Research On Knowledge Graph Complex Question Answering For Bridge Maintenance

The Knowledge Based Question Answering(KBQA)technique leverages knowledge graphs as a primary source of information for generating answers to user questions.Presently,KBQA has undergone rapid development in specific domains,such as healthcare,finance,and intelligent customer service.Simple question answering systems have exhibited satisfactory performance;yet,for intricate and various real-world scenarios,simple question answering is inadequate to fully satisfy user requirements.Due to the intricacies involved in parsing questions and the complexity of logical relations in complex question answering,numerous issues and challenges arise that necessitate investigation and resolution.Currently,there is a lack of research focused on complex question answering tailored to the characteristics of bridge maintenance in China.Additionally,subtask technologies related to entity linking and relation matching for complex question answering in this field require further organization and refinement.This thesis focuses on the task of complex question answering for bridge maintenance and combines the structural characteristics of bridge components and the data characteristics of the bridge maintenance field.Based on the construction and annotation of a knowledge graph and a complex question answering dataset,this thesis conducts research on a complex question answering method for bridge maintenance.(1)To address the problem of lacking a knowledge graph and dataset for complex question answering in bridge maintenance,this thesis proposed a knowledge graph threelevel relation architecture for bridge maintenance.Meanwhile,it adopts the naming method of “component name＿(center pile number)” to distinguish the same components of different bridges,and stores the data after knowledge fusion in the g Store graph database,realizing the knowledge graph representation.Based on this knowledge graph,we had constructed a dataset that comprises information on bridge basic profiles,structural damage situations,disease analyses,and maintenance recommendations.This dataset essentially satisfies the practical demands of conducting complex question answering experiments over a knowledge graph in the context of bridge maintenance.(2)This thesis proposed an entity linking method with multi-semantic features designed to address the problem of poor recognition and disambiguation effects of entity linking methods for domain-specific terms in existing question answering systems for bridge maintenance.This approach initially employs the LERT+GRU+CRF joint learning model for the accurate recognition of mentions,specifically addressing the challenge of identifying domain-specific terminology in bridge maintenance.Subsequently,a multifeature entity disambiguation method was proposed,which integrates mention features,entity-linking dictionary features,as well as triple-context semantic information.This method specifically targets the challenge of applying general-domain entity disambiguation in bridge maintenance.The experimental findings reveal that the mention recognition module exhibits high levels of precision,recall,and F1 score,specifically reaching 88.74%,87.27%,and 88.00%,respectively.Moreover,the entity disambiguation module achieves a recall of 79.44%.(3)In order to address the issue of relation-matching subtasks in question answering for handling multi-entity multi-relation questions,we proposed a relation-matching method based on question entity classification,which solves the problems of inaccurate implicit relation prediction and erroneous chain relation prediction in KBQA.This approach categorizes questions based on entity number,distinguishing between singleentity multi-relation and multi-entity multi-relation queries.It then measures the degree of similarity between candidate relations and the original question,based on different classification outcomes.Ultimately,the method identifies the optimal relation for the given query,based on calculated results.The feasibility of the model was verified by experiments,and the F1 score reaches 80.25%.Lastly,all the subtasks were integrated to form a complete question-answering system which facilitates the conversion of a question to SPARQL structured query language and retrieves answers from the g Store graph database.The final KBQA system attains an average F1 score of 68.94%,providing a valuable reference point for other specialized domains to conduct related research.