Syntactic Representation Learning for Open Information Extraction on Web

Abstract

This paper proposes a representation learning based method to discover new relations between entities from web, which is more general than existing Open Information Extraction(OIE) methods. Given dependency sequences on the expandPath as input, a convolutional neural network(CNN) is adopted to learn the representation layer features of the syntactic dependency patterns which indicate the relations. Experimental results show that compared with the state-of-art OIE methods, the proposed method obviously improves recall without much expense of precision.