Abstract
We present a robust topic modeling framework that mitigates overfitting while capturing the evolving nature of discourse. Our approach proposes a dynamic, temporal network-based model that adapts to emerging topics while maintaining semantic stability through alignment with predefined themes. To complement this adaptability, we introduce a static novelty detection method, catering to audiences favoring consistent topic structures. By balancing flexibility and stability, our framework enhances the interpretability and reliability of topic modeling in dynamic environments. Our findings significantly improve the capacity to monitor and analyze novel topics across various applications, including news tracking and social media analysis, ultimately providing a more robust framework for understanding the evolution of textual data over time. We apply our analysis framework on the coverage by brokers of the French snap elections of 2024.
Authors
