Research On The Construction Of Knowledge Graph In The Field Of Military Equipment

The effective organization and storage of military equipment data is an important cornerstone for constructing intelligent military equipment knowledge system.In the field of military equipment,a large amount of data such as equipment attributes and relationships among the equipment exist,which have important research and application values.However,due to the lack of effective data organization and storage structure,it is difficult for relevant users to obtain equipment information quickly and accurately when facing the massive and scattered military equipment data.Knowledge graph is a kind of data storage structure with a triplet as the unit,and it is a semantic network in essence,which can effectively describe the relationship between entities and the attribute information of entities.To address the problem of lack of effective organization and utilization of data in the field of military equipment,a knowledge graph in the field of military equipment is constructed to realize the structured storage of military equipment data,in which the key technology points are studied,such as relation extraction,attributes extraction and attributes fusion.A method that fuses BERT with the relation position features is proposed for joint entity relationship extraction in the military equipment domain.The joint entity relationship extraction is transformed into a hierarchical sequence labeling task through a hierarchical reinforcement learning framework.BERT is used as an encoder of the input text to obtain its hidden layer state vector.The hierarchical reinforcement learning method is used to recognize the relationship trigger words and their corresponding entities separately,and the relationship location features are fused to identify the relative position information of other words and the relationship trigger words in the entity decoding process.The experiments prove that the method can effectively perform joint entity relationship extraction in the field of military equipment,and the F1 value reaches 82.2%on the military weaponry dataset,which is about 8%improvement compared with the comparison method.Meanwhile,the method has certain generalization ability,and the F1 value reaches 71.8%on the publicly available NYT10 dataset.In terms of attribute extraction in military equipment domain,an attribute extraction dataset in military equipment domain is constructed based on remote supervision idea,and an attribute extraction method based on RoBERTa-BiLSTM-CRF structure and combined with entity boundary prediction layer is proposed.In the model encoding layer,RoBERTa is used to obtain the vector representation of the text,and then it is input to the entity boundary prediction layer for the boundary prediction of entity and attribute values.After obtaining the boundary prediction results,they are used as features to concatenate with the hidden layer vectors output by RoBERTa and input to the bidirectional LSTM and CRF layers for prediction of attribute label sequences.In the model optimization process,the loss values of entity boundary prediction and attribute label prediction are considered together and weighted summed as the overall loss of the model.Experiments show that the method can effectively perform attribute extraction in the field of military equipment,and the F1 value of the model reaches 77%on the constructed attribute extraction dataset.In the context of attribute alignment in the military equipment domain,an attribute synonym expansion method based on edit distance and knowledge expression model based on TransH is proposed for the fusion of multiple attribute expressions in encyclopedic data.The edit distance based method discovers the set of synonyms of attributes by calculating the edit distance between the attribute names and the words in the domain word list.The knowledge representation model-based approach first obtains the attribute triples in the encyclopedia pages,then inputs them to the knowledge representation model TransH for training,obtains the attribute vectors,and obtains the set of synonyms of the attributes by calculating the Euclidean distance between the vectors.The synonyms extended by the above two methods are taken and combined as the final result of attribute synonym extension.The experimental results show that the methods are feasible for the synonym expansion of attributes,and the F1 value of the model reaches 69.8%.In summary,the knowledge graph in the field of military equipment is constructed through relationship extracting,attribute extracting and property alignment,effectively organizes and stores military equipment data,and builds a cornerstone for intelligent military construction.