Competing Narratives in Political Tweets: What Unsupervised Event Linking Adds to Topic Models

Why topic models stop short of competing narratives

Topic modeling has long been the default way to map political discourse on social media. The anchor paper's own BERTopic run over 189,850 German politicians' tweets produced 57 topics covering the war in Ukraine, COVID-19 vaccination, energy prices, migration, and more, with 43% of documents left as outliers [1]. That output is useful for orientation, but a topic like 'energy, gas, renewable, electricity, price' bundles together months of debate without indicating which events belong together or how different actors framed them [1]. The same limitation appears in adjacent work: bibliometric and LDA approaches to TikTok research discourse treat topics as the endpoint of analysis rather than as raw material for narrative reconstruction [2]. Earlier event-detection work went further by detecting spikes and linking them to news articles [6], or by tracking named-entity co-occurrence networks to locate important events in news streams [7], but these pipelines were designed for situational awareness and news summarization, not for separating ideologically opposed tellings of the same event.

The anchor paper's move is to treat topics as a first pass and then build narrative structure on top. Events are defined as sets of documents discussing the same issue in close temporal proximity, clustered with affinity propagation over a combined semantic and temporal similarity [1]. For the energy topic alone, this yielded 155 events [1]. Events are then linked into stories via an event similarity graph and community detection, producing 12 stories for that topic [1]. This is the step that topic models cannot perform: it converts a flat keyword cluster into a sequence of related episodes that can be compared across groups.

Separating narratives by stance rather than by party label

The framework deliberately ignores party affiliation during extraction, using it only for evaluation [1]. Instead, each user is represented by a global embedding built from their averaged topic-level positions relative to topic centroids, and users are clustered with HDBSCAN into stance communities [1]. A 2D projection of these embeddings shows Green party members clustering on one side and Alternative for Germany members clearly separated and opposed, with CDU/CSU and SPD distributed more centrally [1]. This matters because it means the method can, in principle, be applied where party labels are unavailable, which is most social media contexts. The validation is partial: the authors focus on The Greens and AfD because those parties have relatively polarized and objectively opposing positions, and they explicitly note that they do not expect clean separation of individual parties given heterogeneous views within larger centrist parties [1].

Once communities are identified, each story's posts are split by community, and the structure formed by one community's contribution to one story is treated as a competing narrative [1]. In the energy story on gas storage versus price control measures, communities 1 and 5 both opposed the gas surcharge and supported relief for households, but community 1 emphasized government intervention, renewable transition, and even nationalization of energy companies, while community 5 emphasized government inefficiency, lack of transparency, and a disconnect from citizens' interests [1]. Sentiment analysis with XLM-T showed community 5 using noticeably more negative language [1]. This is the kind of divergence that a topic model would flatten into a single 'energy' cluster.

What event linking adds beyond detection and framing analysis

Prior Twitter event detection work established that events can be detected and linked to news. Sem-TED used Sentence Transformer embeddings, UMAP, and HDBSCAN to detect popular topics and a single-layer perceptron to judge relevance to news stories, reporting an average 8% improvement over entity-based methods on a small sample of two datasets [3]. Thapen et al. filtered bio-surveillance algorithms to detect localised spikes in Twitter activity and linked events to news by identifying terms more frequent in event tweets than in a baseline [6]. Moutidis and Williams built time-stamped entity co-occurrence networks and used change-point detection plus community detection on KeyGraphs to characterize events in news streams [7]. These are detection and summarization pipelines; none of them separates competing political narratives within the same event.

The anchor paper's event-linking step is what enables that separation. By grouping events into stories and then splitting stories by user community, it produces something closer to a comparative framing analysis at scale. This connects to a different literature: Kiyak et al. used retweet network analysis and discourse analysis on German-language X data about Syrian and Ukrainian refugees, finding a polarized crowd structure, rising anti-refugee community size, declining neutral opinion leaders, and increasing distance between camps from 2015 to 2022 [4]. Their approach is network-based and manual-interpretive; the anchor paper's is embedding-based and unsupervised. The two agree on the underlying phenomenon of polarized German-language political Twitter, but they measure it differently and neither directly validates the other's narrative boundaries.

Migration case shows framing differences within a single event

The migration case study moves from story-level to event-level analysis. Within the story 'Modern immigration law and civil equality in Germany' in the asylum and migration topic, the anchor paper shows representative tweets per community for a single event [1]. Communities 0, 3, and 4 expressed generally positive views on work migration and integration; communities 1 and 2 were positive with limitations; community 5 rejected the policy outright [1]. Communities 0 and 2 framed the issue in terms of economic needs and labor shortage, while communities 1 and 5 focused on security and preventing illegal immigration [1]. This is a concrete demonstration that the same event can carry different framings that are visible only when posts are split by stance community.

The authors also suggest that this event-level community distribution can reveal which events different narratives emphasize or skip, and in what order competing narratives cover events in a story [1]. That is a structural claim about narrative strategy rather than a validated finding; the paper presents it as an analytical possibility, not as a tested result. The migration case is also where the evidence boundary is clearest: it is one event within one story within one topic in one country's politician tweets.

What the framework does not yet establish

The evaluation is primarily qualitative. The authors state that existing datasets for narrative extraction often lack annotations related to competing narratives, making it difficult to train and evaluate models that distinguish them, and that robust metrics for narrative quality and framing accuracy remain an open challenge [1]. They rely on illustrative case studies and party-affiliation validation rather than a quantitative benchmark [1]. The user-community validation is also partial: the authors deliberately focus on The Greens and AfD because those parties are well-separated, and they acknowledge that centrist parties are distributed across the embedding space [1]. This means the method's ability to separate narratives within less polarized or internally heterogeneous groups is not demonstrated.

The dataset is 189,850 tweets from 786 German politician accounts, collected January 2022 to June 2023, excluding retweets, replies, and quotes [1]. Conclusions are therefore bounded to German politicians' tweets on two polarizing issues, and cannot be generalized to other languages, platforms, or election contexts without further validation. The broader literature reinforces this caution: cross-lingual stance and frame detection remains harder than topic detection, with frame detection especially difficult due to conceptual overlaps, and most stance and discourse corpora remain English-centric and biased toward the Global North [8]. Unsupervised counterfactual story rewriting work also shows that coherence and minimal-edit trade-offs in narrative generation are non-trivial, which is a reminder that automated story construction is not a solved problem [5]. Finally, work on digital news captions shows that intertextuality and cohesion are what let condensed texts connect to broader narratives [9]; the anchor paper's stories are built from tweets that lack those explicit cohesive devices, which may limit how coherent the extracted stories actually are.

About These Sources

This research page is built on 9 peer-reviewed studies — published from 2015 to 2026, 5 from 2024 or later, collectively cited 67 times — selected as the most relevant from 9 studies that passed quality screening, drawn from 72 papers retrieved from a database of over 500 million.

Sources used in this answer