A bidirectional adversarial neural topic model based on mixed von Mises-Fisher distributions
Topic models serve as a textual analysis tool that automatically mine latent topics or semantic information from textual data.However,existing topic models often assume inappropriate priors and struggle to leverage external semantic knowledge to enhance the quality of topics,resulting in insufficient topic coherence.Targeting these limitations,this paper proposes a bidirectional adversarial neural topic model based on mixed von Mises-Fisher(vMF)distributions.This model performs topic inference through an encoder while introducing external semantic knowledge into the topic modeling process.Specifically,it suggests modeling topics as mixed vMF distributions in the word embedding space within the generator network,and a discriminator network is trained to distinguish between real and fake samples.Experimental results on four public text corpora show that the proposed model achieves higher topic coherence compared to other baseline topic models,and effectively improves the accuracy on text clustering experiments based on extracted topics.
国家科技期刊平台
NSTL
万方数据
