Identifying news stories within a large corpus of news texts is not a trivial task. An ideal automatic system should be able to identify the linguistic mentions of the reported events and of the associated actors in documents published at a time t, and then connect them to other mentions in other documents published at a later time, t+1, through a cross-document coreference relation [Lee et al. 2012]. Due to the complexity of the task and the low performance of such automatic systems [Beheshti et al. 2017], previous work has mainly employed topic detection and tracking techniques through the use of clustering. Clustering is an unsupervised machine learning technique that automatically organizes a collection of objects into smaller coherent groups. The end goal is to have clusters whose elements are as similar as possible between them (intra-cluster similarity) and as dissimilar as possible from the elements in the other clusters (inter-cluster similarity). This implies that we cannot determine a priori which criteria are going to define the different clusters.

Clustering has been mainly employed to divide a set of news according to their topic (e.g. politics, sports, etc.) [Huang 2008]. However, the use of clustering to identify news stories is very challenging. Some approaches have tried to improve the performance of this task by incorporating properties of news stories, such as timing features [Khy et al. 2008] [Yang et al. 1998] [Trilling and van Hoof 2020], named entities [Kumaran and Allan 2004] and burstiness [He et al. 2007]. In our case, timing and burstiness are measures we are trying to obtain from our system and, consequently, they cannot be part of the design. Named entities have also limited value, since the actors or places appearing in different news stories can be common between stories.

Later approaches have used Latent Dirichlet Allocation models to deal with news streams [Guo et al. 2013] [Hong et al. 2016] [Rao et al. 2016]. These techniques, however, focus on the extraction of topics rather than stories.

Another approach is timeline generation: automatically generated chronological sequences of events. This is done by using the temporal information attached or contained in single news documents. The timelines are usually centred on a target entity, giving rise to so-called entity-centric timelines [Shahaf and Guestrin 2010] [Baraniak 2019] [Minard et al. 2015]. Other work has proposed to go from timelines to storylines by taking into account the timelines of the actors, i.e. entities, co-participating in event mentions [Laparra et al. 2015] [Laban and Hearst 2017]. Storylines are thus represented as intersections (i.e. co-participation) of independent entity-centric timelines. Vossen, Caselli, and Kontzopoulou (2015) used the basic concepts of the narratology framework to bridge the relations between events of different timelines, starting with the most salient event of the storyline, i.e. the climax point.

A successful model for story tracking is the European Media Monitor [Gey et al. 2009]. This system runs every 10 minutes, and clusters all the news articles published during the last 4 hours. In every run, the system picks the clusters with the biggest number of articles, and compares them to the clusters that already existed in the last run. If two clusters have at least 10% of their articles in common, they are merged, becoming a single story. This system allows for a flexible time span and is language independent.

In this contribution, we use a similar model to the one in Gey, Karlgren, and Kando (2009). However, given that our data is not real-time, we employ a similar approach to Vossen, Caselli, and Kontzopoulou (2015) to account for time. For every pre-defined time window, we identify the most salient stories by clustering the corresponding news articles. Once this is completed, we approximate Vossen, Caselli, and Kontzopoulou (2015)’s proposal of bridging relations across time windows by means of similarity scores between pairs of story clusters (see Subsection B) to reconstruct the full story timeline.

Our pipeline for story extraction consists of (1) clustering the news articles of every week (i.e. the time window) and (2) connecting the clusters across relevant time spans, i.e. weeks, using each of the valid clusters as the climax of a story. The method is language independent, as clustering does not require any language-specific resource, such as lexicons or parsing models.[1]

Given a set of news articles from a fixed span (one year), we divide the corpus into smaller units (weeks). For each week, k-means clustering is applied.[2] The purpose of the clustering process is to detect the most salient stories of each week, spotting both short and long-span stories. We assign the number of clusters in a semi-automated way by manually reviewing the results of each week.

The value of the number of clusters per week (k) is assigned using the Elbow method. This method computes the clustering for k in the range from 1 to 30 (the range has been empirically chosen through a set of preliminary experiments) and stores the sum of squared errors (SSE) for every set of clusters. The optimal number of clusters is the one after which the SSE starts decreasing slower. This number can be easily detected by plotting the SSE in the y-axis and the number of k in the x-axis, which creates an elbow shape in the desired value.

In a collection of news articles, it is to be expected that not all articles are going to be part of a story. These isolated articles are gathered by the algorithm in one or two general clusters with very low SSE. These general clusters are detected and deleted using Silhouette analysis [Rouseeuw 1987]. This method plots all the documents in a cluster according to the silhouette coefficient. This coefficient compares the average distance from all data points in the same cluster and the average distance from all data points to the closest cluster. The resulting plot shows which clusters have gathered the articles that do not belong to any story. In Figure 1, clusters 8 and 10 are general clusters and will be removed from our data for further analysis.

In some cases, the same story is split into more than one cluster. To deal with this issue, we manually merge these clusters by comparing the keywords of their centroids. The centroids are the articles closest to the centre of the cluster and consequently the most representative of its content. In this step, some data cleaning is performed: clusters of international news are excluded (see Subsection C for details), clusters with non-political topics are excluded (e.g. the lottery, Christmas celebrations), and clusters with less than 15 articles are deleted. The process of manual revision takes around 5 minutes per week.

The clusters thus extracted are assumed to be the climax of different stories: they have been detected in weeks in which they are salient enough to form a cluster of over 15 articles. Nonetheless, stories are not limited to one week. Hereafter, for each climax (i.e. each extracted cluster), we link articles of the colliding weeks to the climax using cosine similarity. Henceforth, we will refer to the remaining clusters as climaxes (w).

It is often the case that the same story is detected as a climax in more than one week. In these cases, it is reasonable to expect that, after tracing the articles of the colliding weeks, an important proportion of the articles included in these stories are going to be part of both clusters. If more than 33% of the articles contained by a new story are already part of a previous story, and vice versa, the system merges them. The threshold of 0.33 is chosen based on empirical validation on our data.

This last step runs for the 52 weeks contained in the study, resulting in stories with very different lifespans and intensities that are the basis of our analysis. The method allows for one article to be part of more than one story. Nonetheless, this happens rarely.[3]

To extract the lifespan and intensity of the news stories, it is important to collect an exhaustive dataset of news that gathers all the articles published by one or more media outlets on certain topics during a fixed period. Available news corpora are usually from several media, time periods, and rarely care about exhaustivity. For this reason, we created our own corpus by scraping all articles in 2018 from four major Spanish newspaper outlets: El Pais, El Mundo, ABC, and El Periódico. We narrowed the focus to news about politics, collecting for El Mundo and ABC news from the section España, and for El Pais and El Periódico from the section Política. Political news about Spain are included in these categories, with the main difference that El Pais and El Periódico also include news about international politics in it. International news articles are removed during the cleaning of clusters (see Subsection B). An overview of the resulting corpus is illustrated in Fig. 3.

Overall, the corpus contains 50,385 news articles: 15,245 from El Pais, 7,401 from El Mundo, 20,448 from ABC, and 7,291 from El Periódico. For each news article, the following data are collected: the date of publication, the title, the text, the URL, and the keywords (as specified by each news outlet). In total, an average of 950 articles per week were published. ABC is the biggest publisher, with an average of 385 articles per week. El Pais follows with an average of 293. El Periódico and El Mundo published on average 137 and 139 articles a week, respectively.[4]

The manual annotation of the news stories was performed following a decision tree based on four questions. The annotation was conducted by one of the authors who is a Spanish native speaker and has extensive knowledge on Spanish politics. Figure 4 graphically illustrates the decision tree used for the manual annotation of the stories into issue types.

The first question of the tree distinguishes the actor(s) who put(s) the news in the public agenda (Q1:’“Who put the news on the public agenda in the first place?”’). If the answer to this question is “Media”, “Political Actors” or “Real-World Conditions/Citizens”, then the story is classified as sensational, governmental or prominent, respectively. In case the answer cannot be clearly identified, the annotator has to move to the second branch of the tree, which checks for the impact of the story in the public by assessing the obtrusiveness of the story (Q2:’“Is a large part of the population going to perceive the consequences of the outcome of this story?”’). A negative answer restricts the choice either to sensational or governmental issues. The exact labelling is triggered by a further question using the dramatism attribute (Q3:’“Is it a dramatic issue?”’). A positive answer leads to a sensational story, otherwise the story is labelled as governmental. In case of a positive answer to Q2 the annotator is posed with another question focused on the concreteness of the story (Q4: ‘“Is it a concrete issue?”’). In case of a positive answer, the story is labelled as prominent, otherwise it is considered as a governmental one.

The performance of the semi-automated system for news stories extraction has been evaluated with a subset of 2,666 manually annotated news articles. The gold standard was created by one of the authors by inspecting the main issue in each article and assigning each of them to a story. The measures used to evaluate are Purity and Normalized Mutual Information (NMI).

Purity is a measure of the extent to which clusters are well formed, i.e. contain a single class. To compute Purity, each of the clusters is assigned to its most frequent class, according to the gold standard. Then, the accuracy of this assignment is measured. The Mutual Information score is a complementary measure that assesses the quality of the clustering. In particular, MI provides a measure of the reduction in the entropy of clusters that we compare to the gold labels. This measure is sensitive to the number of clusters. NMI is the normalization of this value between 0 and 1, where 1 means a perfect classification.

The evaluation scores resulted in a Purity of 0.896 and NMI of 0.625. To gain a better understanding of the limitations and reliability of our approach we conducted a manual exploration of the aggregated stories. In particular, we used a contingency matrix to see the distribution of errors among the gold and automatically generated clusters of stories. Figure 5 visually illustrates the contingency matrix-based analysis of the errors against the manually aggregated stories.

Overall the proposed system works quite well, with very few errors. However, major classification issues concern the story labelled as Cluster 4. This cluster corresponds to the story about the referendum for the independence of Catalonia. This was the largest story of 2018 in Spain, as the referendum that took place at the end of 2017 had a lot of political consequences in 2018. This story represents a challenge for our methodology, as it spanned over the entire year, with articles appearing almost every day. This results in a very large story with multiple climaxes. As the cluster gets bigger (by iteratively adding new articles from succeeding weeks), it becomes more difficult for the clusters emerging from new climaxes to share 33% of the articles with the existing bigger cluster (see end of Section 3.B). As a result, this story is represented with multiple clusters instead of just one. We considered applying a more fine-grained clustering to distinguish more specific sub-stories from the more general story Catalan independence referendum. However, this harms the identification of smaller stories about other topics, making them disappear. A single story being the main story throughout the year is not a common situation. But it can happen in cases of big events like Brexit in the UK in 2018 or Covid-19 in 2020. To obviate these problems and also to be able to better generalize our observations, we decided to exclude this large story from our analysis.

When excluding this cluster, we observe an improvement of the NMI (from 0.625 to 0.754). Finally, in Figure 6, we observe that all the differences between the gold data and the output of our semi-automated system are articles that the system ignored (articles in 0 were assigned by the annotator but not by the system). This means that there are no articles assigned to the wrong cluster; the system only missed articles. We therefore conclude that the results are reliable enough to carry on our analysis.

Sensational stories are, in general, ephemeral and intense. This can be easily observed in Figure 7, where the smoothed conditional mean of the intensity of each sensational story is illustrated, with the climax in x=0. The rising action of sensational stories is the shortest of the three types. On average, there are 3 days from the first day of reporting to the climax of a sensational story.

The most common narrative pattern for sensational stories is the following: a rising action of 0 to 1 days and a falling action of 3 to 5 days (median). There are 13 out of 40 stories following this structure.

In Figure 8, we have labelled the 13 stories by topic and adjusted the size of the shapes by average intensity. The topics that more often recur with this narrative pattern are weather incidents, social injustice, and corruption. These are topics that trigger indignation, but they directly affect very few people. They generally span for one single day.

We further observe that stories about weather incidents get to its climax the same day that they emerge, having 0 days of rising action.[7] This is not always the case in stories about corruption and social injustice.

Three stories are clear instances of social injustice, all concerning court decisions about political freedoms. Two of them are decisions from the European Court of Human Rights (ECHR) that contradict previous decisions of a Spanish court[8], and a third one is the final resolution from a Spanish court, which will later be appealed to ECHR.[9] These three stories follow a similar content structure. They start with an article that explains the court decision, with very similar headlines in all newspapers. Then, some newspapers report about the reactions of the government and the political actors in favour or against the decision. In the case of the two European court decisions (in which the tribunal ruled in favour of political freedom), the reactions are published the same day. In the one coming from a Spanish court (which was accused of ruling against freedom of speech), the reactions are not reported until the day after, making the second day of the story its climax. This shows either less previous attention to the case, or a more unexpected decision compared to the other two court cases, which had already been in the news in the past. The days of the falling action mainly consist of articles about the precedents and the consequences, and opinion pieces.

We observe a similar pattern in stories about corruption. On the first day of reporting, multiple articles about the event itself are published. These are followed by reactions from other political actors, and some articles about the context. Finally, in the last days, the articles about the consequences and the opinion pieces are the most common. It can be noticed that two of the stories get to their climax on the second day of reporting. In the first case, the corruption case forces the resignation of the politician involved, making this second event the main climax of the story.[10] This story spans two days, and there are two days of climax. However, in the second story of this kind,[11] reporting starts the day before the court decision. This is an interesting case, as one newspaper published several articles about the background of the corruption case the day before the final sentence. This unusual behaviour seems to indicate a willingness to emphasize this particular story by this media outlet.

Governmental stories are often political discussions or legal debates, which have set timings, often known in advance by journalists. The planned nature of governmental stories gives space for a longer build-up to the story, which results in a longer rising action. In Figure 9, it can be observed that governmental stories tend to have more than one peak (we will also call these sub-climax). For this typology, the curve of the smoothed conditional mean in the main climax is less pronounced and multiple peaks can be observed along the span of the stories (thin lines), mainly in the rising action, but also in the falling action. The rising action of governmental stories takes on average 6.5 days, a difference statistically significant with respect to sensational stories. The average falling action, on the other hand, is similar to sensational stories, lasting around 4 days.

At first glance, there are no clear narrative patterns for stories belonging to this issue type. In Figure 10, it can be observed that stories with 3 to 6 days of rising action and 0 to 2 days of falling action are common. These stories have generally one single peak (one climax). However, the big majority of the stories in this typology do not fit in the main pattern represented in Fig. 10. We do observe a common feature between the non-represented stories: most of them have 2 to 4 different peaks (1 climax and 1 to 3 sub-climax). In Figure 10, these stories are identified in dark red. We observe two kinds of narrative patterns in governmental stories, (1) short-spanning stories with one climax, and (2) multiple sub-climax stories.

The first kind of stories are generally planned events, with a relatively low media interest, such as the birthday of the King, a vote on the internal statutes of a party, or the visit of the President to Cuba. In the days previous to the climax, these stories start growing by publishing articles about the context, previous similar events and the details on how the central event is going to take place. Also, some opinions of known political actors and some speculation about unknown facts. On the central day of the event, detailed descriptions about how it went are published. In most of the cases, the effects or possible consequences of the central event are published the same day, leaving 0 to 1 days of falling action. If there are opinion pieces, which is not always the case, they are part of the rising action.

The stories with multiple climaxes are discussions from legislative bodies, ideological struggles between parties, and law approval debates at regional, municipal, or state level. The multiple peaks in this kind of stories are often due to disagreements on the issue by different actors. These stories start like the first kind: they are planned, and the press knows when the involved characters are going to talk about the issue. However, after the first peak, in which a personality has shown its position/decision, new peaks come from other actors disagreeing and asking for an alternative decision.

An example of this pattern can be found in the story about the renovation of the General Council of the Judiciary. One first peak emerges when the new members of the council are announced.[13] Though, a larger peak is created when the person who was appointed to be the president of the council announces he will not take the position.[14]

Another example of this pattern emerges when the government announces that children in Catalan schools will be able to study in Spanish.[15] The announcement of the government generates the main climax of the story. Some days after, a smaller sub-climax emerges when a court declares a previous measure to facilitate the schooling in Spanish in Catalonia unconstitutional.[16] Finally, a third smaller sub-climax closes the story with the announcement that the Government will take some time to rethink their policies on the regularisation of Spanish in Catalan schools.[17]

Since there are only 9 stories of the type prominent, it is risky to extract many conclusions from them. Figure 11 illustrates the smoothed conditional mean of the intensity of stories belonging to this type, excluding one case: the story of pensions, which is further illustrated in Fig. 12.

It is safe to observe that 6 out of 9 of the prominent stories have a falling action of 1 to 3 days. The rising action, on the other hand, is more varied. These stories are mostly social demands expressed as demonstrations and protests, and the development of policies related to these demands. These observations suggest that media attention to prominent issues is infrequent and fades short after its climax.

We have excluded the story about pensions from Fig. 11 because the plot of this story is very different from the other prominent stories. This story is unique in the whole corpus: it has the longest rising action (38 days) and the longest falling action (24 days). It has 8 different peaks, and the highest one has 34 articles published on a single day. The pensions system is a recurrent topic in Spanish politics, as it is public and universal, and the biggest expense of the national budget.

In 2018, the yearly uprising of pensions was 0.25%, while prices had risen around 3%. This caused indignation, and the first demonstrations took place in the middle of January. However, Pensions does not emerge as a news story until the 8th of February, the day in which the Prime Minister announced some measures to incentivize private retirement funds. From then on, there were demonstrations every week in many cities, with a large one on the 22th of February. After this date, the opposition parties started discussing the topic and the media coverage focused on the politicians, even though the demonstrations kept taking place. On the 14th of March, there was a Parliament session focusing exclusively on the topic of pensions. This day is the highest peak of the story. The Parliament session did not have a clear outcome, and even bigger demonstrations took place after that. By that point, the topic of pensions had become the main point in the political agenda, becoming the trading issue for parties negotiating the national budget (which is a governmental topic).

In the story about pensions, we observe that the constant efforts of the social movement brought a prominent issue into the space that is usually taken by governmental issues. This suggests that prominent stories can in fact become very salient, but they require a sustained mobilization.