A Language Model for Extracting Implicit Relations

Open Information Extraction has shown promise of overcoming a knowledge engineering bottleneck, but has a fundamental limitation. It is unable to extract implicit relations, where the sentence lacks an explicit relation phrase. We present IMPLIE (Implicit relation Information Extraction) that uses an open-domain syntactic language model and user-supplied semantic taggers to overcome this limitation. IMPLIE can extract an implicit has nationality relation, has job title, and has city relation from “French journalist Paul Legall reported ... at Athens International Airport.”. Formal evaluations of IMPLIE show high precision, over 0.90 for nationality and job title on newswire text. IMPLIE nearly doubles recall for 2013 KBP Slot Filling queries and more than doubles it for 2014 when combined with an Open IE-based system, maintaining precision of 0.58 and 0.59 respectively.