Chinese Triple Extraction Based on BERT Model

Information extraction (IE) plays a crucial role in natural language processing, which extracts structured facts like entities, attributes, relations and events from unstructured text. The results of information extraction can be applied in many fields including information retrieval, intelligent QA system, to name a few. We define a pair of entities and their relation from a sentence as a triple. Different from most relation extraction tasks, which only extract one relation from a sentence of known entities, we achieved that extracting both relation and entities(a triple, as defined above), from a plain sentence. Until now, there are so many methods proposed to solve information extraction problem and deep learning has made great progress last several years. Among the field of deep learning, the pretrained model BERT has achieved greatly successful results in a lot of NLP tasks. So we divide our triple extraction task into two sub-tasks, relation classification and entity tagging, and design two models based on BERT for these two sub-tasks, including a CNN-BERT and a Simple BERT. We experimented our models on DuIE Chinese dataset and achieved excellent results.