Source-linked AI summary
Automated Identification of Competing Narratives in Political Discourse on Social Media
Sergej Wildemann, Erick Elejalde
TL;DR
Competing political narratives on social media are difficult to identify and analyze because they reflect divergent ideological perspectives across political communities. This paper introduces an unsupervised pipeline that organizes topics and events into stories and extracts community viewpoints, revealing divergent interpretations in polarized German political discourse. The framework is demonstrated through energy-crisis and migration-policy examples, while the evaluation remains primarily qualitative.
Problem
The paper addresses the challenge of identifying and analyzing competing narratives that emerge across ideological groups in political social media discourse.
Method
An unsupervised multi-stage pipeline models topics, detects and links events into coherent stories, and uses user communities to extract distinct narrative viewpoints.
Results
The framework separates divergent narratives by organizing topics and events into stories and distinguishing politically opposed community perspectives in energy-crisis and migration-policy cases.
Takeaways & Limitations
Automatic narrative framing can help monitor how political actors promote divergent interpretations of events in large-scale social media discourse.
Takeaways & Limitations
The study relies primarily on qualitative evaluation because suitable competing-narrative datasets and robust evaluation metrics remain limited.
Abstract
from arXiv · showhide
Social media platforms have become central to shaping political discourse, serving as arenas where narratives form and evolve, influencing public opinion. Identifying and analyzing these narratives, particularly when they compete across different 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 conflicts surrounding trending political topics. Two case studies on polarizing political issues demonstrate the efficacy of the methodology, showcasing its ability to uncover and analyze divergent viewpoints. The findings contribute to the broader understanding of how narratives propagate within the digital public sphere and offer insights for policymakers, social media platforms, and researchers interested in monitoring political discourse.
1. Introduction
Political narratives on social media shape public opinion and compete across ideological groups. The paper addresses this challenge by proposing computational methods to identify and analyze competing narratives among German politicians on Twitter.
- Social media enables politicians to disseminate narratives directly to citizens, where competing ideological framings can shape political discourse and public opinion.
- The paper proposes an NLP- and clustering-based framework for automatically identifying and analyzing competing narratives in German politicians’ Twitter discourse.
- The framework targets narratives that develop across multiple tweets and potentially involve collaborative contributions from multiple sources.
- Analyzing competing narratives can support media literacy, critical thinking, constructive dialogue, misinformation efforts, and more informed engagement with media.
2. Background and Related Work
Related work treats narratives as connected event sequences and develops computational methods to extract their elements, frames, and competing perspectives. These methods support scalable analysis of how narratives emerge, differ, and evolve over time.
- Narratives are connected successions of events that may combine reality and fiction, function socially, and contain competing storylines.
- Computational narrative understanding extracts actors, actions, events, settings, and relationships, using temporal, causal, semantic, and coherence-based representations.
- Computational framing analysis detects linguistic and rhetorical devices, including word choice, metaphors, and evidence presentation, that influence interpretation.
- Studying competing narratives can help monitor community interests, counter propaganda, and model misinformation by tracking narrative evolution and its drivers.
- Competing narratives can be distinguished and tracked through differences in actors, actions, framing devices, language, sentiment, and message prominence over time.
3. Dataset
The dataset comprises German politicians and their Twitter activity identified through Wikidata and collected through Twitter’s API. Party affiliations are resolved using available current or listed affiliation information, and the tweets are analyzed primarily in their original language.
- Wikidata identifies German politicians by occupation and citizenship, while party affiliations are obtained from the member-of-political-party property.
- When politicians have multiple affiliations, the dataset uses a qualified current party when available or otherwise the first listed party without an end date.
- Twitter’s API supplied 189,850 tweets from 786 accounts collected between January 1, 2022, and June 24, 2023, excluding retweets, replies, and quotes.
- Tweets were translated into English for presentation, but analyses were conducted on the original-language content.
4. Methodology
The methodology uses an unsupervised, multi-stage pipeline that identifies topics, detects temporally coherent events, links them into stories, and separates user viewpoints into competing narratives. It evaluates narrative separation without using political affiliations during extraction, while topic and story examples illustrate the resulting structure.
- Pipeline overview: The pipeline identifies topics, detects events, links events into coherent stories, and extracts different user viewpoints as competing narratives.The stages progress from broad discussion strands to event-based stories and then viewpoint separation.
- User grouping: Political affiliations are disregarded during narrative extraction and reserved for evaluation, while global user grouping is used to distinguish viewpoints within stories.The grouping strategy aims to include users who are less active in a particular topic by using their broader activity trace.
- 4.1. Topic Modeling: Topic modeling with BERTopic groups semantically similar tweets, producing 57 topics from 108,063 non-outlier documents after 43% of tweets are excluded.The embedding-based approach is selected for the short and noisy nature of tweets.
- 4.2. Event Detection: Events group tweets discussing the same issue in close temporal proximity, combining semantic similarity with temporal proximity to separate distinct discussions.The method uses affinity propagation over an affinity matrix combining cosine and temporal similarity; tweets are the analysis unit rather than individual sentences.
- 4.3. Story Generation: Event linking builds stories with an event-similarity graph that penalizes temporal distance and retains each event’s highest-weighted outgoing edge to focus on the main storyline.The graph is directed chronologically, and similarity combines event embeddings with a temporal penalty.
- 4.3. Story Generation: For Topic 11, the story extraction process produces 12 stories from 155 events, with each story represented as a sequence of temporally bounded events.Figure 2 represents stories as rows and events as bars whose widths correspond to start and end dates.
5. Results
The framework is evaluated on German politicians’ tweets by separating user communities and illustrating how opposing groups frame shared political stories differently.
- Results: The framework separates politically opposed communities and reveals divergent interpretations of the same events in polarized discourse.Validation focuses on user-position coherence and two illustrative cases involving energy and migration.
- 5.1. Evaluation of Uncovered User Communities: Global user embeddings place Greens and AfD politicians in clearly separated clusters, while CDU/CSU and SPD users occupy more central, heterogeneous positions.The analysis uses 758 politicians from major parties represented in the German parliament.
- 5.2. Narratives Around Germany’s 2022 Energy Crisis: In the energy-crisis story, Communities 1 and 5 agree on household relief and opposing the surcharge but diverge in their broader framing and proposed solutions.Community 1 supports intervention and long-term renewable or structural measures, whereas Community 5 is skeptical of the government and offers fewer far-reaching solutions.
- 5.3. Law on Better Residence Opportunities for Migrants: In the migration-law event, communities range from generally positive views to rejection, sharing labor-shortage and regulation concerns while differing over inclusion, security, and deportation.Communities 0 and 2 emphasize economic needs; Communities 1 and 5 emphasize security and illegal immigration; Communities 3 and 4 highlight the Residence Opportunities Act.
- Results: Tracking community participation and event order can expose which events competing narratives emphasize, omit, or address at different times.This offers a structural view of narrative dynamics within stories.
6. Conclusions and Future Work
The paper presents an unsupervised framework for extracting competing political narratives from German politicians’ tweets and demonstrates its use on polarized energy and migration discussions. It concludes that qualitative results are promising, while stronger datasets and evaluation metrics remain necessary.
- 6. Conclusions and Future Work: The framework organizes topics and events into stories and uses politician embeddings to identify community perspectives that separate competing narratives.The authors demonstrate this separation through Energy Crisis and Migration Policy examples.
- 6. Conclusions and Future Work: The findings offer insights for policymakers, platforms, and researchers studying political discourse, polarization, and more balanced public discussions.This is the paper’s stated scope of practical relevance.
- 6. Conclusions and Future Work: The evaluation is primarily qualitative because datasets lack competing-narrative annotations and robust metrics for narrative quality, coherence, and framing accuracy remain open challenges.The authors propose correlating event-to-story linking with user streams to strengthen coherence measurement.
Declaration on Generative AI
The authors used Claude 3 Haiku and Grammarly during preparation for grammar, spelling, paraphrasing, and rewording, then reviewed and edited the content.
- Declaration on Generative AI: Claude 3 Haiku and Grammarly were used for grammar, spelling, paraphrasing, and rewording during manuscript preparation.The authors state that they reviewed and edited the resulting content.