Variational Deep Semantic Hashing for Text Documents

• DOI: 10.48550/arxiv.1708.03436
• Publication Date: 2017-01-01
• Authors: Suthee Chaidaroon; Yi Fang
• Keywords: Autoencoder; Interpretability; Generative model

As the amount of textual data has been rapidly increasing over past decade, efficient similarity search methods have become a crucial component large-scale information retrieval systems. A popular strategy is to represent original samples by compact binary codes through hashing. spectrum machine learning utilized, but they often lack expressiveness and flexibility in modeling learn effective representations. The recent advances deep wide range applications demonstrated its capability robust powerful feature representations for complex data. Especially, generative models naturally combine probabilistic with high capacity neural networks, which very suitable text modeling. However, little work leveraged progress In this paper, we propose series novel document first proposed model unsupervised while second one supervised utilizing labels/tags third further considers document-specific factors that affect generation words. formulation provides principled framework extension, uncertainty estimation, simulation, interpretability. Based on variational inference reparameterization, can be interpreted as encoder-decoder networks thus are capable nonlinear distributed documents. We conduct comprehensive set experiments four public testbeds. experimental results effectiveness