Anthropogenic climate change is one of the greatest challenges to date¹. Societal understanding and attitudes toward climate change are critical to successfully implementing both mitigation and adaptation measures. In democratic systems, public policy is largely shaped by voter preferences, as electoral dynamics play an important role in policy-making². For example, evidence shows that in Europe, faster adoption of renewable energy policies has been driven by a shift in public opinion in favor of environmental issues³.

Mass media play a key role in that context as public, policy, and media agenda all influence one another^4,5. On the one hand, media are responsive to policymakers, who are a source of information, and to the general public as their clients, both effects being reinforced by the emergence of social media^6,7. On the other hand, media can also direct public opinion and policymakers’ attention towards selected issues by prioritizing them over others and by framing them in a certain way^8,9. Accordingly, media coverage of climate change is central to informing the public with the potential to shape perceptions of climate change and thus the societal and political discourse^10,11. For example, in the most extreme cases, framing climate change as a controversial issue can lead to opinion polarization, especially on social media^12,13. However, public awareness is strengthened by frequent exposure to environmental media reports^10,14 and by coverage that frames climate change as a widespread concern and emphasizes its personal implications^15,16. Especially, the way the media present climate policy issues influences how people perceive their ability to engage in climate politics¹⁷. For example, framing climate change as an international problem can lead people to feel powerless, resulting in lower civic participation and engagement⁹. By contrast, highlighting the health impacts of climate change has been shown to increase public engagement¹⁸ and support for climate change mitigation and adaptation initiatives¹⁹. In a similar vein, media coverage of climate change has been found to influence human behavior, such as mobility choices²⁰ or investment decisions^21,22. At the same time, the discussion of solutions can also be lopsided. For instance, in the context of agricultural emissions, mitigation strategies holding individuals responsible, such as dietary changes, were discussed more prominently in the US and UK elite media than solutions requiring government policies and changes in agricultural practices¹¹.

Given that currently implemented climate policies still fall short of stabilizing warming to well below 2 °C as agreed upon in the Paris Climate Agreement and as necessary to reduce some of the most severe climatic risks²³, the question of if and how climate change is covered in the mass media—especially in comparison to other current issues—is more important than ever.

Previous studies have predominantly focused on the analysis of textual news data to assess the prevalence of climate change coverage, for instance, newspaper data²⁴. They find the frequency of climate change coverage to be volatile and driven by individual events, such as the Conference of Parties, reports by the Intergovernmental Panel on Climate Change, or local extreme weather events^25,26,27,28. In addition to the frequency of climate change reporting, the positioning of the reporting determines the prevalence. Sampei et al.¹⁴ find that the number of articles on global warming on the front pages of newspapers strongly influences public concern for global environmental issues in Japan. Nevertheless, only a few studies investigate prevalence measures other than the overall frequency of climate change-related articles. For instance, one study finds that only a small margin of climate change topics make it onto the front page of Tanzanian newspapers²⁹. The focus on textual news data, coupled with the narrow focus on coverage frequency, leaves a gap in our understanding of the prevalence of climate change media coverage.

In this study, we address this gap by moving beyond newspaper-based studies and focusing our analysis on climate change coverage on television (TV). TV news is particularly relevant, since they are the most used source of climate change news across markets worldwide^30,31,32. We focus on Germany as one of the world’s largest emitters of greenhouse gases³³ and on the coverage of the leading German TV news program Tagesschau³⁴, viewed daily by a broad audience of more than nine million people (in 2023)³⁵, about 11% of Germany’s population. This constitutes a market share of roughly 40% for TV news in Germany³⁵. Additionally, Tagesschau is considered to follow high journalistic standards and reports on a wide range of topics to a broad audience, making it the leading newscast on German TV³⁴.

The prevalence of climate change coverage relative to other news topics is assessed using two metrics. First, the frequency of climate change coverage is leveraged, a commonly used metric in the literature^25,26,27,36. Second, the time positioning of climate change coverage relative to the other news topics is used, as it reflects the editorial decision on the ranking of news items in Tagesschau, which typically begins with the most important story of the day³⁷. This ranking indicates the news value assigned to each news item.

We analyze a total of 3267 shows aired daily at 8 p.m. between 2015 and 2023. Of these, 2888 shows include subtitles, which allow us to extract the spoken text and transform it into a format suitable for text-based machine learning approaches. Each 15 min news show covers around ten separate news items, which we refer to as “news stories” from hereon. The individual news stories are inferred from the subtitles by matching the beginning of each news story to keywords from the news story headlines using natural language processing (NLP) techniques, such as lemmatization (see Methods for details). Then, the topics of the individual news stories are classified using a combination of a dictionary-based approach and neural topic modeling. Essentially, news stories are classified as climate change coverage if they include a climate change-related keyword. The remaining news stories are classified into general topic clusters using the state-of-the-art neural topic modeling framework BERTopic³⁸. Currently, conventional topic modeling approaches such as Latent Dirichlet Allocations^36,39,40,41 or Structural Topic Models^26,42,43 dominate in the climate change media coverage literature. These methods sort documents into topic clusters and account for their semantic context based on probabilistic bag-of-words approaches⁴⁴. More recently developed neural topic modeling algorithms, like BERTopic, account for semantic context by leveraging pre-trained transformer models^45,46, but have been used only sparingly so far^47,48.

The processed data comprises over 28,000 TV news stories, classified into 35 general topic clusters and five climate change-specific subtopic clusters. This allows us to examine how climate change is covered thematically and in relation to other news topics. Additionally, we gain insights into which climate change-specific subtopics are covered most prevalently. Our analysis reveals that climate change coverage accounts for roughly 4% of total news coverage, surpassed, for example, by sport coverage (9%). Acute crises, in particular, are covered more prevalently. Overall, climate change coverage is dominated by the discourse on climate policy (80%), while only 10% report on climate impacts. Moreover, climate impacts tend to be discussed in later news slots, indicating a lower news value.

Climate change coverage is increasing

Assessing the prevalence of climate change coverage in terms of frequency, our results indicate that climate change coverage accounts for about 4.4% of the total news coverage in Tagesschau over the observation period. Examining the frequency over time, we find climate change coverage to increase from about 2% in 2015 to about 7.8% in 2023. This trend is further evaluated by performing a linear regression on the monthly aggregated coverage data, which yields 0.156 additional climate change-related news stories per month (Supplementary Fig. B.7). That means about 1.9 additional climate change-related news stories are presented in Tagesschau each year. In total, about one third of Tagesschau shows contain climate change coverage, but climate change is only mentioned explicitly in 11.9% of the shows (Supplementary Fig. B.8). Climate change-related news stories that do not mention climate change explicitly often refer to it as “climate crises” or they mention other related terms, such as “climate protection” or “greenhouse gases” (Supplementary Fig. B.4).

To contextualize these findings, we compare the coverage frequency of climate change-related issues with the coverage frequency of sports, a chronic topic that is present in everyday life, and with the coverage of an acutely emerging topic, the war between Russia on Ukraine (Fig. 1). Both topics are identified using the General Tagesschau Topic Model (GTTM) (see Methods for more details). We find that sports coverage is consistently more frequent than climate change coverage, having a coverage share of about 9% between 2015 and 2023. Climate change coverage only exceeded sports coverage occasionally, for instance, in individual months of the year 2019 during large Fridays for Future protests, once in 2021 during the German federal election, and once in 2023 during the debate about the new German energy law for buildings. Ukraine–Russia-related news accounts for about 4.67% of the total news coverage over the whole period. Although this level is comparable to the level of climate change coverage, it considerably exceeds it in the months following the outbreak of the war in 2022, when it accounts for more than 45% of the news stories in the first month. Before 2022, Ukraine-Russia coverage accounted for around 2.1% of coverage monthly and, therefore, is less frequent than climate change coverage. More examples of climate change coverage in the context of acute and chronic news topics are provided in Supplementary Figs. B.10 and B.11, showing comparable issue dynamics. The coverage share of all GTTM topics over the whole observation period is presented in Supplementary Fig. B.12.

Monthly coverage of climate change (orange) in the context of the monthly coverage of one chronic news topic, here Sports (green), and one acutely emerging news topic, here Ukraine and Russia (blue), presented for the years 2015-2023. Topic coverage is measured as the monthly share of topic-related news stories out of the total number of news stories in that month (see Methods, Equation 1). Climate change-related news stories are classified using a dictionary-based approach. Sports and Ukraine and Russia coverage are classified using the topic model GTTM.

Full size image

Overall, the coverage of climate change in Tagesschau is in line with the average level of coverage in nine major German newspapers between 2015 and 2022, at about 3.8% monthly, as a supplementary analysis of data from a previous study²⁷ shows (Supplementary Fig. B.9).

All investigated topic clusters are manually validated (Supplementary Table C.4). Moreover, the results show that they are robust to classifying climate change coverage with variants of the used dictionary (Supplementary Figs. B.5, B.6, and Supplementary Table C.3). Further, they can be validated using a more simplistic dictionary-based approach for the identification of coverage related to sports, as well as Ukraine and Russia (Supplementary Fig. B.25).

Acute crises are covered more prominently than climate change

TV requires viewers to watch the news in sequence. Consequently, news at the beginning of a program can influence attention to subsequent news, as well as their reception^49,50. Therefore, not only the frequency of climate change coverage but also the positioning of the news slot within the show is of great importance. In Tagesschau news stories with the highest news value and therefore highest relevance are covered first, starting with the most important one³⁷. Hence, we extend the frequency-based evaluation of the prevalence of climate change coverage in Tagesschau by examining the time positioning of news on climate change within the individual shows. We find that climate change-related news constitutes 7.55% of all top stories in Tagesschau, i.e., the first story of the show. Yet, only 1.87% of all top stories explicitly mention climate change. Of all climate change-related news stories, 17.41% are top stories. On average, climate change-related news is covered as the fourth news story (median) and starts at around minute 6 of Tagesschau (Supplementary Fig. B.15).

To evaluate what topics are ranked as more relevant than climate change, we estimate the positioning of 35 news topics other than climate change within Tagesschau shows (Fig. 2). The 35 news topics reflect the topics identified using the topic model GTTM (see Methods and Supplementary Table C.12 for details). Each topic’s position is derived from the relative position of the associated news stories. Their position is classified as either before or after the first climate change-related news story of the show, excluding any later news stories related to climate change (in case of multiple).

The coverage share of GTTM topics relative to the total news coverage either before (orange bars, n = 2614) or after (purple bars, n = 4489) climate change-related news stories. Only GTTM topics that hold a share of ≥3% of the total coverage in either group (before/after) are displayed here. In case of multiple climate change-related news stories in one Tagesschau show, we use the first one and exclude the later ones. The topics “Natural hazards and disasters” and “Environment, climate and agriculture” do not include coverage of climate change. The topic that most frequently outranks climate change is shown on the far left, with the other topics following in descending order. Statistical tests assessing the significance of the start time differences between GTTM topics and climate change coverage are presented in Supplementary Table C.7. Additional information on the topics is provided in Supplementary Table C.12.

Full size image

We find that news about acute crises constitutes a substantial portion of the reporting preceding climate change coverage, indicating that acute crises are ranked as more relevant by the editors. These news stories encompass news about the COVID-19 pandemic, news about Ukraine and Russia, for instance, the war in Ukraine since 2022, news about acts of terrorism, or news about migration, for instance, in association with the so-called “migration crisis” in 2015/2016. In fact, we find that topics relating to these crises are frequently discussed as the first news story in Tagesschau (Supplementary Fig. B.14). Overall, those crises are covered significantly earlier in Tagesschau compared to climate change (Supplementary Table C.7). Topics covered after climate change, i.e., ranked less important, differ strongly from the topics prioritized over climate change (Fig. 2 and Supplementary Fig. B.13). Those topics include “Sports,” “Culture, society, and science,” and “Natural hazards and humanitarian crises.” Sports and culture are covered quite consistently and usually at the end of Tagesschau news, while natural hazards, such as floods, may be thematically related to climate change and therefore covered consecutively, even if they do not explicitly mention climate change.

As with the preceding results, the topic clusters were validated through manual reviews of random news story samples (Supplementary Table C.4). Moreover, the results are qualitatively robust when replacing the topic models with a more simplistic dictionary-based multi-class classification approach (Supplementary Fig. B.26). Additionally, we see qualitatively similar results when looking at a general ranking of the Tagesschau topics by their median and mean start time (Supplementary Fig. B.15).

Climate politics are covered most prevalently

Climate change coverage comprises various topics. To assess their prevalence separately, all climate change-related news stories are classified into five topic clusters, using the Tagesschau Climate Topic Model (TCTM) (see Methods for more detail). The topic clusters comprise German climate politics, international climate politics, climate impacts, climate activism, and religion. The religion cluster contains 19 news stories about climate change addressed in a religious context, for example, comments and warnings about global warming from the pope, such as the Laudatio Si encyclical letter in 2015.

Tagesschau news stories related to climate change most frequently deal with (national) climate politics (Fig. 3A). Collectively, both national and international climate politics (e.g., United Nations Climate Change Conferences or climate protection laws) constitute over 79% of the overall climate change coverage in Tagesschau. Less frequently covered are climate impacts, which constitute about 10.3% of climate change coverage. News about climate impact includes, for instance, news on coral bleaching or wildfires. Climate activism is covered in about 8.47% of the climate change-related news (e.g., Fridays for Future protests), and climate change news related to religion encompasses about 1.5%.

A Thematic composition of climate change coverage in Tagesschau. The coverage frequency of climate change-related topics is depicted as a pie chart showing the share of topic-specific coverage compared to total climate change coverage (n = 1252). Topics are classified by the topic model TCTM (see Methods). B Time positioning of climate change-related topics in Tagesschau depicted as a bar chart. For each topic, the height of the bar indicates the share of the topic-related news stories that begin during (left) or after (right) the first 5 min of the program. Statistical tests assessing the significance of the start time differences are presented in Supplementary Table C.8.

Full size image

Investigating the time positioning of the topics, we find climate politics to be positioned significantly earlier in the show than climate impacts, climate activism, and religion (Fig. 3B, Supplementary Table C.8). Approximately half of climate change coverage associated with national or international climate politics starts within the first five minutes of Tagesschau, 19% are top stories of the program (Supplementary Fig. B.18). In contrast, less than one third of climate activism, climate impacts, or religion start within the first 5 min, and the majority is covered toward the middle and end of the newscast. Less than 11% of news stories associated with climate activism, impacts, and religion are the program’s first story (Supplementary Fig. B.18).

As with the preceding results, the results are robust to the replacement of the topic model with a pure dictionary-based approach (Supplementary Fig. B.27), as well as replacing TCTM with GTTM (Supplementary Fig. B.16). The TCTM topic assignments are validated by manually reviewing random samples of news stories (Supplementary Table C.4). The positioning of topics holds for the median and mean start time of the TCTM topics (Supplementary Fig. B.17).

Floods and storms are the most covered climate impacts

News about climate change covers a wide range of subtopics, one of which is climate impacts. We find that climate impacts constitute a small part of the news, usually reported toward the end of the program. This contrasts with the humanitarian and financial damage they can cause. Furthermore, previous literature finds that coverage of climate impacts, especially of extreme weather events, has a strong impact on the perception of climate change⁵¹. Hence, under-reporting of climate impacts may lead to an underestimation of the extent of climate change.

We examine climate impact coverage in Tagesschau by dividing it into subtopics using the Tagesschau Climate Impact Topic Model (TCITM) (see Methods for details). The subtopics encompass floods and storms, heat, polar warming, climate impact reports, food security, impacts on water bodies and the marine ecosystem, and wildfires. Both coverage of national and international impacts is included.

About a quarter of all climate impact coverage is addressing floods and storms worldwide (Fig. 4A). Polar warming (17%) and heat (16%) are covered in about equal shares. Furthermore, a large share of climate impact coverage discusses climate change impact reports (16%), such as reports by the United Nations or the World Meteorological Organization. Other climate impact coverage addresses food security (11%) and impacts on water bodies and the marine ecosystem (9%). Wildfires receive the least coverage in Tagesschau, about 7%.

A Thematic composition of climate impact coverage. The coverage frequency of climate impact topics is depicted as a pie chart indicating the share of topic-specific coverage compared to total climate impact coverage (n = 129). Topics are classified using the topic model TCITM (see Methods). B Time positioning of the climate impact topics in Tagesschau based on their start minutes presented as heat maps. For ten equally long segments of the program (0: beginning; 1: end), the respective share of coverage that starts within the segment is indicated as color saturation. Darker color indicates higher shares. The relative placement of topics is determined by the relative start minute of their associated news stories (see Methods, Equation 2) rounded down to the first digit after the decimal point. News stories starting in the last minute of Tagesschau (edge case) are merged into the last time segment (0.9). In other words, the color saturation in the heat maps illustrates in which time position climate impact topics usually start throughout Tagesschau shows.

Full size image

Evaluating the individual time positioning of these topics within the shows (Fig. 4B), we find most climate impacts to be covered towards the end of the show. In particular, news about heat and polar warming rarely makes it into the first half of the show. In contrast, certain climate impact topics, such as floods or climate impact reports, are also shown at the beginning of Tagesschau (for details see Supplementary Fig. B.19, Supplementary Tables C.10 and C.9). Notably, individual events, such as the 2021 Ahrtal valley flood receive prominent coverage in both frequency and time positioning, similar to other acute crises. However, most extreme events are covered less prevalently than non-climate-related acute crises. One explanation may be that two-thirds of these events did not occur in Germany directly, reducing their perceived immediacy. It is also important to note that the dataset on climate impacts is quite small, as many reports on extreme events, such as floods, do not draw a connection to climate change explicitly²⁷. In this case, they are represented in the topic cluster “Natural hazards and disasters” of GTTM (see Fig. 2) and are not accounted for here.

The classifications of the TCITM topic model are validated through full manual reviews of the topic assignments (Supplementary Table C.4). The results are robust to the manual review of topic assignments (Supplementary Fig. B.28).

The analysis presented in this study provides insights into the prevalence of climate change coverage from 2015 to 2023 in Tagesschau, one of Germany’s leading TV news shows. Classifying the news stories’ topics by utilizing state-of-the-art NLP methods, our analysis suggests that the frequency of climate change coverage is increasing but still falls behind individual chronic news topics, such as sports, and acute news topics, such as the Ukrainian-Russian war. Moreover, we find climate change to be positioned on average in an intermediate news slot in the program, surpassed by acute crises such as the COVID-19 pandemic. Within climate change coverage, climate policy is covered most prevalently, while climate impacts and climate activism receive less attention.

Our results are plausibly explained by the competition for the limited airtime between climate change and other news topics, with editors deciding what news to show at which position in the program. This selection and ranking procedure determines the prevalence of news topics and is strongly dependent on the editors’ assessment of the stories’ news value. Acute crises, such as the Ukrainian–Russian war and the COVID-19 pandemic, are likely to be considered of high news value because they occur unexpectedly, affect a large audience, and have particularly bad consequences, common criteria of high news value⁸. Similarly, we find sports coverage to be particularly frequent in Tagesschau, consistent with previous studies⁵². Sport events are likely to be considered of high news value, given their societal relevance. This contrasts with climate change issues. Climate change itself is likely to be seen as less newsworthy because it is chronic in nature and its negative impacts build up slowly over several years, potentially leading to news fatigue, as observed on social media⁵³. The damage caused is often indirect, for example, by reducing economic growth^54,55 or affecting mental health^56,57, making it difficult to capture in a news story and typically reducing its news value. More direct damages, such as those caused by extreme weather events, tend to be covered with a focus on the acute harm rather than the causes, i.e., climate change²⁷.

The prioritization of acute crises carries the risk of myopic coverage. Long-term, chronic crises, such as climate change, the effects of which lie at least partly in the future²³ but need to be mitigated today⁵⁸, may not be given sufficient consideration. A potential consideration would be the implementation of a regular, fixed broadcasting slot for chronic issues, such as climate change, before or after Tagesschau or throughout the day. An example of a topic with a 5-min-long, fixed broadcasting slot right before Tagesschau is finance, which is devoid of competitive coverage dynamics. This could also prevent issue inertia as the consumption of climate change news is optional and targeted towards interested audiences⁵⁹.

If climate change is covered in Tagesschau, this is most often in the context of the political discourse regarding adaptation and mitigation measures at the national and international level⁶⁰. This finding aligns with the news agenda of Tagesschau³⁷, as well as with results from a study investigating climate change coverage in Tagesschau for the year 2022³⁴. Given Germans’ strong interest in both political and climate news³¹, focusing climate change coverage on politics integrates both informational priorities.

Compared to climate policy, climate impacts are less prominently covered. While a difference in the number of news stories might be plausible, as climate politics is thematically broader than climate impacts, this does not explain the considerably later positioning of climate impacts in the show. One possible mechanism could be that they are grouped thematically with the weather forecast at the end of the program³⁷. However, especially for extreme weather events, the time positioning does not align with the ranking of other crises. This could be due to a lack of spatial proximity to many of these climate impacts, which may limit their relevance for the audience and therefore their news value. However, as climate impact coverage was found to be linked with public concern about climate change⁵¹, this may reinforce the perception of climate change as spatio-temporally distant.

There are several limitations to this study that should be acknowledged. First, the data are focused on a single TV newscast in Germany, which limits the generalizability of the results. However, the findings are in broad agreement with previous studies that use traditional approaches to compare multiple German newscasts^60,61 and newspapers²⁷, which strengthens our confidence that our findings are not exclusive to Tagesschau. An assessment of whether our results are generalizable in the sense that they also hold for TV media in other countries presents a promising avenue for future research. Previous research suggests that there might be cross-national differences in framing⁶², especially compared to the global south²⁶. The algorithmic idea for news story identification presented here may be transferable onto other news programs with some adaptations if the necessary NLP tools are available in the target language and data are available in the required format, i.e., for each news edition a textual transcript of the show (e.g., subtitles) and an ordered list of the news story headlines covered.

Second, we classify the topics of the news stories using a combination of a dictionary-based approach and a neural topic model, both established^{25,27,48,51,63,64,65} but not perfect approaches. The dictionary-based approach may miss some climate change-related news because of lacking keywords. It may also classify tangentially related news stories as climate change-related, potentially leading to over- or underestimation of coverage. We show that the results are robust to different climate change dictionaries (Supplementary Fig. B.5). Moreover, the manual review of topic assignments on randomly sampled data yields a high accuracy and F1-Score for the climate change dictionary (Supplementary Table C.4). Through the application of transformer models, neural topic modeling is, at least partially, a black box algorithm. Hence, its performance is hard to validate and tune. Due to the limited prior work on appropriate performance metrics, we relied heavily on manual topic labeling, as well as manual evaluation of topic interpretability and similarity. These constitute established approaches⁶⁶, the latter is used, for example, by Barbera et al.⁶. Based on topic similarity, we aggregated the identified topics into broader topic clusters (Supplementary Tables C.12–C.17 for the merged and the original topics, respectively). However, these subjective judgments may introduce some bias to the results. In addition, the topic models use hard topic assignments, i.e., each news story is assigned only to its primary topic, which may underestimate topics that are only tangentially addressed in the news stories. While the complementary approach, soft topic assignments, may be closer to reality, as it allows for overlapping topics in individual news stories, it also creates many ambiguous topic assignments, in our case, likely due to the high density of data in the embedding space. To ensure that these factors do not bias our results, two reviewers manually validated the topic classifications of a random sample of news stories (Supplementary Table C.4), with high agreement (≥0.7 for all classification approaches). Furthermore, we provide some general metrics for model performance (Supplementary Tables C.5 and C.6). Beyond neural topic modeling, sentiment analysis of climate change coverage on TV would be an interesting path to study the framing of climate change coverage further, potentially incorporating visual and auditory information.

Third, to analyze the positioning of climate change coverage relative to other news topics (see Fig. 2), only days with climate change coverage in Tagesschau are considered. Our data and methodology do not allow proper identification of days with unreported climate change-related news. Therefore, when there is no climate change coverage on a given day, due to the prioritization of other news stories, the content from the whole day cannot be included. This highlights a gap in this analysis and could be subject to future research.

Finally, we measure the prevalence of climate change coverage using two metrics: frequency and time positioning. Another potentially interesting metric would be news story duration. Due to algorithmic limitations, this is not studied here, but it states an interesting extension of our work on the news story identification.

Data and code used in this study are available in accordance with the data and code availability statements on Zenodo⁶⁷.

Tagesschau data

For this study, news data from the German TV newscast Tagesschau are used. Tagesschau is a daily, 15 min TV newscast on German public TV that is considered objective^68,69,70. It is the biggest newscast on German TV³⁵. The 8 p.m. Tagesschau edition is the most widely watched, attracting around nine million viewers daily (in 2023)³⁵.

The Tagesschau program data are compiled from the publicly accessible Tagesschau online archive. The final dataset comprises 2888 Tagesschau shows from 2015 to 2023 that were broadcast at 8 p.m. For each Tagesschau show in the dataset, the subtitles and a program description are collected. The program description includes the publication date and an ordered list of news story headlines covered in the show. The subtitle file of one show contains all subtitles along with the corresponding start and end timestamps of their display in the video. Out of all available Tagesschau shows (3267 total), 379 lack subtitle data (about 12%) and are therefore not included in the dataset. Upon examination of the excluded shows (Supplementary Table C.1), we find that they are not correlated in time, suggesting that the introduced errors are random and unlikely to bias our results. The final data are evenly distributed over the time frame, about 300 shows per year (Supplementary Fig. B.1).

News story identification

In order to analyze the content of the individual news stories in each Tagesschau show, the news stories are inferred from the subtitles. This is achieved by comparing the subtitle sentences with the headlines of the news stories, known from the program description, using a keyword-matching approach. The first sentence in the subtitles that matches important keywords from a news story headline usually corresponds to the first sentence of that news story. The end of a news story is implicitly derived by identifying the beginning of the following news story. In other words, the last sentence of a news story is the one preceding the first sentence of the following news story, which is identified using the same procedure. This method allows us to identify text passages for 28,461 news stories in total. Furthermore, the order of the news stories and their respective start and end times are determined using time stamps in the subtitle data.

Matching subtitles and headlines

To effectively match the news stories’ headlines to the subtitles, we developed a custom algorithm. It combines multiple methods from NLP to allow advanced processing of textual data. The algorithm is outlined in pseudo code in Boxes 1 and 2. It essentially iterates through the subtitle sentences, trying to match a set of keywords from the news story headlines to similar words in the subtitles. We call this fuzzy matching. We assume that the headlines are provided in the correct order and match them consecutively. A threshold is set to define when a sentence from the subtitles is considered to be the start of a new news story, based on the number of keyword matches. Its value is chosen based on the number of available keywords from the headline. A set of keywords, containing each keyword only once, is extracted for each headline to avoid unintentionally weighing certain keywords higher. Keywords from this set that have an entity type are then intentionally weighted higher, since they tend to carry important information. Before starting the matching procedure, we try to limit the subtitles to news story content only by removing the Tagesschau intro and outro. This step helps to prevent noise in the news story classification and is based on the recognition of common phrases mentioned at the beginning and end of Tagesschau.

Validation

The algorithmic performance was assessed using the first sentence accuracy ({{{mathcal{A}}}}) on manually curated test data. The test data consist of 200 randomly sampled Tagesschau shows from the dataset with a total of 2046 manually classified news stories. The first sentence accuracy is given as

A classification is counted as correct if the matched first sentence corresponds to the first sentence of the test data with a margin of error of  ±150 characters. This metric captures the accuracy of the subtitle partitioning (further information in Supplementary Note A.2). We find 77.7% of all first sentences to be accurately matched by the algorithm. Overall, several approaches for news story identification are investigated, including two different algorithmic approaches in combination with three different keyword matching strategies (for more details see Supplementary Note A.1). The best performing approach (Supplementary Fig. B.20) is presented, discussed and used here.

Applying the algorithm to the full dataset, we are able to identify a matching text passage for about 95% of listed news story headlines. The remaining 5% cannot be identified due to errors when inferring the necessary information from the subtitle data. News stories prone to failed identification are those with headlines that do not provide meaningful keywords, that are listed in the wrong order or that are not listed in the program’s description at all. Investigating the headlines of unidentified news stories in the data, we find that they were uncorrelated both thematically and temporally (Supplementary Fig. B.2). This suggests that the errors introduced by this procedure are random and unlikely to bias our results. The algorithmic output is further validated with automatic sanity checks (Supplementary Notes A.3 and Supplementary Fig. B.21). News stories in the subtitles that do not correspond to a headline cannot be identified or analyzed further.

Topic classification

The topics of individual news stories are classified using a combination of a dictionary-based approach and neural topic modeling. The topic of all news stories is classified using the state-of-the-art transformer-based topic modeling framework BERTopic³⁸. However, climate change coverage spans over multiple topics that were discovered in the data by the topic model (Supplementary Fig. B.16), such as energy politics, agriculture, and transport. Therefore, climate change coverage is classified using a climate change dictionary. Hence, all news stories that include at least one climate change-related keyword from the climate change dictionary are reassigned to the topic “Climate change.” The combination of dictionary-based and topic modeling approaches, therefore, allows a flexible classification of the main topic cluster, namely climate change, while the unsupervised learning process of the topic model ensures that no prior knowledge is required for the other topic clusters.

Dictionary-based classification

To identify news stories related to climate change, a dictionary containing keywords related to climate change is utilized. News stories are classified and referred to as climate change-related if they include at least one keyword from this dictionary. The keyword selection of this dictionary is informed by previous literature^27,51, and presented in the Supplementary Material (Supplementary Table C.2). The utilization of dictionaries to classify climate change coverage is a well-established approach within the domain (e.g., see refs. ^{25,26,27,51,63,71}). The approach yields 1252 news stories related to climate change in the dataset, with 361 explicitly containing “Klimawandel” (German for “climate change”). To ascertain the importance of individual keywords in the dictionary for the classification, the frequency with which each keyword appears in the data is evaluated in Supplementary Fig. B.3.

Neural topic modeling

The topics discussed in the non-climate-change-related coverage, i.e., the coverage not matched using the dictionary-based approach, are assessed using neural topic modeling. Specifically, the neural topic modeling framework BERTopic³⁸ is utilized, as it was consistently delivering the most interpretable results, similar to findings from refs. ^47,64,65.

BERTopic is built in a modular way, allowing customization of the topic model (e.g., through the selection of hyperparameters or algorithms). The steps applied to train the topic models that are used in this work are based on the implementation of BERTopic and displayed in Fig. 5. First, embeddings for the news stories are generated by transforming them into multidimensional vectors in the embedding space of a pre-trained language model. These embeddings facilitate the comparison of news stories, as similar news stories tend to be co-located. For this task, BERTopic leverages Sentence-BERT⁷². In contrast to bag-of-word approaches, Sentence-BERT is designed to maintain the relations between words and sentences, enhancing the contextual understanding of the document^72,73. Its performance depends on the length of the input texts, as explained in the documentation⁷⁴. It generates more discriminative embeddings for short text passages because the topics are more distinct. Longer texts increase the topical noise, resulting in a higher density of input documents in the embedding space, making it harder to differentiate between topics effectively. In our data, the average length of input documents is 1183 characters (including spaces). We find the embeddings to be clearly distinguishable, but not highly fine-grained, as shown in Supplementary Tables C.5 and C.6, and Supplementary Figs. B.22–B.24. In the second step of BERTopic, the dimensionality of the embeddings is reduced using the Uniform Manifold Approximation and Projection algorithm. This step is crucial, as distances in lower-dimensional spaces tend to offer more meaningful and interpretable insights⁷⁵. Next, a standard clustering algorithm is applied on these lower-dimensional embeddings, each cluster representing one topic in the news data. We chose to substitute the BERTopic standard clustering algorithm Hierarchical Density-Based Spatial Clustering (HDBScan) with the k-Means algorithm. k-Means forces a topic assignment for all news stories, whereas HDBScan generates an excessive amount of outliers. In step three of BERTopic, information about the topics is inferred from the news stories assigned to the clusters. First, the news stories are vectorized, including stop word removal. Next, the Term Frequency-Inverse Document Frequency (TF-IDF) Matrix is computed, enabling the identification of important keywords within each topic. These keywords constitute the topic representation. Finally, the Maximal Marginal Relevance is computed to fine-tune the topic representations, increasing the diversity and reducing the redundancies of the keywords in the topic representation⁷⁶. This yields a distinctive and comprehensive topic representation that allows for a sound understanding of the topics and themes in the data. In an additional, fourth step, overlapping, highly related and very similar topics are merged together into more comprehensive and interpretable topics (Supplementary Tables C.12–C.17). These merges are informed by manual reviews of the output topics, as well as a topic hierarchy based on Ward⁷⁷ (more details in Supplementary Note A.5). In the final step, the topics are labeled manually based on a review of the topic representations and assigned news stories by two researches (four-eye-principle). Exact specifications used during model training can be found in Supplementary Table C.11. More information on the selection of algorithms and their hyperparameters is provided in Supplementary Note A.4.

This figure provides an overview of the steps and algorithms applied to train the topic models used in this study.

Full size image

This neural topic modeling method is used to train three topic models for our analysis: the GTTM with 35 topics (trained on all 28,461 Tagesschau news stories), the TCTM with five topics (trained on all 1252 news stories classified as climate change-related based on the climate change dictionary), and the TCITM with seven topics (trained on all 129 news stories classified as climate impacts by the TCTM). More details on the topics are provided in Supplementary Table C.12, C.14, C.16. The daily weather forecast (n = 2752) was excluded from model training and the subsequent data analysis. It was identified by its headline “Das Wetter” (en: “The weather”).

Validation

Currently, no uniform standard for automatic topic model evaluation exists for neural topic models^46,78. Therefore, the topic interpretability is assessed manually during model selection using the topic representations and the three most representative news stories assigned to the topic. The most representative news stories of a topic are those with the highest cosine similarity to the topic representation using their respective c-TF-IDF representations. Nonetheless, there can still be outliers in the data, creating noise in the topic model classifications. To address that, the classifications of both, the dictionary-based approach and neural topic modeling, are validated through a manual review of a random sample of news stories (Supplementary Table C.4). To prevent bias through personal judgment by a single reviewer, we also compute the agreement of two reviewers on classification accuracy using Cohen’s Kappa. Additionally, the distinguishability and similarity of topics in the embedding space is assessed using the cosine distance (Supplementary Figs. B.22–B.24). Further traditional standard topic modeling evaluation metrics are provided and discussed in the SI (Supplementary Note A.6, Supplementary Tables C.5 and C.6). We find high topic diversity across the models, high silhouette scores for the majority of the topics, as well as high accuracy and F1-scores for the topic classifications. As an additional robustness check of our models and results, we use topic-specific dictionaries to classify the topics discovered by the topic models. These robustness checks confirm our findings obtained through neural topic modeling.

Topic prevalence measures

As outlined in the introduction, the prevalence of climate change coverage strongly impacts the public perception of climate change. The extent to which climate change is covered in Tagesschau is gauged by two metrics: the frequency of climate change reporting and the positioning of climate change coverage within Tagesschau.

Topic frequency

Topic frequency TF(x, t) is computed as the number of news stories N(x, t) related to a given topic x within a month t normalized with the total number of news stories N(t) in the same time period t. Formally, TF is defined as:

Topic positioning

News stories in Tagesschau are sorted by their news value³⁷ such that the most relevant news stories are broadcasted first. Due to the linear format of TV programs, news stories at the beginning of a program may impact interest and viewer engagement with later content⁷⁹. Hence, the positioning of a topic within Tagesschau is important to consider when estimating topic prevalence. Given that the length of each Tagesschau edition varies, the relative start minute (RSM) of news stories is computed to quantify their positioning within Tagesschau. The relative start minute RSM(s, i) is the start minute of a news story s in a Tagesschau show T normalized by the total length of T in minutes. Formally, RSM is given as:

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.