| The purpose of the Web Person Abstract is to construct domain knowledge graph about persons from the web pages.The main steps include web content filtering,relation extraction based on the domain ontology model,and web page name disambiguation.There is a large amount of research on ontology modeling,but there is still a lack of standardized guidance in domain modeling,especially the modeling of n-ary relations.The main data for constructing the person abstract ontology comes from using a search engine to search for name.While existing template matching and text extraction algorithms are not directly applicable to extracting person information from different type web pages.The main problem of using the deep learning to extract relation triples is the lack of labeled samples,especially n-ary relations.The web page retrieved by the name searching may correspond to a plurality of different entities.In order to avoid information interference between different entities,the name disambiguation algorithm needs to be performed.However,the existing disambiguation algorithm mainly relies on text hierarchical clustering,and the cluster similar thresholds of different names are different.In order to improve the accuracy of automatically constructing person abstract ontology from the Internet,this paper focuses on the four issues from the following:First,this paper researches on the modeling method of ontology model in the field of person abstraction.The ontology model design language can only represent the unary and binary relations.This paper proposes to map n-ary relations into relational databases and optimize the definition of n-ary relations by using N normal from principles.There are few guiding researches on the construction of domain ontology which mainly relying on the personal experience of domain experts.Based on the Human ontology in Wikidata,this paper combines the entity categories such as institutions and events to provide a generic model of ontology pattern design for person.In addition,this article builds hierarchical relation for location and time entity to facilitate optimization in situations involving high frequency queries and retrieval.Second,this paper proposes a method for extracting personal information from different types web pages.At present,web page information extraction mainly uses template matching extraction and text extraction.Considering the web pages from name searching have different webpage structures,the cost of extracting templates one by one is huge.Only part of the personal information appears in the body area of a web page.Only part of the personal information appears in the body area of a web page.Using the text extraction algorithm for personal information may lose a lot of valid information.The labeled data experiments prove that the personal information extraction algorithm proposed in this paper can greatly improve the recall rate of related information.Compared with other traditional classification algorithms,the deep learning model based on sequence prediction works best.Third,this paper proposes an algorithm for n-ary relation labeling.As deep learning algorithms gradually becomes the mainstream algorithm of relation extraction,the lack of labeled samples in domain ontology model becomes the bottleneck for the automatic generation of domain knowledge.In this paper,the annotation information of the public encyclopedia personal information introduction part is firstly obtained,and then the corresponding sentences are obtained by the person name and attribute searching.Then,the sentence is using for attribute extraction model training.The results show that the deep learning algorithm can effectively solve the relation extraction issue when the sample is sufficient.Compared with binary relations,there are few related researches on n-ary relation extraction.Test samples for n-ary relation are not easy to obtain automatically.The labeling process n-ary relations is much higher than that of binary relations.Manual labeling also has a high error rate.In this paper,the self-supervising of deep learning and active learning method are combining to improve the accuracy of small samples by constructing unbalanced noise marker samples.The algorithms proposed by this paper realizes automatic iterative active learning.Fourth,this paper optimizes the name disambiguation algorithm.Name disambiguation is to classify web pages belonging to the same person into the same category,which belongs to unsupervised clustering issue.Due to the clustering set of each person’s name is different.The number of specific result sets and the thresholds in the cluster are different.The existing methods mainly focus on the estimation of the number of sets and the dynamics of the threshold.In this paper,the dirichlet process hybrid model are used to realize the disambiguation clustering.Compared with the hierarchical clustering method,only the maximum number of categories needs to be adjusted in the dirichlet process hybrid model,which the maximum number of categories is larger than and close to the largest in the cluster.The algorithm can be automatically optimized the number of categories according to the data,and the results show that the algorithm proposed in this paper can be effectively applied to the Web name disambiguation problem.In summary,this paper researches on the automatically constructing person abstract knowledge graph from Internet.This paper focuses on ontology model design,web page personal information extraction,n-art relation extraction and web person name disambiguation.Finally,a person abstract system is designed in this paper.By using the algorithm proposed in this paper and manual correction,a high-quality person abstract knowledge graph is generated. |
