Hydraulic machinery feasibility research report grading review method based on large language model

Abstract

The invention discloses a hydraulic machinery feasibility research report grading review method based on a large language model, comprising the following steps: collecting and analyzing an original hydraulic machinery feasibility research report to generate a logic structure tree and a structured text; based on the logic structure tree, hierarchical gating and strategy scheduling are executed, and gating scheduling data are generated; according to review strategy parameters in the gating scheduling data, executing three-constraint parallel partitioning on the structured text to obtain a sub-document set; utilizing the gating scheduling data to execute auditing adversarial compliance review on each sub-document of the sub-document set to form a preliminary review conclusion set; and aggregating the preliminary review conclusion set, executing cross-subdocument consistency judgment and result summarization, and generating a final review conclusion. According to the method, the compliance problem of the hydraulic machinery research-length document report can be completely and accurately identified.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of water conservancy engineering and artificial intelligence, and relates to a hydraulic machinery feasibility study report hierarchical review method based on a large language model. BACKGROUND

[0002] The feasibility study report is a basis document for engineering demonstration, and report review can ensure the correctness of basic engineering data and the rationality of economy. The completeness, compliance and correctness of the feasibility study report are the focus of review. However, the feasibility study report has the characteristics of long length, wide knowledge coverage and strong professional nature. The traditional review process must be based on a large number of relevant standards, and often consumes a large amount of time and human resources, and may have problems such as omissions or deviations. Therefore, it is of great significance to study how to introduce an advanced report review method to improve the completeness, compliance and correctness of the feasibility study report, and to ensure the safety of the project and improve the management efficiency.

[0003] The emergence of the large language model brings a new solution to report review. The large language model has a huge number of parameters, which enables it to store a large amount of knowledge. In addition, the large language model is trained and learned on a large amount of corpus, has strong generalization performance and can easily understand the content of the text, and can output smoothly when conversing with humans. This makes it possible for the large language model to process text reports. Currently, in other fields, the large language model is widely used, such as building construction, grouting construction, aerospace, file review, power system, etc. These cases and practices show the broad prospects of the large language model.

[0004] However, the feasibility study report document involves a wide range of content, and the content is complex and related to a large number of specification clauses. The input review content often exceeds the context window limit of a single input of the large language model. Therefore, the excess part needs to be truncated and discarded, which cannot guarantee the integrity of the document content, and thus the review result of the large language model on the report deviates from the expectation. At the same time, the document review involves complex tasks, and the use of the truncated original report to input the large model in several times for review is also prone to loss of context information relevance. In addition, too long context can cause the model to easily ignore the middle content, thereby causing information loss during review.

[0005] Therefore, how to solve the review of the long document report of the hydraulic machinery has become a technical challenge to be solved.

[0006] The present application proposes a hydraulic machinery feasibility study report hierarchical review method based on a large language model to solve the above problems existing at present, reduce the labor intensity of the reviewer, and improve the accuracy of the review. SUMMARY

[0007] The application aims to provide a hydraulic machinery feasibility study report hierarchical review method based on a large language model to solve the above problems existing in the prior art.

[0008] The technical scheme is a hydraulic machinery feasibility study report hierarchical review method based on a large language model, comprising the following steps:

[0009] Step S1, collecting an original hydraulic machinery feasibility study report and parsing to generate a logical structure tree and structured text;

[0010] Step S2, based on the logical structure tree, performing hierarchical gating and strategy scheduling to generate gating scheduling data;

[0011] Step S3, according to the review strategy parameters in the gating scheduling data, performing three-constraint parallel partitioning on the structured text to obtain a sub-document set;

[0012] Step S4, using the gating scheduling data, performing an auditable adversarial compliance review on each sub-document of the sub-document set to form a preliminary review conclusion set;

[0013] Step S5, aggregating the preliminary review conclusion set, performing cross-sub-document consistency adjudication and result summarization, and generating a final review conclusion.

[0014] According to one aspect of the application, the step S2 is further:

[0015] Step S21, based on the standard template library constructed according to the hydraulic machinery industry specifications, quantitatively evaluating the integrity and standardization of the logical structure tree from multiple preset dimensions by comparing the standard template library, synthesizing a structure compliance score, and identifying a missing item list;

[0016] Step S22, setting a review threshold and determining whether the structure compliance score reaches the preset review threshold, and if not, terminating the subsequent process;

[0017] Step S23, extracting the logical structure tree that reaches the preset review threshold, and mapping the structure compliance score and the missing item list corresponding to the extracted logical structure tree to core parameters and resource budgets of downstream review tasks to generate the review strategy parameters.

[0018] According to one aspect of the application, the step S21 is further:

[0019] Step S21a, checking the coverage of the logical structure tree on the essential chapters in the standard template library to obtain a coverage score;

[0020] Step S21b, reviewing the chapter hierarchical relationship of the logical structure tree, and determining a hierarchical correctness score according to the structure violation;

[0021] Step S21c, measure the difference between the chapter order of the logical structure tree and the standard template library, and calculate the order consistency score;

[0022] Step S21d, traverse the content of the key chapter in the logical structure tree, evaluate the completeness of the internal technical elements, and obtain the element completeness score;

[0023] Step S21e, weight and sum the coverage score, hierarchical correctness score, order consistency score and element completeness score to generate the final structure compliance score, mark the missing items that do not appear in the structure tree, the key sub-items that are not effectively covered due to hierarchical attribution errors, and the missing elements that do not conform to the industry standard expression, and form a missing item list.

[0024] According to an aspect of the present application, the step S22 further comprises:

[0025] Step S22a, collect historical hydraulic mechanical feasibility study reports, and construct a sample set containing structure compliance scores of historical hydraulic mechanical feasibility study reports and their corresponding two-class artificial labeling labels, the two-class artificial labeling labels being binary qualitative judgment labels of whether the logical structure tree of the historical hydraulic mechanical feasibility study report meets the actual application requirements;

[0026] Step S22b, based on the sample set, generate a receiver operating characteristic curve, the curve representing the relationship between sensitivity and specificity at different score thresholds;

[0027] Step S22c, optimize on the receiver operating characteristic curve to maximize the Youden index, and determine the score value corresponding to the maximum point of the index as the review threshold.

[0028] According to an aspect of the present application, the step S23 further comprises:

[0029] Step S23a, according to the content of the missing item list, adjust the penalty weight of the subsequent document blocking step and the rearrangement weight of the specification retrieval step;

[0030] Step S23b, according to the score interval where the structure compliance score is located, set the semantic similarity threshold in the document blocking step, the retrieval depth in the specification retrieval step, and the budget allocation of the adversarial counter-retrieval step;

[0031] Step S23c, based on the hydraulic mechanical review priority, give priority weight and conflict to the verification correction parameter according to the missing item type and the score interval, and generate the review strategy parameter.

[0032] According to an aspect of the present application, the step S3 further comprises:

[0033] Step S31, calculate the semantic similarity between adjacent sentences by using cosine similarity method, and identify the semantic breakpoints in the structured text based on the semantic similarity between adjacent sentences and the similarity threshold in the review strategy parameter;

[0034] Step S32, filter the chapter boundary points based on the semantic breakpoints in the structured text, evaluate the credibility of the chapter boundary points in the structured text, and filter out the effective boundary points;

[0035] Step S33, consider the semantic breakpoints, the effective boundary points and the input length upper limit set by the review strategy parameter in parallel, perform conflict resolution when there is a constraint conflict among the three, and determine the final segmentation point;

[0036] Step S34, apply the final segmentation point to the structured text to generate the sub-document set.

[0037] According to an aspect of the present application, the step S33 is further:

[0038] Step S33a, follow a preset priority sequence, in which the priority of meeting the effective boundary point constraint is the highest, the priority of meeting the input length upper limit constraint is the second, and the priority of meeting the semantic breakpoint constraint is the lowest, identify the directory page and the special content of the chart block of the structured text, and apply exception rules to the directory page or the special content area of the chart block to overall retain the structural integrity thereof;

[0039] Step S33b, construct a target function for evaluating the segmentation quality, the target function is to minimize the deviation of the segmentation result from the ideal structured text, the constraint conditions include: numbering level logical constraint, chapter length effective constraint, segmentation point uniqueness constraint, professional term consistency constraint, format feature consistency constraint and cross-document consistency constraint, and the optimal final segmentation point is determined by optimizing the function value of the target function to minimize the function value.

[0040] According to an aspect of the present application, the step S4 is further:

[0041] Step S41, collect the hydraulic machinery technical document, construct the specification vector library based on the hydraulic machinery industry standard and the hydraulic machinery technical document, retrieve from the specification vector library and perform weighted rearrangement on the recall result according to the missing item list in the gating scheduling data, and select a candidate clause set for the sub-document;

[0042] Step S42, set a field template, extract compliance constraint information from the candidate clause set according to the field template, and construct a structured specification extraction object;

[0043] Step S43, comparing the sub-document with the specification extraction object, determining a preliminary compliance conclusion, and encapsulating evidence links containing specification clause identifiers;

[0044] Step S44, initiating differential adversarial re-search for the preliminary compliance conclusion to search for counter-evidence clauses, generating a conflict report, and integrating the preliminary compliance conclusion, evidence links, and conflict report to form the preliminary review conclusion set.

[0045] According to an aspect of the present application, the step S42 is further:

[0046] Step S42a, constructing a structured specification extraction object, including: limiting the specification extraction object to contain only pre-set structured fields, the structured fields including clause identifiers, mandatory levels, condition predicates, and quantitative constraints, and prohibiting subjective suggestions in the specification extraction object;

[0047] Step S42b, encapsulating evidence links containing specification clause identifiers, including: for each compliance determination, generating an evidence triple containing a specification clause identifier, a sub-document segment identifier, and a similarity score between the two, establishing traceable association between the sub-document and the specification clause.

[0048] According to an aspect of the present application, the step S42b is further:

[0049] Step S42b1, extracting the clause identifier in the specification extraction object as the first constituent element of the triple;

[0050] Step S42b2, locating the content segment corresponding to the clause identifier in the sub-document, and obtaining the segment identifier of the content segment as the second constituent element of the triple;

[0051] Step S42b3, calculating the semantic similarity between the content segment and the clause identifier, and taking the quantitative score of the similarity as the third constituent element of the triple, to represent the strength of evidence association.

[0052] According to an aspect of the present application, the step S44 is further:

[0053] Step S44a, when the preliminary compliance conclusion is compliance, searching for relevant exception clauses or mutually exclusive clauses in the specification vector library as counter-evidence;

[0054] Step S44b, when the preliminary compliance conclusion is non-compliance, actively retrieving possible remedial clauses or exemption clauses in the specification vector library as counter-evidence;

[0055] Step S44c, dynamically adjust the similarity threshold and the synonym set in the re-search process according to the gating scheduling data, generate a conflict report, and integrate the preliminary compliance conclusion, the evidence link and the conflict report to form the preliminary review conclusion set.

[0056] According to an aspect of the present application, the step S5 is further:

[0057] Step S51, in the aggregated preliminary review conclusion set, identify and enumerate potential conflicts across different sub-document conclusions, the potential conflicts including numerical conflicts, reference conflicts and mandatory level conflicts;

[0058] Step S52, set a ruling rule, and follow the priority sequence defined in the rule to resolve the potential conflicts, the priority sequence including that the conclusion of mandatory provisions is prior to the conclusion of recommended provisions, and the conclusion of the latest version of the standard is prior to the conclusion of the old version of the standard, to obtain the final review conclusion.

[0059] According to an aspect of the present application, the step S3 can also be:

[0060] Read the structured text data and the logical structure tree data, call the pre-stored specification clause base library data for subject identification and specification association analysis, and generate chapter specification association feature data;

[0061] Fuse the chapter specification association feature data and the specification path risk weight data updated based on historical review statistics, and perform weighted summary on each paragraph to generate paragraph specification risk score data;

[0062] Based on the paragraph specification risk score data, construct a candidate block scheme set, use a block objective function containing a specification path break penalty term to evaluate and sort each candidate block scheme, select the scheme with the minimum objective function value as the final block scheme, and generate sub-document set data and specification risk summary data.

[0063] Optionally, the block objective function includes a semantic coherence loss term, an effective boundary break penalty term, a length overflow penalty term and a specification path break penalty term; the specification path break penalty term is used to read the paragraph specification risk score data, identify paragraphs with risk scores higher than a preset value, if the paragraphs are distributed in different sub-documents across a distance exceeding a preset threshold in the candidate block scheme, then an additional penalty weight is imposed on the block objective function according to the risk score to reduce the dispersion of engineering facts corresponding to high-risk specification paths in the sub-document set data.

[0064] According to an aspect of the present application, the step S4 can also be:

[0065] Analyzing the constraint conditions, exception conditions and inheritance relationships of the clauses in the specification clause base library data, and constructing a specification constraint graph data for representing the specification hierarchy through the combination of nodes and edges;

[0066] Extracting engineering fact information from the sub-document set data and establishing a source mapping with the original text, and constructing a fact evidence graph data containing fact nodes and reference relationships;

[0067] Forward traversal along the mandatory constraint edges in the specification constraint graph data, matching the associated fact nodes in the fact evidence graph data to determine the constraint satisfaction, and generating a specification reasoning path set data containing a satisfaction conclusion or a non-satisfaction conclusion and a specification compliance local conclusion data.

[0068] Optionally, for the paths in the specification compliance local conclusion data that are determined to satisfy the conclusion, starting from the corresponding nodes in the specification constraint graph data, retrieving the matching exception fact nodes in the fact evidence graph data along the exception condition edges or the relaxation condition edges; for the paths determined to not satisfy the conclusion, starting from the corresponding nodes in the specification constraint graph data, retrieving the matching remediation fact nodes in the fact evidence graph data along the remediation condition edges or the exemption condition edges, generating an adversarial reasoning path set data; based on the adversarial reasoning path set data, revising the specification compliance local conclusion data to generate a final compliance review conclusion data, and combining the specification reasoning path set data and the adversarial reasoning path set data to construct an evidence path set data.

[0069] Optionally, the step S3 further comprises, after generating the paragraph specification risk score data:

[0070] Filtering the historical review task archive data for paths that have appeared in the non-satisfaction conclusion or have been adversarially revised, clustering and inducing the specification clause node sequences and fact node distribution patterns therein to construct a high-risk specification path template library data containing risk level labels;

[0071] Reading the paragraph specification risk score data and the sub-document set data, performing pattern matching on the current paragraph sequence and the high-risk specification path template library data, correcting and improving the specification risk score of the corresponding paragraph for the matching area, and generating corrected paragraph specification risk score data.

[0072] Optionally, the step S3 further comprises:

[0073] Based on the corrected paragraph specification risk score data, identifying areas with significantly increased specification risks, and performing merging or splitting operations on the block boundaries between high-risk paragraphs based on the sub-document set data to generate optimized sub-document set data;

[0074] Based on the optimized sub-document set data, the high-risk specification path coverage is calculated, the updated specification risk summary data is generated, and the optimized sub-document set data and the updated specification risk summary data are taken as inputs of the subsequent steps.

[0075] Optionally, the step S4 further comprises:

[0076] The high-risk specification reasoning mode is extracted from the historical archived evidence path, and the historical evidence path library data containing the specification path mode and the fact distribution mode are constructed;

[0077] The specification reasoning path set data is read, matched with the templates in the historical evidence path library data, and the high-risk specification path candidate set data is identified;

[0078] According to the risk level label in the high-risk specification path candidate set data, the corresponding sub-document is assigned with the countermeasure reasoning budget containing the extended exception clause layer and the relaxed similarity threshold.

[0079] Optionally, the step S4 further comprises:

[0080] Based on the countermeasure reasoning budget, the exception clause node and the remedial clause node combination consistent with the historical evidence path library data record are preferentially expanded in the specification constraint graph data, and the directional matching is performed in the fact evidence graph data, to generate the enhanced countermeasure reasoning path data;

[0081] The enhanced countermeasure reasoning path data and the countermeasure reasoning path set data are merged, and the specification compliance local conclusion data is revised based on the merged result, to generate the updated compliance review conclusion data and the evidence path set data.

[0082] According to an aspect of the present application, it further comprises the following steps:

[0083] Throughout the execution of the whole process of the method, a traceable parameter log is synchronously recorded and archived;

[0084] The traceable parameter log at least contains:

[0085] The name and version fingerprint of the algorithm model used in each step;

[0086] All threshold values, weights and budget parameters set in the hierarchical gating, parallel partitioning and countermeasure re-retrieval;

[0087] And the rule version and decision trajectory relied on in conflict resolution and consistency resolution.

[0088] Beneficial effects: a large language model-based hydraulic mechanical feasibility study report hierarchical review method is adopted, through hierarchical review and automatic rule retrieval, the compliance problems of the hydraulic mechanical feasibility study long document report are accurately identified, the labor intensity of the reviewer is reduced, and the accuracy of the review is improved. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is a flowchart of the present application.

[0090] Figure 2 is a flowchart of step S2 of the present application.

[0091] Figure 3 is a flowchart of step S3 of the present application.

[0092] Figure 4 is a flowchart of step S4 of the present application.

[0093] Figure 5 is a flowchart of step S5 of the present application. DETAILED DESCRIPTION

[0094] As Figure 1 shown, the following technical solutions are proposed. According to one aspect of the present application, a large language model-based hydraulic mechanical feasibility study report hierarchical review method is provided, characterized in that it comprises the following steps:

[0095] Step S1, collect the original hydraulic mechanical feasibility study report and parse it to generate a logical structure tree and a structured text;

[0096] In this step, the logical structure tree is a data structure used to represent the chapter-level relationship of the feasibility study report, which maps the chapter-section-clause of the report into the nodes of the tree, accurately reflecting the organization architecture of the content, rather than a simple text stream. The structured text refers to the text content that has been cleaned up, formatted uniformly, and associated with the nodes of the logical structure tree.

[0097] Specifically, the system receives an original hydraulic mechanical feasibility study report, which can be in the format of PDF, Word document or scan. First, through the layout analysis module, the document is divided into pages, paragraphs are recognized and charts are extracted. If it is a scan, an optical character recognition (OCR) engine is called to convert the text. On this basis, the system uses regular expressions and pattern matching to extract chapter titles and numbers from the text, and cross-verify them with the table of contents information of the document to accurately locate the chapter boundaries and repair the title or content breakage problems caused by pagination. Finally, a structured text is formed that is free of format noise and complete in content. Correspondingly, according to the recognized chapter title hierarchy, a logical structure tree is constructed that can reflect the logical context of the full text.

[0098] The purpose of performing this step is to convert the unstructured or semi-structured original report into standardized data objects that can be processed and calculated in subsequent steps, which is the basis for achieving automated review;

[0099] Step S2, based on the logical structure tree, performing hierarchical gating and policy scheduling to generate gating scheduling data;

[0100] The gating scheduling data includes structure conformity score, missing item list and review policy parameters.

[0101] Step S3, according to the review policy parameters in the gating scheduling data, performing three-constraint parallel partitioning on the structured text to obtain a sub-document set;

[0102] Step S4, using the gating scheduling data, performing an auditable adversarial compliance review on each sub-document of the sub-document set to form a preliminary review conclusion set;

[0103] The auditable adversarial compliance review includes specification retrieval, controlled abstraction, compliance determination and adversarial reverse retrieval, and the micro review conclusion set includes sub-document compliance conclusion, evidence link and conflict report.

[0104] Step S5, aggregating the preliminary review conclusion set to perform cross-sub-document consistency ruling and result summarization to generate a final review conclusion, wherein the preliminary review conclusion set includes the micro review conclusion set, and the structure conformity score and the missing item list in the gating scheduling data.

[0105] The method described in the application can be implemented based on a computer system, and an exemplary operating environment can include one or more servers, specifically, the servers can be configured with high-performance central processing units (CPUs), such as multi-core processors of Intel Xeon series, for performing general computing tasks and process scheduling, in order to accelerate the calculation of deep learning models, especially steps involving large-scale vector operations, the servers are preferably configured with graphics processing units (GPUs), such as NVIDIA Tesla or Ampere architecture computing cards.

[0106] In terms of storage, the servers can be configured with not less than 128GB of system memory (RAM) to support in-memory processing of large-scale data, and equipped with high-speed solid state drives (SSDs) or network storage for persistent storage of original hydraulic machinery research reports, processed intermediate data, model files and final review conclusions.

[0107] The implementation of the method can be based on a software program. In the present embodiment, the operating system of the server can be a Linux distribution, such as Ubuntu Server 20.04 LTS or later, on which a programming language environment is deployed, such as Python 3.8 or later. Further, the algorithm model involved in the method can rely on multiple open source frameworks and libraries, for example, using the PyTorch or TensorFlow framework for neural network model training and inference; using the Transformers library to load and use pre-trained large language models; using the Faiss vector retrieval library to build and query the standard vector library; and using the scikit-learn library to perform ROC curve generation, Youden index calculation, and other classic machine learning tasks. It can be understood that the above-mentioned software and hardware configuration is only an example, and those skilled in the art can adjust the configuration according to the performance requirements, cost budget and other factors of the actual application scene, which does not deviate from the protection scope of the present application.

[0108] As shown in Figure 2 According to one aspect of the present application, the step S2 is further characterized by:

[0109] Step S21, constructing a standard template library based on the hydraulic machinery industry specification, comparing the standard template library, quantitatively evaluating the integrity and standardization of the logical structure tree from multiple preset dimensions, synthesizing the structure compliance score, and identifying the missing item list;

[0110] Step S22, setting an examination threshold and determining whether the structure compliance score reaches the preset examination threshold, if not, terminating the subsequent process;

[0111] Step S23, extracting the logical structure tree that reaches the preset examination threshold, and mapping the structure compliance score and the missing item list corresponding to the extracted logical structure tree to the core parameters and resource budget of the downstream review task, generating the review strategy parameters.

[0112] Hierarchical gating is a kind of pre-posed quality evaluation mechanism, which first quantitatively evaluates the macro-structural integrity and standardization of the report before detailed content-level compliance review.

[0113] Specifically, the system will compare the logical structure tree generated in the previous step with a preset standard template library, which defines the mandatory chapters, recommended chapters and standard order between chapters that the hydraulic machinery feasibility study report should contain. The comparison process will be carried out from multiple dimensions such as coverage, hierarchical relationship, chapter order and key technical element completeness, and a final structure compliance score will be synthesized. Reports with a score below the preset threshold will be marked as unqualified, and a detailed missing item list will be attached to prompt users to supplement the content. For reports that pass the gating, the system will dynamically generate a set of review strategy parameters based on the structure compliance score and the specific content of the missing item list. For example, a report with complete structure and high score may be assigned more stringent review parameters, while a report missing only non-core chapters may have the weight of related modules adjusted in subsequent steps.

[0114] In this embodiment, the motivation for setting hierarchical gating is to achieve efficient use of resources. By filtering out reports with serious structural defects early, the waste of computing resources in subsequent complex and expensive microscopic review steps is avoided. At the same time, through strategy scheduling, the subsequent review is more targeted, realizing the intelligent change from one-size-fits-all review to individualized review.

[0115] According to one aspect of the present application, the step S21 is further characterized in that:

[0116] Step S21a, checking the coverage of the mandatory chapters in the standard template library by the logical structure tree, obtaining a coverage score;

[0117] Step S21b, reviewing the chapter hierarchical relationship of the logical structure tree, and determining the hierarchical correctness score according to the structure violation;

[0118] Step S21c, measuring the difference between the chapter order of the logical structure tree and the standard template library, and calculating the order consistency score;

[0119] Step S21d, traversing the contents of the key chapters in the logical structure tree, evaluating the completeness of the internal technical elements, and obtaining an element completeness score;

[0120] Step S21e, weighting and summing the coverage score, hierarchical correctness score, order consistency score and element completeness score to generate the final structure compliance score. Mark the missing items that do not appear in the structure tree, the key sub-items that are not effectively covered due to hierarchical attribution errors, and the missing elements that do not conform to the industry standard expression, and form a missing item list.

[0121] In this embodiment, in order to convert the ambiguous qualitative judgment of the report structure in the traditional review into a calculable and comparable quantitative index, the system introduces a multi-dimensional scoring system. Specifically, this step performs a fine evaluation on the logical structure tree Tree_struct of the input report by calling a pre-set standard template library Template_lib, which stores the standard directory structure, the list of mandatory chapters, and the recommended order of chapters for different types of hydraulic machinery projects (such as pumped storage and conventional hydropower stations).

[0122] The quantitative evaluation is realized by scoring and weighted sum of four dimensions, specifically:

[0123] The system first checks whether the Tree_struct contains all the mandatory chapters defined in the Template_lib (e.g., engineering geology, hydrology, engineering tasks and scale, etc.), and calculates the score based on the hit ratio. Optionally, the coverage of recommended chapters is also included with a lower weight, for example, if 9 out of 10 mandatory chapters are hit, the basic coverage score is 0.9;

[0124] The system traverses the nodes of Tree_struct to check for structural violations, such as the appearance of a cross-level phenomenon of 1.1.1 under 1 chapter, or the layer error of 2.1 section under 1 chapter. Each occurrence of structural violation deducts a fixed penalty score p_h (e.g., p_h = 0.05) from the full score 1.0;

[0125] The system extracts the main chapter sequence of Tree_struct and compares it with the standard sequence in Template_lib. Preferably, Kendall's Tau is used to measure the difference between the two sequences, and the result is normalized to the [0, 1] interval to obtain the sequence consistency score.

[0126] This dimension performs a more in-depth content check. The system checks whether the key technical elements are described in the core chapters (such as unit selection) using rules or lightweight models, for example, it checks whether the text fragments in the unit selection chapter mention key information such as rated head, single machine capacity, and unit number. The score is obtained based on the completeness of the elements.

[0127] Finally, the final structural compliance score Score_struct is generated by weighted summation of the following equations: Score_struct = w_coverage*Score_coverage + w_hierarchy*Score_hierarchy + w_order*Score_order + w_elements*Score_elements;

[0128] wherein Score_struct represents the final structural compliance score; Score_coverage represents the coverage score obtained in step S21a; Score_hierarchy represents the hierarchy correctness score determined in step S21b; Score_order represents the order consistency score calculated in step S21c; Score_elements represents the elements completeness score derived in step S21d; w_coverage, w_hierarchy, w_order, w_elements are the weight coefficients of each sub-score, which can be adjusted according to the focus of the review, and in a preferred embodiment, their default values can be set to 0.40, 0.25, 0.15, 0.20 respectively, while all the missed chapters or elements will be recorded to form a structured missing item list Missing_set.

[0129] According to an aspect of the present application, the step S22 is further characterized in that:

[0130] Step S22a, collecting historical hydraulic mechanical feasibility study reports, and constructing a sample set containing the structural compliance scores of the historical hydraulic mechanical feasibility study reports and their corresponding binary artificial labeling tags, which are binary qualitative judgment tags of whether the logical structure tree of the historical hydraulic mechanical feasibility study report meets the requirements of actual application;

[0131] Step S22b, generating a receiver operating characteristic curve based on the sample set, which represents the relationship between sensitivity and specificity at different score thresholds;

[0132] Step S22c, optimizing on the receiver operating characteristic curve to maximize the Youden index, and determining the score value corresponding to the maximum point of the index as the review threshold.

[0133] The motivation of this step is to avoid investing valuable computing resources into structurally incomplete and low-quality reports. However, a fixed review threshold T set by artificial experience (e.g. 0.7) often lacks universality and optimality.

[0134] To achieve the adaptability and scientificity of the threshold, the embodiment adopts a calibration method based on historical data driving. Specifically, the system maintains a sample set containing historical report Score_struct and its corresponding artificial labeling expert conclusion (two-class label: can enter the examination / need to correct and return), based on which a receiver operating characteristic curve (ROC curve) can be drawn, which intuitively shows the trade-off relationship between the sensitivity (the proportion of correctly identifying the reports that can enter the examination) and the specificity (the proportion of correctly identifying the reports that need to be corrected and returned) of the model under different score thresholds.

[0135] To find the best balance point, the system calculates the Youden_index corresponding to each threshold, which is defined as: Youden_index=sensitivity+specificity-1The system selects the score value that makes Youden_index reach the maximum value as the final examination threshold T. This method ensures that the selected threshold is the point with the strongest ability to distinguish between the two types of reports in a statistical sense. Further, to increase the robustness of the mechanism, when multiple maximum points of Youden_index exist, the embodiment preferably adopts a preset arbitration mechanism, that is, the point with higher specificity (more strictly intercepts unqualified reports) is preferred, and if the specificity is still the same, the point with higher threshold is selected.

[0136] According to one aspect of the present application, the step S23 is further characterized in that:

[0137] Step S23a, according to the content of the missing item list, adjust the penalty weight of the subsequent document blocking step and the rearrangement weight of the specification retrieval step;

[0138] Step S23b, according to the score interval where the structure compliance score is located, set the semantic similarity threshold in the document blocking step, the retrieval depth in the specification retrieval step, and the budget allocation of the adversarial anti-retrieval step;

[0139] Step S23c, based on the hydraulic mechanical review priority, give priority weight and conflict correction parameters according to the missing item type and the score interval, and generate the review strategy parameters.

[0140] Through the gated report, its score and missing item list are not discarded, but are used as valuable prior knowledge for fine-grained strategy scheduling of subsequent review steps.

[0141] The system converts Score_struct and Missing_set into specific execution parameters for downstream steps through a preset mapping module, specifically:

[0142] If Missing_set shows that a chapter is missing, the segmentation penalty weight of the context area involving the missing chapter in the subsequent S3 segmentation step can be appropriately reduced.

[0143] If Missing_set shows that environmental protection related content is missing, the rearrangement weight of the environmental protection related specification clauses in the subsequent S4 specification retrieval step will be dynamically adjusted upwards to actively find potential compliance problems.

[0144] For reports with a very high Score_struct (such as >0.9), it indicates that the structure is clear and reliable, and the system can set a higher semantic similarity threshold sim_threshold, and tend to segment at the explicit chapter boundary. Conversely, for reports with a lower score (such as 0.75-0.8), the sim_threshold can be appropriately reduced to allow more fine-grained semantic segmentation.

[0145] Reports with high scores can be assigned a smaller search depth top_k, and reports with low scores may require a larger top_k and a stronger adversarial anti-retrieval budget for more detailed verification.

[0146] As an optional implementation, the system can also dynamically allocate resources according to Score_struct, for example, differentially allocate the overall computing token quota and time quota according to the score, and ensure that more review resources are invested in reports with complex structures and more potential problems. When the budget is exhausted, the unfinished review tasks can be sent to a manual review queue Review_queue.

[0147] In this embodiment, the "general semantic similarity" trigger logic is abandoned, and a hydraulic machinery special compliance benchmark library is constructed with "professional structure compliance + core element completeness" as the trigger basis. The core standards are integrated, three types of rules including the essential chapter list, the core parameter threshold and the chapter logic relationship are extracted, a two-dimensional gating trigger index is designed, and a three-level gating threshold is set. The two-dimensional gating trigger index includes a structure compliance rate and a core element completeness rate. The coverage rate of the actual chapter to the "essential chapter list" is calculated, and the structure compliance rate is obtained by combining the chapter level correctness and weighting. The content corresponding to the "core parameter threshold" in the report is scanned, the proportion of parameter missing and parameter abnormality is counted, and the core element completeness rate is obtained.

[0148] The three-level gating threshold is specifically:

[0149] High compliance (structure compliance rate ≥ 90% and element completeness rate ≥ 85%): trigger "fast pass gating", directly enter the refined review;

[0150] Medium compliance (structure compliance rate 70%-90% or element completeness rate 70%-85%): trigger "condition passing control", mark missing / abnormal items, and focus on subsequent review;

[0151] Low compliance (structure compliance rate <70% or element completeness rate <70%): trigger "interception return control", generate a correction list (such as "missing structure strength review chapter" "safety factor not up to standard"), and terminate subsequent review.

[0152] As shown in Figure 3 According to one aspect of the present application, the step S3 is further characterized by:

[0153] Step S31, using cosine similarity method to calculate the semantic similarity between adjacent sentences, based on the semantic similarity between adjacent sentences and the similarity threshold in the review strategy parameter, identifying the semantic breakpoints in the structured text;

[0154] The system first performs sentence processing on the structured text Text_struct, then calls a pre-trained sentence vector embedding model to convert each sentence into a high-dimensional semantic vector. In this embodiment, a Chinese sentence vector embedding model with comparable effect to BGE (BAAI General Embedding) model is preferred. This kind of model performs well in Chinese semantic similarity calculation task. After obtaining the sequence of sentence vectors, the system calculates the cosine similarity between adjacent sentence vectors one by one to obtain a similarity change curve. When the similarity score is lower than a dynamically set semantic similarity threshold sim_threshold (the threshold is derived from the review strategy parameter generated in embodiment three), the position is marked as a candidate semantic breakpoint Cohesion_breaks. Optionally, to avoid false breakpoints caused by individual sentence noise, median filtering or other smoothing operations can be performed on the similarity sequence to enhance the robustness of breakpoint identification.

[0155] Step S32, based on the semantic breakpoints in the structured text, screening chapter boundary points, evaluating the credibility of the chapter boundary points in the structured text, and screening out effective boundary points;

[0156] The system does not regard all chapter boundaries in Tree_struct as equally important segmentation points. It integrates four types of evidence to comprehensively score the credibility score_bdry of each boundary point:

[0157] The standardization of title mode and numbering;

[0158] The mutual verification strength of table of contents information and main text information;

[0159] The location characteristics of the boundary in the page layout (such as whether it is top-aligned or page-changed).

[0160] the subject consistency of the context before and after the boundary point;

[0161] According to the comprehensive score, only those boundary points higher than the boundary threshold bdry_threshold are screened as valid boundary points Valid_boundaries;

[0162] Step S33, considering the semantic breakpoints, the valid boundary points, and the input length upper limit set by the review strategy parameter in parallel, when there is a constraint conflict, performing conflict arbitration to determine the final segmentation point;

[0163] Step S34, applying the final segmentation point to the structured text to generate the sub-document set.

[0164] After determining all the final segmentation points, the system applies them to the structured text to complete the segmentation operation, generates a series of sub-documents, constitutes a sub-document set Subdocs, and passes them to the subsequent microscopic review step. At the same time, detailed segmentation logs (including the basis for the decision of each segmentation point, abandoned candidate points, etc.) are recorded for tracing.

[0165] This step aims to segment a complete structured report text, which may be very long, into a series of sub-documents with moderate length and semantic cohesion, so as to facilitate input to the large language model for processing.

[0166] Specifically, the segmentation process is not simply performed according to a fixed length, but three core constraints are considered in parallel: A-semantic cohesion, i.e., the semantic correlation between adjacent sentences; B-integrity of chapter boundaries, i.e., try to segment at the end of a natural chapter; C-input length upper limit, i.e., ensure that the total length of the segmented sub-document plus the necessary prompt words does not exceed the upper limit of the large language model processing. When the three constraints conflict, for example, a semantic breakpoint happens to be in the middle of a chapter, the system will make a decision according to the preset priority (such as chapter boundary priority). The review strategy parameter plays a role in this step, for example, the structure conformity score obtained in the previous step will affect the sensitivity threshold of semantic segmentation.

[0167] This step effectively decomposes long text into a granularity suitable for model processing, while through a sophisticated arbitration mechanism, it maximizes the preservation of the structure information and semantic coherence of the original text, providing a guarantee for the accuracy of subsequent microscopic review.

[0168] According to one aspect of the present application, the step S33 further comprises:

[0169] Step S33a, follow the pre-set priority sequence, in which the priority of meeting the valid boundary point constraint is the highest, the priority of meeting the input length upper limit constraint is the second, and the priority of meeting the semantic break constraint is the lowest, identify the directory page and chart block special content of the structured text, and apply exception rules to the directory page or chart block special content area to overall retain its structural integrity;

[0170] Step S33b, construct a target function for evaluating the segmentation quality, which is to minimize the deviation of the segmentation result from the ideal structured text, and the constraint conditions include: numbered hierarchical logical constraint, chapter length validity constraint, segmentation point uniqueness constraint, professional term consistency constraint, format feature consistency constraint and cross-document consistency constraint, and determine the optimal final segmentation point by optimizing the function value of the target function. When determining the final segmentation point, the system will face three possible conflicting constraints at the same time:

[0171] Constraint A: the segmentation point should be located on the semantic break Cohesion_breaks;

[0172] Constraint B: the segmentation point should be located on the valid boundary point Valid_boundaries;

[0173] Constraint C: the length of the sub-document formed after segmentation, plus the reserved cue length, cannot exceed the input length upper limit token_cap of the model.

[0174] When the three constraints cannot be met at the same time, the system performs a priority-based conflict resolver, and in this embodiment, the priority sequence is pre-set as: B (valid boundary point) > C (input length upper limit) > A (semantic break), in other words, the system will give priority to ensuring segmentation at the natural structure boundary of the document, because this maximizes the integrity of the context, in the case where B cannot be met, the system will give priority to meeting the input length upper limit C of the model, because a super-long input will cause the model to be unable to process, and finally, under the premise of meeting B and C, the system will try to choose a semantic break A for segmentation, for example, when a valid boundary point B is too far from the last segmentation point, which will cause the length upper limit C to be exceeded, the system will give up segmentation at the B point and instead look for a segmentation point within the interval that meets C, at this time, the semantic break A will be given priority, further, the resolver also includes exception rules, for example, for special content areas such as directory pages and chart blocks, an overall retention strategy will be applied to avoid mistakenly segmenting them, at the same time, in order to optimize the segmentation effect, the system can also introduce a local backtracking correction mechanism, that is, after a segmentation point is initially determined, a limited backtracking search will be performed within a small window range (such as 5 sentences before and after) around it to find a better segmentation point that better meets the secondary constraints.

[0175] As an alternative or supplement to the hard rule priority arbitration, the final segmentation point can also be determined by a local optimization process, specifically, a target function J for evaluating the segmentation quality is constructed, and an optimization algorithm (such as a greedy search within a sliding window) is used to find the segmentation point that minimizes J, which can be a weighted sum of the following three components:

[0176] J = alpha * Cohesion_loss + beta * Boundary_break_penalty + gamma * Token_overflow_penalty;

[0177] Where Cohesion_loss represents the loss of semantic cohesion within the subdocument, the smaller the value, the more coherent the semantics within the subdocument after segmentation; Boundary_break_penalty represents the structural damage penalty caused by segmentation at non-preferred boundary positions, the value is negatively correlated with the boundary credibility score_bdry; Token_overflow_penalty represents the penalty for subdocument length exceeding the input length upper limit, the weight coefficients alpha, beta, gamma are the weight coefficients of semantic coherence loss, boundary break penalty, length overflow penalty, which can be calibrated on the validation set, by minimizing J, the system can more flexibly balance the three constraints to find a comprehensive optimal segmentation scheme.

[0178] The three constraints are not equal, and follow the priority "professional structural integrity > model adaptability > semantic coherence":

[0179] The first priority is the effective boundary point constraint, which prioritizes the structural integrity of the core chapter of hydraulic machinery (such as unit selection, hydraulic calculation), which is the basis of professional review, and can avoid breaking the chapter logic loop due to segmentation;

[0180] The second priority is the input length upper limit constraint, when the effective boundary point constraint cannot be met (such as a core chapter being too long), the model processing capacity is used as the bottom line to ensure that the subdocument does not exceed the token limit, avoiding review bias caused by content truncation;

[0181] The third priority is the semantic breakpoint constraint, under the premise of meeting the first two constraints, try to choose semantic breakpoint segmentation to reduce the fragmentation of professional semantics, and consider the review accuracy.

[0182] When the three constraints conflict, for example, in some embodiments, the distance between the chapter boundary point and the split point is too far, and the split will exceed the upper limit of the input length → abandon the boundary point, find the semantic breakpoint within the chapter, ensure the sub-document length compliance, and maximize the continuity of the calculation logic;

[0183] In another embodiment, the semantic breakpoint coincides with the valid boundary point, but the sub-document length after splitting exceeds the upper limit → preferentially compress the token occupation of non-core description content (such as background introduction), and if it still cannot meet the requirements, fine-tune the splitting position near the breakpoint to balance the length and semantic integrity;

[0184] Special professional areas (such as complete hydraulic formula derivation) do not have valid boundary points, but the semantic coherence is strong → trigger the exception rule, and the entire area is retained, even if the sub-document length is close to the upper limit, the professional logic chain is not split.

[0185] The three are considered in parallel and dynamically coordinated to finally realize the "three guarantees": guaranteeing the adaptation of the sub-document to the processing capacity of the large language model (input length upper limit constraint), guaranteeing the integrity of the professional structure and core chapter (valid boundary point constraint), and guaranteeing the coherence of the professional semantics and logic chain (semantic breakpoint constraint), solving the pain points of traditional blocking "either cutting professional content or destroying semantics".

[0186] In the three-constraint parallel blocking mechanism, the valid boundary point constraint is the core constraint to ensure the integrity of the professional structure of the hydraulic machinery feasibility study report. Its screening logic revolves around the "natural structure boundary of the hydraulic machinery field document", and through multi-dimensional evidence verification and field adaptation rules, it accurately identifies and retains boundary points that meet the professional review requirements. The specific process is as follows:

[0187] Check if the title number meets the specifications of the hydraulic machinery field, identify the binding relationship between professional keywords (such as "unit rated parameters" "hydraulic efficiency calculation" "safety redundancy design") and title format, exclude the interference of non-professional titles (such as "report preparation instructions" "references"), and directly reduce the boundary credibility score of titles that do not meet the professional number specifications or lack core professional keywords (such as from full score 1.0 to 0.3 or below);

[0188] Extract the title, page number, and brief description of the hydraulic machinery core chapter in the table of contents, and compare the content at the corresponding position with the text. If the professional keywords in the text content match the brief description in the table of contents with a matching degree of ≥80% and a page number deviation of ≤2 pages, the boundary credibility score is increased. If there is a title in the table of contents but no corresponding professional content in the text, or the text content is irrelevant to the brief description in the table of contents (such as the table of contents marked "unit parameters" and the text "project background"), it is marked as "invalid boundary point" and the credibility score is 0;

[0189] If the title is aligned to the top of the page and the first line of the new page, the confidence score of the boundary is increased (e.g. by 0.15). If the title is in the middle of the page without any space, the score is decreased (e.g. by 0.2).

[0190] For the 500 words before and after the boundary point, extract the keywords related to the water power machinery field (e.g. "water head", "flow", "compressive strength", "fatigue life"). If the overlap of the keywords before and after is less than or equal to 30% (e.g. "hydraulic parameters" before and "structural parameters" after), the effectiveness of the boundary point is confirmed and the score is increased (e.g. by 0.25). If the overlap of the keywords before and after is greater than or equal to 60% (e.g. "efficiency calculation steps" before and "efficiency calculation examples" after), it is determined that it is a false boundary of semantically cohesive content and the score is decreased (e.g. by 0.3).

[0191] The effective boundary point constraint is the first priority constraint in the "three constraints", and its screening result directly affects the block logic:

[0192] If the length of the sub-document corresponding to the effective boundary point does not exceed the "input length upper limit constraint" (e.g. the sub-document contains the complete content of "3.2 water power unit selection", and the length is less than or equal to the upper limit of the model token), the boundary point is directly used as the cutting point, and the "semantic breakpoint constraint" is not needed.

[0193] If the length of the sub-document corresponding to the effective boundary point exceeds the "input length upper limit constraint" (e.g. the "4 structural strength checking" chapter is too long), the "semantic breakpoint constraint" is combined within the chapter (e.g. cutting at the semantic breakpoint between "4.1 checking parameters" and "4.2 calculation process") to balance the professional structure integrity and model adaptability.

[0194] If there is no effective boundary point (e.g. a continuous water power calculation process without a title), the "semantic breakpoint constraint" is used as the main constraint to ensure that the professional logic is not fragmented.

[0195] The core value of the water power machinery feasibility report lies in "professional data integrity" and "technical logic continuity". Traditional block division (e.g. fixed length cutting, general semantic cutting) often ignores the industry characteristics, resulting in fragmentation of professional content. This mechanism realizes "lossless cutting" of professional documents through field adaptation design.

[0196] The effective boundary point constraint prioritizes the natural boundary of the hydraulic machinery core chapter as the basis for segmentation, combined with "title standardization + directory mutual verification + theme consistency" multi-dimensional verification, to avoid splitting the core chapter, for example, the "4.2 turbine rated parameter demonstration" chapter in a certain report contains complete parameter calculation logic and charts. Traditional fixed-length segmentation may split "parameter derivation process" and "conclusion" into two sub-documents, while the mechanism ensures the integrity of the chapter through effective boundary point identification, ensuring that the parameter demonstration logic can be evaluated completely during review;

[0197] The semantic breakpoint constraint is based on a pre-trained model in the field of hydraulic machinery, which identifies strong association logic of professional terms such as "rated water head-flow calculation-efficiency check". Only at the semantic natural breaking point, a breakpoint is marked, for example, when segmenting the "hydraulic design" chapter, the continuous logical chain of "formula derivation-data substitution-result analysis" is avoided, and the problem of misjudgment of "data source and calculation logic mismatch" by the reviewer or model is avoided due to segmentation, especially suitable for the text structure of the hydraulic machinery report with a large number of "calculation process + professional conclusion" bindings;

[0198] For "hydraulic calculation chart blocks", "unit parameter comparison tables", and "complete formula derivation segments" in hydraulic machinery reports, the mechanism enforces overall segmentation through exception rules to avoid the problem of traditional segmentation splitting "charts and explanatory text" and "formulas and derivation steps", for example, in a report, "head-flow characteristic curve" and "note: the curve is calculated based on GB / T15468-2013 standard" below, traditional segmentation may split the chart and notes into different sub-documents, while the mechanism will mark the two as "special area boundaries".

[0199] The core requirement of the hydraulic machinery feasibility study report review is "no omission of compliance issues", and traditional segmentation often leads to review result deviation (such as missing parameter conflict and misjudging specification reference) due to "context loss" and "professional association break". The mechanism provides "high-quality sub-documents" for subsequent review steps through constraint coordination, directly improving review accuracy;

[0200] The segmentation preserves the professional logical coherence, making the "parameter calculation-result reference-specification correlation" in the sub-document form a complete closed loop, for example, the "5.3 structural load-bearing capacity calculation" chapter in a certain report contains "material strength parameters-load calculation-safety factor check" continuous logic, the mechanism will segment it as a whole into a sub-document, and subsequent compliance review can directly associate "whether the material parameters meet GB50204-2015 standard" and "whether the safety factor calculation matches the load condition", avoiding the misjudgment of "parameter source unknown" and "calculation logic break" caused by traditional segmentation due to splitting;

[0201] The effective boundary point constraint combined with "catalog and text mutual verification" can accurately locate the key areas such as "specification reference list" to ensure that the sub-documents contain complete "specification number-clause content-reference scene" information. For example, in a report, "6.1 Safety Protection Design" refers to Article 5.2 of "Hydraulic Machinery Safety Design Specification" (SL316-2022). Traditional segmentation may separate "specification number" and "reference clause content", leading to missed "clause applicability" during review. However, the mechanism uses effective boundary point identification to segment "specification reference + design content" as a whole, improving the completeness of specification compliance review.

[0202] Through "constraint coordination", the mechanism reduces "semantic redundancy" and "logical conflicts" between sub-documents. During subsequent cross-sub-document consistency adjudication, only the "parameter consistency of different chapters" (such as "unit set rated power" in "selection chapter" and "energy consumption chapter") needs to be focused on, without dealing with "semantic ambiguity conflicts" caused by improper segmentation. For example, traditional segmentation may separate "unit set number" into two sub-documents with inconsistent expressions (one writes "4 units" and the other writes "four units"). However, the mechanism uses semantic breakpoint constraint to unify professional terms.

[0203] Hydraulic machinery feasibility reports are usually long (more than 500 pages) and complex in terms of profession. Traditional segmentation either wastes computing resources due to "too many sub-documents" or exceeds model processing capacity due to "too long sub-documents". However, the mechanism optimizes resource efficiency through dynamic constraint coordination.

[0204] The upper limit of input length constraint strictly matches the context window of large language models, combined with "professional semantic compression" (prioritizing core parameters and calculation processes, and appropriately simplifying non-professional descriptions) to ensure that sub-document length neither exceeds model limits nor causes excessive segmentation leading to a large number of sub-documents. For example, a 1000-page large hydraulic machinery report may generate 50+ sub-documents through traditional fixed-length segmentation. However, the mechanism uses effective boundary points and semantic breakpoints to generate only 20-25 sub-documents, reducing model inference time by 40% and keeping memory usage below 64GB (compatible with mainstream GPU configurations).

[0205] The mechanism filters out invalid segmentation through "multi-constraint checking" to reduce subsequent manual intervention costs. For example, traditional segmentation may mix "non-core chapters (such as project background)" with "core chapters (such as hydraulic calculation)", requiring manual selection of valid content during review. However, the mechanism uses "professional importance sorting" of effective boundary points to separate core chapters from non-core chapters. Non-core chapters can use a lightweight review strategy (such as quick compliance scanning), while core chapters use in-depth review. Manual work focuses only on core chapter issues, reducing labor costs by more than 30%.

[0206] Combined with the "structure conformity score" output by the upstream "hierarchical gating and strategy scheduling", the mechanism dynamically adjusts the blocking parameters: for reports with high structural integrity (score ≥ 0.9), increase the semantic similarity threshold to reduce the number of sub-documents; for reports with incomplete structure (score 0.7-0.8) and containing complex hydraulic calculation, reduce the threshold and increase the effective boundary point check frequency to ensure the accuracy of blocking, for example, a regular hydropower station report with complete structure can be preferentially divided by chapter boundary, resulting in fewer sub-documents and higher processing efficiency; for a pumped storage power station report with incomplete structure (containing complex two-way flow calculation), fine-grained semantic breakpoint division is used to ensure that the calculation process is not split.

[0207] Step S3 can also be:

[0208] Step S3a, for performing adaptive blocking based on specification path risk, to concentrate engineering facts related to high-risk specification constraints in as few sub-documents as possible, facilitating subsequent reasoning processing on specification constraint graphs and fact evidence graphs, including the following sub-steps.

[0209] Step S3a1, read the structured text data, logical structure tree data and gating strategy parameter data output by the previous steps, and call the pre-stored specification clause basic library data;

[0210] Based on the chapter level information in the logical structure tree data and the clause theme, applicable object and key terms in the specification clause basic library data, perform theme recognition and specification association analysis on each paragraph or sentence in the structured text data, establish an association relationship between each paragraph and one or more specification constraint themes, and obtain chapter specification association feature data for representing the correspondence between chapters and specification constraints, wherein each entry in the chapter specification association feature data includes at least paragraph identification, chapter identification and specification constraint theme identification related to the paragraph, for calculating the influence of specification path risk on blocking in subsequent steps;

[0211] Step S3a2, based on the specification clause basic library data and the pre-configured risk rule library, assign a basic risk weight to each specification constraint theme according to the mandatory level of the specification clause, the constraint object category, and the situation involving sensitive indicators such as safety and environmental protection, to form specification basic risk weight data for describing the importance of specification constraints;

[0212] On this basis, combined with the historical review statistical information accumulated in the preceding steps, the risk weight of the specification constraint theme that frequently does not meet the conclusion or is frequently corrected by the conclusion of the antagonistic reasoning in the past review tasks is appropriately increased, to obtain updated specification path risk weight data, and then, the chapter specification association feature data and the specification path risk weight data are fused. For each paragraph, the specification constraint theme associated therewith and the corresponding risk weight are weighted and summarized to generate paragraph specification risk score data reflecting the importance of each paragraph in the specification level review, wherein each entry in the paragraph specification risk score data at least includes a paragraph identifier and a paragraph specification risk score, for imposing differentiated constraints in the subsequent block target function;

[0213] Step S3a3, reading the structured text data, the logical structure tree data, the paragraph specification risk score data, and the gating strategy parameter data, first constructing a set of candidate block boundary set according to the natural chapter boundary, the title level and the format information in the logical structure tree data, combining the semantic similarity threshold and the maximum input length constraint, to obtain a set of preliminary candidate block scheme data. On this basis, for each candidate block scheme, according to the preset block target function, the semantic coherence loss, the effective boundary break penalty and the length overflow penalty under the block scheme are calculated, and at the same time, the specification path-based breaking penalty term is introduced, that is, for the paragraphs with high risk scores in the paragraph specification risk score data, if these paragraphs are distributed in multiple sub-documents and the distance between the sub-documents exceeds the preset threshold, an additional penalty is added to the block scheme according to the specification risk score of the corresponding paragraph, so as to obtain the block target function evaluation data for measuring the advantages and disadvantages of each candidate block scheme.

[0214] In a preferred embodiment, the block target function can be represented as a target function J_star in the form of a formula, wherein the semantic coherence loss term, the effective boundary break penalty term, the length overflow penalty term and the specification path breaking penalty term correspond to different weight coefficients respectively, and the weight coefficients are obtained by parameter calibration on the labeled data set to ensure that the paragraphs corresponding to the high-risk specification path have greater influence in the target function;

[0215] The specific form of the block target function J_star is as follows:

[0216] J_star=λ1×L_cohesion+λ2×L_boundary+λ3×L_overflow+λ4×L_path;

[0217] Wherein: J_star represents the overall objective function value of the block scheme, the smaller the value, the higher the block quality; L cohesion represents the semantic coherence loss term, calculated as the negative mean value of the semantic similarity of adjacent sentences within the sub-document; L boundary represents the effective boundary break penalty term, which takes the value of 1 if the split point is not located on the effective boundary point, otherwise 0; L overflow represents the length overflow penalty term, calculated as max(0, sub-document length-input length upper limit) / input length upper limit; L path represents the standard path breaking penalty term, calculated as the degree of dispersion of high-risk paragraphs into different sub-documents; λ1, λ2, λ3, λ4 are the weight coefficients of each penalty term, the default values are 0.3, 0.3, 0.2, 0.2 respectively.

[0218] Step S3a4, based on the candidate block scheme set data and the corresponding block objective function evaluation data, all candidate block schemes are sorted according to the value of the objective function J_star, and the candidate block scheme with the minimum value of the objective function J_star is selected as the final block scheme, and the sub-document set data for subsequent standard review is obtained.

[0219] At the same time of selecting the final block scheme, according to the distribution of the paragraph specification risk score data in each sub-document under the scheme, the total score of the specification risk contained in each sub-document and the coverage of the high-risk specification path are calculated, and the specification risk summary data for guiding the subsequent allocation of adversarial reasoning depth and search budget are formed. Then, the sub-document set data is provided to the subsequent specification constraint graph reasoning step for constructing the fact evidence graph, and the specification risk summary data is incorporated into the gating strategy parameter data to update the new gating strategy parameter data, so that the subsequent specification reasoning and adversarial search can allocate resources based on more detailed risk distribution information when reviewing high-risk specification paths.

[0220] In an alternative scheme of the above step S3a, the functions of the parent step S3 are fully implemented through the sub-steps S3a1 to S3a4, that is, starting from the structured text data and the logical structure tree data, combining the specification clause basic library data and the gating strategy parameter data, gradually generating the chapter specification association feature data, the specification path risk weight data and the paragraph specification risk score data, on this basis, constructing the candidate block scheme set data and the block objective function evaluation data, finally obtaining the sub-document set data for the subsequent S4 step and the specification risk summary data for updating the gating strategy, all intermediate data are used in the subsequent steps, there is no isolated data.

[0221] Step S3 can also be S3b, for introducing patterned risk correction based on historical specification reasoning path templates on the basis of performing specification path risk adaptive chunking, so that the chunking strategy can further improve the examination sensitivity for specification paths that have historically been prone to problems.

[0222] The specific steps of step S3b include the following sub-steps:

[0223] Step S3b1, on the basis of performing steps S3a1 to S3a4, from the evidence path set data, compliance examination conclusion data and corresponding gating strategy parameter data that have been archived in historical examination tasks, filter out specification paths that have not met the conclusion or have been significantly corrected by adversarial reasoning many times, cluster and induce the specification clause node sequence, fact node distribution pattern and corresponding sub-document distribution characteristics appearing in these specification paths, and build a high-risk specification path template library data for depicting high-risk specification reasoning patterns;

[0224] Among them, each template in the high-risk specification path template library data includes specification clause node sequence pattern, common fact element combination pattern, typical sub-document distribution range and corresponding risk level label, which is used to provide patterned risk prompts when new reports are chunked;

[0225] Step S3b2, read the paragraph specification risk score data, sub-document set data and specification risk summary data obtained in step S3a of the current report and the offline built high-risk specification path template library data, and perform pattern matching on the paragraph sequence and sub-document division result in the current report. For regions that can reach a preset similarity threshold in specification clause node sequence and fact distribution pattern with a certain template in the high-risk specification path template library data, the specification risk score of the corresponding paragraph and sub-document is corrected and improved, and the corrected paragraph specification risk score data and updated specification risk summary data are generated.

[0226] On this basis, for the region where the corrected specification risk is significantly increased, the local chunking boundary is allowed to be fine-tuned on the basis of the original sub-document set data, including merging or splitting the boundary between high-risk paragraphs, so as to form optimized sub-document set data for replacing the original chunking result. Then, the optimized sub-document set data is used to replace the original sub-document set data to pass to the subsequent specification constraint graph reasoning step, and the updated specification risk summary data is written back to the gating strategy parameter data to guide the subsequent adversarial reasoning budget allocation for high-risk specification paths. As Figure 4 As shown, according to one aspect of the present application, the step S4 is further characterized by:

[0227] Step S41, collect hydraulic machinery technical documents, build a specification vector library based on hydraulic machinery industry standards and hydraulic machinery technical documents, retrieve from the specification vector library and according to the missing item list in the gate scheduling data, perform weighted rearrangement on the recall results, and select a candidate clause set for the sub-document;

[0228] The specification vector library is a pre-constructed knowledge base containing a large number of national standards, industry specifications, technical guidelines and other related information in the field of hydraulic machinery. All clauses have been converted into high-dimensional semantic vectors to form an index for efficient retrieval.

[0229] The specific implementation is: a sub-document Subdoc from Example Four is used as a query (Query), and approximate nearest neighbor (ANN) retrieval is performed in the specification vector library to preliminarily recall a batch of semantically most relevant candidate specification clauses. However, pure semantic relevance is not enough to deal with the complexity of the review. In order to make the retrieval more targeted, the system introduces a rearrangement mechanism driven by the previous step. Specifically, the system reads the missing item list Missing_set generated in Example Three. For example, if Missing_set indicates that the report has structural missing in dam seismic design, the rearrangement module will dynamically increase the sorting weight of all clauses related to seismic, seismic intensity, dynamic stability, etc. in the preliminary recall results.

[0230] In this way, the macro-structural insight of the review (from S2 gate) is effectively transmitted to the micro-level clause-level review, realizing intelligent linkage across stages. This step finally generates an optimized and highly relevant candidate clause set Candidate_norms for each sub-document.

[0231] As an optional implementation, the rearrangement module can be a rule-based weighting function, or a more complex learning-to-rank model (Learning-to-Rank Model), such as LambdaMART, which can integrate semantic relevance, missing item correlation, clause timeliness and other features for comprehensive sorting.

[0232] Step S42, set a field template, extract compliance constraint information from the candidate clause set according to the field template, and build a structured specification extraction object;

[0233] The system uses a strict, pre-set field template to build the specification extraction object Compliance_basis, which is a structured data entity rather than natural language text. Preferably, the template limits the following seven fields that must be included and only include the following seven fields:

[0234] clause_id: unique identifier of the norm clause;

[0235] mandatory_level: mandatory level, such as mandatory, recommended, or reference;

[0236] condition_predicates: condition predicates applicable to the clause, i.e., the if part of the if...then... rule;

[0237] quantity_constraints: quantity constraints, which is a substructure containing (variable_name, unit, comparator, value / interval);

[0238] Exceptions: exception situation description of the clause;

[0239] cross_refs: cross-referenced clause IDs in the clause;

[0240] source_span_id: location identification of the clause in the original specification document.

[0241] The system will construct a specific prompt instruction (Prompt) to guide the large language model to parse information from the text of Candidate_norms and fill in the seven fields like filling in a form, and any subjective suggestions or explanations outside the template will be prohibited from being generated.

[0242] Further, in order to ensure the robustness of the system, the embodiment also designs a fault-tolerant fallback mechanism. If the model fails to successfully extract the key fields (such as quantity_constraints) from a certain clause, the system will not report an error and interrupt, but will directly store the original text segment of the clause in Compliance_basis and attach a label to be supplemented, and at the same time, send the entry to the artificial review queue Review_queue mentioned in Embodiment Three and mark the failure reason as EXTRACTION_FAIL.

[0243] Step S43, comparing the sub-document with the specification extraction object, determining a preliminary compliance conclusion, and encapsulating evidence links containing norm clause identifiers;

[0244] The system will compare the content in the sub-document Subdoc with the constraints in Compliance_basis one by one, for example, check whether the working condition described in Subdoc meets the condition_predicates in Compliance_basis, and whether the technical parameter value reported in Subdoc meets the quantity_constraints in Compliance_basis (for example, the reported compressive strength of concrete is 35 MPa, and the specification constraint is comparator='>=' and value=30 MPa, so it is determined to be met).

[0245] In order to build traceable association, each judgment generates an evidence triple and is integrated into Evidence_links, which is a bridge connecting facts (from sub-documents) and rules (from specifications), and its specific composition is:

[0246] The first component of the triple: extract the clause_id in the specification extraction object as the specification clause identifier; the second component of the triple: locate the specific content segment in the sub-document related to the clause judgment and obtain the unique identifier span_id of the segment as the sub-document segment identifier;

[0247] The third component of the triple: calculate the semantic similarity (such as cosine similarity) between the sentence vector of the content segment and the sentence vector of the specification clause, and take the quantitative score (such as 0.92) as the confidence score representing the strength of evidence association.

[0248] Step S44, for the preliminary compliance conclusion, initiate differential and counter-research to search for counter-evidence clauses, generate conflict reports, and integrate the preliminary compliance conclusion, evidence links, and conflict reports to form the micro-examination conclusion set.

[0249] Using the gate scheduling data, perform an auditable and counter-researched compliance review on each sub-document of the sub-document set to output a micro-examination conclusion set, wherein the auditable and counter-researched compliance review includes specification retrieval, controlled extraction, compliance judgment, and counter-researched retrieval, and the micro-examination conclusion set includes sub-document compliance conclusions, evidence links, and conflict reports;

[0250] This is the core link of the review, and auditable means that each review conclusion must have clear and traceable evidence support, and counter-research means that the review process is not one-way to find supporting evidence, but actively and dialectically to find potential counter-evidence.

[0251] For example, for a sub-document, the system first retrieves relevant clauses from a large corpus of specification vectors containing a large number of design specifications, standards, then through a controlled abstraction process, the model is forced to extract structured constraint information (such as a certain parameter must be greater than 10 MPa) from the clauses, rather than generating subjective descriptions, then the system conducts compliance judgment on the sub-document content and these structured constraints, and draws a preliminary conclusion, and most importantly, the adversarial re-search: if the preliminary conclusion is in compliance, the system will conduct directional search to see if there are exemption or exception clauses; if the conclusion is not in compliance, it will actively search for whether there are remedial or special circumstances exemption clauses, and all judgment processes will generate evidence links containing specification sources, original text fragments and similarity scores.

[0252] This step effectively suppresses the illusion of the model by building a fully traceable calculation chain, and the unique adversarial review mechanism makes the review conclusion more reliable and withstands scrutiny, which is significantly different from the traditional retrieval augmented generation (RAG) method.

[0253] According to an aspect of the present application, the step S42 is further characterized in that:

[0254] Step S42a, constructing a structured specification abstraction object, comprising: limiting the specification abstraction object to contain only pre-set structured fields, and the structured fields include clause identification, mandatory level, condition predicate and quantitative constraint, and prohibiting the generation of subjective suggestions in the specification abstraction object;

[0255] Step S42b, packaging evidence links containing specification clause identification, comprising: for each compliance judgment, generating an evidence triple containing specification clause identification, sub-document fragment identification and similarity score between them, and establishing traceable association between the sub-document and the specification clause.

[0256] According to an aspect of the present application, the step S42b is further characterized in that:

[0257] Step S42b1, extracting the clause identification in the specification abstraction object as the first constituent element of the triple;

[0258] Step S42b2, locating the content fragment corresponding to the clause identification in the sub-document, and obtaining the fragment identification of the content fragment as the second constituent element of the triple;

[0259] Step S42b3, calculating the semantic similarity between the content fragment and the clause identification, and taking the quantitative score of the similarity as the third constituent element of the triple, to represent the strength of evidence association.

[0260] According to one aspect of the present application, the step S44 is further characterized in that:

[0261] Step S44a, when the preliminary compliance conclusion is compliance, searching for relevant exception clauses or mutually exclusive clauses in the norm vector library as counter-evidence;

[0262] Step S44b, when the preliminary compliance conclusion is non-compliance, actively retrieving possible applicable remediation clauses or exemption clauses in the norm vector library as counter-evidence;

[0263] Step S44c, and dynamically adjusting the similarity threshold and synonym set in the counter-retrieval process according to the gating scheduling data, generating a conflict report, and integrating the preliminary compliance conclusion, evidence link and conflict report to form the preliminary review conclusion set.

[0264] The system here plays the role of the devil's advocate, actively and strategically challenging its own preliminary conclusion.

[0265] The system initiates completely different counter-retrieval strategies according to the type of the preliminary compliance conclusion:

[0266] When the preliminary compliance conclusion is compliance, the system considers that the current evidence chain may be incomplete, at which point it will search for exception clauses or mutually exclusive clauses related to the current clause in the norm vector library as counter-evidence, for example, the design described in the sub-document meets the general safety specification, but the counter-retrieval finds a supplementary clause stating that in high seismic intensity areas, the general specification does not apply, and a special anti-seismic specification must be followed, while the report does not mention compliance with the anti-seismic specification, which constitutes a conflict.

[0267] When the preliminary compliance conclusion is non-compliance, the system attempts to provide constructive help, it will actively retrieve possible applicable remediation clauses or exemption clauses in the norm vector library as counter-evidence, for example, a report shows that a certain material index is not up to standard, but counter-retrieval finds a specification that allows the index to be appropriately relaxed under the premise of taking certain strengthening measures.

[0268] In addition, the similarity threshold and synonym set parameters in the counter-retrieval process will be dynamically adjusted according to the review strategy parameters generated in Example Three, achieving more precise control, all the counter-evidence clauses found will be recorded in the conflict report, and finally, the review results of each sub-document will integrate the preliminary conclusion, evidence link and conflict report to form a comprehensive micro review conclusion set. Step S4 can also be:

[0269] Step S4a, for generating traceable compliance review results based on the norm constraint graph and the fact evidence graph, specifically including the following sub-steps:

[0270] Step S4a1, read the sub-document set data, structured text data, logical structure tree data and pre-stored specification clause base library data output by the preceding steps, parse each specification clause based on the clause number, applicable scope, constraint condition, exception condition, relaxation condition and effectiveness level information in the specification clause base library data, extract the constraint condition nodes, exception condition nodes, relaxation condition nodes and inheritance relationship and mutual exclusion relationship between clauses, and construct a specification constraint graph data for representing the specification hierarchical structure and constraint relationship through the combination of nodes and edges;

[0271] Among them, each constraint condition node in the specification constraint graph data has a connection relationship with one or more specification clause numbers, which is used to backtrack to the corresponding specification clause text in subsequent steps;

[0272] Step S4a2, based on the sub-document set data and the structured text data, according to the chapter level information in the logical structure tree data, perform sentence-level and paragraph-level semantic recognition and element extraction on the sub-document content, identify engineering fact information such as unit parameter, hydraulic parameter, material parameter and operation condition parameter, associate these engineering fact information with the corresponding sub-document identification, and construct fact element set data containing fact nodes and text block nodes, and on this basis, establish edge connection relationship between each fact node and its source sub-document, reference source and derivation source, form fact evidence graph data for representing the support relationship and reference relationship between facts;

[0273] Among them, each fact node in the fact evidence graph data retains the mapping information with the original text segment, which is used to output traceable evidence path in subsequent steps;

[0274] Step S4a3, read the specification constraint graph data, fact evidence graph data and gating strategy parameter data obtained by the preceding gating step, according to the review depth requirements of different level specification clauses in the gating strategy parameter data, select the high effectiveness level specification clause nodes that need to be reviewed as starting nodes from the specification constraint graph data, and perform forward traversal along the mandatory constraint edges and inheritance relationship edges in the specification constraint graph, for each path from the high-level clause to the specific constraint condition node, match the fact nodes related to the constraint condition in the fact evidence graph data in turn, judge whether the corresponding fact satisfies the specification constraint, generate specification reasoning path set data for describing the forward satisfaction of each specification constraint chain, and at the same time, according to the matching result, mark the corresponding satisfaction conclusion, non-satisfaction conclusion or missing conclusion on each specification reasoning path, obtain specification compliance local conclusion data corresponding to each specification reasoning path one by one;

[0275] Each path in the specification reasoning path set data is associated with a specification clause node and a fact node participating in the judgment, used for subsequent steps of adversarial reasoning and evidence output;

[0276] In step S4a4, based on the specification reasoning path set data, the specification constraint graph data, the fact evidence graph data, and the gating strategy parameter data, for the specification reasoning paths determined to meet the conclusion in the specification compliance local conclusion data, starting from the corresponding constraint condition node, expanding along the edges representing exception conditions and edges representing relaxation conditions in the specification constraint graph, retrieving the specification clause nodes associated with these exception conditions and relaxation conditions, and querying the fact nodes matching these exception conditions or relaxation conditions in the fact evidence graph data to identify exception facts or relaxation facts that may negate the original conclusion meeting; for the specification reasoning paths determined not to meet the conclusion, starting from the corresponding constraint condition node, expanding along the edges representing remediation conditions or exemption conditions in the specification constraint graph, retrieving the fact nodes matching these remediation conditions or exemption conditions in the fact evidence graph data to identify remediation facts or exemption facts that may eliminate the non-conclusion meeting, through the above-mentioned adversarial expansion and fact matching performed on the conclusion meeting path and the non-conclusion meeting path respectively, generating adversarial reasoning path set data for recording the situation that each specification path positive conclusion is challenged or corrected;

[0277] For each path in the adversarial reasoning path set data, the corresponding initial specification reasoning path identifier and the exception clause node or remediation clause node participating in the confrontation are retained for subsequent comprehensive adjudication;

[0278] In step S4a5, based on the specification reasoning path set data, the adversarial reasoning path set data, and the specification compliance local conclusion data, for each specification reasoning path, first combine the specification clause nodes, fact nodes, and corresponding sub-document identifiers in the specification reasoning path in order to form a positive evidence segment sequence for describing the positive compliance judgment process, then add the exception clause nodes, remediation clause nodes, and supporting fact nodes involved in the adversarial paths in the adversarial reasoning path set data associated with the specification reasoning path to the same evidence sequence, and construct evidence path set data for recording the whole process of positive reasoning and adversarial reasoning.

[0279] On this basis, according to whether the adversarial reasoning finds effective exception facts or remediation facts, the initial conclusion in the specification compliance local conclusion data is modified to obtain the final compliance conclusion for each specification path, the final conclusions of each specification path are aggregated according to the sub-document dimension to generate compliance review conclusion data for subsequent cross-sub-document consistency adjudication, and the evidence path set data and the compliance review conclusion data are provided to the subsequent steps.

[0280] Step S4 can also be:

[0281] Step S4b, for introducing patterned search based on historical evidence path and risk adaptive control on the basis of performing specification constraint graph adversarial reasoning, to further improve the review intensity of high-risk specification path;

[0282] The specific steps of step S4b include the following sub-steps:

[0283] Step S4b1, on the basis of performing steps S4a1 to S4a5, from the evidence path set data and compliance review conclusion data already archived in the previous review task, extract high-risk specification paths that have ever appeared not to meet the conclusion or have been substantially revised by adversarial reasoning. Induce the specification clause node sequence, fact node pattern and common exception or remedial fact combination form involved in these specification paths, and construct a historical evidence path library data for depicting high-risk specification reasoning patterns;

[0284] Each path template in the historical evidence path library data contains a specification path pattern, a fact distribution pattern and a corresponding risk level label, which is used to guide the subsequent key review of new reports;

[0285] Step S4b2, read the specification constraint graph data, fact evidence graph data, historical evidence path library data and gating strategy parameter data. For the specification reasoning path set data formed by the current report, according to the specification clause node sequence and fact node distribution pattern in the path, match each specification reasoning path with the path template in the historical evidence path library data, to obtain the high-risk specification path candidate set data in the current report;

[0286] For the specification path matched to the high-risk template and with higher risk level, according to the configuration in the gating strategy parameter data, allocate higher adversarial reasoning budget to the sub-documents in the sub-document set data corresponding to the fact nodes of the specification path, including increasing the number of expandable exception clause layers, relaxing the similarity threshold for matching exception facts, and expanding the synonym set allowed in the search range, etc., so as to perform more in-depth counterargument retrieval on these high-risk paths in the subsequent adversarial reasoning process;

[0287] Step S4b3, based on the high-risk specification path candidate set data, the specification constraint graph data, the fact evidence graph data, and the adjusted gating strategy parameter data, re-performs the adversarial reasoning process on the high-risk specification paths, and for each high-risk specification path, prioritizes the search in the specification constraint graph according to the combination of the exception clause nodes and the remedy clause nodes recorded in the historical high-risk path templates, and performs directional matching on the fact nodes in the fact evidence graph corresponding to these clauses, so as to more concentratedly search for the adversarial paths prone to problems within the limited computing budget.

[0288] The path results obtained by the reinforcement adversarial reasoning are merged with the initial adversarial reasoning path set data in step S4a4 to generate updated adversarial reasoning path set data, and based on the updated adversarial reasoning path set data, the specification compliance local conclusion data is re-corrected to obtain a more stringent final compliance conclusion for the high-risk specification paths, and then the evidence path set data and the compliance review conclusion data are updated, and the updated evidence path set data and the compliance review conclusion data are provided for subsequent cross-sub-document consistency ruling steps and parameter logging steps. As shown in Figure 5 According to one aspect of the present application, the step S5 further comprises:

[0289] Step S51, in the aggregated preliminary review conclusion set, identify and enumerate potential conflicts across different sub-document conclusions, including numerical conflicts, reference conflicts, and mandatory level conflicts;

[0290] Before performing conflict ruling, the system will perform an important pre-step, i.e., term unification and dimension alignment.

[0291] Call the built Glossary term table, traverse all micro review conclusion sets, and unify the synonyms (such as machine set, generator) appearing therein into standard terms, and unify different units of measurement (such as MW, ten kilowatts) into international standard units.

[0292] On this basis, the system runs a conflict enumerator to identify the following types of potential conflicts in the conclusions of all sub-documents:

[0293] Numerical conflict: for example, sub-document A mentions the total installed capacity of the project as 500 MW, while sub-document B mentions the single machine capacity as 150 MW, and there are 4 units installed (i.e., the total capacity is 600 MW);

[0294] Reference conflict: for example, different sub-documents refer to two different versions of the same national standard (such as GB50201-2012 and GB50201-2018) when performing calculations;

[0295] Forced vs. Recommended: e.g. one conclusion is compliant according to a recommended provision, but another conclusion is non-compliant according to a mandatory provision.

[0296] For enumerated conflicts, the system loads a pre-defined rule base for automatic resolution, which defines an explicit priority sequence, preferably including:

[0297] Mandatory vs. Recommended: conclusions of mandatory provisions take precedence over conclusions of recommended provisions, and conclusions of recommended provisions take precedence over conclusions of reference provisions;

[0298] Versioning: conclusions of the latest version of a standard take precedence over conclusions of older versions of the standard;

[0299] Standard Hierarchy: conclusions of higher-level standards (e.g. national standards) take precedence over conclusions of lower-level standards (e.g. industry standards).

[0300] As an optional solution, when automatic resolution rules are insufficient to uniquely determine a conclusion, the system can trigger a human-in-the-loop interface to push the conflict details, relevant evidence chains, and rule suggestions to domain experts for final adjudication, and write the adjudication results and their reasons back to the system.

[0301] Step S52, set the resolution rules, and follow the priority sequence defined in the rules to resolve the potential conflicts, the priority sequence including that conclusions of mandatory provisions take precedence over conclusions of recommended provisions, and conclusions of the latest version of a standard take precedence over conclusions of older versions of the standard, to obtain a final review conclusion.

[0302] In order to realize the complete transparency and reproducibility of the review process, the present application designs a traceable parameter log Param_log throughout the whole process, which is not a simple program running log, but a structured digital archive recording all the information of key decision points in the whole review process.

[0303] Specifically, Param_log at least contains the following contents:

[0304] Model and version fingerprint: name and version hash value of the algorithm model used in each step (e.g. sentence vector model: bge-large-zh-v1.5, git_hash: xxxxx).

[0305] Specifically, the system loads the pre-trained BGE-large-zh-v1.5 model, the input of which is Chinese sentence text, and the output of which is a 1024-dimensional semantic vector, and for each sentence, the following inference steps are performed:

[0306] Tokenization and tokenization processing are performed on the input sentence;

[0307] inputting the token sequence into an encoder of the BGE model;

[0308] extracting a hidden state of a [CLS] position of a last layer of the model as a semantic vector of the sentence;

[0309] calculating a cosine similarity of the semantic vectors of adjacent sentences:

[0310] sim(s_i, s_{i+1}) = (v_i · v_{i+1}) / (||v_i|| × ||v_{i+1}||);

[0311] where v_i and v_{i+1} represent the semantic vectors of adjacent sentences s_i and s_{i+1}, respectively.

[0312] All threshold values, weights, and budget parameters: Score_struct calculated in hierarchical gating, calibrated threshold T; sim_threshold and bdry_threshold used in parallel blocking; threshold values and budget allocation dynamically adjusted in adversarial re-ranking.

[0313] Rules and ruling tracks: rule base versions relied on in conflict resolution and consistency resolution, and specific ruling processes and reasons for each conflict.

[0314] This Param_log is archived together with the original report, and its significance lies in that at any time, regulatory agencies, auditors, or users themselves can use this log to fully reproduce every detail of the review, understand why the system arrived at the final conclusion, which provides the strongest technical guarantee for the reliability, fairness, and traceability of the review results.

[0315] Specifically, the system aggregates the micro review conclusion set of all sub-documents, and conducts consistency checking based thereon, for example, checking whether there are numerical conflicts in the description of the same technical parameter in different sub-documents, or whether there are conflicts in the cited specification versions, for the conflicts found, the system will load the preset ruling rules (such as mandatory provisions are prior to recommended provisions, and new version standards are prior to old version standards) for automatic resolution, finally, the system will integrate the macro structural review conclusion (structure conformity score and missing item list) and the micro review conclusion after consistency ruling, and bind all related evidence links, to generate a final review report with clear structure, clear conclusion, and traceable evidence.

[0316] This step ensures the global consistency of the review conclusion and the integrity of the final report, and organically integrates the review results of each stage into a whole.

[0317] According to an aspect of the present application, it is characterized in that it further comprises the following steps:

[0318] Throughout the execution of the method, a traceable parameter log is synchronously recorded and archived to ensure end-to-end reproducibility of the review process;

[0319] The traceable parameter log at least includes:

[0320] The name and version fingerprint of the algorithm model used to execute each step;

[0321] All threshold values, weights and budget parameters set in hierarchical gating, parallel chunking and adversarial re-ranking;

[0322] And the rule version and decision track followed in conflict resolution and consistency resolution. The preferred embodiments of the present application are described in detail above, but the present application is not limited to the specific details of the above-described embodiments, and various equivalent transformations of the technical solutions of the present application can be made within the technical concept of the present application, and these equivalent transformations all belong to the protection scope of the present application.

Claims

1. A hierarchical review method for feasibility study reports of hydraulic machinery based on a large language model, characterized in that, Includes the following steps: Step S1: Collect and parse the original hydraulic machinery feasibility study report to generate a logical structure tree and structured text; Step S2: Based on the logical structure tree, perform hierarchical gating and policy scheduling to generate gating scheduling data; Step S3: Based on the review strategy parameters in the gating scheduling data, perform three-constraint parallel block division on the structured text to obtain a set of sub-documents; Step S4: Using the gating scheduling data, perform an auditable adversarial compliance review on each sub-document of the sub-document set to form a preliminary review conclusion set; Step S5: Aggregate the preliminary review conclusions set, perform cross-sub-document consistency adjudication and result summary, and generate the final review conclusions.

2. The method for hierarchical review of hydraulic machinery feasibility study reports based on a large language model as described in claim 1, characterized in that, Step S2 further comprises: Step S21: Construct a standard template library based on the hydraulic machinery industry standards. Compare the standard template library with the completeness and standardization of the logical structure tree from multiple preset dimensions, synthesize the structure compliance score, and identify the list of missing items. Step S22: Set the review threshold and determine whether the structural compliance score has reached the preset review threshold. If it has not been reached, terminate the subsequent process. Step S23: Extract the logical structure tree that reaches the preset review threshold, and map the structure compliance score and missing item list corresponding to the extracted logical structure tree to the core parameters and resource budget of the downstream review task to generate the review strategy parameters.

3. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 2, characterized in that, Step S21 further comprises: Step S21a: Check the coverage rate of the logical structure tree to the required chapters in the standard template library and obtain the coverage score; Step S21b: Examine the chapter hierarchy of the logical structure tree and determine the hierarchy correctness score based on the structural violations; Step S21c: Measure the difference between the chapter order of the logical structure tree and the standard template library, and calculate the order consistency score; Step S21d: Traverse the contents of key chapters in the logical structure tree, evaluate the completeness rate of their internal technical elements, and obtain the element completeness score. Step S21e: The coverage score, hierarchical correctness score, sequence consistency score, and element completeness score are weighted and summed to generate the final structure conformity score. The required items that do not appear in the structure tree, key sub-items that are not effectively covered due to hierarchical misclassification, and missing elements that do not conform to industry standard descriptions are marked and summarized to form a missing item list.

4. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 2, characterized in that, Step S22 further comprises: Step S22a: Collect historical hydraulic machinery feasibility study reports and construct a sample set containing the structural compliance scores of historical hydraulic machinery feasibility study reports and their corresponding two types of manual annotation labels. The two types of manual annotation labels are binary qualitative judgment labels for whether the logical structure tree of the historical hydraulic machinery feasibility study reports meets the requirements of actual application. Step S22b: Based on the sample set, generate a subject operating characteristic curve, which represents the relationship between sensitivity and specificity at different scoring thresholds; Step S22c: Optimize the Youden index on the subject operating characteristic curve to maximize the index, and determine the score value corresponding to the maximum point of the index as the review threshold.

5. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 2, characterized in that, Step S23 further comprises: Step S23a: Based on the contents of the missing item list, adjust the penalty weight of the subsequent document segmentation step and the reordering weight of the standardized retrieval step; Step S23b: Based on the score range of the structure conformity score, set the semantic similarity threshold in the document segmentation step, the retrieval depth in the standard retrieval step, and the budget allocation for the adversarial anti-retrieval step. Step S23c: Based on the review priority of hydraulic machinery, assign priority weights according to the type of missing items and the scoring range, and verify and correct parameter conflicts to generate the review strategy parameters.

6. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 1, characterized in that, Step S3 further comprises: Step S31: Calculate the semantic similarity between adjacent sentences using the cosine similarity method. Based on the semantic similarity between adjacent sentences and the similarity threshold in the review strategy parameters, identify semantic breakpoints in the structured text. Step S32: Filter chapter boundary points based on semantic breakpoints in the structured text, evaluate the credibility of chapter boundary points in the structured text, and filter out valid boundary points; Step S33: Consider the semantic breakpoint, the effective boundary point, and the upper limit of the input length set by the review strategy parameters in parallel. When there is a constraint conflict among the three, execute the conflict resolution to determine the final split point. Step S34: Apply the final segmentation point to the structured text to generate the sub-document set.

7. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 6, characterized in that, Step S33 further comprises: Step S33a: Follow the preset priority sequence. In this sequence, the priority of satisfying the valid boundary point constraint is the highest, the priority of satisfying the upper limit of input length constraint is the second highest, and the priority of satisfying the semantic breakpoint constraint is the lowest. Identify the special content of the table of contents and chart blocks of the structured text, and apply exception rules to the special content areas of the table of contents or chart blocks to preserve the overall structural integrity. Step S33b: Construct an objective function to evaluate the segmentation quality. The objective function is to minimize the deviation between the segmentation result and the ideal structured text. The constraints include: numbering hierarchy logical constraints, chapter length valid constraints, segmentation point uniqueness constraints, professional terminology consistency constraints, format feature consistency constraints, and cross-document consistency constraints. The optimal final segmentation point is determined by minimizing the function value of the objective function through optimization.

8. The method for hierarchical review of hydraulic machinery feasibility study reports based on a large language model as described in claim 1, characterized in that, Step S4 further comprises: Step S41: Collect hydraulic machinery technical documents, construct a specification vector library based on hydraulic machinery industry standards and hydraulic machinery technical documents, retrieve from the specification vector library and perform weighted sorting on the recall results according to the list of missing items in the gated scheduling data, and select a set of candidate clauses for the sub-document; Step S42: Set field templates, extract compliance constraint information from the candidate clause set according to the field templates, and construct a structured specification extraction object; Step S43: Compare the sub-document with the extracted specification object, determine the preliminary compliance conclusion, and encapsulate the evidence link containing the specification clause identifier; Step S44: Based on the preliminary compliance conclusion, initiate a differentiated adversarial counter-search to search for counter-evidence clauses, generate a conflict report, and integrate the preliminary compliance conclusion, evidence links, and conflict report to form the preliminary review conclusion set.

9. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 8, characterized in that, Step S42 further comprises: Step S42a: Constructing a structured specification extraction object, including: limiting the specification extraction object to contain only preset structured fields, the structured fields including clause identifiers, mandatory levels, conditional predicates and quantitative constraints, and prohibiting the generation of subjective suggestions in the specification extraction object; Step S42b: Encapsulate evidence links containing regulatory clause identifiers, including: for each compliance determination, generate an evidence triple containing a regulatory clause identifier, a sub-document fragment identifier, and a similarity score between the two, and establish a traceable association between the sub-document and the regulatory clause.

10. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 9, characterized in that, Step S42b further comprises: Step S42b1: Extract the clause identifier from the specification extraction object and use it as the first constituent element of the triple; Step S42b2: Locate the content segment in the sub-document that corresponds to the clause identifier, and obtain the segment identifier of the content segment as the second constituent element of the triple; Step S42b3: Calculate the semantic similarity between the content fragment and the clause identifier, and use the quantified score of the similarity as the third component of the triple to characterize the strength of the evidence association.

11. The hierarchical review method for hydraulic machinery feasibility study reports based on a large language model as described in claim 8, characterized in that, Step S44 further comprises: Step S44a: When the preliminary compliance conclusion is compliant, conduct a targeted search in the specification vector library for relevant exception clauses or mutually exclusive clauses as counter-evidence; Step S44b: When the preliminary compliance conclusion is non-compliance, actively search the normative vector library for applicable remedies or exemptions as counter-evidence. Step S44c: Based on the gating scheduling data, dynamically adjust the similarity threshold and synonym set in the reverse retrieval process, generate a conflict report, and integrate the preliminary compliance conclusion, evidence link, and conflict report to form the preliminary review conclusion set.

12. The method for hierarchical review of hydraulic machinery feasibility study reports based on a large language model as described in claim 1, characterized in that, Step S5 further comprises: Step S51: In the aggregated set of preliminary review conclusions, identify and enumerate potential conflicts across conclusions in different sub-documents, including numerical conflicts, citation conflicts, and mandatory level conflicts. Step S52: Set the adjudication rules and follow the priority sequence defined in the rules to resolve the potential conflict. The priority sequence includes that the conclusion of mandatory clauses takes precedence over the conclusion of recommended clauses, and that the conclusion of the latest version of the standard takes precedence over the conclusion of the old version of the standard, to obtain the final review conclusion.

13. The method for hierarchical review of hydraulic machinery feasibility study reports based on a large language model as described in claim 1, characterized in that, It also includes the following steps: Throughout the entire execution process of the method, a traceable parameter log is recorded and archived synchronously; The traceable parameter log contains at least: The name and version fingerprint of the algorithm model used in each step; All thresholds, weights, and budget parameters set in hierarchical gating, parallel block partitioning, and adversarial reverse retrieval; And the version of rules and the trajectory of decisions based on conflict and consistency decisions.

Images