Research On The Construction Of Industrial Knowledge Graph And Its Application In The Automotive Fiel

Due to the complex technical route of building a knowledge graph(KG),related researches mainly focus on the technological innovation,and there are relatively few studies on the realization of complete technical route and the construction of knowledge graph.In addition,previous works are generally based on standard datasets.Such type of data has consistent text style and regular expression,while in practice,the majority of the textual data is multivariate,heterogeneous and noisy.Thus,achieving highquality knowledge extraction at a low cost is still a hard problem.Furthermore,industrial knowledge graph is one of the hottest research fields,which aims at extracting industry-related structural knowledge and optimizing the downstream tasks.Actually,researchers have made great progress in the extraction of specific knowledge types,but there is still a lack of research on complex entities especially for products and technologies,while such knowledge is of great significance for industry chain analysis.In view of the above challenges,this paper conducts research on the technology and application of industrial knowledge graph.Firstly,we devise several data characteristics oriented knowledge extraction algorithms,including named entity recognition and relation extraction.Secondly,we apply these algorithms to automotive field to extract entities such as companies,products,and technologies,as well as relationships such as R & D,cooperation and competition,so as to build the automotive industrial knowledge graph.Thirdly,we exploit KG based supply chain analysis to measure the applicative value of the industrial knowledge graph.The main work and contributions of this paper are as follows:First,this paper proposes a high-quality context-awareness named entity recognition method.The descriptions of domain entities are relatively complex,but their correlated corpora,such as news documents,usually have rich contextual information.Since named entities tend to occur in a document repeatedly,we assume that frequently occurring strings are more likely to be entities.Therefore,we use the common substrings between the given sample and its document to find high-quality contexts via semantic computing.Then we design a deep network architecture to connect the entity boundary information with the positions of common substrings,and integrate the sentence representations enhanced by multiple contexts,so as to learn the named entity recognizer.Second,this paper proposes a method for semi-supervised relation extraction guided by the hierarchical pattern quality.In order to alleviate the dependency on labeled dataset,we first apply syntactic tools to automatically mine the expression patterns between entities,and then use a small number of labeled templates to predict the relation types of unlabeled ones.Furthermore,to solve the noisy labelling problem,we devise a pattern quality assessment algorithm to grade the quality of pseudo-labels,thereby obtaining pseudo-labels and quality ratings of corresponding samples.From this,a multi-objective relation extraction model is proposed,which aims at exploiting label and semantic information sufficiently.Third,this paper proposes a method for constructing automotive industrial knowledge graph based on the entity fusion.Specifically,the method first uses the proposed entity recognition and relation extraction model to extract automotive knowledge triples from unstructured text,and then designs a prompt learning based unsupervised entity linking algorithm to map the extracted entities to encyclopedia entries,as so to obtain a standard automotive industry graph.The KG contains 5 types of entities,9 types of relationships,4761 nodes and 12571 edges.The results of knowledge assessment and knowledge retrieval demonstrate its quality and analytical capabilities.Fourth,this paper proposes an industrial chain construction and analysis method which integrates automotive knowledge graph.In order to verify the application value of the industrial knowledge graph,this paper proposes a set of semi-automatic industrial chain construction methods,which combine external knowledge to obtain attributes of entities and relationships so as to divide them into separate categories,and then transforms the industrial knowledge graph into a clearly structured industry chain.Based on it,we compare the automotive layout differences between China and other countries from global industry and subindustry perspectives respectively,and deeply analyze the strengths and weaknesses of China’s automotive industry.To summarize,this paper conducts all-round KG research,including algorithm design,KG construction and application exploitation.The experiments show that knowledge extraction algorithms can be easily transferred to other fields,while the applicative exploration of industrial knowledge graph is promising.