PageRank Algorithm-Based Recommendation System for Construction Safety Guidelines

Abstract

The construction industry faces significant challenges with frequent accidents, largely due to the inefficient use of safety guidelines. These guidelines, which are often text and figure heavy, demand substantial human effort to identify the most relevant items for specific tasks and conditions. Additionally, the guidelines contain both central and peripheral elements, and central items are critical yet difficult to identify without extensive domain knowledge. This study proposes a novel recommendation framework to enhance the usability of these safety guidelines. By leveraging natural language processing (NLP) and knowledge graph (KG) modeling techniques, unstructured safety texts are transformed into a structured, interconnected KG. The PageRank and Louvain Clustering algorithm is then employed to rank guidelines by their relevance and importance. A case study on “High-rise Building Construction (General) Safety and Health Guidelines”, using ‘scaffolding’ as the keyword, demonstrates the framework’s effectiveness in improving retrieval efficiency and practical application. The analysis highlighted key clusters such as ‘fall’, ‘drop’, and ‘scaffolding’, with critical safety measures identified through their interconnections. This research not only overcomes the fragmentation of safety management documents but also contributes to advancing hazard analysis and risk prevention practices in construction management.

1. Introduction

The construction industry is a massive and labor-intensive sector [1,2]. It makes significant contributions to the economy, particularly in terms of the GDP [3,4,5]. However, the construction industry is inherently high risk, characterized by frequent and severe accidents [6,7,8]. The likelihood of fatal accidents is higher compared to other industries [9,10]. In Korea, the frequency of construction accidents is increasing despite strengthened safety regulations such as the “Serious Accidents Punishment Act” enacted in 2021 [11,12]. Small- and medium-sized construction sites are particularly affected due to insufficient safety management capabilities and a lack of dedicated on-site safety managers, compounded by rising construction costs [13,14,15]. Additionally, there is minimal compliance with safety documentation requirements as outlined in the “Occupational Safety and Health Act” and the “Construction Technology Promotion Act” [16,17].

Advanced construction safety practices in countries like the United States, Australia, and the United Kingdom mandate comprehensive Job Hazard Analysis (JHA) to document risk reduction measures at the task level, thereby enhancing safety in high-risk environments [18,19]. Government agencies in these countries develop and publish standardized safety guidelines to promote best practices across the industry [20]. For example, the Occupational Safety and Health Administration (OSHA) in the United States publishes detailed standards and regulations, such as those for fall protection and scaffolding, to ensure worker safety at construction sites [21]. Similarly, Safe Work Australia publishes Codes of Practice for various sectors, providing clear safety management techniques and criteria [22]. The Korea Occupational Safety and Health Agency (KOSHA) has also been publishing technical guidelines for JHA since 2010, emphasizing risk identification and mitigation during job execution [23].

Despite these efforts, a significant challenge lies in the fragmented and unstructured nature of these safety guidelines, which are typically presented in lengthy PDF documents [24,25]. This unstructured format complicates the understanding of the relationships between guidelines and their practical applications [26,27,28]. The extensive length and unstructured format of these documents pose a barrier for professionals trying to implement the guidelines effectively in their daily work. Consequently, there is a need for more streamlined and accessible methods for the dissemination and utilization of these guidelines to achieve their intended goals [29,30].

This research aims to address the challenge of fragmented safety management guidelines by integrating them into a unified knowledge graph (KG). Using natural language processing (NLP) and KG modeling techniques, the text in safety guideline documents is transformed into a structured, interconnected KG. PageRank is applied to rank the guidelines by their relevance and importance. The primary goal of this paper is to develop a framework for a recommendation system that enables safety managers to easily access the most relevant and central guideline items for specific construction work types and conditions, thereby improving the efficiency of construction hazard analysis and risk prevention practices.

The subsequent sections of this paper are organized as follows: a literature review of relevant studies, an explanation of the NLP and KG modeling processes, the construction and validation of the recommendation system through case studies, and a discussion of the research’s significance and limitations. This structured approach ensures a comprehensive understanding of the proposed system and its potential impact on construction safety management.

2. Literature Review

2.1. Previous Works on Ontologies and Knowledge Graphs for Construction Safety Management

In the realm of construction safety, several researchers have undertaken studies on ontology modeling and knowledge base development to enhance the efficiency and effectiveness of safety management processes [31,32,33]. Wang & Boukamp (2011) [34] were among the first to introduce a novel framework for JHA. Their framework restructured JHA documents hierarchically and developed a semantic extraction process to automate construction safety management. This framework incorporated ontology modeling to transform structured knowledge into machine-readable formats using RDF graph modeling. Additionally, the inference process facilitated the rapid generation of new JHAs, providing timely solutions for field safety managers. This approach enabled the reuse and transformation of existing construction safety knowledge into new formats, marking a significant advancement in safety knowledge management.

Zhong et al. (2015) [35] advanced the field further by combining ontology modeling and semantic inference to enhance construction risk and safety management. They introduced a meta-ontology model that integrated knowledge domains across different classes and utilized Semantic Web technologies to understand and infer relationships among various knowledge elements. By constructing a network model that encompassed risk factors and construction process elements, their framework contributed to improved construction process management and facilitated more effective construction risk analysis. This work underscored the importance of contextualized information and the power of the Semantic Web in leveraging knowledge resources within the construction domain.

Gao et al. (2022) [36] took a different approach by introducing a KG model to address construction safety and health issues. Their research focused on creating an ontology model-based pattern-matching system for decision-making purposes. They established the structure of the ontology model and stored a comprehensive set of 34 technical standards and manuals under respective labels for practical use. The pattern-matching techniques enabled the extraction of relevant information from the existing database, thus providing valuable assistance in safety management decision-making processes. These studies highlight the application of ontology modeling and KG models in addressing construction safety and management challenges, integrating diverse knowledge domains, and facilitating inference mechanisms to support decision making within the industry.

More recent research has seen the integration of NLP, machine learning, computer vision [37], and deep learning with KGs in construction safety management [32,38]. Zhong et al. (2020) [39] developed an automated model that combined computer vision and knowledge safety modeling. Their research identified risks through image processing and modeled risk scenarios using attribute-based graphs, incorporating over 3000 safety hazard reports. By leveraging the information extracted from image processing, they formulated KG queries to acquire risk-related information and identify risk factors. Similarly, Wang and El-Gohary (2023) [40] proposed an on-site suitability inspection framework using deep learning techniques to automatically extract domain-specific relationships from construction safety regulations. They represented the extracted requirements as query graphs, employing deep learning models to understand word relationships and structure the query graph accordingly. This framework led to the development of software that aids in evaluating compliance with safety regulations at construction sites.

Below,

Table 1. Cases of ontological modeling of construction safety management knowledge.

Authors	Data	Methodology	Purpose
Wang and Boukamp (2011) [34]	JHA document	Directly modeled JHA documents in XML format	Transform JHA concepts into ontology models and utilize these models for guidance inference
Zhang et al. (2022) [41]	Construction site photos	Leveraged the BERT model for constructing a graph-based model, establishing relationships between entities identified using computer vision techniques	Analyze objects extracted through computer vision using network modeling to identify potential hazards to workers
Gao et al. (2022) [36]	34 standards and technical manuals, including domestic and international design, construction, and management regulations	Constructed a KG model tailored to the labeling requirements of the network model, accurately representing the manual’s contents	Develop an ontology-based network model to facilitate knowledge sharing and reuse, deriving construction safety guideline outcomes through rule-based graph pattern matching
Wang and El-Gohary (2023) [40]	A set of 20 OSHA safety regulations specifically focused on fall prevention	Performed KG modeling following classification using a deep learning model, assigning specific classes to edges in network relationships	Automatically extract relationships between pieces of information using a KG and generate new information from the stored database
Zhong et al. (2020) [39]	Chinese Specification Quality and Safety Inspection Guide of Urban Rail Transit Engineering	Generated semantic annotations on conceptual entities for computer comprehension of images, processing these annotations using natural language techniques, and further developing the KG	Provide diverse reports and predictions regarding risk situations
Lu et al. (2015) [42]	Typical provisions from Occupational Safety and Health Administration	Constructed an ontology model through rule-based reasoning to capture domain knowledge and facilitate automated inference in the Semantic Web context	Store safety regulation clauses to assist in safety management decision-making systems and implement a real-time alert system for workers to enhance safety awareness and response
Adebayo Isah and Kim (2023) [43]	The construction risk registers of the Townsville Ocean Terminal Project in Queensland, Australia	Modeled hazardous knowledge in a KG to capture intricate relationships and interconnections within this domain	Facilitate informed decision-making processes regarding risk assessment and mitigation strategies by providing a comprehensive inventory of potential hazards

presents a summary of the most relevant research studies that have introduced KGs and ontology models to construct and utilize databases for construction safety guidelines.

2.2. Previous Works on Text Mining and Natural Language Processing Approaches for Construction Safety Guidelines

The application of NLP to construction safety guidelines has significantly improved how safety information is managed in the construction industry [44,45]. Zhang et al. (2019) [46] used NLP to verify compliance with construction safety guidelines by extracting features based on grammatical and semantic characteristics of the text. Their approach identified key attributes such as comparative relations, quantities, subjects, and compliance-checking attributes within sentences, facilitating a more systematic analysis of safety documents.

Chi et al. (2016) [47] developed a machine learning-based text classification model to address fall prevention topics. By analyzing data from OSHA standards and field records, their model effectively categorized safety-related information, enhancing the ability to predict and prevent fall-related incidents. This work highlighted the importance of leveraging machine learning techniques to improve the classification and utilization of safety guidelines.

Further advancements include the integration of NLP with other technologies to enhance the analysis of safety guidelines. Xu et al. (2022) [48] utilized text mining techniques combined with deep learning to analyze construction incident reports, specifically focusing on hot work accidents. Their study aimed to uncover recurring themes and patterns to provide valuable insights into common safety issues. The research emphasized the importance of systematically reviewing incident data to inform safety practices and mitigate future risks. They used the Latent Dirichlet Allocation model to extract key topics from accident reports and convolutional neural networks to predict accident causes. This work underscores the potential of NLP to transform unstructured incident reports into structured data, enhancing safety management practices by providing actionable insights for improving process safety management in hot work operations.

2.3. Knowledge Gaps

Despite significant advancements in construction safety management, the existing research has primarily concentrated on the extraction and analysis of safety information from documents using NLP and KG modeling. These efforts have enhanced the accessibility and organization of safety information but have not adequately addressed the recommendation of specific safety guidelines based on the extracted data. Most studies focus on transforming declarative and explicit knowledge into machine-readable formats without providing actionable insights or prioritizing guidelines for safety managers.

The inherent limitations of current ontology models and KGs, often represented using RDF syntax, include a lack of contextual understanding and the inability to capture shared elements across different knowledge domains. This limitation hinders the development of robust inference and recommendation processes that can assist on-site construction managers with essential tasks such as JHA and regular safety inspections.

Moreover, the absence of integrated recommendation systems that prioritize the most critical safety guidelines limits the practical utility of the extensive safety documents developed through significant effort and resources. While the existing research on construction safety guidelines primarily focuses on information extraction and classification, it often overlooks the necessity of building an integrated database that can support the recommendation of specific safety measures or the identification of the most relevant guideline items for various construction tasks. Furthermore, current approaches tend to overlook the need for interconnected safety guidelines, where individual safety measures are treated in isolation rather than as part of a larger, inter-related system of practices. The absence of this interconnected approach prevents safety managers from recognizing patterns of risk and implementing preventive measures that are informed by multiple safety considerations simultaneously.

To address these deficiencies, this paper proposes a novel framework for a recommendation system designed to select the most relevant construction safety guidelines for various work types and conditions.

3. Proposed Framework

3.1. Outline of the Proposed Framework

Figure 1 depicts a flowchart outlining the framework for the KG modeling and recommendation system proposed in this study. The framework encompasses three main steps: preprocessing of the construction safety guidelines (Step 1), KG model creation (Step 2), and the application of ranking and clustering algorithms for finding the safety guideline items that are most important and relevant (Step 3). By following this process, a comprehensive KG model based on relationships between guideline items within the construction safety guidelines can be built, aiding in the identification of key guideline items. Each step is detailed in the subsequent subsections, describing the specific actions taken according to the proposed framework.

3.2. Step 1: Preprocessing the Guideline Text

The text data used in this research were extracted from eighty-six construction safety guideline documents provided by the KOHSA. These guidelines are categorized by the type of work, such as bridge construction, building demolition, etc., and are originally available in a PDF.

Table 2. Example of title/category/content structure (content translated from the original Korean text for presentation).

Title	Category	Content
Safety and Health Guidelines for Suspension Bridge Construction Works	Review points for establishing a work plan	Once the overall conditions of the construction site and the specifications outlined in the design documents have been examined, it is necessary to develop a specific work plan that aligns with the designated work procedures for each phase. It is important to ensure that there are no issues with the use of construction equipment. Afterward, this detailed work plan must be prepared and submitted for approval by the supervisory agency.
		The work plan document should be prepared by someone with ample experience and knowledge in the construction of bridges. It is crucial to periodically check and confirm whether the contents of the plan are being properly executed during the construction phase.
	Anchorage safety operations	The scale of concrete pouring should be planned, taking into account the production and pouring capacity of concrete, aiming to prevent the occurrence of detrimental cracks in the structure.
		As the structure involves reinforced concrete, it is important to consider workability and temperature crack control when deciding the vertical and horizontal divisions of the concrete.

shows the construction safety guideline data converted into CSV format, containing the title, category, and content columns, separated by statements. A total of 86 construction safety guideline documents were converted into a table format consisting of 5988 rows.

Next, the Soynlp library was used to tokenize the content column. Soynlp, a Korean NLP library, provides functions for extracting words and nouns using unsupervised learning approaches [49]. The tokenized items were divided into L-tokens (nouns) and R-tokens (conjunctions, complements, particles, etc.), with the L-tokens selected for constructing the KG, as they include key concepts and technical terms from the guidelines.

The TF-IDF weight function was then used to extract the index of each token (i.e., keywords) within the sentences based on their TF-IDF scores [50,51]. The n_gram parameter was configured as 1 to 3 to account for frequently recurring key terms that form a cohesive meaning at the sentence level. For each entry in the content column, the TF-IDF function calculated the weights of individual tokens, with higher-weighted tokens indicating their importance. These tokens were stored as indices in the index column, facilitating the mapping of links between items in the KG.

3.3. Step 2: Constructing a Knowledge Graph Representing the Guidelines

A KG is a data model that uses nodes, edges, and properties to store information in a network-like structure [52]. In this research, Neo4j, a widely used platform for graph data science and analytics, was employed to build the KG and apply algorithms using the Cypher language. Neo4j Desktop version 5.15.0 was used in this study, and the Graph Data Science Library version 2.7.0 was utilized to implement the graph algorithms. Broadly speaking, there are two types of KG models: property graph models and RDF (Resource Description Framework, Triplet Schema) graphs [53]. The RDF graph model is beneficial for representing complex concepts by using the RDF triple structure to express subjects, predicates, and objects [54]. On the other hand, the property graph model is used to model relationships between data, making it advantageous for querying and data analysis tasks based on these relationships [55]. Cypher is one of the most popular languages for creating and querying property graph models [56].

The KG representing the KOSHA Safety Guidelines was constructed by linking all items based on the hierarchy of the documents and the keywords shared among the items (Figure 2) [57]. The KG has three types of nodes: category, content, and index. Category nodes represent the subtitles of each guideline, forming the highest level in the hierarchical structure. Content nodes capture specific guideline items under each category, while tokens extracted from each content node through the TF-IDF process create individual index nodes. These nodes are hierarchically connected based on the inclusion relation, with the edges titled “includes”.

3.4. Step 3: Developing a Safety Guideline Recommendation Mechanism Based on PageRank Algorithm

To identify the most relevant guideline items, a graph-based analysis algorithm was applied to the KG. The recommendation system aimed to select central guideline items within the network, implying their relevance and importance due to their inter-related nature. The PageRank algorithm and Louvain Clustering were used to achieve this objective.

The PageRank algorithm, which is widely used for ranking web pages based on their importance within a network, assigns a numerical value (PageRank) to each node in the graph [58,59,60]. It calculates the PageRank value based on incoming links from other nodes, with higher values indicating greater importance. This iterative algorithm updates PageRank values until convergence, redistributing the importance across the nodes based on the graph’s link structure.

Specifically, a PageRank score can be calculated using the following equation:

$$PR\left(A\right)=\left(1-d\right)+d({\displaystyle \frac{PR\left(T_1\right)}{C\left(T_1\right)}}+\cdots +{\displaystyle \frac{PR\left(T_n\right)}{C\left(T_n\right)}})$$

(1)

where

• $PR\left(A\right)$: the PageRank value of node A;
• $PR\left(T_n\right)$: the PageRank value of a target node of node A;
• $C\left(T_n\right)$: the number of outgoing links from each node;
• $d$: 0.85 (damping factor).

In construction safety management, organizing and categorizing safety guidelines into coherent groups is useful for practical application. A clustering algorithm is essential to effectively group related guidelines, revealing underlying structures and relationships within the KG. This facilitates the identification of key guideline items and improves the recommendation process, ensuring safety managers can access and apply the most relevant safety measures for specific construction scenarios. This approach enhances the retrieval efficiency and ensures comprehensive, contextually relevant guideline recommendations.

Louvain Clustering, used for this purpose in this research, partitions nodes into multiple communities, maximizing internal connectivity within each community. The algorithm optimizes a quality function known as modularity, which evaluates the degree of community structure division. Through iterative evaluation and node movement, the Louvain Clustering algorithm achieves the partition of the graph into communities with high internal connectivity and low connectivity between different communities [61,62].

The modularity (Q) is calculated using the following equations.

$$Q={\displaystyle \frac{1}{2m}}\sum _{i,j}\left[A_{ij}-{\displaystyle \frac{k_i{k}_j}{2m}}\right]\delta \left(c_i,c_j\right)$$

(2)

$$∆Q=\left[{\displaystyle \frac{\sum _{in}+2k_{i,in}}{2m}}-{\left({\displaystyle \frac{\sum _{tot}+k_i}{2m}}\right)}^2\right]-[{\displaystyle \frac{{\mathsf{\Sigma }}_{in}}{2m}}-{\left({\displaystyle \frac{{\mathsf{\Sigma }}_{tot}}{2m}}\right)}^2-{\left({\displaystyle \frac{k_i}{2m}}\right)}^2]$$

(3)

where

• $A_{ij}$: the weight of the link between nodes i and j;
• $k_i$: the sum of weights of all links connected to node I;
• $\delta \left(c_i,c_j\right)$: indicator function that equals 1 if nodes i and j belongs to the same community and 0 otherwise;
• $m$: the total sum of link weights in the graph;
• $∆Q$: change in modularity.

To explore communities within a network, the algorithm searches for optimal modularity. However, the direct calculation of modularity can be computationally expensive. Therefore, a hierarchical approach is often used, where modularity is evaluated at different levels of community structure. The Louvain Clustering algorithm iteratively evaluates ΔQ (cf. Equation (3)) for each node and moves nodes between communities if it results in a positive change in modularity. This process is repeated until no further improvement in modularity can be achieved. The result of this iterative operation is a partition of the graph into communities that have relatively high internal connectivity and low connectivity between different communities [61,62].

4. Illustrative Case Study

4.1. Knowledge Graph Generation

The generated KG (Figure 3), constructed using the Neo4j Python library, is structured based on predefined relationships [63]. In Figure 3, the category nodes are depicted in yellow, content nodes in blue, and index nodes in red. The relationships between nodes are represented by the include edges. Overall, this KG structure offers a systematic and interconnected representation of guidelines, enabling effective traversal and exploration within the domain it encompasses. It organizes information hierarchically, with category nodes at the top level representing the subtitles of each guideline. Each category node is linked to relevant content through content nodes. Within the content nodes, essential details such as the title, category, content, and index are stored as property values, providing important metadata that facilitate efficient retrieval and analysis. To establish connections between the various elements of the KG, tokens extracted from each content node are used to create index nodes. These index nodes serve as references and are linked to their corresponding content nodes through include edges.

During the graph generation process, a total of 669 category nodes, 5988 content nodes, and 102,923 index nodes are extracted. To enhance the graph’s connectivity and prepare it for further analysis, a preprocessing step is conducted. This step involves establishing connections between content nodes, creating a web of relationships among the information entities. These connections are represented by Relate edges, which signify shared indexes or commonalities between content nodes. The weight of each Relate edge is determined by the number of shared indexes, indicating the strength of the relationship. Upon completion of the graph construction and preprocessing, the resulting network comprises 217,220 include edges, representing inclusions or associations between different elements, such as including a specific guideline within a broader category. Additionally, there are 13,301,903 Relate edges, reflecting the interconnectedness between content nodes based on shared indices. This graph structure enables advanced analysis, such as information retrieval, similarity-based recommendation systems, and graph-based algorithms like PageRank, to extract valuable insights and support decision-making processes in construction safety and risk management domains.

Hypothetically, a field manager selects a suitable title pertaining to the task they are currently involved in or planning to undertake. They proceed by conducting a search using a keyword relevant to the task. Subsequently, the graph retrieves the corresponding category, content, index entries from the KG based on the provided title and keyword. Furthermore, it also projects the relationship between each content (Relate) node and the relationships connecting category– content– index (include). The PageRank algorithm is employed to analyze the relationship between content – [:Relate] – content within the current projected graph.

4.2. Projected Knowledge Graph and Calculation of Node Attributes

Before applying the graph algorithm, one scholarly assumption made was that highly shared indexes within the projected graph are deemed significant. The provided diagram represents a graph projection of the content and index associated with ‘scaffolding’ within the “High-rise Construction (General) Safety and Health Work Guidelines”. Subsequently, the graph algorithm is applied to this projected graph (Figure 4).

Then, PageRank is computed based on the shared keyword count, which serves as the weight for the Relate relationship within the same label. The PageRank values in the projected graph range from a minimum of 0.5 to a maximum of 1.44. A low PageRank value indicates a weaker connectivity within the projected graph, implying a lower number of shared indexes. Content items with a limited number of shared indexes are likely to contain more localized or specific information. As a consequence, they may have limited representation in the search results, as they may not align with the broader or more universal safety guidelines sought by field managers.

Next, the Louvain algorithm is utilized to project the graph using the content – [:Include] – index relationship. Each cluster primarily represents the modularity among the contents, indicating that contents within the same cluster possess a significant number of shared keywords and exhibit high similarity. In the current projected graph, a total of 26 contents have been detected, forming 12 clusters. The results of each graph algorithm are typically stored as properties of the content nodes. In the projected graph, the number of contents within each cluster ranges from a minimum of 1 to a maximum of 6.

Clusters with fewer contents generally exhibit a lower number of shared indices within the projected graph, indicating reduced similarity with other content items. Consequently, clusters containing only one content item are likely to include more specialized or specific instructions, which diverges from the study’s goal of delivering more universal safety guidelines. To meet the objective of providing field managers with the most relevant and central safety guidelines, it is essential to concentrate on clusters with a higher number of content items. These clusters are more likely to feature shared indices and demonstrate greater similarity among their contents.

4.3. Safety Guidelines Recommended Based on Ranking by PageRank Algorithm

Table 3. List of recommended safety guideline items ranked by PageRank algorithm for “High-rise Building Construction”.

Cluster	PageRank	Content
4098	1.14818	To prevent tripping hazards, it is necessary to install toe boards around the working platform and seal the gap between the working platform and the concrete surface with rubber mats or similar materials to prevent any gaps from occurring. This helps to prevent accidents and ensure safety on the job site.
4098	1.09787	When there is a concern that the instantaneous wind speed may exceed 30 m per second, measures should be taken to minimize the wind pressure on the scaffolding. This includes removing any windbreakers or barriers installed on the sides to ensure the minimum wind pressure acting on the scaffolding. Additionally, any materials or objects loaded on the working platform should be securely fastened or removed to prevent them from being blown away. However, if the structural calculations indicate that the scaffolding can withstand higher wind speeds, exceptions can be made in such cases.
4098	1.08622	During the ascent of the scaffolding, it is important to adhere to the specified hoisting speed and lifting length as stated in the manual. Before commencing the ascent, the condition of the climbing shoe and wall shoe anchor installations, as well as the hydraulic cylinder, should be checked for any abnormalities. If any issues are identified, they should be rectified before proceeding with the work. Once the ascent is completed, the installation condition should be inspected and recorded for documentation purposes.
4098	1.08519	Around the scaffolding and on the working platform, the storage of flammable materials and the use of open flames should be prohibited.
4098	1.0307	During the ascent of the scaffolding, it is essential to follow the prescribed work sequence and methods outlined in the manual. Additionally, it is important to designate a competent supervisor or manager to oversee and supervise the progress of the work
4098	0.97971	The initial risk assessment results for falling objects should be incorporated into the installation plans and design of safety facilities, and feedback should be provided.
3821	1.44173	On the topmost level of external structures, it is recommended to install primary preventive measures such as a cocoon system to prevent falling incidents. Alternatively, if a single-level fall protection system, such as a fall prevention net or guardrail, is installed within the falling risk area based on the review of falling scenarios and simulations, the installation of fall protection systems can be omitted in lower levels. In such cases, vertical protection nets or other fall prevention facilities can be installed instead. For high-rise construction projects, the fall prevention net should be installed within a vertical distance of 9 m from the working level of the concrete structure, steel frame floor, or the lower level of the dismantling work area. The net should extend horizontally for a minimum distance of 3 m. It should be designed and installed in advance to maintain sufficient protection against falling objects such as materials and tools and to withstand wind loads safely. The vertical protection nets on each level should be installed tightly without any gaps to eliminate the risk of falling or slipping due to gusty winds. In the case of pre–emptive measures such as an SCN (Safety Coverage Net) installed on the uppermost or upper levels, integral falling object prevention nets or guardrails can be attached to the upper portion, ensuring compliance with the minimum horizontal distance within the falling risk area. In such cases, a structural examination and assembly drawing should be prepared for the entire system, including the attached preventive measures.
3821	1.38124	For the installation plan of a fall prevention net for high-rise construction, the following guidelines should be followed: A fall prevention net for high-rise construction should be installed when carrying out construction activities (new construction, renovation, dismantling, etc.) on the external walls of buildings at a height of 6 floors or 22 m and above. However, if pre–emptive measures for falling prevention have already been installed prior to the work on that specific level, the installation of the fall prevention net can be excluded. The fall prevention net should be installed within a distance of 9 m below each floor where concrete structure scaffolding assembly/dismantling work or steel frame structure concrete pouring work has been completed. The horizontal extension length of the fall prevention net should be at least 3 m from the edge of the building.
3821	1.3252	The fall prevention and protection facilities should be designed based on the safety verification documentation provided by the manufacturer in their manual. An architect or structural engineer should refer to this documentation, along with considering the site conditions, results of fall simulations, wind loads, and other load conditions, to conduct a structural review. Based on this review, a detailed design document should be prepared, which includes assembly drawings, safety procedures, and other relevant information. Furthermore, the design document should undergo a secondary review by an architect or structural engineer, who may be a construction company or a supervisor (including a construction manager). After their approval, the installation of the fall prevention and protection facilities can proceed
3750	1.07323	Compliance with site conditions: It should be ensured that that the applied load conditions in the structural analysis align with the actual site conditions. This includes considering factors such as the weight of the scaffolding, live loads during work, applied loads from materials, wind pressure, horizontal loads, temperature loads, etc. Safety of members in load combinations: The safety of structural members should be assessed by comparing the maximum stress generated in each member with the material-specific safety factors, considering uncertainties related to material properties and repetitive use, among other factors. Structural review at each ascending stage: Structural reviews should be conducted at each stage of scaffolding ascent, ensuring that the structural integrity is maintained and addressing any potential issues or concerns Safety assurance in dismantling: Safety should be ensured during the dismantling process by considering the methods, sequence, and proper order of dismantling to avoid any hazards or instability
3750	1.0307	When ascending the scaffolding, it is crucial to confirm whether the concrete at the anchor bolt locations has achieved a strength equal to or greater than the required strength specified in the structural review. Only after confirming the concrete strength should the scaffolding ascent proceed.
2701	1.25972	The scaffolding manual must include the following information: Wind speed used in the structural calculations, excluding the major equipment. Permissible load for users, excluding the major equipment. Dimensions, materials, capacity, and quantities of key components such as anchors, shoes, rails, hydraulic cylinders, etc. Scaffolding ascent procedures, including ascent speed, length of each ascent, and the method and sequence of ascent. Required concrete strength for anchor support points and reinforcement methods if necessary. Installation and dismantling procedures for scaffolding. Other safety precautions and guidelines that users should be aware of.
2701	1.09534	The fall prevention and protection plan should prioritize the installation of pre–emptive measures such as the cocoon system, which is a hypothetical assembly placed on the topmost level to prevent falling objects from occurring. This includes considering the installation of pre–emptive measures for fall prevention as a primary approach in the planning process.

presents the results of applying the PageRank algorithm to the projected graph derived from the “High–rise Building Construction (General) Safety and Health Guidelines” using the keyword ‘scaffolding.’ The table highlights content items with PageRank values of 0.9 or higher within clusters that contain at least two content items. Among the 26 content items extracted from the graph, a subset of 13 was selected based on a combination of PageRank values and Louvain algorithm results. The PageRank algorithm assessed each content item’s importance and connectivity based on shared keyword counts, while the Louvain algorithm identified clusters with high modularity, indicating strong similarities among content items. This integration of PageRank and Louvain algorithms resulted in a refined selection of 13 content items, highlighting their significance within the graph and their association with relevant clusters. Content items with high PageRank values often cover prevalent safety topics such as ‘fall,’ ‘drop,’ and ‘scaffolding,’ which are crucial in various safety contexts. Content within the same cluster typically shares themes related to ‘fall,’ ‘scaffolding,’ and ‘structure,’ reflecting thematic coherence. By leveraging PageRank and clustering techniques, a more targeted and relevant subset of content items was extracted, facilitating efficient information retrieval and a deeper understanding of key topics in the “High-rise Building Construction” safety guidelines.

5. Discussion

The conversion of construction safety guidelines into a KG offers significant advantages by integrating diverse PDF documents into a unified database. This approach enhances the efficiency of retrieving relevant safety information and facilitates the incorporation of various types of data that may be otherwise inaccessible to on-site safety managers. The structured nature of KGs allows for the establishment of relationships between different safety guidelines and measures, resulting in a more comprehensive and interconnected understanding of safety procedures. This structure supports the generation of precise and contextually relevant recommendations tailored to specific construction projects.

The problem of identifying the most central and relevant safety guideline items among a vast collection of documents is akin to the challenge of finding the most relevant webpages among numerous options on the Internet. The PageRank algorithm, originally developed by Google’s founders for ranking webpages by their relevance and importance, is adapted in this research to address a similar problem within the context of construction safety management. Unlike the World Wide Web, where webpages are interconnected via hyperlinks, construction safety guidelines are primarily text based, often in a PDF format. Therefore, it is necessary to first construct a web-like structure (i.e., knowledge graph) of safety guideline items. In this KG, nodes representing guideline items are connected by indexes, similar to keywords in hyperlinks. By applying the PageRank algorithm to this network, the central and relevant safety guideline items can be identified, employing a comparable methodology to that used for ranking webpages.

The case study results show that items with high PageRank values focus on key aspects of accident prevention and worksite safety. For instance, measures to prevent tripping hazards around scaffolding, minimize risks from wind pressure, and ensure adherence to safety procedures during scaffold ascent are highlighted. These high-PageRank items are directly related to preventing accidents in high-rise construction and are strongly connected to various other guidelines. In particular, guidelines for fall prevention, wind hazard protection, and safety checks during scaffold installation and dismantling rank highly, indicating their importance due to their strong interconnections with other guidelines. By applying PageRank, users can quickly identify the most critical safety measures, which are highly linked to multiple other guidelines, allowing for better focus on key safety topics such as fall prevention, wind protection, and scaffolding stability.

The Louvain Clustering algorithm groups similar safety guidelines into clusters based on their content and context. In this analysis, 13 key clusters were identified, with keywords like ‘fall,’ ‘scaffolding,’ and ‘structure’ frequently appearing together. For example, Cluster 4098 focuses on scaffolding safety, including guidelines for safe scaffold ascent and the prevention of storing flammable materials near scaffolds. Cluster 3821, on the other hand, emphasizes fall prevention, particularly the installation of fall protection systems in high-rise construction. These clusters highlight how closely related topics, such as scaffolding safety and fall protection, are interconnected. By grouping related guidelines, the Louvain algorithm provides a systematic way to organize safety content, making it easier for users to navigate through similar topics. This ensures that crucial safety measures, like those addressing scaffolding hazards and fall protection, are comprehensively addressed within their respective clusters, enhancing the overall safety management process.

This approach is particularly applicable to construction safety guidelines, where important terms are frequently repeated without significant omissions. By providing specific keywords and guideline titles, users enable the analysis of relevant content, categories, and indexes. The graph projection, obtained from the provided data, is subjected to the PageRank algorithm and the Louvain Clustering algorithm. The Louvain Clustering algorithm effectively groups inter-related graph components, while the PageRank algorithm quantifies the importance of nodes within each cluster. This process facilitates the identification of key content within each cluster. By employing these algorithms, the recommendation and search system effectively identifies and groups related guidelines, enabling access to pertinent content based on their input. This scholarly approach enhances the accessibility and usability of construction safety guidelines, contributing to improved decision making and enhanced safety management in the construction industry.

The system’s capacity to dynamically update with new guidelines and best practices ensures scalability and adaptability, offering access to the latest safety information. In a large construction site, safety managers can monitor real-time conditions and respond immediately using the KG-based recommendation system. For instance, when newly installed scaffolding is exposed to higher-than-expected winds, the system, leveraging the PageRank algorithm, can prioritize safety guidelines related to wind hazards, such as removing windbreaks or checking the scaffolding’s secure fastening. This allows managers to take immediate preventive action, reducing the risk of accidents. Additionally, the system provides tailored safety guidelines based on the specific phase of a project. For example, during the installation of steel structures in high-rise buildings, the Louvain Clustering algorithm can group relevant guidelines on fall prevention and recommend the most suitable safety measures for that particular phase, ensuring optimal safety and efficiency. In other words, this system fosters continuous improvement by updating managers with current best practices and safety technologies, promoting collaboration and knowledge sharing among safety professionals, and ultimately raising safety standards across the construction industry.

This study acknowledges several limitations that may affect the reliability of its findings. Firstly, the construction safety guidelines used were derived from data spanning 2010 to 2023, which, while comprehensive across various construction types and scenarios, may not fully incorporate the most recent safety measures. Consequently, the extracted statements might not reflect the latest and most specific safety practices. Secondly, the use of index-based algorithms poses inherent challenges. The indexes generated using Korean NLP tools might not accurately capture the specialized terminology used in the construction industry. Additionally, the frequent occurrence of less significant indexes could lead to inflated PageRank values or inaccurate groupings in the Louvain Clustering algorithm, potentially skewing the identification of key statements. These limitations suggest the need for caution in interpreting the results and highlight the importance of integrating up-to-date data and more precise NLP techniques tailored to the construction sector in future research.

6. Conclusions

This study introduces a novel KG approach aimed at organizing and systematizing construction safety guidelines within construction projects, with the additional application of this approach to a recommendation system. Leveraging the PageRank algorithm and the Louvain Clustering algorithm, the KG is designed to efficiently extract essential information from the graph database, thereby enabling the retrieval of safety-related information relevant to construction trades. Through the implementation of the proposed KG, effective collaboration between on-site safety managers and the database is facilitated, subsequently offering valuable support in safety management processes.

In line with this approach, the study further highlights the increased importance of safety managers’ roles following the implementation of the Major Accident Punishment Act in Korea. The PageRank algorithm, in particular, identifies high-priority items essential for accident prevention and worksite safety in high-rise construction, such as mitigating tripping hazards around scaffolding, managing wind pressure risks, and ensuring safety protocols are followed during scaffold ascent. These high-PageRank items are directly linked to accident prevention and are closely connected to other safety guidelines. Additionally, the Louvain Clustering algorithm organizes similar safety statements into clusters based on modularity, such as Cluster 4098, which focuses on scaffolding safety, and Cluster 3821, which centers on fall prevention measures in high-rise construction. This clustering enables the systematic navigation of related safety guidelines, providing comprehensive coverage of interconnected risks. By transforming construction safety guidelines into a knowledge graph and utilizing graph algorithms for recommendations, the study offers a promising solution to reduce the workload of safety managers, streamlining the retrieval of relevant guidelines and saving time in decision making and operational activities.

While the proposed recommendation system provides valuable insights, it requires more rigorous field testing to assess its real-world applicability. Future work should focus on enhancing the recommendation algorithms to account for contextual factors, such as the type and conditions of construction tasks, rather than solely relying on inter-relationships between safety guideline items. Improving the system’s contextual awareness will enable more precise recommendations tailored to specific construction scenarios. Moreover, future research could benefit from integrating heterogeneous data sources, such as accident reports and legal provisions, into the KG database. This integration would enrich the recommendation system, providing a more comprehensive and contextually relevant set of guidelines. Future advancements in AI-based NLP, including more sophisticated morphological analysis and tokenization techniques, as well as AI-driven question-answering processes utilizing Retrieval-Augmented Generation (RAG) pipelines, could further enhance the accuracy and functionality of the recommendation system.