Topic detection based on BERT and seed LDA clustering model
Jing Jing Wu, Bicheng Li, Qilong Liu · 2023
Aiming at the problem that the LDA model is not effective for short text topic extraction, this paper proposes a topic detection method based on BERT and seed LDA clustering model. Firstly, the seed LDA model (sLDA) is designed for optimize the LDA model. A seed word set is introduced for the LDA model. The words generated under a certain probability will be generated from the seed word set, thus guiding topic generation; Secondly, sLDA and BERT were combined to construct the B-sLDA model. The sLDA model was used to obtain topic features, and the BERT model was used to obtain text features. The result of fusing two features were input into the K-means clustering model to obtain topic clusters. Finally, to complete the topic detection , each cluster was input into the TF-IDF model, and the first 10 output words were used as the topic of the corresponding cluster. In this paper, the official dataset 20 newsgroups and the real public opinion dataset about "Shanghai COVID-19" obtained by crawler were selected for the experiment. The experimental results show that sLDA is superior to the original LDA model in terms of perplexity and coherence, which proves that the seed word set has a certain guiding effect on the LDA model. B-sLDA clustering model has significant improvement in Silhouette Coefficient and Calinski-Harabasz.