Relation Extraction Methods With Knowledge Enhancement

With the continuous development of the Internet,network information is showing an explosive growth trend,and a lot of useful information is contained in it.Relying on manual filtering and editing of knowledge can no longer meet the needs of processing massive text data.How to automatically extract structured information from massive text and improve the efficiency of screening and acquisition of information is an urgent problem to be solved at present.Entity relation extraction,which is an effective solution to this problem,aims to automatically identify entities and the hidden semantic relations between entities in unstructured text data.Deep learning models have shown good performance in entity relation extraction tasks.However,the current entity relation extraction models mostly use pipeline methods or simple joint extraction models,ignoring the clear and necessary support between the two subtasks of entity recognition and relation extraction.Meanwhile,entity relation extraction has the characteristics of less sample data and high labeling cost in practical applications.The relation extraction model tends to struggle when the supervised training data is few.To tackle above limitations,this thesis studies entity relation extraction from the perspective of knowledge enhancement,explores two effective entity relation extraction models,validates the proposed method through experiments,and explores the advantages and disadvantages of the models.The main contributions of this thesis include:· Propose a novel contrastive student-teacher learning framework for joint extraction of entities and relations.In view of the problem that the joint model is insufficient in modeling the potential interaction features between the subtasks of entity recognition and relation extraction,this thesis defines the privileged information in the relation extraction task,and proposes a contrastive teacher-student learning framework that encourages the student model to fuse the knowledge of two expert teacher models,enhancing the latent features needed by the model to generate specific entity-relation triples.Experimental results on benchmark datasets such as COLIEE,ADE,and SciERC demonstrate the effectiveness of the proposed algorithm.· Propose a few-shot relation extraction model with automatically generated prompts.In order to deal with the challenges brought by insufficient training data to the relation extraction model in real scenarios,this thesis uses the T5 model to automatically generate multiple prompt templates,converts the relation extraction task into a prediction task based on prompt learning,and proposes a template regularization network.Use prompt information to stimulate the latent knowledge in the pre-trained language model and improve the effect of small sample relation extraction.Experiments on Few Rel and NYT-25 datasets show that the proposed model is significantly better than the baseline model.· Develop an online relation extraction system.Based on the current needs of Internet users to obtain information efficiently,this thesis implements a prototype system of relation extraction through modular packaging of the aforementioned algorithms.The system applies the knowledge enhancement scheme to relation extraction,which not only achieves good extraction results,but also presents structured textual relations in a visual way,improving the efficiency of information acquisition by users.