Natural Language Processing Risk Assessment Application Developed for Marble Quarries

Abstract

In this study, by using the texts describing the hazards and precautions taken during text mining, the necessary processes were carried out to first estimate the probability value and severity value of the risk and then calculate the risk values by Natural Language Processing analysis. In order to be used within the scope of the study, two data sets were generated from the data in the risk assessment report prepared by applying the L-type matrix risk assessment in marble quarries between 2015 and 2021. Stochastic Gradient Descent (SGD) was used for classification and prediction by analyzing text data. One data set was used to analyze the probability value of the risk and the other was used to analyze the severity value of the risk. In light of the results, when a text containing hazard and precaution information was entered, a system was developed that analyzed this text, estimated the probability and severity values, and calculated the risk assessment score. The application of the SGD algorithm to learning models developed on text data yielded an accuracy rate of 91.2% in the risk probability data set and 97.5% in the risk severity data set. The results indicated that the models were capable of conducting automatic risk assessment on text data and of effectively predicting the requisite probability and severity values. Due to the high accuracy rates obtained during the study, this risk assessment software was recommended for use in marble quarries.

1. Introduction

Marble mining, a substantial sub-branch of the mining sector, has an important place in Türkiye. There are more than 2000 mines in Türkiye, and more than 600 kinds of marble have been found. In 2018, Türkiye had 40 unique kinds of marble and the largest variety of marble in the world [1]. Furthermore, marble mining has made significant contributions to Türkiye’s economy in terms of both employment and foreign exchange earnings [2]. Nevertheless, along with these positive effects, marble quarrying involves various dangers and risks: collapse, falls, accidents caused by machinery and equipment, explosions, respiratory problems caused by dust, and exposure to loud noise [3,4]. Risk is defined as the possibility of loss, injury or other harmful consequences arising from a hazard [5]. The types of occupational accidents in marble quarries generally vary depending on the working conditions of the quarries, the equipment utilized, and the occupational safety standards [6,7,8]. The aforementioned dangers and risks need to be identified and assessed, and precautions need to be proactively taken before an occupational accident occurred [9,10]. For this purpose, a risk assessment method needs to be used [11]. In studies performed in the mining sector, decision matrix risk assessment was used to assess hazard risks in underground mining by Janjuhah et al. [12]. In one of the recent studies, the method proposed by Spanidis et al. [13] was a combination of the analytical hierarchy process (AHP), the quantification of risk factors, and the preference ranking technique based on the similarity to the ideal solution (TOPSIS) method for ranking restoration alternatives based on the low-risk approach. In another recent study, Spanidis et al. [14] proposed a methodology for risk assessment of natural hazards in surface mining projects, using the triangular Fuzzy Analytic Hierarchy Process (FAHP) to determine the probability of risk occurrence, the Expected Value (EV) function, the Monte Carlo simulation, and the Program Evaluation Review Technique (PERT) method to make cost and time overrun estimates. Korshunow et al. [15] improved an occupational risk management approach and basically tried to put forward a matrix method for occupational risk assessment based on a new mining enterprise. Tripathy [16] intended to identify and evaluate mining hazards and risks using different risk assessment techniques, such as Failure Mode and Effect Analysis (FMEA) and Workplace Risk Assessment and Control (WRAC), to improve mining safety. Kemajl et al. [17] outlined a methodology for accident prediction and risk assessment as a case study at Stanterg Mine. Matloob and Khan [18] emphasized the unidentified hazards during mining and found solutions to these hazards using Artificial Intelligence and Machine Learning. Esmaeilzadeh et al. [19] investigated and performed hazard risk assessment utilizing the Failure Modes and Effects Analysis (FMEA) method for incidents and risks emerging in quarries in West Azerbaijan. Stojance et al. [20] used the fuzzy TOPSIS method for workplace risk assessment in an underground lead and zinc mine and compared the results with the numbers of injuries and accidents occurring at individual workplaces to assess its accuracy.

In studies carried out on risk assessment in the literature, it was noted that subjects such as Probability [21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36], Severity [25,30,37,38,39,40,41,42,43,44], Detectability [29,44,45,46,47], Exposure [21,22,31,46,48], Consequence [21,22,31,49,50], Undetectability [33,49,51,52,53], Cost [29,30,45,54] and Impact [55,56,57] were considered to be risk criteria. The methods utilized to determine the risk criteria in these studies were the Fuzzy AHP, AHP, BWM, SWARA, SWARA, Expert judgment, Fuzzy cognitive map, Direct given, ITARA, Set-valued statistics, Combination weighting, QFD, Multi-criteria Weighting, and BRAW methods (1) [5]. In addition, the methods used in risk assessment methods were Fuzzy sets (TFNs) [24,29,33,37,39,40,46,48,58,59,60,61,62,63,64,65,66], Pythagorean fuzzy sets [52,67,68,69], Interval-valued Pythagorean fuzzy sets [23,26,70], Fuzzy sets (Trapezoidal fuzzy numbers [71,72,73] Interval type-2 fuzzy sets [38,74], Hesitant fuzzy linguistic term sets [34,75], Grey linguistic term sets [76], Hesitant fuzzy sets [75], Bipolar fuzzy sets [77], Q-rung orthopair uncertain linguistic sets [75], Fermatean fuzzy sets in 2019 [78], Spherical fuzzy sets [46], Linguistic spherical fuzzy sets [53], Picture fuzzy sets [53], D numbers [79], Neutrosophic sets [47], Intuitionistic fuzzy sets [80], Interval-valued intuitionistic fuzzy sets [36], Interval 2-Tuple linguistic variables [31], 3,4-quasirung fuzzy sets [39], Fermatean fuzzy linguistic sets [43], and Rough sets (1) [81]. In these studies, risk assessment, Fine Kinney [21,22,33,36,62,82,83,84,85,86,87,88,89,90], Failure mode and effect analysis (FMEA) [26,44,91,92,93,94,95,96], L type matrix [97,98], Hazard and Operability Methodology (HAZOP) [32,35,99,100,101,102], Fault Tree Analysis (FTA) [100,103,104], Bayesian network [29,88,105,106,107], Event tree [105,108,109], and Bow-tie [110] methods were used. Studies in which risk score was estimated were Adem et al. in 2018 [111], Gnoni and Bragatto in 2013 [112], Janjuhah et al. [12], Moatari-Kazerouni et al., Zhou et al. [113,114], Oliveira et al., Ozbakır [115,116], Halabi et al. [117], Ridwan et al. [118], Tadic et al., Jafari et al. [66,119], Oturakci, Yang [64,120], Mohandes and Zhang, Mohandes et al. [48,121], Golinko et al. [122], Mo [29,79], and Gul and Ak [27]. Reviewing the studies, it became clear that there were not many studies on risk assessment methods in marble quarries.

In this study, taking into consideration the risk assessment report made by using the L-type risk assessment method in a marble quarry, the required processes were conducted to estimate the probability and severity values that constituted the risk with Natural Language Processing analysis and to calculate the risk scores, utilizing the texts identifying the hazards and the precautions taken.

2. Materials and Methods

2.1. Risk Assessment Process

In the marble quarry risk assessment process used within the scope of the study area, primarily workplace-specific hazards and risks were established. Afterward, field inspections and interviews with employees were performed and the outcomes obtained were evaluated.

The L-type matrix (5 × 5) method was used as the risk assessment method. In this method, the risk value was found by multiplying the probability and severity values (Equilibrium (1)):

Risk (R) = Likelihood (P) × Severity (S)

(1)

Using the L-type matrix method, scoring was performed in accordance with the information presented in

Table 1. Likelihood ratings (S) [122].

Score	Probability	Categorization
5	Very frequent (Once a week, every day), under normal working conditions	Almost certain
4	Frequent (Monthly)	Likely
3	Occasional (A few events in a year)	Possible
2	Remote (Once a year), only in abnormal conditions	Unlikely
1	Hardly ever (Once a year)	Rare

and

Table 2. Severity (degree of damage) ratings (P) [122].

Score	Severity	Result
5	Death, permanent total disability	Catastrophic
4	Major injury, requiring long-term treatment and therapy, occupational disease	Major
3	Minor injury, requiring inpatient treatment	Moderate
2	No loss of working days, requiring outpatient treatment without a lasting impact and requiring first aid	Minor
1	No loss of working hours and requiring first aid	Insignificant

.

As seen in Table 1, in the L-type matrix method, the probability of the damage occurring was ranked with scores from 1 to 5. The probabilities were described as daily, weekly, monthly, every three to six months, and very low. The explanations made against these probabilities were stated as very high, high, medium, low, and almost never. In Table 2, in the L-type matrix method, the effects of the hazard are listed from Severity 1 to 5. The severity was grouped as multiple deaths, death or permanent disability, very large material and environmental pollution, severe injury, occupational disease, major material damage, injuries requiring at least three days of rest, material damage, minor injuries requiring first aid, material damage beyond repair, accidents that did not lead to damage or injury, and no loss of working hours. The severity of the hazard was determined to be very serious, serious, medium, slight, or very slight. Afterward, in the L-type matrix method, the probability value for each hazard in the study area was determined with the assistance of Table 1, the severity value with Table 2, and the risk scores were presented with the assistance of Equilibrium (1), as seen in

Table 3. 5 × 5 L-type risk decision matrix [97,98].

Risk Matrix		Severity (S)
		Insignificant	Minor	Moderate	Major	Catastrophic
Likelihood (P)	Rare 1	Low 1	Low 2	Low 3	Low 4	Low 5
	Unlikely 2	Low 2	Low 4	Low 6	Moderate 8	Moderate 10
	Possible 3	Low 3	Low 6	Moderate 9	Moderate 12	Extreme 15
	Likely 4	Low 4	Moderate 8	Moderate 12	Extreme 16	Extreme 20
	Almost certain 5	Low 5	Moderate 10	Extreme 15	Extreme 20	Extreme 25

. Hence, the damage effect degree was found by multiplying the probability and severity levels. In addition, the approval of sector experts was obtained in the creation of the matrix and the risk assessment.

In the L-type matrix method, the risk value action and timetable are presented in

Table 4. Risk value, action, and time planning [123].

Risk Value	Action and Time Planning
25	Not tolerable risk. It is the risk group that it is not accepted to start work without any measures.
15–20	Significant risk. It is the risk group that is extremely important and should be taken measures immediately
8–12	Moderate risk. They are significant risks that need to be taken measures in the short term
4–6	Tolerable risk. It is a tolerable risk group that requires attention in the long term.
1–3	Insignificant risk. Risks that do not matter much and can be accepted.

. The risk values were numbered from 1 to 25. Explanatory information was given for each value in the action and time planning.

Unacceptable Risk: The activity cannot be continued without improvement in the subjects that scored 15, 16, 20, 25 points as a result of the evaluation. Strict measures shall be taken, and intensive training shall be implemented. In case of a violation, a fine will be applied.

Considerable Risk: Improvement is absolutely required in the subjects that scored 8, 9, 10, or 12 points as a result of the evaluation. The activity may be continued with temporary measures where the improvement is performed. Precautionary measures and policies shall be determined.

Acceptable Risk: Improvements in subjects that scored 1, 2, 3, 4, 5, or 6 points as a result of the evaluation are evaluated within the framework of financial and technological means. Improvements are made in subjects that are deemed appropriate. They are among the acceptable risks. Regular training shall be procured on the subject.

Insignificant risks: 1. There is no need to plan control processes to eliminate the identified risks and to keep echoes of the activities to be carried out.

2.2. Data Science Process

The aim of this study was to estimate the risk probability value and the risk severity value by analyzing the texts defining the hazards and the precautions taken utilizing natural language processing techniques. Therefore, the processes needed for the calculation of risk values were conducted meticulously.

Within the extent of the study, data collected from risk assessment reports prepared by a marble quarry were brought together, and two different data sets were created. The first of these data sets was used for the analysis of the risk probability value, and the other was reserved for the analysis of the risk severity value.

The risk probability value specified the probability of a risk occurring, whereas the risk severity value measured the level of damage that could occur.

3. Results

3.1. Experimental Study

In order to be utilized within the scope of the study, two data sets were produced by combining the data in the risk assessment reports made by occupational safety experts in a marble quarry between 2015 and 2021. One of the data sets was employed for the analysis of the risk probability value, and the other for the analysis of the risk severity value. Summary representations of the data sets are demonstrated in

Table 5. Summary of the data set produced for risk probability value analysis.

Id	Text	Class
1	Garbage should be collected in leak-proof garbage bags and lidded garbage bins. The lids of garbage bins should always be kept closed.	1
2	PSYCHOLOGICAL RISK FACTORS—Unplanned work.	2
3	PRODUCTION AREA—Machinery equipment falling at height stone falling from height.	2
…	…	…
305	There should be enough plates, forks, spoons, glasses, etc. for all employees using the cafeteria at the same time. The use of uncleaned materials by employees should be prevented.	1
306	PRODUCTION AREA/WORKING WITH THE DRILLING MACHINE—Stone flying metal spatter.	3

and

Table 6. Summary of the data set produced for the analysis of risk severity values.

Id	Text	Class
1	PSYCHOLOGICAL RISK FACTORS—Task uncertainty.	2
2	The insulation of mobile electrical cables should not be with a damaged or attached device, and terminal connections should not be left open and should be properly insulated.	5
3	BUSINESS GARDEN—Inappropriate parking spaces in the business/improper parking.	4
…	…	…
305	CHEMICAL RISK FACTORS—Waste oils.	3
306	Before starting, drilling machine care should be taken to ensure that they are fixed correctly, and the strengths of the nails and ropes should be checked.	4

.

As seen from Table 5 and Table 6, both data sets consist of 306 rows of data. To ensure consistency in the analysis outcomes, the same hazard and precaution expressions were utilized as independent attributes in the data sets. The dependent attribute was marked with the Class label in the data sets. As the probability and severity values would be estimated separately according to the specified hazard and precaution expressions, these values were planned to form class attributes in two separate data sets.

To establish the risk score of a hazard before and after the precaution, the probability and severity values had to be identified. The multiplication operation was implemented to the acquired probability and severity values and the risk score, and the risk level was increased to Moderate. For this reason, the objective of the study was to predict the risk probability and risk severity values as accurately as possible. To estimate the probability and severity values, the preprocessing and creation of data, the testing of the learning algorithm, and the estimation of the results were performed in the data analysis phase.

3.2. Data Preprocessing

Data preprocessing, also named data preparation, is the process of processing raw data and cleaning, manipulating, and reorganizing it prior to analysis. This step typically includes formatting, adjusting, and integrating to improve the information contained in data sets. Data preprocessing is an essential step to prepare data for processing and to mitigate the probability of bias; however, it might be a laborious task [124].

To make our data ready for natural language processing analysis, the data were subjected to various preprocessing processes:

• Our data were gathered in a way that creates data sets with certain orders.
• Data were augmented to enable enrichment and inter-class balance adjustment in the collected data.
• A data cleaning process was implemented on text data by eliminating non-word and non-space punctuation marks and special characters, converting all letters to lowercase, identifying unnecessary words, and removing them.

At the end of these processes, the data were made ready for analysis.

Figure 1 and Figure 2 were reviewed as examples of how text data looked before and after preprocessing.

To create a near-normal order in the distribution of class labels in the data sets, following the data augmentation processes applied to both data sets, the numbers of rows in the data sets were updated to 454 for the risk probability data set and 669 for the risk severity data set. The distributions of class labels in the current data sets are given in Figure 3 and Figure 4.

With the other preprocessing steps applied to the text data, the number of words in the text data to be analyzed in the current version of the Risk Probability data set was reduced from 9968 words to 5786 words. The number of words to be analyzed in the current version of the Risk Severity data set was reduced from 16,692 words to 9388 words. This data reduction process was conducted by eliminating stop words, which were considered ineffective elements in the text data.

3.3. Data Analysis

By analyzing the independent variable text data available in the data sets, the Natural Language Processing (NLP) technique was utilized for the process of estimating class labels. Natural Language Processing (NLP) is a field focusing on the interactions between computers and human language, allowing computers to understand, interpret, and produce natural language [125]. Recently, NLP is one of the most substantial technologies. Comprehending complex structures in language and extracting insights and actions from them is of vital importance for artificial intelligence. The significance of natural language processing in various areas is enormous and constantly increasing because digital information in the form of language is widespread [126].

In the data sets, the Stochastic Gradient Descent (SGD) algorithm was used to carry out classifications and predictions by analyzing text data. The SGD algorithm performed a great simplification process in gradient descent processes and worked depending on randomly selected examples in each iteration. SGD algorithm did not require remembering the examples observed in previous iteration stages. Therefore, the algorithm could analyze all the examples in the deployed system instantly. In this case, SGD could optimize the expected error rate, as the examples to be analyzed were randomly selected from the basic distribution [127,128]. SGD has become quite popular in machine learning applications recently. It has a serious role in performing convex and non-convex optimizations. The SGD algorithm could be utilized in machine learning applications; it has a lower cost than the stochastic approach due to the extreme costs of full gradient calculations in objective functions [129]. In this study, the learning models for NLP were created on the generated data sets using the SGD algorithm. The purpose of utilizing these models was to predict the risk probability and risk severity values by inserting the hazard or precaution information into the system. After these values were estimated, the risk score could be calculated, and the risk could be graded.

Table 7. Complexity matrix of the risk probability data set analysis results.

	Forecast Values
Actual Values			2	3	4
		2	49	1	2
		3	0	37	7
		4	0	2	39

and

Table 8. Complexity matrix of the risk severity data set analysis results.

	Forecast Values
Actual Values		2	3	4	5
	2	41	0	0	0
	3	0	48	3	0
	4	0	0	54	0
	5	0	0	2	53

demonstrate the complexity matrices of the learning models created as an outcome of the analysis processes performed on the data sets, whereas

Table 9. Risk probability data set analysis results.

Class	Precision	Recall	F1-Score
1	1.00	0.94	0.97
2	0.93	0.84	0.88
3	0.81	0.95	0.88
Accuracy:			0.912

and

Table 10. Risk severity data set analysis results.

Class	Precision	Recall	F1-Score
2	1.00	1.00	1.00
3	1.00	0.94	0.97
4	0.92	1.00	0.96
5	1.00	0.96	0.98
Accuracy:			0.975

demonstrate the analysis outcomes of the data sets.

Based on an examination of the outcomes in the tables showing the complexity matrices and analysis outcomes, the learning models created on text data with the SGD algorithm ensured an accuracy rate of 91.2% in the risk probability data set and 97.5% in the risk severity data set. These results indicated that the created learning models successfully estimated the probability and severity values required for automatic risk assessment on text data.

In line with the results obtained, a structure was created that predicted the likelihood and severity values by performing an analysis by inserting the text expression indicating the hazard and precaution statuses. Then, the risk assessment score was calculated, and the risk level of the entered text expression was changed to Moderate (Figure 5). Within the scope of the study, the attained risk assessment software was recommended for use in marble quarries due to the Extreme success in the accuracy rates of the outcomes.

4. Discussion

The results of this study demonstrate the successful application of Natural Language Processing (NLP) and Text Mining techniques for risk assessment in marble quarries. The study developed a machine learning-based system that predicts the probability and severity of risks based on hazard and precaution descriptions, achieving high accuracy rates of 91.2% for probability and 97.5% for severity. These results demonstrate the potential of NLP-driven risk assessment systems to improve occupational safety by automating the identification and assessment of hazards.

Compared to traditional risk assessment methods, such as manual inspections and expert judgements, the NLP-based system reduces the reliance on human expertise and the time required to perform the assessment. Previous studies in the mining industry have used various risk assessment approaches such as the Fine–Kinney [21,33,36,62,82,83,84,85,86,87,88,89,90], Failure Mode and Effects Analysis (FMEA) [26,44,91,92,93,94,95,96], and Fuzzy Analytical Hierarchy Process (FAHP) [13,14] methods. However, these methods often involve manual analysis or expert judgement, which can create subjectivity and inconsistency in risk assessment. The integration of NLP into risk assessment offers a more objective, consistent and scalable solution for analyzing data sets with large textual information.

The engineering applications of this system can be extended beyond marble quarries to other sub-areas of mining, including underground mining and coal mining. This wider applicability is supported by the system’s ability to process textual data from any risk assessment report, making it versatile in different sectors where occupational hazards need to be systematically assessed.

After evaluating the outcomes obtained, a marble quarry in Türkiye was analyzed, and risk assessment reports made with the L-type matrix method were employed within the scope of the study. In the studies where other risk values were addressed [12,39,48,64,65,79,85,111,113,115,122,123], no study was detected investigating a marble quarry, which is a mining sector. The studies mostly considered petrochemical plants [112,130], underground mines [12], the production sector in 2017 [113], hydroelectric power plants [123,131], the processing industry [65,66,132], textile enterprises [29,64], gas plants [88], tank production processes [26], hospitals [48], oil stations [86], construction sites [45,66], bus drivers [122], wind turbines [79], and marble factories [39].

Moreover, in this study, probability and severity values were estimated using NLP and Text Mining methods. In previous studies, a decision matrix risk assessment [12,48,115], which is a hybrid estimation tool for production systems, was developed by combining the characteristics and strengths of a systematic risk analysis method to control major accident hazards in an industrial environment [117]. Furthermore, a hazard identification risk assessment risk control (HIRARC) model [118,123] was made. A fuzzy model for risk assessment was designed [65,66,132]. In addition, a model was created to determine the ranking of hazards with the AHP method [64,120]. A comprehensive model using fuzzy ANP, grey relational analysis, and interval-valued Pythagorean fuzzy TOPSIS [70] was designed. Finally, a modified risk score method [122] and BWM method [39] were employed.

Studies in which the risk score was predicted include Adem et al. in 2018 [111], Gnoni and Bragatto in 2013 [113], Janjuhah et al. [12], Moatari-Kazerouni et al., Zhou et al. [113,114], Oliveira et al., Ozbakir [115,116], Halabi et al. [117], Saedi et al., Ridwan et al. [118,123], Tadic et al., Jafari et al. [66,121], Oturakci, Yang [64,120], Mohandes and Zhang, Mohandes et al. [70,121], Golinko et al. [122], Mo, [79], Gul and Ak [39], and Spanidis et al. [13,14].

Adem et al. in 2018 [111], Moatari-Kazerouni et al., and Zhou et al. [113,114] used the hesitant fuzzy linguistic term sets. Oturakcı, Yang [64,120], Mohandes and Zhang [70], Stefanovic et al. in 2017 [65], Gul and Ak [39], and Gul et al. in 2017 [51] used fuzzy sets (TFNs). Mo [79] used D numbers by utilizing the risk assessment reports made by the authorized engineers of the marble quarry in this study.

Oturakçı, Yang [64,120], Golinko et al. [122], and Mo [79] estimated only hazard values, and Gul and Ak [37] utilized severity values in their studies. This study estimated both probability and severity values that created the risk score.

Kantas Yılmaz and Karakus [98] and Ates and Albayrak [97] implemented risk assessment in two different laboratories by using the L-type matrix method, whereas in this study, the data in a risk assessment method were considered and the probability and severity values were estimated by using NLP and Text Mining methods. Shekhar et al. [133] also analyzed the Directorate General of Mine Safety (DGMS) fatality reports for non-coal mines in India using NLP and Text mining. This method eliminated the large number of labor hours spent on reviewing each report and the technical expertise required to analyze mining accidents. Similarly, with the method used in this study, risk assessment values and probability and severity values were estimated with high accuracy to carry out the risk assessment study in the marble quarry without the need for much time or an expert.

Moatari-Kazerouni et al. [113] reached approximately 70% accuracy in the outcomes attained with the risk estimation tool, while this study reached 91% accuracy in risk probability and 97% accuracy in risk severity.

A notable advantage of this system is its capacity for continuous improvement with the addition of new data. As more hazard and precaution reports are processed, the model can be refined to potentially improve the prediction accuracy even further. Furthermore, as only Stochastic Gradient Descent (SGD) was used in the current study, future work could focus on comparing different machine learning algorithms to optimize performance.

Despite the high accuracy rates achieved, the study has limitations. The current application is specific to marble quarries, and the extent to which the system can generalize to other industries requires further investigation. Additionally, while the NLP-based model eliminates some of the manual efforts associated with traditional methods, it still requires high-quality textual data, which might not always be readily available or well-documented in every industrial setting. Addressing this limitation will require the development of a more robust data collection process or the integration of additional data sources to enrich the model’s learning capability.

As a result, compared to other traditional methods, the inaccurate or incomplete assessments made by the L-type matrix risk assessment method using NLP and Text Mining methods are eliminated; thus, it helps to perform accurate and detailed risk assessment, which is easier to apply in the sub-branches of any mining industry.

After the identification of risks in the risk assessment process in the engineering application, the risk scores of the risks identified using the proposed model can be calculated easily, with high accuracy and without wasting time. Accordingly, the planning phase of the measures to be taken by the relevant authorities can be started immediately. Once the actions to be taken have been determined, the proposed model can be re-run on the action expressions, and the potential impact on the risk score can be calculated quickly with high accuracy. As a result, with the use of the proposed model, the ratings of the risks that may occur in mines can be calculated automatically much faster and with a higher level of accuracy than manual calculations. In this process, it can significantly benefit the authorities in terms of time and accuracy.

Furthermore, future work is planned to establish a program to facilitate the application of this method.

5. Conclusions

By analyzing the texts identifying the hazards and precautions taken in risk assessment reports of a marble quarry utilizing natural language processing techniques, the probability and severity values creating the risk were predicted. Within this study, the data attained from the risk assessment reports prepared by occupational safety experts and employees in the marble quarry between 2015 and 2021 were brought together, and two different data sets were created. The first of these data sets was used for the analysis of the probability value, whereas the other was reserved for the analysis of the risk severity value. A model that could estimate risk scores was created due to the gathered data. Learning models developed on text data using the SGD algorithm achieved an accuracy rate of 91.2% in the risk probability data set and 97.5% in the risk severity data set. These outcomes suggested that the models had the ability to conduct automatic risk assessment on text data and could effectively predict the required probability and severity values. Hence, when a text expression indicating the hazard and precaution statuses was inserted, a system was developed that predicted the probability and severity values. The risk assessment score was calculated by performing model analysis. This system classified the risk level of the entered text expression as Moderate. Taking into consideration the outcomes obtained in the study, the developed risk assessment software was recommended for use in marble quarries due to the extreme success in accuracy rates. In future studies, this model will be applied to other risk assessment methods.