Knowledge Support for Emergency Response During Construction Safety Accidents

Abstract

Emergency response to construction safety accidents is the focus of this study. Despite the abundance of data and materials available for emergency response in construction safety, the unstructured nature of the knowledge and the disordered state of storage have limited the timely application of this knowledge in decision-making for emergency response. In this study, scenario-response theory, natural language processing, and deep learning technologies were employed to construct a domain knowledge graph for emergency response in the field of safety accidents. First, based on scenario-response theory and domain-specific materials, four categories of scenario domains and 14 types of scenario elements were identified. Second, according to the mapping relationships between scenario elements and emergency response knowledge, 14 entity types and 10 relationship types were determined, thereby forming the knowledge structure pattern of this field. Subsequently, 4877 entities and 5783 relationships were extracted by means of the BERT-BiLSTM-CRF model and the BERT-CNN model, with F1 values reaching approximately 0.8. Finally, the Neo4j graph database was adopted for data storage, and a domain knowledge graph was constructed. Based on this graph, services such as knowledge association, knowledge retrieval, and intelligent question-answering were implemented. These services effectively addressed key challenges in information acquisition and decision support for on-site safety management, thereby significantly enhancing response efficiency and quality while strengthening overall safety management practices within the construction industry.

1. Introduction

The construction industry is recognized as a vital component of the national economy and has long been regarded as a key driver of economic growth. While significant accomplishments have been achieved in the industry, it should be noted that safety conditions have become increasingly severe [1]. Due to the large number of workers in the construction industry and varying levels of skill [2], safety accidents have continued to occur frequently despite the enhancement of safety management and training measures, as a result of various complex factors [3]. According to statistics, the average annual number of fatalities from housing and municipal safety accidents in China from 2020 to 2024 was 682. In 2021, the number of fatalities reached the highest level in the past five years, with 823 deaths were reported. In recent years, both the number of housing and municipal safety accidents and the associated fatalities have remained at high levels, further emphasizing the urgency of the safety situation.

The effective application of information technology has been shown to significantly enhance safety management and knowledge management in the construction industry [4], thereby reducing personnel injuries and economic losses caused by safety accidents. Although the importance of emergency response knowledge has been acknowledged by project managers and technical personnel, the emergency response to construction safety accidents in China has long been dominated by a forecast-response model [5]. However, such plans have often failed to cover all possible scenarios at the accident site, making it difficult for flexible responses and targeted measures to be achieved. The present study combines scenario-response theory with knowledge graph technology to dynamically generate emergency response strategies, enabling real-time responses to sudden incidents at the scene.

Scenario-response emergency response strategies can be developed based on immediate scenario elements exhibited by construction safety accidents. By analysing the knowledge and material requirements of the immediate scenario, corresponding emergency response measures can then be formulated. Knowledge graphs, as an advanced data storage and processing method, are capable of structurally organizing and storing domain knowledge through rich semantic forms, thereby clearly illustrating the relationships between different pieces of knowledge and revealing complex knowledge structures [6]. When guided by scenario–response theory, the construction of knowledge graphs enables the identification and analysis of knowledge response requirements under specific accident scenarios, allowing knowledge to be more accurately integrated across multiple aspects such as accident control and recovery. For emergency response to construction safety accidents, this approach can significantly improve the efficiency of acquiring safety knowledge, thereby reducing personnel injuries and economic losses.

The primary objective of this study is to construct an emergency response knowledge graph for construction safety accidents, based on the situational response theory. This initiative aims to integrate domain-specific emergency response knowledge, thereby providing knowledge support and question-answering services for incident management, with the goal of enhancing the efficiency and accuracy of emergency response actions. To achieve this objective, the following research tasks were formulated and completed: (1) Identification of scenario domains and scenario elements that can describe construction safety accident scenarios; (2) Development of a domain knowledge structure model; (3) Extraction of entity types and relationships through the analysis of collected relevant materials; (4) Selection of the Neo4j graph database for data storage to construct the domain knowledge graph; (5) Implementation of a large language model-enhanced question-answering system for the domain knowledge graph. In this context, this study next reviews scenario-response approaches and the application of knowledge graphs in construction safety.

2. Literature Review

2.1. Current Status of Research on Scenario-Response Emergency Response Theory

Construction safety accidents are characterized by uncertainty, high frequency, and severe consequences. When accidents are addressed under various complex scenarios, traditional forecast-response emergency response methods have been found to exhibit significant limitations [5]. Existing knowledge-assisted emergency management frameworks primarily rely on rule-based reasoning and historical data analysis to generate emergency response measures. However, these methods are generally lacking in the ability to respond dynamically to real-time developments at the accident site, especially when faced with complex incidents, where flexibility is often limited. Zichen et al. [7] proposed a gas accident emergency response method integrating pre-trained deep learning models and knowledge graphs to enhance response efficiency. However, this approach relies heavily on large volumes of standardized textual data, limiting its flexibility and real-time capabilities when confronting complex emergencies. Lee et al. [8] proposed a method based on BERT and graph models to efficiently extract accident risk knowledge from unstructured data. However, this approach heavily relies on specific construction safety cases, potentially introducing bias and limiting its applicability to novel or unseen construction safety data. In contrast, scenario-response emergency response methods are designed to analyze the knowledge and material requirements of the immediate scenario based on the scenario elements exhibited by construction safety accidents and to formulate corresponding emergency response measures. Xie et al. [9] defined scenario-response as a method in which models are constructed based on the real-time and evolving characteristics of sudden events such as rainstorms and food incidents, with targeted dynamic response measures being implemented to improve the quality and efficiency of emergency management. Shi et al. [10] examined a scenario-based emergency management method for railways, proposing three scenario simulation models: dynamic Bayesian networks, fuzzy neural networks, and convolutional neural networks. Suo et al. [11] conducted a comprehensive analysis from three aspects: scenario elements, scenario evolution, and scenario effects, and a scenario-driven model was constructed. Niyazi and Behnamian [12] developed an accident emergency response tool based on scenario analysis for flood emergency logistics planning. Pang et al. [13] extracted key scenario elements from multiple dimensions of accidents, employed a predictive model, and the determination of scenario state probabilities was achieved. She J et al. [3] constructed a scenario simulation and emergency decision-making evaluation framework for construction safety accidents, in which Bayesian networks were integrated with entropy-weighted TOPSIS. Liu et al. [14] proposed a multi-scenario simulation method based on knowledge metatheory and dynamic Bayesian network, and its effectiveness in railway emergency decision-making was validated through a train derailment case study. Bian et al. [15] proposed a probabilistic risk analysis method based on fault tree analysis and Bayesian networks to assess the risk of power transmission line tripping accidents caused by natural hazards such as wildfires, lightning, strong winds, and ice storms. Zeng et al. [16] combined knowledge elements and dynamic Bayesian networks to investigate emergency scenario deduction for oil spill accidents caused by tanker collisions. Song et al. [17] proposed a modeling and analysis approach for emergency scenario evolution systems based on generalized stochastic Petri nets, aimed at improving the “scenario-response” decision-making process. Collectively, these studies have demonstrated the broad applicability and practical value of scenario-based emergency response models across multiple fields, such as natural disasters and railway transportation, providing scientific decision support for effective responses to safety accidents.

In this mode, emergency response to construction safety accidents is understood as involving “the analysis of the impact of the current scenario elements on the accident based on real-time information, and the subsequent determination of appropriate emergency response measures”. The situational information of construction safety accidents represents the real-time status of the accident, so the scenario-response emergency response can be regarded as a dynamic process. Based on the different situational factors exhibited by the accident, accident response personnel conduct response analyses under the influence of various factors and promptly implement effective response measures to mitigate the consequences of the safety accident.

2.2. Application of Knowledge Graphs in Construction Safety

A knowledge graph is defined as a knowledge base that is organized and stored in a graphical format [18], which enables the integration of domain entities and relationships in the form of a graph. Knowledge graphs are categorized into two types according to their application domains: general-purpose and domain-specific. General domain knowledge graphs are designed to emphasize domain coverage and the breadth of knowledge, whereas specialized domain knowledge graphs focus on the depth of knowledge in specific domains to meet specialized needs. The knowledge graph constructed in this study is classified as a domain-specific knowledge graph for the construction vertical domain.

In research on knowledge extraction for the construction of knowledge graphs in early-stage construction safety, rule-based methods are generally employed [19]. For instance, in studies on the management of high-frequency, low-consequence general safety accidents, manually coded rules were utilized by Tixier et al. to extract domain attributes and entities from unstructured accident report texts [20]. With the continuous development of natural language processing technology, pre-trained models based on deep learning algorithms have begun to be applied to knowledge extraction [21]. Zhong et al. [22] employed a deep learning model combining BiLSTM and CNN to automatically extract construction process constraint knowledge from construction specification texts. Moon et al. [23] developed a named entity recognition model based on word vectors, through which key entities were extracted from construction specifications for construction safety management. Feng et al. [24] applied a BiLSTM-CRF model to extract accident information from accident investigation reports, and reliance on large-scale training datasets was reduced through a proposed data augmentation small-sample training framework. Xu et al. [25] applied the BERT-BiLSTM-CRF method to knowledge entity recognition in the field of coal mine construction safety, and higher accuracy than traditional methods was achieved.

With the development of natural language processing and machine vision, an increasing number of researchers have applied these technologies to extract information from construction standard specification texts and construction scene images [8]. Fang et al. [26] adopted recurrent neural network methods to extract entity objects from construction images, through which spatial distance information related to accidents was obtained and unsafe behaviors at construction sites were identified. Pan et al. [27] applied deep learning technology to extract entities from videos, through which domain knowledge graphs were continuously updated. These studies were considered to provide new solutions for real-time monitoring and accident prevention. Shuai [28] proposed an enhanced reasoning-based natural language processing framework that simulates human reasoning processes to analyze contract texts and identify unilateral contractual change risks in construction projects. Li T et al. [29] reviewed the application of digital twin technology in intelligent tunnel construction, suggesting that it optimizes tunnel engineering decision support through real-time data acquisition and multi-physical modeling. Borjigin S.G et al. [30] introduced a semi-automated textual requirements development framework based on natural language processing and template matching, aimed at improving the quality of construction industry documentation and reducing communication barriers and delays. These studies were found to improve not only the automation and intelligence of construction safety management but also the efficiency and effectiveness of accident handling and prevention. However, research on the construction of knowledge graphs in the field of emergency response to construction safety accidents remains in its infancy, and numerous challenges are still encountered in the actual construction process, requiring further in-depth exploration by scholars.

3. Research Methods

3.1. Research Framework

In this study, a knowledge graph was constructed for the targeted research domain. Methods such as scenario–response theory and knowledge mapping were employed to determine the pattern layer of the domain knowledge graph. Subsequently, based on the characteristics of the selected domain knowledge data sources, natural language processing and deep learning techniques were utilized to extract entities and relationships from the domain knowledge graph. The extracted results were stored in a Neo4j graph database, through which a visualizable domain knowledge graph was formed to serve as the data source for knowledge-based question-answering systems. During the implementation of knowledge-based intelligent question-answering for the domain, the knowledge graph stored in Neo4j was used as the data source, and the query results were provided as input to a large language model to generate answers with richer semantics and greater accuracy. By integrating and leveraging domain-specific knowledge, the knowledge graph-based question-answering method was designed to provide relevant professionals with a rapid and accurate means of knowledge acquisition, thereby enhancing the efficiency and quality of accident emergency response. The overall research framework is illustrated in Figure 1. The model proposed in this paper is not only applicable to Chinese text, but also be suitable to others after optimization, which provided a reference for recognizing named entities in multiple languages. With the overall framework specified, the next step is to explain how scenario elements are mapped to control and recovery knowledge.

3.2. Emergency Response Knowledge Mapping Based on Safety Accident Scenario Elements

The emergency response process for construction safety accidents is initiated with the collection and analysis of accident scenario elements, through which the current immediate situation of the accident is determined, and the essential characteristics of the accident are reflected by depiction key situational factors. The classic “triangle model” theory in the domain of public safety proposes that public safety incidents can be divided into sudden events, disaster-bearing entities, and emergency management [31]. Based on this theoretical framework, the scenario domain of safety accidents is typically divided into events, entities, and activities, while emergency management cases are further classified into disaster-causing factors, emergency activities, and disaster-bearing entities. By integrating the “triangular model” and conducting an in-depth analysis of relevant investigation reports and other materials related to construction safety accidents, the situational domain of construction safety accidents in this study was subdivided into four main categories: accident events, accident causation, accident consequences, and emergency activities, encompassing a total of 14 scenario elements.

The scenario elements included in the accident events scenario domain consist of accident type, time information, spatial information, environmental information, and accident severity. The elements of the accident causation scenario domain comprise person factors, equipment factors, and management factors. The elements of the accident consequences domain include personal harm and property damage. The elements of the accident emergency activities domain cover the materials required for emergency response, the personnel organization responsible for executing emergency response, and the knowledge required to control and recover from the accident.

As shown in Figure 2, the safety accident control phase is divided into six sub-tasks, such as personnel rescue. The knowledge requirements for personnel rescue are considered to include safety accident identification, preliminary assessment, information collection, rescue techniques and evacuation, and the use of rescue equipment. Material supply is required to be supported by knowledge of management, scheduling, and demand assessment. Communication and coordination are defined to encompass internal and external coordination and communication, as well as communication support technologies. On-site control is characterized by the establishment of security perimeters, evidence protection, and vehicle evacuation. Engineering rescue operations are necessitated by the need for expertise in analyzing the causes of property damage and implementing rescue strategies. Medical rescue operations are required to be guided by knowledge of injury assessment, basic first aid, tool usage, referral and follow-up procedures. The accident recovery phase is composed of four sub-tasks, such as accident investigation. The knowledge required for accident investigation is identified to include evidence collection and analysis, accident report writing, and relevant legal knowledge. Site cleanup work is understood to involve debris clearance and restoration, waste classification and disposal, cleaning and disinfection, and material needs analysis. Post-accident handling is considered to encompass economic and insurance compensation, mental health counselling, and public opinion management. The lifting of alerts is required to be supported by environmental safety assessment confirmation, site environmental restoration standards, and knowledge of security facility removal.

During the rapid emergency response to safety accidents, the necessary knowledge is transmitted quickly and in an orderly manner among various emergency response organizations. The process of emergency response knowledge mapping analysis, based on safety accident scenario elements, directly involves the analysis of knowledge requirements through the description of the accident scenario elements. An emergency response knowledge mapping system was established in this study to enable rapid knowledge provision. The principle of knowledge mapping is illustrated in Figure 3, and a detailed analysis of the mapping between scenario elements and emergency response knowledge is presented in

Table 1. Mapping analysis of scenario elements and emergency response knowledge.

Scenario Elements	Knowledge Mapping
	Knowledge of Accident Control	Accident Recovery Knowledge
Accident type	Accident identification; Initial assessment of the accident	/
Time information	Assessment of material requirements	Assessment of material requirements
Spatial information	Rescue techniques and evacuation; The use of rescue equipment; Assessment of material requirements; Accident scene security perimeters; Vehicle evacuation at the accident site	/
Environmental information	Assessment of material requirements	Assessment of material requirements
Accident severity	Assessment of material requirements; Materials management and scheduling; Internal and external coordination and communication; Accident scene security perimeters; Vehicle evacuation at the accident site	Accident report writing; Relevant legal knowledge; Public relations and public opinion management
Person factors	/	Evidence collection and analysis; Responsibility Consequence Analysis
Equipment factors	Initial assessment of the accident; Rescue techniques and evacuation; Assessment of material requirements; Accident scene security perimeters;	Evidence collection and analysis; Responsibility Consequence Analysis
Management factors	/	Evidence collection and analysis; Responsibility Consequence Analysis
Personal harm	The knowledge of injury assessment; the knowledge of basic first aid; Use of medical rescue equipment; Referral and follow-up procedures	Economic and insurance compensation; Mental health counselling; Relevant legal knowledge; Accident report writing
Property damage	Causes of property damage; Emergency response technical strategies	Accident report writing; Economic and insurance compensation; Relevant legal knowledge; Debris clearance and restoration; Waste classification and disposal

. These mapping relationships facilitate the dynamic integration of knowledge, ensuring the precise alignment of situational elements with the most appropriate emergency response measures. For instance, during an incident, the identification of the accident and the assessment of its severity result in the mapping of relevant information to specific rescue protocols and resource requirements, such as the provision of medical equipment or the deployment of rescue gear. This knowledge mapping mechanism allows the knowledge graph to provide rapid, precise, and actionable information to field personnel, thereby significantly enhancing emergency response times and optimizing the overall efficiency of emergency management.

When constructing the domain knowledge graph model layer, a top-down top-level design strategy was adopted. A scenario-response emergency response strategy was employed in this study to construct the domain emergency response knowledge graph, and scenario elements were directly applied as entity types within the domain. Subsequently, based on the analysis of entity relationship types across different dimensions, 10 relationship types, including occurrence time relationships and control knowledge relationships, and 30 domain entity relationships were identified. Considering the knowledge characteristics of the construction safety accident emergency response domain, and an in-depth analysis of scenario elements and their interactive relationships, the knowledge graph model layer structure for the construction safety accident emergency response domain was constructed. This structure is illustrated in Figure 4, where circular boxes represent various entity types and arrows indicate the relationship types between different entity types.

3.3. Knowledge Extraction Model

3.3.1. Entity Recognition Model

Compared to traditional machine learning methods, the continuous development of word embedding techniques and deep learning models has led to deep learning models achieving superior performance in various natural language processing tasks, making them the mainstream approach in current named entity recognition research. The knowledge in this study was primarily derived from accident investigation reports, emergency response plans, and response protocols, which are often unstructured and contain multiple types of entities. Based on the number of instances and the clarity of their definitions, domain entity types were classified into two categories. The first category consists of entities with fewer instances and clear definitions, such as accident type, accident level, time information, environmental information, and emergency disposal organizations, which were directly determined according to existing standards. The second category instances entities with numerous instances and vague definitions, for which deep learning models were required for named entity recognition. For entities in the second category, the BERT-BiLSTM-CRF model was employed in this study. The model is composed of three components: BERT, which is used to generate dynamic word vector representations; BiLSTM, which is applied to capture context-dependent features; and CRF, which is employed to perform global optimization [32].

While BERT is a powerful model for understanding context in text, alternative transformer-based models, such as RoBERTa and ERNIE, were also considered. RoBERTa has shown superior performance in large-scale data tasks due to its extended pretraining strategy, yet it does not significantly outperform BERT in tasks like entity recognition and complex relationship extraction. Furthermore, ERNIE, which incorporates external knowledge through knowledge graphs, excels in tasks heavily reliant on structured external knowledge. However, for the purpose of emergency response system development, where textual data is largely unstructured, BERT-BiLSTM-CRF and BERT-CNN models offer better performance. The combination of BiLSTM for capturing long-term dependencies and CRF for ensuring label consistency in entity recognition provides a robust approach for processing accident reports, where context and coherence are critical. Additionally, the CNN layer in BERT-CNN excels in extracting local relationships between entities, making it well-suited for relationship extraction in emergency response data.

The BERT component was capable of obtaining context-sensitive dynamic word vectors through pre-training, which enabled the effective handling of the semantic complexity inherent in emergency response to construction safety accidents. For example, in the sentence “Workers fell from scaffolding due to a lack of anti-slip measures in rainy weather”, BERT was able to accurately identify “rainy weather”, “anti-slip measures”, and “fall” to recognize the semantic relationships among them, and to correctly label the accident cause and the accident event. The BiLSTM component was applied to capture contextual dependencies by means of bidirectional sequence modeling. In accident investigation reports, sentences such as “Due to inadequate shoring, the excavation pit collapsed, trapping three workers” often appear. BiLSTM was able to correlate “inadequate support” with “collapse of the excavation pit” while simultaneously considering the consequence information “three workers were trapped”, thereby achieving comprehensive identification of the accident cause, accident event, and accident consequence. Although BiLSTM demonstrated strength in processing long-distance dependencies, it was unable to adequately capture dependencies between adjacent labels. To address this limitation, the CRF component was applied to optimize transition probabilities between adjacent labels [33], thereby ensuring the coherence and logical consistency of extracted entities. For example, in the sentence “After an electric shock accident occurs, immediately cut off the power supply and perform cardiopulmonary resuscitation”, “cutting off the power supply” was not incorrectly labeled as the cause of the accident but was uniformly labeled as “emergency response”, thereby ensuring consistency in knowledge extraction.

3.3.2. Relationship Extraction Model

In relation extraction, two main approaches are typically distinguished: rule-based extraction and deep learning-based extraction. Rule-based methods require the manual construction of a large number of rules, which not only consumes significant human and temporal resources but also struggles to cover all possible scenarios. Given its complexity and diversity, the BERT-CNN model was chosen for relation extraction. Among the ten types of entity relationships, the occurrence time relationship and occurrence environment relationship were determined directly by existing standards due to their limited number. For the remaining eight types of relationships, extraction was conducted using the BERT-CNN deep learning model. This model is composed of two components: BERT, which is employed to generate context-sensitive dynamic word vectors, and CNN, which is applied to extract local features and classify entity relationships.

Similarly, the BERT model was used to generate dynamic word vectors, thereby enhancing the context sensitivity of semantic representations. For instance, in the sentence “The crane cable broke, causing the heavy object to fall, injuring two workers”, BERT can link “cable breakage” with “heavy object falling” as “accident cause-accident type” and link “heavy object falling” with “workers injured” as “accident type-accident consequence”. The CNN model was particularly effective in capturing local features and was suitable for identifying localized dependencies in accident statements. For example, in the sentence “After the excavation pit collapsed, the rescue team used support frames to stabilize the soil”, CNN can accurately extract the “use relationship” between “rescue team” and “support frames”. Likewise, in the sentence “The tower crane overturned, causing construction materials to scatter”, CNN can accurately identify the causal relationship between “tower crane overturning-causing-materials scattering”.

4. Experimental Analysis

4.1. Data Collection and Preprocessing

To ensure the comprehensiveness and accuracy of the domain knowledge graph content, 15 laws and regulations in the field of construction safety management, 409 safety accident investigation reports (covering the six most frequent accident types in recent years: falls from height, struck by objects, earthwork/excavation collapses, lifting injuries, machinery injuries, and electric shocks), 32 safety accident emergency response plans and on-site disposal plans, and 2 textbooks—Emergency Management and Plan Compilation for Construction Accidents and Emergency Response and Rescue for Work Safety Accidents—were used as data sources to build the corpus for the domain knowledge graph. The legal regulations primarily serve as the source of knowledge for accident investigation and other aspects of accident recovery. Accident reports are mainly used as sources for knowledge related to the location of the accident, human factors, equipment factors, management factors, and certain aspects of accident recovery. Emergency response plans and textbooks primarily serve as sources for knowledge on accident control, the selection of emergency response materials, and accident recovery.

To reduce noise interference, the original dataset was preprocessed to entity extraction. Given the large scale of the data, data cleaning was conducted through a human–machine collaboration approach. Multiple sets of cleaning rules were proposed and implemented for different types of data. For accident investigation reports, documents containing fewer than 100 words or lacking content related to “emergency response processes” or “liability determination and penalties” were excluded. For textbooks, emergency response plans, and on-site disposal plans, non-textual information such as charts and images was removed or converted, and core content was extracted using key information matching technology. If two documents contained identical key information, they were deemed duplicates and excluded. For legal and regulatory texts, redundant content was removed, including generic template phrases, repeated clauses and definitions, legal citations and cross-references, appendices, and indexes. Following data cleaning and normalization, human-in-the-loop annotation is conducted.

4.2. Entity and Relationship Annotation

4.2.1. Entity Annotation

For data requiring entity extraction using deep learning models, the BIO + ES method was employed for annotation. This approach was designed to enhance annotation quality by introducing additional label types, thereby optimizing the training performance of subsequent models. Specifically, the boundary identifiers “O−” and “O+” were introduced to mark characters immediately adjacent to entities, rather than uniformly labeling all characters outside entities as “O”. In the actual annotation process, the Label-Studio platform was utilized to complete the data annotation work. The annotation category definitions are provided in

Table 2. Definition of entity annotation labels.

Entity Category	Tags Included (Initial Letter, Non-Initial Letter, Ending)
Spatial information entity	B- Spa, I- Spa, E- Spa
Person factor entity	B- Per, I- Per, E- Per
Equipment factor entity	B- Equ, I- Equ, E- Equ
Management factor entity	B- Man, I- Man, E- Man
Personal harm entity	B- Har, I- Har, E- Har
Property damage entity	B- Pro, I- Pro, E- Pro
Emergency disposal material entity	B- Mat, I- Mat, E- Mat
Knowledge of accident control entity	B- Ack, I- Ack, E- Ack
Knowledge of accident recovery entity	B- Ark, I- Ark, E- Ark

Notes: Tags follow the BIO + ES scheme, where “B/I/E” denote the beginning, inside, and end positions of an entity, respectively. “O⁺” and “O⁻” indicate boundary characters before or after the entity, “S” represents single-character entities and “O” represents non-physical characters.

. Label-Studio can collaborate with machine learning models to implement an automated labeling process, thereby reducing the burden of manual annotation and improving data processing efficiency. To ensure the consistency of the annotations, detailed training was provided to the annotators, ensuring that each annotator fully understands the specific meanings of each label. During the annotation process, a cross-validation method was employed, where two annotators independently annotated the same data, and the annotation results were compared to ensure the reliability of the annotations. Additionally, regular checks of the annotation results were conducted to ensure that the labeling of each entity category met the required standards. If any errors were identified, they were corrected through a feedback mechanism, and the relevant annotators underwent retraining to maintain ongoing accuracy in the annotations. Annotation examples are illustrated in Figure 5.

4.2.2. Relation Annotation

In the annotating of entity relationships, the most important step was to accurately identify the head entities, tail entities, and their mutual relationships contained in the text sentences. In this study, approximately 1500 sentences were manually annotated, after which computational methods were applied to further complete the annotation process. A total of 7143 instances were annotated. In the annotated corpora, the data items were arranged from left to right as: head entity, tail entity, relationship, and text sentence. Detailed examples of relationship annotations are provided in

Table 3. Relationship annotation methods (Example).

Head Entity	Tail Entity	Relationship	Text Statement
Medical Rescue Team	Insulated gloves/ Dry clothes, scarves, hats, and other insulated items	Usage relationship	The medical rescue team should wear insulated gloves or wrap dry clothes, scarves, hats, or other insulated items around their hands to drag the electrocuted person away from the power source.
Medical Rescue Team	Dry wooden boards	Usage relationship	The medical rescue team can first use dry wooden boards to insulate the electrocuted person from the ground and interrupt the current flowing into the ground.
Medical Rescue Team	Do not enter within 8–10 m of the location where the wire fell to prevent step voltage electrocution	Demand relationship	Medical rescue teams must not enter within 8–10 m of the landing site of a fallen power line to prevent step voltage electrocution.

.

4.3. Data Segmentation

The annotated entity recognition data was divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model training, the validation set was employed to optimize hyperparameters, and the test set was used to evaluate the model’s generalization ability. Through a combination of manual and machine-based organization and annotation work, an entity-relationship recognition experimental corpus containing approximately 5143 entries was constructed, and the annotated corpus was further divided according to the 8:1:1 rule, 4115 training sets, 514 validation sets, and 514 sets.

4.4. Model Training and Result Analysis

The hardware and software configurations used for model training are summarized in

Table 4. Training Environment Configuration.

Label	Configuration
Device memory	32 G
Operation system	Windows11
CPU	intel(R)core(TM)i7-11800H
GPU	NVIDIA GeForce RTX 3070
CUDA	12.1
Python	3.8
PyTorch	1.9.0

. Knowledge extraction tasks were implemented using the PyTorch1.9.0 deep learning framework and the Python3.8 programming language, and training was conducted in PyCharm Professional. In the specific model parameter settings, the maximum sequence length (max_seq_length) defines the maximum text length the model can process. If this value is too short, important information may be cut off. If it is too long, the computational burden increases. Based on the distribution of sentence lengths in the dataset, it was set to 256. The learning rate influences the training speed and convergence of the model. If the learning rate is too high, training may become unstable. If it is too low, training may proceed slowly or get stuck in a local minimum. For the BERT layer, a smaller learning rate is typically used. It was set to 5 × 10⁻⁵. The hidden size (hidden_size) and number of layers (num_layers) in the BiLSTM layer affect the model’s capacity and complexity. Larger or more layers can improve performance. However, they also increase the risk of overfitting and computational requirements. In this study, the hidden size was set to 256 and the number of layers to 2, which are commonly used parameter sizes. The batch size (batch_size) affects training stability and memory usage. A larger batch size can increase stability but requires more memory. Based on the experimental hardware constraints, it was set to 64. The number of epochs (epochs) determines how thoroughly the model is trained. Too few epochs may lead to underfitting, while too many may result in overfitting. However, overfitting can be prevented with the early stopping strategy. Early stopping monitors the performance on the validation set. Training stops when performance does not significantly improve over a set number of epochs. If an effective early stopping strategy is implemented, a higher number of epochs (e.g., 100) can provide ample training space. The actual training epochs will be determined by early stopping. The dropout parameter is used to prevent overfitting during training. Dropout is applied to the BiLSTM layer to improve the model’s generalization ability. A typical dropout value ranges from 0.1 to 0.5. In this study, it was set to 0.3.

Since the same corpus is used in this study, the max_seq_length of the BERT-CNN model was set to 256, in line with the entity extraction model parameters. The batch size was set to 64. The model achieved the best performance when the number of epochs was set to 10. To improve training speed, the learning rate was set to 3 × 10⁻⁵. In the convolutional layer configuration, kernel sizes were set to [3,4,5]. This helps the model capture dependencies in texts of varying lengths. To reduce overfitting, the dropout rate was set to 0.3, and the regularization coefficient was set to 0.005.

To validate the superiority of the BERT-BiLSTM-CRF model, a comparison was made with the traditional BiLSTM-CRF model to analyze the performance of named entity recognition. In this study, entity recognition is performed for nine entity types using both models, and the corresponding precision (P), recall (R), and F1 scores are calculated. The comparison results reveal that the BERT-BiLSTM-CRF model outperforms the BiLSTM-CRF model in entity recognition. Except for the entities related to equipment, property damage, and personal injury, the precision of the BERT-BiLSTM-CRF model is higher by more than 0.06 in all other cases. This improvement is attributed to the fact that these three entity types have standardized nominal expressions, with less ambiguity. This suggests that the BERT-BiLSTM-CRF model demonstrates superior performance, as shown in

Table 5. Experimental results of named entity recognition models.

Entity Type	BERT-BiLSTM-CRF			BiLSTM-CRF
	Precision	Recall Rate	F1	Precision	Recall Rate	F1
Spatial information	0.783	0.792	0.787	0.704	0.759	0.730
Person factors	0.823	0.835	0.829	0.718	0.517	0.601
Equipment factors	0.776	0.793	0.784	0.770	0.876	0.820
Management factors	0.798	0.770	0.784	0.699	0.656	0.677
Property damage	0.867	0.879	0.873	0.847	0.847	0.847
Personal harm	0.881	0.893	0.887	0.862	0.852	0.857
Knowledge of accident control	0.867	0.879	0.873	0.736	0.691	0.713
Knowledge of accident recovery	0.804	0.816	0.810	0.738	0.831	0.782
Emergency disposal materials	0.842	0.854	0.848	0.634	0.665	0.649

. The BERT- CNN model for entity relationship classification is presented in

Table 6. Experimental results of entity relationship classification recognition model.

Relationship Type	Precision	Recall Rate	F1
Location of occurrence	0.819	0.826	0.823
Cause	0.814	0.755	0.785
Lead to	0.847	0.821	0.834
Decide	0.827	0.872	0.849
Use	0.788	0.782	0.785
Control knowledge	0.781	0.816	0.798
Recover Knowledge	0.756	0.831	0.792
Demand	0.808	0.790	0.799

. Based on the results returned by the model, it can be concluded that overall, satisfactory performance was achieved, with an F1 score of approximately 0.8

The following is an example that was used to verify the predictive capability of the model. Due to space constraints, only the sequence of events is provided: At 17:30 on × Month ×, 2008, an intermediate ventilation shaft excavation and transport operation was underway at a subway construction site, alongside jet grouting pile construction. The rear support equipment for the jet grouting pile operation was mixing cement slurry. Large quantities of cement were utilized at the site, resulting in numerous empty cement bags. Rainy conditions made the ground slippery. Worker A, who was responsible for clearing debris, slipped while barefoot during the cleanup of empty cement bags. Other workers noticed that A did not immediately stand up and rushed to assist. At that point, A was still conscious. After workers reported the incident to the project department, personnel immediately dialed 120 for emergency medical assistance. On-site CPR was initiated, while simultaneously notifying the project manager and contacting A’s relatives. At 17:50, hospital paramedics arrived at the site, administered emergency care, and transported A to the nearest hospital. The ambulance delivered the patient at 18:15, where A was pronounced dead after unsuccessful resuscitation efforts. The following day, the hospital issued a death certificate citing sudden death due to electric shock.

The entity extraction results from the model were as follows:

(1) Entities: Subway construction site, worker A (barefoot), improperly matched residual current device, inadequate temporary electrical management by project department, sudden death by electric shock, assisting, resuscitation, rescue.

(2) Relationships: Location of occurrence, caused, control knowledge.

The above example demonstrates that the entity recognition model performs well in identifying entities such as “personnel harm,” but struggles with entities such as “management factors.” This discrepancy may have stemmed from the former being more intuitive, while the latter often involved multiple factors and specialized technical terminology. Regarding the entity-relationship recognition model, it likely excelled at identifying direct relationships like “location of occurrence” and “caused,” as these typically involved clear entities and causal connections. However, the model may have faced challenges with abstract relationships like “control knowledge” or “cause,” particularly when descriptions were non-standard or contained negative information, which may have impacted accuracy. Overall, the model’s performance validated the feasibility of the knowledge extraction method proposed in this study, providing reliable data input for subsequent knowledge graph construction.

5. Results and Applications

5.1. Extraction Results

The BERT-BiLSTM-CRF model was employed to identify entity types that are numerous and primarily distributed in unstructured text. For entity relationship classification and extraction, the BERT-CNN was applied. As a result, 4877 entities and 5783 relationships were extracted. Building on the extracted triples, the subsequent subsections will sequentially demonstrate Knowledge graph visualization, knowledge association, Knowledge retrieval, and question and answer service.

5.2. Knowledge Graph Visualization

Among numerous graph database tools, Neo4j is recognized as the most widely used graph database system worldwide [34]. It is characterized by several advantages, including high read/write efficiency, a relatively gentle learning curve, a mature and user-friendly interface, abundant learning resources, and support for REST API interfaces. Based on these advantages, Neo4j was selected as the storage and query platform for the emergency response knowledge graph of safety accidents, as shown in Figure 6.

5.3. Knowledge Association

Knowledge graphs are known for their powerful visualization capabilities, as they can clearly present the overall knowledge characteristics of the emergency response to safety accidents and the knowledge representation structure of the relationships between nodes in the form of directed graphs. Through the nodes and relationships of knowledge graphs, knowledge units such as accident types, times, and locations contained in safety accidents are linked together, thereby transforming flat, static, scattered knowledge units into three-dimensional, dynamic knowledge networks that facilitate the mining of implicit knowledge. In this study, knowledge graphs were utilized to achieve instance-based visualization of emergency response knowledge. By clicking on nodes, users can view attribute information, obtain knowledge associations efficiently, and thereby gain a deeper understanding of the knowledge.

5.4. Knowledge Retrieval

The emergency response knowledge graph was designed to provide retrieval services, with different storage methods resulting in different retrieval approaches. In Neo4j, retrieval is performed using the Cypher query language, which is regarded as more convenient and efficient compared with the structured query language used in traditional relational databases. The retrieval functionality of the knowledge graph enabled users to construct appropriate query statements according to their needs, thereby allowing them to obtain the required nodes and relationships accurately and efficiently.

As an illustrative example, the case of “electrocution accidents caused by human factors” was retrieved using the following Cypher query: MATCH (humanFactor: human factor)-[r: cause]->(electricShock: electrocution accident) RETURN humanFactor, electricShock. The relevant nodes and relationships were directly displayed, and the query results are presented in Figure 7. By examining the human factors contributing to electric shock incidents and identifying the causes of these factors, construction practices could be standardized, and managers were enabled to detect and control unsafe behaviors that might lead to accidents. In addition, the knowledge graph was demonstrated to provide users with multi-dimensional entry points for retrieval, thereby enabling comprehensive and multi-perspective access to emergency response information. This functionality contributed to the enhancement of safety management levels.

5.5. Question and Answer Service

This study used large language models as an enhancement tool for knowledge graph-based question answering. In a knowledge graph-based question answering system, the system architecture is divided into three main layers: the interaction layer, the logic layer, and the data layer. The interaction layer is responsible for providing the user interface, enabling seamless interaction between users and system functions through front-end design. The logic layer includes various methods and models, such as entity recognition models, question intent recognition models, graph reasoning and retrieval methods, and graph database query retrieval statements. Additionally, it involves invoking the ChatGLM2-6B large language model to process user queries and provide necessary API service interfaces to the application layer. The Data Layer consists of the non-relational database Neo4j and the relational database MySQL. The former is responsible for the persistent storage of knowledge graphs, while the latter handles the storage of user information. The collaborative operation of these three layers ensures that the system can respond to user needs in an efficient and accurate manner.

To validate the practical utility of the question-answering system developed in this study, it was evaluated using the M Construction Project. Examples of the knowledge requirements for various emergency response organizations within the project are presented in

Table 7. Overview of Knowledge Requirements for Emergency Response Organizations (Example).

No.	Emergency Response Organizations	Knowledge Requirements (Example)
1	On-site Emergency Command Team	What government departments should be reported to when an accident occurs on-site?
2	Hazard Rescue and Assistance Team	What supplies do the staff at the scene of the accident require for rescue?
3	Medical Rescue Team	How should the injuries caused by the accident be treated?
4	Safety and Security Team	What supplies are needed for establishing the on-site emergency command and the rescue operation?
5	Communications Liaison Team	What information should be reported when coordinating with local medical and emergency services?
6	Post-accident Prevention Team	How should rescue work be carried out in a low-temperature environment?
7	Post-accident Handling Team	How should the injury diagnosis be carried out?
8	Production Recovery Team	What tools are required for on-site cleaning after the accident?

. For identical emergency response knowledge needs, three distinct methods were employed to acquire information: the question-answering system developed in this study, a traditional search engine, and manual retrieval of accident contingency plans and on-site handling protocols. The query times and the validity of evaluation results for all three methods were statistically compared, with the results summarized in

Table 8. Statistical Analysis of Case Study Results.

No.	Response Time (s)			Validity of Answers
	Q&A System	Search Engine	Textual Materials	Q&A System	Search Engine	Textual Materials
1	<30 s	<30 s	>60 s	Accurate	Accurate	Accurate
2	<30 s	>60 s	>60 s	Accurate	Accurate	Accurate
3	<30 s	<30 s	>60 s	Accurate	Accurate	Accurate
4	<30 s	>60 s	>60 s	Accurate	Relatively accurate	Accurate
5	<30 s	>60 s	>60 s	Accurate	Accurate	Accurate
6	<30 s	>60 s	>60 s	Accurate	Accurate	Accurate
7	<30 s	>60 s	>60 s	Accurate	Relatively accurate	Accurate
8	<30 s	>60 s	>60 s	Accurate	Relatively accurate	Accurate

. Under the same query conditions, the response time of the question-answering system is on average less than 30 s, while traditional search engines and manual retrieval mostly require more than 60 s to return results. The findings demonstrate that the developed question-answering system exhibits significant expertise and efficiency in the domain of emergency response to construction safety accidents. Furthermore, the system supports graph-based relationship queries, assisting personnel who face challenges in articulating problems using specific terminology. In comparison with traditional web search engines, this system offers clear advantages in both query speed and accuracy. However, due to the inherent limitations of the knowledge graph, the system is unable to provide answers when knowledge demands exceed its predefined scope.

6. Conclusions

The construction industry, as a labor-intensive sector, employs many workers; however, safety management practices often remain inadequate, leading to frequent accidents that pose serious threats to the lives and property of workers. The fragmentation of emergency response knowledge and the inefficiency of knowledge acquisition among relevant personnel have significantly hindered the effectiveness of emergency response in construction safety accidents. To address these challenges, scenario-response theory and knowledge graph technology were applied in this study to the domain of emergency response for construction safety accidents. Guided by scenario-response theory, the construction of the knowledge graph was achieved, enabling the integration of emergency response knowledge in the field.

Specifically, the scenario elements and emergency response processes of construction safety accidents were analyzed, and a domain knowledge graph for construction safety accident emergency response was constructed. First, based on relevant theoretical research, 4 categories of scenario domains describing construction safety accidents were identified, and 14 influential scenario elements were extracted from domain-specific materials. Second, by decomposing the emergency response process into tasks, the knowledge requirements for emergency response were examined at the accident control and recovery stages. Based on the scenario-response theory, a mapping relationship between scenario elements and emergency response knowledge was established. Using the seven-step knowledge graph modeling method, the 14 scenario elements were defined as entity types, and 10 relationship types and 30 relationships were determined, thereby forming a domain knowledge structure model that provided the theoretical foundation for knowledge graph construction. In terms of knowledge extraction, a combined manual and machine-based method was adopted according to the characteristics of different entity types, whereas for entity types with broad instance ranges and less clear definitions, the BERT-BiLSTM-CRF model was employed for entity extraction. In addition, the BERT- CNN model was utilized for entity relationship classification. The results of the model performance verification indicate that high accuracy was demonstrated by both models in the recognition task. Although the BERT-BiLSTM-CRF and BERT-CNN models generally perform well in most cases, certain limitations remain when processing cross-sentence dependencies and longer texts. Particularly in accident reports involving multi-level contextual information, the models are prone to errors when capturing long-range dependencies and complex relationships. Future improvements could involve the integration of Transformer models to better capture long-distance dependencies in the text. Finally, based on the extracted knowledge triples, the Neo4j graph database was selected for data storage, and a domain knowledge graph was constructed to present knowledge in a structured and visual manner, thereby providing a reliable foundation for knowledge application in emergency response to construction safety accidents. It validates the feasibility and effectiveness of the proposed method in practical emergency response scenarios. The question-answering system developed in this study utilizes a knowledge graph-based retrieval mechanism. Based on the experimental results from the M Construction Project, more precise and rapid emergency response support is provided to relevant personnel by this system.

Due to various constraints, certain unavoidable limitations are present in this study. Regarding the scope of construction safety accident types, only the six most frequent accident categories reported by the Ministry of Housing and Urban-Rural Development were selected, ensuring practical relevance and focus. However, this approach overlooks other accident types, particularly rare yet severe incidents. This omission may affect the comprehensiveness of the knowledge graph and the applicability of the emergency response Q&A system, limiting its capability to address uncommon accident scenarios. The scope of accident types covered could be expanded in future research to accommodate a broader range of application scenarios.

The knowledge extraction employed in this study primarily targeted textual materials and has not yet incorporated multimodal data, such as images and videos. This limitation reduces the system’s ability to process diverse forms of information in actual on-site emergency situations. Therefore, future research could focus on refining multimodal knowledge extraction methods for domain-specific image or audio-visual data sources, thereby broadening the scope of knowledge and enhancing domain expertise.