Temporal Narrative Monitoring in Dynamic Information Environments

1. Problem Statement

Understanding information environments (IE) during crisis events is difficult due to the rapid evolution of online discourse. Existing analytical approaches suffer from significant limitations in dynamic settings:

• Classification-based methods: Rely on predefined labels and static categories, failing to capture narratives that emerge, transform, or dissipate unpredictably. They often lack visibility into how meaning changes over time.
• Network-based methods: Focus on interaction structures (e.g., retweets, replies) to map information flow but provide limited insight into the evolution of the semantic content of narratives.
• Temporal Topic Modeling: While methods like Dynamic Topic Models (DTM) or BERTopic track topic shifts, they often produce independent temporal snapshots. They lack mechanisms to link topics across time windows, making it difficult to track a specific narrative as a continuous, evolving entity.

The core challenge is to model narratives not as static labels or latent topics, but as temporally evolving semantic structures that persist and drift within a shared semantic space without requiring prior label specification.

2. Methodology

The authors propose a system-oriented framework grounded in Situational Awareness theory (specifically Levels 1: Perception and 2: Comprehension). The system transforms unstructured social media streams into interpretable, temporally structured representations.

System Architecture & Pipeline

1. Data Ingestion & Embedding:
   • Social media posts are ingested and embedded using a sentence-level transformer model (all-mpnet-base-v2).
   • This creates a shared semantic space where both historical and new data points can be compared using Cosine Similarity.
2. Rolling Temporal Clustering:
   • The system uses HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) to avoid forcing assignments on sparse or unrelated data (treating them as noise).
   • Rolling Windows: Data is processed in rolling temporal windows (e.g., 12-hour windows advancing in 4-hour increments).
   • Temporal Linkage: Clusters are linked across windows based on centroid similarity. If a new window's cluster centroid has a cosine similarity > 0.85 with a previous window's centroid, the narrative is considered persistent.
   • Lifecycle Management:
     • Emergence: New coherent groupings in the embedding space that don't match existing clusters form new narratives.
     • Dissipation: Clusters receiving no assignments in a subsequent window are marked as dissipated.
     • Drift: Shifts in cluster centroids over time measure narrative drift.
3. Representative Post Selection:
   • To manage computational costs and redundancy, the system selects a small set of representative posts per cluster.
   • Selection is based on proximity to the cluster centroid (semantic centrality) with a diversity constraint (minimum pairwise semantic distance) to avoid selecting paraphrased content or reposts.
4. Generative Summarization:
   • Representative posts are passed to a generative language model (e.g., GPT-4.1-mini) to produce cluster-level summaries and thematic labels.
   • These outputs serve as interpretive abstractions for human analysts rather than definitive classifications.

3. Key Contributions

1. Scalable NLP System: A system for detecting and tracking narratives in non-stationary social media streams without predefined labels.
2. Temporal Modeling Approach: A clustering-based method that enables the measurement of narrative emergence, persistence, and drift by maintaining explicit cluster identities across rolling windows.
3. Empirical Evaluation: A rigorous evaluation of the system's behavior during a real-world crisis event (the U.S. capture of Nicolás Maduro in Venezuela, Jan 2026), demonstrating high cluster coherence and revealing heterogeneous narrative lifecycles.

4. Results

The system was tested on 9,914 X (Twitter) posts collected during the 2026 Maduro capture event.

• Cluster Coherence & Accuracy:
  • Stratified human annotation of 1,033 posts revealed that 91% of posts were assigned to appropriate clusters.
  • Centroid Distance was a significant predictor of classification correctness; posts closer to the centroid aligned more strongly with the dominant theme.
  • Larger clusters exhibited greater internal thematic consistency.
• Narrative Drift & Lifecycle:
  • Drift: The median cosine drift between consecutive windows was 0.32. The distribution was right-skewed, indicating that while some narratives remained stable (minimal drift), others underwent rapid restructuring (drift > 0.50).
  • Lifecycle: The environment contained a mix of transient fragments and stable anchors.
    • Median persistence: 2 temporal windows (8 hours).
    • Longest persistence: 22 windows (88 hours), corresponding to mainstream reporting of the event.
• Noise vs. Structure:
  • The system dynamically adapted to discourse structure. Early in the event, a high proportion of content (up to 100% initially) was classified as "noise" (fragmented discourse).
  • As the event progressed, the noise fraction declined (mean 0.49) while the number of detected clusters increased, indicating the consolidation of discourse into coherent narratives.

5. Significance and Implications

• Situational Awareness Support: The framework directly supports Level 1 (Perception) and Level 2 (Comprehension) of situational awareness by filtering noise and organizing dynamic data into evolving narrative structures.
• Cognitive Load Reduction: By transforming raw data streams into persistent, interpretable clusters with generative summaries, the system reduces the cognitive burden on analysts, allowing them to track narrative evolution rather than reconstructing categories from scratch.
• Dynamic Monitoring: Unlike static analysis, this approach treats narrative structure as a system state, enabling the monitoring of continuity, drift, and stability as observable properties.
• Human-in-the-Loop: The system is designed to augment, not replace, human analysts. It surfaces emergent clusters for sense-making rather than imposing rigid taxonomies, making it suitable for complex, non-stationary environments where ground truth is unknown or shifting.

Limitations: The approach requires pre-specification of temporal window sizes and similarity thresholds. It prioritizes high-level structural views over granular qualitative nuances and relies on modular components (embeddings, clustering, LLMs) which introduce computational overhead and sensitivity to upstream model choices.