Researchers have developed BERTilda, a framework for tracking topic lifecycles in longitudinal text streams. This involves identifying how topics split into subtopics or merge into broader narratives over time. The system uses an embedding-based topic model to discover topics independently in each window and then constructs a temporal graph linking topics across adjacent windows. Two signals support the links: semantic similarity between topic representations and a bidirectional coverage signal that estimates document outflow and inflow. The framework labels continuations, splits, merges, disappearances, and unclear transitions using graph-based rules. --- Why it matters: BERTilda is relevant to researchers in AI because it provides an explainable way to track topic lifecycles in text streams, which can be useful for applications such as information retrieval, sentiment analysis, and event detection. The framework's ability to identify splits, merges, and disappearances can also help improve the accuracy of dynamic topic models. Source: https://arxiv.org/abs/2608.18101