1 Introduction

The election to the 20th German Bundestag in 2021 can be considered a historic election. After 16 years, Angela Merkel, who is a member of the party Christian Democratic Union (CDU), did not run again for the office of chancellor. The sister parties CDU and Christian Social Union (CSU), who constitute a parliamentary group in the Bundestag, received fewer votes compared to the Social Democratic Party (SPD), of which Olaf Scholz is a member, who was subsequently elected as new chancellor. After the election, the SPD formed a government with the Green Party (Bündnis 90/Die Grünen) and the Free Democratic Party (Liberals, FDP).

Polls of eligible voters showed that topics such as social security, climate, economy, labor and the management of the ongoing COVID-19 pandemic were decisive for them in deciding which party to vote for in this election.Footnote 1 Given the continued presence of state-level pandemic restrictions during the election year, campaigning on social networks played an important role. One of the most popular social networks used by many politicians to disseminate political content is Twitter.Footnote 2 Twitter enables users around the world to express themselves and interact with others through short messages called “tweets”. In the election year 2021, there was still a limitation so that a tweet could have a maximum of 280 characters. This limitation encourages users to express themselves briefly and clearly. Twitter also supports the sharing of images, GIFs, and videos alongside text messages.

Due to the large amount of publicly available data, Twitter is often used in research to gauge political sentiment by analyzing both: the tweets of politicians and party accounts, as well as tweets by the general public (Budiharto & Meiliana, 2018; Costa et al., 2021; Hellwig et al., 2023; Schmidt et al., 2022). Furthermore, it is possible to gain an understanding of what political topics are being discussed at a given time. An established method for this task in natural language processing (NLP) to identify topics in text, is to apply topic modeling. Topic modeling is a statistical technique to identify latent topics or themes within a collection of documents, such as tweets (Hong & Davison, 2010). As an illustration, terms such as “covid”, “vaccine” and “lockdown” might be grouped together to capture the thematic emphasis on COVID-19 pandemic-related matters. Topic modeling have been employed extensively over the past decade to uncover latent structures within a corpus, enabling a broad spectrum of applications (cf. Boyd-Graber et al., 2014). In this paper, we apply topic modeling to explore topics that were discussed on Twitter during the 2021 federal election campaign. We examined two different type of German language tweets: (1) the tweets of a fixed set of accounts of politicians and political parties and (2) tweets from users who mentioned those accounts with the @-sign to gain further insights into the general public’s perspective. In addition to the text of the tweets, the images included in these were analyzed as well. The research questions are as follows:

  • What are the major topics considering the tweets of the entire election year 2021?

  • How do the topics addressed in the tweets of the political actors differ from the topics addressed in the tweets of general users who mention these accounts?

  • How do the number of tweets for each identified topic change over the course of the election year?

  • How does the sentiment of tweets assigned to specific topics evolve throughout the course of an election year?

  • Which topics are addressed by both political parties and the general public?

The main contributions to the research area are as follows:

  • The extension of a pre-existing corpus of 58,864 tweets by 11,817 images posted by 89 Twitter accounts of the major German political parties for the election year 2021.

  • The extension of a pre-existing corpus of 707,241 tweets by 26,371 images that were included in tweets that mentioned (using @-sign) the 89 Twitter accounts of the major German parties for the election year 2021.

  • Application of topic modeling via BERTopic as proposed by Grootendorst (2022) to identify topics in both corpora.

  • Analysis of topics in context of sentiment analysis results throughout the year.

Resources related to this work such as programming code, visualizations and corpus information are publicly available on GitHub.Footnote 3

2 Related work

In this section, we first elaborate on the current state-of-the-art of topic modeling, and then summarizes the research in the context of topic modeling on political Twitter.

2.1 Methods for topic modeling

Topic modeling has proven to be an effective, unsupervised method to identify common patterns and relationships in textual data (cf. Jelodar et al., 2019). Many previously introduced machine learning approaches for topic modeling are based on Latent Dirichlet Allocation (LDA) (Jelodar et al., 2019). LDA is a generative probabilistic model first introduced by Blei et al. (2001). Documents are treated as combinations of different underlying topics. Each topic is defined by a collection of words that are likely to appear together (Jelodar et al., 2019). By examining the words with the highest probabilities in each topic, LDA helps one to understand the main themes or subjects discussed in the documents (Jelodar et al., 2019). Topic models based on LDA have been applied in various fields such as medical science (Paul & Dredze, 2011), software engineering (Asuncion et al., 2010), social media analysis (Moßburger et al., 2020; Schmidt et al., 2020b), digital humanities (Schmidt et al., 2020a) and political science (Dahal et al., 2019; Karami et al., 2018; Xue et al., 2020).

A limitation of methods based on LDA is that they overlook the semantic connections between words by utilizing bag-of-words representations (Grootendorst, 2022). By disregarding the contextual information of words within a sentence, the bag-of-words approach may not effectively represent the documents (Grootendorst, 2022). To address this problem, text embedding techniques such as Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., 2019) have gained popularity. Unlike most previous models that treat words in isolation, BERT processes words in relation to their surrounding context, enabling a deeper understanding of words and capturing their semantic relationships (Devlin et al., 2019). One way to use such language models for topic modeling is BERTopic which was introduced by Grootendorst (2022) and gained attention recently. A document is first converted into a dense vector representation using Sentence-BERT (Grootendorst, 2022). The dimensionality of the document embeddings is then reduced with the help of Uniform Manifold Approximation and Projection (UMAP) (Grootendorst, 2022). Finally, clustering algorithms such as k-Means or Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) can be used to identify clusters or topics, respectively (Grootendorst, 2022). Three corpora were considered for evaluation by Grootendorst (2022), including a corpus consisting of 44,253 tweets posted by Donald Trump. When comparing BERTopic with LDA in terms of the common evaluation metrics topic diversity and topic coherence, BERTopic always outperformed LDA in the case of all three corpora, including the Twitter corpus.

2.2 Topic modeling on Twitter for political research

Topic modeling has proven to be an effective approach for extracting information within various political contexts on the Twitter. Previous research has applied topic modeling techniques to identify different topics within a broader political scope. For example, Miller (2019) examined a publicly available corpus of IRA/Russian Federation associated tweets from July 2014 to September 2017 related to the 2016 US presidential election to identify key topics being discussed in such tweets. Among the 56 topics identified, seven were found to be associated with the campaign of presidential candidate Hillary Clinton, and these were all identified as being mostly negative in terms of their sentiment. In contrast, there were ten topics that referenced presidential candidate Donald Trump, with four of them expressing support. Additionally, topics emerged that involved President Putin and the Russian Federation in general, but no consistently positive or negative sentiment of these topics was discernible. The author’s conclusion emphasizes that topic modeling uncovered the heterogeneous and contradictory nature of the corpus, as tweets addressed a wide range of political positions and issues.

Dahal et al. (2019) collected 309,016 geotagged tweets related to climate change from July 1, 2016, to February 28, 2018. They identified topics such as transportation, energy, fossil fuel industry, and international agreement within this context. They showed that certain topics are discussed more frequently at times when they are also of importance in the political debate. For example, the topic of international agreement was extensively discussed at the time when the US declared the withdrawal from the Paris Climate Accords. When comparing different countries in which tweets were posted, it was observed that for Australia, nearly 50% of the collected tweets were assigned to the topic of energy, whereas in other countries like Canada, the United Kingdom, and the United States, this proportion was significantly lower (below 25%).

Achmann and Wolff (2023) applied BERTopic to explore topics in posts and stories (temporary available photo or video content) posted by political parties and politicians on Instagram in the context of the 2021 German federal election campaign. Achmann and Wolff (2023) employed BERTopic to derive 25 topics each from the posts and stories. The majority of posts dealt with policy issues, while the majority of stories did not deal with policy issues. Instead, they were concerned with the documentation of the rallies and campaign trail or call-to-action content, i.e. content that call on people to vote for a certain party or politician.

Furthermore, there have been studies investigating elections and the topics discussed by users on Twitter. Karami et al. (2018) collected 24 million tweets on the two candidates Barack Obama and Mitt Romney in the context of the 2012 US presidential election. Topics with words in the context of economy, jobs, budget deficit, healthcare and tax were identified. Taking the sentiment of tweets into account, it was shown that the topic of budget deficit was significantly more negative in tweets collected for Mitt Romney compared to those collected for Barack Obama.

Overall, topic modeling is widely utilized in the political context of Twitter to identify the wide range of topics discussed in tweets. Once topics are identified, it is common to examine, variations in topic prevalence over time and in different geographic regions as well as differences in sentiment of tweets on specific topics.

3 Methodology

3.1 Data acquisition

3.1.1 Tweets by political party accounts

In order to examine the topics addressed in tweets posted by political party accounts during the election year 2021, we used a Twitter corpus by Schmidt et al. (2022). The corpus comprises 58,864 tweets (see Table 1), whereby these were posted by 89 politicians and political party accounts of the seven largest parties represented in the Bundestag before and after the election (see Table 6 for the full list of accounts). The CDU and its sister party, the Bavarian regional party CSU, were considered as one party. For every party, the ten politician accounts and the four official party accounts with the highest number of followers as of January 2022 were considered. Furthermore, for each tweet in the corpus, there is already a sentiment classification for the text of the tweet, which can be either positive, negative or neutral. The sentiment was determined using a BERT model, which was fine-tuned by Schmidt et al. (2022).

Table 1 Statistics on the entire corpus collected by Schmidt et al. (2022): Tweets by politicians and official party accounts of the respective party
Full size table
Fig. 1
figure 1

Examples of images posted by political party accounts (see Table 8 for details on example images)

Full size image

We will include the text of the tweets into our upcoming topic modeling analysis. The tweets of this corpus by Schmidt et al. (2022) are referenced in the following as tweets by political party accounts. In addition to analyzing the tweet’s text, we also examine the text in the images posted in tweets. In order to do so, we subsequently extracted the text of the images via OCR. Version 5.3.1 of TesseractFootnote 4 was used, and we were able to extract text from around half of the images (21,959) posted by political party accounts. Some images are mainly photographs of the politicians of the respective party and thus offer no text content. In case text is discernible within an image, the text often comprises quotations or political statements (see Fig. 1 for some examples).

3.1.2 Tweets mentioning politicians

In order to analyze which topics were addressed by accounts that mentioned political party accounts on Twitter in the election year, topic modeling was carried out on the corpus already acquired by Hellwig et al. (2023) in a previous study. The corpus comprises 707,241 tweets (see Table 2) that mentioned (using @-sign) the 89 party and politician accounts considered by Schmidt et al. (2022). Similar to the tweets in the corpus by Schmidt et al. (2022), there is a sentiment classification for the text of each tweet, which can be either positive, negative or neutral. Again, the sentiment was identified with the help of a fine-tuned BERT model (Hellwig et al., 2023).

Table 2 Statistics on the entire corpus collected by Hellwig et al. (2023): Tweets that mentioning accounts of the respective parties
Full size table
Fig. 2
figure 2

Examples of images posted by users mentioning political party accounts (see Table 9 for details on example images)

Full size image

Again, by using Tesseract, text was extracted from the images, although the number of images that could be extracted from the tweets was lower in proportion to the total number of tweets comparing it with the tweets posted by the political party accounts. When looking at some of the images, it can be noticed that the type of images is very different (see Fig. 2) to the ones by parties and politicians. Commonly, the visual content includes screenshots, such as those from news articles, as well as photographs depicting politicians affiliated with a respective political party.

3.2 Analyzing political topics in tweets using BERTopic

3.2.1 Topic modeling

Topic modeling was carried out for both the corpus with tweets from the political party accounts and the corpus with tweets that mentioned these accounts. A model was fitted for each of the following sub-corpora separately:

  • Text extracted from the tweets posted by political party accounts.

  • Text extracted from the images posted by political party accounts.

  • Text extracted from tweets mentioning political party accounts.

  • Text extracted from images of tweets mentioning political party accounts.

BERTopic was chosen because it incorporates the semantic relations between words by using embedding representation of documents, which helps to generate more meaningful topics (as already outlined in Sect. 2). Furthermore, it enabled us to analyze the frequency of documents on specific topics over time and the comparison of the topics identified in each sub-corpora, which is relevant to answer the research questions. Apart from utilizing BERTopic, we also employed Latent Dirichlet Allocation (LDA) to identify topics in the sub-corpora. However, the LDA models produced a relatively fewer number of meaningful topics. Often, there was no distinct thematic pattern that could be discerned based on the most prevalent tokens.Footnote 5

For the concrete application of BERTopic, we used the official implementation for PythonFootnote 6, which uses HDBSCAN as a clustering algorithm. Without limiting the minimum number of documents per cluster, well over 50 clusters which are regarded as topics were identified for each of the four sub-corpora, however with some topics having a very small number of documents associated with them. In order to obtain more meaningful and important topics, a minimum number of documents per topic was defined based on the size of the sub-corpora. We determined that a topic must encompass a minimum of 1/250 of the total documents in the sub-corpus. Taking this limit into account, we were able to generate 25 topics for each sub-corpus.

3.2.2 Preprocessing and representation

The text of both tweets and the extracted text from images underwent preprocessing, which included removing punctuation and stop words, converting the text to lowercase as well as lemmatization via the Python package SpacyFootnote 7 which proved to be beneficial preprocessing steps for German language text (Fehle et al., 2021). Additionally, documents with fewer than five words were excluded. To represent the preprocessed tweets, a Sentence-BERT model, specifically “paraphrase-multilingual-MiniLM-L12-v2”Footnote 8, which is suggested for multilingual documents or any other language than English.Footnote 9 All tweets were transformed into embeddings using an NVIDIA GeForce GTX 1080Ti GPU with 11 GB VRAM. As suggested by Grootendorst (2022), UMAP is used to reduce the dimensionality of the embeddings, which can improve the performance of clustering algorithms such as k-Means or HDBSCAN in terms of clustering accuracy and time (Allaoui et al., 2020).

4 Results

4.1 Identified topics

In this section, we present the topics derived from tweets originating from the political party accounts, as well as those mentioning these accounts. The appendix D encompasses diagrams illustrating the prevalence of the topics identified for each of the four sub-corpora over the course of the election year. For the topics identified within the sub-corpora of tweet texts, supplementary diagrams are provided, offering insights into the sentiment trends associated with each topic throughout the election year.

4.1.1 Tweets of politicians

Topics identified in tweets posted by political party accounts are presented in Table 3. An English translation for all tokens can be found in the Table 10. For both sub-corpora, it can be noticed that there is one topic to which a comparatively large number of tweets were assigned (topic 1). These were mainly tweets that could not be assigned to a policy-focused topic, but referred to politicians, parties and their programs in general. In this context, some identified topics focused specifically on individual parties (e.g. tweet text: topic 6; tweet image: topic 7, topic 14, topic 15).

Table 3 Identified topics in tweets posted by political party accounts
Full size table

Looking at the policy-focused topics, some were debated throughout the election year. An extensively discussed topic is the COVID-19 pandemic and the challenges it posed to the society (tweet text: topic 2; tweet image: topic 6). However, topics related to COVID-19, such as masks (tweet text: topic 22) or vaccinations (German: “Impfung”) (tweet image: topic 11) were also identified. In the majority of tweets assigned to this topic, a negative sentiment was expressed. Furthermore, the topic was less frequently discussed between May and October compared to the other months. Further policy-related topics that endured throughout the election year include climate policy (tweet text: topic 3; tweet image: topic 4), digitization (tweet text: topic 4; tweet image: topic 3) and finance policy (tweet text: topic 5; tweet image: topic 5).

Furthermore, some topics were primarily discussed at certain points in time. As an example, there’s a topic surrounding Afghanistan and the return to power of the Taliban in August 2021, which was identified in both sub-corpora (tweet text: topic 12; tweet image: topic 16) and was discussed the most in August 2021. Similarly, a topic focused on Israel and antisemitism garnered significant attention in May 2021 (tweet text: topic 12).

4.1.2 Tweets mentioning accounts by politicians

All topics that were identified in the two sub-corpora of tweets mentioning the political party accounts are presented in Table 4). An English translation for all tokens can be found in the Table 11. As for the tweets posted by the political party accounts, a comparatively large topic could be identified for both sub-corpora to which tweets were assigned that were not policy-focused (tweet text: topic 1, tweet images: topic 1).

Table 4 Identified topics in tweets mentioning political party accounts
Full size table

Policy-focused topics that were debated throughout the election year were identified. As for the tweets of the political party accounts, topics focusing on the COVID-19 pandemic were identified. There are several topics related to different aspects of the COVID-19 pandemic, such as vaccinations (tweet text: topic 2, tweet images: topic 5), masks (tweet text: topic 18, tweet image: topic 18) or lockdowns (tweet text: topic 24).

Other policy-focused topics include finance policy (tweet text: topic 4) or social issues like gender equality and racism (tweet text: topic 11, tweet image: topic 14). In both sub-corpora, we identified a prevalent topic focusing on cannabis (tweet text: topic 17, tweet image: topic 11), which remained a prominent subject of discussion throughout the entire election year. It is worth mentioning that as of April 1st, 2024, the possession of small quantities has been legalized in Germany. However, there was an outlier in October for both sub-corpora, meaning that the topic was debated comparatively much in that month (see Fig. 6).

In addition, it can be noticed that within the sub-corpus containing tweet texts, two distinct topics emerged, encompassing tweets that not only garnered significant attention throughout the election year but almost exclusively express a negative sentiment. This is on the one hand a topic focusing on the party AfD (tweet text: topic 13) and on the other hand a topic focusing on corruption scandals of the parties CDU and CSU and corruption in general (tweet text: topic 16).

As for the tweets posted by political party accounts, we identified topics that underwent extensive debates on Twitter at specific points in time. As for the political party accounts, a topic with regard to Afghanistan and the return to power of the Taliban in August 2021 (tweet image: topic 13) was identified but only for the sub-corpus of images. Equally, we identified a topic focusing on Israel and antisemitism (tweet text: topic 10, tweet image: topic 19), which was extensively discussed in May 2021.

Fig. 3
figure 3

Tweet text: Cosine similarity matrix for topic similarity comparison of topics identified in the corpus of tweets posted by political party accounts and tweets mentioning political party accounts

Full size image

4.2 Topic similarity between corpora

To further illustrate the presence of topics addressed in both tweets of political party accounts and those mentioning them, we identified similar topics between the sub-corpora using their corresponding topic embeddings. Subsequently, we compared these topics addressed by these two perspectives using cosine similarity, and for visualization purposes, we utilized a cosine similarity matrix.

We considered on the one hand topics that could be identified in the tweet text (see Fig. 3), but on the other hand also topics that were identified in the text extracted from images (see Fig. 8). Upon examining the two resulting cosine similarity matrices, several topics are being discussed by both perspectives, such as COVID-19, climate policy, financial policy, antisemitism, and social issues like gender equality and racism. Topics like cannabis (text of mentions: topic 17, text extracted from images: topic 11) or humor and satire (text of mentions: topic 23) could not be identified on the side of political party tweets. On the other hand, there are topics like Turkey (text of political party tweets: topic 24) for which no equivalent topic could be identified on the side of tweets that mentioned political party accounts.

5 Discussion

In the following section, we discuss and interpret the overall results, highlighting notable findings that emerged from our analysis. First, we collected the images posted in tweets from the corpora curated by Schmidt et al. (2022) and Hellwig et al. (2023) and used OCR to extract their texts. The average number of images per tweet was higher in tweets posted by political party accounts compared to tweets that mentioned political party accounts.

We employed BERTopic to identify topics in four sub-corpora: the two newly curated sub-corpora and the existing sub-corpora containing the tweet texts curated by Schmidt et al. (2022) and Hellwig et al. (2023). Similar to the findings by Achmann and Wolff (2023), we observed the presence of a comparatively large topic within all sub-corpora we examined, which does not focus on specific policies, but concentrates on politicians, parties and their programs in general. Both on the side of tweets that were posted by political party accounts and on the side of tweets that mentioned political party accounts, the COVID-19 pandemic was an intensively discussed topic. Additionally, various other topics also addressed different aspects partly related to COVID-19. Schmidt et al. (2022) speculated that the COVID-19 pandemic might be a reason for the overall negative sentiment in tweets from political party accounts. Our findings support this hypothesis, as we observed that tweets related to COVID-19 were predominantly associated with negative sentiment.

Furthermore, a topic surrounding Afghanistan and the return to power of the Taliban was identified in the tweets of political party accounts. Since the sentiment is primarily negative, this debated topic could indeed be a cause for the overall negative sentiment of tweets from some parties in August, as suggested by Schmidt et al. (2022). Other debated topics include finance policy, social issues and energy policy, which have been subjects of debate on Twitter in previous elections as well (Karami et al., 2018; Miller, 2019).

We proceeded with a comprehensive comparison of topic embeddings to ascertain the presence of topics addressed in both tweets by political party accounts and tweets that mentioned those. Topics such as the COVID-19 pandemic, climate policy, financial policy, antisemitism and social issues were discussed by both sides. Nevertheless, certain topics, such as cannabis, were exclusively addressed in tweets mentioning political party accounts. Notably, it is likely that these topics can also be identified in tweets of the other perspective by increasing the number of topics to be identified beyond 25.

Overall, BERTopic proved to be effective in identifying meaningful topics, although, especially considering the image sub-corpora, there were topics for which the thematic context can not be clearly discerned based on the most frequent tokens and tweets assigned to them (e.g. Table 3b, topic 22).

6 Limitations

Our work provides valuable insights into the topics debated in both tweets by political party accounts and tweets mentioning them on Twitter in the election year 2021. However, there are certain limitations of our work that we intend to address: While we attempted to capture the topics addressed through images by extracting text from them, we need to acknowledge that this approach may only partially capture the range of topics conveyed via this medium. In future work, we intend to enhance the analysis by incorporating automated image captioning methods or human annotations to gain a more comprehensive understanding of the topics addressed through images and how they are portrayed in terms of sentiment and presentation. In addition, it is important to note that, occasionally, the applied OCR might misinterpret text in the images, particularly in the case of proper nouns or unique words. Another limitation is the restriction of identifying a maximum of 25 topics for each sub-corpus. As a result, there are inevitably other topics that were discussed in tweets of a certain sub-corpus, but do not represent an independent topic because not enough tweets were assigned to them. Finally, it is worth noting that Twitter’s popularity and usage in Germany is not as widespread as in other countries. Only 10% of Germans regularly use Twitter,Footnote 10 in contrast to 23% of U.S. adults.Footnote 11 As a consequence, the corpora curated only represent a limited subsection of public (social media) sentiment.

7 Conclusion and future work

In conclusion, this study provides insights into the Twitter discourse surrounding the 2021 German federal election by utilizing BERTopic for topic modeling. Examining both the tweet text and the text extracted from images allowed to analyze a wide range of topics discussed on Twitter throughout the election year. Our analysis of tweets posted by accounts of the major German political parties and those mentioning them revealed that the most discussed topics and sentiment trends throughout the election year. In all sub-corpora examined, we noted the prevalence of a larger topic not focussing on policy-oriented topics. Instead, that topic tends to focus on politicians, parties, and their general campaigns. Looking at policy-focused topics, we found that the COVID-19 pandemic, climate policy, finance policy and social issues like racism and gender equality were among the most prominent topics discussed on Twitter during the election.

In future work, we plan to conduct a more in-depth analysis of visual content such as images and videos shared in the context of the election. In addition to extracting text from images, we intend to explore the emotion tone and sentiment conveyed through these using computer vision techniques (Schmidt et al., 2021b; El-Keilany et al., 2022). We also see potential in switching from the basic sentiment concept to more fine-grained emotion analysis (Schmidt et al., 2021a; Dennerlein et al., 2023) and also include methods of aspect based sentiment analysis to analyse the cause-effect relation of sentiments and emotions (Fehle et al., 2023). Finally, one could investigate the alignment between topics discussed on Twitter and the debates in the German Bundestag. By exploring potential overlaps between political discourse on social media platforms and formal legislative discussions, we seek to explore the extent to which public concerns find resonance in the legislative process.Footnote 12