Neural topic modeling on hyperspheres: Spherical representation learning with von Mises-Fisher mixtures

Neural topic models (NTMs) based on variational autoencoders (VAEs) have emerged as a scalable and flexible alternative to classical probabilistic models for uncovering latent thematic structures in text corpora. However, most existing NTMs either overlook the geometric structure of word embeddings or rely on Euclidean priors that are poorly aligned with L₂-normalized semantic spaces. Moreover, VAE-based NTMs often suffer from Kullback-Leibler (KL) divergence collapse and fail to capture directional semantics critical for coherent topic discovery. In this work, we propose the von Mises-Fisher Mixture Neural Variational Topic Model (vMNVTM), a hyperspherical neural topic modeling framework that explicitly models both document representations and topic-word relationship using the von Mises-Fisher (vMF) distribution. Our approach leverages the hyperspherical nature of vMF distribution to capture directional alignment and angular dispersion in the latent space, and enables more coherent and interpretable topic representations in word embedding space. We further introduce a vMF-aware embedding clustering (vEC) loss to align topics with semantically meaningful regions in the word embedding space and a temperature-controlled concentration mechanism to enhance topic separability. Extensive experiments on benchmark datasets demonstrate that vMNVTM outperforms state-of-the-art NTMs in terms of topic coherence, diversity, and interpretability, highlighting the importance of directional modeling in neural topic inference. The source code of our model is publicly accessible at https://github.com/gvbdxgvc/vMNVTM.