Neural Relation Extraction Based On Distant Supervision Approaches

With the advent of the era of big data,artificial intelligence technology has continued to develop and has begun to affect and change human social production and daily life.The Knowledge Graph(KG)is the basis for machines to understand human knowledge,so it has become an important research topic in the field of artificial intelligence.In the real world,human knowledge is complicated and fast to renew.Building the knowledge graph manually cannot guarantee the comprehensiveness and timeliness.Therefore,it is very meaningful to study the technology of automatic construction for Knowledge Graph.Relation extraction is one of the crucial sub-tasks in the construction of Knowledge Graph,which aims to automatically extract the relation between entities from unstructured text.Currently,the mainstream solutions of relation extraction are to train the neural relation extraction model to learn the relation patterns in sentences.Due to the high cost of manual annotation,distant supervision is widely used in the relation extraction task.Distant supervision constructs the training data automatically by aligning the Knowledge Graph with the unsupervised corpora,which can reduce the cost of manual annotation.However,there exist problems of long-tail distribution and mislabeling in the training data constructed by distant supervision,which makes it difficult to train the relation extraction models.Therefore this thesis proposes two novel relation extraction algorithms that can alleviate the problem of insufficient training,thereby improving the performance of relation extraction significantly.The main contributions of this thesis are as follows:· Mine the semantic information of entities from unsupervised corpora: To solve the problem of insufficient training appeared in the existing works due to imbal-anced training data in distant supervised learning.We construct an entity proximity graph from an unsupervised corpora,and further mine the implicit mutual relations between entity pairs.This information cannot be obtained from the training data,so it is important supplementary information for the relation extraction task,and can be applied to further alleviate the problem of insufficient training.· Propose a novel relation extraction algorithm based on the implicit mutual relation: To solve the problem of insufficient training of entity pairs,this thesis utilizes the implicit mutual relations to model the proximity between entity pairs,so that each entity pair can benefit from the training data of other entity pairs.For introducing the implicit mutual relations,an extensible relation extraction model is designed,which can integrate the implicit mutual relations,entity types and training sentences.Experiments on two widely used datasets NYT and GDS show that our proposed method is better than the state-of-the-art models,and it can alleviate the problem of insufficient training existed in distant supervision approaches.· Propose a novel relation extraction algorithm based on the relation prototype:For the problem of insufficient training of relation in distant supervision,this thesis proposes a novel relation extraction algorithm based on the relation prototype.The relation prototype is learned from unsupervised corpora and training data,and can capture the correlation between the relations,so that the knowledge can be transferred from the relations with sufficient training data to the long-tail relations,thereby alleviating the problem of insufficient training.For introducing the relation prototype,this novel algorithm integrates the relation prototype,entity type information and training data information into relation extraction model.Experiments show that this novel relation extraction model significantly outperforms the stateof-the-art models,and it can alleviate the problem of insufficient training.