The energy transition as a global challenge is widely discussed in Germany and shaped by diferent narratives. We investigate how newspapers, comments and social media posts difer between sources, platforms, and entities to visualize and monitor respective indicators of the debate over time. For this purpose, we propose a framework of an unsupervised and automated topic model-based approach, the results of which are visualized in a dashboard in conjunction with results from a reliable and supervised classification model using parameter-eficient fine-tuning methods and language models. The analysis of the classification results is refined by integrating the information on annotator disagreement which is captured in a new human-coder labeled sample of our text corpus on the German energy transition. We place a particular emphasis on highlighting the issue of social (in)justice in the debate on energy transition. narrative, energy transition, topic model, parameter-eficient fine-tuning, annotation The energy transition is a critical global issue, addressing environmental, economic, and social challenges that afect and involve everyone. As a transformative process aimed at shifting from fossil fuels to renewable energy, it has become a central topic in (social) media and public debates. The respective coverage not only reflects public opinion but also plays a significant role in shaping it, influencing how society perceives and engages with the transition. Understanding and tracking these discussions is therefore essential to support informed, inclusive, and efective energy policy development.
CEUR
ceur-ws.org
more equitable and inclusive policy development in the energy transition. As a complementary and
expanding approach to Loschke et al. [
For the automated representation of the debate we will make use of continuous, i.e. updateable topic models, while for the snapshot analysis we will utilize parameter-eficient fine-tuning (PEFT) in combination with few-shot methods on (large) language models (LLMs). Our text corpus will primarily comprises national and regional newspapers, to which we selectively add online comments and social media content.
In the following, we present our methodological workflow, which is based on the findings of Schofield et al. (2015), that “suggest not automating these workflows but adding transparency and easy-to-use tools to support practitioners’ context-specific judgments and interventions” [ 3]. This paper aims to provide a concrete description of a best practice case study for the continuous and reliable extraction and presentation of narrative elements from real world application scenarios — and to open it up for further discussion and improvement. To provide a first glimpse of the possibilities of our project, we present preliminary results of the our continuous topic models with integrated change detection as an example analysis. Further elements of the (later) presented framework are currently still under development. The database and the parameters used for the presented results are preliminary as well. However, their heuristic choice illustrates the potential of our methods for an even more sophisticated parameter choice. We provide the programming code, analyses and (permitted) data of the project in the GitHub repository https://github.com/LarsG321/ Dissemination-Monitoring-narratives-about-the-energy-transition-in-Germany. We welcome any type of questions, suggestions or criticisms via the issue tracker of the respective repository.
Latent Dirichlet allocation (LDA) [4], estimated with a Gibbs sampler [5], is supposedly (still) the most frequently used topic model variant. In particular in the social sciences [6] it still enjoys great popularity years after the publication of the structural topic model (STM) [ 7], probably due to the simplicity of application, the modest model assumptions and the meaningful performance, which cannot be outperformed consistently even by neural topic models [8, 9].
After the publication of the seminal paper of the attention algorithm [ 10] and the resulting performance boost for the corresponding transformer-based language models, BERTopic [11] as a neural topic model (NTM) based on language models (LMs) has also become quite popular. That is likely because it is particularly convenient to use, in fact BERTopic is even more convenient than LDA as it estimates the number of topics data-driven by default. However, it has already been shown in a few analyses that BERTopic performs inferior to Gibbs LDA in terms of several (in particular those based on human judgments) performance measures [8, 9]. In combination with the ongoing scrutiny of existing automated evaluation methods for topic models [12, 13, 14], even if the contextualized topic model [15] works better for well-chosen parameters, overall LDA serves as a robust, stable, reliable universal tool — and many NTMs perform worse in comparison. In our project, since we want to update the resulting topic models continuously, we evaluate which of the two methods RollingLDA [16], including a dedicated monitoring tool to analyze the topic evolutions [17], or BERTrend [18], that is based on BERTopic, is better suited for our use case and select the more suitable model for application.
The research fields of parameter-eficient fine-tuning (PEFT) and mixture of experts (MoE) are very volatile and at the same time strongly overlapping in both aims and architectures, so that a number of publications also explicitly consider combinations. While PEFT in its origin (and name) aims at eficiency and implements this by keeping a large part of the pre-trained parameters fixed, MoE (according to its name) primarily targets performance improvement through input-dependent (de-)activation or weighting of parameters. However, implicitly — or explicitly — the respective other objective is usually achieved as well. In recent years, many innovations, mostly with incremental improvements in certain domains, have been published. The supposedly most well-known method in this area is LoRA [19], which has subsequently undergone various further developments, such as PRILoRA [20], VeRA [21], DoRA [22], RoSA [23] and the quantized — or quantization-aware — versions QLoRA [24] and QA-LoRA [25]. Moreover, several studies were able to show that the combination of PEFT methods in MoE architectures proves efective. These include mixture of vectors (MoV) and mixture of LoRA (MoLoRA) [26] or parameter-eficient routed fine-tuning (PERFT) [ 27]. Furthermore, it was shown that in MoE PEFT architectures it is advantageous to activate only the PEFT modules that are most relevant for the task, while freezing the other experts for fine-tuning [ 28].
For our use case, we rely on the well-established PEFT variant LoRA, which has been proven to
perform well across diverse tasks. In fact, we compare LoRA to its quantized version QLoRA and
combine both with the language modeling head PETapter [29]. Thus, we unite the advantages of PEFT,
i.e. eficiency, robustness, modularity, ease of use and sharing, with the additional performance gain
by using a few-shot method that integrates the task description into the input utilizing the masked
language modeling objective in a PET-alike manner [30]. For the pretrained base (L)LM, we are focusing
on German or multilingual models since the corpus is in German language. Options include the models
German BERT [31], LeoLM [
For the design of our framework, we follow a human-in-the-loop best practice [
In Figure 1, our project’s methodological workflow is illustrated. As a database, we focus on national
and regional German newspapers. We provide a list of example newspapers that we have access to and
are considering to include in Appendix A. The resulting corpus is complemented with commentary
content from the websites of, e.g., Bild, Spiegel, Süddeutsche Zeitung, Welt, taz, and Zeit, for which we
need to evaluate the full public availability of the respective community comments. Further, we collect
social media posts utilizing promising software packages for the platforms X [
Next, we define a filter that primarily aims for a recall, i.e. relevance with regard to the energy
transition, close to 100% while maintaining the highest possible precision. In Section 4, we provide
an example of a filter for the newspapers Bild, Süddeutsche Zeitung, and Welt. The filtered and
preprocessed corpora are modeled using RollingLDA or BERTrend, respectively. Every week, the
procedure is repeated with the newly added data. After postprocessing, the results of the modeling
are made publicly available in our dashboard on an ongoing weekly basis. For this, we make use of
automated topic labeling [
The interpretation of the results is used for the development of the codebook [cf. 46] for supervised classification, which we use for a restrospective snapshot analysis. In the definition of the codebook, we will (need to) focus on the operationalizable definition of what social justice means in order to avoid too much disagreement in the labels due to insuficiently precise definitions [cf. 47, 48, 49]). We aim for the resulting labeled data set to provide evidence for (at least) the following hypotheses:
s g n i d n i F
(Gold Standard)
• The (media-perceptible) acceptance for the German energy transition dropped in the last two
years (2023, 2024).
• The negative aspects regarding the energy transition are disseminated disproportionately often
by certain media.
• The argument of social (in)justice is more often used as a counter-argument for the German
energy transition as a whole than as an argument for a more socially just implementation of it.
The annotation is done by five independent coders in order to be able to provide not only a (reliable) gold
standard label but also topic-based uncertainty estimates through (reasonable) annotator disagreement
for our evaluations. In addition to the evaluation of the annotator disagreement [
Supposing a suficiently high annotator agreement, we employ the fine-tuned model based on the codebook-labeled partial dataset to predict and analyze a variety of descriptive characteristics, e.g., polarization and subjectivity scores for topics, arguments, newspapers and media types, respectively, over time. These findings are in turn fed back into our dashboard to provide an comprehensive overview of the German news media debate on the energy transition and its social (in)justice.
This first analysis is concerned with the left half of our framework in Figure 1 up to the development of the dashboard and presents first results for the RollingLDA methodology.
In order to investigate the German energy transition and its accompanying narratives, we retrieve articles from three major German newspapers Bild, Süddeutsche Zeitung (SZ), and Welt from 2014 to 2024. These outlets are selected to capture a broad ideological and stylistic range: SZ is generally considered left-leaning, Welt is viewed as conservative, and Bild is known for its more opinion-based and sensationalist approach. We apply a pre-filtering step to retrieve only relevant articles, i.e. containing terms explicitly related to the German energy transition. For this, we make use of the keywords • Bild: (wende | gesetz | politik | habeck | hammer) & (waerme | energie | klima | heiz) • Süddeutsche Zeitung and Welt: (wende | gesetz | politik ) & (waerme | energie | klima | heizung) In addition, for Bild, we exclude articles belonging to the Sports section according to meta data as well as all articles containing the pattern leserbriefe or seite in the title since these articles revealed to be noisy and seem to be falsely extracted by the data provider. These steps result in a dataset of 21,046 articles from SZ, 21,107 from Welt, and 2,263 from Bild.
To explore the thematic structure of the filtered articles, we applied RollingLDA [ 16] for various topic
numbers in combination with a change detection procedure [17] to identify changes in topics and the
most influential words for these respective changes. The topic models were trained independently
on each newspaper corpus. In Appendix B, we provide details about the preprocessing and heuristic
hyperparameter choices, i.e. a more sophisticated choice of parameters could lead to even clearer (and
better interpretable) results.. As explained for our framework in Figure 1, we employed automated
topic labeling leveraging LLMs [
The basic topic model results, i.e. top ten explanatory words per topic and the monthly topic frequencies, can bee accessed in Appendix C. Within the SZ corpus, the topics covered both the German energy transition itself and international political afairs, with frequent mentions of Ukraine and Russia. Additional thematic clusters included climate protection and protest (Topic 4) and climate change in relation to agriculture (Topic 2), indicating that the filtering process efectively isolated content directly relevant to the German energy transition. Notably, topics related to sustainability and renewable energies were found consistently across multiple topic number configurations. A similarly broad thematic distribution emerged within the Welt corpus. The identified topics ranged from Russia-EU relations in the context of the Ukraine conflict (Topic 8) and German energy policy (e.g., topics 1, 4, 6, 7) to real estate and infrastructure (e.g., topics 5, 10). As in the SZ corpus, themes tied to sustainability, transportation, and the German energy transition were prevalent and showed consistent patterns across diferent topic number configurations. Furthermore, comparative analyses between the two newspapers revealed that certain topic clusters — particularly those pertaining to the Ukraine conflict, sustainable energy, and transportation — were present in both corpora, with occasional co-moving patterns in topic prevalence over time. In contrast, the Bild corpus exhibited a more pronounced focus on national politics and prominent public figures. The most frequently identified topics included German domestic politics in the context of the Ukraine conflict (Topic 1), Thuringian state-level politics (Topic 2) and German politicians and public figures (nearly all topics). This orientation aligns with Bild’s reputation for spotlighting highprofile personalities and emphasizing sensationalist content. Despite these diferences, Bild’s articles also contained relevant discussions on the German energy transition, demonstrating the efectiveness of the pre-filtering approach in capturing pertinent material, even within a more personality-driven news environment.
Overall, the RollingLDA analysis successfully extracted thematically coherent topics that illuminate the debate on the German energy transition. Süddeutsche Zeitung and Welt both covered a broad range of energy- and sustainability-related issues alongside international political afairs, while Bild’s coverage skewed more heavily toward national politics and individual political figures. These findings ofer an important foundation for understanding how diferent newspaper orientations shape narratives surrounding the German energy transition, laying the groundwork for deeper investigations into public debate and policy implications in this domain.
Moving on to the change detection, several meaningful changes were identified. Figure 2 displays some results for the change detection. Blue curves indicate intra-topic similarities over time, while the red curves represent statistically motivated bootstrapped thresholds [17]. That is, vertical lines mark detected changes in time (top row). For illustration, the plots start in 2018 displaying five distinct changes for topic 5 “German Energy Transition” of SZ. We select the first of those five changes for a more comprehensive analysis, focusing on the most influential words associated with this change (bottom row). Accordingly, it was triggered by an increased usage of terms related to fossil fuels (e.g., gas, oel) and the word russland. This aligns with the start of the Russian invasion of Ukraine in February 2022, which significantly impacted the fossil fuel crisis in Germany. For Welt, the third topic emerged as the most relevant to the energy transition. Here, the most pertinent change occurred in July 2021, attributed to the catastrophic flooding (180 deaths) in the German Ahrtal, which received extensive media attention. The increased frequency of terms such as klimawandel (climate change) and katastrophenschutz (disaster prevention) supports this interpretation. For Bild, it is noteworthy that the intra-topic cosine similarity are rather low, indicating that topic 5 for Bild, due to the smaller number of articles, demonstrates less stability over time compared to the other two corpora. This volatility leads to a reduced cosine similarity, suggesting a less persistent thematic presence within this corpus. Nonetheless, the method identified the beginning of the Russia-Ukraine conflict as a relevant change, leveraged by an increased usage of the terms Putin, Ukraine, and Russischen.
To summarize the preliminary findings, it becomes evident that the Süddeutsche Zeitung and Welt appear to contain several relevant topics related to the German energy transition, whereas the analysis of Bild reveals challenges due to a relative lack of energy-related topics. In addition, we observed the cross theme of social injustice to be barely represented in the current corpus, indicating areas for further refinement and investigation for future analyses.
Despite our concise approach, there are still several challenges that we will briefly address here. It is likely to be easily possible to classify thematic aspects, to depict sentiment over time, as well as to present key actors in the debate. However, it is still challenging to render entire narratives (over time) beyond individual elements [cf., e.g., 50, 51]. Narratives are complex to extract (even if, by nature, they rely on content simplification) and have several components, whereby there are already many diferent concepts in the (theoretical) definition and hence also in the (operationalized) extraction.
Regarding data, it should be noted that social media data availability (e.g., X as most prominent
example) is increasingly burdening analyses utilizing such data. Furthermore, Instagram, TikTok, but
also X contain ever-increasing video and image content, posing the challenge to appropriately combine
these diferent data types for modeling (we limit ourselves to text data). This is already evident for the
articles of the Bild newspaper, which, due to their heavy use of visual content, but also due to their use
of neologisms (especially in connection with the 2023 Gebäudeenergiegesetz GEG, “Buildings Energy
Act”, which coined the term Heizungshammer, “heating hammer”), require special treatment when it
comes to preprocessing and keyword searches (cf. Section 4). Further, it is unclear how to adequately
take into account diferent, e.g., age-related, media consumption habits when evaluating perception. In
case we get access to regional newspaper data, moreover, we will investigate the regional diferences in
reporting on the German energy transition, similar to Ozgun and Broekel [
In our project “Diskurs Energiewende”, additionally to the presented framework, we aim to contribute to the development of NLP methods. For RollingLDA, we plan to implement an extension of emerging and fading topics, for which we will evaluate the positive and negative aspects of how BERTrend handles them. This might be especially helpful for the RollingLDA results presented above, since the method ifxates the topics in the first iteration. This leads to the efect that non-present topics in later iterations are mapped to topics of the first iteration. Thus, if the narrative of social injustice gets important in the year 2020, for example, and was not present in the first iteration in year 2018 it does not receive its own topic. Further, we intend to develop advancements in the overlapping area of narrative detection using PEFT and the use of advanced learning methods making special use of annotator disagreement. These would allow us to analyze the dissemination and timeline of narratives through diferent media sources. In turn, it will be challenging to develop a fair performance score to compare models fine-tuned with classical gold standard labels and those incorporating annotator disagreement [cf. 47, 48, 49]).
While the current preliminary analysis is capable of detecting topics with respect to the energy transition, certain limitations with regard to the narrative detection of social injustice apply. To overcome this problem, we will extend the RollingLDA analyses and compare them with other topic model results. We will extend the keyword search to obtain a larger corpus, i.e. a higher recall. In addition, a focused sub-corpus analysis of social injustice related articles within the corpus of the energy transition articles will be conducted. Further, as incremental improvements we will incorporate lemmatization and improve our preprocessing function to exclude non-alphanumeric letters (e.g. “€”), which further normalizes the text corpus and might improve the narrative extraction. Based on the improved corpus, we will evaluate the usefulness of RollingLDA, BERTrend [18] and BERTopic [11] to identify narratives in our data and to implement it in our final publicly available dashboard.
This work was funded by the German Federal Ministry for Economic Afairs and Climate Action (Bundesministerium für Wirtschaft und Klimaschutz, BMWK) as part of the joint project Social (in)justice in the energy transition - from the digital debate to the living world (project no. 03EI5267B). [3] A. Schofield, S. Wu, T. Bayard de Volo, T. Kuze, A. Gomez, S. Sultana, ”My very subjective human interpretation”: Domain expert perspectives on navigating the text analysis loop for topic models, Proceedings of the ACM on Human-Computer Interaction 9 (2025). doi:10.1145/3701201. [4] D. M. Blei, A. Y. Ng, M. I. Jordan, Latent Dirichlet allocation, Journal of Machine Learning Research 3 (2003) 993–1022. doi:10.1162/jmlr.2003.3.4-5.993. [5] T. L. Grifiths, M. Steyvers, Finding scientific topics, Proceedings of the National Academy of
Sciences 101 (2004) 5228–5235. doi:10.1073/pnas.0307752101. [6] D. Ramage, E. Rosen, J. Chuang, C. D. Manning, D. A. McFarland, Topic modeling for the social sciences, in: Workshop on Applications for Topic Models, NIPS, 2009. URL: http://vis.stanford. edu/papers/topic-modeling-social-sciences. [7] M. E. Roberts, B. M. Stewart, D. Tingley, E. M. Airoldi, The structural topic model and applied social science, in: NIPS-Workshop on Topic Models: Computation, Application, and Evaluation, 2013. URL: https://www.wcfia.harvard.edu/files/wcfia/files/stmnips2013.pdf. [8] A. M. Hoyle, R. Sarkar, P. Goel, P. Resnik, Are neural topic models broken?, in: Findings of the Association for Computational Linguistics: EMNLP 2022, Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 5321–5344. doi:10.18653/v1/2022.findings-emnlp. 390. [9] Z. Li, A. Mao, D. Stephens, P. Goel, E. Walpole, A. Dima, J. Fung, J. Boyd-Graber, Improving the TENOR of labeling: Re-evaluating topic models for content analysis, in: Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, St. Julian’s, Malta, 2024, pp. 840–859.
URL: https://aclanthology.org/2024.eacl-long.51/. [10] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need, in: I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Systems, volume 30, Curran Associates, Inc., 2017. URL: https://proceedings.neurips.cc/paper_files/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. [11] M. Grootendorst, BERTopic: Neural topic modeling with a class-based tf-idf procedure, 2022.
arXiv:2203.05794. [12] C. Doogan, W. Buntine, Topic model or topic twaddle? Re-evaluating semantic interpretability measures, in: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Online, 2021, pp. 3824–3848. doi:10.18653/v1/2021.naacl-main.300. [13] A. Hoyle, P. Goel, A. Hian-Cheong, D. Peskov, J. Boyd-Graber, P. Resnik, Is automated topic model evaluation broken? The incoherence of coherence, in: Advances in Neural Information Processing Systems, volume 34, Curran Associates, Inc., 2021, pp. 2018–2033. URL: https://proceedings.neurips. cc/paper_files/paper/2021/file/0f83556a305d789b1d71815e8ea4f4b0-Paper.pdf. [14] D. Stammbach, V. Zouhar, A. Hoyle, M. Sachan, E. Ash, Revisiting automated topic model evaluation with large language models, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Singapore, 2023, pp. 9348–9357. doi:10.18653/v1/2023.emnlp-main.581. [15] F. Bianchi, S. Terragni, D. Hovy, Pre-training is a hot topic: Contextualized document embeddings improve topic coherence, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), Association for Computational Linguistics, Online, 2021, pp. 759–766. doi:10.18653/v1/2021.acl-short.96. [16] J. Rieger, C. Jentsch, J. Rahnenführer, RollingLDA: An update algorithm of Latent Dirichlet Allocation to construct consistent time series from textual data, in: Findings of the Association for Computational Linguistics: EMNLP 2021, Association for Computational Linguistics, Punta Cana, Dominican Republic, 2021, pp. 2337–2347. doi:10.18653/v1/2021.findings-emnlp.201. [17] J. Rieger, K.-R. Lange, J. Flossdorf, C. Jentsch, Dynamic change detection in topics based on rolling LDAs, in: Proceedings of the Text2Story’22 Workshop, CEUR-WS 3117, 2022, pp. 5–13. URL: https://ceur-ws.org/Vol-3117/paper1.pdf. [18] A. Boutaleb, J. Picault, G. Grosjean, BERTrend: Neural topic modeling for emerging trends detection, in: Proceedings of the Workshop on the Future of Event Detection (FuturED), Association for Computational Linguistics, Miami, Florida, USA, 2024, pp. 1–17. doi:10.18653/v1/2024. futured-1.1. [19] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, LoRA: Low-rank adaptation of large language models, in: International Conference on Learning Representations, 2022. URL: https://openreview.net/forum?id=nZeVKeeFYf9. [20] N. Benedek, L. Wolf, PRILoRA: Pruned and rank-increasing low-rank adaptation, in: Y. Graham, M. Purver (Eds.), Findings of the Association for Computational Linguistics: EACL 2024, Association for Computational Linguistics, St. Julian’s, Malta, 2024, pp. 252–263. URL: https: //aclanthology.org/2024.findings-eacl.18/. [21] D. J. Kopiczko, T. Blankevoort, Y. M. Asano, VeRA: Vector-based random matrix adaptation, in: The Twelfth International Conference on Learning Representations, 2024. URL: https://openreview. net/forum?id=NjNfLdxr3A. [22] S.-Y. Liu, C.-Y. Wang, H. Yin, P. Molchanov, Y.-C. F. Wang, K.-T. Cheng, M.-H. Chen, DoRA: Weight-decomposed low-rank adaptation, in: R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, F. Berkenkamp (Eds.), Proceedings of the 41st International Conference on Machine Learning, volume 235, PMLR, 2024, pp. 32100–32121. URL: https://proceedings.mlr.press/ v235/liu24bn.html. [23] M. Nikdan, S. Tabesh, E. Crnčević, D. Alistarh, RoSA: Accurate parameter-eficient fine-tuning via robust adaptation, in: Proceedings of the 41st International Conference on Machine Learning, ICML’24, JMLR.org, 2024. URL: https://dl.acm.org/doi/10.5555/3692070.3693618. [24] T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer, QLoRA: Eficient finetuning of quantized LLMs, in: Advances in Neural Information Processing Systems, volume 36, Curran Associates, Inc., 2023, pp. 10088–10115. URL: https://proceedings.neurips.cc/paper_files/paper/2023/ file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf. [25] Y. Xu, L. Xie, X. Gu, X. Chen, H. Chang, H. Zhang, Z. Chen, X. Zhang, Q. Tian, QA-LoRA: Quantization-aware low-rank adaptation of large language models, in: The Twelfth International Conference on Learning Representations, 2024. URL: https://openreview.net/forum?id= WvFoJccpo8. [26] T. Zadouri, A. Üstün, A. Ahmadian, B. Ermis, A. Locatelli, S. Hooker, Pushing mixture of experts to the limit: Extremely parameter eficient MoE for instruction tuning, in: The Twelfth International Conference on Learning Representations, 2024. URL: https://openreview.net/forum?id= EvDeiLv7qc. [27] Anonymous, PERFT: Parameter-eficient routed fine-tuning for mixture-of-expert model, 2025.
URL: https://openreview.net/forum?id=PPjpGTPG5K. [28] Z. Wang, D. Chen, D. Dai, R. Xu, Z. Li, Y. Wu, Let the expert stick to his last: Expert-specialized ifne-tuning for sparse architectural large language models, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Miami, Florida, USA, 2024, pp. 784–801. doi:10.18653/v1/2024.emnlp-main.46. [29] J. Rieger, M. Ruckdeschel, G. Wiedemann, PETapter: Leveraging PET-style classification heads for modular few-shot parameter-eficient fine-tuning, 2024. arXiv:2412.04975. [30] T. Schick, H. Schütze, Exploiting cloze-questions for few-shot text classification and natural language inference, in: P. Merlo, J. Tiedemann, R. Tsarfaty (Eds.), Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, Association for Computational Linguistics, Online, 2021, pp. 255–269. doi:10.18653/v1/ 2021.eacl-main.20. [31] B. Chan, S. Schweter, T. Möller, German‘s next language model, in: D. Scott, N. Bel, C. Zong (Eds.), Proceedings of the 28th International Conference on Computational Linguistics, International Committee on Computational Linguistics, Barcelona, Spain (Online), 2020, pp. 6788–6796. doi:10.
A. Potential national and regional newspapers under study The following newspapers are considered to be taken into account in our analyses: • national newspapers: Bild, Der Tagesspiegel, Die Welt, Süddeutsche Zeitung, taz — die
Tageszeitung • magazines/weekly newspapers: Der Spiegel, Stern • regional newspapers: Reutlinger Nachrichten, Schwarzwälder Bote, Stuttgarter Nachrichten,
Stuttgarter Zeitung, Südwest Presse • thematic outlets: EID Energie Informationsdienst, Energie & Management/Powernews.org, Handelsblatt, KI — Kälte Luft Klimatechnik,
The programming code is publicly available via our GitHub project repository https://github.com/ LarsG321/Dissemination-Monitoring-narratives-about-the-energy-transition-in-Germany. Due to license restrictions, we are unable to share the raw data. The following preprocessing steps are conducted for all corpora: • Filter for texts between 2014-01-01 and 2024-03-31 • Remove duplicated texts • Remove umlauts, HTML and XML nodes • Remove special characters (apparently we missed removing the € sign) • Replaced line breaks with spaces • Lowercase the text • Remove German stopwords (leveraging a common and rather conservative list) • Remove punctuation • Remove numbers LDA hyperparameters used for all corpora:
• = 5, 6, 7, 8, 9, 10, 15, 20 ; alpha = eta = 1/K RollingLDA hyperparameters used for all corpora (if not specified, we use the default parameters from the function RollingLDA in the software package rollinglda [16]): • init = "2014-02-01" (first month of data) • chunks = "month" (monthly updates) • memory = "month" (one month of memory) • memory.fallback = 0, 1, 2 (number of documents used as memory if there are no documents in the “original” memory time period) Change detection [17] hyperparameters used for all corpora: • = {1, 2, 3, 4} (maximum number of months in the reference period) • = {0.80, 0.81, 0.82, 0.83, 0.84, 0.85, 0.86, 0.87, 0.88, 0.89, 0.90} (quantile to determine threshold) Final parameters for RollingLDA and change detection: • Bild: memory.fallback=0, = 10, = 1, = 0.85 • SZ: memory.fallback=0, = 10, = 1, = 0.85 • Welt: memory.fallback=2, = 10, = 1, = 0.80
BILD − Topic 1 to Topic 5 BILD − Topic 6 to Topic 10 1500 : 1 ic m p re o
T G afd trump roy putin rechts land lindner putins s t c i l f n o d n a s c i t i l o P n a : 3 ic m p re o
T G luetzerath graichen krim hitler steinmeier flick s c i t i l o P e t a t S n a i : g 2 n ic ir p u o h
T T euro kemmerich manta tina haut schweiger goetz inflation syrien adresse frodeno s e i t i l a n o s r e P d n a s c i t i l o P n a : 4 ic m p re o
T G scholz kohl laschet habeck robert heynckes olaf jupp eier y c i l o P y m o n o c E n e e r G n a : 5 ic m p re o
T G habeck gruene robert trump minister waermepumpe gasheizungen akw heizhammer 2014 2016 2018
2020 Date 2022 2024 2014 2016 2018
2020 Date 2022 2024 cdu neubauer ukraine luisa druckerei klamroth newsprint erdogan putin tipp olaf kueblboeck h t l a e H d n a , s b : o 9 J c , i o p r o u
T E jobgeld gesundheit tipp ps geld zaman y m o n o c E d n a s c i t i l o P n a : 0 1 ic m p re o
T G tagestrend kohl jobgeld charles neuinfektionen gorbatschow euro rasen muskeln baerbock euro € tagestrend helmut 10 r e c c o S h c i n u :ic1opT reaynBM spd gruenen partei cdu fdp scholz csu koalition ki afd fridays m s i v i t c
A :4 te c a ip im oT lC stadt schueler meter wasser for greta future thunberg autos n o i t i s n a r T y g r e n
E :5 an icopT reGm strom euro wasserstof gas kosten deutschland bundesregierung pro energien s e u s s I l a i c o S d n :ic6p liitsca oT oP trump biden polizei aktivisten generation israel hamas e c n a n i F d n a y m o n o c
E :7 an icopT reGm unternehmen euro inflation milliarden firmen preise wirtschaft demokratie geld men- ezb schen gesellschaft banken putin china trump usa praesident europa russischen russland museum kinder e r u t l u :ic9 adnC p t oT rA kunst buch menschen the welt leben kuenstler e f i L d :0 an 1 ly c i iopT aFm frau leben mutter eltern mann leute familie geschichte frauen gesellschaft kind s t r o p S d n a s c i t i l o
P : n 1 a ic m op re
T G seite lesen fc dax innenpolitik sport polizei angaben t n e m n o r i v n E d n a e g n a h
C :3 te c a iop lim
T C erde co wissenschaftler meter klimawandel holz y l i m a
F : d 2 n c a i op ife
T L leben mann buch frau eltern mutter menschen stadt geschichte grad s e i t r a P l a c i t i l o
P : n 4 a ic m op re
T G spd cdu afd fdp union laschet scholz ezb zinsen s c i t i l o P d n a y g r e n
E : n 6 a c i i s p s o u
T R trump euro putin strom Date forscher partei s c i t i l o
P : n 7 a ic m op re
T G spd partei cdu csu gruenen merkel afd fdp seehofer usa russische geht trump politik freiheit jungen men- eu schen gesellschaft milliarden leben euro gibt unternehmen kinder dollar e d a r T d n a y m :0 on 1 o ic cE p o U
T E europaeischen europa china