KEML: A Knowledge-Enriched Meta-Learning Framework for Lexical Relation Classification

Lexical relations describe how concepts are semantically related, in the form of relation triples. The accurate prediction lexical between is challenging, due to sparsity patterns indicating existence such relations. We propose Knowledge-Enriched Meta-Learning (KEML) framework address classification. In KEML, LKB-BERT (Lexical Knowledge Base-BERT) model first presented learn concept representations from text corpora, with rich knowledge injected by distant supervision. A probabilistic distribution auxiliary tasks defined increase model's ability recognize different types further a neural classifier integrated special recognition cells, order combine meta-learning over task and supervised learning for LRC. Experiments multiple datasets show KEML outperforms state-of-the-art methods.