Details of Research Outputs

Open information extraction (Open IE), aiming at distilling structured, machine-readable triples from natural language text, plays an important role in various applications, including natural language understanding, knowledge graph construction, etc. Previous supervised Open IE approaches are mostly tailored to extract predicate-argument triples, in which the predicate is usually limited to verb phrases, whereas, the semantic relations expressed within noun phrases are being neglected. However, identifying semantic relation between entities is no trivial task due to the implicit and complex relation expressions. To address the above issue, we present ReadOIE, a framework for coarse-to-fine Open IE via relation oriented reading comprehension, to extract relation-entity triples. In our framework, all entity pairs are extracted to generate structured questions and the input sentence is regarded as the context passage. Semantic relations that best answer the questions are then extracted by comprehending the given context. Moreover, in order to identify the non-existence relations between entities, we design a coarse-to-fine relation extraction approach consisted of an extensive detection module and an intensive extraction module. The extensive detection performs relation existence judgement on a coarse level and intensive extraction identifies the relation on a fine-grained level. Extensive experiments on benchmark datasets demonstrate that ReadOIE outperforms the state-of-the-art baselines.