Deep Representation Learning for Effective Clustering of Short Persian Texts
Mahdi Molaei, Mohammad‐Reza Feizi‐Derakhshi, Ali-Akbar Rasooly, Cina Motamed · Engineering Technology & Applied Science Research · 2025
Short text clustering poses several challenges because of the limited contextual information available, especially for low-resource languages such as Persian. This study proposes a novel deep clustering architecture that consists of an RNN-based autoencoder to learn the latent representation of the text in preserving the rich structural features. This architecture involves a second network, the Representation network, to maximize the existing distance between clusters, minimize the overall cluster overlapping, and improve clustering in the latent space. The two-phase training approach first involved training using autoencoder reconstruction loss and then jointly optimizing for improved cluster separation. Experiments with different embedding types were carried out, and the evaluation results showed that the proposed method outperformed previous approaches. The proposed model provides an impactful advancement in representation learning and training for the short-text domain.