Reasearch On Relation Extraction Based On Distant Supervision

In the information age,a large amount of text information is generated every day.Therefore,information extraction plays an increasingly important role in the field of natural language processing.Relation extraction is an important research topic in information extraction.It is a key step in the process of knowledge graph construction and completion,but the acquisition of training data is usually difficult.A common method is to use the distant supervision method to obtain training data according to the existing knowledge graph.However,this method of obtaining training data has the problem of error labeling.To solve the problem of error labeling in the relation extraction method based on distant supervision,we studied the following aspects:1.Research on distant supervision relation extraction method based on the shortest dependency path.Among the relation extraction tasks on the hand-labeled corpora,the shortest dependency path has a good performance.Therefore,the shortest dependency path is introduced into the relation extraction task based on distant supervision,which is used to eliminate the interference of unrelated words in the sentence.We use the deep learning model to model the shortest dependency path,and multi-instance learning or attention mechanism to get the relation of the entity pairs.2.Research on distant supervision relation extraction method based on feature fusion.In the process of obtaining the shortest dependency path,a large amount of information is removed.In order to avoid mistakes in the relation extraction due to incomplete information,the deep learning model is used to fuse the shortest dependency path with the sentence to obtain the fusion feature.And the relation extraction based on the fusion feature is performed using multiinstance learning or attention mechanism.3.Research on distant supervision relation extraction method based on reinforcement learning.When using multi-instance learning to get the relation of an entity pair,we use an instance to represent a Bag,so there will be incomplete problem of the Bag information.When using the attention mechanism to obtain the relation of an entity pair,since all the instances in the Bag are integrated,the information of the error labeling instance is introduced in the Bag information.Therefore,the instance selector is constructed based on the reinforcement learning to filter the instance of the incorrect annotation.The data filtered by the instance selector is used for the relation extraction model for training to obtain a better relation extraction model.4.Design and implementation of relation extraction system based on distant supervision.The relation extraction system is built based on the web framework for easy access.Through the actual use of the relation extraction system,a more intuitive relation extraction effect is obtained.