The Technique Of Data Exploration Based On Natural Language Via Knowledge Reasoning

With the increasingly scale of data accumulated by human society,the cost of data analysis and processing are also increasing.Whether knowledge can be efficiently ob-tained from massive data has become one of the key factors affecting the efficiency of data analysis.Data exploration refers to the process of acquiring knowledge from source data;data exploration based on natural language can provide a natural language interface for data,lower the cost for data analysis,and improve the efficiency of data analysis.Knowledge reasoning can mine the potential data semantics in existing knowledge,thus complementing the knowledge base and knowledge graph.Aiming at the problem of incomplete metadata in natural language data exploration tasks,this thesis designs a composite metadata,and accordingly designs a natural language data exploration method based on knowledge reasoning technology.The main work is as follows:(1)A new composite metadata is proposed,and a new construction method of re-lational database composite metadata based on knowledge reasoning technology is de-signed.Based on the multi-source knowledge such as the schema of the relational database,the simple metadata of the relational database is first defined.After building knowledge graphs based on simple metadata,aggregate the graphs to form a more complete knowl-edge graph.Then the knowledge reasoning is performed to complete the knowledge graph,so as to fully mine the rich data semantics in the graph.At last,the final graph serves as the composite metadata.The experimental results show that through knowledge reasoning,the final graph includes more knowledge,and the data semantics contained in the metadata is more complete.This method can deal with the problem of incomplete metadata.(2)A deep learning model based on graph computing,feature enhancement and some other methods is proposed to accomplish a natural language data exploration task for re-lational database based on the composite metadata: translating natural language question into database query SQL.The model adopts the encoder-decoder architecture.The en-coder uses the graph convolutional network and the Transformer model to encode the joint input composed of the knowledge graph and natural language;the decoder uses the abstract syntax tree as the intermediate representation to complete the translation from nat-ural language to SQL query.The experimental results show that the model can effectively finish the natural language data exploration task,generaing corresponding SQL query.(3)Based on the above research work,a software system prototype of natural lan-guage data exploration based on knowledge reasoning is designed and implemented.The system mainly includes natural language data exploration function,metadata construction and management function,database real-time interaction function,corpus management function,etc.