Knowledge Base Evaluation for Semantic Knowledge Discovery

Abstract

Semantic knowledge discovery has traditionally been evaluated at the text level. For example, evaluations such as MUC and ACE evaluate the information extraction of particular types of semantic roles and relations primarily at the mention level. We suggest that evaluating at the level of a knowledge base (KB) extracted from the text has significant advantages over evaluation at the text level. By knowledge base, we mean the combination of a database, a descriptive schema for the contents of the database, a collection of background knowledge, and an inference capability.