Publication

Abstract : Novelty detection is a classical one-class classification problem in machine learning that attempts to categorize the ‘known’ data inputs from the ‘unknown’ inputs mainly using unsupervised learning approaches. There are many approaches available for detecting the novelty from the text documents, primarily on the sentence level, with varying degrees of success based on the methodology they adopt. The advancements in deep learning have contributed significantly for developing more effective algorithms for detecting the novelty. This work primarily focuses on finding sentence-level document novelty using latent Dirichlet allocation (LDA), which is one of the globally accepted and extensively used topic modeling algorithms by combining it with auto-encoders. Experiments on benchmarked datasets indicate that the proposed approach outperforms some of the state-of-the-art approaches in novelty detection.