Www 2022 | taxoenrich: self supervised taxonomy complemented by Structural Semantics

Article address ：https://arxiv.org/abs/2202.04887

Classification is the basis of many real-world applications in various fields , It can be used as a structural representation of knowledge . In order to deal with more and more new concepts that need to be organized as a taxonomy , Researchers turned to using new concepts to automatically complement existing taxonomies . In this paper, a new classification completion framework is proposed TaxoEnrich, The framework makes effective use of the semantic features and structural information in the existing classification , And provides a better candidate location representation , To improve the performance of classification completion . To be specific ,TaxoEnrich It consists of four parts :(1) Language model based on powerful pre training , The classification context that combines conceptual semantics and classification relations is embedded ;(2) Classification aware sequence encoder , The candidate position representation is learned by encoding the classification structure information ;(3) Query perception brother encoder , According to the importance of candidate brother code to query location matching , Adaptively aggregate candidate sibling codes to enhance candidate position representation ;(4) We use our new candidate location to represent queries that extend existing work - Location matching model . A large number of experiments on four large real data sets from different fields show that ,TaxoEnrich The best performance of all evaluation indexes , And it is much more advanced than the previous methods .