Chapitre D'ouvrage Année : 2024

ANTM: Aligned Neural Topic Models for Exploring Evolving Topics

Hamed Rahimi
Hubert Naacke
Bernd Amann

Résumé

In today’s world, where the amount of textual information generated by humans and machines is rapidly growing, computational methods for organizing, summarizing, and tracking textual information and its evolution are becoming increasingly important. This paper introduces ANTM, an algorithmic family of dynamic topic models that combines novel techniques for discovering evolving topics in large corpora. ANTM preserves the temporal continuity of evolving topics by extracting temporal features from documents with advanced pre-trained large language models and by employing an overlapping sliding window algorithm for sequential aligned document clustering. This clustering method identifies different numbers of topics within each time frame and aligns semantically similar document clusters across time periods. It captures emerging and fading topics over time and allows for a more diverse and coherent representation of evolving topics. We evaluate ANTM against four other dynamic topic models on three datasets and conclude that it outperforms the state-of-the-art approaches in terms of interpretability and diversity. Furthermore, we demonstrate its effectiveness in handling large corpora, while improving the scalability and adaptability of dynamic topic models to different domains.

Fichier non déposé

Dates et versions

hal-04842482 , version 1 (17-12-2024)

Identifiants

Citer

Hamed Rahimi, Hubert Naacke, Camelia Constantin, Bernd Amann. ANTM: Aligned Neural Topic Models for Exploring Evolving Topics. Abdelkader Hameurlain; A Min Tjoa; Reza Akbarinia; Angela Bonifati. Transactions on Large-Scale Data- and Knowledge-Centered Systems LVI: Special Issue on Data Management - Principles, Technologies, and Applications, 14790, Springer Berlin Heidelberg, pp.76-97, 2024, Lecture Notes in Computer Science (TLDKS), 978-3-662-69603-3. ⟨10.1007/978-3-662-69603-3_3⟩. ⟨hal-04842482⟩
188 Consultations
0 Téléchargements

Altmetric

Partager

  • More