Standard Text Semantic Analysis and Key Clause Extraction Methods and Systems

Abstract

This invention discloses a method and system for semantic analysis and key clause extraction of standard text. The method includes: performing multi-level semantic parsing on the standard text to identify the clauses to be evaluated and their corresponding types, extracting key parameter entities from the clauses to be evaluated, and constructing a clause-level semantic association structure; based on the semantic association structure, constructing a standard clause parameter knowledge graph, which supports semantic querying of technical parameters; based on the standard clause parameter knowledge graph, calculating and filtering the weights of the clauses to be evaluated, and generating a structured list of standard requirements summaries based on several selected key clauses and the parameter information associated with them in the corresponding knowledge graph. By combining multi-level semantic parsing and parameterized knowledge graph construction, the method achieves automated and structured extraction of the core requirements of standard text, solving the technical problems of low efficiency, incomplete coverage of key information, and lack of semantic association caused by relying on manual or shallow text processing.

Description

Technical Field

[0001] This invention relates to the field of text information processing technology, and in particular to a method and system for standard text semantic analysis and key clause extraction. Background Technology

[0002] Standard texts, such as technical specifications, quality requirements, and industry regulations, serve as crucial bases for industrial production and compliance management. These texts typically contain numerous clearly binding clauses that detail product technical parameters, testing conditions, and acceptance criteria. Quickly and accurately extracting these key clauses from a vast amount of standard texts is a critical step for companies in product development, quality control, and compliance self-inspection, significantly contributing to improved production efficiency and ensured product quality.

[0003] Currently, the interpretation and information extraction of standard texts still largely rely on manual work by domain experts. Experts manually read through the texts, identifying binding statements such as "should," "must not," and "must," and extracting specific technical parameters and thresholds. This method is not only inefficient—processing a complex standard document often takes several days—but the completeness and accuracy of the extracted results also heavily depend on the expert's personal experience and condition, making it prone to omissions or misinterpretations of key clauses due to subjective negligence. This makes it difficult to meet the demands of modern industry for rapid response and comprehensive coverage of standard applications.

[0004] To alleviate the burden on manual labor, several automated or semi-automated text mining methods have emerged in existing technologies. One common approach uses rule engines based on keyword matching, locating potential key clauses through a predefined vocabulary list. However, standard text is flexible and diverse; the same constraint semantics may have multiple expressions. Simple keyword matching struggles to handle complex linguistic phenomena, resulting in unsatisfactory recall and precision. Another approach attempts to utilize general natural language processing models for text analysis, but these models are not specifically optimized for the structured and logical features of standard text. They often fail to effectively distinguish the type of enforcement in clauses (such as "requirement" versus "prohibition"), and cannot accurately separate and associate the two core parameter entities scattered throughout the clauses: "test conditions" and "judgment criteria," leading to fragmented information and a lack of semantic coherence.

[0005] In recent years, knowledge graph technology has been widely used in the field of information organization due to its powerful semantic association and reasoning capabilities. However, in the specific field of standard text processing, how to automatically transform unstructured text into structured knowledge graphs rich in parametric semantics remains a significant technical bottleneck. The core difficulty lies in the lack of a semantic parsing mechanism that can deeply integrate clause type recognition and deep parameter extraction as a prerequisite for building high-quality knowledge graphs. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for standard text semantic analysis and key clause extraction. By combining multi-level semantic parsing, parameterized knowledge graph construction and intelligent summary generation, it realizes the automated and structured extraction of the core requirements of standard text, and solves the technical problems of low efficiency, incomplete coverage of key information and lack of semantic association caused by the reliance on manual or shallow text processing in traditional methods.

[0007] To address the aforementioned technical problems, a first aspect of this invention provides a method for standard text semantic analysis and key clause extraction, comprising the following steps:

[0008] Multi-level semantic parsing is performed on the standard text to identify the clauses to be evaluated and their corresponding types in the standard text, and the key parameter entities of the clauses to be evaluated are extracted. A clause-level semantic association structure is constructed based on the clause type and the key parameter entities.

[0009] Based on the semantic association structure, a standard clause parameter knowledge graph is constructed, which supports semantic querying of technical parameters;

[0010] Based on the knowledge graph of the standard clause parameters, the clauses to be evaluated are weighted and screened to obtain several key clauses.

[0011] Based on the selected key clauses and the corresponding parameter information associated in the knowledge graph, a structured list of standard requirements summaries is generated.

[0012] Furthermore, the step of performing multi-level semantic parsing on the standard text, identifying the clauses to be evaluated and their corresponding types in the standard text, extracting key parameter entities from the clauses to be evaluated, and constructing a clause-level semantic association structure based on the clause type and the key parameter entities includes:

[0013] By using a pre-defined clause classification model, the standard text is classified at the clause level to identify the clauses to be evaluated and determine their respective clause types, which include requirement clauses and prohibition clauses.

[0014] Using the named entity recognition model, the identified clauses to be evaluated are processed by parameter entity extraction, and condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds are extracted. The condition entities and the judgment criterion entities together constitute the key parameter entities.

[0015] Based on the clause type, the condition entity, and the judgment criterion entity, a tree-shaped semantic association structure is established with the clause to be evaluated as the root node and linked to the condition entity and the judgment criterion entity respectively, thus obtaining the clause-level semantic association structure.

[0016] Furthermore, the step of performing clause-level classification processing on the standard text using a preset clause classification model to identify the clauses to be evaluated and determine their respective clause types includes:

[0017] The standard text is segmented at the sentence level, and the segmentation results are corrected based on semantic integrity rules to obtain independent text units to be classified.

[0018] The independent text units are input into the preset clause classification model for semantic feature extraction and classification calculation. The preset clause classification model outputs a classification probability distribution representing the clause type.

[0019] Based on the classification probability distribution, text units containing preset constraint semantic patterns are identified as the clauses to be evaluated, and the corresponding clause types are determined.

[0020] The identified clauses to be evaluated are type-verified by a preset pattern matching rule base. When the clause type is inconsistent with the binding semantic pattern, the clause type is corrected according to the output of the pattern matching rule base.

[0021] Furthermore, the modification of the segmentation results based on semantic integrity rules includes:

[0022] Detect whether there are referential words in each text unit obtained from the initial segmentation. When the text unit contains a pronoun pointing to a technical entity, mark the text unit as the current text unit with referential dependency.

[0023] In the text sequence preceding the current text unit, find the text unit containing a description of the specific technical object referred to by the pronoun, and determine it as the antecedent text unit;

[0024] Based on the syntactic dependency relationship between the current text unit and the antecedent text unit, they are merged into a complete semantic unit;

[0025] Component analysis is performed on the semantic units to verify whether they constitute a declarative sentence structure that fully expresses the technical requirements;

[0026] The semantic units that pass structural verification are treated as independent text units that meet the requirements of semantic integrity.

[0027] Furthermore, the process of extracting parameter entities from the identified terms to be evaluated using the named entity recognition model, extracting condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds, includes:

[0028] The term to be evaluated is processed by word sequence labeling using a named entity recognition model based on sequence labeling architecture. Numerical value entities and their corresponding physical unit entities are identified and labeled. The numerical value entities and physical unit entities together constitute candidate parameter entities.

[0029] Based on the preset entity context dependency rules, the candidate parameter entities are classified into semantic roles. Candidate parameter entities that represent the test environment or operation state are classified as condition entities, and candidate parameter entities that represent the qualification judgment boundary are classified as judgment criterion entities.

[0030] The logical relationship between the condition entity and the decision criterion entity is established through dependency parsing. When both the condition entity and the decision criterion entity exist in the same clause to be evaluated, the correspondence between the condition entity and the decision criterion entity is constructed based on the syntactic parsing results, forming a complete parameter constraint pair.

[0031] Furthermore, the step of using a named entity recognition model based on sequence labeling architecture to perform word sequence labeling on the terms to be evaluated, identifying and labeling numerical value entities and their corresponding physical unit entities, includes:

[0032] The terms to be evaluated are segmented into words and tagged with parts of speech to generate a word sequence with grammatical tags;

[0033] The word sequence is input into the named entity recognition model, and the model outputs the label corresponding to each word in the preset entity label set through its sequence labeling architecture to obtain the entity label sequence.

[0034] Based on the entity tag sequence, identify and extract numerical value entities that represent specific quantities and their directly modified physical unit entities from the text;

[0035] The numerical value entity is parsed using compound expression parsing, and the continuous numerical range containing connectors is parsed into a standardized numerical representation with clear upper and lower limits;

[0036] Based on preset unit mapping rules, physical unit entities are standardized, and similar units with different expressions are uniformly converted into the standard unit form.

[0037] The standardized numerical value entity and the physical unit entity are semantically combined to generate a candidate parameter entity with complete dimensional information.

[0038] Further, the step of establishing a tree-like semantic association structure based on the clause type, the condition entity, and the judgment criterion entity, with the clause to be evaluated as the root node and linked to the condition entity and the judgment criterion entity respectively, includes:

[0039] The clause to be evaluated is taken as the root node of the tree structure, and a clause type attribute representing the constraint nature is added to the root node;

[0040] Using the conditional entity as the first type of child node, the conditional entity is linked to the corresponding root node by testing the conditional relationship edge;

[0041] Using the judgment criterion entity as the second type of child node, the judgment criterion entity is linked to the corresponding root node through the judgment basis relationship edge;

[0042] Based on the logical dependency relationship between the condition entity and the judgment criterion entity in the original text, constraint logic edges are established between the first type of child nodes and the second type of child nodes, thereby forming a tree structure with multi-layer semantic association.

[0043] Furthermore, based on the semantic association structure, a standard clause parameter knowledge graph is constructed. This knowledge graph supports semantic queries of technical parameters, including:

[0044] Map the clause nodes to be evaluated, condition entity nodes, and judgment criterion entity nodes in the clause-level semantic association structure to entity nodes in the knowledge graph.

[0045] Map the test condition relation edges, judgment basis relation edges, and constraint logic edges in the clause-level semantic association structure to relation edges in the knowledge graph;

[0046] Add type attributes representing test environment features to condition entity nodes in the knowledge graph, and add type attributes representing threshold features to judgment criterion entity nodes;

[0047] Based on the logical dependency relationships established in the semantic association structure, multi-hop semantic paths between entity nodes are constructed in the knowledge graph;

[0048] Establish an interface for converting natural language queries to knowledge graph queries, and map the technical parameter requirements in natural language queries to entity nodes and relation edges in the knowledge graph through semantic similarity calculation;

[0049] Based on the mapping results, semantic queries are performed in the knowledge graph, returning standard terms that meet the query conditions and their associated parameter entity information.

[0050] Furthermore, based on the knowledge graph of the standard clause parameters, the clauses to be evaluated are weighted and filtered to obtain several key clauses, including:

[0051] Based on the association structure features of the nodes of the terms to be evaluated in the knowledge graph, the initial weight value of each term to be evaluated is calculated. The association structure features include the number of connected parameter entities and the distribution of relation edge types.

[0052] The initial weight value is first corrected by combining the clause type attribute of the clause to be evaluated, and the type-corrected weight is obtained.

[0053] The frequency of occurrence of various parameter entities in the knowledge graph is statistically analyzed, and the type correction weights are subjected to a second enhancement process based on the frequency analysis results to obtain frequency enhancement weights.

[0054] Based on a pre-defined list of important parameter entities, the frequency enhancement weights are subjected to a third adjustment process to obtain the final weight value for each of the terms to be evaluated.

[0055] Based on the final weight values, the clauses to be evaluated are sorted, and the clauses with weight values ​​higher than a preset threshold are selected as key clauses.

[0056] Accordingly, a second aspect of the present invention provides a standard text semantic analysis and key clause extraction system, which performs text semantic analysis and key clause extraction based on the above-described standard text semantic analysis and key clause extraction method, including:

[0057] The text parsing module is used to perform multi-level semantic parsing on standard text, identify the clauses to be evaluated and their corresponding types in the standard text, extract the key parameter entities of the clauses to be evaluated, and construct a clause-level semantic association structure based on the clause type and the key parameter entities.

[0058] The graph construction module is used to construct a standard clause parameter knowledge graph based on the semantic association structure, and the knowledge graph supports semantic querying of technical parameters;

[0059] The clause filtering module is used to perform weight calculation and filtering on the clauses to be evaluated based on the standard clause parameter knowledge graph, and obtain several key clauses.

[0060] The list generation module is used to generate a structured list of standard requirements summaries based on several selected key clauses and the corresponding parameter information associated with the knowledge graph.

[0061] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects:

[0062] 1. By introducing a multi-level semantic parsing mechanism that integrates clause type identification and parameter entity extraction, and constructing a tree-like association structure of "clause-condition-criteria", the machine-readable parsing of deep semantics of standard text was realized for the first time. It effectively solved the core problem that traditional single keyword matching or general NLP models could not accurately distinguish the mandatory force of clauses and could not associate fragmented technical parameters, laying a high-quality structured data foundation for subsequent intelligent applications.

[0063] 2. By further constructing the semantic association structure into a knowledge graph rich in parameter attributes and establishing a conversion interface from natural language to graph query, users can perform complex condition searches in an intuitive semantic way; this fundamentally solves the technical bottleneck of traditional document retrieval or database query methods being unable to understand user intent or perform parameter-based logical association reasoning, thus resulting in insufficient depth and accuracy of information retrieval.

[0064] 3. By calculating comprehensive weights based on the knowledge graph topology, parameter entity frequency, and importance, and generating a structured summary list accordingly, a leap from "information extraction" to "knowledge refinement" is achieved. This completely solves the long-standing problems of strong subjectivity and low efficiency in manual interpretation of standards, as well as the fact that existing automated methods can only list text fragments and cannot provide prioritization, resulting in poor decision support. Attached Figure Description

[0065] Figure 1 This is a flowchart of the standard text semantic analysis and key clause extraction method provided in the embodiments of the present invention;

[0066] Figure 2 This is a block diagram of the standard text semantic analysis and key clause extraction system provided in the embodiments of the present invention.

[0067] Figure label:

[0068] 1. Text parsing module, 2. Graph construction module, 3. Clause filtering module, 4. List generation module. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0070] Please refer to Figure 1 The first aspect of this invention provides a method for standard text semantic analysis and key clause extraction, comprising the following steps:

[0071] Step S100: Perform multi-level semantic parsing on the standard text, identify the clauses to be evaluated and their corresponding types in the standard text, extract the key parameter entities of the clauses to be evaluated, and construct a clause-level semantic association structure based on the clause type and key parameter entities.

[0072] Multi-level semantic parsing of the standard text is the fundamental processing stage of this invention. This stage first performs clause-level classification processing on the standard text using a pre-set clause classification model. This model, built on a pre-trained language model, can identify sentences or paragraphs with binding semantics and classify them as requirement clauses or prohibition clauses. For example, in electrical safety standards, the model will identify the text "The insulation resistance of electrical equipment should not be less than 100MΩ after a damp heat test" as a requirement clause. Simultaneously, a named entity recognition model extracts parameter entities from the identified clauses to be evaluated. This model uses a sequence labeling architecture to accurately locate and extract the specific values, units, and constraints representing technical indicators. From the above clauses, the model extracts "damp heat test" as a condition entity to describe the test environment; and extracts "100MΩ" as a judgment criterion entity to define the pass / fail threshold.

[0073] After completing clause type identification and parameter entity extraction, a clause-level semantic association structure is constructed based on these elements. This structure uses the clause to be evaluated as the root node, the clause type as an attribute of this node, and the condition entities and judgment criterion entities as child nodes, connected to the root node through specific semantic relationship edges. Taking the aforementioned electrical equipment insulation resistance clause as an example, "the insulation resistance of electrical equipment shall not be less than 100MΩ" is the root node, linked to the "damp heat test" child node through the "test conditions" relationship edge, and linked to the "100MΩ" child node through the "judgment basis" relationship edge. This tree structure clearly expresses the complete logic of "under what conditions what standard needs to be met," transforming the originally unstructured text into a semantic network that machines can understand and process.

[0074] Step S200: Based on the semantic association structure, construct a knowledge graph of standard clause parameters. The knowledge graph supports semantic querying of technical parameters.

[0075] Based on the clause-level semantic association structure output in step S100, a standard clause parameter knowledge graph is constructed. The construction process first maps all nodes in the semantic association structure (clauses to be evaluated, condition entities, and judgment criterion entities) to entity nodes in the knowledge graph, and maps the semantic relationships between nodes (test condition relationship edges, judgment basis relationship edges, etc.) to relationship edges in the knowledge graph. Each entity node is assigned rich attributes. For example, in electrical safety standards, attributes such as "test voltage," "duration," and "ambient humidity" are added to the "withstand voltage test" condition entity node; and attributes such as "lower resistance limit," "measured voltage," and "pass / fail standard" are added to the "insulation resistance" judgment criterion entity node.

[0076] This knowledge graph supports semantic queries of technical parameters, establishing an interface for converting natural language to graph query statements. Through semantic similarity calculation, it maps the user's query intent in natural language (such as "find all electrical equipment standard clauses requiring insulation resistance greater than 10MΩ") to corresponding entity nodes and relation edges in the knowledge graph. The graph engine then executes the query, not simply matching keywords, but reasoning based on the graph's semantic network, returning all standard clauses that meet the parameter condition and their complete technical requirement context. This allows users to quickly and accurately locate the required technical constraint information from massive amounts of standard documents.

[0077] Step S300: Based on the standard clause parameter knowledge graph, the weights of the clauses to be evaluated are calculated and screened to obtain several key clauses.

[0078] The knowledge graph constructed in step S200 is used to achieve intelligent weight calculation and screening of clauses to be evaluated. The weight calculation is not based on a single factor, but comprehensively considers the multi-dimensional information provided by the knowledge graph. The initial weight is calculated based on the relational structure characteristics of the clause to be evaluated in the graph. For example, in electrical safety standards, if a clause is associated with multiple key safety parameter entities such as "dielectric strength," "insulation resistance," and "leakage current," and the relationship edge types are complex, its initial weight is higher. Subsequently, it is modified in combination with the clause type attributes. Generally, requirement clauses are given a higher weight coefficient than prohibition clauses.

[0079] Furthermore, the frequency of occurrence of various parameter entities throughout the knowledge graph is statistically analyzed. Clauses containing high-frequency parameter entities are weighted more heavily, as these parameters are often the focus of attention in that technical field. For example, in electrical safety standards, parameters such as "insulation resistance," "withstand voltage," and "grounding resistance" typically appear frequently. Simultaneously, a list of important parameter entities (such as core indicators related to electrical safety) is pre-defined, and clauses associated with entities on this list undergo additional weight adjustments. Through this series of progressive calculations and adjustments, a comprehensive final weight value is obtained for each clause to be evaluated. Based on this, clauses are ranked, and those with weight values ​​higher than a preset threshold are selected as key clauses for this analysis.

[0080] Step S400: Based on the selected key clauses and the parameter information associated with them in the corresponding knowledge graph, generate a structured list of standard requirements summaries.

[0081] The structured list of standard requirements summaries is not simply a list of the selected key clauses; rather, it is based on a standard clause parameter knowledge graph, extracting key clauses and all their associated parameter information. The summary generator will organize the text content of the key clauses, all their test conditions (condition entities), and qualification criteria (judgment criterion entities) according to a preset template, forming a clear and complete structured entry.

[0082] For example, in electrical safety standards, a key clause regarding "electrical clearance" would generate a summary entry that clearly states the clause content: "The electrical clearance between live parts shall not be less than 3 mm," with sub-items listing the test conditions "under pollution level 2" and the judgment criterion "Electrical clearance ≥ 3 mm." Ultimately, all such summary entries are compiled into a unified list of standard requirements summaries. This list significantly improves the readability and usability of the core requirements of the standard, enabling engineers, quality control personnel, or compliance experts to quickly grasp the essence of the standard and use it to guide product design, testing and verification, or compliance reviews.

[0083] The above steps enable a leap from automated deep analysis of standard texts to knowledge-based and structured applications, ultimately significantly improving the efficiency and accuracy of standard interpretation, ensuring the coverage of key clause identification, and providing direct and reliable data support for standard-based intelligent decision-making.

[0084] Specifically, step S100 involves performing multi-level semantic parsing on the standard text, identifying the clauses to be evaluated and their corresponding types within the standard text, extracting key parameter entities from the clauses to be evaluated, and constructing a clause-level semantic association structure based on the clause type and key parameter entities, including:

[0085] Step S110: The standard text is classified at the clause level using a preset clause classification model to identify the clauses to be evaluated and determine their clause types, which include requirement clauses and prohibition clauses.

[0086] The input standard text is automatically processed using a pre-defined clause classification model to identify binding clauses to be evaluated and determine their types. This process first performs sentence-level segmentation, breaking down the complete document into independent text units. This segmentation is not simply done by period; it incorporates semantic integrity rules for correction. For example, in electrical safety standards, when a sentence contains referential terms such as "its insulation performance" or "the electrical clearance," the system automatically searches for the sentences containing these antecedents and merges multiple sentences with referential dependencies into a single complete semantic unit, ensuring that each text unit to be evaluated is semantically independent and complete.

[0087] Next, these independent text units are input into a pre-defined clause classification model for processing. This model, built on a pre-trained language model, is capable of deeply understanding the semantic meaning of the text. The model extracts semantic features and performs classification calculations on the text units, outputting a classification probability distribution representing the probability of each clause type. Based on this probability distribution, and combined with a pre-defined constraint semantic pattern library (e.g., containing key patterns such as "should," "must," "must not," and "prohibited"), a comprehensive judgment is made, identifying text units containing strong constraint semantics as clauses to be evaluated, and simultaneously determining their specific type as "requirement type" or "prohibition type." In electrical safety standards, the clause "the protection level of the equipment enclosure should not be lower than IP54" is accurately identified as a requirement type clause. To further improve classification accuracy, a pre-defined pattern matching rule library is introduced for secondary verification. When the clause type initially determined by the model is inconsistent with the deep constraint pattern contained in the clause text in the rule library, the type is corrected based on the output of the rule library, thus forming a final reliable clause type determination result.

[0088] Step S120: Using the named entity recognition model, parameter entity extraction processing is performed on the identified clauses to be evaluated. Condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds are extracted. The condition entities and judgment criterion entities together constitute key parameter entities.

[0089] Building upon step S110, the focus shifts to precisely extracting key technical parameter entities from the identified terms to be evaluated. This process is accomplished using a specialized named entity recognition model, which typically employs a sequence labeling architecture. The model first performs word segmentation and part-of-speech tagging on the terms to be evaluated, generating a sequence of words with grammatical tags. Subsequently, the model, through its internal sequence labeling mechanism, assigns a tag from a pre-defined entity tag set to each word in the sequence, thereby accurately locating numerical value entities (such as "100MΩ", "3mm", "5mA") representing specific quantities in the text, as well as their directly modified, indivisible physical unit entities (such as "MΩ", "mm", "mA").

[0090] For the identified basic entities, deep standardization processing is performed. For example, for a compound expression like "1000V±5%", the model will parse it into a standardized numerical representation with clear upper and lower limits (950V to 1050V). Simultaneously, based on a pre-defined unit mapping rule base, physical unit entities are standardized and transformed, for example, "kilovolt" and "kV" are uniformly converted to "kV", and "milliampere" and "mA" are uniformly converted to "mA", ensuring consistency in unit representation. Finally, the standardized numerical quantity entities and physical unit entities are semantically combined to generate candidate parameter entities with clear numerical and dimensional information. These candidate parameter entities are then classified into semantic roles according to pre-defined entity context dependency rules. Entities representing the test environment, operating state, or preconditions are classified as condition entities (e.g., "damp heat test", "under rated voltage"), while entities representing the qualification judgment boundary are classified as judgment criterion entities (e.g., "insulation resistance ≥100MΩ", "leakage current ≤0.5mA"). Together, they constitute the key parameter entities.

[0091] Step S130: Based on the clause type, condition entity, and judgment criterion entity, establish a tree-shaped semantic association structure with the clause to be evaluated as the root node and linked to the condition entity and judgment criterion entity respectively, to obtain the clause-level semantic association structure.

[0092] The outputs of the preceding steps are integrated into a machine-understandable and semantically rich structured representation. This step constructs a tree-like semantic association structure based on the identified clause types, condition entities, and decision criterion entities. This structure uses the text content of the clause to be evaluated itself as the root node of the entire tree, and attaches a key attribute to this root node—namely, the clause type (requirement type or prohibition type) determined in S110.

[0093] Then, all condition entities extracted from S120 under this clause are treated as first-type child nodes, and connected to their respective root nodes one by one through a link named "Test Condition Relationship Edge". Similarly, all decision criterion entities are treated as second-type child nodes, and connected to the root node through a link named "Decision Basis Relationship Edge". In the electrical safety standard, for the clause "At altitudes below 2000m, electrical clearances shall not be less than 3mm", a "Above 2000m altitude" condition entity node and a "3mm" decision criterion entity node are established, and connected to the root node through the test condition relationship edge and the decision basis relationship edge, respectively.

[0094] To more accurately reflect the technical logic, "constraint logic edges" are established between condition entities and decision criterion entities based on the dependency parsing results of the original text. These edges explicitly indicate the specific conditions under which a particular decision criterion takes effect. For example, in complex clauses of electrical safety standards, the logical relationship "under pollution level 2 conditions, the electrical clearance must meet X mm; under pollution level 3 conditions, the electrical clearance must meet Y mm" can be precisely expressed. The resulting tree-like semantic association structure fully encapsulates the constraints, preconditions, and compliance criteria of a standard clause, providing direct and standardized input for subsequent knowledge graph construction.

[0095] The above process enables a leap from shallow lexical analysis to deep semantic structure construction of standard text, transforming unstructured natural language clauses into machine-readable structured data with clear type attributes, parameter composition, and internal logical relationships, laying a precise and reliable data foundation for subsequent graph-based storage, intelligent querying, and knowledge application.

[0096] Further, step S110 involves performing clause-level classification processing on the standard text using a preset clause classification model to identify the clauses to be evaluated and determine their respective clause types, including:

[0097] Step S111: Perform sentence-level segmentation on the standard text and correct the segmentation results based on semantic integrity rules to obtain independent text units to be classified.

[0098] Sentence-level segmentation of standard text is a fundamental step in semantic parsing. This process first employs a basic segmentation algorithm based on punctuation and line breaks to divide the complete standard document content into independent text units. These segmented text units then undergo further verification using semantic integrity rules to ensure that each unit is semantically self-contained.

[0099] The core function of semantic integrity rules is to identify and process referential dependencies in text. In electrical safety standard applications, when an incomplete reference is detected in a text unit, such as "its insulation strength" or "this protection level," the unit is automatically marked as a current text unit with referential dependencies. Subsequently, the text sequence preceding the current text unit is scanned to locate the antecedent text units containing descriptions of the specific technical object referred to.

[0100] After determining the syntactic relationship between the current text unit and its antecedent text unit through syntactic dependency analysis, these two text units are semantically merged to form a complete semantic unit. The merged semantic unit then undergoes subject-verb-object analysis to verify whether it constitutes a declarative sentence structure expressing a complete technical requirement. For example, in electrical safety standards, for a text unit like "its insulation resistance value should not be less than 100MΩ," it is verified whether it contains a clear technical subject, technical action, and technical indicator. Only semantic units that pass structural verification are ultimately determined as independent text units that meet the semantic integrity requirements and proceed to subsequent processing.

[0101] Step S112: Input the independent text units into the preset clause classification model for semantic feature extraction and classification calculation. The preset clause classification model outputs the classification probability distribution representing the clause type attribution.

[0102] Inputting individual text units into a pre-defined clause classification model is a crucial step in clause recognition. This classification model is built upon a pre-trained language model and possesses deep semantic understanding capabilities. The model first performs word embedding representation on the input text units, converting them into numerical representations in a high-dimensional vector space.

[0103] In the semantic feature extraction stage, the model captures key semantic features in the text through a multi-layer self-attention mechanism. These features include not only the surface meaning of words but also their deeper semantic meaning in different contexts. In the analysis of electrical safety standards, the model pays particular attention to the binding expression patterns in the text, such as linguistic features with electrical safety constraints, such as "should be grounded," "must be insulated," and "must not be short-circuited."

[0104] The classification calculation stage utilizes fully connected layers and a softmax activation function to map the extracted semantic features onto the probability distribution of clause types. The classification probability distribution output by the model accurately reflects the likelihood of a text unit belonging to each clause type. For example, for the text unit "the enclosure of electrical equipment must be reliably grounded," the model will output a higher probability of a requirement clause. The entire processing fully considers the complex sentence structures and technical terms commonly found in standard texts.

[0105] Step S113: Based on the classification probability distribution, text units containing preset constraint semantic patterns are identified as clauses to be evaluated, and the corresponding clause types are determined.

[0106] Identifying clauses to be evaluated based on classification probability distribution is a multi-factor decision-making process. First, a probability threshold is set, and text units with a classification probability higher than this threshold are initially marked as candidate clauses to be evaluated. This threshold setting needs to strike a balance between recall and precision, ensuring that as many potential key clauses as possible are captured.

[0107] A pre-defined constraint semantic pattern library plays a crucial role in this step. This library systematically includes common constraint expressions found in standard texts. In electrical safety standards, this includes explicit requirement patterns such as "must be reliably grounded" and "must be insulated," as well as prohibition patterns such as "must not be operated while energized" and "must not be short-circuited." Matching candidate clauses to be evaluated with the expression patterns in the library further confirms their constraint characteristics.

[0108] For successfully matched text units, the final clause type is determined by combining classification probability distribution and semantic pattern features. For example, in electrical safety standards, "all exposed conductive parts shall be reliably grounded" is identified as a requirement clause, while "it is strictly forbidden to open the casing of equipment while it is energized" is identified as a prohibition clause. This comprehensive judgment mechanism effectively improves the accuracy of clause type identification.

[0109] Step S114: The identified clauses to be evaluated are verified by using a preset pattern matching rule base. When the clause type is inconsistent with the binding semantic pattern, the clause type is corrected according to the output of the pattern matching rule base.

[0110] Type validation using a pre-defined pattern matching rule base is crucial for ensuring the accuracy of classification results. This rule base contains a large number of validated semantic pattern instances, covering various complex electrical safety expressions in electrical safety standards, such as special expressions like double insulation requirements and grounding continuity tests.

[0111] When an inconsistency is detected between the clause type and the binding semantic pattern, a type correction process is initiated. For example, in electrical safety standards, certain complex sentences containing negative words, such as "non-double-insulated equipment shall not be used in damp environments," may be misclassified. Correction rules stored in the rule base will provide processing guidance for these specific cases.

[0112] Automatically correct the clause types in question based on the output result of the pattern matching rule library. The correction process may involve re - analyzing the text structure, deeply analyzing the semantics of keywords, or comprehensively considering the context. In the analysis of electrical safety standards, this verification and correction mechanism can accurately identify the differences between professional concepts such as "functional grounding" and "protective grounding", significantly improving the reliability of clause type determination and providing an accurate basis for subsequent processing.

[0113] Through the complete processing flow composed of the above four sub - steps, the accurate identification and classification of key clauses in the electrical safety standard text are realized, converting unstructured text information into structured data with clear type labels, laying a solid foundation for subsequent parameter extraction and knowledge graph construction, and effectively improving the automation level and accuracy of electrical safety standard text processing.

[0114] Furthermore, the correction of the segmentation result based on the semantic integrity rule in step S111 includes:

[0115] Step S1111, detect whether there are referential words in each text unit obtained by primary segmentation. When the text unit contains a pronoun referring to a technical entity, mark this text unit as the current text unit with referential dependence.

[0116] Detecting referential words in the text units obtained by primary segmentation is the primary link of semantic integrity processing. Adopt a method combining rules and statistics to conduct lexical - level analysis on each text unit to identify possible referential phenomena. This process not only focuses on common pronouns such as "its", "the", "this", etc., but also includes demonstrative pronouns and noun phrases with referential functions.

[0117] In the application scenario of electrical safety standards, maintain a dedicated referential vocabulary library in the field of electrical safety, which contains common referential expressions in the standard text of this field, such as "this insulation level", "its protection performance", "this electrical clearance", etc. When these words appear in a text unit, further analyze their context to determine whether they actually refer to a certain technical entity. This judgment process comprehensively considers the grammatical role, semantic features of the word and other words co - occurring with it.

[0118] For the referential situation that is confirmed to refer to a technical entity, mark this text unit as the current text unit with referential dependence. For example, in the electrical safety standard, for the text unit containing "its dielectric strength", mark this unit as having referential dependence and record that it refers to a specific electrical insulation component. This marking not only records the existence of the reference, but also records the type of reference, its position in the text, and the possible reference range. These information provide the necessary basic data for subsequent antecedent word search and semantic merging.

[0119] Step S1112: In the text sequence preceding the current text unit, find the text unit containing a description of the specific technical object referred to by the pronoun, and determine it as the antecedent text unit.

[0120] After identifying the current text unit with a referential dependency, it is necessary to search for the corresponding antecedent text unit in the preceding text sequence. The search process employs a dual constraint strategy based on syntax and semantics to ensure that the found antecedents are grammatically sound and semantically matching.

[0121] First, the search is performed within a limited window preceding the current text unit. This window size is typically set to the first 3-5 sentences to balance accuracy and efficiency. In electrical safety standard analysis, special attention is paid to statements containing definitions of electrical equipment components, descriptions of insulation material properties, or specifications of protection levels.

[0122] For each candidate antecedent text unit, matching is performed by calculating the semantic similarity between the pronoun and the candidate antecedent. This calculation considers not only the similarity on the lexical surface but also evaluates their distance in the semantic space using a word vector model. For example, in electrical safety standards, the semantic relevance between "this insulation class" and "polyvinyl chloride insulation material" among the candidate antecedents is verified. Simultaneously, the consistency between the candidate antecedents and the pronouns in grammatical features such as gender, number, and case is examined.

[0123] The final antecedent text units must meet multiple conditions, including high semantic relevance, matching grammatical features, and relatively close position. In electrical safety standards, for the current text unit referring to "its withstand voltage level," text units containing the description of "primary winding of a power transformer" are identified as its antecedents. This strict matching mechanism ensures the accuracy of reference resolution and lays the foundation for subsequent semantic unit merging.

[0124] Step S1113: Based on the syntactic dependency relationship between the current text unit and the antecedent text unit, merge them into a complete semantic unit.

[0125] Based on the syntactic dependency relationship between the current text unit and the antecedent text unit, these two text units are merged into a complete semantic unit. The merging process first performs syntactic analysis on the two text units to build their respective dependency relationship trees, and then identifies the key dependency relationships connecting the two units.

[0126] During the syntactic analysis phase, a dependency parser is used to parse the grammatical structure of text units, identifying key components such as core predicates, subjects, and objects. In electrical safety standards, this analysis helps to understand the logical connections between two text units, especially how antecedents are associated with electrical equipment such as "circuit breaker" or "fuse" in the current text unit, and with predicate verbs such as "should be disconnected" or "must be melted."

[0127] Merging text units requires maintaining semantic coherence and grammatical correctness. Based on the grammatical structure of the two text units, an appropriate connection method should be selected, such as using relative pronouns, parallel connections, or complex sentences. In electrical safety standards, for merging "low-voltage distribution cabinet" and "its protection level should reach IP44," an appropriate complex sentence connection method should be chosen. During the merging process, duplicate sentence components should be eliminated, and word order should be adjusted to ensure that the merged semantic unit conforms to Chinese expression habits.

[0128] The merged semantic units need to undergo a consistency check to ensure that the referential relationships are correctly resolved and that the merged sentences are grammatically and semantically complete. For example, in electrical safety standards, the merged statement "the protection level of low-voltage distribution cabinets should reach IP44" is a complete and semantically clear expression. The output of this step is a text unit with a complete semantic expression, which prepares the text for subsequent syntactic structure verification.

[0129] Step S1114: Perform component analysis on the semantic unit to verify whether it constitutes a declarative sentence structure that fully expresses the technical requirements.

[0130] Component analysis of the merged semantic units is a crucial step in verifying their structural integrity. A method combining grammatical rules and statistical models is employed to perform deep syntactic analysis on the semantic units. This analysis process first identifies the main sentence components within the semantic units, including subject, predicate, object, attributive, and adverbial modifiers.

[0131] During component analysis, special attention is paid to the unique expression methods of electrical safety standard texts. For technical requirement statements, it is verified whether they contain clearly defined electrical safety technical subjects, safety technical actions, and safety technical indicators. The completeness of these elements is an important basis for judging whether a semantic unit expresses complete technical requirements.

[0132] Using a parsing model trained on a corpus of electrical safety standard texts, the model can accurately identify common sentence structures in electrical safety standards. For example, for a statement such as "the enclosure of electrical equipment must have a grounding mark in a conspicuous location," it is necessary to identify "enclosure of electrical equipment" as the subject, "have" as the predicate, "grounding mark" as the object, and "in a conspicuous location" as the adverbial, and verify that these components together constitute a complete statement of electrical safety requirements.

[0133] The verification process also includes checking the logical integrity of the semantic unit. It assesses whether the statement expresses complete electrical safety constraints, including the range of safety parameters, the explicit specification of test conditions, and the quantitative requirements of safety performance indicators. In electrical safety standards, only semantic units that pass all these verifications are considered structurally complete, such as statements like "the electrical clearance between live parts and accessible surfaces should be no less than 6 mm."

[0134] Step S1115: The semantic units that pass the structural verification are treated as independent text units that meet the semantic integrity requirements.

[0135] Treating the semantically validated units as independent text units that meet semantic integrity requirements is the final step in semantic integrity processing. These validated text units possess complete syntactic structures and clear electrical safety technical semantics, and can serve as reliable input for subsequent clause classification and parameter extraction.

[0136] A standardized representation is generated for each passed semantic unit, including its text content, syntactic structure information, semantic role annotation, and related metadata. In electrical safety standard processing, this metadata includes electrical safety-specific information such as equipment type, safety level, and applicable environment. This information will be used in subsequent processing flows to ensure that the results of semantic integrity processing are effectively utilized.

[0137] Before output, the approved semantic units undergo a quality assessment. Assessment metrics include syntactic complexity, semantic clarity, and the completeness of electrical safety technical elements. For example, for a semantic unit involving "double insulation" requirements, the assessment evaluates whether it fully includes key elements such as insulation structure, testing methods, and acceptance criteria. This assessment helps subsequent processing workflows adopt appropriate processing strategies based on the characteristics of the text units.

[0138] The final set of independent text units constitutes a semantically complete representation of the electrical safety standard text, providing high-quality foundational data for subsequent clause type identification and parameter entity extraction. These text units retain the technical semantics of the original electrical safety standard text while possessing good machine-processable structural characteristics.

[0139] Further, in step S120, the identified clauses to be evaluated are processed by extracting parameter entities using the named entity recognition model. This process extracts condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds, including:

[0140] Step S121: Using a named entity recognition model based on sequence labeling architecture, word sequence labeling is performed on the terms to be evaluated to identify and label numerical value entities and their corresponding physical unit entities. The numerical value entities and physical unit entities together constitute candidate parameter entities.

[0141] The named entity recognition model based on sequence labeling architecture employs a hierarchical processing mechanism to achieve accurate identification of parametric entities. The model input uses a word embedding layer finely tuned to a standard text corpus in the electrical safety field, enabling accurate mapping of electrical safety terminology to a high-dimensional semantic space. During processing, the model's bidirectional long short-term memory network layer fully captures contextual dependencies within the text sequence through forward and backward information transmission. For example, when analyzing the statement "the dielectric strength test voltage should be 2.5kV±5%" in electrical safety standards, the model can understand the modifying relationship between "±5%" and "2.5kV," as well as the limiting effect of "test voltage" on electrical safety performance.

[0142] During entity recognition, the model must handle not only simple combinations of numerical units but also complex parametric expressions. A multi-labeling strategy is employed, which, in addition to labeling basic numerical values ​​and physical units, also identifies key information such as operators (e.g., ">" in withstand voltage testing, "≥" in insulation resistance, and "<" in leakage current), tolerance symbols (e.g., ±), and range symbols (e.g., "~" in electrical parameters). For complex parameter expressions, such as "insulation resistance test voltage 500V / 1000V," the model decomposes it into voltage values ​​under different test conditions. Furthermore, a dedicated post-processing module is established to validate and standardize the model output using regular expressions and a rule engine, ensuring consistent processing of different expressions in the electrical safety field, such as "kilovolt" versus "kV" and "megaohm" versus "MΩ."

[0143] Step S122: Based on the preset entity context dependency rules, the candidate parameter entities are classified into semantic roles. Candidate parameter entities that represent the test environment or operation state are classified as condition entities, and candidate parameter entities that represent the qualification judgment boundary are classified as judgment criterion entities.

[0144] The pre-defined entity context dependency rule base employs a dual mechanism based on semantic role labeling and syntactic structure analysis to achieve accurate classification of parameter entities. The rule base not only includes lexical-level matching patterns but also defines complete syntactic templates to describe the typical occurrence environments of conditional entities and decision criterion entities in electrical safety standards. For example, in electrical safety standards, conditional entities typically appear as adverbial phrases such as "under the condition of..." or "when...", while decision criterion entities often appear in predicate nominative positions such as "should not be lower than" or "must not exceed" or as verb complements. A comprehensive judgment is made by analyzing the entity's position in the dependency syntax tree and its syntactic relationships with other components, combined with semantic information from an electrical safety domain dictionary.

[0145] In practical applications, complex scenarios involving multiple conditions and criteria need to be handled. Taking electrical equipment safety standards as an example, in the clause "Under the conditions of an ambient temperature of 40℃ and a relative humidity of 90%, the insulation resistance value shall not be less than 100MΩ," it is possible to accurately identify that "ambient temperature of 40℃" and "relative humidity of 90%" are conditional entities, while "100MΩ" is the judgment criterion entity. By analyzing the parallel structure and logical relationships in the sentence, the correspondence between multiple conditional entities and a single judgment criterion entity can be established. For more complex cases, such as "The leakage current shall not exceed 0.5mA at AC 50Hz and shall not exceed 0.25mA at DC," the correct pairing relationship between the power supply type condition and the corresponding leakage current criterion can be established by analyzing the structure of the conditional adverbial clause.

[0146] Step S123: Establish the logical relationship between the condition entity and the decision criterion entity through dependency parsing. When both the condition entity and the decision criterion entity exist in the same clause to be evaluated, construct the correspondence between the condition entity and the decision criterion entity based on the syntactic parsing results to form a complete parameter constraint pair.

[0147] Dependency parsing employs deep semantic parsing techniques to establish logical relationships between parameter entities. Using a high-precision dependency parser based on neural networks, it can identify long-distance dependencies and complex sentence structures in Chinese. During the analysis, it focuses on syntactic components closely related to parameter constraints, such as conditional adverbs, comparative adverbs, and result complements commonly found in electrical safety standards, accurately extracting the modification relationships and logical connections between conditional entities and decision criterion entities.

[0148] When processing electrical safety technical clauses, it can identify various types of logical association patterns. For example, in the electrical safety standard clause "Electrical equipment operating at altitudes below 2000m and ambient temperatures ranging from -25℃ to +40℃ shall have electrical clearances of not less than 6mm," dependency analysis identifies that "altitude below 2000m" and "ambient temperature ranging from -25℃ to +40℃" together constitute the operating environment conditions, and these two condition entities jointly constrain the judgment criterion entity of "6mm electrical clearance." It also supports handling conditional branching cases, such as "when the pollution level is 2, the creepage distance shall be not less than 4mm; when the pollution level is 3, the creepage distance shall be not less than 8mm." In this case, it can establish a precise matching relationship between the pollution level conditions and the corresponding creepage distance requirements, forming a complete parameter constraint pair.

[0149] The above steps realize the full-process processing of accurately extracting technical parameter entities from the clauses to be evaluated, accurately classifying parameter types, and establishing complete parameter logical relationships, forming structured parameter constraint knowledge. This provides a reliable data foundation for building a high-quality standard clause parameter knowledge graph, and significantly improves the machine understandability and processability of technical requirements in standard texts.

[0150] Further, in step S121, the named entity recognition model based on sequence labeling architecture performs word sequence labeling processing on the terms to be evaluated, identifying and labeling numerical value entities and their corresponding physical unit entities, including:

[0151] Step S1211: Perform word segmentation and part-of-speech tagging on the terms to be evaluated to generate a word sequence with grammatical tags.

[0152] Lexical segmentation and part-of-speech tagging of the terms to be evaluated are fundamental preprocessing steps for parameter entity recognition. A deep learning-based word segmentation model, trained on standard text corpora in the field of electrical safety, is employed to accurately identify the boundaries of electrical safety terminology. During word segmentation, special attention is paid to common compound technical terms in electrical safety standard texts, such as "insulation resistance," "withstand voltage," and "leakage current," ensuring that these terms with specific technical meanings are fully preserved. The segmented word sequence is then further input into a part-of-speech tagging module, which uses a combination of Hidden Markov Models and rules to assign accurate part-of-speech tags to each word.

[0153] During part-of-speech tagging, basic parts of speech such as nouns, verbs, adjectives, numerals, and quantifiers are identified, while special processing is applied to the characteristics of electrical safety standard texts. For example, composite units like "kV / mm" are identified as complete unit nouns; special symbols such as "≥" and "±" are labeled as comparison operators or tolerance symbols. The lexical sequences with completed part-of-speech tagging constitute the basic data with rich grammatical markers, which provide important feature basis for subsequent sequence tagging. The entire preprocessing process ensures that subsequent entity recognition can be based on accurate linguistic foundations.

[0154] Step S1212: Input the word sequence into the named entity recognition model, and output the label corresponding to each word in the preset entity label set through its sequence labeling architecture to obtain the entity label sequence.

[0155] Inputting a sequence of words into the named entity recognition model is the core step in achieving parametric entity recognition. This model employs a sequence labeling architecture combining a bidirectional long short-term memory network and a conditional random field, enabling it to fully utilize contextual information for entity boundary recognition and type determination. The model's input layer embeds words, their part-of-speech tags, location information, and other features into a high-dimensional vector representation, extracting deep semantic features through a multi-layer neural network. During training, the model learns the expression patterns of parametric entities in a large number of electrical safety standard texts, accurately recognizing various complex electrical safety parameter expressions.

[0156] The entity label sequence output by the model adopts the BIO annotation system, clearly indicating the position of each word within the entity. For example, in the expression "test voltage 2.5kV", "2.5" is labeled as "B-VALUE" (start of numerical value), and "kV" is labeled as "B-UNIT" (start of physical unit). For more complex expressions such as "insulation resistance range 100MΩ to 500MΩ", the model can accurately label all components of the starting value, ending value, and unit. The model also has the ability to handle nested and overlapping entities, ensuring accurate annotation results under various complex conditions.

[0157] Step S1213: Based on the entity label sequence, identify and extract the numerical value entities that represent specific quantities and the physical unit entities that they directly modify in the text.

[0158] Extracting numerical entities and physical unit entities from entity label sequences is a crucial step in entity recognition post-processing. By parsing entity boundary information in the labeled sequence, complete parametric entity fragments are identified. The extraction process not only focuses on the textual content of the entities themselves but also records the position, length, and relationships of each entity within the original text. A rule-based post-processing algorithm is employed to validate and correct the labeled sequence output by the model, ensuring the accuracy and completeness of entity extraction.

[0159] The entity extraction process needs to handle various complex linguistic phenomena. For example, for scattered numerical values ​​and units, such as "withstand voltage 3kV, insulation resistance 100MΩ", it can correctly match the numerical values ​​and corresponding units. For special cases where units are omitted, it infers from the context, such as automatically completing the unit information when "electrical clearance not less than 6" is followed by "unit millimeters". A detection mechanism for common error patterns has also been established, which can identify and correct problems such as miswriting "megohm" as "megohmm" or "kV" as "KV", ensuring that the extracted parameter entities meet the specifications.

[0160] Step S1214: Perform compound expression parsing on the numerical value entity, and parse the continuous numerical range containing the connector into a standardized numerical representation with clear upper and lower limits.

[0161] Parsing compound expressions of numerical entities is a crucial step in handling complex parameter representations. A grammar-rule-based parsing method can identify and process various complex numerical ranges, tolerance expressions, and conditional expressions. For example, for a tolerance expression like "2.5kV ± 5%", the parsing process yields a base value of 2.5kV, a tolerance of ± 5%, and generates a standardized numerical range representation [2.375kV, 2.625kV]. For "operating voltage range 220V~240V", the parsing process yields a lower limit of 220V and an upper limit of 240V.

[0162] During the parsing process, complex expressions with comparison operators are also handled. For example, "">100MΩ" is parsed as the open interval (100MΩ, ∞), and "≤0.5mA" is parsed as the closed interval [0, 0.5mA]. For fractional expressions such as "1 / 2 times the rated voltage", its decimal value of 0.5 is calculated and stored in a standardized manner. Parsing compound unit expressions is also supported; for example, "kV·A" is decomposed into basic units and its dimensional relationship is recorded. All parsing results use a unified standardized representation, facilitating subsequent knowledge graph construction and semantic querying.

[0163] Step S1215: Based on the preset unit mapping rules, the physical unit entities are standardized, and the same type of units with different expression forms are uniformly converted into the reference unit form.

[0164] Standardizing physical unit entities based on preset unit mapping rules is a crucial step in ensuring parameter consistency. Maintaining a complete knowledge base for unit mapping in the electrical safety field is essential, encompassing the complete mapping relationships of the International System of Units (SI), commonly used units, and electrical safety-specific units. This knowledge base not only records conversion relationships between units but also includes dimensional information and usage scenarios. For example, different expressions such as "kilovolt," "kV," and "Kv" are uniformly mapped to the standard unit "kV," while simultaneously recording their voltage dimensional characteristics.

[0165] The standardization process requires handling unit conversions and equivalent transformations. For example, converting "milliampere" to "A" (ampere) and "megaohm" to "Ω" (ohm). For composite units, such as "kilovolt per millimeter," they are broken down into basic units and conversion factors with the International System of Units (SI) are calculated. A unit compatibility check mechanism has also been established to detect unreasonable unit usage, such as directly adding voltage and current units, ensuring the correct physical meaning of parameter entities.

[0166] Step S1216: Semantically combine the standardized numerical value entity with the physical unit entity to generate a candidate parameter entity with complete dimensional information.

[0167] The final step in generating candidate parameter entities is to semantically combine the standardized numerical value entities with the physical unit entities. Based on the syntactic relationships and semantic associations of the parameters in the original text, the numerical values ​​are correctly paired with their corresponding physical units. During the combination process, factors such as the distance between the numerical value and the unit, and the syntactic role of intermediate words are considered to ensure the accuracy of the pairing. For example, in "withstanding pressure strength 3kV", the combination relationship between the numerical value "3" and the unit "kV" is correctly established.

[0168] After pairing, structured candidate parameter entities are generated. Each entity contains complete attributes such as standardized numerical values, standard units, original text, and location information. For complex parameters, such as "test voltage AC500V / DC1000V", multiple candidate entities are generated to represent parameter values ​​under different conditions. Contextual information of the parameter entities is recorded, such as modifiers like "not less than" and conditions like "at rated frequency," which are crucial for subsequent parameter type classification and logical association establishment. The final set of candidate parameter entities provides high-quality input data for subsequent semantic role classification.

[0169] Further, step S130 involves establishing a tree-like semantic association structure based on the clause type, condition entity, and judgment criterion entity, with the clause to be evaluated as the root node and linked to the condition entity and judgment criterion entity respectively, including:

[0170] Step S131: The clause to be evaluated is taken as the root node of the tree structure, and a clause type attribute representing the constraint nature is added to the root node.

[0171] Using the clause to be evaluated as the root node of the tree structure is the foundation for constructing a semantic association structure. The root node carries the complete text content of the clause and clarifies its binding nature by adding a clause type attribute. The clause type attribute uses standardized enumerated values, including basic types such as requirement and prohibition, while also supporting extended subtypes to meet the needs of different standard systems. Each root node generates a unique identifier upon creation, used to establish cross-document relationships in the knowledge graph.

[0172] In the application scenarios of electrical safety standards, for the clause "Electrical clearance shall not be less than 6mm", a root node is created to store the complete text and a "requirement" type attribute is added. For the clause "Maintenance work shall not be performed while the circuit is energized" in the electrical safety standard, the root node, in addition to storing the text content, will also be marked as "prohibited". The source information of the clause is recorded, such as the standard number GB 4706.1, chapter number 29.1, and other metadata. This information, together with the type attribute, constitutes the complete descriptive system of the root node.

[0173] Step S132: Using the conditional entity as the first type of child node, link the conditional entity to the corresponding root node by testing the conditional relationship edge.

[0174] Conditional entities, as the first type of child nodes, are connected to the root node through test conditional relationship edges, forming conditional branches in the semantic network. Test conditional relationship edges are directional, pointing from the root node to the conditional entity, indicating that these conditional entities constrain and limit the technical requirements stated by the root node. Each conditional entity node contains complete parameter information, including structured data such as parameter name, numerical range, and unit of measurement. By analyzing the syntactic role of the conditional entity in the original text, corresponding semantic tags are added to the relationship edges, such as environmental conditions, electrical conditions, and installation conditions.

[0175] In the application scenario of electrical safety standards, for the clause "After damp heat test (temperature 40±2℃, relative humidity 93±3%), the insulation resistance shall not be less than 100MΩ", two condition entity sub-nodes are created to represent the temperature and humidity conditions respectively, and connected to the root node through test condition relationship edges. In the electrical performance testing scenario, for the clause "Perform dielectric strength test under rated voltage of 220V", a rated voltage condition entity node is created, and its semantic role as an electrical condition is marked in the relationship edge. This structured representation allows each element of the test condition to be accurately understood and processed by the machine.

[0176] Step S133: Using the decision criterion entity as the second type of child node, link the decision criterion entity to the corresponding root node through the decision basis relation edge.

[0177] The judgment criterion entity, as a second type of child node, is connected to the root node through judgment basis relation edges, forming a judgment branch in the semantic network. The judgment basis relation edges carry the logical semantics of compliance judgment, explicitly indicating that the parameter values ​​in the child nodes are the direct basis for judging whether the technical requirements of the root node are met. Each judgment criterion entity node records detailed information about the threshold parameters, including comparison operators, baseline values, allowable deviations, and other key data. By analyzing the syntactic relationship between the judgment criteria and the core predicate, different logical semantics are assigned to the relation edges, such as "must meet," "not allowed to exceed," and "should meet requirements."

[0178] In the specific implementation of electrical safety standards, for the clause "insulation resistance shall not be less than 100MΩ", a judgment criterion entity node is created to store the threshold "100MΩ", and connected to the root node through a judgment basis relationship edge, recording the comparison relationship "not less than". In the field of electrical safety, for the requirement "leakage current shall not exceed 0.5mA", a corresponding judgment criterion entity node is established, and the constraint property of "shall not exceed" is clearly marked in the relationship edge. This refined representation method ensures accurate modeling of electrical safety compliance requirements.

[0179] Step S134: Based on the logical dependency relationship between the condition entity and the judgment criterion entity in the original text, a constraint logic edge is established between the first type of child node and the second type of child node, thereby forming a tree structure with multi-layer semantic association.

[0180] Establishing constraint logic edges between condition entities and decision criterion entities is a crucial step in constructing complete semantic relationships. These constraint logic edges are created based on the logical dependencies in the original text, accurately reflecting the specific range of conditions to which a particular decision criterion applies. Deep semantic analysis identifies implicit relationships between condition entities and decision criterion entities, particularly handling complex cases where multiple conditions correspond to multiple criteria. Constraint logic edges can carry rich logical operator information, such as AND, OR, and NOT, supporting the construction of complex combinations of condition and criterion relationships.

[0181] In the practical application of electrical safety standards, for the clause "the electrical clearance should be no less than 4 mm under pollution level 2 conditions and no less than 8 mm under pollution level 3 conditions," two sets of constraint logic edges are established to associate the pollution level 2 condition with the 4 mm criterion and the pollution level 3 condition with the 8 mm criterion, respectively. In electrical equipment safety standards, for the clause "the leakage current should not exceed 0.5 mA at AC 50 Hz and 0.25 mA at DC," corresponding constraint logic edges are created to maintain the correct correspondence between the power supply type conditions and the corresponding leakage current requirements. This meticulous correlation modeling ensures the completeness and accuracy of electrical safety technical requirements.

[0182] The above steps realize the full-process processing of accurately extracting technical parameter entities from the clauses to be evaluated in electrical safety standards, accurately classifying parameter types, and establishing complete logical relationships between parameters, forming structured electrical safety parameter constraint knowledge. This provides a reliable data foundation for building a high-quality knowledge graph of electrical safety standard clause parameters, and significantly improves the machine understandability and processability of technical requirements in electrical safety standard texts.

[0183] Specifically, in step S200, a knowledge graph of standard clause parameters is constructed based on the semantic association structure. The knowledge graph supports semantic queries of technical parameters, including:

[0184] Step S210: Map the clause node to be evaluated, condition entity node, and judgment criterion entity node in the clause-level semantic association structure to entity nodes in the knowledge graph.

[0185] Mapping nodes in the clause-level semantic association structure to knowledge graph entity nodes is a core step in achieving structured knowledge storage. This process requires establishing a complete node type identification mechanism. By analyzing the attribute characteristics and topological position of each node in the semantic association structure, its corresponding entity type in the knowledge graph is determined. For clause nodes to be evaluated, core attributes such as text content, clause type, and standard source are extracted to generate standard clause entities in the knowledge graph. For condition entity nodes, technical attributes such as parameter names, numerical ranges, and units of measurement are parsed to construct test condition entities. Judgment criterion entity nodes are formed by parsing elements such as threshold type, comparison operators, and benchmark values ​​to create qualified judgment criterion entities.

[0186] During the mapping process, a unified entity identifier generation strategy is adopted to ensure that each entity node has a globally unique identifier. The generation of entity identifiers comprehensively considers entity content features and contextual information to avoid conflicts between different entities. Simultaneously, an entity disambiguation mechanism is established to merge entities with the same semantics but different expressions, such as identifying "working voltage" and "rated voltage" as the same type of entity. This refined entity mapping strategy guarantees the consistency and accuracy of entities in the knowledge graph.

[0187] In the implementation case of the electrical safety standard, for the clause "the electrical clearance shall not be less than 4 mm under pollution level 2 conditions", three entity nodes are created: the standard clause entity stores the complete text content and is marked as a requirement; the condition entity records the environmental condition "pollution level 2"; and the judgment criterion entity stores the size requirement "4 mm". The three entities are connected through specific semantic relationships to form a complete knowledge representation.

[0188] Step S220: Map the test condition relation edges, decision basis relation edges, and constraint logic edges in the clause-level semantic association structure to relation edges in the knowledge graph.

[0189] Mapping the relation edges in the semantic association structure to knowledge graph relation edges is a crucial step in constructing a knowledge network. By analyzing the semantic features and connection patterns of various relation edges in the semantic association structure, a precise mapping to the knowledge graph relation model is established. Test condition relation edges are mapped to the "has_test_condition" relation, carrying the constraint relationship between the standard clause and its test environment conditions. Judgment criterion relation edges are mapped to the "has_criteria" relation, expressing the verification relationship between the standard clause and the pass / fail criterion. Constraint logic edges are mapped to the "constrains" relation, reflecting the logical dependency between the condition entity and the judgment criterion entity.

[0190] During the relation mapping process, rich attribute information is added to each relation edge. For test condition relations, metadata such as the applicable scenarios (e.g., humid and hot environments, polluted environments) and the degree of enforceability are recorded. For judgment basis relations, information such as the verification method of the criterion (e.g., dielectric strength test, insulation resistance test) and confidence level is labeled. Constraint logic relations carry details such as logical operators and condition combination methods. These relation attributes enhance the semantic expressive power of the knowledge graph, supporting more precise queries and reasoning.

[0191] A mechanism for calculating relation weights was implemented, assigning different weight values ​​to relations in the knowledge graph based on their explicitness and syntactic tightness in the original text. For example, in electrical safety standards, explicitly stated test condition relations (such as "after damp heat testing") receive higher weights than implicit condition relations. This weighting mechanism provides an important basis for subsequent similarity calculations and relevance ranking.

[0192] In the specific application of electrical safety standards, for the clause "the product should operate normally within a temperature range of -25℃ to +40℃", a "has_test_condition" relationship is established between the standard clause entity and the temperature condition entity, and the temperature range parameter is marked in the relationship. Simultaneously, a "has_criteria" relationship is established between the standard clause entity and the functional requirement entity to record the criterion of "operating normally". When multiple test conditions exist, the logical combination method between the conditions is recorded through relationship attributes.

[0193] Step S230: Add type attributes representing test environment features to condition entity nodes in the knowledge graph, and add type attributes representing threshold features to judgment criterion entity nodes.

[0194] Adding type attributes to knowledge graph entities is a crucial means of achieving refined knowledge management. Based on an ontology library in the electrical safety domain and industry classification standards, a complete classification system for test environment features is established for condition entity nodes. This system covers major categories such as electrical environment, mechanical environment, and climatic environment, with each major category further subdivided into specific environmental parameter types. For example, the electrical environment includes subcategories such as voltage, current, and frequency; the mechanical environment includes subcategories such as vibration and shock; and the climatic environment includes subcategories such as temperature and humidity. Each condition entity is assigned a corresponding type label based on its parameter characteristics.

[0195] For entities subject to the judgment criteria, a classification system based on threshold features was constructed. This system categorizes thresholds into interval thresholds, limit thresholds, and enumeration thresholds according to their mathematical properties. Interval thresholds are suitable for range requirements, such as "insulation resistance 100-500MΩ"; limit thresholds are suitable for extreme value limitations, such as "leakage current not exceeding 0.5mA"; and enumeration thresholds are suitable for discrete values, such as "protection level should be IP54". Each threshold type has a corresponding attribute template to guide the extraction and storage of relevant attributes.

[0196] An inheritance and reasoning mechanism for type attributes has been established. Once the entity type is determined, typical attribute fields corresponding to that type are automatically added to it. For example, for a voltage-type conditional entity, extended attribute fields such as "voltage type (AC / DC)" and "rated value" are automatically added. For a limit-type threshold entity, attributes such as "limit direction" and "allowable deviation" are automatically added. This mechanism ensures the consistency of attributes for entities of the same type.

[0197] In the practical application of electrical safety standards, the condition entity "dielectric strength of insulating material under rated voltage" is identified as an electrical environment type, and attributes such as voltage type and frequency are automatically added. For the judgment criterion entity "grounding resistance not exceeding 0.1Ω", it is identified as a limit threshold, and the limit direction is recorded as "not exceeding", while test condition attributes such as test current and measurement method are added.

[0198] Step S240: Based on the logical dependency relationship established in the semantic association structure, construct multi-hop semantic paths between entity nodes in the knowledge graph.

[0199] Constructing multi-hop semantic paths between entity nodes is fundamental to deep knowledge retrieval. Based on the logical dependencies established in the semantic association structure, direct and indirect relationships between entities are analyzed to generate semantically meaningful connection paths. The construction of multi-hop semantic paths employs a graph traversal-based algorithm, starting from the initial entity, exploring reachable entity nodes along relational edges, recording the traversed paths, and evaluating the semantic rationality of the paths.

[0200] Several types of semantic path patterns are defined, including conditional reasoning paths, safety requirement tracing paths, and standard association paths. Conditional reasoning paths connect related test condition entities, supporting reasoning queries based on combined conditions. Safety requirement tracing paths connect safety requirement entities with their dependent condition entities, supporting complete tracing of safety requirements. Standard association paths connect related entities in different electrical safety standard documents, supporting cross-standard knowledge discovery. Each path type has corresponding weight calculation rules, comprehensively evaluating the importance of the path based on factors such as path length, relationship strength, and entity relevance.

[0201] During the path construction process, a path optimization mechanism is implemented to improve query efficiency by pre-compiling high-frequency paths and establishing path indexes. For complex multi-hop paths, a strategy of segmented storage and dynamic combination is adopted to balance storage overhead and query performance. Simultaneously, a path verification mechanism is established to ensure the correctness and completeness of the generated semantic paths in terms of electrical safety technical logic.

[0202] In the application scenarios of electrical safety standards, multi-hop semantic paths can be constructed from the entity of "insulating material" to the entity of "withstand voltage test conditions" and then to the entity of "test voltage threshold". When a user queries the test requirements for a certain insulating material, the complete test specification can be returned through this path. In the electrical safety management system, a semantic path can be established from "electrical hazard type" to "protective measures" and then to "safety requirements", supporting comprehensive analysis of safety requirements.

[0203] Step S250: Establish a conversion interface from natural language query to knowledge graph query statement, and map the technical parameter requirements in the natural language query to entity nodes and relation edges in the knowledge graph through semantic similarity calculation.

[0204] Establishing an interface for converting natural language queries to knowledge graph queries is a key technology for achieving intelligent retrieval. This interface, based on a deep semantic understanding model, parses user-input natural language questions into structured query intents. The parsing process includes entity recognition, relation extraction, and condition analysis. Using a named entity recognition model adaptively trained in the electrical safety domain, it accurately identifies key entities in the query, such as electrical safety parameters, standard clauses, and test conditions.

[0205] In terms of semantic similarity calculation, a multi-level similarity evaluation strategy is adopted. For entity-level similarity, word vector similarity and electrical safety ontology concept distance are combined to calculate the semantic association between the query entity and the knowledge graph entity. For relation-level similarity, the degree of matching between the relational semantics expressed in the query and the relational patterns in the knowledge graph is analyzed. Simultaneously, the logical structure and complexity of the query conditions are also considered to ensure accurate mapping of complex queries.

[0206] The query conversion interface supports various types of natural language queries, including direct queries, conditional queries, and comparison queries. Direct queries include "find insulation resistance requirements," conditional queries include "dielectric strength requirements in humid and hot environments," and comparison queries include "compare the withstand voltage ratings of different insulation materials." For each query type, corresponding query templates and mapping rules have been established to ensure the generation of standardized knowledge graph query statements.

[0207] In its implementation, the system also provides query clarification and expansion mechanisms. When a user's query is ambiguous or lacks sufficient information, supplementary information can be obtained interactively. For example, when a user queries "pressure withstand requirements," they can be asked about the specific equipment type or application scenario. The system also supports progressive refinement of query results, allowing users to add new conditions to the initial results for secondary queries.

[0208] Step S260: Perform a semantic query in the knowledge graph based on the mapping result, and return the standard terms that meet the query conditions and their associated parameter entity information.

[0209] Performing semantic queries within the knowledge graph based on the mapping results is the final implementation step of knowledge services. The system receives a transformed structured query statement and uses the graph database's query engine to search for matching entities and relationships within the knowledge graph. Query execution employs an index-based optimization algorithm to quickly locate relevant entity nodes, and then performs graph traversal along semantic relationship edges to collect a set of entities that satisfy the query conditions.

[0210] The query process employs a multi-stage result filtering and ranking mechanism. First, precise matching is performed based on the query conditions to identify entities that fully meet the requirements. Then, similarity expansion is performed, using semantic similarity calculations to find related entities. Finally, the results are comprehensively ranked based on factors such as entity importance, relationship strength, and timeliness, ensuring that the most relevant results are presented first. For complex multi-condition queries, faceted browsing of results is also supported, allowing users to filter and explore query results from different dimensions.

[0211] It offers a variety of result display formats, including structured lists, knowledge graph visualizations, and association path displays. Structured lists clearly present the key attributes and source information of each matching entity. Knowledge graph visualizations display the network of relationships between entities in a graph format, supporting interactive exploration. Association path displays highlight the semantic connections between query conditions and results, helping users understand the context of technical requirements.

[0212] In an application example of electrical safety standards, when a user queries "insulation resistance test requirements," a list of relevant standard clause entities is returned. Each entity includes detailed information such as specific resistance value requirements, test method conditions, and pass / fail criteria. Users can view the specific technical requirements corresponding to different insulation classes through an interactive interface, compare the applicable conditions of different test methods, trace the evolution history of relevant standards, and obtain comprehensive technical guidance.

[0213] Specifically, in step S300, based on the standard clause parameter knowledge graph, the clauses to be evaluated are weighted and screened to obtain several key clauses, including:

[0214] Step S310: Based on the association structure features of the nodes to be evaluated in the knowledge graph, calculate the initial weight value of each node to be evaluated. The association structure features include the number of connected parameter entities and the distribution of relation edge types.

[0215] Calculating initial weights based on the relational structure features of the nodes to be evaluated in the knowledge graph is a fundamental step in assessing the importance of a clause. This calculation process comprehensively considers the topological characteristics of nodes in the network, including the number of connected parameter entities, the type distribution of relation edges, and the centrality index of the nodes. A weight calculation model based on a graph neural network is adopted. This model learns the representation vector of each node to be evaluated by aggregating the feature information of neighboring nodes, and then calculates the initial weights based on the dimensional features of the representation vectors. A clause node with a larger number of connected parameter entities generally indicates that the technical elements involved in the clause are more complex, and its structural importance in the knowledge graph is correspondingly higher.

[0216] The distribution of relation edge types is another important consideration. Analyzing the semantic type of each relation edge, such as test condition relations, decision basis relations, and constraint logic relations, different types of relation edges are assigned different weight coefficients. For example, clause nodes with multiple test condition relations and decision basis relations often represent clauses containing a complete technical requirement system, and their initial weight values ​​are correspondingly higher. Considering the directionality and multiplicity of relations, a weight propagation model based on the relation network is constructed to ensure the accuracy and rationality of weight calculation.

[0217] In the practical application of electrical safety standards, the clauses related to "electrical clearances and creepage distances" are found to connect multiple parameter entities such as pollution level, material group, and operating voltage, and also have relationships between test conditions and judgment criteria, thus receiving a higher initial weight. In contrast, the "equipment identification" clause in the general requirements for electrical equipment typically only connects to simple judgment criteria entities, with a single relationship type, resulting in a relatively lower initial weight. This weight calculation method based on the characteristics of the associated structure can objectively reflect the structural importance of clauses within the electrical safety standard system.

[0218] Step S320: Perform a first correction on the initial weight value based on the clause type attribute of the clause to be evaluated to obtain the type-corrected weight.

[0219] The first adjustment process, based on the clause type attributes of the clauses to be evaluated, is an important adjustment mechanism based on the strength of semantic constraints. A clause type weighting coefficient table is established, assigning corresponding adjustment coefficients according to the strength of the technical constraints of different clause types. Requirement clauses typically receive higher positive coefficients because they explicitly stipulate the technical requirements that must be met; prohibition clauses are assigned different coefficient levels based on the importance of their prohibited content. For advisory and optional clauses, their weighting coefficients are appropriately reduced to accurately reflect their actual binding force within the standard system.

[0220] The revision process employs a tiered adjustment strategy. Initial revisions are made based on the fundamental coefficients of the clause type, followed by fine-tuning considering the context of the clause. For example, prohibitory clauses in the Basic Requirements for Electrical Safety section have a higher weighting coefficient than similar clauses in the appendix. Rules for handling combinations of clause types are also established. For composite clauses containing both requirement and prohibition statements, a weighted average is used to calculate the comprehensive revision coefficient. This refined revision mechanism ensures that the clause type attributes are reasonably reflected in the weighting calculation.

[0221] In the implementation of electrical safety standards, a higher positive coefficient is applied to requirement clauses such as "protective grounding shall be reliably connected." For prohibition clauses such as "aluminum wire shall not be used as grounding conductors," a higher correction coefficient is assigned based on its position in the safety grounding section. In equipment installation standards, recommendation clauses such as "cable tray wiring is recommended" are subject to a relatively lower correction coefficient. This type of correction approach accurately captures the differences in the importance of different clauses in the implementation of electrical safety standards.

[0222] Step S330: Calculate the frequency of occurrence of various parameter entities in the knowledge graph, and perform a second enhancement process on the type correction weights based on the frequency analysis results to obtain the frequency enhancement weights.

[0223] The frequency of occurrence of various parameter entities in a statistical knowledge graph is a significant enhancement mechanism based on industry attention. By analyzing the distribution of parameter entities throughout the knowledge graph, the frequency and density of each parameter category are calculated. High-frequency parameter entities typically represent the broad applicability and importance of the technical element within the standards system, and their associated clauses receive corresponding weight enhancement. The frequency analysis employs a dynamic statistical method based on time windows, considering both the cumulative frequency of parameter occurrence and its trend in the latest standards, ensuring the timeliness and accuracy of weight enhancement.

[0224] The enhancement process employs a frequency-segmented nonlinear enhancement strategy, dividing parameter entities into multiple levels based on their frequency of occurrence, with each level corresponding to a different enhancement coefficient. For parameter categories with extremely high frequencies, the enhancement amplitude is appropriately controlled to avoid excessive weight concentration. Simultaneously, the uniformity of parameter entity distribution within the standard system is considered; high-frequency parameters appearing in multiple professional fields are given additional enhancement weights. This frequency-analysis-based enhancement mechanism can effectively identify technical requirements with strong industry consensus.

[0225] In the application scenarios of electrical safety standards, statistics show that parameters such as "insulation resistance," "withstand voltage," and "leakage current" appear frequently in multiple standards. Therefore, clauses related to these parameters are given significantly increased weight. In low-voltage electrical appliance standards, basic performance parameters such as "breaking capacity" and "making capacity" appear frequently, and related clauses are also strengthened accordingly. This frequency-driven enhancement approach allows for an accurate grasp of the hot topics and key directions of technological development in the field of electrical safety.

[0226] Step S340: Based on the preset list of important parameter entities, the frequency enhancement weights are adjusted in a third way to obtain the final weight value of each clause to be evaluated.

[0227] The third adjustment process, based on a pre-defined list of key parameter entities, is a crucial customized step tailored to specific business needs. This list is compiled by domain experts based on industry development plans, technological breakthroughs, and key product quality control points, and includes core technical parameters. By establishing a mapping relationship between the list's parameters and parameter entities in a knowledge graph, the terms to be evaluated related to key parameters are identified, and differentiated weight adjustments are made based on the parameter's importance level.

[0228] The adjustment process employs a coefficient mapping mechanism based on importance hierarchy, categorizing important parameter entities into multiple levels such as critical, important, and general, with each level corresponding to a different adjustment coefficient. For clauses related to multiple important parameters, the overall adjustment effect is calculated by superimposing the coefficients. Simultaneously, a constraint mechanism for the adjustment magnitude is established to ensure that the weight adjustments are within a reasonable range and to avoid extreme values. This adjustment process based on expert knowledge effectively compensates for the limitations of purely data-driven methods.

[0229] In the implementation of electrical safety standards, the list of important parameters includes key parameters related to electrical safety such as "insulation coordination," "protection against electric shock," and "fire protection," with clauses related to these parameters receiving significantly increased weight. In the field of explosion-proof electrical standards, core parameters related to explosion-proof safety, such as "maximum surface temperature" and "protection level," are included in the important list, and the weight of related clauses is significantly strengthened. This targeted adjustment based on the list of important parameters ensures that critical technical requirements for electrical safety are given sufficient consideration during the selection process.

[0230] Step S350: Sort the clauses to be evaluated based on the final weight values, and select the clauses to be evaluated with weight values ​​higher than the preset threshold as key clauses.

[0231] The final decision-making step in determining key clauses involves ranking and screening the clauses to be evaluated based on their final weight values. A multi-dimensional ranking strategy is employed. First, clauses are ranked in descending order according to their final weight values. Then, auxiliary indicators such as the timeliness and applicability of the clauses are considered for fine-tuning. For clauses with the same weight value, secondary indicators such as the complexity of their related structures and the importance of their parameter entities are further compared to ensure the rationality and stability of the ranking results. The ranking process uses a segmented ranking algorithm based on a sliding window, ensuring both ranking efficiency and accuracy.

[0232] The threshold setting employs a dynamic adjustment mechanism, determining an appropriate threshold level based on the specific needs of each application scenario. For standard analysis in the electrical equipment design phase, a relatively lenient threshold is used to retain more key clauses to support comprehensive design; for standard applications in the electrical safety certification phase, a stricter threshold is adopted, focusing on the most critical safety technical requirements. A percentage-based threshold setting method is also provided, determining the selection range of key clauses according to a weighted percentage distribution. This flexible threshold setting mechanism ensures that the screening results can adapt to different business needs.

[0233] In the application of electrical equipment safety standards, the top 15% of clauses with the highest final weight values ​​are selected as key clauses, covering core safety aspects such as insulation requirements, protection levels, and grounding protection. In explosion-proof electrical acceptance standards, an absolute threshold method is used, selecting clauses with weight values ​​exceeding 0.85 as key acceptance criteria.

[0234] Accordingly, please refer to Figure 2 A second aspect of this invention provides a standard text semantic analysis and key clause extraction system, which performs text semantic analysis and key clause extraction based on the above-mentioned standard text semantic analysis and key clause extraction method, including:

[0235] Text parsing module 1 is used to perform multi-level semantic parsing on standard text, identify the clauses to be evaluated and their corresponding types in the standard text, extract the key parameter entities of the clauses to be evaluated, and construct a clause-level semantic association structure based on the clause type and key parameter entities.

[0236] Graph construction module 2 is used to build a standard clause parameter knowledge graph based on semantic association structure. The knowledge graph supports semantic query of technical parameters.

[0237] Clause screening module 3 is used to perform weight calculation and screening of clauses to be evaluated based on the standard clause parameter knowledge graph, and obtain several key clauses.

[0238] The list generation module 4 is used to generate a structured list of standard requirements summaries based on several key clauses selected and the parameter information associated with them in the corresponding knowledge graph.

[0239] The embodiments of this invention aim to protect a standard text semantic analysis and key clause extraction method and system, which has the following effects:

[0240] 1. By introducing a multi-level semantic parsing mechanism that integrates clause type identification and parameter entity extraction, and constructing a tree-like association structure of "clause-condition-criteria", the machine-readable parsing of deep semantics of standard text was realized for the first time. It effectively solved the core problem that traditional single keyword matching or general NLP models could not accurately distinguish the mandatory force of clauses and could not associate fragmented technical parameters, laying a high-quality structured data foundation for subsequent intelligent applications.

[0241] 2. By further constructing the semantic association structure into a knowledge graph rich in parameter attributes and establishing a conversion interface from natural language to graph query, users can perform complex condition searches in an intuitive semantic way; this fundamentally solves the technical bottleneck of traditional document retrieval or database query methods being unable to understand user intent or perform parameter-based logical association reasoning, thus resulting in insufficient depth and accuracy of information retrieval.

[0242] 3. By calculating comprehensive weights based on the knowledge graph topology, parameter entity frequency, and importance, and generating a structured summary list accordingly, a leap from "information extraction" to "knowledge refinement" is achieved. This completely solves the long-standing problems of strong subjectivity and low efficiency in manual interpretation of standards, as well as the fact that existing automated methods can only list text fragments and cannot provide prioritization, resulting in poor decision support.

[0243] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A standard text semantic analysis and key clause extraction method, characterized in that, Includes the following steps: Multi-level semantic parsing is performed on the standard text to identify the clauses to be evaluated and their corresponding types in the standard text, and the key parameter entities of the clauses to be evaluated are extracted. A clause-level semantic association structure is constructed based on the clause type and the key parameter entities. Based on the semantic association structure, a standard clause parameter knowledge graph is constructed, which supports semantic querying of technical parameters; Based on the knowledge graph of the standard clause parameters, the clauses to be evaluated are weighted and screened to obtain several key clauses. Based on the selected key clauses and the corresponding parameter information associated in the knowledge graph, a structured list of standard requirements summaries is generated. The process involves performing multi-level semantic parsing on the standard text, identifying the clauses to be evaluated and their corresponding types within the standard text, extracting key parameter entities from the clauses to be evaluated, and constructing a clause-level semantic association structure based on the clause type and the key parameter entities. This includes: By using a pre-defined clause classification model, the standard text is classified at the clause level to identify the clauses to be evaluated and determine their respective clause types, which include requirement clauses and prohibition clauses. Using the named entity recognition model, the identified clauses to be evaluated are processed by parameter entity extraction, and condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds are extracted. The condition entities and the judgment criterion entities together constitute the key parameter entities. Based on the clause type, the condition entity, and the judgment criterion entity, a tree-shaped semantic association structure is established with the clause to be evaluated as the root node and linked to the condition entity and the judgment criterion entity respectively, thus obtaining the clause-level semantic association structure; The process of extracting parameter entities from the identified terms to be evaluated using a named entity recognition model extracts condition entities representing technical indicator constraints and judgment criterion entities representing qualification thresholds, including: The term to be evaluated is processed by word sequence labeling using a named entity recognition model based on sequence labeling architecture. Numerical value entities and their corresponding physical unit entities are identified and labeled. The numerical value entities and physical unit entities together constitute candidate parameter entities. Based on the preset entity context dependency rules, the candidate parameter entities are classified into semantic roles. Candidate parameter entities that represent the test environment or operation state are classified as condition entities, and candidate parameter entities that represent the qualification judgment boundary are classified as judgment criterion entities. The logical relationship between the condition entity and the decision criterion entity is established through dependency parsing. When both the condition entity and the decision criterion entity exist in the same clause to be evaluated, the correspondence between the condition entity and the decision criterion entity is constructed based on the syntactic parsing results, forming a complete parameter constraint pair.

2. The standard text semantic analysis and key clause extraction method according to claim 1, characterized in that, The step of performing clause-level classification processing on the standard text using a preset clause classification model to identify the clauses to be evaluated and determine their clause types includes: The standard text is segmented at the sentence level, and the segmentation results are corrected based on semantic integrity rules to obtain independent text units to be classified. The independent text units are input into the preset clause classification model for semantic feature extraction and classification calculation. The preset clause classification model outputs a classification probability distribution that represents the clause type. Based on the classification probability distribution, text units containing preset constraint semantic patterns are identified as the clauses to be evaluated, and the corresponding clause types are determined. The identified clauses to be evaluated are type-verified by a preset pattern matching rule base. When the clause type is inconsistent with the binding semantic pattern, the clause type is corrected according to the output of the pattern matching rule base.

3. The standard text semantic analysis and key clause extraction method according to claim 2, characterized in that, The correction of the segmentation results based on semantic integrity rules includes: Detect whether there are referential words in each text unit obtained from the initial segmentation. When the text unit contains a pronoun pointing to a technical entity, mark the text unit as the current text unit with referential dependency. In the text sequence preceding the current text unit, find the text unit containing a description of the specific technical object referred to by the pronoun, and determine it as the antecedent text unit; Based on the syntactic dependency relationship between the current text unit and the antecedent text unit, they are merged into a complete semantic unit; Component analysis is performed on the semantic units to verify whether they constitute a declarative sentence structure that fully expresses the technical requirements; The semantic units that pass structural verification are treated as independent text units that meet the requirements of semantic integrity.

4. The standard text semantic analysis and key clause extraction method according to claim 1, characterized in that, The step involves using a named entity recognition model based on sequence labeling architecture to perform word sequence labeling on the terms to be evaluated, identifying and labeling numerical value entities and their corresponding physical unit entities, including: The terms to be evaluated are segmented into words and tagged with parts of speech to generate a word sequence with grammatical tags; The word sequence is input into the named entity recognition model, and the model outputs the label corresponding to each word in the preset entity label set through its sequence labeling architecture to obtain the entity label sequence. Based on the entity tag sequence, identify and extract numerical value entities that represent specific quantities and their directly modified physical unit entities from the text; The numerical value entity is parsed using compound expression parsing, and the continuous numerical range containing connectors is parsed into a standardized numerical representation with clear upper and lower limits; Based on preset unit mapping rules, physical unit entities are standardized, and similar units with different expressions are uniformly converted into the standard unit form. The standardized numerical value entity and the physical unit entity are semantically combined to generate a candidate parameter entity with complete dimensional information.

5. The standard text semantic analysis and key clause extraction method according to claim 1, characterized in that, The step of establishing a tree-like semantic association structure based on the clause type, the condition entity, and the judgment criterion entity, with the clause to be evaluated as the root node and linked to the condition entity and the judgment criterion entity respectively, includes: The clause to be evaluated is taken as the root node of the tree structure, and a clause type attribute representing the constraint nature is added to the root node; Using the conditional entity as the first type of child node, the conditional entity is linked to the corresponding root node by testing the conditional relationship edge; Using the judgment criterion entity as the second type of child node, the judgment criterion entity is linked to the corresponding root node through the judgment basis relationship edge; Based on the logical dependency relationship between the condition entity and the judgment criterion entity in the original text, constraint logic edges are established between the first type of child nodes and the second type of child nodes, thereby forming a tree structure with multi-layer semantic association.

6. The standard text semantic analysis and key clause extraction method according to any one of claims 1-5, characterized in that, Based on the semantic association structure, a standard clause parameter knowledge graph is constructed. This knowledge graph supports semantic queries of technical parameters, including: Map the clause nodes to be evaluated, condition entity nodes, and judgment criterion entity nodes in the clause-level semantic association structure to entity nodes in the knowledge graph. Map the test condition relation edges, judgment basis relation edges, and constraint logic edges in the aforementioned clause-level semantic association structure to relation edges in the knowledge graph; Add type attributes representing test environment features to condition entity nodes in the knowledge graph, and add type attributes representing threshold features to judgment criterion entity nodes; Based on the logical dependency relationships established in the semantic association structure, multi-hop semantic paths between entity nodes are constructed in the knowledge graph; Establish an interface for converting natural language queries to knowledge graph queries, and map the technical parameter requirements in natural language queries to entity nodes and relation edges in the knowledge graph through semantic similarity calculation; Based on the mapping results, semantic queries are performed in the knowledge graph, returning standard terms that meet the query conditions and their associated parameter entity information.

7. The standard text semantic analysis and key clause extraction method according to any one of claims 1-5, characterized in that, Based on the knowledge graph of the standard clause parameters, the clauses to be evaluated are weighted and screened to obtain several key clauses, including: Based on the association structure features of the nodes of the terms to be evaluated in the knowledge graph, the initial weight value of each term to be evaluated is calculated. The association structure features include the number of connected parameter entities and the distribution of relation edge types. The initial weight value is first corrected by combining the clause type attribute of the clause to be evaluated, and the type-corrected weight is obtained. The frequency of occurrence of various parameter entities in the knowledge graph is statistically analyzed, and the type correction weights are subjected to a second enhancement process based on the frequency analysis results to obtain frequency enhancement weights. Based on a pre-defined list of important parameter entities, the frequency enhancement weights are subjected to a third adjustment process to obtain the final weight value for each of the terms to be evaluated. Based on the final weight values, the clauses to be evaluated are sorted, and the clauses with weight values ​​higher than a preset threshold are selected as key clauses.

8. A standard text semantic analysis and key clause extraction system, characterized in that, Text semantic analysis and key clause extraction are performed based on the standard text semantic analysis and key clause extraction method as described in any one of claims 1-7, including: The text parsing module is used to perform multi-level semantic parsing on standard text, identify the clauses to be evaluated and their corresponding types in the standard text, extract the key parameter entities of the clauses to be evaluated, and construct a clause-level semantic association structure based on the clause type and the key parameter entities. The graph construction module is used to construct a standard clause parameter knowledge graph based on the semantic association structure, and the knowledge graph supports semantic querying of technical parameters; The clause filtering module is used to perform weight calculation and filtering on the clauses to be evaluated based on the standard clause parameter knowledge graph, and obtain several key clauses. The list generation module is used to generate a structured list of standard requirements summaries based on several selected key clauses and the corresponding parameter information associated with the knowledge graph.

Images