Demonyms and Compound Relational Nouns in Nominal Open IE

Extracting open relational tuples that are mediated by nouns (instead of verbs) is important since titles and entity attributes are often expressed nominally. While appositives and possessives are easy to handle, a difficult and important class of nominal extractions requires interpreting compound noun phrases (e.g., “Google CEO Larry Page”). We substantially improve the quality of Open IE from compound noun phrases by focusing on phenomena like demonyms and compound relational nouns. We release RELNOUN 2.2, which obtains 3.5 times yield with over 15 point improvement in precision compared to RELNOUN 1.1, a publicly available nominal Open IE system.