DNA-HINT Domain-novelty Aware Hierarchical Introspection for Hierarchical Novelty Detection

Anonymous

Paper Link: https://openreview.net/forum?id=gJ5BiQ0R_th

Paper Type: Long paper (up to eight pages of content + unlimited references and appendices)

Abstract: Deep neural networks have achieved impressive performance for text classification that recognizes a predefined set of classes. However, recognizing texts of novel classes unseen during training is not well explored. It is desirable for large-scale text datasets to augment a function of detecting the novelty of a newly-joined text, especially in practical application scenarios such as an e-commerce system. We aim to achieve a hierarchical novelty detection that predicts the closest known class in the taxonomy for a text of a novel class. Furthermore, existing approaches typically encounter issues, such as (i) the inconsistency problem that the predictions in any pair of parent-child nodes are not successive; (ii) the blocking problem that the prediction at a certain level is not confident and unable to be passed downward in the taxonomy; (iii) the overconfidence problem of a softmax classifier that predicts high confidence regardless of whether a text is a known or novel class. In this paper, we propose a novel model, Domain-Novelty Aware Hierarchical Introspection (DNA-HINT), to achieve the goal without those problematic issues. Extensive experiments conducted on four benchmark datasets show that DNA-HINT is effective particularly for deep levels that are often considered in realistic scenarios.