@article{Meng2026, 
author = {Lingwu Meng and Guangchi Liu and Junzhou Luo},
title = {Distributed cross-modal enhanced bertopic model for large-scale event evolution analysis},
year = {2026},
journal = {Big Data Mining and Analytics},
keywords = {Cross Modal, Enhanced BERTopic, Interaction influence, Dynamic Topic Model, Distributed computing},
url = {https://www.sciopen.com/article/10.26599/BDMA.2026.9020028},
doi = {10.26599/BDMA.2026.9020028},
abstract = {The spatiotemporal evolution analysis of social media event themes provides important support for network public opinion monitoring and emergency management. However, the cross modal, interactive, and time sensitive features of social media data pose significant challenges to existing topic mining methods. Traditional methods have shortcomings in cross modal semantic representation, information influence modeling, time decay mechanisms, and large-scale data processing, which limit the accuracy and stability of topic evolution analysis. Therefore, this article proposes a distributed cross modal dynamic topic evolution analysis framework, Cross-Modal Enhanced BERTopic (CME-BERTopic). This framework is based on Chinese-CLIP with Self Attention and Cross Attention (CN-CLIP-SA-CA) to extract deep cross modal semantic features, and enhances the HDBSCAN clustering algorithm with information interaction influence to optimize event topic representation. Meanwhile, Top-K Compression with Influence Decay Dynamic Topic Model (TCI-DTM) was designed to depict the dynamic evolution of topics over time. Finally, integrate each module into a distributed computing framework to support efficient processing of large-scale social media data. The experimental results based on the Flickr and DeepSeek datasets show that Distribute CME-BERTopic (DCME-BERTopic) and TCI-BERTopic is significantly better than traditional BERTopic in topic recognition and evolution tracking, with stronger adaptability, robustness, and computational efficiency.}
}