, 2012}).

There are different types of sentiment analysis (^{Feldman, 2013}), but their general aim is to make a machine able to process and evaluate sentiments (^{Kechaou, Ammar & Alimi, 2013}). Notwithstanding, although limited (^{Stieglitz & Dang-Xuan, 2013}), their use has been propagated in the analysis of various types of content such as blogs, review sites, datasets and microblogging (^{Vinodhini & Chandrasekaran, 2012}) and have diversified (^{Meena & Prabhakar, 2007}; ^{Turney, 2002}), unifying it with topic modelling (^{Cai, Spangler, Chen & Zhang, 2010}). The tone or sentiment within a message can be clearly identified as an innovative characteristic or property of the content. Knowing that the characteristics of the innovations are related to the process of diffusion in social media (^{Peslak, Ceccucci & Sendall, 2010}) as well as to any news item or news topic (^{Rogers, 2003}), we ask:

RQ1b: What were the characteristics or innovative properties of the news topics disseminated?

RQ2a: Did these characteristics influence the diffusion of news topics?

Social media has created a more complex ecosystem in terms of duration and distribution of news products (^{Newman, Dutton & Blank, 2012}), and therefore debates have been opened on the emotional responses as a function of the diffusion of news (^{Ibrahim, Ye & Hoffner, 2008}), the topics, entities and relationships exposed in news articles and social media (^{Kang, O’Donovan & Höllerer, 2012}; ^{Newman, Chemudugunta, Smyth & Steyvers, 2006}), the temporal dynamics of the messages with respect to specific events (^{Jungherr, 2014}) and the acceptance or rejection of the information (^{Emery, Szczypka, Abril, Kim & Vera, 2014}).

One of the challenges in addressing these topics when one works with large scale data set from a long period of time comes from the identification of the diverse stages of the process. That includes the discovery of the topics, finding those that are similar and group them, finding short-term issues within the topics and identifying how these changes as a function of time (^{Kim & Oh, 2011}). In this sense, this research study will try to learn:

RQ1c: How are the news topics disseminated through time?

RQ2b: What was the relationship between the time or moment of production of the message and the appearance of news topics?

Lastly, although there have been many studies on social media with analysis techniques of large quantities of data (^{Asfari, Hannachi, Bentayeb & Boussaid, 2013}; ^{Bogdanov, Busch, Moehlis, Singh & Szymanski, 2013}; ^{Gerber, 2014}; ^{Ghosh & Guha, 2013}; ^{Ferrari, Rosi, Mamei & Zambonelli, 2011}; ^{Michelson & Macskassy, 2010}; ^{Paul & Dredze, 2014}), little has been done to address the importance of the message’s authorship. In platforms such as Twitter, the channel could be referring to the author that emits the message, having as a possibility that the message was created by the account holder or by a third party that the account holder simply re-tweets (RT), which can have implications on the dynamics of the production of information. Therefore, it is necessary to determine:

RQ1d: What were the channels or types of message authorship that originated the news topics?

RQ2c: In what way did the channel or type of authorship affect the appearance of news topics?

Method

Findings

The number of tweets emitted by the newspaper El Tiempo in Twitter in the year 2013 (N= 54 878) was lowest in the months of June and August (2 544 and 4 133, respectively), and highest in the month of February (5 266 messages) (January= 5 122, March= 4 414, April= 5 085, May= 4 722, July= 4 557, September= 4 280, October= 5 249, November= 4 769, December= 4 501). From June 9th to 20th, there were no tweets found due to a technical error in the application used to capture the messages from the @ElTiempo account, which explains the atypical results in this month.

The three days of the week with the highest number of tweets (emitted by @ElTiempo in 2013) were Thursday (n=9 155), Tuesday (n=9 114) and Wednesday (n=8 675). Figure 1 shows the visualization of the diffusion of tweets throughout an average week in 2013, considering that Monday starts at hour 0 and Sunday ends at hour 166. We can see that each day of the week had a similar behavior, with a peak of message emission around 10:00 in the morning, and a notable decrease starting at 8:00 in the evening. These results allowed us to clearly see that the diffusion process corresponded to the working days of the outlet.

Source: the authors.

Figure 1 Total number of tweets per hour during the week. Cumulative for 2013 Profile @ElTiempo 

To answer the question of which news topics emerged from the @ElTiempo Twitter channel in 2013 and how they were diffused (RQ1a), the underlying topics were modeled with unsupervised topic modeling, by which keywords were repeatedly clustered, and then the predominance of each topic in each message was determined:

The results showed 39.1% of the tweets were mainly related to the topic Politics/National Interest/Conflict, which is composed of themes such as armed conflict, violence, national events of great scale and incidents corresponding to political figures and governmental institutions. In second place, we found messages associated to the topic Sports/Entertainment/General (27.7%), followed by Venezuela/International (10.3%) and Residual topics (23%). In this last classification, messages could refer to localized social issues, gender, “isolated” news or other topics that, due to their nature and frequency, could not be unequivocally associated with any of the major themes identified in the corpus. In this group we found examples like the following:

Visualizations of how these topics were spread in time in the year 2013 (RQ1c) are shown in Figure 2. These time series clearly show that there is no underlying regularity in the appearance of topics, so we cannot speak of a seasonal series, which suggests that other characteristics -different to temporality itself- are responsible for the variation in each one of the four topics.

Source: the authors.

Figure 2 Temporal evolution of topics in the year 2013 

With the aim of describing innovative characteristics or properties of the news topics disseminated in the year 2013 (RQ1b), the tone of the message as well as its formal aspects were identified. In the first place, we found that an important number of the messages containing some of the four news topics did not possess indications of subjectivity. This means that they could be classified as neutral (39.2%). However, a clear sentiment in the content was detected in most of them, with 19.8% of the messages having a predominantly positive tone, and 19.7% having a predominantly negative tone.

Likewise, the formal characteristics of the messages were analyzed, and the words that were most used in the news topics were found. Amongthese, there is a clear predominance of the words “Colombia” (2 971 cases) and “Bogotá” (2 391 cases) in the corpus of the tweets. These were listed along with words that referred to conflict, delinquency related events or violence (“paz” [peace], “gobierno” [government], “farc”, “presidente” [president], “policia” [police], “Santos”). It should be noted that the presence of the words “mundial” [world wide] (723 cases) and “mujer” [woman] (482 cases) were also found among the most used words. When looking at the average number of characters (M= 93.91, SD= 21.25), we can see that the messages did not use the maximum number allowed at that moment (140 characters), using an average of 12 words per tweet (M= 11.65, SD= 3.98). The data reveal that most of the messages included links (M= 0.88, SD= 0.35) to other websites, and in lesser number, mentions to other users (M= 0.35, SD= 0.63) and labels or hashtags (M= 0.33, SD= 0.54).

An analysis of the mentions (@) found in the messages revealed that the most-mentioned profiles (@Portafolioco, @ElTiempo, @FUTBOLRED, @CityTV) belonged to the same publisher as the newspaper (Casa Editorial El Tiempo). Likewise, it was observed that the most-mentioned public figure accounts were @JUANMANSANTOS, @PETROGUSTAVO, @FALCAO and @BARACKOBAMA. To understand the type of authorship or channel used to disseminate the news topics (RQ1d), an analysis of the retweets (RT) was performed on the @ElTiempo profile. We found that 18.6% of the news topics disseminated all through the year 2013 originated from sources other than the account holder (retweets), while the percentage of their own tweets was 81.4%.

Next, with the objective of verifying the relationship between the appearance of the news topics and: a) the innovative characteristics or properties of the message referring to the tone as well as formal aspects (RQ2a); b) the time or moment of production (RQ2b); and c) the channel or authorship of the message (RQ2c), a series of linear regressions were performed. We verified the absence of multicollinearity (with tolerance values close to 1 and the variance inflation factors (VIF) below 5). The results are summarized in Table 1.

The analysis shows that only regression models for topics 2 (F(4.54872)= 84.483, p< .000), 3 (F(4.54872)= 11.731, p< .000) and 4(F(4.54872)= 44.187, p< .000) can explain part of the variance of the appearance of news items as a whole, although that explained variance is low (6%, 1% and 3%, respectively). Notwithstanding the data reveal that the innovative properties of the message, meaning their formal characteristics and tone, were linked to the dissemination of topics 2, 3 and 4 with low but statistically significant coefficients. Similarly, we see that time or moment of production was a determinant factor in the same topics. In the case of message authorship (RT or not), we observed that it only explained the dissemination of news topics 3 and 4. Even though the influence of these factors is low, the fact that they appear to model the appearance of almost all the news topics is significant. This suggests that even if the link is weak, the diffusion of news topics is not fortuitous, but responds to these and other patterns that should be studied.

Discussion and conclusion

In this study we have examined the dissemination of news topics on Twitter based on the modeling of underlying patterns in a longitudinal corpus of informative messages of the social account of a communication outlet, which has allowed us to explore the variables that influence the appearance of themes. The search for patterns can be applied both in social media and in online media, making it possible to explore media agenda problems and, even, to manifest certain practices that are difficult to attend, such as “fake news” or false news, since the algorithms of detection of underlying topics can be modeled with different approaches.

According to the results of the study, the topic Politics/National Interest/Conflict was found in the largest part of the messages published in the Twitter account of the Colombian newspaper El Tiempo in the year 2013. If we examine the most important news events in Colombia in 2013 (such as the start of peace talks between Santos’ Government and the Colombian Revolutionary Armed Forces (FARC-EP) at the end of 2012 and the elimination rounds of the FIFA World Cup Brazil 2014), we see that they correspond, besides the topics found, with keywords (peace, government, FARC, president, police, Santos, against, world and worldwide, among others).

The data suggests that there is a relationship between the mostmentioned profiles, the more frequent keywords and the main underlying news topics (topics 1 and 2) found with the topic modeling algorithm, as the results of all these items fit within the categories politics or sports.

When analyzing the number of characters as an innovative property, it is important to note that the tweets that use all of the space provided by the platform (140 characters) are merely 4.5%, a total of 2 495 cases. This could suggest that the tendency of the outlets is to publish messages that are relatively short, summarizing information using just a few words, and giving concise answers to other users of the platform. We could also highlight the fact that 71.4% of the tweets lacked mentions, and that 86.5% had links. The first could suggest a low degree of interaction with other users in the platform or the possibility that the other people/entities named in the messages did not have a Twitter account, and the second could point to a high rate of re-directing from this platform.

The range of published tweets varied between 4 000 and 5 000 (permonth), evidencing a notable diminishing in the production of tweets on Saturdays (5 585) and Sundays (5 873) with respect to the other days of the week. This situation changed with the increase shown on Mondays (there were 2 136 more tweets than on Sunday), which could be linked to journalistic dynamics and the volume of information generated during the weekend. Previous literature has suggested that time lapses (as far as social and mental constructs) help to create a reality (^{Stubbs, 2001}). Therefore, the days of the week could be associated with certain types of messages or emotions that are transmitted, by a person or, in this case, a medium of news diffusion (relaxation or sports associated to Sundays, leisure with the other weekend days, etc.).

On the other hand, the low frequency of retweets (18.6%) indicated low levels of content replication. This means that the newspaper generates scarce moments of visibility for messages from other accounts. While maintaining a high percentage of originality of its publications, the account studied redirected the users to their website or other online resources, turning the external websites into the frame of reference for reading their tweets. Then, even if their own messages are short, they provide more information that is hosted in other domains. Despite this, we can highlight that 70% of the tweets did not have hashtags. This limits in some way the inclusion of the messages in conversations different from those generated by the news diffusion medium.

Although the statistical models used only explained little variance found in the news topics, we can highlight that the innovative properties or characteristics and the time or moment of production of the message appeared to be significant predictors in 3 out of 4 news topics modeled. The channel or message authorship appeared to be a significant predictor in only two topics, one of them referring to political and conflict subjects, and the other to residual topics. This association could be due to the fact that generally, in order to report on political or conflict subjects, the sources (defined as channels in this study) are usually cited verbatim to back the information. This could invite us to think about using the messages from other profiles (i.e. political figures) to communicate this type of news events. However, as these data were not consistent enough, we believe that other complementary analyses are necessary.

Considering what was presented previously, we can argue that these results shed light, using the theory proposed by ^{Rogers (2003)} and the advanced computational methods, on some characteristic elements of the diffusion of news in social media. Also, the automated analysis of content methods used, topic modeling and the automated sentiment analysis, among others, were shown to be convenient for this kind of work, by becoming a theoretical-practical support and being helpful in the study of journalistic dynamics. Having found a similarity in numbers between messages classified as positive and negative, as well as a good part of the corpus labeled as neutral, we can infer that there is a distance in terms of linguistic subjectivity in the studied messages. This implies that the media outlet, at least in its Twitter profile, avoids adjectives when reporting events.

One of the limitations of the study was the lack of data from June 9th to June 20th, 2013, caused by a technical problem in the application that was used to compile the tweets. This could have caused a slight bias when observing that month’s data, but we must consider that it is impossible to recover the messages that were lost in that timeline, unless the source (in this case El Tiempo newspaper) provides them. Another limitation was the scarcity of previous research on Latin America that could have had similar characteristics such as the size of the sample as well as the techniques used. This is the reason why the comparison of empirical data with other equivalent explorations was not possible.

Besides what was already mentioned, we can add that as Twitter is a platform where the messages have a short lifetime (as the oldest messages disappear as new tweets are published). Another limitation was the gathering of an archive that encompassed a longer timeframe, as this implies having open access to the content stored by the owners of the profiles. Lastly, and with respect to the theory by ^{Rogers (2003)} in our research, a limiting factor appeared when we could not include the social system as a category of analysis, due to the characteristics of Twitter and the perspective we used.

In future research studies, it would be relevant to ask about the inner workings of the process of adoption of these news topics in social media from the perspective of the subject, how short-term news topics appear and disappear as a function of time, and what models could be created to try to explain the variance of other message characteristics from the messages found in platforms such as Twitter.