Span Based Open Information Extraction

Open information extraction (Open IE) is achallenging task especially due to its brittledata basis. Most of Open IE systems have to betrained on automatically built corpus and eval-uated on inaccurate test set. In this work, wefirst alleviate this difficulty from both sides oftraining and test sets. For the former, we pro-pose an improved model design to more suffi-ciently exploit training dataset. For the latter,we present our accurately re-annotated bench-mark test set (Re-OIE6) according to a seriesof linguistic observation and analysis. Then,we introduce a span model instead of previousadopted sequence labeling formulization forn-ary Open IE. Our newly introduced modelachieves new state-of-the-art performance onboth benchmark evaluation datasets.