Major obligation extraction and retrieval weighting method for regulation clauses

Abstract

The invention provides a subject obligation extraction and retrieval weighting method for regulation and clauses, which comprises the following steps of: obtaining a regulation and clause text, and preprocessing the regulation and clause text to obtain preprocessed data; performing modal discrimination according to the preprocessing; carrying out ontology alignment and pointer network model positioning processing according to the preprocessed data and the modal category; identifying association conditions and exception fragments in the preprocessed data, and regularizing the association conditions and exception fragments into atomic predicates to construct a logic expression; and constructing a tetrad and carrying out retrieval sorting processing. Compared with the prior art, the subject obligation extraction and retrieval weighting method for the regulation clauses has the following advantages that modal information of obligation, prohibition and permission can be accurately recognized in a regulation clause text, and a complete semantic structure is formed by combining subject and behavior extraction; and through logic modeling of association conditions and exception fragments, the application and exclusion range of terms is defined, so that the result has operability.

Description

Technical Field

[0001] This invention applies to the field of law, and particularly relates to a weighted method for extracting and retrieving the main obligations of regulations and clauses. Background Technology

[0002] With the development of natural language processing and information retrieval technologies, the industry is gradually attempting to apply these methods to the parsing of regulations and industry standard texts. Existing methods generally revolve around keyword retrieval, rule templates, or general machine learning models, aiming to identify semantic information related to obligations, prohibitions, or permits in large-scale regulatory documents. However, these technologies have significant limitations when dealing with high-risk, highly constrained fields such as the power industry.

[0003] Existing methods typically rely on fixed modal trigger words for discrimination, such as "must," "should," "must not," and "may." While this literal matching approach is simple to implement, it is easily limited by the diversity of expressions and the complexity of context. When regulations and clauses use variant or implicit expressions, simply relying on trigger words often fails to identify them correctly. For example, phrases like "should be given priority" and "strictly control the frequency of operations," although semantically carrying a mandatory or restrictive meaning, may be missed because they are not in the trigger word list. Furthermore, when negative elements, conditional constraints, or exceptional cases exist, simple keyword detection is more prone to misjudgment, leading to system outputs that contradict the original requirements.

[0004] The neglect of syntactic aspects is also a major limitation of existing methods. Most keyword- and template-driven systems lack the ability to model syntactic scope, making it impossible to distinguish the conditions under which constraint clauses take effect, or to determine the logical relationship between exception clauses and main clauses. For example, in a sentence like "Operators shall take immediate measures when the voltage exceeds U0, unless a backup protection device has been activated," both conditions and exceptions are included. If only the trigger word "shall" is captured without analyzing its corresponding conditions and exceptions, the extraction results will lack accuracy and completeness, leading to biases in subsequent applications.

[0005] Another prominent issue is the lack of clearly defined responsible parties. Regulations typically require that responsibility be explicitly assigned to a specific position, organization, or piece of equipment. However, existing automated extraction methods often only focus on modal words and action verbs, failing to systematically model the responsible parties. Consequently, the generated results lack clear attribution of responsibility, making them ineffective in specific search or question-answering scenarios. This is particularly problematic in the power industry, where the division of responsibilities among entities is very strict; without accurately identifying the responsible party, the extracted clauses will lose their practical value.

[0006] In the field of information retrieval, commonly used Retrieval Augmentation (RAG) frameworks or re-ranking models generally treat candidate clauses in a relatively uniform manner. In other words, the system does not explicitly differentiate between clauses, whether they are mandatory or merely advisory. This approach weakens the usability of search results because, in real-world business scenarios, obligatory and prohibitive clauses are often more critical and prioritized than permissive clauses, requiring higher weight in search ranking. The lack of this differentiated processing mechanism makes it difficult for existing systems to provide reliable support in mission-critical tasks.

[0007] At the generation and output levels, current automated methods also have shortcomings. Most systems output either the original clauses directly or a concisely compressed summary text. This unstructured result not only lacks readability but also hinders quick understanding and comparison by users. More importantly, users find it difficult to directly obtain structured information such as "responsible party—modal type—specific behavior—applicable conditions" from the results, and it's also difficult to directly trace the source of the clauses and their applicable validity. For industries requiring precise enforcement of regulations, this output method lacks verifiability and auditability, posing significant application risks.

[0008] Existing methods perform poorly when dealing with complex variations of regulatory texts. Regulatory clauses often contain long sentences, cross-sentence references, or contextual dependencies; without a text-level processing mechanism, the accuracy of automated extraction drops significantly. For example, if a subject appears in one sentence but the constrained behavior is only explicitly stated in the next, common methods often fail to connect the two, leading to information fragmentation. Furthermore, for clauses involving timeliness or version differences, the retrieval and generation results may be distorted without an effective time modeling mechanism.

[0009] Therefore, there is an urgent need for a new method for extracting and retrieving weighted obligations from the subjects of regulations and clauses to solve the above-mentioned technical problems. Summary of the Invention

[0010] This invention provides a method for extracting and retrieving weighted obligations of regulatory clauses, aiming to solve the problems of inaccurate identification of obligation modalities, missing conditional exceptions, unclear responsible parties, and lack of differentiation and verifiability of retrieval results in the processing of existing regulatory clauses, and to achieve high-precision extraction, weight optimization, and traceable generation.

[0011] This invention provides a method for extracting and retrieving weighted obligations of regulatory clauses, the method comprising the following steps: S1. Obtain the text of the rules and regulations, and preprocess the text of the rules and regulations to obtain preprocessed data; wherein, the preprocessed data includes a set of sentences, a dependency parsing graph, candidate trigger words, and location information; S2. Based on the preprocessing, perform modal discrimination to obtain the modal category and modal intensity; S3. Based on the preprocessed data and the modality category, perform ontology alignment and pointer network model localization to obtain the responsible entity; S4. Identify the associated conditions and exception fragments in the preprocessed data and regularize them into atomic predicates to construct logical expressions, thereby obtaining Boolean expressions and condition satisfaction levels; S5. Construct quadruples based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction degree, and perform retrieval and sorting processing to obtain quadruple data; S6. Generate structured output text based on the quadruple data.

[0012] Preferably, step S1 includes the following sub-steps: S11. Segment the text of the regulations and clauses according to the clause numbers and punctuation marks to obtain the set of sentences; S12. Establish document identifiers, clause identifiers, and sentence order indexes in the sentence set, and perform word segmentation and part-of-speech tagging on each sentence in the sentence set, thereby constructing a dependency syntax graph; S13. Identify the trigger words in the text of the regulations and clauses to obtain the candidate trigger words.

[0013] Preferably, step S2 includes the following sub-steps: S21. Extract the vector of the subsequent trigger word as a context representation based on the location information; S22. On the dependency syntax graph, with the candidate trigger word as the center, the domain node features are aggregated through a graph attention mechanism to obtain the syntax domain; S23. Concatenate the context representation and the syntactic domain to obtain a comprehensive feature; S24. Classify the comprehensive features to obtain the modal category and the modal intensity.

[0014] Preferably, step S3 includes the following sub-steps: S31. Based on the dependency parsing graph, label the candidate subjects in the sentence set; S32. Backtrack upwards along the edges of the dependency syntax graph based on the candidate trigger words to the noun phrases they govern; S33. Calculate the semantic similarity between the candidate subject and the standard items in the industry ontology library, and determine whether the semantic similarity is greater than a preset threshold: if so, proceed to step S34. S34. Determine the predicate span margin of the candidate subject based on the pointer network model to obtain the responsible subject.

[0015] Preferably, step S4 includes the following sub-steps: S41. Extract the association conditions and exception fragments associated with the candidate trigger words in the dependency syntax graph based on the candidate trigger words, extract the predicates and parameters based on the fragments, perform regularization and standardization processing, and uniformly formalize them into the Boolean expression; S42. Calculate the satisfaction degree of the association condition and the exception fragment to obtain the condition satisfaction degree.

[0016] Preferably, the Boolean expression satisfies the following condition: ; in, Indicates the applicable logic, This represents the conjunction operator. Represents the disjunction operator. Indicates logical negation; Represents a conditional index set. Indicates the set of exception indexes.

[0017] Preferably, step S5 includes the following sub-steps: S51. Construct the quadruple based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction degree; S52. Calculate the confidence level of the quadruple; S53. Perform retrieval and sorting processing on the quadruplets, and update them according to the confidence level; to obtain the quadruplet data.

[0018] Preferably, the confidence level satisfies the following condition: ; in, Indicates the modal intensity; Indicates the semantic similarity; Indicates the degree to which the condition is satisfied; The predicate span margin is represented by Conf, and the confidence level is represented by Conf. , , as well as All of these represent weighting coefficients.

[0019] Compared with existing technologies, the weighted method for extracting and retrieving the main obligations of regulatory clauses proposed in this invention has the following advantages: (1) It can accurately identify modal information of obligations, prohibitions and permissions in the text of rules and regulations, and combine the subject and behavior to form a complete semantic structure; (2) By using logical modeling of conditions and exceptions, the scope of application and exclusion of clauses is clarified, so that the results are operable and have the ability to determine the scenario; (3) Introduce a quadruple confidence and retrieval weighting mechanism, combined with structured generation and clause traceability information, to improve the accuracy, interpretability and compliance value of retrieval results. Attached Figure Description

[0020] The present invention will now be described in detail with reference to the accompanying drawings. The above and other aspects of the present invention will become clearer and more readily understood through the detailed description following the accompanying drawings. In the drawings: Figure 1 This is a flowchart of the method for extracting and retrieving weighted obligations of the main subjects of regulations and clauses provided in the embodiments of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] Please refer to Figure 1 This invention provides a method for extracting and retrieving weighted obligations of regulatory clauses, mainly used for extracting and retrieving the obligations of regulatory clauses in the power industry. The method includes the following steps: S1. Obtain the text of the rules and regulations, and preprocess the text of the rules and regulations to obtain preprocessed data; wherein, the preprocessed data includes a set of sentences, a dependency parsing graph, candidate trigger words, and location information.

[0023] In this embodiment of the invention, step S1 includes the following sub-steps: S11. Segment the text of the regulations and clauses according to the clause numbers and punctuation marks to obtain the set of sentences; S12. Establish document identifiers, clause identifiers, and sentence order indexes in the sentence set, and perform word segmentation and part-of-speech tagging on each sentence in the sentence set, thereby constructing a dependency syntax graph; S13. Identify the trigger words in the text of the regulations and clauses to obtain the candidate trigger words.

[0024] Specifically, when preprocessing the text of the aforementioned regulations and clauses, it is necessary to first standardize the encoding and clean the format (remove footnotes, headers and footers, watermarks and table remnants), and then follow the clause numbering pattern (e.g., "Chapter..."). strip"" (a) ") is used to divide sentences into clauses, resulting in a set of sentences: ; in, Indicates the first One sentence.

[0025] To enable traceable retrieval, an index is created for each sentence: doc_id, clause_id, .

[0026] Where doc_id is the unique identifier for the document, and clause_id is the unique identifier for the clause. This is the sentence order. Then... Perform word segmentation and part-of-speech tagging, and construct a dependency syntax graph: ; in, For word nodes, It is a dependency relationship (subject-predicate, verb-object, adverbial-head, prepositional-object, etc.).

[0027] By combining industry vocabularies, terminology standardization and unit normalization (e.g., unifying voltage / current / frequency to specified units and precision) are achieved, laying a consistent foundation for subsequent numerical condition parsing. A linear scanning method using left-to-right token order is employed within sentences. Identify deontic triggers. Deontic indicates a scope of obligation / prohibition / permission. The corresponding vocabulary is as follows: ; If the current word If the word list is explicitly hit, then the explicit score is given by the indicator function: ; in, 1 for a hit, 0 otherwise. To cover implicit expressions where no explicit trigger word appears, calculate the maximum cosine similarity between the context embedding and the category prototype: ; in, For words The context vector, The prototype vectors are categorized into three types: obligatory, prohibited, and permitted. Measure semantic proximity. Define trigger credibility by combining explicitness, semantics, and syntactic position: ; ; in, Intensity is assigned based on the dependent role (e.g., modal auxiliary verbs directly dominate predicates, which take higher values). The weights of explicit hit, semantic approximation, and syntactic position are controlled respectively.

[0028] Set threshold ,when When the word and its context constitute a valid normative tone trigger candidate, its metadata and location information are saved.

[0029] ; in, For the position of a word within a sentence, To predict the tone category; if If the current triggering candidate and its related context information are discarded, then the word will not be retained as a candidate for triggering the normative tone.

[0030] To reduce redundancy within the same clause, non-maximum suppression is applied to candidates with overlapping spans or excessively close adjacency, retaining only the highest-scoring candidate and merging context windows. Finally, the candidates, along with metadata and location information, are archived: preprocessed data includes doc_id, clause_id, and sentence order. Word segmentation and parts of speech, dependency syntax diagram The design includes clause hierarchy path, version validity, unit-normalized numerical fields, trigger categories, and scores for each item; location information includes character and token offsets, relative / absolute page numbers (if available), and number anchors. This design ensures that trigger detection is verifiable and traceable while also covering both explicit and implicit regulatory language.

[0031] S2. Based on the preprocessing, perform modal discrimination to obtain the modal category and modal intensity.

[0032] In this embodiment of the invention, the modality discrimination uses a concatenated vector of context representation and dependency graph aggregation representation as classification input, and the output probability is calibrated for confidence through temperature scaling.

[0033] Step S2 includes the following sub-steps: S21. Extract the vector of the subsequent trigger word as a context representation based on the location information; S22. On the dependency syntax graph, with the candidate trigger word as the center, the domain node features are aggregated through a graph attention mechanism to obtain the syntax domain; S23. Concatenate the context representation and the syntactic domain to obtain a comprehensive feature; S24. Classify the comprehensive features to obtain the modal category and the modal intensity.

[0034] Specifically, after completing the trigger candidate detection, it is necessary to identify the candidate trigger words. The normative mood category. Encoded at the sentence level by a pre-trained Transformer model, that is, the sentence containing the trigger word. The entire sentence is input to an encoder (such as BERT / RoBERTa) to obtain the representation of each word in the context of the entire sentence; the vector of the trigger word position is taken as the context representation. , denoted as:

[0035] Meanwhile, in dependency syntax graphs Above, with Centered on a node, features of its neighboring (i.e., directly adjacent) nodes are aggregated to form a syntactic neighborhood representation. Graph attention mechanism is used: ; in, This represents the neighborhood set of candidate trigger words in the dependency syntax graph; For candidate trigger word node vectors, It is the vector of its neighboring nodes; Given a linear transformation parameter matrix, learn the projection onto neighborhood features; An attention scoring vector is used to calculate... The correlation; It is a nonlinear function.

[0036] By combining context and syntax, we obtain a comprehensive feature:

[0037] The classification uses an external fully connected layer and a Softmax layer (not the internal structure of the base Transformer), outputting the conditional probabilities of the three modalities: ; in, Indicates that in a given trigger instance Below, it belongs to the modal category. The probability of; It is a discrete categorical variable, belonging to one of the following categories: Obligation-Based (OBL), Prohibition-Based (BAN), or Permission-Based (PER). To improve confidence stability, unnormalized logits are... Perform temperature scaling ( ): ; With: ; As modality strength, it reflects the certainty of the trigger for the corresponding modality in the current sentence, and can be used for subsequent quadruple construction and retrieval weighting.

[0038] S3. Based on the preprocessed data and the modality category, perform ontology alignment and pointer network model localization to obtain the responsible entity.

[0039] In this embodiment of the invention, step S3 includes the following sub-steps: S31. Based on the dependency parsing graph, label the candidate subjects in the sentence set; S32. Backtrack upwards along the edges of the dependency syntax graph based on the candidate trigger words to the noun phrases they govern; S33. Calculate the semantic similarity between the candidate subject and the standard items in the industry ontology library, and determine whether the semantic similarity is greater than a preset threshold: if so, proceed to step S34. S34. Determine the predicate span margin of the candidate subject based on the pointer network model to obtain the responsible subject.

[0040] Specifically, after determining modality categories and strengths, the system further identifies the responsible parties and corresponding behaviors constrained by the textual rules and regulations, ensuring a one-to-one correspondence between modalities and execution objects and actions, thus supporting subsequent retrieval weighting and generation. Subject identification utilizes dependency syntax graphs. As the framework: First use the domain NER in the sentence The text identifies potentially responsible positions, organizations, or equipment entities (such as operators, dispatching agencies, relay protection devices, maintenance units, substations, shift supervisors, etc.), and then selects candidate trigger words. Backtracking upwards along dependency edges to the noun phrases they govern ensures a tight syntactic connection between the subject and the trigger. To unify synonyms and abbreviations, and during initialization / periodic updates, a set of standard ontology items is loaded from the industry ontology repository to calculate candidate subjects. With a certain standard item Semantic similarity: ; when Standardized main body adopted at that time If the threshold is not reached, the fallback is either (i) retaining the original description and marking it as unstandardized, or (ii) mapping it to the nearest parent class (such as operators - electrical class) to avoid erroneous merging.

[0041] Behavior extraction based on candidate trigger words Using the predicate as a pivot point, the action is located along the predicate-object path and extended to a complete predicate phrase containing the necessary object / modifier (e.g., maintaining voltage stability, checking protection devices). To accurately determine the boundaries, a pointer network model based on Transformer encoding is used to measure the predicate span. Joint estimation means that, given a sentence and after triggering, the start and end positions are... The probability of:

[0042] in It is the first Conditional probability starting with a word. After the starting point is determined, the first Conditional probability of a word as the endpoint. Training objective: cross-entropy.

[0043] The constraint pointer network accurately locates the predicate boundaries. Through this step, the intra-sentence modality, subject, and behavior are stably aligned, providing a verifiable and traceable structural foundation for subsequent condition / exception resolution and the complete construction of quadruples.

[0044] S4. Identify the associated conditions and exception fragments in the preprocessed data and regularize them into atomic predicates to construct logical expressions, thereby obtaining Boolean expressions and condition satisfaction levels.

[0045] In this embodiment of the invention, step S4 includes the following sub-steps: S41. Extract the association conditions and exception fragments associated with the candidate trigger words in the dependency syntax graph based on the candidate trigger words, extract the predicates and parameters based on the fragments, perform regularization and standardization processing, and uniformly formalize them into the Boolean expression; S42. Calculate the satisfaction degree of the association condition and the exception fragment to obtain the condition satisfaction degree.

[0046] Specifically, after clarifying the subject and behavior, it is necessary to depict the limiting conditions and exceptions for the clause to take effect. (In dependency syntax diagrams) In this process, the associated conditions and exception fragments are extracted from the trigger word as the root. The exception fragments are first regularized, and then the predicates and parameters are extracted based on them. That is, the predicates and the value terms and comparison terms they participate in are identified after standardization (denoising, disambiguation, and unit unification).

[0047] Parameters are the value dimensions bound to predicates, such as voltage value, protection device status, load type, and whether the device is activated / locked. The resulting fragments are then merged into a set of atomic predicates. For example, voltage Relay protection has been activated, etc. The associated conditions and exception fragments are uniformly formalized into Boolean expressions: ; in This indicates the applicable logic of the rule clause text. For the conjunction operator that takes each element of a set as a pair, For disjunction operators, Indicates logical negation; For a conditional index set (which must satisfy the condition). This is an exception set of indexes (not applicable if certain conditions are met). This is to be considered in conjunction with specific query / execution scenarios. (e.g., equipment type, voltage level, load category, etc.), define satisfaction level: ; In the formula Indicates in the scene The overall applicability of the following clauses, This is a chain multiplication of the elements of a set. Matching degree of atomic predicates: numerical predicates (including thresholds / ranges, such as "voltage at...") The condition is continuously evaluated using interval / membership functions; discrete predicates (enumerating states / categories, such as device = started, load = class I) are calculated using Boolean indices or class membership degrees. Through this step, the conditions and exceptions in the text are extracted into verifiable fragments and condensed into computable ones. and This allows for the precise determination of the applicable boundaries and exclusion scope of obligations / prohibitions / permissions—subjects—behaviors, serving subsequent quadruple construction and retrieval weighting.

[0048] S5. Construct quadruples based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction, and perform retrieval and sorting processing to obtain quadruple data.

[0049] In this embodiment of the invention, step S5 includes the following sub-steps: S51. Construct the quadruple based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction degree; S52. Calculate the confidence level of the quadruple; S53. Perform retrieval and sorting processing on the quadruplets, and update them according to the confidence level; to obtain the quadruplet data.

[0050] Specifically, after completing the subject, action, and condition / exception analysis, the sentence will be... Abstracted into quadruples: ; in, As the responsible party, Modality (obligation / prohibition / permission). It is a predicate phrase. This is a logical expression for conditions / exceptions. To measure extraction reliability, the confidence level is defined to satisfy the following conditions: ; in, The modal intensity (maximum class probability) after temperature scaling in step one; The similarity between the main body and the industry ontology standard items; For condition satisfaction; The predicate span margin is defined as the normalized sum of the differences between the Top-1 and Top-2 confidence levels for both the starting and ending points, used to characterize boundary certainty. An empirical threshold is set. (e.g., 0.70), when Conf This indicates a high level of confidence.

[0051] During the retrieval phase, in the baseline rearranger (This can be achieved by injecting a quadrupole signal into a BM25+LR, LambdaMART learning sorter, or BERT cross encoder, etc., to obtain an updated score.) ; in, The subject extracted from the query. For subject semantic similarity, To query the scene metadata carried; The weighting coefficients for subject matching, modality strength, condition satisfaction, and overall confidence are measured to determine the contribution of each factor to the ranking results (these weights can be trained or set according to business needs). Updated Used for reordering candidate terms and selecting the Top-K: firstly, as evidentiary context for the RAG / reader; secondly, as interpretable retrieval logs (including...) This mechanism is used for traceability and auditing. It ensures that clauses that are more closely matched to the query subject, have clearer modalities, more suitable conditions, and are extracted with greater credibility are presented first, thereby improving the relevance and interpretability of the results.

[0052] S6. Generate structured output text based on the quadruple data.

[0053] In this embodiment of the invention, the structured output text includes four fields: subject, modality, behavior, and condition, and automatically generates a conflict prompt and clause number when a conflict modality is detected.

[0054] Specifically, after constructing the four-tuples and weighting the retrieval, the system outputs the top-ranked clauses in a readable and verifiable manner. When determining the optimal clause, it doesn't directly return the original text, but instead generates a structured answer based on the four-tuple data, mandating the inclusion of the three elements of "subject, modality, and behavior." Simultaneously, it automatically concatenates corresponding phrases to accurately present the scope of application and exclusion scenarios when conditions / exceptions exist. Example: Original clause text (excerpt) in When the load is running, the operator should maintain the bus voltage at [value missing]. Scope; unless a backup protection device has been activated. The extracted quaternion is ,obligation Maintain bus voltage at ,in (load kind) (Backup protection device) (Already started). The structured answer generated based on this is: the subject is the operator, the mode is obligation (should), and the behavior is to maintain the bus voltage at... Range, and automatically supplement only when This method is applicable during load testing; it is not applicable when backup protection devices are activated. It also outputs verification information to support backtracking, including the source clause number, hierarchical path, and evidence pinpoint (character or token offset). Users can easily jump back to the original text for location and review from the answer. This approach uses a subject-modality-behavior framework and explicitly injects conditions / exceptions, achieving a balance between readability, execution guidance, and audit interpretability.

[0055] If the system detects opposing modalities under the same topic (e.g., one clause "allows" and another "prohibits"), it adds a conflict warning to the result, clearly providing the relevant clause number and version information to facilitate comparison and adjudication. The generated answer is displayed in structured text format, including fields such as the subject's standardized name, modal type, behavioral predicate, applicable conditions, and clause ID. It is both directly readable and usable for subsequent auditing and traceability. This design makes the output not only understandable but also verifiable and legally compliant.

[0056] Therefore, in the method flow of this invention, each step is interconnected and progressively advances: text preprocessing and trigger detection establish standardized corpora and potential modal signals for subsequent processing, ensuring the standardization and completeness of the input; modality discrimination accurately distinguishes the semantic color of obligations, prohibitions, and permissions, providing a reliable basis for the binding nature of the clauses; subject and behavior extraction further clarifies the attribution of responsibility and operational actions, so that modalities no longer exist in isolation but are bound to specific execution objects and behaviors; condition and exception modeling formalizes the premises and exclusions of the clauses into logical expressions and can calculate applicability in different scenarios, thereby avoiding semantics from being out of context. Subsequently, the four-tuple construction unifies "subject-modality-behavior-condition" into a structured representation, and through confidence calculation and retrieval weighting mechanisms, clauses with greater binding force, applicability, and credibility are given priority in the ranking; finally, the structured generation stage transforms the results into verifiable text with clause numbers, timeliness information, and logical conditions, which is easy to understand and can also be used for traceability and compliance auditing. This forms a complete technical chain from text parsing and semantic modeling to retrieval generation, ensuring that the extraction and retrieval processes are accurate, interpretable, and practical.

[0057] Compared with existing technologies, the weighted method for extracting and retrieving the main obligations of regulatory clauses proposed in this invention has the following advantages: (1) It can accurately identify modal information of obligations, prohibitions and permissions in the text of rules and regulations, and combine the subject and behavior to form a complete semantic structure; (2) By using logical modeling of conditions and exceptions, the scope of application and exclusion of clauses is clarified, so that the results are operable and have the ability to determine the scenario; (3) Introduce a quadruple confidence and retrieval weighting mechanism, combined with structured generation and clause traceability information, to improve the accuracy, interpretability and compliance value of retrieval results.

[0058] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0059] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form without departing from the spirit and scope of the claims of the present invention, and all such changes are within the protection scope of the present invention.

Claims

1. A weighted method for extracting and retrieving the main obligations of regulatory clauses, characterized in that, The method for extracting and retrieving the weighted principal obligations includes the following steps: S1. Obtain the text of the rules and regulations, and preprocess the text of the rules and regulations to obtain preprocessed data; wherein, the preprocessed data includes a set of sentences, a dependency parsing graph, candidate trigger words, and location information; S2. Based on the preprocessing, perform modal discrimination to obtain the modal category and modal intensity; S3. Based on the preprocessed data and the modality category, perform ontology alignment and pointer network model localization to obtain the responsible entity; S4. Identify the associated conditions and exception fragments in the preprocessed data and regularize them into atomic predicates to construct logical expressions, thereby obtaining Boolean expressions and condition satisfaction levels; S5. Construct quadruples based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction, and perform retrieval and sorting processing to obtain quadruple data; S6. Generate structured output text based on the quadruple data.

2. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S11. Segment the text of the regulations and clauses according to the clause numbers and punctuation marks to obtain the set of sentences; S12. Establish document identifiers, clause identifiers, and sentence order indexes in the sentence set, and perform word segmentation and part-of-speech tagging on each sentence in the sentence set, thereby constructing a dependency syntax graph; S13. Identify the trigger words in the text of the regulations and clauses to obtain the candidate trigger words.

3. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 2, characterized in that, Step S2 includes the following sub-steps: S21. Extract the vector of the subsequent trigger word as a context representation based on the location information; S22. On the dependency syntax graph, with the candidate trigger word as the center, the domain node features are aggregated through a graph attention mechanism to obtain the syntax domain; S23. Concatenate the context representation and the syntactic domain to obtain a comprehensive feature; S24. Classify the comprehensive features to obtain the modal category and the modal intensity.

4. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 3, characterized in that, Step S3 includes the following sub-steps: S31. Based on the dependency parsing graph, label the candidate subjects in the sentence set; S32. Backtrack upwards along the edges of the dependency syntax graph based on the candidate trigger words to the noun phrases they govern; S33. Calculate the semantic similarity between the candidate subject and the standard items in the industry ontology library, and determine whether the semantic similarity is greater than a preset threshold: if so, proceed to step S34. S34. Determine the predicate span margin of the candidate subject based on the pointer network model to obtain the responsible subject.

5. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 4, characterized in that, Step S4 includes the following sub-steps: S41. Extract the association conditions and exception fragments associated with the candidate trigger words in the dependency syntax graph based on the candidate trigger words, extract the predicates and parameters based on the fragments, perform regularization and standardization processing, and uniformly formalize them into the Boolean expression; S42. Calculate the satisfaction degree of the association condition and the exception fragment to obtain the condition satisfaction degree.

6. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 5, characterized in that, The Boolean expression satisfies the following condition: ； in, Indicates the applicable logic, This represents the conjunction operator. Represents the disjunction operator. Indicates logical negation; Represents a conditional index set. Indicates the set of exception indexes.

7. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 6, characterized in that, Step S5 includes the following sub-steps: S51. Construct the quadruple based on the modal strength, the responsible subject, the Boolean logic expression, and the condition satisfaction degree; S52. Calculate the confidence level of the quadruple; S53. Perform retrieval and sorting processing on the quadruplets, and update them according to the confidence level; to obtain the quadruplet data.

8. The method for extracting and retrieving weighted obligations of regulatory clauses as described in claim 7, characterized in that, The confidence level satisfies the following condition: ; in, Indicates the modal intensity; Indicates the semantic similarity; This indicates the degree to which the condition is satisfied; The predicate span margin is represented by Conf, and the confidence level is represented by Conf. , , as well as All of these represent weighting coefficients.

Images