Method And Practice Of Domain Knowledge Graph Construction

With the rapid development of neural networks,deep learning and other technologies,the research and application of perceptual intelligence have made great progress.The other aspect of artificial intelligence: cognitive intelligence,has also attracted more and more attention.Knowledge graph(hereinafter referred to as KG),as an important part of the realization of cognitive intelligence,its research and application have also been paid more and more attention by academia and industry.At present,the related technologies of KG have been applied to many fields such as search engines,intelligent question answering,language understanding,recommendation systems,etc.,bringing great convenience to people's production and life.According to the scope of knowledge involved in the KG,we can divide the KGs into multiple categories such as General KG,Industry KG,and Domain KG.The purpose of large-scale KG such as General KG is to include all aspects of knowledge as much as possible.They need to be constructed from massive amounts of data,and the construction process is extremely difficult and complex.Therefore,it will consume a lot of manpower and material resources,and only large Internet vendors such as Google and Baidu have the needs and capabilities to realize it.Compared with general KGs,Domain KGs are relatively less difficult to construct on the one hand,and only need to involve knowledge in the field.On the other hand,many industries and companies currently have certain requirements for constructing their Domain KGs.Although the construction of a single Domain KG is less difficult than the construction of General KG,there are also various problems in the current research on the construction of Domain KG:? Building Domain KG from scratch: Currently,the construction of Domain KG is concentrated in popular fields such as finance,medical and education,while the construction of KG in many other fields needs to start from scratch,and it is difficult to use existing KG construction methods.The content involved in different fields is very different,and the granularity,breadth,and depth of the knowledge to be processed are also very different.Therefore,it is difficult for the construction methods of KGs in different domains to be useful for reference,which makes it difficult to build a Domain KG from scratch.? Construction of Domain KG faces technical difficulties: Relationship extraction lacks annotation data when performing supervised/semi-supervised extraction,and it is necessary to balance the relationship between annotation cost and extraction accuracy;The definition and construction of domain ontology rely heavily on manual labor.In most cases,the ontology is constructed manually,and even often requires the participation of experts,which makes it difficult to automate.This paper studies the problems in the process of constructing Domain KGs mentioned above,and the main research results are as follows:? Domain KG construction Standardization: In view of the situation where it is difficult to learn from and reuse KGs in other fields when building a KG from scratch.This paper summarizes the methods and processes of Domain KG construction and summarized them as: Data acquisition,Domain phrase mining,Ontology construction,Entity-Relation extraction,Data storage,Query/visualization.Each step is specifically designed and implemented to form a Domain KG construction platform,which is suitable for any Domain KG construction.? Improving relationship extraction scheme: This paper provides specific solutions for the two situations where the accuracy requirement of relationship extraction is high and the accuracy requirement is ordinary.Aiming at entityrelation extraction with high precision requirements,this paper builds a entityrelation annotating platform and website for the supervised relationship extraction method,providing online annotation,multi-person collaboration,adding notes,progress statistics,and one Key export data and other functions greatly reduce the cost and difficulty of supervised relationship extraction.Aiming at entity-relation extraction with ordinary accuracy requirements,this paper proposes to improve Bootstrapping relation extraction and introduces word embedding to evaluate extraction results to reduce error propagation.? Automatically build ontology from scratch: In view of the unreusable ontology faced by many Domain KG construction,the problem of high human participation in ontology construction,this paper proposes the use of encyclopedia data and Word Net auxiliary domain phrases to automatically/semi-automatically construct domain ontology method.? Constructing a military KG: Using the above-mentioned technologies and solutions,this paper uses the military field as an example to construct a knowledge map in the military field.