Research Of Joint Entity And Relation Extraction Based On Multi-layer Binary Tagging Schema And Multi-task Learning

Information extraction refers to extracting specific factual information such as entities,relationships,events,etc.from natural language texts.It is a technology that transforms un-structured or semi-structured text data into structured information.With the popularity of the Internet,there are more and more data in the form of text.How to convert unstructured text data into structured information and provide data support for downstream applications is what information extraction needs to solve.The specific tasks of information extraction include named entity recognition,entity re-lation extraction,event extraction,etc.Typically,named entity recognition task and entity relation extraction task are combined into a joint entity relation extraction task,which is the core task in information extraction.Utilize deep neural network models,named entities,entity types,and relations between entities in the text can be automatically extracted.The extraction results are usually presented in the form of entity relation triples(head entity,relationship,tail entity).With the rapid development of deep learning,more and more entity relationship extrac-tion models have begun to be based on deep neural networks.We proposed an entity extraction model and a relation extraction model based on the multi-layer binary tagging schema.Then,we proposed a joint entity relation extraction model based on multi-task learning.The main work is as follows:(1)We proposed an entity extraction model based on the multi-layer binary tagging schema,which is aiming at overlapping entity problems where an entity belongs to multiple entity types and two different entities share some common words.With the help of multi-layer binary taggers to indicate the entity position and entity type in the text sequence,it can effectively solve the overlapping problem that cannot be solved by the traditional sequence labeling model.(2)We also proposed an entity relation extraction model based on the multi-layer binary tagging schema,which was proven to be very effective in SEO(single entity overlap)and EPO(entity pair overlap)problems.Based on the head entities extracted in the previous steps,a multi-layer binary tagging schema was used to identify all tail entities under different relation categories.(3)To alleviate the error propagation problem in the pipeline approach of relational triple extraction,we proposed a joint entity relation extraction model based on multi-task learning.The joint model uses a unified feature encoding neural network for the entity extraction model and the relation extraction model and optimizes them jointly.(4)We also studied the information interaction mechanism between the entity extraction model and the relation extraction model.We proposed two ways of information transmission,namely the input layer-based method and the attention mechanism-based method.Based on the above two information interaction methods and multi-task learning,we proposed two joint entity relation extraction models.We call them the Joint-Input Model and the Joint-Attention model respectively.