Proceedings of the First International Workshop on Formalisms and Methodology for Learning by Reading

A Hybrid Approach to Unsupervised Relation Discovery Based on Linguistic Analysis and Semantic Typing


This paper describes a hybrid approach for unsupervised and unrestricted relation discovery between entities using output from linguistic analysis and semantic typing information from a knowledge base. We use Factz (encoded as subject, predicate and object triples) produced by Powerset as a result of linguistic analysis. A particular relation may be expressed in a variety of ways in text and hence have multiple facts associated with it. We present an unsupervised approach for collapsing multiple facts which represent the same kind of semantic relation between entities. Then a label is selected for the relation based on the input facts and entropy based label ranking of context words. Finally, we demonstrate relation discovery between entities at different levels of abstraction by leveraging semantic typing information from a knowledge base.

  • 269656 bytes

information extraction, learning, natural language processing, natural language processing, powerset


Association for Computational Linguistics

Downloads: 1917 downloads

UMBC ebiquity