Coal mine safety production field knowledge graph construction method

Abstract

The application discloses a knowledge graph construction method in the field of coal mine safety production, and relates to the technical field of coal mine safety production, and comprises the following steps: performing causal semantic analysis on a coal mine safety production accident text, constructing a speculative expression judgment semantic matrix containing modality word intensity value, causal trigger word weight value and sentence limited range value, and judging whether the coal mine safety production accident text adopts speculative expression based on a semantic vector distribution result; when the judgment result is that speculative expression is adopted, extracting a modality word intensity gradient, a semantic dependence span and a sentence limited convergence degree, and performing unified scale mapping processing to generate a causal semantic uncertainty parameter. The application solves the problem that the existing coal mine safety production knowledge graph construction technology cannot distinguish the uncertainty degree of causal semantics under speculative expression and perform differentiated intensity labeling, and realizes fine modeling and dynamic updating of the causal relationship intensity.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety production technology, specifically to a method for constructing a knowledge graph in the field of coal mine safety production. Background Technology

[0002] The construction of knowledge graphs in the field of coal mine safety production refers to the process of establishing a structured semantic network around equipment, personnel, working environment, types of hidden dangers, accidents, safety regulations and their interrelationships in coal mine safety production activities. Its core objective is to unify the modeling and association of multi-source heterogeneous data scattered in monitoring systems, accident reports, management systems and expert experience, thereby achieving systematic cognition and intelligent analysis of safety risks. Existing knowledge graph construction technologies in the field of coal mine safety production typically begin with domain ontology design, determining entity types and hierarchical relationships such as "equipment-component-status," "hazard-risk level-response measures," and "accident-cause-effect." Next, structured monitoring data and unstructured text data are acquired through data collection and preprocessing. Entity recognition and relation extraction are then performed using rule matching, natural language processing, or machine learning methods. The extracted results are then subjected to entity alignment, deduplication and fusion, and consistency verification to form standardized triplet data. These triples are then stored in a graph database to construct a graph structure model. Finally, risk association analysis and knowledge updates are achieved by combining inference rules or graph computation algorithms, thus completing the closed loop of knowledge graph construction and application in the field of coal mine safety production.

[0003] The existing technology has the following shortcomings: In the process of constructing knowledge graphs in the field of coal mine safety production, when accident investigation texts use speculative expressions such as "may lead to" or "there is an induced risk," differences in the degree of causal semantic uncertainty may occur. Because the causal relationship between hidden dangers and accident outcomes is not expressed in definitive statements but rather described through biased or inferential language, the causal relationship exhibits varying degrees of semantic strength. Existing knowledge graph construction technologies in the field of coal mine safety production cannot differentiate between high-certainty and low-certainty causal relationships based on the degree of causal semantic uncertainty in the use of speculative expressions in coal mine safety accident texts. Furthermore, they cannot perform differentiated intensity labeling, only identifying the existence of causality and uniformly assigning the same intensity label. This results in speculative and certain causal relationships being treated equally in the graph structure, amplifying the impact of low-certainty causal relationships during risk propagation calculations. This causes risk assessment results to deviate from the actual accident mechanism, reducing the accuracy of coal mine safety production knowledge graphs in accident analysis and risk prediction.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing knowledge graphs in the field of coal mine safety production, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a knowledge graph in the field of coal mine safety production, specifically including the following steps: S1. Perform causal semantic analysis on the text of coal mine safety production accidents, construct a speculative expression judgment semantic matrix including modal word intensity value, causal trigger word weight value and sentence limitation range value, and determine whether the text of coal mine safety production accidents uses speculative expression based on the semantic vector distribution results. S2. When the judgment result is to use speculative expression, extract the modal word intensity gradient, semantic dependency span and sentence constraint convergence, and perform unified scale mapping to generate causal semantic uncertainty parameters. S3. Based on the causal semantic uncertainty parameter, divide the interval and map the causal semantic uncertainty parameter to the deterministic interval and the non-deterministic interval. Based on the mapping result, distinguish between high deterministic causal relationships and low deterministic causal relationships. S4. Based on the differentiation results, perform intensity mapping calculations on high-determinism causal relationships and low-determinism causal relationships respectively, generate causal intensity coefficients, and write the causal intensity coefficients into the corresponding causal directed edge weight values; S5. During the knowledge graph update process, acquire accident verification data. Based on the deviation between the accident verification data and the causal strength coefficient, adjust the causal semantic uncertainty parameter and the causal strength coefficient proportionally to achieve dynamic updating of the causal relationship strength.

[0007] Preferably, S1 specifically includes the following steps: S101. Perform causal semantic analysis on coal mine safety production accident texts, including sentence segmentation, part-of-speech tagging and dependency calculation of coal mine safety production accident texts, locate causal relationship statements between hidden danger entities and accident result entities, and extract modal words, causal trigger words and sentence limiting components from causal relationship statements to form a set of causal semantic elements. S102. Based on the set of causal semantic elements, intensity values ​​are assigned to modal words, directional weight values ​​are assigned to causal trigger words, and range values ​​are marked for sentence limiting components. The intensity values ​​of modal words, weight values ​​of causal trigger words, and sentence limiting range values ​​are vectorized and encoded according to a unified dimension to construct a speculative expression judgment semantic matrix. S103. Calculate the semantic vector distribution of the semantic matrix for speculative expression judgment, determine the judgment distance value between the deterministic semantic center and the speculative semantic center, compare the judgment distance value with the preset judgment threshold, and determine whether the coal mine safety production accident text uses speculative expression based on the comparison result.

[0008] Preferably, S103 is as follows: The semantic vectors in each row of the speculative expression judgment semantic matrix are standardized, and the mean and variance of the vectors are calculated according to the unified vector dimension. Based on the standardization results, deterministic semantic center vectors and speculative semantic center vectors are constructed. The deterministic semantic center vectors and speculative semantic center vectors are both calculated from a pre-labeled set of coal mine safety production accident text sample vectors. The semantic vectors in the speculative expression judgment semantic matrix are distanced to the deterministic semantic center vector and the speculative semantic center vector, respectively. The first distance value from the semantic vector to the deterministic semantic center vector and the second distance value from the semantic vector to the speculative semantic center vector are obtained. The difference between the first distance value and the second distance value is calculated as the judgment distance value. The judgment distance value is compared with the preset judgment threshold. When the judgment distance value is greater than the preset judgment threshold, the text of the coal mine safety production accident is judged to use speculative expression. When the judgment distance value is less than or equal to the preset judgment threshold, the text of the coal mine safety production accident is judged not to use speculative expression.

[0009] Preferably, S2 specifically includes the following steps: S201. When the judgment result is a speculative expression, the modal word intensity gradient, semantic dependency span, and statement constraint convergence are extracted from the causal relationship statement. The modal word intensity gradient is obtained by numerically sorting the level positions of modal words in the preset semantic intensity sequence and calculating the difference between adjacent levels. The semantic dependency span is obtained by calculating the node hierarchy distance between the hidden danger entity and the accident result entity in the dependency relationship path. The statement constraint convergence is obtained by statistically analyzing the number of semantic nodes and the hierarchy range covered by the statement constraint component in the dependency structure and performing numerical quantification. S202. Perform unified scale mapping processing on the extracted modal word intensity gradient, semantic dependency span and sentence constraint convergence, including interval normalization and distribution rearrangement of each semantic feature value, so that the modal word intensity gradient, semantic dependency span and sentence constraint convergence are mapped to a unified numerical range, and feature vectors are constructed according to a fixed dimension order. S203. Based on the feature vectors after unified scale mapping, perform weighted combination calculation, and numerically fuse the modal word intensity gradient, semantic dependency span and sentence constraint convergence according to the preset weight ratio to generate the causal semantic uncertainty parameter of the corresponding causal relationship sentence.

[0010] Preferably, S201 specifically refers to: When the judgment result is to use speculative expression, all modal words are located in the causal relationship statement and mapped to the preset semantic intensity sequence. The modal words are encoded in order of semantic intensity from low to high. Multiple modal words appearing in the same causal relationship statement are sorted in the order of semantic intensity level. The numerical difference between adjacent levels is calculated and accumulated to obtain the modal word intensity gradient. Based on the dependency relationship path, the shortest dependency path between the hidden danger entity and the accident result entity is determined. The number of nodes and the path level depth in the shortest dependency path are counted. The number of nodes and the path level depth are weighted and summed to obtain the semantic dependency span. In the dependency structure, identify the set of modifier nodes corresponding to the qualifiers of the statement, count the number of semantic nodes covered by the set of modifier nodes and the corresponding hierarchical range, and quantify the numerical value according to the product of the number of covered nodes and the hierarchical range to obtain the statement qualifier convergence. The modal word intensity gradient, semantic dependency span and statement qualifier convergence are output as independent feature values.

[0011] Preferably, S3 is specifically as follows: A parameter distribution sequence is constructed based on the causal semantic uncertainty parameters corresponding to multiple causal relation statements. The causal semantic uncertainty parameters are sorted according to their numerical values, and the interval division boundary value is determined according to the preset distribution statistical interval. The interval division is completed by using the interval division boundary value as the dividing point. Based on the interval division boundary value, the causal semantic uncertainty parameter is mapped to a deterministic interval and an indeterminate interval. The causal semantic uncertainty parameter less than or equal to the interval division boundary value is mapped to the deterministic interval, and the causal semantic uncertainty parameter greater than the interval division boundary value is mapped to the indeterminate interval. A corresponding interval identifier is generated for each causal semantic uncertainty parameter. The causal relationships of corresponding causal association statements are distinguished based on the interval identifier. Causal relationships mapped to deterministic intervals are marked as high-deterministic causal relationships, and causal relationships mapped to non-deterministic intervals are marked as low-deterministic causal relationships. The high-deterministic and low-deterministic causal relationships are then written into the causal relationship attribute field in the knowledge graph construction process.

[0012] Preferably, S4 is as follows: Based on the differentiation results, the causal semantic uncertainty parameters corresponding to high-determinism causal relationships and low-determinism causal relationships are obtained respectively. Based on the interval division boundary value, the high-determinism intensity mapping interval and the low-determinism intensity mapping interval are determined. The high-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is less than or equal to the interval division boundary value, and the low-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is greater than the interval division boundary value. Preset high-determinism mapping ratios and preset low-determinism mapping ratios are set for the high-determinism mapping interval and the low-determinism mapping interval, respectively. The preset high-determinism mapping ratio is a proportional coefficient determined based on sample statistics and is used to perform proportional transformation calculations on the causal semantic uncertainty parameters within the high-determinism mapping interval. The preset low-determinism mapping ratio is a proportional coefficient determined based on sample distribution decay rules and is used to perform decreasing proportional transformation calculations on the causal semantic uncertainty parameters within the low-determinism mapping interval to generate the corresponding causal intensity coefficient. When generating causal directed edges between potential hazard entities and accident result entities, the corresponding causal strength coefficient is written into the corresponding causal directed edge weight value field, and the causal strength coefficient is associated with the classification results of high-deterministic causal relationships or low-deterministic causal relationships and stored together.

[0013] Preferably, S5 is as follows: During the knowledge graph update process, newly occurring coal mine safety production accident investigation texts and corresponding accident actual handling result records are collected. Causal relationship matching and positioning are performed on the accident investigation texts. The accident actual handling result records are correlated with the causal directed edges between the corresponding hidden danger entities and accident result entities in the knowledge graph, generating accident verification data that corresponds one-to-one with the causal directed edges. The actual occurrence result intensity value recorded in the accident verification data is compared with the causal intensity coefficient stored in the corresponding causal directed edge. The difference between the actual occurrence result intensity value and the causal intensity coefficient is calculated to form the deviation result. The proportional adjustment coefficient is determined according to the numerical direction and magnitude of the deviation result. Based on the proportional adjustment coefficient, the corresponding causal semantic uncertainty parameter and causal intensity coefficient are proportionally adjusted and calculated. The adjusted causal semantic uncertainty parameter is then re-involved in the interval division and intensity mapping calculation, and the adjusted causal intensity coefficient is written into the corresponding causal directed edge weight value field to realize the dynamic update of the causal relationship strength.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention achieves a structured and quantitative expression of the degree of causal semantic uncertainty in coal mine safety accident texts by constructing a speculative expression judgment semantic matrix, causal semantic uncertainty parameters, and an interval mapping mechanism. Compared with traditional knowledge graph construction techniques that only identify the existence of causality, this technical solution can perform semantic space discrimination on speculative expressions such as "may lead to" and "existence of inducing risk," and generate continuous causal semantic uncertainty parameters through multi-dimensional joint modeling of modal word intensity gradient, semantic dependency span, and sentence-limited convergence, thereby transforming the differences in semantic strength into calculable numerical intervals. Through interval partitioning and intensity mapping calculation, high-deterministic causal relationships and low-deterministic causal relationships are expressed with different weights in the graph structure, avoiding the equivalent treatment of speculative and deterministic causal relationships in the knowledge graph, and improving the refinement and numerical distinguishability of causal relationship expression.

[0015] 2. This invention introduces a ratio adjustment mechanism between accident verification data and deviation results, enabling dynamic correction of causal semantic uncertainty parameters and causal strength coefficients during knowledge graph updates. This forms a closed-loop control structure between semantic modeling, strength mapping, and actual feedback. By comparing and analyzing the actual accident handling results with the causal strength coefficients stored in the graph structure, the weights of causal relationships can be continuously corrected, gradually bringing the strength of causal relationships in the graph closer to the actual accident mechanism. This suppresses the over-amplification of low-deterministic causal relationships in risk propagation calculations. This dynamic update mechanism enhances the stability and adaptability of the coal mine safety production knowledge graph in accident analysis and risk prediction, improving the consistency between risk assessment results and actual accident evolution paths. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart illustrating the knowledge graph construction method in the field of coal mine safety production according to the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] This invention provides, for example Figure 1 The method for constructing a knowledge graph in the field of coal mine safety production, as shown, specifically includes the following steps: S1. Perform causal semantic analysis on the text of coal mine safety production accidents, construct a speculative expression judgment semantic matrix including modal word intensity value, causal trigger word weight value and sentence limitation range value, and determine whether the text of coal mine safety production accidents uses speculative expression based on the semantic vector distribution results. In this embodiment, S1 specifically includes the following steps: S101. Perform causal semantic analysis on coal mine safety production accident texts, including sentence segmentation, part-of-speech tagging and dependency calculation of coal mine safety production accident texts, locate causal relationship statements between hidden danger entities and accident result entities, and extract modal words, causal trigger words and sentence limiting components from causal relationship statements to form a set of causal semantic elements. When performing causal semantic analysis on texts related to coal mine safety accidents, the original accident text is first processed through rule-driven sentence segmentation. Using punctuation and semantic breakpoint recognition algorithms, long paragraphs are broken down into semantically complete short sentence units. Then, part-of-speech tagging is performed on each short sentence. Sequence tagging models based on conditional random fields or bidirectional long short-term memory networks can be used to categorize nouns, verbs, modal verbs, adverbs, etc. Based on this, dependency relation calculation is performed. By constructing a dependency syntax tree, subject-predicate, verb-object, and modification relations are determined, locating the connection paths between the hazard entity and the accident result entity from the syntactic structure. For example, in the sentence "abnormal local ventilation may lead to gas accumulation and trigger an explosion," "abnormal local ventilation" is identified as the hazard entity, "explosion" as the accident result entity, and "lead to" as the causal trigger word, "may" as the modal word, and "trigger" as the secondary causal trigger word in the dependency path. This extracts modal words, causal trigger words, and sentence limiting components. These elements are then aggregated according to entity alignment rules to form a set of causal semantic elements containing semantic roles and connection weight information. The purpose of this approach is to avoid judging causal relationships solely through keyword matching, but rather to perform precise calculations based on syntactic structure and semantic dependency paths, so that the causal expression between potential hazards and accidents has a structured semantic basis.

[0020] Coal mine safety production accident text refers to the natural language description content derived from accident investigation reports, hidden danger investigation records, or safety analysis materials; causal semantic parsing refers to the semantic computation process of identifying the causal relationship between hidden dangers and accident results from the text; sentence segmentation refers to breaking down complex compound sentences into semantically complete syntactic units; part-of-speech tagging refers to assigning part-of-speech tags to each word to distinguish modal words, verbs, and modifiers; dependency relation calculation refers to determining the dominance and modification relationships between words through syntactic dependency analysis; causal relationship statements between hidden danger entities and accident result entities refer to statements that simultaneously contain descriptions of hidden dangers and accident results and have causal connection paths; modal words are used to express probability or degree of certainty; causal trigger words are used to indicate causal pointing relationships; sentence limiting components are used to limit the scope or conditions of causal relationships; the causal semantic element set is a data set after uniformly encoding modal words, causal trigger words, sentence limiting components, and their positions in dependency paths, used for subsequent semantic strength calculation and uncertainty quantification.

[0021] S102. Based on the set of causal semantic elements, intensity values ​​are assigned to modal words, directional weight values ​​are assigned to causal trigger words, and range values ​​are marked for sentence limiting components. The intensity values ​​of modal words, weight values ​​of causal trigger words, and sentence limiting range values ​​are vectorized and encoded according to a unified dimension to construct a speculative expression judgment semantic matrix. After the set of causal semantic elements is formed, modal words, causal trigger words, and sentence limiting components need to be transformed into calculable numerical features. The numerical assignment of modal word intensity can be based on corpus statistics and semantic level classification. For example, "inevitably" and "certainly" can be marked as high-intensity intervals, while "possibly" and "perhaps" can be marked as medium-to-low-intensity intervals. The intensity values ​​can be further refined by combining the confidence distribution of word vectors in the semantic space. The numerical assignment of causal trigger word directional weights can be calculated based on the dependency path length and the dominance level of the trigger word in the syntactic structure. For example, direct subject-verb dominance relationships are assigned higher weights, while indirect clause dominance relationships are assigned lower weights. The numerical labeling of sentence limiting component ranges can be quantified based on the semantic span covered by the limiting words. For example, "under specific conditions" has a narrower limiting range and is assigned a smaller range value, while "generally" has a wider limiting range and is assigned a larger range value. After numericalization, the modal word intensity values, causal trigger word weight values, and sentence limiting range values ​​are vectorized and encoded according to a unified dimension. Each causal relationship sentence is represented as a fixed-dimensional numerical vector, and these vectors are arranged in sentence order to form a speculative expression judgment semantic matrix. This process can transform the originally discrete semantic differences into distributional differences in a continuous numerical space, providing a structured input basis for subsequent semantic vector distribution calculations.

[0022] Intensity numerical assignment refers to assigning continuous numerical values ​​to modal words based on semantic level and statistical distribution to represent the degree of causal certainty; directional weight numerical assignment refers to assigning weight values ​​based on the directional relationship and hierarchical position of causal trigger words in the dependency path to represent the intensity of causal direction; range numerical labeling refers to quantifying the semantic range covered by the limiting components of a statement to characterize the scope of application of causal relationships; vectorization encoding according to a unified dimension refers to mapping different types of semantic numerical features to a numerical vector space of uniform length and order to ensure comparability between statement vectors; the speculative expression judgment semantic matrix is ​​a two-dimensional numerical matrix formed by arranging multiple statement vectors according to a fixed dimension to characterize the overall distribution of causal semantic features in the entire text of a coal mine safety production accident.

[0023] S103. Calculate the semantic vector distribution of the semantic matrix for speculative expression judgment, determine the judgment distance value between the deterministic semantic center and the speculative semantic center, compare the judgment distance value with the preset judgment threshold, and determine whether the coal mine safety production accident text uses speculative expression based on the comparison result.

[0024] This paper calculates the semantic vector distribution of the semantic matrix for speculative expression judgment and determines the judgment distance between the semantic vector and the deterministic semantic center. This distance is then compared with a preset judgment threshold. The aim is to transform the expression type recognition process, which originally relied on intuitive linguistic judgment, into a numerical judgment process based on the geometric distribution of semantic space. Speculative expressions in coal mine safety accident texts often manifest in the combination differences of modal word intensity, causal trigger word directionality, and sentence limitation range. These differences are difficult to distinguish accurately at the single feature level, but they form clustering distribution features in the multidimensional semantic vector space. By constructing deterministic and speculative semantic centers, two typical expression forms can be abstracted as reference positions in the semantic space. By calculating the judgment distance between the semantic vector to be judged and the two semantic centers, its proximity to the two typical expressions can be quantified. Furthermore, by comparing with the preset judgment threshold, continuous distance differences are transformed into clear category boundaries, thereby avoiding the risk of misjudgment caused by judging solely based on keywords or local features. This allows speculative expression recognition to be based on the overall semantic distribution, providing a stable data premise for subsequent quantification and hierarchical modeling of causal semantic uncertainty.

[0025] In this embodiment, S103 specifically refers to: The semantic vectors in each row of the speculative expression judgment semantic matrix are standardized, and the mean and variance of the vectors are calculated according to the unified vector dimension. Based on the standardization results, deterministic semantic center vectors and speculative semantic center vectors are constructed. The deterministic semantic center vectors and speculative semantic center vectors are both calculated from a pre-labeled set of coal mine safety production accident text sample vectors. Standardizing each row of the semantic vector in the semantic matrix for speculative expression judgment refers to normalizing the numerical scale of the multidimensional semantic vector, composed of modal word intensity values, causal trigger word weight values, and sentence limitation range values, under the condition of a unified vector dimension, so that features of different dimensions are in a comparable range. Specifically, the original values ​​can be converted into standard distribution values ​​based on the mean center by calculating the mean and variance of the sample set for each feature, thereby eliminating the bias caused by the difference in the magnitude of values ​​of different dimensions. For example, the distribution of modal word intensity values ​​may be concentrated in a small range, while the distribution of causal trigger word weight values ​​may be more dispersed. By calculating the mean and variance of each feature in all semantic vectors, features of different dimensions can be mapped to a unified scale space. Based on the standardization results, two types of center vectors are constructed using a pre-labeled set of coal mine safety production accident text sample vectors. The pre-labeled set of coal mine safety production accident text sample vectors includes text vectors that have been manually confirmed to belong to deterministic and speculative expressions. By calculating the mean of the two types of sample vectors according to their dimensions, deterministic semantic center vectors and speculative semantic center vectors are obtained. The deterministic semantic center vector represents the central location of typical deterministic causal semantics in the multidimensional space, while the speculative semantic center vector represents the central location of typical speculative causal semantics in the multidimensional space. By constructing these two semantic centers, a reference benchmark can be provided for subsequent semantic vector distribution calculations, enabling precise measurement of the positional relationship of the text vectors to be judged in the semantic space, thereby achieving expression category determination based on statistical distribution features.

[0026] The semantic vectors in the speculative expression judgment semantic matrix are distanced to the deterministic semantic center vector and the speculative semantic center vector, respectively. The first distance value from the semantic vector to the deterministic semantic center vector and the second distance value from the semantic vector to the speculative semantic center vector are obtained. The difference between the first distance value and the second distance value is calculated as the judgment distance value. Calculating the distances between the semantic vectors in the speculative expression judgment semantic matrix and both the deterministic and speculative semantic center vectors is a metric analysis of the position of each row of semantic vectors in a unified semantic space. The semantic vectors in the speculative expression judgment semantic matrix are multi-dimensional numerical vectors composed of modal word intensity values, causal trigger word weight values, and statement limitation range values, used to characterize the semantic feature distribution of a single causal relationship statement. Distance calculation can be implemented using Euclidean distance, weighted Euclidean distance, or Mahalanobis distance, where the difference in each dimension represents the degree of deviation between the semantic vector to be judged and the semantic center in the corresponding semantic feature. For example, the difference in modal word intensity reflects the difference in the degree of certainty, and the difference in causal trigger word weight reflects the difference in causal orientation. By calculating the first distance value from the semantic vector to the deterministic semantic center vector and the second distance value from the semantic vector to the speculative semantic center vector, the closeness of the semantic vector to be judged in the deterministic and speculative semantic spaces can be obtained, respectively. The difference between the first and second distance values ​​is further calculated as the decision distance value. When the first distance value is less than the second distance value, the decision distance value shows a negative shift, indicating that the semantic vector is closer to the deterministic semantic center; when the second distance value is less than the first distance value, the decision distance value shows a positive shift, indicating that the semantic vector is closer to the speculative semantic center. Through this dual-center distance comparison calculation, the relative positional relationship between a single vector and two semantic category centers is transformed into a measurable decision distance value. This allows the judgment of causal semantic expression type to be based on the geometric distance in the semantic space, rather than a single threshold matching, thereby improving the accuracy and stability of expression category classification.

[0027] The judgment distance value is compared with the preset judgment threshold. When the judgment distance value is greater than the preset judgment threshold, the text of the coal mine safety production accident is judged to use speculative expression. When the judgment distance value is less than or equal to the preset judgment threshold, the text of the coal mine safety production accident is judged not to use speculative expression.

[0028] Comparing the judgment distance value with a preset judgment threshold is the final determination of semantic attribution after calculating the distances from the semantic vector to both the deterministic and speculative semantic center vectors. The judgment distance value is derived from the difference between a first distance value and a second distance value, representing the relative closeness of the semantic vector to be judged within the deterministic and speculative semantic spaces. The preset judgment threshold is a critical boundary value statistically obtained based on a pre-labeled set of coal mine safety production accident text sample vectors. It can be determined by analyzing the distribution of judgment distance values ​​between a large number of known deterministic and speculative expression texts, selecting the median or optimal classification boundary point of the intersection of the distance distributions of the two types of samples as the preset judgment threshold. In actual calculations, when the judgment distance value is greater than the preset judgment threshold, it indicates that the semantic vector is closer to the speculative semantic center vector in the numerical space, thus determining that the coal mine safety production accident text uses speculative expression; when the judgment distance value is less than or equal to the preset judgment threshold, it indicates that the semantic vector is closer to the deterministic semantic center vector, thus determining that the coal mine safety production accident text does not use speculative expression. For example, if the calculated decision distance value in a certain accident text is positively offset and exceeds the threshold boundary, it indicates that the combined features of modal word strength and causal trigger word weight are closer to the distribution of speculative semantic samples, and therefore it is classified as a speculative expression. By introducing a preset decision threshold, continuous decision distance values ​​can be transformed into discrete category judgments, making semantic classification based on statistical distribution and avoiding interference from misjudgments of a single feature on the overall expression category judgment.

[0029] S2. When the judgment result is to use speculative expression, extract the modal word intensity gradient, semantic dependency span and sentence constraint convergence, and perform unified scale mapping to generate causal semantic uncertainty parameters. In this embodiment, S2 specifically includes the following steps: S201. When the judgment result is a speculative expression, the modal word intensity gradient, semantic dependency span, and statement constraint convergence are extracted from the causal relationship statement. The modal word intensity gradient is obtained by numerically sorting the level positions of modal words in the preset semantic intensity sequence and calculating the difference between adjacent levels. The semantic dependency span is obtained by calculating the node hierarchy distance between the hidden danger entity and the accident result entity in the dependency relationship path. The statement constraint convergence is obtained by statistically analyzing the number of semantic nodes and the hierarchy range covered by the statement constraint component in the dependency structure and performing numerical quantification. S202. Perform unified scale mapping processing on the extracted modal word intensity gradient, semantic dependency span and sentence constraint convergence, including interval normalization and distribution rearrangement of each semantic feature value, so that the modal word intensity gradient, semantic dependency span and sentence constraint convergence are mapped to a unified numerical range, and feature vectors are constructed according to a fixed dimension order. Applying a unified scale mapping to the extracted modal intensity gradient, semantic dependency span, and sentence constraint convergence aims to eliminate differences in the numerical range and distribution of different semantic features, enabling comparison and fusion of multidimensional semantic features within the same numerical space. Specifically, this involves first calculating the minimum and maximum values ​​of the modal intensity gradient, semantic dependency span, and sentence constraint convergence in the sample set. Then, interval normalization is performed on each type of semantic feature based on its original value range, mapping feature values ​​of different dimensions to a unified numerical interval. For example, a semantic dependency span with a large original span is compressed to the same numerical range as the modal intensity gradient. Subsequently, the normalized feature distribution undergoes distribution rearrangement. By sorting the sample distribution and remapping it to equally spaced intervals, the impact of extreme values ​​on the overall distribution is reduced. After completing interval normalization and distribution rearrangement, the modal intensity gradient, semantic dependency span, and sentence constraint convergence are arranged in a fixed dimension order to construct a feature vector. For example, the first dimension corresponds to the modal intensity gradient, the second dimension corresponds to the semantic dependency span, and the third dimension corresponds to the sentence constraint convergence. This ensures that each causal relationship sentence corresponds to a numerical vector with a consistent structure, providing a unified input basis for subsequent weighted fusion calculations.

[0030] Unified scaling mapping refers to the numerical standardization of semantic feature values ​​from different sources and with different dimensions, so that they fall into the same numerical range. Each semantic feature value includes three types of numerical features: modal word intensity gradient, semantic dependency span, and sentence constraint convergence. Interval normalization maps the original feature values ​​to a preset interval according to their distribution range to eliminate differences in dimensions. Distribution rearrangement reorders the normalized values ​​according to the statistical distribution to make the feature distribution more balanced. Unified numerical interval refers to the common value range of various semantic features after mapping. Fixed dimension order means that each semantic feature is arranged in a preset position when constructing the feature vector to ensure that the vector structure corresponding to different sentences is consistent. The feature vector is a numerical array composed of multiple semantic feature values ​​processed by unified scaling in a fixed order, used to represent the semantic feature combination state of a single causal relationship sentence.

[0031] S203. Based on the feature vectors after unified scale mapping, perform weighted combination calculation, and numerically fuse the modal word intensity gradient, semantic dependency span and sentence constraint convergence according to the preset weight ratio to generate the causal semantic uncertainty parameter of the corresponding causal relationship sentence.

[0032] The weighted combination calculation based on feature vectors mapped to a unified scale aims to comprehensively quantify the modal intensity gradient, semantic dependency span, and sentence constraint convergence within a unified numerical space, thereby forming a single parameter that can holistically reflect the uncertainty of causal expression. In practice, the contribution of the three types of semantic features to the uncertainty can be pre-analyzed based on labeled samples, and preset weight ratios can be set accordingly. For example, the influence ratio of modal intensity gradient on uncertainty, the influence ratio of semantic dependency span on indirectness, and the influence ratio of sentence constraint convergence on the narrowing of the applicable scope can be determined based on the sample distribution. Subsequently, the feature vectors mapped to the unified scale are numerically fused. Each feature value is multiplied according to its corresponding preset weight ratio and summed to obtain the comprehensive value of the corresponding causal relationship statement. This comprehensive value is the causal semantic uncertainty parameter. For example, in a causal relationship statement, a high modal intensity gradient, a large semantic dependency span, and a low sentence constraint convergence result in a high causal semantic uncertainty parameter after weighted combination, indicating a high degree of uncertainty in the causal relationship within the statement. This weighted combination calculation avoids the dominance of a single semantic feature in the result, allowing the uncertainty parameter to comprehensively reflect the influence of multiple semantic dimensions.

[0033] Weighted combination calculation refers to the proportional superposition of the numerical values ​​of each dimension according to a preset weight ratio on the basis of the multidimensional feature vector after uniform scale mapping. The preset weight ratio is the influence coefficient of each semantic feature determined in advance based on the sample statistical results or expert annotation results. Numerical fusion refers to the integration of three feature values, namely modal word intensity gradient, semantic dependency span and sentence constraint convergence, into a single numerical value through weight allocation. The causal semantic uncertainty parameter of the corresponding causal relationship statement is a quantitative representation of the uncertainty of a single causal relationship statement under the comprehensive effect of multidimensional semantic features, which is used for subsequent interval mapping and causal relationship classification processing.

[0034] In this embodiment, S201 specifically refers to: When the judgment result is to use speculative expression, all modal words are located in the causal relationship statement and mapped to the preset semantic intensity sequence. The modal words are encoded in order of semantic intensity from low to high. Multiple modal words appearing in the same causal relationship statement are sorted in the order of semantic intensity level. The numerical difference between adjacent levels is calculated and accumulated to obtain the modal word intensity gradient. When the judgment result is a speculative expression, locating all modal words in the causal relationship statement and calculating the modal word intensity gradient is to transform the differences in the degree of certainty in the expression into quantifiable gradient change values. Specifically, this can be achieved first by identifying modal verbs and adverbs indicating probability, such as "maybe," "perhaps," "basically certain," and "inevitably," using a part-of-speech tagging model. Then, a pre-defined semantic intensity sequence is constructed. This sequence is a set of modal words ranked by degree of certainty obtained from a large number of coal mine safety accident texts. The modal words are arranged in an ordered list from low to high according to the strength of their expressed certainty, and each modal word in the sequence is assigned a corresponding level code; for example, the lowest degree of certainty is assigned a lower level value, and the highest degree of certainty is assigned a higher level value. For multiple modal words appearing in the same causal relationship statement, they are sorted from low to high according to their level codes, and the numerical differences between adjacent levels are calculated and accumulated to obtain the modal word intensity gradient. For example, in the statement "Local ventilation abnormalities may further lead to gas accumulation and ultimately inevitably cause an explosion," "may" and "inevitably" correspond to different level codes. By calculating and summing the level differences, a modal intensity gradient reflecting the magnitude of semantic intensity changes can be obtained. A pre-defined semantic intensity sequence is used to establish a unified benchmark for ranking the degree of certainty, level codes are used to convert linguistic expressions into numerical representations, and modal intensity gradients are used to characterize the changing trend of the degree of certainty within the same statement, thus providing continuous intensity change features for subsequent calculations of causal semantic uncertainty.

[0035] Based on the dependency relationship path, the shortest dependency path between the hidden danger entity and the accident result entity is determined. The number of nodes and the path level depth in the shortest dependency path are counted. The number of nodes and the path level depth are weighted and summed to obtain the semantic dependency span. Determining the shortest dependency path between the potential hazard entity and the accident result entity based on dependency relationship paths aims to quantify the semantic tightness between the two types of entities at the syntactic structure level. In practice, a dependency tree of the entire sentence can be constructed first using a dependency parsing algorithm, with each word in the sentence as a node and dominance relations as directed edges, thus forming a graph structure. In this dependency graph, the node containing the potential hazard entity and the node containing the accident result entity are taken as the start and end points, respectively. A graph traversal algorithm is used to search for all dependency paths between the two nodes, and the path with the fewest nodes is selected as the shortest dependency path. A dependency relationship path refers to the sequence of nodes traversed in the dependency syntactic tree from one node to another along a dominance or subordinate relation; the shortest dependency path between the potential hazard entity and the accident result entity represents the syntactic pathway with the tightest connection between them in the syntactic structure. The number of nodes refers to the total number of word nodes included in the shortest dependency path, and the path level depth refers to the maximum dependency level distance in the path, i.e., the number of levels from the root node to the deepest node in the path. By weighted summing of the number of nodes and the path hierarchy depth, the path length and structural complexity can be combined into a single value. For example, when there are many nodes and a deep hierarchy, it indicates that there are many mediating semantic components between the potential hazard entity and the accident outcome entity, and the semantic connection is relatively indirect. When there are few nodes and a shallow hierarchy, it indicates that the causal expression between the two is more direct. The semantic dependency span is this weighted result, used to characterize the extent of causal semantic extension in the syntactic structure, providing structural features reflecting semantic indirectness for subsequent calculation of causal semantic uncertainty.

[0036] In the dependency structure, identify the set of modifier nodes corresponding to the qualifiers of the statement, count the number of semantic nodes covered by the set of modifier nodes and the corresponding hierarchical range, and quantify the numerical value according to the product of the number of covered nodes and the hierarchical range to obtain the statement qualifier convergence. The modal word intensity gradient, semantic dependency span and statement qualifier convergence are output as independent feature values.

[0037] Identifying the set of modifying nodes corresponding to the limiting elements of a statement in a dependency structure is to characterize the degree to which the limiting expression constrains the scope of the causal relationship. A dependency structure is a directed tree structure constructed through dependency parsing, where each word is a node, and nodes are connected by dominion relationships to form a hierarchy. In this structure, by identifying limiting words that indicate conditions, scope, or degree of restriction, such as "under specific conditions," "only in a local area," and "usually," the node containing the limiting element can be determined. Then, by traversing downwards along the dependency relationship, all semantic nodes it governs or modifies are obtained, forming a set of modifying nodes. The number of semantic nodes covered by the set of modifying nodes refers to the total number of word nodes directly or indirectly affected by the limiting element, and the corresponding hierarchical range refers to the maximum hierarchical span of these nodes in the dependency structure. By statistically analyzing the number of semantic nodes covered by the set of modifying nodes and the corresponding hierarchical range, the breadth and depth of the constraint imposed by the limiting element on the causal expression can be reflected. Subsequently, numerical quantization is performed based on the product of the number of covered nodes and the hierarchical range. A higher convergence value is obtained when the coverage area is smaller and the hierarchical range is more concentrated, while a lower convergence value is obtained when the coverage area is larger and the hierarchical range is more dispersed. The statement-constrained convergence is this quantization result, used to represent the degree of concentration of causal semantics after being constrained by the constrained components. It is output as an independent feature value along with the modal word intensity gradient and semantic dependency span, for subsequent uniform scaling and calculation of causal semantic uncertainty parameters.

[0038] S3. Based on the causal semantic uncertainty parameter, divide the interval and map the causal semantic uncertainty parameter to the deterministic interval and the non-deterministic interval. Based on the mapping result, distinguish between high deterministic causal relationships and low deterministic causal relationships. In this embodiment, S3 is specifically: A parameter distribution sequence is constructed based on the causal semantic uncertainty parameters corresponding to multiple causal relation statements. The causal semantic uncertainty parameters are sorted according to their numerical values, and the interval division boundary value is determined according to the preset distribution statistical interval. The interval division is completed by using the interval division boundary value as the dividing point. Constructing a parameter distribution sequence based on causal semantic uncertainty parameters corresponding to multiple causal-related statements aims to determine reasonable interval boundaries at the overall sample level. Specifically, this involves first collecting all causal semantic uncertainty parameters corresponding to causal-related statements from the same batch of coal mine safety accident texts, sorting these parameters by value to form a parameter distribution sequence arranged from low to high. Then, statistical analysis is performed on the sorted parameter sequence, such as analyzing the sample frequency distribution within different numerical ranges, identifying dense and sparse regions in the parameter distribution, and determining interval boundary values ​​based on preset distribution statistical intervals. These preset distribution statistical intervals can be set based on the overall sample distribution characteristics, such as determining the boundaries according to the median position of the parameter distribution or the inflection point of sample aggregation. When a significant change in parameter distribution density is observed before and after a certain numerical position after sorting, the value corresponding to that position can be used as the interval boundary value. Using the interval boundary value as the dividing point, smaller values ​​are assigned to one interval, and larger values ​​to another continuous interval, thus completing the interval division and ensuring that interval division is based on overall distribution characteristics rather than a single fixed value.

[0039] A parameter distribution sequence is an ordered set of numerical values ​​formed by arranging the causal semantic uncertainty parameters corresponding to multiple causal relationship statements according to their numerical magnitude. It reflects the overall distribution of the sample. Preset statistical intervals are numerical segmentation rules determined based on sample statistical results, used to analyze the concentration of parameters within different segments. Interval division boundary values ​​are boundary values ​​selected based on parameter distribution characteristics, used to divide continuous uncertainty parameters into different interval categories. By constructing the parameter distribution sequence and determining the interval division boundary values, it is ensured that interval division is based on statistical distribution, allowing the distinction between high-certainty and low-certainty causal relationships to be completed based on the overall numerical characteristics of the sample, providing a stable numerical foundation for subsequent interval mapping and causal relationship classification.

[0040] Based on the interval division boundary value, the causal semantic uncertainty parameter is mapped to a deterministic interval and an indeterminate interval. The causal semantic uncertainty parameter less than or equal to the interval division boundary value is mapped to the deterministic interval, and the causal semantic uncertainty parameter greater than the interval division boundary value is mapped to the indeterminate interval. A corresponding interval identifier is generated for each causal semantic uncertainty parameter. Mapping causal semantic uncertainty parameters to deterministic and non-deterministic intervals based on interval division boundary values ​​is a process of classifying individual causal semantic uncertainty parameters after completing the statistical analysis of the parameter distribution sequence. Specifically, each causal semantic uncertainty parameter is numerically compared with the interval division boundary value. When the causal semantic uncertainty parameter is less than or equal to the interval division boundary value, it is classified into a deterministic interval; when it is greater than the interval division boundary value, it is classified into a non-deterministic interval. After interval classification, a corresponding interval identifier is generated for each causal semantic uncertainty parameter, such as a binary identifier or interval number to record its interval category. For example, if the interval division boundary value is located in the middle of the sample distribution, if the causal semantic uncertainty parameter value corresponding to a causal relationship statement is lower than the boundary value, it indicates that its semantic features are closer to a deterministic expression distribution, and it is marked as a deterministic interval; if the value is higher than the boundary value, it indicates that its semantic features are closer to a speculative expression distribution, and it is marked as a non-deterministic interval. Through this boundary value-based mapping process, continuous numerical parameters can be transformed into discrete interval categories, providing a clear basis for subsequent causal relationship classification.

[0041] The deterministic and non-deterministic intervals are two continuous numerical segments formed by dividing the causal semantic uncertainty parameter within the overall sample distribution. The deterministic interval corresponds to the range of lower uncertainty parameters, while the non-deterministic interval corresponds to the range of higher uncertainty parameters. The corresponding interval label is a category label generated for each causal semantic uncertainty parameter, used to distinguish different types of causal relationships during knowledge graph construction. The interval mapping process refers to the computational process of classifying causal semantic uncertainty parameters into preset interval categories through numerical comparison. By setting deterministic and non-deterministic intervals and generating corresponding interval labels, structured classification of causal relationships can be achieved during the knowledge graph construction stage, ensuring that the subsequent distinction between high-determinism and low-determinism causal relationships is based on a clear interval division.

[0042] The causal relationships of corresponding causal association statements are distinguished based on the interval identifier. Causal relationships mapped to deterministic intervals are marked as high-deterministic causal relationships, and causal relationships mapped to non-deterministic intervals are marked as low-deterministic causal relationships. The high-deterministic and low-deterministic causal relationships are then written into the causal relationship attribute field in the knowledge graph construction process.

[0043] Distinguishing causal relationships based on interval identifiers involves transforming numerical classification results into structured relational attributes in a knowledge graph after mapping the causal semantic uncertainty parameter intervals. Specifically, this involves iterating through the interval identifiers corresponding to each causal relationship. When the interval identifier corresponds to a deterministic interval, the causal relationship in that relationship is defined as a high-determinism causal relationship; when it corresponds to a non-deterministic interval, it is defined as a low-determinism causal relationship. Subsequently, during knowledge graph construction, when generating causal directed edges between the potential hazard entity and the accident result entity, the high-determinism or low-determinism causal relationship is written into the causal relationship attribute field. For example, a causal relationship type field or intensity category field can be set in the graph database to distinguish causal relationships with different degrees of certainty. For instance, when a causal semantic uncertainty parameter is mapped to a deterministic interval, its corresponding "potential hazard-accident result" relationship is marked as a high-determinism causal relationship in the graph structure; when the parameter is mapped to a non-deterministic interval, it is marked as a low-determinism causal relationship, thus achieving differentiated classification at the graph structure level.

[0044] High-determinism causal relationships refer to causal relationship types where the causal semantic uncertainty parameter lies within the deterministic interval, indicating that the semantic expression between the potential hazard and the accident outcome is close to a deterministic statement. Low-determinism causal relationships refer to causal relationship types where the causal semantic uncertainty parameter lies within the uncertain interval, indicating that the semantic expression between the potential hazard and the accident outcome has a high degree of speculation. Interval identifiers are category labels generated by the interval mapping process, used to identify the numerical range to which the causal semantic uncertainty parameter belongs. Causal relationship attribute fields are data fields in the knowledge graph used to store causal relationship type or category information. By incorporating high-determinism and low-determinism causal relationships into the causal relationship attribute fields during the knowledge graph construction process, the degree of certainty of causal relationships can be preserved at the graph structure level, providing a classification basis for subsequent causal strength labeling and risk propagation calculation.

[0045] S4. Based on the differentiation results, perform intensity mapping calculations on high-determinism causal relationships and low-determinism causal relationships respectively, generate causal intensity coefficients, and write the causal intensity coefficients into the corresponding causal directed edge weight values; In this embodiment, S4 specifically refers to: Based on the differentiation results, the causal semantic uncertainty parameters corresponding to high-determinism causal relationships and low-determinism causal relationships are obtained respectively. Based on the interval division boundary value, the high-determinism intensity mapping interval and the low-determinism intensity mapping interval are determined. The high-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is less than or equal to the interval division boundary value, and the low-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is greater than the interval division boundary value. Based on the differentiation results, the causal semantic uncertainty parameters corresponding to high-determinism and low-determinism causal relationships are obtained separately. Then, based on the interval division boundary values, the high-determinism intensity mapping interval and the low-determinism intensity mapping interval are determined. This establishes a numerical segmentation basis for subsequent intensity mapping calculations after the causal relationship classification is completed. In specific implementation, the causal semantic uncertainty parameters corresponding to each causal association statement can be read first, and the parameters can be numerically compared according to the interval division boundary values. Parameters with values ​​less than or equal to the interval division boundary values ​​are assigned to one continuous numerical interval, while parameters with values ​​greater than the interval division boundary values ​​are assigned to another continuous numerical interval. For example, when the interval division boundary values ​​are derived from the median position of the overall parameter distribution, if a causal semantic uncertainty parameter value is below the median value, it is assigned to the high-determinism intensity mapping interval; if it is above the median value, it is assigned to the low-determinism intensity mapping interval. This interval division based on boundary values ​​allows subsequent intensity mapping calculations to use different mapping rules in different numerical segments, thereby avoiding the use of a uniform intensity transformation strategy for all causal relationships.

[0046] The high-determinism intensity mapping interval and the low-determinism intensity mapping interval are two continuous numerical segments formed based on the interval division boundary values. The high-determinism intensity mapping interval corresponds to the numerical range where the causal semantic uncertainty parameter is less than or equal to the interval division boundary value, while the low-determinism intensity mapping interval corresponds to the numerical range where the causal semantic uncertainty parameter is greater than the interval division boundary value. The interval division boundary value is a boundary value determined based on the overall distribution characteristics of the causal semantic uncertainty parameter, used to realize the conversion of continuous parameters into discrete segments. By constructing the high-determinism and low-determinism intensity mapping intervals, segmented numerical processing basis for different categories of causal relationships can be provided during the knowledge graph construction process. This ensures that the intensity mapping is based on interval segmentation rather than a simple proportional scaling, thereby enhancing the hierarchical nature and controllability of intensity calculation.

[0047] Preset high-determinism mapping ratios and preset low-determinism mapping ratios are set for the high-determinism mapping interval and the low-determinism mapping interval, respectively. The preset high-determinism mapping ratio is a proportional coefficient determined based on sample statistics and is used to perform proportional transformation calculations on the causal semantic uncertainty parameters within the high-determinism mapping interval. The preset low-determinism mapping ratio is a proportional coefficient determined based on sample distribution decay rules and is used to perform decreasing proportional transformation calculations on the causal semantic uncertainty parameters within the low-determinism mapping interval to generate the corresponding causal intensity coefficient. Setting preset high-certainty mapping ratios and preset low-certainty mapping ratios for high-certainty and low-certainty intensity mapping intervals respectively establishes differentiated intensity conversion rules for different types of causal relationships after interval division. Specifically, based on historical coal mine safety accident text samples, the distribution characteristics of causal semantic uncertainty parameters corresponding to high-certainty causal relationships can be statistically analyzed to calculate their average impact level in risk propagation calculations. Based on this, a preset high-certainty mapping ratio is determined, allowing causal semantic uncertainty parameters within the high-certainty intensity mapping interval to be converted and calculated at a fixed ratio, thereby generating a relatively stable causal intensity coefficient. For low-certainty intensity mapping intervals, the risk propagation attenuation trend of low-certainty causal relationships in the samples can be analyzed, establishing a ratio coefficient determined based on sample distribution attenuation rules. This allows causal semantic uncertainty parameters within the low-certainty intensity mapping interval to be converted and calculated at a decreasing ratio; for example, a lower ratio coefficient is used as the parameter value approaches the upper limit of the interval, resulting in a relatively smaller causal intensity coefficient. Through this interval-differentiated ratio conversion, high-certainty causal relationships maintain a higher weight in the graph structure, while low-certainty causal relationships exhibit a progressively decreasing weight distribution.

[0048] The preset high-determinism mapping ratio and the preset low-determinism mapping ratio are pre-determined ratio coefficients for different intensity mapping intervals. The ratio coefficient determined based on sample statistics is a numerical ratio obtained by statistically analyzing the causal semantic uncertainty parameter and the actual risk impact in the labeled samples. The ratio conversion calculation refers to performing numerical operations on the causal semantic uncertainty parameter and the corresponding ratio coefficient to generate a new intensity value. The ratio coefficient determined based on the sample distribution decay rule is a ratio parameter set according to the distribution characteristics of low-determinism causal relationships in the samples where the risk impact decreases as uncertainty increases. The decreasing ratio conversion calculation refers to performing numerical operations after adjusting the ratio coefficient according to the parameter value within the low-determinism intensity mapping interval. The causal intensity coefficient is a value obtained after ratio conversion calculation and is used to represent the weight of the causal relationship between the hidden danger entity and the accident result entity in the knowledge graph.

[0049] When generating causal directed edges between potential hazard entities and accident result entities, the corresponding causal strength coefficient is written into the corresponding causal directed edge weight value field, and the causal strength coefficient is associated with the classification results of high-deterministic causal relationships or low-deterministic causal relationships and stored together.

[0050] When generating causal directed edges between potential hazard entities and accident result entities, the corresponding causal strength coefficient is written into the weight value field of the corresponding causal directed edge. This embeds the numerical strength of causal relationships into the graph structure during the knowledge graph construction phase. Specifically, a weight value field can be set for each causal directed edge between a potential hazard entity and an accident result entity in the graph database. Simultaneously, the causal strength coefficient is written as the numerical content of this weight value field while constructing the causal directed edge, and the classification result (high-determinism or low-determinism) of the causal relationship is recorded. For example, when a causal relationship is classified as a high-determinism causal relationship and the corresponding causal strength coefficient is relatively high, a directed edge from the potential hazard entity to the accident result entity is created in the graph structure, and this value is written into the weight value field. The classification type is also marked in the relationship attributes. When the causal relationship is classified as a low-determinism causal relationship and the corresponding causal strength coefficient is relatively low, a causal directed edge is also generated, but with a smaller weight value and marked as low-determinism. Through this writing and association process, weight values ​​can be directly read for path calculation during graph computation or risk propagation analysis, and different strategies can be applied based on the classification results.

[0051] The causal directed edges between potential hazard entities and accident outcome entities are graph structure elements in the knowledge graph representing causal direction relationships. The starting point of each edge is a potential hazard entity node, and the ending point is an accident outcome entity node. The corresponding weight value field of the causal directed edge is a data field in the graph structure used to store the numerical value of the causal relationship strength. Association storage refers to simultaneously storing the causal strength coefficient and the causal relationship classification result in the attribute set of the same causal directed edge, establishing a correspondence between numerical and category information. By associating the causal strength coefficient and classification result at the graph structure level, it is possible to ensure that high-deterministic and low-deterministic causal relationships have distinguishable numerical representations in the graph model, providing a data foundation for subsequent weight-based path analysis and risk propagation calculations.

[0052] S5. During the knowledge graph update process, acquire accident verification data. Based on the deviation between the accident verification data and the causal strength coefficient, adjust the causal semantic uncertainty parameter and the causal strength coefficient proportionally to achieve dynamic updating of the causal relationship strength.

[0053] In this embodiment, S5 specifically refers to: During the knowledge graph update process, newly occurring coal mine safety production accident investigation texts and corresponding accident actual handling result records are collected. Causal relationship matching and positioning are performed on the accident investigation texts. The accident actual handling result records are correlated with the causal directed edges between the corresponding hidden danger entities and accident result entities in the knowledge graph, generating accident verification data that corresponds one-to-one with the causal directed edges. The purpose of collecting newly occurring coal mine safety accident investigation texts and corresponding actual accident handling records during the knowledge graph update process is to calibrate the strength of existing causal relationships using real accident feedback information. Specifically, the coal mine safety production management information platform can receive newly generated accident investigation texts and handling result records in real time. The accident investigation texts are then subjected to causal semantic analysis to identify hidden danger entities and accident result entities, and the existence of corresponding causal directed edges is searched in the knowledge graph. Subsequently, the result type, impact level, or actual consequence degree in the actual accident handling result records are numerically processed and a unique correspondence is established with the corresponding causal directed edges in the knowledge graph. For example, when an accident record shows that a specific hidden danger indeed caused a specific accident result and had a certain impact level, this actual impact level can be converted into an actual result strength value and matched with the corresponding causal directed edges in the graph, thus forming accident verification data that corresponds one-to-one with the causal directed edges for subsequent deviation analysis.

[0054] Newly occurring coal mine safety production accident investigation texts and corresponding accident actual handling results records refer to the newly added accident analysis documents and their final processing results data after the knowledge graph is constructed. Causal relationship matching and localization refers to the process of aligning the causal relationships in the accident text with the existing causal directed edges in the knowledge graph at the entity level and matching them at the relation level by identifying hidden danger entities and accident result entities. The causal directed edges between hidden danger entities and accident result entities are graph structure edges in the knowledge graph representing causal direction relationships. Accident verification data, corresponding one-to-one with each causal directed edge, refers to each causal directed edge being associated with a verification record derived from the actual accident result, used to reflect the correspondence between predicted intensity and actual results. By constructing accident verification data, a real data foundation can be provided for subsequent proportional adjustments based on deviation results.

[0055] The actual occurrence result intensity value recorded in the accident verification data is compared with the causal intensity coefficient stored in the corresponding causal directed edge. The difference between the actual occurrence result intensity value and the causal intensity coefficient is calculated to form the deviation result. The proportional adjustment coefficient is determined according to the numerical direction and magnitude of the deviation result. Comparing the actual occurrence intensity value recorded in the accident verification data with the causal intensity coefficient stored in the corresponding causal directed edges is a quantitative analysis of the consistency between predicted intensity and actual feedback during the dynamic update phase of the knowledge graph. Specifically, the actual occurrence intensity value can be extracted from the accident verification data, and this value can be numerically encoded based on the actual impact level, loss extent, or handling level of the accident. Then, the causal intensity coefficient stored in the corresponding causal directed edges is read from the knowledge graph, and the two are numerically compared to calculate the difference between the actual occurrence intensity value and the causal intensity coefficient, forming a deviation result. For example, when the causal intensity coefficient indicates a high intensity of an accident caused by a potential hazard, but the actual occurrence intensity value is low, a positive or negative difference is generated between the two. By analyzing the direction of the deviation result, it can be determined whether the predicted intensity is too high or too low; by analyzing the magnitude of the deviation result, it can be determined the degree of deviation. Furthermore, a proportional adjustment coefficient is determined based on the direction and magnitude of the deviation result, making subsequent adjustments to the causal semantic uncertainty parameter and the causal intensity coefficient more targeted and quantitatively based.

[0056] The intensity value of the actual occurrence result recorded in the accident verification data is an intensity representation obtained by quantifying the actual accident handling results; the causal intensity coefficient stored in the corresponding causal directed edge is a value in the knowledge graph used to represent the weight of the causal relationship between the hidden danger entity and the accident result entity; numerical comparison refers to directly calculating the difference between two values ​​to obtain the difference value; the difference between the intensity value of the actual occurrence result and the causal intensity coefficient is the deviation result, which is used to reflect the degree of difference between the predicted intensity and the actual result; the numerical direction of the deviation result indicates whether the predicted value is higher or lower than the actual value, and the magnitude of the deviation result indicates the degree of deviation; the proportional adjustment coefficient is a numerical parameter calculated based on the deviation result, which is used to guide the subsequent proportional correction of the causal semantic uncertainty parameter and the causal intensity coefficient, thereby realizing the dynamic correction of the causal relationship intensity.

[0057] Based on the proportional adjustment coefficient, the corresponding causal semantic uncertainty parameter and causal intensity coefficient are proportionally adjusted and calculated. The adjusted causal semantic uncertainty parameter is then re-involved in the interval division and intensity mapping calculation, and the adjusted causal intensity coefficient is written into the corresponding causal directed edge weight value field to realize the dynamic update of the causal relationship strength.

[0058] The proportional adjustment calculation, based on the proportional adjustment coefficient, for the corresponding causal semantic uncertainty parameter and causal intensity coefficient is a process of adaptively correcting the existing causal relationship values ​​after obtaining the deviation results. In practice, the proportional adjustment coefficient can be used as an adjustment factor, numerically calculated with the original causal semantic uncertainty parameter to obtain the adjusted causal semantic uncertainty parameter; simultaneously, the proportional adjustment coefficient is applied to the corresponding causal intensity coefficient, causing the causal intensity coefficient to be corrected synchronously in the direction of the deviation. For example, when the deviation results indicate that the original causal intensity coefficient is higher than the actual occurrence intensity value, the proportional adjustment coefficient can be set to a value less than the benchmark proportion, adjusting the original causal semantic uncertainty parameter upwards and the causal intensity coefficient downwards; when the deviation results indicate that the original causal intensity coefficient is lower than the actual occurrence intensity value, the proportional adjustment coefficient can be set to a value greater than the benchmark proportion, adjusting the causal semantic uncertainty parameter downwards and the causal intensity coefficient upwards. The adjusted causal semantic uncertainty parameter re-participates in the interval partitioning and intensity mapping calculation process, thereby generating a new causal intensity coefficient. The updated causal intensity coefficient is then written into the corresponding causal directed edge weight value field, so that the causal relationship strength in the knowledge graph is continuously updated with the feedback of real accidents.

[0059] Proportional adjustment calculation refers to the numerical operation process of proportionally correcting the original numerical parameters using proportional adjustment coefficients; causal semantic uncertainty parameter is a numerical variable used to represent the degree of uncertainty in causal expression; causal intensity coefficient is a numerical variable used to represent the weight of the causal relationship between the potential hazard entity and the accident result entity; interval partitioning and intensity mapping calculation is the process of re-determining high-determinism and low-determinism causal relationships based on the adjusted causal semantic uncertainty parameter and calculating the corresponding causal intensity coefficients; the corresponding causal directed edge weight value field is a numerical field in the knowledge graph that stores the weights of causal relationships. By performing proportional adjustment calculation and re-participating in interval partitioning and intensity mapping calculation, a closed-loop update logic can be formed, enabling the causal relationship strength to be continuously and dynamically corrected during the knowledge graph update process.

[0060] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0061] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0062] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0063] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0066] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a knowledge graph in the field of coal mine safety production, characterized in that, Specifically, the following steps are included: S1. Perform causal semantic analysis on the text of coal mine safety production accidents, construct a speculative expression judgment semantic matrix including modal word intensity value, causal trigger word weight value and sentence limitation range value, and determine whether the text of coal mine safety production accidents uses speculative expression based on the semantic vector distribution results. S1 specifically includes the following steps: S101. Perform causal semantic analysis on coal mine safety production accident texts, including sentence segmentation, part-of-speech tagging and dependency calculation of coal mine safety production accident texts, locate causal relationship statements between hidden danger entities and accident result entities, and extract modal words, causal trigger words and sentence limiting components from causal relationship statements to form a set of causal semantic elements. S102. Based on the set of causal semantic elements, intensity values ​​are assigned to modal words, directional weight values ​​are assigned to causal trigger words, and range values ​​are marked for sentence limiting components. The intensity values ​​of modal words, weight values ​​of causal trigger words, and sentence limiting range values ​​are vectorized and encoded according to a unified dimension to construct a speculative expression judgment semantic matrix. S103. Calculate the semantic vector distribution of the semantic matrix for speculative expression judgment, determine the judgment distance value between the deterministic semantic center and the speculative semantic center, compare the judgment distance value with the preset judgment threshold, and judge whether the coal mine safety production accident text uses speculative expression based on the comparison result. S2. When the judgment result is to use speculative expression, extract the modal word intensity gradient, semantic dependency span and sentence constraint convergence, and perform unified scale mapping to generate causal semantic uncertainty parameters. S3. Based on the causal semantic uncertainty parameter, divide the interval and map the causal semantic uncertainty parameter to the deterministic interval and the non-deterministic interval. Based on the mapping result, distinguish between high deterministic causal relationships and low deterministic causal relationships. S4. Based on the differentiation results, perform intensity mapping calculations on high-determinism causal relationships and low-determinism causal relationships respectively, generate causal intensity coefficients, and write the causal intensity coefficients into the corresponding causal directed edge weight values; S5. During the knowledge graph update process, acquire accident verification data. Based on the deviation between the accident verification data and the causal strength coefficient, adjust the causal semantic uncertainty parameter and the causal strength coefficient proportionally to achieve dynamic updating of the causal relationship strength.

2. The knowledge graph construction method in the field of coal mine safety production according to claim 1, characterized in that, S103 specifically refers to: The semantic vectors in each row of the speculative expression judgment semantic matrix are standardized, and the mean and variance of the vectors are calculated according to the unified vector dimension. Based on the standardization results, deterministic semantic center vectors and speculative semantic center vectors are constructed. The deterministic semantic center vectors and speculative semantic center vectors are both calculated from a pre-labeled set of coal mine safety production accident text sample vectors. The semantic vectors in the speculative expression judgment semantic matrix are distanced to the deterministic semantic center vector and the speculative semantic center vector, respectively. The first distance value from the semantic vector to the deterministic semantic center vector and the second distance value from the semantic vector to the speculative semantic center vector are obtained. The difference between the first distance value and the second distance value is calculated as the judgment distance value. The judgment distance value is compared with the preset judgment threshold. When the judgment distance value is greater than the preset judgment threshold, the text of the coal mine safety production accident is judged to use speculative expression. When the judgment distance value is less than or equal to the preset judgment threshold, the text of the coal mine safety production accident is judged not to use speculative expression.

3. The knowledge graph construction method in the field of coal mine safety production according to claim 1, characterized in that, S2 specifically includes the following steps: S201. When the judgment result is a speculative expression, the modal word intensity gradient, semantic dependency span, and statement constraint convergence are extracted from the causal relationship statement. The modal word intensity gradient is obtained by numerically sorting the level positions of modal words in the preset semantic intensity sequence and calculating the difference between adjacent levels. The semantic dependency span is obtained by calculating the node hierarchy distance between the hidden danger entity and the accident result entity in the dependency relationship path. The statement constraint convergence is obtained by statistically analyzing the number of semantic nodes and the hierarchy range covered by the statement constraint component in the dependency structure and performing numerical quantification. S202. Perform unified scale mapping processing on the extracted modal word intensity gradient, semantic dependency span and sentence constraint convergence, including interval normalization and distribution rearrangement of each semantic feature value, so that the modal word intensity gradient, semantic dependency span and sentence constraint convergence are mapped to a unified numerical range, and feature vectors are constructed according to a fixed dimension order. S203. Based on the feature vectors after unified scale mapping, perform weighted combination calculation, and numerically fuse the modal word intensity gradient, semantic dependency span and sentence constraint convergence according to the preset weight ratio to generate the causal semantic uncertainty parameter of the corresponding causal relationship sentence.

4. The knowledge graph construction method in the field of coal mine safety production according to claim 3, characterized in that, S201 specifically refers to: When the judgment result is to use speculative expression, all modal words are located in the causal relationship statement and mapped to the preset semantic intensity sequence. The modal words are encoded in order of semantic intensity from low to high. Multiple modal words appearing in the same causal relationship statement are sorted in the order of semantic intensity level. The numerical difference between adjacent levels is calculated and accumulated to obtain the modal word intensity gradient. Based on the dependency relationship path, the shortest dependency path between the hidden danger entity and the accident result entity is determined. The number of nodes and the path level depth in the shortest dependency path are counted. The number of nodes and the path level depth are weighted and summed to obtain the semantic dependency span. In the dependency structure, identify the set of modifier nodes corresponding to the qualifiers of the statement, count the number of semantic nodes covered by the set of modifier nodes and the corresponding hierarchical range, and quantify the numerical value according to the product of the number of covered nodes and the hierarchical range to obtain the statement qualifier convergence. The modal word intensity gradient, semantic dependency span and statement qualifier convergence are output as independent feature values.

5. The knowledge graph construction method in the field of coal mine safety production according to claim 1, characterized in that, S3 specifically refers to A parameter distribution sequence is constructed based on the causal semantic uncertainty parameters corresponding to multiple causal relation statements. The causal semantic uncertainty parameters are sorted according to their numerical values, and the interval division boundary value is determined according to the preset distribution statistical interval. The interval division is completed by using the interval division boundary value as the dividing point. Based on the interval division boundary value, the causal semantic uncertainty parameter is mapped to a deterministic interval and an indeterminate interval. The causal semantic uncertainty parameter less than or equal to the interval division boundary value is mapped to the deterministic interval, and the causal semantic uncertainty parameter greater than the interval division boundary value is mapped to the indeterminate interval. A corresponding interval identifier is generated for each causal semantic uncertainty parameter. The causal relationships of corresponding causal association statements are distinguished based on the interval identifier. Causal relationships mapped to deterministic intervals are marked as high-deterministic causal relationships, and causal relationships mapped to non-deterministic intervals are marked as low-deterministic causal relationships. The high-deterministic and low-deterministic causal relationships are then written into the causal relationship attribute field in the knowledge graph construction process.

6. The knowledge graph construction method in the field of coal mine safety production according to claim 1, characterized in that, S4 specifically refers to: Based on the differentiation results, the causal semantic uncertainty parameters corresponding to high-determinism causal relationships and low-determinism causal relationships are obtained respectively. Based on the interval division boundary value, the high-determinism intensity mapping interval and the low-determinism intensity mapping interval are determined. The high-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is less than or equal to the interval division boundary value, and the low-determinism intensity mapping interval is the numerical interval in which the causal semantic uncertainty parameter is greater than the interval division boundary value. Preset high-determinism mapping ratios and preset low-determinism mapping ratios are set for the high-determinism mapping interval and the low-determinism mapping interval, respectively. The preset high-determinism mapping ratio is a proportional coefficient determined based on sample statistics and is used to perform proportional transformation calculations on the causal semantic uncertainty parameters within the high-determinism mapping interval. The preset low-determinism mapping ratio is a proportional coefficient determined based on sample distribution decay rules and is used to perform decreasing proportional transformation calculations on the causal semantic uncertainty parameters within the low-determinism mapping interval to generate the corresponding causal intensity coefficient. When generating causal directed edges between potential hazard entities and accident result entities, the corresponding causal strength coefficient is written into the corresponding causal directed edge weight value field, and the causal strength coefficient is associated with the classification results of high-deterministic causal relationships or low-deterministic causal relationships and stored together.

7. The knowledge graph construction method in the field of coal mine safety production according to claim 1, characterized in that, S5 specifically refers to: During the knowledge graph update process, newly occurring coal mine safety production accident investigation texts and corresponding accident actual handling result records are collected. Causal relationship matching and positioning are performed on the accident investigation texts. The accident actual handling result records are correlated with the causal directed edges between the corresponding hidden danger entities and accident result entities in the knowledge graph, generating accident verification data that corresponds one-to-one with the causal directed edges. The actual occurrence result intensity value recorded in the accident verification data is compared with the causal intensity coefficient stored in the corresponding causal directed edge. The difference between the actual occurrence result intensity value and the causal intensity coefficient is calculated to form the deviation result. The proportional adjustment coefficient is determined according to the numerical direction and magnitude of the deviation result. Based on the proportional adjustment coefficient, the corresponding causal semantic uncertainty parameter and causal intensity coefficient are proportionally adjusted and calculated. The adjusted causal semantic uncertainty parameter is then re-involved in the interval division and intensity mapping calculation, and the adjusted causal intensity coefficient is written into the corresponding causal directed edge weight value field to realize the dynamic update of the causal relationship strength.

Images