BERTilda: Explainable Topic Lifecycle Tracking with Split/Merge Detection via Similarity-and-Flow Temporal Graphs
BERTilda is an explainable framework for tracking topic lifecycles in longitudinal text streams, detecting structural reorganizations like splits, merges, continuations, and disappearances It combines embedding-based topic modeling per time window with a temporal graph linking topics via semantic similarity and bidirectional document coverage signals (inflow/outflow) Evaluated on U.S. congressional tweets and historical speech datasets, achieving up to 87% majority agreement with human annotator
Analysis
TL;DR
- BERTilda is an explainable framework for tracking topic lifecycles in longitudinal text streams, detecting structural reorganizations like splits, merges, continuations, and disappearances
- It combines embedding-based topic modeling per time window with a temporal graph linking topics via semantic similarity and bidirectional document coverage signals (inflow/outflow)
- Evaluated on U.S. congressional tweets and historical speech datasets, achieving up to 87% majority agreement with human annotators on lifecycle labels
- Outperforms similarity-only and forward-only baselines, with particularly strong performance in detecting topic disappearances
- Addresses the gap between dynamic topic models (which assume smooth drift) and snapshot models (which lack temporal correspondence)
Why It Matters
BERTilda provides a practical, explainable approach to tracking how topics evolve, split, and merge over time in real-world text streams—a capability increasingly critical for monitoring political discourse, social media trends, and emerging narratives. Its dual-signal graph construction (semantic similarity plus document flow) offers a more robust alternative to methods relying on a single temporal signal, making it valuable for researchers and practitioners building topic monitoring systems.
Technical Details
- Two-stage architecture: Topics are discovered independently in each time window using an embedding-based topic model, then linked across adjacent windows via a temporal topic graph
- Dual linking signals: (i) semantic similarity between topic representations and (ii) a bidirectional coverage signal estimating document outflow (where a topic's documents go) and inflow (where they come from) through cross-window tweet-to-topic attribution
- Graph-based lifecycle labeling: Rules applied to the temporal graph classify transitions as continuations, splits, merges, disappearances, or unclear transitions
- Evaluation: Tested on political corpora (U.S. congressional tweets, historical speeches) with topic-quality and temporal-stability diagnostics; lifecycle labels validated against a gold-standard subset annotated by three independent annotators, achieving up to 87% majority agreement and the highest macro-average agreement among compared methods
Industry Insight
- The framework's explainability through graph-based rules makes it suitable for high-stakes applications like political analysis and media monitoring where understanding topic evolution is as important as detecting it
- The strong disappearance detection capability addresses a commonly overlooked aspect of topic modeling—topics that fade rather than drift—enabling more complete narrative tracking
- The dual-signal approach (similarity + document flow) could be adapted to other domains beyond politics, such as tracking technology trends, scientific literature evolution, or brand perception shifts over time
Disclaimer: The above content is generated by AI and is for reference only.