Hybrid neural tagging model for open relation extraction

Open Relation Extraction (ORE) task remains a challenge to obtain semantic representation by discovering arbitrary relations from the unstructured text. Conventional methods heavily depend on feature engineering or syntactic parsing, which are inefficient error-cascading. Recently, leveraging supervised deep learning address ORE is promising way. However, there two main challenges: (1) The lack of enough labeled corpus support training; (2) exploration specific neural architecture that adapts characteristics open relation extracting. In this paper, we build large-scale, high-quality training in fully automated And wedesign tagging scheme assist transforming into sequence processing. Furthermore, propose hybrid network model (HNN4ORT) for tagging. employs Ordered Neurons LSTM encode potential information capture associations among arguments and relations. It also emerges novel Dual Aware Mechanism, including Local-aware Attention Global-aware Convolution. dual awarenesses complement each other. Takes sentence-level semantics as global perspective, at same time, implements salient local features achieve sparse annotation. Experiment results various testing sets show our achieves state-of-the-art performance compared toconventional other models.