Study On The Question Answering Based On Knowledge Graph Of Mining Technology

At present,the total amount of mining science data is huge,but most of them are discretely stored.They have problems of low utilization rate,high redundancy,unstructured,and difficulty in mining.In recent years,with the development of knowledge graph,which can not only organize complex data into an organic system but also provides underlying data support for a variety of computer applications.This paper first studies how to use the machine learning to semi-automatically construct the knowledge graph of mining science,and introduces homogenous network clustering and heterogeneous network clustering into the knowledge graph construction,which greatly reduces the workload.Based on the construction of mining science knowledge graph,this paper combines the current development of question answering based on knowledge graph,and studies the question answering based on knowledge graph of mining science.In this paper,the question answering based on knowledge graph is divided into the following parts:(1)Natural question analysis.(2)Entity and relationship component extraction.(3)Entity and relation linking.(4)Search language generation.(5)Natural answer generation.Based on the above division,this paper studies various technical parts,focusing on natural question analysis,entity and relation linking,and natural answer generation.Natural question analysis,the current methods mainly focus on statistical-based syntactic dependency analysis and sequence-based syntactic dependency analysis methods.The statisticalbased syntactic dependency analysis method relies on a large amount of annotation data,while the sequence model requires extremely high data labeling,moreover low training efficiency and poor interpretability.Therefore,this paper proposes a new integrated model,names flat random forest,and applies the model to syntactic dependency analysis.Experiments show that compared with other methods,the proposed flat random forest acquires not only competitive prediction accuracy,higher training efficiency and better interpretability but also adaptive model size.Entity and relation linking,most of the current research divides the entity linking and relation linking into two separate tasks,which makes it impossible to complement each other.In this paper,based on the GTSP model,the entities and relations are jointly linked.Because GTSP has high time complexity and can only returns the optimal results,this paper further models the link density,XGBoost for predict,so that the model ends up with the ability to return a result sort.Experiments show that the accuracy of both entity and relation linking are significant better than separately linking,and with multiple optimal results returns from XGBoost,the accuracy is further improved.Natural answer generation,this paper focuses on how to generate an answer in the form of natural sentences.This paper combines the sequence model with knowledge graph retrieval by means of the attention mechanism and copy mechanism existing in the current research work to obtain a natural sentence answer generator.Through experiments,the natural sentence answer generator used in this paper has achieved good results both on public and mining science data sets.