Open Domain Entity Relation Extraction Based On Attention Mechanism

Bigcilin is a reticular knowledge graph with automatic construction ability.The so-called reticular structure refers to the horizontal and vertical knowledge.The relationship between entities and the relationship between synonyms of hypernyms belong to the horizontal relationship,while the r elationship between entities and hypernyms and the hierarchical relationship between hypernyms belong to the vertical relationship.The main purpose of this paper is to provide assistance for the Bigcilin,which aims to provide a stable horizontal relation ship supplement for the Bigcilin,that is,the complement of the relationship between entities.1.Aiming at the problem of the disunity between the entity in the vertical relationship and the entity in the horizontal one,that is,entity path disambiguat ion.Firstly,this paper proposes a method based on word vector similarity calculation,which corresponds entity path information to entity semantics,treats entity path information and entity semantics as two strings,and calculates cosine similarity using the word vector provided by Tencent.This method has a good effect in the case of obvious semantic features.But in Chinese,many words have opposite meanings.In order to excavate the "opposition" Association behind the text,this paper introduces the entity path disambiguation based on deep learning,including ELMO model and Bert model,and analyses the results of the two models.2.In the open domain relation extraction task,aiming at the lack of corpus in the Chinese domain relation extraction task,this paper proposes a relation extraction method based on heuristic rules.Four types of heuristic rule templates are used to extract triples from the text,and the search engine results are used to calculate the confidence.The preliminary experimental results are obtained,and the corpus for supervised methods is provided.Then,in view of the difficulty of relation extraction in open domain,the seq2 seq relation extraction method is introduced to improve the weaknesses of the mainstream method which relies on named entity recognition and restricts the type of relation.While obtaining the relation,the model also obtains head entity and tail entity,links named entity recognition with relation extraction task,and finally the experiment results are further enhanced by the search engine,which make the model more suitable for Chinese data sets.3.After obtaining the relation triple from the text,it is necessary to map the head entity and tail entity of the relation triple to the specific entity semantics,that is,entity mapping.In this paper,an entity mapping method combining text information is proposed.The structure of transformer network is used to learn the feature representation of text information and entity semantics.The probability of whether the head entity and tail entity extracted from text belong to the entity semantics is predicted by the feature representation with context and entity semantics information.Experiments show that the method achieves the desired results.