Research On Safety Accident Risk Of Construction Engineering Based On Natural Language Processing Technology

For a long time,the frequent occurrence of safety accidents in construction projects resulting in casualties and property damage has been of wide concern to the state and society.In order to prevent similar accidents from occurring,an accident investigation report is formed after each accident,and by drawing lessons from historical accidents,safety management strategies and preventive measures can be developed for preventing future risks from occurring.In the past,accident analysis and research relied on manual analysis by experts using their expertise,which made the results susceptible to subjective human influence.At the same time,as the volume of accident report data continues to grow,manual analysis is time-consuming and labor-intensive,it also does not meet the actual needs.Therefore,the use of automated methods to analyze construction safety incidents is essential for safety management.In the face of these accident reports,which usually exist in unstructured or semi-structured form,shallow machine learning combined with manual extraction of words,syntax and semantic features in text is widely used in current research.There are problems of poor learning performance and inadequate accuracy of information extraction.To address these problems,this study conduct research on a model for automatic extraction of accident types and risk factors based on natural language processing techniques,and uses it for safety risk analysis.The details of the research are as follows.(1)To address the problem that the dependency information between long-distance words in the report cannot be captured,an accident type extraction model based on graph neural network is proposed.Firstly,the serialized text is transformed into graph structured data based on word co-occurrence relationships in the accident report.Then,an LSTM network is used to interact word nodes with surrounding neighboring nodes to capture information from higher-order neighboring nodes by setting up a multi-layer LSTM.Finally,the word node information is aggregated into a graph representation of the whole report for classification through an attention mechanism.Experiments were conducted on the constructed chapter-level text classification dataset and an average F1 measure of 92% was obtained,validating the effectiveness of our proposed algorithm.(2)To address the problem that existing methods cannot effectively extract risk factors in construction accident reports,a model combining multilevel neural networks and heuristic rules is proposed to achieve the extraction of risk factors in accident reports.Firstly,the character-level vector representation of the text is obtained using the pre-trained language model BERT.Next,CNN and BILSTM are used to extract local and global feature information of the context,respectively.Then,CRF is used to learn the dependency relationship between tags to complete the extraction of risk entities in the text.Finally,based on this,the risk factors in the accident report are extracted according to seven heuristic rules.Experiments are conducted on the constructed sentence-level entity annotation dataset,and the model achieves 91.4% and 86% F1 measure performance on the risk entity and factor extraction tasks,respectively,indicating that the model has good competitiveness.(3)Focusing on construction safety management needs,a construction safety accident risk analysis system is constructed based on the correlation analysis of the above accident types and risk factor extraction results.Depending on the input,the system mainly implements the risk analysis and database update functions.Among them,the risk analysis function gives the known risk factors,potential risk factors and accident types that may lead to accidents based on the inputted hidden danger ranking text;the database update function automatically updates the association rule base based on the inputted accident report set.By extracting accident types and risk factors from the collected accident report dataset and analyzing the association between them,256 association rules are obtained.The relevant tests show that the constructed system can meet the practical application requirements.This study combines natural language processing technology for the practical needs of the construction engineering safety management environment,and effectively improves the problem of difficult information extraction and insufficient accuracy rate in accident risk analysis.The test results show that the extraction of accident types and risk factors in safety accident reports has been improved to a certain extent,providing an effective help to the field of construction safety management.