UMP-MG: A Uni-directed Message-Passing Multi-label Generation Model for Hierarchical Text Classification

Abstract Hierarchical Text Classification (HTC) is a formidable task which involves classifying textual descriptions into taxonomic hierarchy. Existing methods, however, have difficulty in adequately modeling the hierarchical label structures, because they tend to focus on employing graph embedding methods encode structure while disregarding fact that HTC labels are rooted tree structure. This significant because, unlike graph, inherently has directive ordains information flow from one node another—a critical factor when applying task. But structure, message-passing undirected, will lead imbalance of message transmission between nodes applied HTC. To this end, we propose unidirectional multi-label generation model for HTC, referred as UMP-MG. Instead viewing classification problem previous done, novel approach conceptualizes it sequence task, introducing prior during decoding process. further enables blocking direction ensure method better suited and thus resulted enhanced representation. Results obtained through experimentation both public WOS dataset an E-commerce user intent demonstrate our proposed can achieve superlative results.