Research On Entity Relationship Extraction Of Power Safety Operation Based On Semantic Representation Technolog

Power-related texts have an increasingly important role as one of the power resources,and more data with potential value can be found from the power texts.Electric power safety is very important,and the scale of the generated electric power safety operation texts is huge and the value density is low,which is practically difficult to utilize.Therefore,it is needed that entities and their relations can be extracted to build a knowledge graph of electric power safety operations,which can support electric power field maintenance,safety maintenance of grid equipment,electric power safety training and safety inspectors in scenarios such as knowledge guidance,response and query,and has an important role for electric power intelligence.Entity relationship extraction,as a key step in natural language processing,has carried out many researches in many aspects.The current mainstream approach uses deep learning models for information extraction work.Combining different deep learning approaches can overcome the limitations of individual deep learning models in information extraction in terms of feature representation and insufficient learning,and the main problems in the process of entity relationship extraction can be solved by using the advantages of deep learning models.In this paper,we use the semantic representation technique based on deep learning to carry out the work of entity relationship extraction of electric power safety operation text.For the characteristics of large data volume and many data types in the field of electric power safety,the data are cleaned and pre-processed using conventional and targeted means,and the redundant and wrong formats and characters in the text are processed using regular matching and manual removal,etc.According to the characteristics of the data,the phrases that cannot extract complete entity relations are removed and the annotation of the extraction types is performed using manual methods.Then,according to the analysis of the data set,there are statements that cannot be extracted to the entity relationship related to power safety operations,i.e.,irrelevant statements have an impact on the effect and performance of extraction,and this paper proposes to use the sentence bag attention mechanism for noise reduction processing of the data set,dividing the text into words and vectorization,calculating the sentence similarity,dividing the sentences with high similarity into several sentence bags,and using the neural network that incorporates the attention mechanism for the sentence bags The computation is performed by using a neural network with a fused attention mechanism,and the sentences with low attention weights are removed proportionally.The Bert preprocessing model allows semantic characterization of complex semantics in the text data of power safety operations,so as to better learn the semantic features of sentences,location features,etc.,and to be able to grasp the contextual relevant content features well,by using a two-way long and short term memory neural network technique to learn semantic features from both forward and opposite directions,and finally discriminate using a conditional random field model.Thus,the entity relationship triad is extracted.Compared with the traditional methods,the model proposed in this paper performs well on the text dataset of power safety operations,with the F1 value of relationship extraction statistics above 90.0%.By applying the sentence pocket attention noise reduction technique,the extraction efficiency of the textual dataset of power safety operations can be significantly improved.To address the phenomena of long entities,coreference and misreference in the extracted entity relations,dependency analysis based on deep graph coding is used for dependency tree generation,combined with double-layer labeling for labeling word position relations to perform reference disambiguation,and the F1 value is improved by more than 3.0% compared with the traditional labeling method after testing.Finally,the graph database Neo4 j was used to generate the knowledge map.