Social media platforms have become central to shaping political discourse, serving as arenas where narratives form and evolve, in!uencing public opinion. Identifying and analyzing these narratives, particularly when they compete across di"erent political ideologies, is crucial for understanding the dynamics of modern political communication. This paper presents an unsupervised framework for identifying and characterizing competing narratives in political discourse on social media, focusing on German politicians' tweets. The framework employs a multi-stage pipeline that integrates natural language processing techniques such as topic modeling, event detection, and event linking. By forming data into coherent stories and uncovering the distinct perspectives of user communities, the system is able to detect the key competing narratives, highlighting the divergent framings and con!icts surrounding trending political topics. Two case studies on polarizing political issues demonstrate the e#cacy of the methodology, showcasing its ability to uncover and analyze divergent viewpoints. The $ndings contribute to the broader understanding of how narratives propagate within the digital public sphere and o"er insights for policymakers, social media platforms, and researchers interested in monitoring political discourse.
Social media has become a central arena for political discourse, shaping public opinion. Platforms like X
(formerly Twitter) are key outlets for politicians to communicate directly with citizens [
Our work contributes to the development of computational narrative framing analysis [
This research area holds promise for fostering evidence-based approaches to societal challenges.
For example, promoting media literacy and critical thinking through exposure to diverse viewpoints
can empower individuals to evaluate information more critically, identify biases, and make informed
decisions [
Narrative studies have a long history across various disciplines, including literary studies, psychology,
and sociology. Narratives are typically de$ned as representations of a connected succession of events
[
A crucial aspect of computational narrative understanding is the automatic extraction of narrative
elements from text data. This includes identifying actors, actions, events, settings, and the relationships
between them [
Framing understanding is another crucial aspect in shaping the reader’s perception of a narrative.
Computational framing analysis aims to automatically detect framing devices and understand how they
are used to in!uence interpretation [
Finally, studying competing narratives is essential for understanding how di"erent perspectives and
interpretations of events emerge and spread. This could help to monitor community interests, to counter
propaganda campaigns, and to model the spread of (mis-)information [
For our study, we focus on the narratives within the political discourse of German politicians on Twitter/X. Before collecting the corresponding tweets, the relevant politicians and their Twitter accounts must $rst be identi$ed. For this task, we leverage the structured database Wikidata, which provides information on politicians’ a#liations with political parties and their country of citizenship.
We $lter Wikidata entities by the occupation property matching “politician”. Subsequently, the country of citizenship is utilized to $lter German politicians. Their party a#liation is provided by the member of political party property. However, this property may list multiple parties, as politicians may change their a#liation over time. In certain instances, the current party is marked by a quali$er. Otherwise, we consider the $rst party listed, ignoring any a#liation with an end date. Extracting Twitter usernames from Wikidata with the above $lters yielded 1,324 accounts belonging to 1,300 politicians.
Using Twitter’s API, we collected all tweets of these politicians from January 1st, 2022, to June 24th, 2023. The resultant dataset comprised 189,850 tweets from 786 accounts, excluding retweets, replies, and quotes. For presentation purposes, we translated the tweets into English using the NLLB model [26]. Otherwise, all the analyses are conducted on the original content (i.e., original language).
We operationalize narratives as semantically and temporarily coherent stories represented by structured interpretations of events and issues, providing meaning and context to political communication. Starting with the social media posts, we sought to identify these stories and the competing narratives of di"erent user groups within them. Given a single tweet’s brevity and limited scope, we follow a corpus-level approach to identify narratives rather than trying to extract them from individual tweets.
Our methodology consist of a multi-stage pipeline (see Fig. 1). After identifying general topics within the data, we then detect events that form the building blocks of our stories. Events are linked into coherent stories from which we extract di"erent viewpoints, which we refer to as competing narratives.
Note that, during the narrative extraction process, we disregard tweet authors’ political a#liations, as this information is unique to our dataset and may not be available in other contexts. However, it will be used during the evaluation to assess the e"ectiveness of our method to identify competing narratives.
We are publicly releasing the full source code for our data processing pipeline1 to promote open science and ensure reproducibility. 1https://github.com/fjen/competing-narratives
Given the diversity of topics in the dataset, the $rst step is to identify individual topics and remove noise. An unsupervised topic model is trained to group documents based on their semantic similarity. Here, we choose an embedding-based approach, as it is more robust to the short and noisy nature of tweets than traditional topic models like LDA. In particular, we rely on the modular BERTopic framework [27], which essentially clusters documents based on their semantic embeddings and generates text representations for each topic. Notably, we use Sentence-BERT with the paraphrase-multilingual-MiniLM-L12-v2 model [28] for embedding generation and set a minimum cluster size of 500 documents to avoid overly $ne-grained topics.
After disregarding outliers (i.e., not clustered tweets), we identify 57 topics in the dataset containing 108,063 documents. A signi$cant portion of the documents are labeled as outliers (43%), which is expected given the nature of the data. As shown in Table 1, the remaining documents are grouped into topics that successfully capture the main discussion strands during the observation period with examples like the war in Ukraine, the COVID-19 pandemic, or Europe’s energy crisis.
We de$ne an event as a set of documents that discuss the same issue in close temporal proximity [
For the extraction of events, we borrow parts of the proposed Multi-TimeLine Summarization (MTLS)
approach by Yu et al. [
Tweets are clustered into events using the a#nity propagation (AP) algorithm [29]. It operates on an a#nity matrix S1, where S1(i, j) represents the similarity between tweets i and j based on a linear combination of their semantic similarity and temporal proximity:
S1(i, j) = · Stemp(i, j) + (1 → ) · Scos(i, j) where Scos is the cosine similarity of the tweet embeddings and Stemp is the temporal similarity de$ned by an exponential function (Eq. 2) with a decay factor :
Stemp(i, j) =
1 exp ·|ti↑ tj|
Our time unit is in days and we set to 0.05 as suggested by Yu et al.. Also, is set to 0.4 to balance the similarity measures. For example, we obtain 155 events for the "energy" topic (Topic 11 in Table 1).
Next, we group events into coherent stories, creating sequences of events that are meaningfully related.
Since we are working at the document level and not at the sentence level, we cannot use a procedure
similar to MTLS for event linking, which is based on sentence co-occurrence. Our approach is based on
an event similarity graph where we $lter the most signi$cant edges to partition the events into stories.
Event Similarity Measure Since each event contains multiple tweets, we must de$ne the semantic
similarity measure on the inter-event level. The AP algorithm used to cluster those tweets also returns
documents at the cluster centers as representatives of the events [
Similarly to event detection, an a#nity matrix S2 is created based on the similarity of event embeddings and a penalty term, temporal distance, to avoid connecting temporally distant events, which in turn could lead to improbable stories spanning the entire time range. We de$ne the event similarity S2(e1, e2) between a pair of events e1 and e2 as:
S2(e1, e2) = Ssem(e1, e2) · exp↑ ·|te1 ↑ te2 | (3) where Ssem is the similarity of the event embeddings, and tei is the timestamp of the representative document within an event. The temporal penalty decreases the similarity score as the time di"erence between the events increases, with the decrease rate controlled by the parameter. We set = 0.01. Forming Storylines To create coherent storylines, we leverage the events similarity matrix S2. We construct a directed graph G = (V, E), where V represents the set of events and E the edges between them. Each edge eij between events i and j is assigned a weight based on their similarity score S2(i, j). The edge direction is determined by the chronological order of the events given by the representative documents, ensuring that each event has a pathway to all subsequent events. To maximize coherence, we retain only the highest-weighted outgoing edge for each event. This helps us focus on the main storyline avoiding diverging substories. Finally, to separate the storylines, we apply the Leiden algorithm [30] for community detection, with each partition representing a di"erent story.
Figure 2 illustrates the outcome of the story extraction process for topic 11 (energy, gas, renewable, electricity, price), which resulted in the creation of 12 stories from 155 events.
To distinguish the di"erent viewpoints expressed in the extracted stories, we aim to identify users who share similar stances and are likely to contribute to the same narratives. Since politicians often hold aligned opinions on various topics, such as party policies, we adopt a global (i.e., not restricted to one particular topic) approach to user grouping. This method further ensures that even users less active within a given topic are appropriately categorized based on their overall trace of tweets.
Using a user’s posts, we aim to create an embedding that captures their overall position. Text embeddings have been shown to successfully quantify the degree of political bias in texts [31] and identify ideological placement [32, 33]. Moreover, they can highlight di"erences in how political groups discuss speci$c policy issues [34]. To determine each user’s overall stance, we $rst average the document embeddings of their posts within each single topic. This gives us a topic-speci$c representation of the user’s perspective. Next, we calculate the user’s relative position within that topic by subtracting the topic centroid from the user’s average embedding. This step helps to situate the user’s stance relative to the broader discourse on that issue. Finally, we average these topic-speci$c user embeddings across all topics. This global embedding represents the user’s overall political stance across various issues.
To unravel the diverse perspectives shaping the narratives within a story, we use the commonalities of users involved in terms of their global embeddings. For this, we cluster the user embeddings (via HDBSCAN) to form distinct communities of users who share comparable political views. With multiple users contributing to a story, we can now split the story posts according to their respective communities. As a result, each event within a story is enriched by the perspectives of the various user communities, re!ecting the collective voices that contribute to the discourse. The structure represented by a given community’s contribution to one particular story constitutes a competing narrative within that story.
We present the results based on applying the proposed framework to the dataset of German politicians’ tweets. We $rst evaluate the division of user positions to assess the overall coherence and consistency of the narratives within each user community. Then, we illustrate, through two examples, how the same story or event can be di"erently perceived and discussed by opposing groups. Our analysis shows how the proposed framework can help reveal distinct narratives emerging from the political discourse. We validate our approach by showing how supporters of di"erent (politically opposed) parties emphasizing divergent interpretations of the same events are segregated by our method into separate narratives. 5.1. EvaluatiUosenrEmbeddings with Parties of Uncovered User Communities
Party Distribution per Community 30 party
Alliance '90/The Greens
Alternative for Germany 25 20 s rseu 15 10 5 0 0 1 2 3 4
5 community (b) Distribution of members from two politically opposed parties within the user communities in Topic 23 (asylum, refugees, migration). y 3 6 5 4 2 1 0
Party
Social Democratic Party of Germany Free Democratic Party ChristianDemocratic Union Alliance '90/The Greens The Left ChristianSocial Unionof Bavaria
Alternative for Germany 0 2 4
6 x (a) User embeddings colored by political party. Markers’
size represents the users’ activity frequency. 0 −1.5 −1 −0.5 0 0.5 1
Sentiment (b) Sentiment Distribution of Communities in Story 10 (a) Story 10 within the directed graph of events
Since the identi$cation of distinct narratives within each story relies on user communities, we $rst verify the validity of the user classi$cation. In our dataset, each user represents a German politician. Using Wikidata, we extracted information about their current party memberships. This allows us to assess the alignment of the identi$ed communities with their respective political parties.
Figure 3a shows a 2D projection of the user embeddings generated in Section 4.4. We focus our analysis on the main parties currently represented in the German parliament with 758 users in our dataset. The $gure indicates a good characterization of politicians based on their party a#liations. For example, The Greens party members form a cluster mostly to the left side of the embedding space, re!ecting their shared ideological positions. In stark contrast, the right-wing Alternative for Germany (AfD) party occupies a cluster clearly separated and opposed to the Greens, highlighting the deep ideological divides between these two political factions. Other major parties, such as the CDU/CSU and SPD, are positioned more centrally, with their members distributed across the embedding space in a manner that also corresponds to their relative ideological stances. Further, we can observe an alignment of vocal members of the liberal party FDP with some of the Greens positions (bottom left).
The generated user communities (see Section 4.4) should express di"erent viewpoints within each topic. We do not expect a clean separation of the individual parties given the heterogeneous views present within the larger, more central parties. Instead, for validation purposes, we focus our analysis on parties with relatively polarized and objectively opposing positions, namely The Greens and the AfD. These two parties appear well-separated within the identi$ed communities, enabling the downstream use for narrative separation. Figure 3b exempli$es the six identi$ed communities and their party a#liations composition for topic 23 (asylum, refugees, migration) when considering the two selected parties. Note, however, as implied by the embeddings overview, that these groups additionally include other members from multiple other parties, covering the sizeable political landscape.
As an insightful example for competing narratives, we focus on the discussions surrounding the energy crisis in Germany due to gas supply issues with Russia in 2022, which is captured by Topic 11 (energy, . . . ). From the story graph in Figure 4a, we select the debate over “Gas Storage vs. Price Control Measures”. This story links the discussion of a planned gas price surcharge due to higher purchase costs with the later adopted “gas price cap” by the ruling “tra#c light” (Ampel) coalition. Further references are made to increased storage costs, related taxation, and the excess pro$ts of energy companies.
To delineate the di"erences in the viewpoints of the extracted communities and to identify contrasts, we $rst use sentiment analysis. Figure 4b shows the sentiment distribution within them, with documents classi$ed using XLM-T [35]. Comparing the dominant communities (i.e., most posts in the story), 1 and 5, the latter uses noticeably more negative language, suggesting di"erent framing.
Further, large language models allow us to summarize the community voices in order to demonstrate tangible di"erences in storytelling. The dominant positions taken by both communities are summarized “Community 1 predominantly expresses urgent concern over rising energy costs, particularly focusing on gas prices, which have spurred calls for immediate relief measures such as the removal of the Gas Umlage (surcharge) and implementation of a Gas Price Cap or Break. There is consensus on the necessity for robust governmental intervention to mitigate economic strain caused by these escalating expenses. Many advocate for suspending fiscal constraints like the debt brake to facilitate necessary spending during this crisis, emphasizing that such financial flexibility is essential to provide timely support to households and businesses. Additionally, there’s a strong demand for long-term solutions including transitioning away from fossil fuels towards sustainable energy alternatives, with some suggesting more radical actions like the nationalization of energy companies to secure public interest over corporate profits. Overall, while immediate relief measures are widely supported, there is also recognition of the need for strategic planning and investment in renewable technologies to address future challenges.” “Community 5 predominantly expresses strong opposition to the proposed Gas Surcharge, viewing it as economically burdensome, poorly conceived, and socially unfair. A dominant sentiment is that this measure was hastily implemented and cra!ed under influence from profit-driven entities rather than considering public welfare, highlighting a significant disconnect between government actions and citizens’ interests. There’s widespread support for the decision to halt the Gas Surcharge and replace it with measures like a Gas Price Brake, which are perceived as more equitable solutions aimed at alleviating the financial strain on households while ensuring economic stability without violating fiscal constraints such as the debt brake. The discourse also underscores frustration with governmental ine"iciency and lack of transparency in decision-making processes, advocating for clearer communication and e"ective crisis management strategies that prioritize public interest over political or corporate gains. Overall, there is a call for more prudent and socially responsible policy-making that adequately addresses energy costs while safeguarding both individuals’ financial well-being and broader economic health.” in Figure 5. Here, we prompt the phi-4 model [36] for all posts per community, gaining insights without sifting through the content.
In the ongoing debate over rising energy costs, the two communities $nd common ground in their advocacy for $nancial relief for households and their opposition to the proposed surcharge, but their narratives diverge as they each shed light on di"erent matters. Community 1 supports government intervention to relieve the economy and also mentions long-term strategies such as the transition to renewable energy or, more radically, the nationalization of energy companies and the lifting of the debt brake. In contrast, Community 5 is mainly skeptical about the current government’s approach and questions its intentions, but fails to mention any far-reaching solutions. The negative sentiment observed is also re!ected in the individual posts, which are characterized by strong language. For example, the crisis is attributed a purely political origin and the main players in the government are defamed, while the planned gas surcharge is described as “truly miserable”, unconstitutional and insane.
In addition to unpacking the community positions in the overall context of a story, we want to address the di"erent facets of an event. For this example, we switch to the polarizing debate on migration in Topic 23, where the planned facilitation of the citizenship of migrant workers is discussed.
Table 2 shows the extracted representative tweets per community within an event. The perspectives on work migration and swifter integration are divided into generally positive statements from Communities 0, 3, and 4, positive with limitations from 1 and 2, and general rejection from Community 5. The main similarities are the recognition of the need for more regulated immigration and the desire to address the labor shortage in Germany. However, their approaches di"er, with some advocating for more open and inclusive policies, while others prioritize security and deportation. Di"erent framings are employed to support the narratives. For example, Communities 0 and 2 frame the issue in terms of economic needs, while Communities 1 and 5 focus on security and preventing illegal immigration. Conversely, Communities 3 and 4 highlight the positive impact of the new Residence Opportunities Act.
Dissecting community involvement within each event in a story can also give analysts insights into the narrative framing at a more structural level. For example, we can track which events di"erent narratives try to emphasize by looking at their community distribution and which events communities select to skip altogether. Moreover, analysts can study the timing and order in which competing narratives cover the events in the story to try to uncover internal dynamics and driving factors.
This work introduces an unsupervised framework for identifying and analyzing competing political narratives on social media, focusing on German politicians’ tweets. We uncovered competing narratives by automatically organizing topics and events into coherent stories and leveraging politician embeddings to identify distinct community perspectives. We validate our method through illustrative examples related to the Energy Crisis and Migration Policy, demonstrating the framework’s e"ectiveness in separating competing narratives within highly polarized political discourse. This approach to large-scale social media analysis o"ers valuable insights, underscoring the potential of automatic narrative framing to monitor how political actors promote divergent interpretations of events on social media. These $ndings provide valuable insights for policymakers, social media platforms, and researchers seeking to better understand political discourse, address polarization, and foster more balanced public discussions.
It is essential to highlight that the development of computational approaches for analyzing competing
narratives relies on suitable datasets and evaluation metrics. Existing datasets for narrative extraction
often lack annotations related to competing narratives, making it challenging to train and evaluate
models that can e"ectively distinguish and analyze them [37]. Furthermore, developing robust metrics
to assess the quality and coherence of extracted narratives and the accuracy of framing analysis remains
an open challenge [
We want to thank our former student Ahmad Hamadeh for the thoughtful discussions - his exploratory work laid an excellent foundation for further development.
During the preparation of this work, the author(s) used Claude 3 Haiku and Grammarly in order to: Grammar and spelling check, paraphrase, and reword. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.