Text processing method and system based on large language model

Abstract

This invention discloses a text processing method and system based on a large language model. The method involves acquiring an electronic document to be processed, inputting it into a document type recognition model built on the large language model, and outputting document type labels. Based on the document type labels, a detection logic generator built on the large language model is invoked to dynamically generate a corresponding set of detection logic. The electronic document and the set of detection logic are input into a multi-level inference engine built on the large language model, where the detection logic is executed sequentially, extracting evidence information, constructing inference paths, and calculating confidence scores to generate single-item detection results with evidence chains and confidence scores. All results are summarized to generate a hierarchical detection report. In response to user interaction with the detection items, a dialogue interpreter built on the large language model outputs natural language explanation information. This invention achieves adaptive and interpretable intelligent text detection.

Description

Technical Field

[0001] This invention relates to the field of electronic document processing technology, specifically to a text processing method and system based on a large language model. Background Technology

[0002] As important information assets, electronic records require rigorous quality control to safeguard their evidentiary value. Current detection methods generally rely on manually pre-set static rule bases for specific document types, with the system simply comparing document content against these fixed rules. However, electronic records are diverse in type and complex in structure, making it difficult for static rules to comprehensively cover all aspects, thus limiting the adaptability and accuracy of the detection. More importantly, this rule-matching-based detection process lacks transparency and interpretability. Users only receive a "pass" or "fail" conclusion, unable to understand the reasoning behind the conclusion, making it difficult to effectively verify and trace questionable results. Summary of the Invention

[0003] In view of the above problems, the present invention provides a text processing method and system based on a large language model, which achieves adaptive and traceable intelligent detection by dynamically generating detection logic that matches document features and outputting interpretable reasoning evidence.

[0004] To achieve the above objectives, in a first aspect, this application provides a text processing method based on a large language model, comprising:

[0005] Obtain the electronic archive file to be processed, which contains metadata information and electronic file information;

[0006] The electronic archive files are input into the document type recognition model, which is built based on a large language model and outputs the document type labels of the electronic archive files.

[0007] Based on document type tags, a detection logic generator is invoked. The detection logic generator is built on a large language model and dynamically generates a set of detection logics corresponding to the electronic archive file according to the document type tags. The set of detection logics includes accuracy detection logic, integrity detection logic, usability detection logic, and security detection logic.

[0008] The electronic archive files and the set of detection logics are input into the multi-level inference engine. The multi-level inference engine is built based on a large language model and executes each detection logic in the set of detection logics in sequence. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs inference paths and calculates confidence scores, and generates a single detection result with evidence chains and confidence scores.

[0009] Summarize all individual test results to generate a hierarchical test report. The hierarchical test report includes a problem description, evidence chain information, confidence score and problem level for each test item.

[0010] In response to the user's interaction with any detection item in the hierarchical detection report, a follow-up question is generated. The follow-up question and the corresponding evidence chain information are input into the dialogue interpreter. The dialogue interpreter is built based on a large language model and outputs natural language explanation information for the detection item. The natural language explanation information includes the detection basis, reasoning process and confidence basis.

[0011] In some embodiments, electronic archive files are input into a document type recognition model, which is built based on a large language model and outputs document type labels for the electronic archive files, including:

[0012] Extract the header feature information of electronic archive files. The header feature information includes file format identifier, metadata field structure, and a summary of the header content.

[0013] The file header feature information is input into the type preclassifier, which is built based on the rule engine. It performs a preliminary match based on the file format identifier and the metadata field structure, and outputs a set of candidate types.

[0014] Extract the full-text content information of electronic archive files. The full-text content information includes the sequence of chapter titles, the distribution of key terms, paragraph structure features, and data table features.

[0015] The full text content information and the candidate type set are input into the semantic type discriminator. The semantic type discriminator is built based on a large language model, performs semantic understanding of the full text content information, performs accurate matching in the candidate type set, and outputs the document type label of the electronic archive file. The document type label includes a primary category label and a secondary category label. The primary category label indicates the business domain category of the electronic archive, and the secondary category label indicates the standard document template category of the electronic archive.

[0016] The document type label is input into the type confidence evaluator, which is built based on a large language model. The type confidence evaluator calculates the type matching confidence for the primary and secondary classification labels respectively. If the type matching confidence is lower than the preset threshold, the type review process is triggered.

[0017] The type review process includes:

[0018] Extract the associated file information of the electronic archive file. The associated file information includes a list of archived files in the same batch and historical archive records. Input the associated file information and the full text content information together into the semantic type discriminator for secondary identification, and output the corrected document type label.

[0019] In some embodiments, a set of detection logic corresponding to electronic archive files is dynamically generated based on document type tags, including:

[0020] Read the business domain category tag and the document template category tag from the document type tags;

[0021] Input the business domain category tags and document template category tags into the detection logic template library, and retrieve the basic detection logic template that matches the document template category tags. The basic detection logic template contains detection item placeholders and detection rule placeholders.

[0022] Extract document content feature information from electronic archive files. Document content feature information includes document structure features and document data features.

[0023] The document content feature information and the basic detection logic template are input into the detection logic instantiator. Based on the document content feature information, the detection item placeholders and detection rule placeholders in the basic detection logic template are filled in, and the instantiated detection logic set is output.

[0024] The instantiated detection logic set is input into the detection logic validator, which performs a logic validity check on the instantiated detection logic set and outputs the final detection logic set.

[0025] In some embodiments, document content feature information and a basic detection logic template are input to the detection logic instantiator. Based on the document content feature information, the detection item placeholders and detection rule placeholders in the basic detection logic template are filled in, and an instantiated detection logic set is output, including:

[0026] The document structure features in the document content feature information are input into the detection item mapper. The detection item mapper is built based on the large language model and performs semantic parsing on the document structure features. The document structure features include chapter title sequence and paragraph structure features. The detection item mapper identifies the functional semantics of each chapter in the chapter title sequence and maps the detection item placeholders in the basic detection logic template to specific detection item descriptions that match the functional semantics of each chapter, thus completing the semantic filling of the detection item placeholders.

[0027] The document data features from the document content feature information are input into the rule parameter generator. The rule parameter generator is built based on a large language model and performs numerical distribution analysis and term frequency analysis on the document data features. The document data features include key term distribution and data table features. The rule parameter generator determines a reasonable range of detection thresholds based on the numerical distribution analysis results and determines the matching mode of detection conditions based on the term frequency analysis results. It replaces the detection rule placeholders in the basic detection logic template with specific detection rule descriptions that include detection thresholds and detection conditions, thus completing the parameterized filling of the detection rule placeholders.

[0028] The descriptions of detection items that have completed semantic filling are associated and paired with the descriptions of detection rules that have completed parameterized filling to generate instantiated detection logic. All instantiated detection logic is then aggregated and a set of instantiated detection logic is output.

[0029] In some embodiments, each detection logic in the detection logic set is executed sequentially. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs an inference path, and calculates a confidence score, generating a single detection result with an evidence chain and a confidence score, including:

[0030] Read the detection logic to be executed from the detection logic set. The detection logic to be executed includes the detection item description, the detection rule description, and the detection target data range.

[0031] Based on the range of the target data, locate and extract the original data fragments from the electronic archive file;

[0032] The original data fragments and the detection rule descriptions are input into the evidence extractor. The evidence extractor is built based on a large language model and performs semantic parsing on the original data fragments according to the detection rule descriptions to extract key evidence information related to the detection rule descriptions from the original data fragments.

[0033] Key evidence information is input into the reasoning path builder, which is built based on a large language model. The reasoning path builder performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to a preset reasoning paradigm. The multi-step reasoning path contains multiple reasoning nodes, and adjacent reasoning nodes are connected through causal relationships.

[0034] Input the multi-step inference path into the confidence calculator, which is built based on a large language model. The calculator evaluates the confidence of the inference nodes in the multi-step inference path and generates the overall confidence score of the detection logic to be executed.

[0035] Key evidence information, multi-step reasoning paths, and overall confidence scores are structured and encapsulated to generate individual detection results. Each individual detection result includes a problem description field, an evidence chain field, a reasoning path field, a confidence score field, and a problem level field.

[0036] In some embodiments, key evidence information is input into a reasoning path builder, which is built based on a large language model. The reasoning path builder performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to a preset reasoning paradigm, including:

[0037] Key evidence information is input into the evidence relationship analyzer, which is built on a large language model. It performs pairwise relationship analysis on numerical evidence, textual evidence, structural evidence and related evidence in the key evidence information, identifies the supporting, contradictory and complementary relationships between various types of evidence, and outputs an evidence relationship graph. The evidence relationship graph is organized with evidence as nodes and relationship type as edges.

[0038] The evidence relationship graph is input into the inference starting point locator, which is built based on a large language model. The inference starting point locator identifies independent evidence nodes that do not have prior supporting evidence as initial evidence nodes in the evidence relationship graph, and identifies core evidence nodes that have the most subsequent supporting evidence as key inference nodes. It outputs the inference starting point set and the inference key node set.

[0039] The inference starting point set, the inference key node set, and the evidence relationship graph are input into the path generator. The path generator is built based on the large language model. It takes the initial evidence node in the inference starting point set as the starting point and the key inference node in the inference key node set as the necessary path point. It performs path search in the evidence relationship graph to generate multiple candidate inference paths from the initial evidence node through the key inference node to the final conclusion node.

[0040] Multiple candidate reasoning paths are input into the path optimizer, which is built based on a large language model. The path optimizer evaluates the path length, logical rigor, and evidence coverage of the multiple candidate reasoning paths. Based on the comprehensive score of path length, logical rigor, and evidence coverage, the optimal reasoning path is selected from the multiple candidate reasoning paths. The optimal reasoning path includes an initial evidence node, intermediate reasoning nodes, and a final conclusion node. Each reasoning node includes input evidence, reasoning rules, and output conclusion.

[0041] The optimal reasoning path is output as a multi-step reasoning path.

[0042] In some embodiments, all individual test results are summarized to generate a hierarchical test report, including:

[0043] Read all individual test results. Each individual test result includes a problem description field, a chain of evidence field, an inference path field, a confidence score field, and a problem level field.

[0044] Individual test results are grouped into primary groups based on the problem level field. The primary groups include the serious problem group, the general problem group, and the warning problem group. Individual test results within the same problem level group are sorted from high to low based on the confidence score field.

[0045] For the individual test results within each issue level group in the first-level group, secondary grouping is performed according to the test logic type. The test logic types include accuracy test logic, integrity test logic, usability test logic, and security test logic, forming a two-dimensional grouping matrix with issue level as the row and test logic type as the column.

[0046] The two-dimensional grouping matrix is ​​input into the report summary generator, which is built based on a large language model. It performs semantic aggregation on the single detection results in each cell of the two-dimensional grouping matrix, extracts the common features and core issues in each cell, and generates cell summary information. At the same time, it performs cross-cell semantic association analysis on the cell summary information of all cells, identifies the causal relationship and logical progression relationship between different cells, and generates cross-domain association summary information.

[0047] The cell summary information and cross-domain association summary information are input into the report structure arranger. The report structure arranger is built based on a large language model and arranges the cell summary information and cross-domain association summary information in a structured manner according to three levels: overall summary, category details, and repair suggestions. The overall summary level includes the overall detection conclusion of the electronic file, the total number of problems, and the core risk warning. The category details level includes a detailed table of detection results with a two-dimensional grouping matrix as the skeleton. The repair suggestions level includes the repair priority ranking and repair operation description for each detection item.

[0048] The output of the report structure arranger is linked and bound to the complete information of all individual test results to generate a hierarchical test report. The hierarchical test report supports layered viewing and interactive navigation from the overall summary level down to the category detail table level and then to the individual test result level.

[0049] In some embodiments, the follow-up question and the corresponding chain of evidence information are input into the dialogue interpreter, which is built based on a large language model and outputs natural language explanation information for the detected item, including:

[0050] Receive follow-up questions initiated by the user. The follow-up questions include the natural language question entered by the user and the identifier of the detection item currently focused by the user.

[0051] Based on the test item identifier, locate and extract the individual test results corresponding to the test item identifier from the hierarchical test report;

[0052] The natural language question in the follow-up request is input into the intent parser. The intent parser is built based on a large language model, performs semantic parsing and intent classification on the natural language question, and outputs the follow-up intent label.

[0053] Based on the follow-up intent label, extract the target information fragment that matches the follow-up intent label from the single detection results;

[0054] The target information fragment and the natural language question are input into the explanation text generator. The explanation text generator is built based on a large language model and responds to the natural language question in a targeted manner according to the target information fragment, generating a preliminary explanation text.

[0055] The initial explanation text is input into the explanation depth controller, which is built based on a large language model. The explanation depth controller evaluates the explanation quality of the initial explanation text, makes adaptive adjustments to the initial explanation text based on the evaluation results, and outputs the final natural language explanation information.

[0056] The final natural language interpretation information is associated and bound with the evidence chain field in the individual detection results to generate an interactive interpretation view.

[0057] In some embodiments, a preliminary explanation text is input to an explanation depth controller, which is built based on a large language model. The explanation depth controller evaluates the explanation quality of the preliminary explanation text, adaptively adjusts the preliminary explanation text based on the evaluation results, and outputs the final natural language explanation information, including:

[0058] The initial explanation text is input into the readability evaluator, which is built based on a large language model. The evaluator performs sentence complexity analysis and term density analysis on the initial explanation text and generates a readability score based on the analysis results.

[0059] The preliminary explanation text is input into the information integrity evaluator, which is built based on a large language model. The evaluator performs semantic alignment comparison between the preliminary explanation text and the target information fragment, identifies the missing key evidence nodes and missing reasoning steps, and generates an information integrity score based on the identification results.

[0060] Readability scores and information integrity scores are input into an adjustment strategy decision-maker. This decision-maker, built on a large language model, selects the appropriate adjustment strategy based on the combination of the readability and information integrity scores. If the readability score is below a first preset threshold and the information integrity score is above a second preset threshold, a simplification adjustment strategy is selected. This includes splitting complex sentences, converting passive voice, and adding annotations to technical terms. If the readability score is above the first preset threshold and the information integrity score is below the second preset threshold, a supplementary adjustment strategy is selected. This includes inserting missing evidence node descriptions and supplementing intermediate conclusions in the reasoning steps. If both the readability score and the information integrity score are below the first and second preset thresholds, a reconstruction adjustment strategy is selected. This includes reorganizing the narrative structure of the interpreted text.

[0061] The adjustment strategy output by the adjustment strategy decision-maker and the initial explanation text are input into the text adjustment executor. The text adjustment executor is built based on the large language model, performs the corresponding adjustment operation on the initial explanation text according to the adjustment strategy, and outputs the final natural language explanation information.

[0062] In a second aspect, the present invention also provides a text processing system based on a large language model, applicable to the method described in the first aspect. The system includes a document type recognition module, a detection logic generation module, a multi-level inference engine module, a report generation module, and a dialogue explanation module. The document type recognition module has a built-in document type recognition model, which is constructed based on a large language model and is used to receive electronic archive files and output document type tags for the electronic archive files. The detection logic generation module is connected to the document type recognition module and has a built-in detection logic generator, which is constructed based on a large language model and is used to dynamically generate a set of detection logics corresponding to the electronic archive files based on the document type tags. The multi-level inference engine module is connected to both the detection logic generation module and the data storage terminal of the electronic archive files. The system includes a built-in multi-level inference engine, constructed based on a large language model. This engine sequentially executes each detection logic in the detection logic set, extracting evidence information, constructing inference paths, and calculating confidence scores during the execution of each logic, generating individual detection results with evidence chains and confidence scores. A report generation module connects to the multi-level inference engine module. This module includes a built-in report summary generator and report structure arranger, both also based on the large language model. These summarize all individual detection results and generate a hierarchical detection report. A dialogue explanation module connects to the report generation module. This module includes a built-in dialogue explainer, also based on the large language model. This explainer receives user requests for follow-up questions on any detection item in the hierarchical detection report and outputs natural language explanations for those items.

[0063] Unlike existing technologies, the above technical solution provides a text processing method and system based on a large language model. It acquires the electronic document to be processed, inputs it into a document type recognition model built on a large language model, and outputs document type labels. Based on the document type labels, it calls a detection logic generator built on a large language model to dynamically generate a corresponding set of detection logic. The electronic document and the set of detection logic are input into a multi-level inference engine built on a large language model, which executes the detection logic sequentially, extracts evidence information, constructs inference paths, and calculates confidence scores, generating single-item detection results with evidence chains and confidence scores. All results are summarized to generate a hierarchical detection report. In response to user interaction with the detection items, a dialogue interpreter built on a large language model outputs natural language explanation information. This invention achieves adaptive and interpretable intelligent text detection.

[0064] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0065] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0066] In the accompanying drawings of the instruction manual:

[0067] Figure 1 This is a schematic diagram illustrating the multi-dimensional detection process of electronic archive documents as described in a specific implementation method;

[0068] Figure 2 This is a flowchart illustrating the multi-path data processing for electronic document information authenticity detection in a specific implementation method.

[0069] Figure 3 This is a schematic diagram illustrating steps S101 to S105 of the method described in the specific implementation embodiment;

[0070] Figure 4 This is a schematic diagram illustrating steps S201 to S206 of the method described in a specific implementation.

[0071] Figure 5 This is a schematic diagram of the structure of the text processing system described in a specific embodiment.

[0072] The reference numerals used in the above figures are explained as follows:

[0073] 1. Word processing system;

[0074] 11. Document type recognition module;

[0075] 12. Detection logic generation module;

[0076] 13. Multi-level inference engine module;

[0077] 14. Report generation module;

[0078] 15. Dialogue Explanation Module. Detailed Implementation

[0079] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0080] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0081] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0082] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0083] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0084] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0085] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0086] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0087] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0088] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0089] Please see Figures 1 to 2 In a first aspect, this embodiment provides a text processing method based on a large language model, including:

[0090] Obtain the electronic archive file to be processed, which contains metadata information and electronic file information;

[0091] The electronic archive files are input into the document type recognition model, which is built based on a large language model and outputs the document type labels of the electronic archive files.

[0092] Based on document type tags, a detection logic generator is invoked. The detection logic generator is built on a large language model and dynamically generates a set of detection logics corresponding to the electronic archive file according to the document type tags. The set of detection logics includes accuracy detection logic, integrity detection logic, usability detection logic, and security detection logic.

[0093] The electronic archive files and the set of detection logics are input into the multi-level inference engine. The multi-level inference engine is built based on a large language model and executes each detection logic in the set of detection logics in sequence. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs inference paths and calculates confidence scores, and generates a single detection result with evidence chains and confidence scores.

[0094] Summarize all individual test results to generate a hierarchical test report. The hierarchical test report includes a problem description, evidence chain information, confidence score and problem level for each test item.

[0095] In response to the user's interaction with any detection item in the hierarchical detection report, a follow-up question is generated. The follow-up question and the corresponding evidence chain information are input into the dialogue interpreter. The dialogue interpreter is built based on a large language model and outputs natural language explanation information for the detection item. The natural language explanation information includes the detection basis, reasoning process and confidence basis.

[0096] In this embodiment, electronic archival files refer to digitized files to be tested, obtained from an archival management system or data acquisition interface. The metadata information contained therein is associated with the electronic file information through a unified identifier. The metadata information records the creation background and management attributes of the archive in the form of structured fields, while the electronic file information contains the main content data of the archive. Together, they form the input basis for subsequent testing.

[0097] The document type recognition model is built upon a large language model. Its construction process involves using massive amounts of electronic archival text data from different business domains to fine-tune the large language model for domain adaptation, enabling it to recognize the structural features and semantic themes of archival documents. The model's training objective is to output labels indicating the business domain category and the standard document template category based on the document's chapter title sequence, key term distribution, and paragraph structure, thereby achieving type discrimination based on content semantics rather than filename.

[0098] The detection logic generator is built upon a large language model. Its construction process involves fine-tuning the large language model using paired data containing detection logic descriptions and corresponding document types. This allows it to autonomously generate detection logic descriptions across four dimensions: accuracy, completeness, usability, and security, based on the input document type tags. Accuracy detection logic focuses on verifying the consistency of data content; completeness detection logic checks the completeness of necessary elements; usability detection logic verifies the readability and accessibility of the file format; and security detection logic focuses on detecting the confidentiality and tamper-proof nature of information. The detection logic generator maps document types to specific detection requirements through semantic understanding, replacing the traditional method of manually configuring rules.

[0099] The multi-level inference engine is built upon a large language model. Its construction process involves fine-tuning the large language model using training data containing detection logic, document fragments, and corresponding inference paths. This enables the model to locate evidence from documents, construct multi-step inference paths, and evaluate the confidence level of conclusions. When executing each detection logic, the multi-level inference engine first locates the original data fragments related to the detection item in the electronic document as the source of evidence. Then, it constructs a logical connection from evidence to conclusion through causal relationships. Finally, it calculates a confidence score based on the sufficiency of the evidence and the rigor of the logic, ultimately outputting a single detection result containing a complete chain of evidence.

[0100] The hierarchical testing report organizes individual test results into layers according to problem level, which is determined based on a combination of confidence score and problem severity. The report includes a problem description, evidence chain information, confidence score, and problem level for each test item. Users can obtain core risk alerts from an overall overview, or drill down to specific test items level by level to view the complete evidence chain and reasoning process.

[0101] The dialogue interpreter is built upon a large language model. Its construction process involves fine-tuning the large language model using paired data including user follow-up questions, corresponding evidence chains, and the desired explanation text. This enables it to generate natural language explanations based on the user's follow-up intent and the evidence chain information. The detection criteria explain the norms or rules referenced in the judgment, the reasoning process demonstrates the logical derivation steps from evidence to conclusion, and the confidence level is based on the reliability of the explanation scoring results. This allows for in-depth questioning by users regarding questionable detection items and targeted responses from the system.

[0102] Combination Figure 1As shown, the multi-dimensional inspection process for electronic archival documents can be broadly divided into three stages: inspection plan formulation, inspection data preparation, and inspection report generation. In the inspection plan formulation stage, the system establishes a set of inspection logics, including accuracy inspection logic, integrity inspection logic, usability inspection logic, and security inspection logic. Specifically, the accuracy inspection logic inspects both metadata and electronic document information. Metadata inspection items include length detection, type detection, value range detection, valid value range detection, special character detection, file number standardization detection, and duplication detection. Electronic document information inspection items include source authenticity detection, content authenticity detection, and correlation consistency detection. In the integrity inspection logic, metadata inspection items include mandatory field detection and continuity detection, while electronic document information inspection items include content integrity detection. In the usability inspection logic, metadata inspection items include metadata usability detection, software environment information detection, and hardware environment information detection, while electronic document information inspection items include file format detection. In the security inspection logic, metadata inspection items include checking whether the title contains classified words, phrases, or sentences, while electronic document information inspection items include top-secret archival document detection. In the data preparation stage, the system prepares the data to be tested by selecting view data and importing archived files; finally, in the report generation stage, the system executes the test and generates an analysis report, which users can use to query historical tests and view test results.

[0103] Combination Figure 2 As shown, when performing accuracy detection logic on electronic document information, one core detection step is source authenticity detection. This detection is achieved through multi-path data processing: on the one hand, the electronic document information to be detected undergoes a business processing process, generating processed electronic document information; on the other hand, this electronic document information is processed by an algorithm to generate an algorithm processing value. This algorithm processing value can be a digital digest calculated using a hash algorithm on the content of the electronic document information, used to uniquely identify the original state of the document content. The processed electronic document information and the algorithm processing value are used together as input to enter the authenticity detection execution step. By comparing whether the algorithm processing value is consistent with the original value stored in the metadata, and whether the processed content meets expectations, a comprehensive verification of the source authenticity and content authenticity of the electronic document information is completed.

[0104] In this embodiment, the document type recognition model outputs type labels based on the semantics of the document content. The detection logic generator dynamically generates a detection scheme based on the type labels. The multi-level inference engine performs the detection and outputs an interpretable result with evidence chains and confidence scores. The dialogue interpreter responds to user follow-up questions and provides natural language explanations based on evidence chains. This embodiment achieves adaptive, interpretable, and interactive intelligent text detection by integrating a large language model throughout the entire process of document type recognition, detection logic generation, multi-level inference, and dialogue interpretation.

[0105] In some embodiments, electronic archive files are input into a document type recognition model, which is built based on a large language model and outputs document type labels for the electronic archive files, including:

[0106] Extract the header feature information of electronic archive files. The header feature information includes file format identifier, metadata field structure, and a summary of the header content.

[0107] The file header feature information is input into the type preclassifier, which is built based on the rule engine. It performs a preliminary match based on the file format identifier and the metadata field structure, and outputs a set of candidate types.

[0108] Extract the full-text content information of electronic archive files. The full-text content information includes the sequence of chapter titles, the distribution of key terms, paragraph structure features, and data table features.

[0109] The full text content information and the candidate type set are input into the semantic type discriminator. The semantic type discriminator is built based on a large language model, performs semantic understanding of the full text content information, performs accurate matching in the candidate type set, and outputs the document type label of the electronic archive file. The document type label includes a primary category label and a secondary category label. The primary category label indicates the business domain category of the electronic archive, and the secondary category label indicates the standard document template category of the electronic archive.

[0110] The document type label is input into the type confidence evaluator, which is built based on a large language model. The type confidence evaluator calculates the type matching confidence for the primary and secondary classification labels respectively. If the type matching confidence is lower than the preset threshold, the type review process is triggered.

[0111] The type review process includes:

[0112] Extract the associated file information of the electronic archive file. The associated file information includes a list of archived files in the same batch and historical archive records. Input the associated file information and the full text content information together into the semantic type discriminator for secondary identification, and output the corrected document type label.

[0113] In this embodiment, the file header feature information is structured data read from the storage area of ​​the beginning of the electronic archive file, used for subsequent rapid type prediction. Specifically, the file format identifier is obtained by parsing the binary header of the file or the format declaration field in the file header; the metadata field structure is obtained by reading the field names, field types, and field order defined in the file header; and the file header content summary is obtained by extracting the first few characters or key descriptive statements from the beginning paragraph of the file.

[0114] The rule engine is built from a set of predefined matching rules. Each rule describes the document type prediction conditions corresponding to the combination of file format identifiers and metadata field structures. Based on the input file format identifiers and metadata field structures, the rule engine matches the conditional statements in the rule base one by one, adding document type identifiers that meet the conditions to the candidate type set. The rule base is constructed from summaries of common file format and document type correspondences by archival experts, as well as statistical analyses of the correlation between file formats and types in historical archived data. The type pre-classifier quickly filters the input file header feature information through conditional matching logic, outputting a candidate type set that meets the matching conditions. This set contains multiple possible document type identifiers to narrow down the search scope for subsequent precise identification.

[0115] The full-text content information is a set of semantic features extracted after parsing the main text of the electronic archive file. Among them, the chapter title sequence is obtained by hierarchical segmentation of the document structure, the distribution of key terms is determined by domain feature word matching and word frequency statistics, the paragraph structure features are obtained by analyzing paragraph length and numbering format, and the data table features are obtained by detecting table areas in the document and extracting cell content.

[0116] The semantic type discriminator is trained using a large amount of electronic archive text data labeled with business domain categories and document template categories for fine-tuning. The training objective is to minimize the difference between predicted labels and true labels. During inference, the semantic type discriminator uses the candidate type set as a constraint to perform semantic understanding of the full text content. By analyzing the document's topic distribution, structural features, and language habits, it matches the type label combination in the candidate type set that is semantically closest to the document content.

[0117] Type matching confidence is determined by analyzing the semantic consistency between the document content and the labeled type tags, and is output as a numerical value between 0 and 1. The preset threshold can be determined based on the accuracy distribution of historical recognition results. When the type matching confidence is lower than the preset threshold, it indicates that the initial recognition result may be unreliable, triggering the type review process.

[0118] In the type verification process, the list of archived files in the same batch in the associated file information is obtained from the current archived batch records in the document management system, while the historical archived records are obtained from the historical database of the document management system. The list of archived files in the same batch provides the relationships between documents, and the historical archived records provide historical type annotations for similar documents. This contextual information helps the semantic type discriminator to re-determine the document type in a richer context during secondary identification.

[0119] This embodiment achieves efficient and accurate identification of electronic document types by constructing a two-level recognition architecture that combines rapid pre-classification by a rule engine with precise identification by a semantic model. The document header feature information is initially matched by the rule engine to generate a candidate type set. The full-text content information is precisely matched within the candidate range by a semantic type discriminator to output two-level labels. A type confidence evaluator quantitatively assesses the reliability of the identification results and initiates a review process by introducing related file information when the confidence level is insufficient. This approach improves processing efficiency while ensuring recognition accuracy and provides a reliable type label foundation for the accurate generation of subsequent detection logic.

[0120] Please see Figure 3 In some embodiments, a set of detection logic corresponding to the electronic archive file is dynamically generated based on the document type tag, including:

[0121] S101. Read the business domain category tag and document template category tag from the document type tag;

[0122] S102. Input the business domain category label and the document template category label into the detection logic template library, and retrieve the basic detection logic template that matches the document template category label. The basic detection logic template contains detection item placeholders and detection rule placeholders.

[0123] S103. Extract the document content feature information of the electronic archive file. The document content feature information includes document structure features and document data features.

[0124] S104. Input the document content feature information and the basic detection logic template into the detection logic instantiator, fill in the detection item placeholders and detection rule placeholders in the basic detection logic template according to the document content feature information, and output the instantiated detection logic set.

[0125] S105. Input the instantiated detection logic set into the detection logic validator, perform logic validity verification on the instantiated detection logic set, and output the final detection logic set.

[0126] In step S101, the business domain category label and the document template category label are derived from the document type label output by the semantic type discriminator in the aforementioned embodiment. Both are read from the data structure storing the document type label through the data interface and are used to indicate the business domain scope to which the electronic file belongs and the standard document template framework it follows, providing a query basis for retrieving the matching basic template from the detection logic template library in the future.

[0127] In step S102, the detection logic template library is a predefined collection of templates, pre-built and stored in the system database by archival experts according to the detection requirements of various standard document templates. Each basic detection logic template contains a detection framework encompassing four dimensions: accuracy, completeness, usability, and security. Placeholders for detection items in the basic detection logic templates are used to reserve positions for descriptions of detection items to be filled, and placeholders for detection rules are used to reserve positions for descriptions of detection rules to be filled. The retrieval process matches and searches the detection logic template library based on the input document template category tag, returning the basic detection logic template corresponding to that document template category tag, which serves as the basic skeleton for subsequent instantiation and filling.

[0128] In step S103, document content feature information is obtained by parsing and extracting the main text of the electronic archive file. Document structure features are obtained by parsing the document's chapter hierarchy, title format, and paragraph organization. Document data features are obtained by identifying and extracting numerical fields, table structures, and key terms in the document. The extracted document content feature information serves as the specific material for placeholder filling in the subsequent detection logic instantiator.

[0129] In step S104, the training process of the detection logic instantiator uses a large amount of paired data containing document content feature information, basic detection logic templates, and corresponding filling results for fine-tuning. The training objective is to accurately fill the placeholders in the basic detection logic template based on the document content feature information. During inference, the detection logic instantiator receives document content feature information and the basic detection logic template as input. Through semantic understanding, it maps document structure features to detection item placeholders, generating detection item descriptions that match the functions of document chapters. At the same time, it maps document data features to detection rule placeholders, generating detection rule descriptions that include detection thresholds and matching conditions. Finally, it associates and pairs the filled detection item descriptions with the detection rule descriptions, outputting an instantiated detection logic set.

[0130] In step S105, the detection logic validator performs a logical validity check on the instantiated detection logic set. The validation rules are predefined based on the syntax specifications and logical constraints of the detection logic and stored in the validation rule base. The validation process includes checking whether the description of each detection item in the detection logic is clear and complete, whether the description of the detection rule is operable, whether the data fields referenced in the detection rule exist in the document, whether the detection threshold is within a reasonable range, and whether there are logical conflicts between the detection items. The instantiated detection logic set that passes the validation is output as the final detection logic set, which is used by the subsequent multi-level inference engine to perform detection.

[0131] This embodiment achieves personalized customization of electronic document detection solutions by constructing a dynamic generation mechanism that combines template retrieval with instantiation filling. Business domain category tags and document template category tags are used to retrieve matching basic detection logic templates from the detection logic template library. Document content feature information is used by the detection logic instantiator to fill placeholders and generate an instantiated detection logic set. The detection logic validator verifies the logic validity of the filling results to ensure executability. This solution enables on-demand dynamic generation of detection logic, providing a precisely adapted detection solution for subsequent multi-level inference execution.

[0132] In some embodiments, document content feature information and a basic detection logic template are input to the detection logic instantiator. Based on the document content feature information, the detection item placeholders and detection rule placeholders in the basic detection logic template are filled in, and an instantiated detection logic set is output, including:

[0133] The document structure features in the document content feature information are input into the detection item mapper. The detection item mapper is built based on the large language model and performs semantic parsing on the document structure features. The document structure features include chapter title sequence and paragraph structure features. The detection item mapper identifies the functional semantics of each chapter in the chapter title sequence and maps the detection item placeholders in the basic detection logic template to specific detection item descriptions that match the functional semantics of each chapter, thus completing the semantic filling of the detection item placeholders.

[0134] The document data features from the document content feature information are input into the rule parameter generator. The rule parameter generator is built based on a large language model and performs numerical distribution analysis and term frequency analysis on the document data features. The document data features include key term distribution and data table features. The rule parameter generator determines a reasonable range of detection thresholds based on the numerical distribution analysis results and determines the matching mode of detection conditions based on the term frequency analysis results. It replaces the detection rule placeholders in the basic detection logic template with specific detection rule descriptions that include detection thresholds and detection conditions, thus completing the parameterized filling of the detection rule placeholders.

[0135] The descriptions of detection items that have completed semantic filling are associated and paired with the descriptions of detection rules that have completed parameterized filling to generate instantiated detection logic. All instantiated detection logic is then aggregated and a set of instantiated detection logic is output.

[0136] In this embodiment, the detection item mapper learns from a large amount of paired data containing document structure features and corresponding detection item descriptions, enabling it to identify the functional semantics of each chapter title based on the input chapter title sequence and paragraph structure features. For example, when a chapter title describes funding or budget, the detection item mapper identifies its functional semantics as financial data verification; when a chapter title describes results or tasks, the detection item mapper identifies its functional semantics as content consistency verification. The chapter title sequence used by the detection item mapper to perform semantic parsing of document structure features comes from the extraction results of various levels of headings in the main text of the electronic archive document, and the paragraph structure features come from the parsing results of the paragraph organization method in the main text.

[0137] The rule parameter generator learns from a large amount of paired data containing document data features and corresponding detection rule parameters, enabling it to determine detection thresholds and conditions based on the input document data features. Numerical distribution analysis statistically analyzes the central tendency and dispersion of numerical fields in the document data features to determine a reasonable range for the detection threshold. For example, based on the deviation distribution between the values ​​of each item and the total value in the budget data, it determines the maximum allowable deviation range for consistency checks. Terminology frequency analysis statistically analyzes the occurrence position and frequency of key terms in each chapter of the document data features to determine the matching pattern for detection conditions. For example, based on whether a key term appears in a specified chapter, it determines the matching condition for field existence checks. The document data features used by the rule parameter generator for numerical distribution analysis are derived from the extraction results of tabular data and numerical fields in electronic archives, while the document data features used for terminology frequency analysis are derived from the statistical results of the distribution of key terms in electronic archives.

[0138] When associating and pairing semantically filled detection item descriptions with parameterized filled detection rule descriptions, the detection item descriptions and detection rule descriptions corresponding to the same detection item are combined into a complete instantiated detection logic based on the predefined correspondence between detection items and detection rules in the basic detection logic template. After summarizing all instantiated detection logic, a set of instantiated detection logic is output.

[0139] This embodiment achieves the transformation from a general detection logic template to personalized detection logic by constructing a detection item mapper and a rule parameter generator. The detection item mapper replaces detection item placeholders with detection item descriptions that match the functional semantics of each chapter based on document structure characteristics. The rule parameter generator replaces detection rule placeholders with detection rule descriptions that include specific detection thresholds and conditions based on document data characteristics. The two work together to ensure that the generated instantiated detection logic set can accurately adapt to the document characteristics of the current electronic archive file, providing a detection scheme that is highly matched to the document content for subsequent multi-level inference execution.

[0140] Please see Figure 4 In some embodiments, each detection logic in the detection logic set is executed sequentially. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs an inference path, and calculates a confidence score, generating a single detection result with an evidence chain and a confidence score, including:

[0141] S201. Read the detection logic to be executed from the detection logic set. The detection logic to be executed includes the detection item description, the detection rule description, and the detection target data range.

[0142] S202. Based on the range of the target data, locate and extract the original data fragments in the electronic archive file;

[0143] S203. Input the original data fragments and the detection rule descriptions into the evidence extractor. The evidence extractor is built based on a large language model. It performs semantic parsing on the original data fragments according to the detection rule descriptions and extracts key evidence information related to the detection rule descriptions from the original data fragments.

[0144] S204. Input the key evidence information into the reasoning path builder. The reasoning path builder is built based on a large language model. It performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to the preset reasoning paradigm. The multi-step reasoning path contains multiple reasoning nodes, and adjacent reasoning nodes are connected through causal relationships.

[0145] S205. Input the multi-step inference path into the confidence calculator. The confidence calculator is built based on the large language model and evaluates the confidence of the inference nodes in the multi-step inference path to generate the overall confidence score of the detection logic to be executed.

[0146] S206. The key evidence information, multi-step reasoning path and overall confidence score are encapsulated in a structured manner to generate a single detection result. The single detection result includes a problem description field, evidence chain field, reasoning path field, confidence score field and problem level field.

[0147] In step S201, the detection target data range is automatically determined by analyzing the data fields or document chapters involved in the detection rule description. This provides coordinate guidance for locating specific data areas in the electronic archive file. For example, when the detection rule description involves funding data, the detection target data range points to the table or chapter area containing funding data in the electronic archive file.

[0148] In step S202, the document structure parsing tool can be used to identify the chapter hierarchy and data area distribution of the electronic archive file. Based on the location markers specified for the target data range, the corresponding data area is located in the document structure, and text content or numerical data is extracted from that area as raw data fragments. The granularity of the raw data fragment extraction matches the precision of the target data range, ensuring that the extracted data fragments cover all the data content required by the detection rule description.

[0149] In step S203, the evidence extractor learns from a large amount of paired data containing original data fragments, detection rule descriptions, and corresponding key evidence information. This enables it to perform semantic parsing of the original data fragments based on the input detection rule description, identifying and extracting data content related to the detection rule description as key evidence information. For example, when the detection rule description involves data consistency verification, the evidence extractor identifies the data items and their values ​​that need to participate in the consistency comparison from the original data fragments, and outputs these data items and values ​​as key evidence information.

[0150] In step S204, the preset inference paradigm is a set of predefined inference logic templates covering common detection scenarios such as numerical comparison, field existence determination, and logical consistency determination. The inference path builder learns from a large amount of paired data containing key evidence information and corresponding multi-step inference paths, enabling it to select a matching inference paradigm based on the input key evidence information and generate a multi-step inference path according to the requirements of the selected inference paradigm. A multi-step inference path can be understood with the following example: for instance, the first inference node extracts values ​​A and B from the key evidence information; the second inference node compares values ​​A and B; and the third inference node concludes whether the data are consistent based on the comparison results.

[0151] In step S205, the confidence calculator learns from a large amount of paired data containing multi-step inference paths and corresponding overall confidence scores, enabling it to evaluate the confidence of each inference node in the input multi-step inference path. The evaluation criteria include the sufficiency of the evidence relied upon by each inference node, the certainty of the inference rules used, and the completeness of the inference steps. The confidence calculator comprehensively calculates the overall confidence score of the detection logic to be executed based on the confidence of each inference node, for example, by merging the confidence scores of each node into an overall confidence score through weighted averaging or taking the minimum value.

[0152] In step S206, the problem description field records a summary of the problem detected, the evidence chain field records key evidence information, the reasoning path field records a multi-step reasoning path, the confidence score field records the overall confidence score, and the problem level field is determined based on the overall confidence score and the preset severity of the detection rule.

[0153] In this embodiment, the evidence extractor extracts key evidence information from the original data fragments according to the detection rule description. The reasoning path builder generates a multi-step reasoning path according to the key evidence information and a preset reasoning paradigm. The confidence calculator evaluates the confidence of the reasoning nodes in the reasoning path and generates an overall confidence score. Finally, through structured encapsulation, a single detection result containing a problem description, evidence chain, reasoning path, confidence score, and problem level is generated. This upgrades the detection process from simple rule matching to traceable multi-step reasoning, providing complete evidence support and confidence quantification for the detection conclusion.

[0154] In some embodiments, key evidence information is input into a reasoning path builder, which is built based on a large language model. The reasoning path builder performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to a preset reasoning paradigm, including:

[0155] Key evidence information is input into the evidence relationship analyzer, which is built on a large language model. It performs pairwise relationship analysis on numerical evidence, textual evidence, structural evidence and related evidence in the key evidence information, identifies the supporting, contradictory and complementary relationships between various types of evidence, and outputs an evidence relationship graph. The evidence relationship graph is organized with evidence as nodes and relationship type as edges.

[0156] The evidence relationship graph is input into the inference starting point locator, which is built based on a large language model. The inference starting point locator identifies independent evidence nodes that do not have prior supporting evidence as initial evidence nodes in the evidence relationship graph, and identifies core evidence nodes that have the most subsequent supporting evidence as key inference nodes. It outputs the inference starting point set and the inference key node set.

[0157] The inference starting point set, the inference key node set, and the evidence relationship graph are input into the path generator. The path generator is built based on the large language model. It takes the initial evidence node in the inference starting point set as the starting point and the key inference node in the inference key node set as the necessary path point. It performs path search in the evidence relationship graph to generate multiple candidate inference paths from the initial evidence node through the key inference node to the final conclusion node.

[0158] Multiple candidate reasoning paths are input into the path optimizer, which is built based on a large language model. The path optimizer evaluates the path length, logical rigor, and evidence coverage of the multiple candidate reasoning paths. Based on the comprehensive score of path length, logical rigor, and evidence coverage, the optimal reasoning path is selected from the multiple candidate reasoning paths. The optimal reasoning path includes an initial evidence node, intermediate reasoning nodes, and a final conclusion node. Each reasoning node includes input evidence, reasoning rules, and output conclusion.

[0159] The optimal reasoning path is output as a multi-step reasoning path.

[0160] In this embodiment, the paired data used for training the evidence relationship analyzer can be constructed as follows: For key evidence information generated in historical detection tasks, annotators label the relationship type between each pair of evidence, forming training samples containing evidence pairs and relationship labels. The inference starting point locator identifies each evidence node by analyzing its in-degree and out-degree in the evidence relationship graph. The in-degree represents the number of pieces of evidence supported by other evidence, and the out-degree represents the number of pieces of evidence supported by other evidence. Evidence nodes with an in-degree of zero are independent evidence nodes, and evidence nodes with the largest out-degree are core evidence nodes.

[0161] When the path generator searches for paths in the evidence relationship graph, it starts from the initial evidence node and takes the key reasoning node as the necessary path point. It expands along the edges of supporting and complementary relationships. At each step of expansion, it selects adjacent nodes with supporting or complementary relationship types from the outgoing edges of the current node to add to the path, until it reaches a node without outgoing edges as the final conclusion node, thereby generating multiple candidate reasoning paths.

[0162] When evaluating path length, the path optimizer counts the number of inference nodes in the inference path; when evaluating logical rigor, it analyzes the completeness of the causal relationship chain between adjacent nodes in the inference path; and when evaluating evidence coverage, it counts the number of evidence types involved in the inference path. The comprehensive score is determined by weighted summation of the three evaluation results, with each weight pre-set based on the degree of influence of different evaluation indicators on the accuracy of the inference results in historical detection tasks.

[0163] In the optimal reasoning path, each reasoning node contains input evidence that indicates the content of the evidence on which the node is based, reasoning rules that indicate the type of logical rules used from the input evidence to the output conclusion, and output conclusion that indicates the reasoning result of the node. Together, these three constitute a traceable record of reasoning steps.

[0164] In this embodiment, the evidence relationship analyzer constructs an evidence relationship graph based on the logical connections between various types of evidence in the key evidence information. The reasoning starting point locator identifies the initial evidence node and key reasoning node in the evidence relationship graph. The path generator searches for multiple candidate reasoning paths with the initial evidence node as the starting point and the key reasoning node as the necessary step point. The path optimizer selects the optimal reasoning path from the candidate paths through three evaluations: path length, logical rigor, and evidence coverage, as the output of the multi-step reasoning path. This realizes the automatic construction of a structured reasoning path from scattered evidence, providing a logically clear and evidence-sufficient reasoning foundation for subsequent confidence assessment.

[0165] In some embodiments, all individual test results are summarized to generate a hierarchical test report, including:

[0166] Read all individual test results. Each individual test result includes a problem description field, a chain of evidence field, an inference path field, a confidence score field, and a problem level field.

[0167] Individual test results are grouped into primary groups based on the problem level field. The primary groups include the serious problem group, the general problem group, and the warning problem group. Individual test results within the same problem level group are sorted from high to low based on the confidence score field.

[0168] For the individual test results within each issue level group in the first-level group, secondary grouping is performed according to the test logic type. The test logic types include accuracy test logic, integrity test logic, usability test logic, and security test logic, forming a two-dimensional grouping matrix with issue level as the row and test logic type as the column.

[0169] The two-dimensional grouping matrix is ​​input into the report summary generator, which is built based on a large language model. It performs semantic aggregation on the single detection results in each cell of the two-dimensional grouping matrix, extracts the common features and core issues in each cell, and generates cell summary information. At the same time, it performs cross-cell semantic association analysis on the cell summary information of all cells, identifies the causal relationship and logical progression relationship between different cells, and generates cross-domain association summary information.

[0170] The cell summary information and cross-domain association summary information are input into the report structure arranger. The report structure arranger is built based on a large language model and arranges the cell summary information and cross-domain association summary information in a structured manner according to three levels: overall summary, category details, and repair suggestions. The overall summary level includes the overall detection conclusion of the electronic file, the total number of problems, and the core risk warning. The category details level includes a detailed table of detection results with a two-dimensional grouping matrix as the skeleton. The repair suggestions level includes the repair priority ranking and repair operation description for each detection item.

[0171] The output of the report structure arranger is linked and bound to the complete information of all individual test results to generate a hierarchical test report. The hierarchical test report supports layered viewing and interactive navigation from the overall summary level down to the category detail table level and then to the individual test result level.

[0172] In this embodiment, the value of the problem level field is derived from the classification result determined by combining the confidence score and the preset severity of the detection rules. The value of the confidence score field is derived from the output of the confidence calculator. During first-level grouping, the division into severe problem group, general problem group, and warning problem group is determined based on the specific value of the problem level field. Within the same problem level group, items can be sorted from high to low according to the confidence score field, with high-confidence detection items listed first.

[0173] When performing semantic aggregation, the report summary generator extracts common semantics from the problem description fields of multiple individual detection results within each cell of the two-dimensional grouping matrix. It identifies recurring detection patterns or problem types and uses the most frequently occurring common features as the core content of the cell summary information. During cross-domain semantic association analysis, the report summary generator performs causal logic analysis on the cell summary information of different cells, identifying whether a problem described in one cell might be a cause or result of a problem described in another cell. The identified causal relationships and logical progression relationships are used as cross-domain association summary information.

[0174] When determining the repair priority, the report structure arranger calculates based on the confidence score and problem level of each test item. Test items with higher confidence scores and more severe problem levels have higher repair priority. The repair operation description is automatically generated based on the test logic type to which the test item belongs and the content of the problem description field. Accuracy test logic corresponds to data correction operations, integrity test logic corresponds to data supplementation operations, usability test logic corresponds to format conversion operations, and security test logic corresponds to de-identification or encryption operations.

[0175] The implementation of hierarchical viewing and interactive navigation is as follows: Each data unit at each level of the hierarchical detection report contains an index identifier pointing to the corresponding data unit at the next level. When a user clicks on a core risk warning in the overall summary level, the system jumps to the corresponding detection result details area at the category detail table level based on the index identifier. When the user continues to click on a detection item in the category detail table level, the system jumps to the complete information of the single detection result of that detection item based on the index identifier.

[0176] In this embodiment, the report summary generator extracts common features of cells through semantic aggregation to generate cell summary information, identifies causal relationships between different cells through cross-domain association analysis to generate cross-domain association summary information, and the report structure arranger determines the repair priority and associates the repair operation description based on the confidence score and problem level. Finally, it generates a hierarchical detection report that supports index jumps and realizes the organization and presentation of scattered detection results into a structured and interactive report.

[0177] In some embodiments, the follow-up question and the corresponding chain of evidence information are input into the dialogue interpreter, which is built based on a large language model and outputs natural language explanation information for the detected item, including:

[0178] Receive follow-up questions initiated by the user. The follow-up questions include the natural language question entered by the user and the identifier of the detection item currently focused by the user.

[0179] Based on the test item identifier, locate and extract the individual test results corresponding to the test item identifier from the hierarchical test report;

[0180] The natural language question in the follow-up request is input into the intent parser. The intent parser is built based on a large language model, performs semantic parsing and intent classification on the natural language question, and outputs the follow-up intent label.

[0181] Based on the follow-up intent label, extract the target information fragment that matches the follow-up intent label from the single detection results;

[0182] The target information fragment and the natural language question are input into the explanation text generator. The explanation text generator is built based on a large language model and responds to the natural language question in a targeted manner according to the target information fragment, generating a preliminary explanation text.

[0183] The initial explanation text is input into the explanation depth controller, which is built based on a large language model. The explanation depth controller evaluates the explanation quality of the initial explanation text, makes adaptive adjustments to the initial explanation text based on the evaluation results, and outputs the final natural language explanation information.

[0184] The final natural language interpretation information is associated and bound with the evidence chain field in the individual detection results to generate an interactive interpretation view.

[0185] In this embodiment, the detection item identifier is a unique index for each detection item in the hierarchical detection report. It originates from the identifier assigned to each individual detection result when the report is generated. When a user clicks on a detection item in the report, the system automatically captures this identifier and includes it in the follow-up request. Based on the detection item identifier, an index lookup is performed in the data structure of the hierarchical detection report, returning the complete individual detection result data that matches the identifier.

[0186] The intent parser is constructed as follows: It collects follow-up questions from historical users regarding detection results. Annotators then label each question with an intent category, covering inquiries about detection basis, reasoning processes, reasons for confidence levels, and remediation methods. The follow-up intent tag is the category identifier output by the intent parser after classifying the natural language questions. Based on the different values ​​of the follow-up intent tag, corresponding field content is selected from individual detection results as target information fragments. For example, when pointing to detection basis, content is extracted from the evidence chain field; when pointing to the reasoning process, content is extracted from the reasoning path field.

[0187] The explanatory text generator is constructed by collecting natural language questions and corresponding target information fragments from historical follow-up requests, and having annotators write targeted response texts as training targets. When generating the initial explanatory text, the generator provides targeted responses to the points of confusion in the natural language questions based on the content of the target information fragments.

[0188] The interpretation depth controller performs readability and information completeness assessments on the initial interpretation text. Readability assessment is determined by analyzing sentence length and terminology density, while information completeness assessment is determined by comparing the content coverage of the interpretation text with the target information fragment. If the readability score is below a preset threshold, the initial interpretation text undergoes sentence segmentation and terminology annotation; if the information completeness score is below a preset threshold, missing key evidence nodes and reasoning steps are added. The preset threshold can be determined based on user satisfaction feedback statistics of historical interpretation texts. The adjusted text is then output as the final natural language interpretation information.

[0189] In this embodiment, the intent parser outputs follow-up intent tags based on the natural language question, extracts target information fragments from the single detection result based on the follow-up intent tags, the explanation text generator generates preliminary explanation text by responding to the natural language question in a targeted manner based on the target information fragments, and the explanation depth controller evaluates the readability and information integrity of the preliminary explanation text and outputs the final natural language explanation information after adaptive adjustment. Finally, it is associated and bound with the evidence chain field to generate an interactive explanation view, realizing the user's in-depth follow-up questioning of the detection result and the system's targeted response.

[0190] In some embodiments, a preliminary explanation text is input to an explanation depth controller, which is built based on a large language model. The explanation depth controller evaluates the explanation quality of the preliminary explanation text, adaptively adjusts the preliminary explanation text based on the evaluation results, and outputs the final natural language explanation information, including:

[0191] The initial explanation text is input into the readability evaluator, which is built based on a large language model. The evaluator performs sentence complexity analysis and term density analysis on the initial explanation text and generates a readability score based on the analysis results.

[0192] The preliminary explanation text is input into the information integrity evaluator, which is built based on a large language model. The evaluator performs semantic alignment comparison between the preliminary explanation text and the target information fragment, identifies the missing key evidence nodes and missing reasoning steps, and generates an information integrity score based on the identification results.

[0193] Readability scores and information integrity scores are input into an adjustment strategy decision-maker. This decision-maker, built on a large language model, selects the appropriate adjustment strategy based on the combination of the readability and information integrity scores. If the readability score is below a first preset threshold and the information integrity score is above a second preset threshold, a simplification adjustment strategy is selected. This includes splitting complex sentences, converting passive voice, and adding annotations to technical terms. If the readability score is above the first preset threshold and the information integrity score is below the second preset threshold, a supplementary adjustment strategy is selected. This includes inserting missing evidence node descriptions and supplementing intermediate conclusions in the reasoning steps. If both the readability score and the information integrity score are below the first and second preset thresholds, a reconstruction adjustment strategy is selected. This includes reorganizing the narrative structure of the interpreted text.

[0194] The adjustment strategy output by the adjustment strategy decision-maker and the initial explanation text are input into the text adjustment executor. The text adjustment executor is built based on the large language model, performs the corresponding adjustment operation on the initial explanation text according to the adjustment strategy, and outputs the final natural language explanation information.

[0195] In this embodiment, the readability evaluator performs sentence complexity analysis by statistically analyzing the average sentence length and the number of nested clauses in the preliminary explanation text; it performs terminology density analysis by statistically analyzing the frequency and density of technical terms in the text. The lower the sentence complexity and the lower the terminology density, the higher the readability score. The information integrity evaluator performs a semantic-level comparison between the preliminary explanation text and the target information fragment, identifies the coverage ratio of key evidence nodes and reasoning steps in the target information fragment in the preliminary explanation text, and quantifies the coverage ratio as an information integrity score.

[0196] The first and second preset thresholds can be determined based on user satisfaction feedback statistics of historical explanation texts. The adjustment strategy decision-maker selects the corresponding adjustment strategy based on the comparison results of readability score and the first preset threshold, and information integrity score and the second preset threshold. The text adjustment executor performs corresponding adjustment operations on the preliminary explanation text according to the adjustment strategy. The simplification adjustment strategy reduces the difficulty of understanding by splitting complex sentences, changing passive voice, and adding annotations to technical terms. The supplementary adjustment strategy improves information coverage by inserting missing evidence node descriptions and supplementing intermediate conclusions of reasoning steps. The reconstruction adjustment strategy improves readability and information integrity by reorganizing the narrative structure of the explanation text, and outputs the final natural language explanation information.

[0197] In this embodiment, the readability evaluator generates a readability score through sentence complexity analysis and term density analysis, the information integrity evaluator generates an information integrity score through semantic alignment comparison, the adjustment strategy decision-maker selects a simplification, supplementation or reconstruction adjustment strategy based on the comparison results of the two scores and a preset threshold, and the text adjustment executor performs the corresponding adjustment operation on the preliminary explanation text according to the selected strategy, thereby realizing the automated evaluation and adaptive optimization of the quality of the explanation text.

[0198] Please see Figure 5In a second aspect, this embodiment also provides a text processing system 1 based on a large language model, applicable to the method described in the first aspect. The system includes a document type recognition module 11, a detection logic generation module 12, a multi-level inference engine module 13, a report generation module 14, and a dialogue explanation module 15. The document type recognition module 11 has a built-in document type recognition model, which is constructed based on a large language model and is used to receive electronic archive files and output document type tags for the electronic archive files. The detection logic generation module 12 is connected to the document type recognition module 11 and has a built-in detection logic generator, which is constructed based on a large language model and is used to dynamically generate a set of detection logics corresponding to the electronic archive files based on the document type tags. The multi-level inference engine module 13 is connected to the detection logic generation module 12 and the data storage terminal of the electronic archive files, respectively. The inference engine module 13 incorporates a multi-level inference engine, which is built based on a large language model. This engine sequentially executes each detection logic in the detection logic set, extracting evidence information, constructing inference paths, and calculating confidence scores during the execution of each logic, generating individual detection results with evidence chains and confidence scores. The report generation module 14 is connected to the multi-level inference engine module 13. It incorporates a report summary generator and a report structure arranger, both built based on the large language model. These are used to summarize all individual detection results and generate a hierarchical detection report. The dialogue explanation module 15 is connected to the report generation module 14. It incorporates a dialogue explainer, also built based on the large language model. This explainer receives user requests for follow-up questions on any detection item in the hierarchical detection report and outputs natural language explanations for those items.

[0199] After the document type recognition module 11 outputs document type tags, the detection logic generation module 12 dynamically generates a set of detection logic based on these tags. The multi-level inference engine module 13 performs the detection and generates individual detection results with evidence chains and confidence scores. The report generation module 14 summarizes and generates a hierarchical detection report. The dialogue explanation module 15 responds to user follow-up questions and outputs natural language explanation information. The modules are connected in sequence to form a complete intelligent detection link.

[0200] By adopting the above technical solutions, this invention differs from existing technologies and has the following beneficial effects: After the document type recognition model outputs document type tags, the detection logic generator dynamically generates a set of detection logics containing four dimensions: accuracy, completeness, usability, and security. When executing each detection logic using a multi-level inference engine, evidence information is extracted, inference paths are constructed, and confidence scores are calculated to generate individual detection results with evidence chains and confidence scores. All individual detection results are summarized to obtain a hierarchical detection report. This report organizes the detection results hierarchically according to the problem level and supports interactive navigation. Furthermore, combined with a dialogue interpreter, it responds to user follow-up questions and outputs natural language explanations containing detection basis, inference process, and confidence basis. This achieves adaptive, interpretable, and interactive intelligent text detection, improving the transparency of the detection process and the credibility of the results.

[0201] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A text processing method based on a large language model, characterized in that, include: Obtain the electronic archive file to be processed, the electronic archive file containing metadata information and electronic file information; The electronic archive file is input into a document type recognition model, which is built based on a large language model, and outputs the document type label of the electronic archive file. Based on the document type tag, a detection logic generator is invoked. The detection logic generator is built based on a large language model and dynamically generates a set of detection logics corresponding to the electronic archive file according to the document type tag. The set of detection logics includes accuracy detection logic, integrity detection logic, usability detection logic, and security detection logic. The electronic archive file and the detection logic set are input into the multi-level inference engine. The multi-level inference engine is built based on a large language model and executes each detection logic in the detection logic set in sequence. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs inference paths, and calculates confidence scores to generate a single detection result with evidence chains and confidence scores. Summarize all individual test results to generate a hierarchical test report, which includes a problem description, evidence chain information, confidence score, and problem level for each test item; In response to a user's interactive operation on any detection item in the hierarchical detection report, a follow-up question is generated. The follow-up question and the corresponding evidence chain information are input into the dialogue interpreter. The dialogue interpreter is built based on a large language model and outputs natural language explanation information for the detection item. The natural language explanation information includes the detection basis, reasoning process and confidence basis.

2. The text processing method based on a large language model according to claim 1, characterized in that, The electronic archive file is input into a document type recognition model, which is built based on a large language model, and outputs the document type label of the electronic archive file, including: Extract the file header feature information of the electronic archive file, which includes file format identifier, metadata field structure, and file header content summary; The file header feature information is input into a type preclassifier, which is built based on a rule engine. The type preclassifier performs a preliminary match between the file format identifier and the metadata field structure, and outputs a set of candidate types. Extract the full-text content information of the electronic archive file, which includes chapter title sequence, key term distribution, paragraph structure features, and data table features; The full-text content information and the candidate type set are input into the semantic type discriminator. The semantic type discriminator is built based on a large language model, performs semantic understanding on the full-text content information, performs precise matching in the candidate type set, and outputs the document type label of the electronic archive file. The document type label includes a primary category label and a secondary category label. The primary category label indicates the business domain category of the electronic archive, and the secondary category label indicates the standard document template category of the electronic archive. The document type label is input into the type confidence evaluator, which is built based on a large language model. The type confidence evaluator calculates the type matching confidence for the primary classification label and the secondary classification label respectively. If the type matching confidence is lower than a preset threshold, the type review process is triggered. The review process for this type includes: Extract the associated file information of the electronic archive file. The associated file information includes a list of archived files in the same batch and historical archived records. Input the associated file information and the full text content information into the semantic type discriminator for secondary identification and output the corrected document type label.

3. The text processing method based on a large language model according to claim 1, characterized in that, Dynamically generate a set of detection logic corresponding to the electronic archive file based on the document type tags, including: Read the business domain category tag and the document template category tag from the document type tags; Input the business domain category label and the document template category label into the detection logic template library, and retrieve the basic detection logic template that matches the document template category label. The basic detection logic template includes detection item placeholders and detection rule placeholders. Extract the document content feature information of the electronic archive file, wherein the document content feature information includes document structure features and document data features; The document content feature information and the basic detection logic template are input into the detection logic instantiator. The detection item placeholders and detection rule placeholders in the basic detection logic template are filled according to the document content feature information, and the instantiated detection logic set is output. The instantiated detection logic set is input into the detection logic validator, which performs a logic validity check on the instantiated detection logic set and outputs the final detection logic set.

4. The text processing method based on a large language model according to claim 3, characterized in that, The document content feature information and the basic detection logic template are input into the detection logic instantiator. Based on the document content feature information, the detection item placeholders and detection rule placeholders in the basic detection logic template are filled in, and an instantiated detection logic set is output, including: The document structure features in the document content feature information are input into the detection item mapper. The detection item mapper is built based on a large language model and performs semantic parsing on the document structure features. The document structure features include chapter title sequences and paragraph structure features. The detection item mapper identifies the functional semantics of each chapter in the chapter title sequence and maps the detection item placeholders in the basic detection logic template to specific detection item descriptions that match the functional semantics of each chapter, thus completing the semantic filling of the detection item placeholders. The document data features in the document content feature information are input into the rule parameter generator. The rule parameter generator is built based on a large language model and performs numerical distribution analysis and term frequency analysis on the document data features. The document data features include key term distribution and data table features. The rule parameter generator determines a reasonable range of detection thresholds based on the numerical distribution analysis results and determines the matching mode of detection conditions based on the term frequency analysis results. The detection rule placeholders in the basic detection logic template are replaced with specific detection rule descriptions that include the detection thresholds and the detection conditions, thus completing the parameterized filling of the detection rule placeholders. The descriptions of detection items that have completed semantic filling are associated and paired with the descriptions of detection rules that have completed parameterized filling to generate instantiated detection logic. All instantiated detection logic is then summarized and the set of instantiated detection logic is output.

5. The text processing method based on a large language model according to claim 1, characterized in that, Each detection logic in the detection logic set is executed sequentially. During the execution of each detection logic, the multi-level inference engine extracts evidence information, constructs an inference path, and calculates a confidence score, generating a single detection result with an evidence chain and a confidence score, including: Read the detection logic to be executed from the set of detection logics. The detection logic to be executed includes a description of the detection item, a description of the detection rule, and a range of detection target data. Based on the range of the target data, locate and extract the original data fragments from the electronic file; The original data fragment and the detection rule description are input into the evidence extractor, which is built based on a large language model. The evidence extractor performs semantic parsing on the original data fragment according to the detection rule description and extracts key evidence information related to the detection rule description from the original data fragment. The key evidence information is input into the reasoning path builder, which is built based on a large language model. The reasoning path builder performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to a preset reasoning paradigm. The multi-step reasoning path contains multiple reasoning nodes, and adjacent reasoning nodes are connected by causal relationships. The multi-step inference path is input into a confidence calculator, which is built based on a large language model. The confidence calculator evaluates the confidence of the inference nodes in the multi-step inference path and generates an overall confidence score for the detection logic to be executed. The key evidence information, the multi-step reasoning path, and the overall confidence score are structured and encapsulated to generate the individual detection result. The individual detection result includes a problem description field, an evidence chain field, a reasoning path field, a confidence score field, and a problem level field.

6. The text processing method based on a large language model according to claim 5, characterized in that, The key evidence information is input into the reasoning path builder, which is built based on a large language model. The reasoning path builder performs logical relationship analysis on the key evidence information and generates a multi-step reasoning path according to a preset reasoning paradigm, including: The key evidence information is input into the evidence relationship analyzer, which is built based on a large language model. It performs pairwise relationship analysis on the numerical evidence, textual evidence, structural evidence and related evidence in the key evidence information, identifies the supporting relationship, contradictory relationship and complementary relationship between various types of evidence, and outputs an evidence relationship graph. The evidence relationship graph is organized with evidence as nodes and relationship type as edges. The evidence relationship graph is input into the inference starting point locator, which is constructed based on a large language model. The inference starting point locator identifies independent evidence nodes without prior supporting evidence in the evidence relationship graph as initial evidence nodes, and identifies core evidence nodes with the most subsequent supporting evidence as key inference nodes. The inference starting point set and the inference key node set are output. The inference starting point set, the inference key node set, and the evidence relationship graph are input into the path generator. The path generator is built based on a large language model. It takes the initial evidence node in the inference starting point set as the starting point and the key inference node in the inference key node set as the necessary path point. It performs path search in the evidence relationship graph to generate multiple candidate inference paths from the initial evidence node through the key inference node to the final conclusion node. The multiple candidate reasoning paths are input into a path optimizer, which is built based on a large language model. The path optimizer evaluates the path length, logical rigor, and evidence coverage of the multiple candidate reasoning paths respectively. Based on the comprehensive score of the path length, logical rigor, and evidence coverage, the optimal reasoning path is selected from the multiple candidate reasoning paths. The optimal reasoning path includes an initial evidence node, an intermediate reasoning node, and a final conclusion node. Each reasoning node includes input evidence, reasoning rules, and an output conclusion. The optimal reasoning path is output as the multi-step reasoning path.

7. The text processing method based on a large language model according to claim 1, characterized in that, Summarize all individual test results and generate a hierarchical test report, including: Read all individual detection results. Each individual detection result includes a problem description field, an evidence chain field, an inference path field, a confidence score field, and a problem level field. The individual test results are grouped into primary groups according to the problem level field. The primary groups include a serious problem group, a general problem group, and a warning problem group. The individual test results within the same problem level group are sorted from high to low according to the confidence score field. For the individual detection results within each problem level group in the first-level group, secondary grouping is performed according to the detection logic type. The detection logic type includes accuracy detection logic, integrity detection logic, usability detection logic, and security detection logic, forming a two-dimensional grouping matrix with problem level as the row and detection logic type as the column. The two-dimensional grouping matrix is ​​input into the report summary generator, which is built based on a large language model. The generator performs semantic aggregation on the single detection results in each cell of the two-dimensional grouping matrix, extracts the common features and core issues in each cell, generates cell summary information, and performs cross-cell semantic association analysis on the cell summary information of all cells to identify the causal relationship and logical progression relationship between different cells and generate cross-domain association summary information. The cell summary information and the cross-domain association summary information are input into the report structure arranger. The report structure arranger is built based on a large language model and arranges the cell summary information and the cross-domain association summary information in a structured manner according to three levels: overall summary, category details, and repair suggestions. The overall summary level includes the overall detection conclusion of the electronic file, the total number of problems, and the core risk warning. The category details level includes a detailed table of detection results with the two-dimensional grouping matrix as the skeleton. The repair suggestions level includes the repair priority ranking and repair operation description for each detection item. The output of the report structure arranger is associated and bound with the complete information of all individual test results to generate the hierarchical test report. The hierarchical test report supports hierarchical viewing and interactive navigation from the overall summary level down to the category detail table level and then to the individual test result level.

8. The text processing method based on a large language model according to claim 1, characterized in that, The follow-up question and the corresponding chain of evidence are input into the dialogue interpreter, which is built based on a large language model and outputs natural language explanation information for the detected item, including: Receive follow-up questions initiated by the user, the follow-up questions including a natural language question input by the user and the identifier of the detection item currently focused by the user; Based on the detection item identifier, locate and extract the individual detection result corresponding to the detection item identifier from the hierarchical detection report; The natural language question in the follow-up request is input into the intent parser, which is built based on a large language model, performs semantic parsing and intent classification on the natural language question, and outputs follow-up intent labels; Based on the questioning intent label, extract the target information fragment that matches the questioning intent label from the single detection result; The target information fragment and the natural language question are input into the explanation text generator, which is built based on a large language model and responds to the natural language question in a targeted manner according to the target information fragment, generating a preliminary explanation text. The preliminary explanation text is input into the explanation depth controller, which is built based on a large language model. The explanation depth controller evaluates the explanation quality of the preliminary explanation text, makes adaptive adjustments to the preliminary explanation text based on the evaluation results, and outputs the final natural language explanation information. The final natural language interpretation information is associated and bound with the evidence chain field in the individual detection result to generate an interactive interpretation view.

9. The text processing method based on a large language model according to claim 8, characterized in that, The preliminary explanation text is input into an explanation depth controller, which is built based on a large language model. The controller evaluates the explanation quality of the preliminary explanation text, adaptively adjusts it based on the evaluation results, and outputs the final natural language explanation information, including: The preliminary explanation text is input into a readability evaluator, which is built based on a large language model. The evaluator performs sentence complexity analysis and term density analysis on the preliminary explanation text and generates a readability score based on the analysis results. The preliminary explanation text is input into the information integrity evaluator, which is built based on a large language model. The evaluator performs semantic alignment comparison between the preliminary explanation text and the target information fragment, identifies the missing key evidence nodes and missing reasoning steps, and generates an information integrity score based on the identification results. The readability score and the information integrity score are input into an adjustment strategy decision-maker, which is constructed based on a large language model. The decision-maker selects a corresponding adjustment strategy based on the combination of the readability score and the information integrity score. If the readability score is lower than a first preset threshold and the information integrity score is higher than a second preset threshold, a simplification adjustment strategy is selected. This simplification adjustment strategy includes splitting complex sentences, converting passive voice, and adding annotations to technical terms. If the readability score is higher than the first preset threshold and the information integrity score is lower than the second preset threshold, a supplementary adjustment strategy is selected. This supplementary adjustment strategy includes inserting missing evidence node descriptions and supplementing intermediate conclusions of the reasoning steps. If the readability score is lower than the first preset threshold and the information integrity score is lower than the second preset threshold, a reconstruction adjustment strategy is selected. This reconstruction adjustment strategy includes reorganizing the narrative structure of the explanatory text. The adjustment strategy output by the adjustment strategy decision-maker and the preliminary explanation text are input into the text adjustment executor. The text adjustment executor is built based on a large language model, performs corresponding adjustment operations on the preliminary explanation text according to the adjustment strategy, and outputs the final natural language explanation information.

10. A text processing system based on a large language model, characterized in that, The system applicable to the method of any one of claims 1 to 9 comprises: A document type recognition module, which has a built-in document type recognition model, is constructed based on a large language model and is used to receive electronic archive files and output the document type tags of the electronic archive files. A detection logic generation module is connected to the document type recognition module. The detection logic generation module has a built-in detection logic generator, which is built based on a large language model and is used to dynamically generate a set of detection logics corresponding to the electronic archive file according to the document type tags. A multi-level inference engine module is connected to the detection logic generation module and the data storage terminal of the electronic archive file. The multi-level inference engine module has a built-in multi-level inference engine, which is built based on a large language model. It is used to execute each detection logic in the detection logic set in sequence. During the execution of each detection logic, it extracts evidence information, constructs inference paths, and calculates confidence scores to generate a single detection result with evidence chains and confidence scores. A report generation module is connected to the multi-level inference engine module. The report generation module has a built-in report summary generator and a report structure arranger. Both the report summary generator and the report structure arranger are built based on a large language model and are used to summarize all individual detection results and generate a hierarchical detection report. The dialogue explanation module is connected to the report generation module. The dialogue explanation module has a built-in dialogue explainer, which is built based on a large language model. It is used to receive follow-up questions from users about any detection item in the hierarchical detection report and output natural language explanation information for the detection item.

Images