Intelligent correction method and device based on semantic analysis, equipment and medium

Abstract

The invention relates to the technical field of semantic analysis, can be applied to business scenes of financial science and technology, medical treatment and health and the like, and discloses an intelligent correction method and device based on semantic analysis, equipment and a medium. The method comprises the following steps: extracting a structured document content from an explanation data source, extracting an auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source, carrying out semantic understanding processing on the structured document content and the auxiliary explanation texts by utilizing a semantic analysis module to generate a semantic analysis result, and executing a correction operation through an intelligent correction module based on the semantic analysis result. And after the correction operation is completed, the structured document content and the auxiliary explanation text are input into a format reconstruction module, and a specification-combined document is generated and output. According to the method, semantic understanding is carried out by combining structured document content and the auxiliary explanation text, and an automatic and standardized document compliance processing flow is realized based on intelligent correction and format reconstruction driven by a compliance semantic analysis result.

Description

Technical Field

[0001] The present invention relates to the field of semantic parsing technology, and in particular to an intelligent correction method, device, equipment and storage medium based on semantic analysis. Background Art

[0002] In the fintech business sector, particularly when promoting and communicating with consumers in consumer-facing products like insurance and wealth management, business personnel often rely on generative systems, template libraries, or manual work to create explanatory documents and supporting materials. While these materials possess strong business expressiveness, they lack systematic semantic understanding and content review mechanisms, often resulting in ambiguous statements, missing risk warnings, and inconsistent formatting. Current content review relies on manual item-by-item checks, which is not only labor-intensive and inefficient, but also prone to omissions, leading to repeated revisions and severely restricting compliance response speed and customer experience.

[0003] In the healthcare sector, scenarios such as health insurance descriptions, disease education content, and health service agreements involve a large number of professional and sensitive terms and expressions, placing extremely high demands on standardization and compliance. Existing technologies mostly rely solely on static content generation or template application, lacking the ability to extract correlations between structured document content and auxiliary text, as well as integrated analysis and automatic correction mechanisms for the overall semantics of the document. Consequently, manual review and formatting of generated content before submission is still required, resulting in low efficiency and difficulty in ensuring consistency.

[0004] Current semantic analysis systems are mainly used in fields such as public opinion monitoring, question-answering systems, or information retrieval. Their analysis granularity is insufficient to support the semantic linkage between multi-level structures and auxiliary explanations in complex documents. In the compliance semantic analysis tasks for structured content and multi-source auxiliary text, existing methods are often unable to identify potential non-standard expressions distributed at different levels or in multiple text paragraphs, and it is also difficult to complete the corresponding correction and structural reconstruction of semantic units. At the same time, format adjustment often relies on manual application of templates or fixed style configuration, lacking the ability to dynamically adapt standard formats according to content semantics, and unable to meet the consistent requirements of different fields for format accuracy and standardization.

[0005] To sum up, existing technologies still have problems such as coarse processing granularity, low intelligence, and inefficiency in supporting high-frequency content iteration in the integrated process of joint processing of structured content and auxiliary text, semantic understanding and automatic correction, and format standard output. These problems make it difficult to meet the integrated needs of intelligent review and standard output in highly compliant industries such as finance and healthcare. Summary of the Invention

[0006] The main purpose of the present invention is to provide an intelligent correction method, device, equipment and storage medium based on semantic analysis, aiming to solve the technical problem that the existing technology cannot achieve semantic joint analysis and synchronous correction of structured document content and auxiliary commentary text, resulting in low content review efficiency and high compliance processing duplication.

[0007] To achieve the above object, the present invention provides an intelligent correction method based on semantic analysis, comprising:

[0008] Extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0009] Performing semantic understanding processing on the structured document content and the auxiliary explanation text through a semantic analysis module to generate a semantic analysis result;

[0010] According to the semantic analysis results, a correction operation is performed on the structured document content and the auxiliary explanation text by an intelligent correction module;

[0011] Input the structured document content and auxiliary explanatory text that have completed the correction operation into the format reconstruction module to generate a compliant format document;

[0012] Output the document in the compliant format.

[0013] Furthermore, to achieve the above-mentioned purpose, the present invention provides an intelligent correction device based on semantic analysis, comprising:

[0014] A structured extraction module, configured to extract structured document content from the solution content and extract auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0015] A semantic analysis module, configured to perform semantic understanding processing on the structured document content and the auxiliary explanation text through the semantic analysis module to generate a semantic analysis result;

[0016] An intelligent correction module, configured to perform correction operations on the structured document content and the auxiliary explanation text according to the semantic analysis result through the intelligent correction module;

[0017] A format reconstruction module is used to input the structured document content and auxiliary explanation text that have completed the correction operation into the format reconstruction module to generate a compliant format document;

[0018] The compliance output module is used to output the compliance format document.

[0019] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and an intelligent correction program based on semantic analysis stored in the memory and runnable on the processor. When the intelligent correction program based on semantic analysis is executed by the processor, the steps of the intelligent correction method based on semantic analysis as described above are implemented.

[0020] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which an intelligent correction program based on semantic analysis is stored. When the intelligent correction program based on semantic analysis is executed by a processor, the steps of the intelligent correction method based on semantic analysis as described above are implemented.

[0021] Beneficial effects: The present invention relates to the field of semantic parsing technology and can be applied to business scenarios such as financial technology and medical health. It discloses an intelligent correction method based on semantic analysis, including: extracting structured document content in the solution content, and extracting auxiliary interpretation text corresponding to each unit in the structured document content from the interpretation data source, using the semantic analysis module to perform semantic understanding processing on the structured document content and the auxiliary interpretation text to generate semantic analysis results, performing correction operations based on the semantic analysis results through the intelligent correction module, and after the correction operation is completed, inputting the structured document content and the auxiliary interpretation text into the format reconstruction module to generate and output a compliant format document. The present invention realizes an automated and standardized document compliance processing process by combining the structured document content and the auxiliary interpretation text for semantic understanding, and based on the intelligent correction and format reconstruction driven by the compliance semantic analysis results, reducing multiple rounds of manual review and rework without changing the original content expression intention, thereby improving compliance processing efficiency and compliance pass rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0023] Figure 1 Schematic diagram of an application environment of an intelligent correction method based on semantic analysis in one embodiment of the present invention;

[0024] Figure 2 1 is a flow chart of an embodiment of an intelligent correction method based on semantic analysis according to the present invention;

[0025] Figure 3 Schematic diagram of functional modules of a preferred embodiment of an intelligent correction device based on semantic analysis of the present invention;

[0026] Figure 4 A schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0027] Figure 5FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] The intelligent correction method based on semantic analysis provided by the embodiment of the present invention can be applied in Figure 1 In an application environment, the user terminal communicates with the server terminal via a network. The server terminal can extract the structured document content in the solution content through the user terminal, and extract the auxiliary interpretation text corresponding to each unit in the structured document content from the interpretation data source, use the semantic analysis module to perform semantic understanding processing on the structured document content and the auxiliary interpretation text, generate a semantic analysis result, perform a correction operation based on the semantic analysis result through the intelligent correction module, and after the correction operation is completed, input the structured document content and the auxiliary interpretation text into the format reconstruction module to generate and output a compliant format document. The present invention realizes an automated and standardized document compliance processing process by combining the structured document content and the auxiliary interpretation text for semantic understanding, and based on the compliance semantic analysis result driven intelligent correction and format reconstruction, thereby reducing multiple rounds of manual review and rework without changing the original content expression intent, thereby improving compliance processing efficiency and compliance pass rate. The user terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices. The server terminal can be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0030] See also Figure 2 , Figure 2 This is a flow chart of an embodiment of an intelligent correction method based on semantic analysis provided by the present invention. It should be noted that although a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0031] like Figure 2 As shown, the intelligent correction method based on semantic analysis proposed by the present invention includes the following steps:

[0032] S10, extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0033] In this embodiment, extracting structured document content from the solution content refers to obtaining information units with hierarchical and tagged features from business solution materials provided to customers or for internal use according to preset structure recognition rules. The solution content usually includes product introductions, service descriptions, data presentations, legal terms and operating instructions, etc., and the structured document content is clearly divided into information components with logical separations such as page-level structure, chapter structure, and chart structure in these materials. The meaning of structured here includes not only the document's visual layout format (such as page numbering, title fonts, table separators, etc.), but also the organization form at the information semantic level, such as "Part I·Product Description", "Part II·Data Source", "Appendix A·Risk Level Comparison" and other labels. The extraction of structured document content depends on the cooperation of format parsing algorithms and label recognition models. Format parsing can be completed by combining document OCR text stream with document DOM structure analysis, and label recognition can be performed by combining semantic classification models to identify title text, thereby cutting the original document into several structural units with identifiable labels.

[0034] An explanatory data source is a collection of natural language annotated data accompanying business materials. Sources may include manual scripts written by document creators, transcribed speech recorded by system backends, spoken text by marketing personnel, or paragraph-level explanatory text extracted through other document parsing tools. Unlike the rigorous logical structure of structured document content, explanatory data sources tend to be more colloquial, figurative, or contextual. Their content often aims to enhance customer understanding, explain technical terms, and supplement background logic. Establishing a one-to-one correspondence between structured documents and explanatory data requires extracting semantic features from each document structural unit and accurately matching them within the explanatory data source using semantic similarity, paragraph alignment rules, or content template rules. Auxiliary explanatory text is a collection of explanatory text extracted, filtered, and organized from the explanatory data source. This content must semantically explain and support each unit in the structured document and ultimately, along with the structured content, can be incorporated into subsequent semantic analysis and compliance assessment processes.

[0035] In practical implementation, document content structuring is first required. Taking PDF documents as an example, a document structure parsing engine analyzes the page structure, extracting each page's title, paragraphs, figure boundaries, and table header information. Regular expressions, machine-learning title recognition models, or image recognition algorithms are used to identify the boundaries and attributes of document units and convert them into structured records. The labels of each structural unit in the document, such as chapter number, title keywords, and page location, are then used as query criteria to search the annotation data source for annotations with high semantic similarity. To achieve this matching, a bidirectional encoder such as BERT can be used to generate embedding vectors between the document unit and the annotation data, and matching is performed using cosine similarity or nearest neighbor search. In some scenarios, if the auxiliary annotation text is derived from a standard template or knowledge base, the corresponding text fragment can be directly indexed based on the structural unit labels. After matching, the system packages the structured document content and the corresponding auxiliary annotation text into a unified structured input object for subsequent semantic analysis.

[0036] In the application scenario, different format parsing methods can be used to extract structured document content. If the source document format is HTML or Markdown, the title and paragraph boundaries can be directly extracted through the DOM tree structure; if the document source is a scanned PDF, the image must first be converted into text through the OCR recognition module, and then the document structure is reconstructed based on the text layout and font features. For the recognition of table content, the position and logical ownership of each cell can be determined through the table boundary detection algorithm, and the table can be matched with the explanatory text as a structural unit. Template matching methods can also be used to define the structure of commonly used solution templates in the industry. When the new solution content meets a certain template, the structural unit is directly segmented according to the template logic.

[0037] During the extraction of auxiliary commentary text, if the source of the commentary data is complex, such as speech transcriptions, timestamps or page references automatically generated in the speech content can be used as matching features to assist in positioning. To improve matching accuracy, a fine-grained semantic classifier can be introduced to cluster or pre-classify the auxiliary commentary text, limiting matching to text of the corresponding type, thereby improving efficiency and accuracy.

[0038] To adapt to different business scenarios, the system can also set up an extensible structural unit labeling system. For example, in the financial industry, structured documents may use "product description," "revenue calculation," and "risk warning" as basic units, while in the medical industry, they may use "clinical indications," "drug mechanism," "adverse reactions," and "precautions for use" as labels. By defining a unified labeling meta-model, structural units in different fields can be modeled in a unified framework, facilitating downstream analysis and corrective processing.

[0039] Example: In a healthcare scenario, a pharmaceutical company is submitting promotional materials for a new drug to a regulatory platform. The promotional plan includes graphical comparative efficacy data, chapter information from the instructions, and oral presentations by physicians. During the extraction phase, the system matches document units in the "Mechanism of Action" section of the instructions with natural language annotations from the physicians explaining the mechanism, packaging them into structured analysis objects. This provides a foundation for subsequent detection of exaggerated efficacy and inaccurate wording.

[0040] In the financial sector, an insurance agent generated an annuity sales proposal. AIGC technology automatically generated a product benefit description, cash flow diagram, and customer benefit comparison table. During the extraction phase, the structuring module identified the product description and risk warning sections. It also mapped the agent's explanations in the recording, such as "This product's annual return is equivalent to 5% interest," into supplementary explanation text, providing accurate input for the semantic analysis phase to determine whether there are any exaggerated or misleading claims. This integration of structuring and supplementary explanations ensures a high degree of contextual restoration throughout the subsequent analysis, thereby enhancing the credibility and accuracy of the correction model.

[0041] By jointly extracting and structurally binding structured document content and supplementary explanatory text, a complete contextual input foundation is provided for subsequent semantic analysis and correction. By structurally decomposing document content and establishing a precise mapping relationship with semantic interpretation, problems such as unclear distribution of original document information, isolated text semantics, and fragmented context can be effectively addressed. This makes semantic analysis more accurate and correction suggestions more targeted, thereby improving the efficiency of compliance reviews.

[0042] S20, performing semantic understanding processing on the structured document content and the auxiliary explanation text by a semantic analysis module to generate a semantic analysis result;

[0043] In this embodiment, the semantic analysis module performs semantic understanding of structured document content and supporting explanatory text. This involves utilizing a language understanding model to identify multi-dimensional semantic features and extract compliance risk information based on the structured corpus of document units and their corresponding supporting explanatory text. In this process, the semantic analysis module undertakes the tasks of text-oriented semantic feature modeling and violation information identification. Its core capabilities include context modeling, compliance semantic recognition, inaccurate description extraction, and missing element detection.

[0044] Structured document content typically has strong structural features and terminology density, while auxiliary commentary text is more colloquial and contextual. In actual processing, it is necessary to first unify the format and lexical boundaries, use a word segmentation engine to segment the Chinese text, and integrate linguistic features such as part-of-speech tagging and named entity recognition to provide a foundation for subsequent deep models. Semantic understanding processing refers to constructing an embedded expression for the above text input and introducing semantic context. Semantic embedding can be performed through word vectors (such as Word2Vec, GloVe) or context-aware language models (such as BERT, RoBERTa, MacBERT) to ensure that the model can understand the hyponymy, negation structure, the semantic position of compliant keywords, and the direction of their modifiers in the expression.

[0045] After the embedding modeling is completed, the semantic analysis module completes targeted semantic detection by loading the preset compliance policy library. The compliance policy library can contain three types of rule bases: a set of illegal keywords, an inaccurate description feature library, and a missing annotation strategy. The illegal keyword set includes typical misleading terms (such as "guaranteed high returns", "risk-free", and "promised returns"). The inaccurate description feature library contains language structure markers such as vague words, absolute expressions, and contrasting rhetoric. The missing annotation strategy is used to detect whether key legal or regulatory requirements are missing in the fields that should be marked (such as unindicated product providers, data sources, risk disclosure statements, etc.).

[0046] The semantic analysis result is a comprehensive output of the above-mentioned multiple detection dimensions. The result can be presented in the form of structured tags, identifying the illegal expressions found in the text, their specific location (page number, paragraph number, line number), the original text of the expression, the type of violation (such as false promises, incomplete information), the severity level (high risk, medium risk, low risk) and the corresponding recommended rule number, etc., for subsequent correction modules to perform operations. The semantic analysis module can also output a "fuzzy term list" and "misleading phrase location" for inaccurate descriptions, and output a "field missing list" for missing annotations to ensure that each type of semantic risk point is completely restored.

[0047] In practical applications, the semantic analysis module can be implemented based on a deep semantic pre-training model. By introducing labeled compliance annotated corpus, BERT-like models can be retrained to identify contextual patterns related to compliance risks. Alternatively, a hybrid recognition strategy can be implemented by combining a rule engine with a deep model. Keyword rule matching can be used to identify high-frequency, fixed risks, while a deep model can be used to identify implicitly expressed risks.

[0048] For Chinese materials, the semantic understanding process often requires the inclusion of custom dictionaries to enhance the ability to identify domain terminology. For example, in the financial industry, specialized terms such as "policy year" and "cash value" are introduced, while in the medical industry, specialized terms such as "efficacy observation period" and "range of indications" are introduced. To improve the effectiveness of misleading judgments, text structure features can also be incorporated as an aid. For example, if a page contains both positive statements and no risk warnings, the model can combine the layout information to make a higher confidence judgment.

[0049] For the processing of auxiliary commentary text, the model needs to enhance fault tolerance and context integration capabilities. A sliding window mechanism can be used to splice it with the structured document unit as a whole analysis object to avoid semantic recognition failure due to the commentary language being too colloquial or interrupted.

[0050] Example: In a healthcare business scenario, a pharmaceutical company plans to recommend a new drug product to doctors. Its literature includes the statement "The efficacy is significantly superior to similar drugs." When processing this text, the semantic analysis module identifies "superior to similar drugs" as an inaccurate comparative description and outputs a violation type of "Unverified comparison," prompting a deletion and revision. The auxiliary explanation text contains the statement "Doctors report that this drug has very few side effects." The model identifies this as an "absolute statement" and fails to detect any risk descriptions. The model outputs "Risk warning missing" and generates a corresponding tag in the semantic analysis results.

[0051] In the financial sector, an agent was selling a pension product. Their structured documents contained phrases such as "principal guaranteed, stable returns," and the supplementary explanation text contained the phrase "annual returns equivalent to 5% annual interest." The semantic analysis module detected the "principal guaranteed" as a false promise and the "equivalent to 5% annual interest" as an exaggeration. It also identified the absence of the phrase "this product does not promise returns" in the supplementary explanation. The analysis output included two illegal statements and a missing annotation, facilitating subsequent targeted corrections.

[0052] The semantic analysis module, through its integrated analysis of structured document content and supplementary explanatory text, accurately identifies multiple language risks and establishes their location and context within the original document, enabling systematic detection of illegal expressions, inaccurate descriptions, and missing key content. This process not only improves the comprehensiveness and accuracy of semantic recognition but also provides actionable structured input for subsequent automated corrections, reducing reliance on manual judgment and improving compliance processing efficiency.

[0053] S30, performing correction operations on the structured document content and the auxiliary explanation text through an intelligent correction module according to the semantic analysis result;

[0054] In this embodiment, based on the results of semantic analysis, an intelligent correction module performs correction operations on the structured document content and auxiliary commentary text. This means that based on the aforementioned semantic analysis identifying issues such as illegal expressions, inaccurate descriptions, and missing annotations in the text, an intelligent module with language reconstruction capabilities is invoked to automatically perform operations such as content replacement, rewriting, and supplementation, thereby modifying non-compliant expressions into compliant and semantically coherent expressions. The core technical features of this operation include the prioritization of correction tasks, the generation and execution of correction operation instructions, the preservation of the semantic context of the text, and the restoration of structural consistency and information integrity.

[0055] The input of the intelligent correction module is the structured semantic analysis result, which usually includes the type of illegal statement, the specific location, the content of the statement, a list of inaccurate descriptions, the identification of missing fields and the recommended handling method. The correction module first prioritizes the correction of different types of semantic anomalies, usually sorting them according to the risk level of the violation. For example, high-risk illegal statements (such as "guaranteed returns" and "risk-free") need to be handled first, while low-risk and ambiguous expressions can be postponed. After the priority division is completed, the system generates a correction operation instruction queue to clarify the execution type of each operation, such as keyword replacement, sentence reconstruction or insertion of standard sentences.

[0056] When performing correction operations, for keyword replacements of illegal expressions, the correction module usually loads the synonym replacement table and context sample database from the compliance policy library to ensure that the semantics after replacement retains the original meaning but avoids violations. For example, "guaranteed annual income" is replaced with "expected annual income" while retaining the time structure and logical direction. For inaccurate descriptions, the system calls the language generation model (such as GPT, T5) to generate a variety of candidate sentences based on the original context, and then uses the sensitive word screening and rule verification module to filter out compliant expressions and complete the replacement. If the sentence contains misleading expressions due to missing modifiers or improper sentence structure, the system can automatically add contextual supplementary words or adjust the sentence structure to make its semantic expression clearer and more standardized.

[0057] When semantic analysis results contain missing annotations, such as missing product providers, data sources, or risk warnings, the correction module locates the missing areas based on the structured document's content type and uses the corresponding standard declaration template in the policy library to supplement them. For example, the module can add text such as "This product is provided by ×× institution. The data is derived from the ×× report of ×× year. Invest with caution" to the end of a product introduction paragraph, and automatically adjust the sentence structure and format to make it fit the context through the natural language generation engine.

[0058] Throughout the correction process, the module must maintain the contextual semantic consistency and structural rationality of the text to avoid information interruption or logical confusion caused by replacement or insertion. After each correction task, the system re-executes a lightweight semantic consistency check process to ensure semantic coherence, consistent style, and clear logic before and after the correction.

[0059] The intelligent correction module can be implemented by combining deep learning models with a rules engine. During the prioritization phase, a classification model can be trained using compliance-annotated data to predict the risk level of violation types. Corrective actions can be generated using a hybrid strategy of template-driven and generative models. For known, frequently violating expressions, a mapping table is used for rapid replacement. For inaccurate descriptions and ambiguous rhetoric, the generative model generates multiple candidate solutions based on context and then screens them based on compliance rules. The execution queue of correction tasks can be controlled by a state machine, with priority executed and the coverage of each action recorded to avoid conflicts caused by multiple modifications to the same area. In cases of semantic ambiguity or style conflicts, example text annotated by compliance personnel can be introduced as a style reference to guide the model in generating expressions consistent with the original industry language standards. When processing auxiliary commentary text, since this type of text often resembles natural spoken language, correction actions require enhanced language style adaptation capabilities. This can be achieved by segmenting the auxiliary commentary text and feeding it into the model along with the corresponding structured document units. A multi-channel attention mechanism is then used to ensure consistency between the correction actions and the main document, thus preventing style gaps.

[0060] Example: In a healthcare scenario, a document paragraph contains phrases such as "This product is suitable for all populations" and "Side effects are minimal." The semantic analysis module identifies these phrases and passes them to the correction module. Based on the system's rules, the system replaces "suitable for all populations" with "suitable for adults over 18 years of age" and rewrites "Side effects are minimal" to "Clinical data shows a low incidence of side effects." Furthermore, the system detects the absence of a risk warning in the paragraph and inserts the following at the end: "Please consult a professional physician before use, as individual risks may vary."

[0061] In practice in the financial field, the document states that "the annual return of this product can reach 7%, and the principal is risk-free." The correction module identifies "the annual return can reach 7%" as an inaccurate prediction and changes it to "the historical annualized rate of return is 7% (not representative of future performance)"; replaces "the principal is risk-free" with "the principal is subject to a certain fluctuation risk", and adds the standard statement "investment needs to be reasonably selected based on one's own risk tolerance" to restore compliance expression.

[0062] By performing correction operations on structured document content and auxiliary explanatory text, it is possible to quickly and accurately locate, effectively replace, and restructure expressions that violate compliance rules without relying on manual sentence-by-sentence verification. This not only ensures the accuracy of compliant expressions, but also preserves the original semantics and industry style, making content modifications more natural and credible.

[0063] S40, inputting the structured document content and auxiliary explanation text that have completed the correction operation into a format reconstruction module to generate a compliant format document;

[0064] In this embodiment, the corrected structured document content and auxiliary explanatory text are input into the format reconstruction module to generate a compliant document. This means that after the compliant content is corrected, the document's overall layout, hierarchical structure, title style, numbering specifications, paragraph layout and other formatting features are further uniformly adjusted to meet the document output standards required by industry or regulatory requirements. This process no longer focuses on the compliance semantics of the content itself, but instead focuses on the standardization and consistency of the document's presentation, ensuring that the document meets the output requirements for submission, archiving, and reviewability in terms of both visual structure and technical format.

[0065] Structured document content refers to a digital document representation structure composed of multi-level units, including document pages, chapters, entries, tables, diagrams, and other content with clear boundaries and semantic divisions. These contents have been semantically corrected in the aforementioned correction link, but the format level may be inconsistent due to the inconsistency of the original generation source (such as the AIGC model) or multiple edits. Auxiliary explanatory text is usually attached to each unit in the above document structure to enhance content understanding, but its format is often more colloquial or loosely structured, and needs to be typeset according to unified rules to facilitate compliance checks or automated recognition of business approval processes.

[0066] The core function of the formatting module is to load a set of preset formatting templates. These templates are typically based on predefined standard document formats for industries such as finance, healthcare, and law. They contain a structured set of rules for parameters such as paragraph indentation, title font and size, paragraph line spacing, item numbering, and figure numbering. The purpose of formatting templates is to provide a set of standard specifications for document expression, enabling formatting discrepancies analysis and structural mapping of input documents.

[0067] After the document is entered into the format reconstruction module, the system first analyzes the current formatting features of the structured document content and auxiliary explanatory text, including the line spacing, indentation, font, heading numbering format, headers and footers of each paragraph. Using a structure recognition algorithm, it creates a mapping table between content and format. The system then compares the current format with the target industry template, identifying formatting discrepancies, including mismatched titles, confusing numbering, incorrect fonts, inconsistent indentation, and unusual margins. These discrepancies are then marked and categorized.

[0068] After the difference items are identified, the format reconstruction module will adjust the format in batches according to the target template rules. The title font will be reset to the industry standard (such as the first-level title uses bold size 3, and the second-level title uses bold size 4), the numbering will be uniformly adjusted to a multi-layer nested format (such as 1, 1.1, 1.1.1), the table style will be adjusted to a unified border and alignment, and the paragraph format of the auxiliary commentary text will also be adjusted in conjunction with the format of the corresponding document unit. Throughout the process, format adjustments are not limited to visual layout, but also include the integrity of the label hierarchy and machine-readable structure to support subsequent automatic recognition and archiving needs.

[0069] The final document generated is a compliant format document, which contains structured content that has undergone dual processing of content correction and format reconstruction, and meets multiple standards of reviewability, archiving, and regulatory auditability.

[0070] The format reconstruction module can be implemented based on a template-driven automatic format adjustment engine. The module first calls the structure parser to perform DOM structure analysis on the input document, and divides the document content into logical nodes and presentation nodes. The system constructs a target format map based on the preset industry format template, and each type of node matches the corresponding format rule. During the adjustment phase, the rule matching and position positioning mechanism are used to replace each format feature that is different. If the title style is inconsistent, the system will replace the title tag with a standard title style tag; if the paragraph indentation is incorrect, the indentation value will be automatically calculated according to the current level and the style rules will be reapplied. For the auxiliary commentary text part, its format adjustment depends on the level of the main document unit to which it is bound, and format synchronization is achieved through inheritance rules.

[0071] In practical applications, this module can integrate technologies such as PDF reconstruction libraries, Word automatic typesetting APIs, and structured document editors (such as XML editors) to achieve automatic format verification and changes. In addition, to support template expansion across multiple industries, format template collections are configured and managed in JSON or XML format, facilitating rapid adaptation to different business scenarios.

[0072] Example: In the healthcare industry, after content correction, AIGC's automatically generated patient education materials, while linguistically compliant, all paragraph headings used a default font, had confusing numbering, and lacked a separate risk statement. The formatting module reorganized the materials, changing the first-level headings to a bold, size 3 font and centering, standardizing the numbering to 1.1 and 1.2, and bolding and separating the "Risks of Use" paragraph. The materials were then exported to PDF format for physician approval and patient distribution.

[0073] In the financial industry, even after content corrections, an investment plan prospectus still doesn't meet regulatory requirements. The formatting module remaps the document structure to a standard financial template, separating the cover page, product details page, and risk disclosure page into separate pages. It also adds unified version information and company statements to the footer, ultimately exporting the document to a compliant format for filing in compliance systems.

[0074] By reformatting the corrected structured document content and supporting explanatory text, we can further unify the document's external structure while ensuring content compliance. This can prevent secondary review failures or archiving difficulties caused by issues such as confusing styles, inconsistent numbering, and broken structures. Reformatting not only improves the document's readability and professionalism, but also provides technical support for compliance approval, historical archiving, and system recognition, ensuring both content quality and output format throughout the automated compliance process.

[0075] S50: Output the compliant format document.

[0076] In this embodiment, outputting a compliant document format means exporting the corrected and reformatted structured document content and supporting explanatory text into a standardized digital document format, ensuring that the document is readable, reviewable, and submittable, and meeting the practical application needs of business departments, compliance review systems, or regulatory platforms. "Output" here encompasses more than just file storage or display; it also encompasses multiple levels of processing, including content structure encapsulation, format specification implementation, data integrity verification, and adaptation of the output medium and format.

[0077] The compliant document is the final version of the structured document content and supporting explanatory text from the previous steps, after content optimization in the semantic correction module and presentation adjustments in the format reconstruction module. This document is structurally stable, semantically compliant, and formatted to industry standards. Its purpose is not only to store data but also to serve as a basis for subsequent delivery and compliant submission.

[0078] The output process is often accompanied by format conversion, such as converting internal editing formats (such as HTML, Markdown, and rich text JSON structures) into externally recognizable standard formats such as PDF, DOCX, ODF, and XML. This conversion not only includes the output of the visual layer of the content, but also includes the binding of structured metadata, such as chapter structure, title index, version information, generation time, signature tags, etc., to support business archiving or compliance audits.

[0079] Furthermore, output actions often involve the document's target location, including the local file system, enterprise content management platforms (such as ECM), compliance archive systems, regulatory reporting platforms, or direct entry into the approval process as an email attachment. To support these diverse scenarios, the output mechanism must possess multi-target adaptability, including file naming strategies, output path planning, format compatibility control, and output logging capabilities to ensure subsequent tracking and control.

[0080] Output modules often use formatting engines or document rendering engines to synthesize formats. For example, they can combine them with the LaTeX typesetting engine to generate high-quality PDF documents, use the Word API to automatically generate editable DOCX documents, or call an HTML renderer to output browser-friendly page documents. During the output process, the system also performs format integrity checks to ensure that there are no missing pages, garbled characters, or formatting errors in the document, ensuring that the output is qualified for formal submission.

[0081] Outputting compliant documents can be achieved through integrated document output engines such as Apache POI, PDFBox, Pandoc, and the LibreOffice SDK. One approach involves converting the formatted structured content and auxiliary text into an internal intermediate structure (such as a Document Object Model tree) before rendering it according to the target output format. In the PDF output, each element, such as paragraphs, headings, and charts, is mapped to a rendering layer, ensuring that margins, fonts, colors, and alignment conform to the template requirements.

[0082] During the output process, you can also add a watermark module, a signature module, or a compliance label module to ensure tamper-proofing, version identification, and audit tracking of output documents. For example, documents in the financial industry can automatically be watermarked "For compliance approval only." Documents in the medical industry can also be electronically stamped with the hospital's compliance department's signature and the identity and timestamp of the document's author.

[0083] For different business system integration scenarios, the output module also supports interface deployment, such as sending the final document to the enterprise DMS (document management system) or approval system through RESTful API to achieve automatic archiving and circulation.

[0084] After being exported through a unified output channel, completed compliant documents can be directly used in practical applications such as submission for approval, distribution to clients, and archiving and storage, eliminating the inefficient manual organization, format conversion, or content splicing required in traditional processes. The output module converts the results of format adjustments and structural packaging into standard format documents, ensuring consistency in document structure, content, and format. This improves the delivery quality and traceability of compliant materials in business, audit, and regulatory processes, significantly enhancing overall automated processing capabilities.

[0085] The present invention relates to the field of semantic parsing technology and can be applied to business scenarios such as financial technology and medical health. It discloses an intelligent correction method based on semantic analysis, including: extracting structured document content in the solution content, and extracting auxiliary interpretation text corresponding to each unit in the structured document content from the interpretation data source, using a semantic analysis module to perform semantic understanding processing on the structured document content and the auxiliary interpretation text to generate a semantic analysis result, performing a correction operation based on the semantic analysis result through an intelligent correction module, and after completing the correction operation, inputting the structured document content and the auxiliary interpretation text into a format reconstruction module to generate and output a compliant format document. The present invention realizes an automated and standardized document compliance processing process by combining structured document content and auxiliary interpretation text for semantic understanding, and based on the intelligent correction and format reconstruction driven by the compliance semantic analysis results, reducing multiple rounds of manual review and rework without changing the original content expression intention, thereby improving compliance processing efficiency and compliance pass rate.

[0086] In one embodiment, the above step S10 includes:

[0087] S101, extracting structured document content from the solution content, wherein the structured document content includes document page units, chapter units, and table units;

[0088] S102, parsing the document format of the structured document content according to the identification tags of the document page unit, chapter unit, and table unit;

[0089] S103, based on the parsed document format, establishing a mapping relationship table between each unit in the structured document content and the auxiliary interpretation text in the interpretation data source;

[0090] S104, extracting, from the explanation data source, page explanation text corresponding to each document page unit, chapter summary text corresponding to each chapter unit, and data annotation text corresponding to each table unit according to the mapping relationship table;

[0091] S105 , classifying and integrating the page explanation text, chapter summary text and data annotation text according to corresponding document page units, chapter units and table units to generate a structured auxiliary explanation text set.

[0092] In this embodiment, solution content refers to a collection of data documents used to present core content such as services, products, and strategies in a business system. It usually comes from manually edited document templates, structured output generated by large models, or content compilations pulled from APIs. In actual applications, solution content may exist in the form of JSON, XML, or HTML structures, or may be stored in the form of editable documents (such as DOCX and PDF). These solution contents contain structured organizational units. In order to facilitate semantic extraction and analysis, the structured document content must be extracted first.

[0093] Structured document content refers to text units that can be clearly divided into different semantic levels, usually including document page units, chapter units and table units. Document page units correspond to the content of one page or one screen, which is the logical division of the document presentation; chapter units are divided according to the title structure, reflecting the thematic organization of the content; table units are used to display data information, usually including indicators, numerical values, annotations, etc. The above units must be identified by identification tags, which may include HTML tags, Markdown syntax, paragraph styles, regular pattern matching results, etc. The process of parsing the document format is to classify, restore the structure and semantically align these identification tags, and extract format elements such as the structural position, parent-child hierarchy, and style features of each unit.

[0094] The explanatory data source refers to a data set used to provide supplementary explanations for the text, and its source can be historical copywriting materials, model output fragments, document template libraries, script databases, etc. These contents usually exist in unstructured or semi-structured forms and need to be structurally mapped through feature comparison or label indexing. The mapping relationship table is a structured data index table used to establish the association between each unit of the structured document and the explanatory data source. The key is to use the unique identifier of the structured document unit (such as ID, title text, location index) as the connection point, and combine it with supplementary features such as domain keywords, timestamps, and file paths for matching.

[0095] After completing the mapping, it is necessary to extract page explanation text, chapter summary text, and data annotation text from the explanation data source, corresponding to document pages, chapters, and table units respectively. Page explanation text is used to supplement the overview description of each page of the document, chapter summary text is used to explain the background and core logic of each chapter, and data annotation text is used to explain the source, indicator definition, calculation method, etc. of each column or row in the table. In order to improve the callability of data, it is necessary to classify and integrate these text contents, and generate a structured auxiliary explanation text collection based on the hierarchical relationship of the structured document. The organizational structure of the collection should be strictly aligned with the document format to facilitate pairing processing by the subsequent semantic analysis module.

[0096] In one implementation, Word or PDF documents can be parsed to identify paragraph styles, headers and footers, and table areas, and the document content can be structurally extracted by combining natural language template matching. Document page units are segmented by page breaks, chapter units are hierarchically judged based on font size and numbering patterns, and table units are identified and structured data is extracted through border lines. Each unit will generate corresponding structural location information, including page index, paragraph number, title path, etc.

[0097] To establish a mapping between structure and explanation, we can pre-encode the features of the text snippets in the explanation data source and perform vector matching with the contextual features of the structured units (title, context, and page position), thereby constructing a one-to-one or one-to-many index mapping table. This mapping strategy can be implemented using static methods such as fuzzy text matching and TF-IDF calculation, or by using semantic embedding models such as BERT and SimCSE for vectorized alignment.

[0098] Finally, the extracted page explanation text, chapter summary text and data annotation text are hierarchically integrated according to the corresponding document pages, chapters and table units, and a nested structure is used to construct an auxiliary explanation text collection. For example, the JSON format is used to identify that a document page contains several chapters, and the chapters have corresponding explanation paragraphs and table annotations. The structure remains consistent with the original document.

[0099] This embodiment maps and integrates structured document content with auxiliary commentary text at the structural level to achieve accurate binding of commentary data to document content, enabling subsequent semantic analysis and intelligent correction processes to be performed based on a complete semantic context, thereby improving the accuracy of illegal content identification and language rewriting, and avoiding analysis misjudgments or content replacement errors due to missing context.

[0100] In one embodiment, the above step S20 includes:

[0101] S201, loading a preset compliance policy library, wherein the compliance policy library includes a set of illegal expressions, an inaccurate description feature library, and a missing annotation policy;

[0102] S202, performing word segmentation processing on the structured document content and the auxiliary explanation text to generate semantically related context sentence fragments;

[0103] S203, matching illegal expressions in the context sentence fragments according to the illegal expression set, and marking the position and type of the illegal expressions;

[0104] S204, identifying ambiguous terms and misleading expressions in the context sentence fragment based on the inaccurate description feature library, and generating an inaccurate description list;

[0105] S205: Detecting unannotated product providers, data sources, and risk warnings in the structured document content according to the missing annotation strategy, and generating a missing annotation list;

[0106] S206 , integrating the location and type of the illegal expressions, the inaccurate description list and the missing annotation list to generate the semantic analysis result.

[0107] In this embodiment, the semantic analysis module is a computational component used to understand the contextual semantic structure, compliance implications, and potential risk factors of structured document content and supporting explanatory text. Its core task is to transform unstructured text into actionable structured semantic information for use by the remediation module. This module's processing is based on a set of rule resources, including a set of illegal expressions, a library of inaccurate description features, and a strategy for annotating missing items. These resources form the semantic rule foundation of the pre-set compliance policy library.

[0108] A collection of illegal expressions refers to keywords or phrases identified as potentially sensitive, misleading, or illegal during historical reviews or compliance assessments. These phrases are typically compiled by compliance personnel based on regulatory policies and internal company regulations. Examples include sensitive phrases such as "guaranteed return," "risk-free return," and "fastest profit." This collection includes not only the original text but also variations in different contexts.

[0109] The inaccurate description feature library is a collection of knowledge summarized from highly ambiguous and ambiguous expression patterns. This may include terms based on lexical ambiguity (such as "higher" and "stable"), compound structures that modify unclear quantitative relationships (such as "expected" and "usually"), or misused industry terms. Its sources may include historical compliance failure cases, corpus analysis statistics, and language model judgment.

[0110] The missing information marking strategy is used to define the key information types that should be clearly marked in the document and their judgment logic. It is mainly used to identify undisclosed or uncited statutory warning information, source statements or subject information, such as "product provider name", "risk warning terms", "data source", etc., and is usually implemented with syntactic templates, semantic assertions or structural position rules.

[0111] The semantic analysis module first performs word segmentation on the input structured document content and auxiliary commentary text, aiming to generate semantically complete contextual sentence fragments. Word segmentation is not limited to Chinese word segmentation but also includes named entity recognition, syntactic structure decomposition, and hyponym binding, ensuring that the analysis units not only have word boundaries but also semantic relevance.

[0112] After generating context sentence fragments, the set of illegal expressions is used to perform pattern matching and context judgment on these fragments. The risky expressions are located through forward rule matching, context semantic enhancement, word form normalization, etc., and their position, type, unit and word context in the original text are recorded.

[0113] Based on the inaccurate description feature library, the semantic analysis module identifies ambiguous terms and misleading expressions in the context. This identification mechanism can use fuzziness metric thresholds, language model scores, or semantic uncertainty analysis combined with the context of the auxiliary explanation text to extract potential inaccuracies and generate an inaccurate description list, including the location, content, ambiguity type, and rewording suggestions (if necessary).

[0114] At the semantic level of structured documents, missing annotation detection is typically performed based on structural pattern analysis. For example, missing risk warnings in a product description or missing data source identifiers in a benefit statement can be detected. This strategy relies on the ability to locate the structure and semantic coverage of document units. By detecting coverage gaps, it generates a list of missing annotations and outputs the contextual location and type of the missing content.

[0115] The semantic analysis results consist of the location and type of illegal expressions, a list of inaccurate descriptions, and a list of missing annotations. They are the basis of the semantic structure input into the subsequent correction module. All recognition results must have structural identification, semantic tags, and traceable locations to support the precise execution of the correction logic.

[0116] In one implementation, the semantic analysis module can be built using a multi-stage pipeline architecture. First, a custom tokenizer is invoked to perform tokenization and syntactic segmentation on the structured document content and auxiliary commentary text, while simultaneously performing named entity recognition and paragraph marker classification. This process can incorporate existing language model toolkits such as spaCy, LTP, and HanLP, supplemented by regular expressions to identify specific terms.

[0117] After the semantic fragment is generated, the module uses a keyword matching algorithm to scan whether there are any illegal expressions defined in the compliance policy library. If a match is found, its position, expression text, and paragraph index are recorded in the illegal expression list; if there is a semantic offset between synonyms or near-synonymous sentences, the similarity is judged through a semantic embedding model such as SimCSE to supplement the recognition.

[0118] Identifying inaccurate descriptions can be accomplished using a fuzzy judgment rule set, consisting of a sentence template and a list of uncertain words, which are combined and matched against the snippets at runtime. Language models can also be used to score snippets and filter out content with ambiguity exceeding a certain threshold. For example, if the word "stable" in "This product is long-term stable" is ambiguous, it may require rewording.

[0119] To detect missing annotations, we can use a predefined paragraph template library for structural matching. For example, if a product description lacks an "Investor Notes" paragraph or a "Risk Warning" phrase, it will be automatically determined to be missing. We also use methods such as paragraph length, keyword density, and semantic analogies to confirm whether the paragraph should contain this information.

[0120] The semantic analysis results are organized and output in a JSON structure. Each violation or missing record contains fields such as type, content, location information, unit type, and recommended processing strategy.

[0121] This embodiment uses a semantic analysis module to perform semantic understanding processing based on structured document content and auxiliary explanatory text under the guidance of compliance policy rules. It can accurately identify illegal expressions, inaccurate descriptions and missing annotations in documents, thereby providing structured and semantically clear operational targets for subsequent corrections, avoiding omissions and misidentifications in manual review, and significantly improving the efficiency and consistency of compliance corrections.

[0122] In one embodiment, the above step S30 includes:

[0123] S301, classifying the illegal expressions in the semantic analysis results into high-risk, medium-risk, and low-risk priority levels according to their severity, and assigning the highest correction priority to the illegal expressions of the high-risk priority level;

[0124] S302, generating a corrective operation execution queue in the order of high-risk, medium-risk, and low-risk priority levels;

[0125] S303, detecting whether there are conflicting tasks in the correction operation execution queue that perform multiple correction operations on the same structured document unit or the same auxiliary interpretation text, deduplicating or merging the conflicting tasks, and generating an optimized correction operation execution queue;

[0126] S304: According to the optimized correction operation execution queue, a correction operation is performed on the structured document content and the auxiliary explanation text by an intelligent correction module.

[0127] In this embodiment, the intelligent correction module is responsible for targeted adjustment, replacement, or supplementation of illegal or inaccurate information in structured document content and auxiliary commentary text, acting as the direct execution vehicle for the semantic analysis results. Based on the list of illegal statements, inaccurate descriptions, and missing annotations output by the semantic analysis module, the intelligent correction process constructs a correction execution system capable of priority control, conflict detection, and correction instruction optimization. Its execution is not limited to simple content replacement but incorporates multiple elements, including policy prioritization, correction type identification, contextual consistency maintenance, and sentence structure adaptation.

[0128] The first core technical feature of corrective actions is the prioritization of remediation based on the severity of the offending language. This severity can be determined based on factors such as regulatory risk level, historical compliance observations, dictionary scores for offending terms, and confidence in model assessment outputs. For example, statements containing promises of absolute returns or misleading investment behavior are classified as high risk, while general marketing exaggerations are medium risk, and inappropriate but lower-risk content is classified as low risk. This grading is the core judgment basis of the corrective action process, determining the urgency and sequence of corrective actions.

[0129] After completing the prioritization process, the system generates a queue of corrective actions based on priority. This queue includes three types of actions: keyword replacement for offending expressions, typically selecting replacements by matching against a library of compliant keywords; rewriting inaccurate phrases, using equivalent rewriting methods based on contextual semantic vectors; and supplementing prompts for missing information, using standardized prompt templates based on product or paragraph type. These actions have clear semantic goals and structural boundaries, with each action tied to the original text location, a suggested modification plan, and an action weight.

[0130] Since there may be multiple correction suggestions in the same document unit or explanatory text paragraph, the intelligent correction module needs to detect whether there are conflicting tasks, that is, executing multiple correction actions within the same text range that may overlap, conflict, or be out of order. For example, an illegal statement may be marked as "keyword replacement" and "phrase rewriting" at the same time. If both are executed directly, the output logic will be confused. Therefore, the system must establish conflict identification rules to determine whether the tasks are mutually exclusive, and merge conflicting tasks (such as giving priority to semantic generation of alternative keywords) or deduplicate (such as retaining high-priority actions), and finally generate an optimized correction operation execution queue.

[0131] The intelligent correction module then executes each optimized correction operation, precisely locating the target segment within the original text and performing replacement, rewriting, or insertion based on the type of modification. This operation must maintain consistency within the text context to avoid grammatical errors, semantic ambiguities, or sudden changes in style. Corrected results should retain identifying information to support subsequent review and version comparison.

[0132] In one implementation, when classifying the offending statements in the semantic analysis results into high, medium, and low risk levels, a weighted risk score can be used to quantify each offending statement. This risk score calculation considers multiple dimensions, including rule matching strength, degree of semantic violation, intensity of behavioral expression, and potential consequences of the violation. First, in the rule matching dimension, the offending statement is matched against prohibited keywords, expression patterns, or structured rules in the compliance policy library, and different risk base values ​​are assigned based on the severity level of the matching entries. For example, statements containing words such as "guaranteed profit," "guaranteed principal," and "return promise" may be assigned a higher risk weight, while more ambiguous terms such as "recommendation" and "better than average" may be assigned a lower weight. Second, in the semantic dimension, the statement is encoded using a pre-trained language model such as BERT or BGE, and semantic similarity is calculated with the vector center of the compliance violation semantic set. A higher similarity indicates a stronger semantic violation, reflecting its proximity to the high-risk statement in the actual semantic space. Third, in the dimension of behavioral expression, the rhetorical tone and structural features of the statement (such as commitments, introductory statements, and exclusive descriptions) are used to assess the intensity of its influence on user decision-making or the risks it implies. These structures often include phrases like "must," "absolutely," and "recommended immediately," and are generally considered triggers for medium- to high-risk behavior. Finally, in the dimension of potential consequences, the risk of misleading, complaints, or regulatory penalties that may result from the statement is estimated based on industry practices. For example, whether the statement involves misleading advertising, financial inducement, or information concealment is considered. Once the above dimensions are scored, the scores are weighted and summed using pre-set weights to form a final risk score. This score is then categorized into intervals: scores exceeding 0.75 are classified as high risk; scores between 0.4 and 0.75 are classified as medium risk; and scores below 0.4 are classified as low risk. This quantitative mechanism enables standardized, automated, and interpretable prioritization of noncompliant statements within large volumes of structured documents, providing a clear basis for subsequent corrective action.

[0133] During the execution queue generation phase, the system selects templates from three categories of corrective action templates that match the current violation. For keyword replacement, the system uses a library of compliant keyword replacements for precise matching. For phrase rewriting, the system uses pre-trained language models (such as BERT and T5) to generate semantically equivalent representations within a limited semantic vector range. For supplementary prompts, the system uses industry-standard prompt templates, such as "Investment is risky, please choose with caution."

[0134] The conflict detection mechanism uses a task conflict identification algorithm based on overlapping text spans to detect whether the target segments of tasks in the execution queue overlap. If multiple tasks are detected targeting the same paragraph or phrase, the system automatically adjusts the execution order or merges the operations based on the correction type priority strategy (for example, phrase rewriting over keyword replacement). The conflict resolution status is also marked for logging and manual review.

[0135] In the final execution stage, the system applies the operations in the task queue to the structured document content and auxiliary commentary text one by one, calls the compliance editor to adjust the sentence structure, supports language fluency detection and grammatical integrity verification, generates the final corrected document version, and enters the subsequent format reconstruction stage.

[0136] This embodiment accurately identifies violation points based on the results of semantic analysis. By setting priorities, building operation queues, detecting task conflicts, and accurately executing operations, it makes the content adjustment process more strategic and the execution more precise, and reduces the number of manual rounds of modification and compliance confirmation, significantly improving compliance processing efficiency, reducing communication costs, and increasing the first-time pass rate of documents.

[0137] In one embodiment, the above step S30 includes:

[0138] S305, extracting a synonym replacement table for the illegal expressions in the semantic analysis result from a preset compliance policy library through an intelligent correction module;

[0139] S306, matching candidate replacement words that meet current grammatical requirements from the synonym replacement table based on the grammatical structure of the structured document content and the auxiliary commentary text;

[0140] S307, synchronously replacing the illegal expressions in the structured document content and the auxiliary commentary text with the candidate replacement words to generate replaced structured document content and auxiliary commentary text;

[0141] S308, extracting contextual semantic vectors of misleading expressions and ambiguous terms in the replaced structured document content and auxiliary commentary text based on the list of inaccurate descriptions in the semantic analysis result;

[0142] S309, generating a plurality of candidate rewriting sentences based on the context semantic vector by calling a pre-trained language model through an intelligent correction module;

[0143] S310, screening candidate sentences without sensitive words from the rewriting candidate sentences according to the compliance policy library to generate a set of compliant rewriting sentences;

[0144] S311, replacing misleading expressions and ambiguous terms in the structured document content and auxiliary explanatory text with statements in the set of compliant rewritten statements to generate updated structured document content and auxiliary explanatory text;

[0145] S312, based on the missing annotations in the semantic analysis result, locating the end of the product description paragraph in the updated structured document content and the auxiliary explanation text by an intelligent correction module;

[0146] S313, matching the standard prompt template of the current product type from the compliance policy library through the intelligent correction module, and inserting the standard prompt template into the end position of the paragraph to generate structured document content and auxiliary explanation text that complete the correction operation.

[0147] In this embodiment, the execution of the corrective operation relies on the structured information set generated in the semantic analysis phase, which includes three categories of content to be processed: illegal expressions, inaccurate descriptions, and missing prompts. Illegal expressions are usually language expressions that are misleading or exceed regulatory boundaries, such as promised returns, exclusive comparisons, etc., and must be detected and replaced first. The system first extracts the synonym replacement table corresponding to these illegal expressions from the policy library. The construction sources of the replacement table include but are not limited to accepted expressions in historical compliance cases, consensus of domain experts, and the summary of semantically equivalent corpus. The entries in the replacement table contain not only the words themselves, but also attribute tags such as part-of-speech tags, context availability constraints, collocation frequency, industry preference expression level, etc., so that the same illegal word can select the closest replacement item according to the context.

[0148] When matching candidate replacement terms, the system performs part-of-speech tagging, syntactic analysis, and dependency analysis on the structured document content and auxiliary commentary text to ensure that the selected replacement term does not cause subject-verb mismatches or logical breaks within the grammatical framework of the original sentence. For example, if the original sentence is "This product guarantees returns," where "guarantee" is a nonsensical term, the system might suggest replacing it with "support" or "provide." However, if the subject of the sentence is "Terms of Service," "support" would be a more appropriate replacement.

[0149] Terms or phrases identified as inaccurate descriptions, such as "stable value-added" and "long-term security," often exhibit semantic ambiguity or pragmatic exaggeration. To avoid disjointed expressions caused by mechanical replacement, the system extracts the sentence and its surrounding context to construct a contextual semantic vector. This vector is typically the result of multi-level feature fusion, including word-level embeddings, syntactic structure representations, and entity context weights. The intelligent correction module inputs this vector into a pre-trained language model, which then generates multiple candidate sentences based on the contextual semantics, replacing the inaccurate content while maintaining the amount of information.

[0150] Generated rewritten sentences are submitted to the compliance policy library for sensitive content review. This review process supports rule set filtering, sentence structure comparison, and intent detection model evaluation to ensure that the selected rewrite candidates strike a balance between wording style, compliance tone, and actual statement boundaries. For example, rewriting "This product is suitable for all customers" to "This product is suitable for eligible customer groups" preserves the core information while eliminating the risk of the statement.

[0151] After replacing misleading content, it is necessary to further complete the missing key compliance information. Missing items marked in the semantic analysis results are usually product responsible parties, data source descriptions, or necessary prompts. For example, if a paragraph only describes product functions without mentioning the data source and scope of application, the system will automatically locate the end of the paragraph and identify the exact insertion point through paragraph structure markers, sentence terminators, and conclusion recognition. Then, based on the current product type and compliance rules, it will select a matching prompt template from the policy library for insertion.

[0152] This replacement and insertion operation not only affects the text content but also synchronously updates the internal structure of the document, including paragraph numbering, highlighting prompts, or annotations, ensuring that the visual structure of the content meets business review standards. For example, "Data source: A third-party investigation report" can be added to the auxiliary explanation text, while footnote indexing can be added to the document structure to maintain document consistency.

[0153] Throughout the entire remediation process, a priority control mechanism for the remediation strategy is implemented. Different issue items are prioritized, such as prioritizing legal risk terms, followed by ambiguous expressions, and finally addressing missing prompts. This remediation process not only reflects the level of compliance risk but also optimizes the subsequent review process for compliance personnel. The final updated content can be submitted directly for compliance review or generated into a standardized output document.

[0154] Example: In the financial sector, an insurance agency platform uses generative AI to assist in the production of customer promotional materials. One automatically generated sales copy contained illegal or exaggerated language, such as "This product offers stable returns, guaranteed principal and interest, and is suitable for all customers." Traditionally, such content requires manual annotation and multiple rounds of revisions, with an average approval cycle of two to three business days. After deploying this system, the system identified the level of violation in these expressions through semantic analysis. Based on historical compliance materials and a policy library, the system replaced "stable" with "provided with a risk diversification mechanism," removed "guaranteed principal and interest" and replaced it with a risk disclosure statement, and reworded "suitable for all customers" to "suitable for eligible customer groups." After the intelligent correction module generated the content, it also added a reminder stating "Data derived from a 2023 customer satisfaction survey" and inserted a standardized reminder: "This product does not guarantee returns; invest with caution." After review and confirmation by compliance personnel, the content was approved immediately, shortening the approval cycle to 0.5 business days, effectively improving compliance efficiency and reducing communication costs.

[0155] In the healthcare sector, a hospital's self-developed intelligent follow-up system uses a question-and-answer model to generate postoperative rehabilitation instructions. Ambiguous or even inaccurate statements such as "postoperative rehabilitation has no side effects" and "this regimen can replace traditional treatments" often appear in the automatically generated content. Through the system's semantic understanding and correction mechanism, the system accurately identifies "no side effects" as an inaccurate description and "replaceable" as a medical compliance risk. It then rewrites the system based on the medical industry knowledge base, generating statements such as "Most patients recover well after surgery, but individual responses vary" and "This regimen can be used as an adjunct treatment pathway after a doctor's evaluation." Furthermore, based on a missing annotation strategy, the system automatically adds the prompt "Please follow your doctor's instructions and do not adjust your treatment on your own" and inserts a source note at the end: "Rehabilitation recommendations are compiled based on the 2022 Clinical Guidelines of the Chinese Rehabilitation Medicine Association." The output text strikes a balance between readability and compliance integrity, while also meeting the requirements of the hospital's internal review mechanism.

[0156] This embodiment uses a three-level intelligent correction mechanism: synonym replacement, contextual rewriting, and missing prompt completion. This mechanism effectively improves the accuracy and coverage of correction suggestions. While ensuring semantic coherence, the correction process also ensures that compliant content expression complies with industry regulatory requirements, avoiding multiple rounds of review due to wording errors. Leveraging the context generation capabilities of pre-trained language models, it can adapt to a variety of document scenarios, achieving compliant expression conversion from fine-grained replacement to full-sentence reconstruction.

[0157] In one embodiment, after the above step S30, the method further includes:

[0158] S314, loading the prohibited content feature library in the preset compliance policy library through the prohibited content processing module;

[0159] S315, performing semantic similarity analysis based on the prohibited content feature library to match prohibited content paragraphs in the structured document content and the auxiliary commentary text;

[0160] S316, deleting the prohibited content paragraphs in the structured document content and the auxiliary commentary text, recording the deleted position of the prohibited content paragraph, and extracting the adjacent paragraphs of the deleted position as the associated context;

[0161] S317, calling a pre-trained language model to generate a coherent rewritten text based on the associated context;

[0162] S318: inserting the coherent rewritten text into the deleted position recorded in the structured document content and the auxiliary commentary text to generate the structured document content and the auxiliary commentary text after the correction operation is completed.

[0163] In this embodiment, the prohibited content processing module is responsible for detecting and replacing potential industry prohibited expressions in structured document content and auxiliary explanation texts. The prohibited content feature library loaded by this module is a subset of the compliance policy library, which includes industry competitor comparison descriptions, product service comparison descriptions, and illegal operation keywords. Industry competitor comparison descriptions may include expressions such as "this product is better than a certain brand" and "the highest profit among similar products", which often constitute expressions that violate the principle of industry neutrality in advertising texts; product service comparison descriptions such as "the only one with dual account function" and "more efficient than traditional products" are also prone to regulatory misjudgments; illegal operation keywords cover high-frequency banned words such as "guaranteed principal and interest", "guaranteed profit without loss", and "no approval required". This type of content has a fixed expression template and high compliance risk, and needs to be defined in advance through feature extraction.

[0164] Based on a prohibited content feature library, the system performs semantic similarity analysis on structured document content and supporting explanatory text. This similarity analysis relies on a sentence vectorization model. After vectorizing the original text and prohibited expressions, the cosine similarity metric is used to determine if there is a high degree of overlap. Paragraphs exceeding this threshold are considered prohibited content. Unlike keyword matching, semantic similarity analysis can identify variant expressions and changes in language style, enabling it to capture indirect violations such as "Our products are far ahead of the industry average."

[0165] Once a prohibited content paragraph is identified, the system removes it from the structured document content and auxiliary commentary text, and records its location. To ensure semantic coherence, the system extracts a certain number of sentences from the paragraphs before and after the deleted paragraph to form a contextual link. This context typically includes the ending sentence before the deleted paragraph and the opening sentence after the deleted paragraph to preserve the logical chain.

[0166] The system then uses a pre-trained language model to automatically generate coherent rewrites based on the context. This model has been fine-tuned using compliant corpus and industry semantic data to ensure the generated content maintains linguistic fluency, contextual cohesion, and compliant expression consistency.

[0167] Finally, the system inserts this coherent rewritten text into the original structured document content and auxiliary commentary at the locations recorded for deletion, while also rearranging the logical connections between adjacent paragraphs to create a compliant version with seamless semantics and more natural expression. This process results in a complete document version that includes corrections and replacements for prohibited content.

[0168] During implementation, feature vector libraries of prohibited content can be configured for different risk levels to match text review standards in different scenarios. For example, in the financial and insurance industry, keywords such as "unique," "guaranteed unchanged," and "extremely low risk" could be defined as high-risk; in the medical field, "can replace traditional treatments" and "significantly higher than the industry average" could be designated as medium-risk. During implementation, the system can dynamically adjust thresholds, for example, identifying a prohibited paragraph as one with a semantic similarity exceeding 0.82. Regarding semantic context construction, the system can default to extracting one paragraph before and one paragraph after as context; more intelligent window selection strategies can also be implemented based on sentence length or subject consistency within a paragraph. The pre-trained language model can utilize a BERT+GPT fusion architecture to enhance the contextual awareness and diversity of generated content. After generation, filtering rules, such as regular expressions, can be used to further enhance the security of the generated content by filtering retained or added words. Furthermore, to adapt to different application scenarios, custom domain-specific prohibited phrases can be added, such as exaggerated descriptions of data transmission security levels in medical and health data processing instructions, or implicitly violating regulations in finance by omitting warnings such as "does not constitute investment advice."

[0169] Example: In the financial sector, a wealth management company generated a client asset allocation proposal that stated, "Compared to other wealth management products on the market, this product offers more stable returns." Using the prohibited content processing module, the system identified this sentence as containing a comparative description of "other wealth management products," with a semantic similarity score of 0.91, meeting the high-risk deletion threshold. The system extracted the two paragraphs before and after this paragraph, constructed the context, and generated the following content using a pre-trained language model: "This product is suitable for clients who prefer stable returns, and the return structure is flexibly adjusted based on market data." After the new text was inserted, the text became logically coherent, without any comparative statements, and met the regulatory compliance review standards.

[0170] In a healthcare setting, a rehabilitation instructional document contained the sentence "This program significantly outperforms traditional rehabilitation methods." The system identified this as medium-risk prohibited content related to "product and service comparison," deleted it at the paragraph level, and generated text based on the relevant context: "The rehabilitation plan will be tailored to the individual patient's condition, incorporating multiple treatments to enhance recovery." This rewording made the content compliant and neutral, avoiding direct comparisons across institutions or programs and enhancing professional credibility.

[0171] This embodiment, by disabling the content processing module, expands compliance corrections from general expression corrections to the precise deletion and intelligent rewriting of high-risk paragraphs. After identifying illegal comparison expressions or misleading marketing language, the system can quickly locate and automatically repair potential compliance risks. This can significantly shorten the content review cycle, reduce compliance errors and omissions, and improve the logical integrity and readability of the text. Through the contextual connection mechanism and the control strategy of the generated language, it takes into account compliance, fluency, and generation flexibility, and has significant application advantages in batch document error correction tasks.

[0172] In one embodiment, the above step S40 includes:

[0173] S401, loading a format template set in a preset compliance policy library through a format reconstruction module;

[0174] S402, parsing the current format elements of the structured document content and the auxiliary commentary text through a format reconstruction module;

[0175] S403, detecting format differences between the current format element and a target industry template in the format template set by a format reconstruction module;

[0176] S404: adjusting, by a format reconstruction module, paragraph spacing, title font style, and numbering sequence rules of the structured document content and the auxiliary commentary text based on the format difference items, so that the structured document content and the auxiliary commentary text match the format rules of the target industry template;

[0177] S405: Output the adjusted structured document content and auxiliary explanation text into a compliant format document through a format reconstruction module.

[0178] In this embodiment, the format reconstruction module is used to perform unified format standardization processing on the structured document content and auxiliary commentary text after the content correction is completed, so as to meet the compliance display standards of a specific industry or organization. The module first loads the format template set, which is a subset of the compliance policy library and contains multiple industry standard document templates. These templates not only define the basic rules for paragraph indentation, but also include the font style requirements of the title (such as font type, font size, bold, alignment, etc.), and the hierarchical rules of the numbering sequence, such as 1.1, 1.2.3 or ABC and other multi-level title numbering structures. The setting of the template set can cover a variety of document style requirements such as financial regulatory reports, medical service instructions, insurance policy text specifications, product description standards, etc., and is highly configurable.

[0179] After loading the template, the system parses the format elements of the current structured document content and auxiliary commentary text. The current format elements refer to the typesetting parameters actually used in the document, including paragraph indentation, title font attributes (such as whether it is Songti, whether it is size 14, whether it is bold), and the structural pattern of the title number. These elements are generally automatically extracted from document structured tags, typesetting attributes or style tag information. If the source is a structured format such as XML, HTML or DOCX, it can be extracted through a tag parser; if it is in plain text or PDF format, it can be reconstructed using a format recognition model or a visual analysis model.

[0180] The system then checks for discrepancies between the current formatting elements and the target industry template. These discrepancies can include indents that differ from the template requirements, title font styles that don't conform to standards, or a confusing numbering hierarchy. These discrepancies can impact the document's compliance approval rate and the accuracy of business processing.

[0181] To resolve the differences, the system makes corresponding adjustments based on the identified format differences. Paragraph spacing adjustment operations may include unifying the left indentation of all paragraphs to the template requirement value (such as 2 character units), and automatically filling in the paragraph spacing before and after paragraphs; title style adjustment includes unifying all first-level titles to 16-point bold, Songti centered, and setting the second-level title to 14-point non-bold left-aligned; the numbering sequence can be rewritten as multi-level ordered numbering, and the hierarchical correspondence is automatically adjusted. The above adjustments rewrite parameters based on the difference item rules, and keep the original text structure unchanged to ensure content consistency.

[0182] After formatting is complete, the system exports the document to a compliant format. This output process supports mainstream document formats such as DOCX, PDF, and HTML, and supports rendering and generation using template engines or style engines when necessary. All formatting elements in the output document are identical to the target template, providing a clear, standardized, and reviewable presentation that meets the needs of review, archiving, and client communication during document circulation.

[0183] In implementation, the format template collection uses a JSON structure to store template rules. Each industry-standard template includes fields such as format name, applicable scenarios, paragraph rules, title rules, and numbering rules. The system automatically selects the appropriate template based on the document's business type tag. For example, medical documents can be loaded with the medical industry standard template, while financial materials can be loaded with the financial industry disclosure format template.

[0184] The current formatting elements can be parsed by reading structural tags (such as paragraph styles in DOCX and class tags in HTML), or by combining NLP document structure recognition methods to extract logical paragraphs. If the input is an image or PDF layer text, the system can use OCR combined with a layout parsing model to identify paragraph blocks and reconstruct the formatting structure.

[0185] When detecting format differences, the system constructs a format attribute vector, compares the current document format with the template format one by one, and calculates similarity or matches along the corresponding dimensions. Examples include font name matching, font size range comparison, and whether indentation values ​​are within a single character. Identified mismatches are considered differences.

[0186] Adjustments can be made using the paragraph style rewriting engine, which supports batch replacement of corresponding attributes at the paragraph and heading levels. For content with a chaotic numbering structure, the numbering relationship at each level can be recalculated by traversing the heading tree, and the overall document numbering can be updated to maintain a clear heading structure. The final document export can use template engines such as Jinja2 or Word style interfaces to restore formatting and uniform rendering.

[0187] Example: A large insurance company used the AIGC tool to generate a preliminary proposal document for a newly launched annuity insurance product. This document included basic product information, benefit demonstrations, risk warnings, and supporting explanatory text written by the marketing department. Because it involved financial products, the content required review and approval by the compliance department. The system first extracted the structured document content from the preliminary proposal, including the PDF file's page structure, chapter hierarchy, and embedded comparison tables. It then identified the corresponding explanatory text for each page, chapter, and table, such as the coverage explanation, demonstration logic explanation, and customer information. It mapped document units to the supporting content using tag information and layout structure. The semantic analysis module then processed the extracted content. Based on the set of illegal expressions in the pre-set compliance policy library, fuzzy description features, and missing prompts, it identified high-risk expressions such as "guaranteed benefits" and "guaranteed return." It also identified ambiguous terms such as "potentially increased benefits," as well as the lack of descriptions of "risk warnings" and "insurance liability exemptions" in the product liability section. Based on the analysis results, the intelligent correction module sets priorities according to the severity of the illegal statements, lists "guaranteed returns" as high-risk items, and arranges for priority replacement; calls the synonym replacement table in the strategy library to replace "return guarantee" with "return plan"; calls the pre-trained model to generate multiple semantically consistent but more clear and compliant expressions for "fuzzy sentences", and selects candidate sentences without sensitive words to update the original text; for missing prompt information, the system identifies the end of the product liability description paragraph and inserts standard risk warning templates such as "This product is a long-term insurance product and does not guarantee the principal and returns." After the correction is completed, the format reconstruction module loads the financial industry report template, adjusts the paragraph indentation and title style of the structured document (for example, the first-level title is 14-point Songti bold and centered), corrects the confusion of chapter numbering, and outputs a final DOCX file with a unified format and compliant content for the marketing team to send to the compliance department for final archiving and filing.

[0188] Similarly, consider an internet hospital that generates personalized chronic disease management recommendations based on user health questionnaires and backend algorithms, including medication usage, dietary advice, and risk warnings. The initially generated document is presented as a structured plan document with supplementary explanations, with the AIGC tool automatically generating the basic content. This type of document must be submitted for medical ethics and compliance review before it can be officially pushed to users. The system first extracts sections from the structured document content, such as "Diagnostic Analysis," "Recommended Treatment Plan," and embedded blood glucose data tables, and simultaneously establishes a corresponding relationship with the semantic explanation content, such as a textual description of dietary recommendations and a concise explanation of the efficacy mechanism. After the semantic analysis module is activated, it loads the medical compliance policy library and identifies high-risk expressions such as "significantly improve the condition" and "ensure recovery" as non-evidence-based, and inaccurate descriptions such as "likely effective" and "improvement tendency." The system also discovered that some recommendation paragraphs lacked the source of the prescription and the basis for the drug's indication. During the correction phase, the system sets the modification order based on the risk level, giving priority to replacing illegal expressions and calling standard medical expression templates. For example, it replaces "significantly improved the condition" with "improvement was observed in some patients in clinical studies"; calls the language model to rewrite the vague text, selects candidate sentences that comply with the principle of medical prudence to complete the replacement; and adds standard prompts such as "Recommended drugs are formulated in accordance with the "Guidelines for the Prevention and Treatment of Type 2 Diabetes in China". Please follow the doctor's instructions for specific medication" at the end of the prescription instructions. Finally, the format reconstruction module loads the electronic medical record standard template, unifies the indentation of all paragraphs to 2 characters, unifies the title font to bold size 14, adjusts the numbering structure to standard styles such as "1.1, 1.2", and outputs it in standard PDF format for compliance, medical ethics committee review, and platform filing before being pushed to the patient end.

[0189] This embodiment uses a format reconstruction module to ensure that after semantic correction and content revision, the text content is not only compliant, but also meets the standardization requirements for structural display. It is particularly suitable for scenarios with high requirements for format consistency, such as regulatory approval, customer delivery, and archiving. Through template comparison and structure-level rewriting operations, the system can automatically identify and correct inconsistencies in content formats from different sources, reduce manual typesetting workload, improve content review pass rate and display clarity, and avoid misunderstandings or document returns due to format confusion. At the same time, the module supports template adaptation and output format switching for multiple industries, and has good industry scalability and deployment flexibility.

[0190] In one embodiment, an intelligent correction device based on semantic analysis is provided, and the intelligent correction device based on semantic analysis corresponds one-to-one to the intelligent correction method based on semantic analysis in the above embodiment. Figure 3 , Figure 3This is a functional module diagram of a preferred embodiment of the intelligent correction device based on semantic analysis of the present invention. It includes a structured extraction module 10, a semantic analysis module 20, an intelligent correction module 30, a format reconstruction module 40, and a compliance output module 50. Each functional module is described in detail as follows:

[0191] A structured extraction module 10 is used to extract the structured document content in the solution content and extract the auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0192] A semantic analysis module 20 is configured to perform semantic understanding processing on the structured document content and the auxiliary explanation text through a semantic analysis module to generate a semantic analysis result;

[0193] An intelligent correction module 30, configured to perform correction operations on the structured document content and the auxiliary explanation text through an intelligent correction module according to the semantic analysis result;

[0194] The format reconstruction module 40 is used to input the structured document content and auxiliary explanation text that have completed the correction operation into the format reconstruction module to generate a compliant format document;

[0195] The compliance output module 50 is used to output the compliance format document.

[0196] In one embodiment, the structured extraction module 10 is specifically configured to:

[0197] Extracting structured document content from the solution content, wherein the structured document content includes document page units, chapter units, and table units;

[0198] parsing the document format of the structured document content according to the identification tags of the document page unit, chapter unit, and table unit;

[0199] Based on the parsed document format, a mapping relationship table between each unit in the structured document content and the auxiliary commentary text is established in the commentary data source;

[0200] Extracting, from the explanation data source, page explanation text corresponding to each document page unit, chapter summary text corresponding to each chapter unit, and data annotation text corresponding to each table unit according to the mapping relationship table;

[0201] The page explanation text, chapter summary text and data annotation text are classified and integrated according to corresponding document page units, chapter units and table units to generate a structured auxiliary explanation text set.

[0202] In one embodiment, the semantic analysis module 20 is specifically configured to:

[0203] Loading a preset compliance policy library, which includes a set of illegal statements, an inaccurate description feature library, and a missing annotation policy;

[0204] Performing word segmentation processing on the structured document content and the auxiliary explanation text to generate semantically related context sentence fragments;

[0205] According to the set of illegal expressions, matching illegal expressions in the context sentence fragments and marking the position and type of the illegal expressions;

[0206] Based on the inaccurate description feature library, identifying ambiguous terms and misleading expressions in the context sentence fragments, and generating an inaccurate description list;

[0207] According to the missing annotation strategy, detect the unannotated product providers, data sources and risk warnings in the structured document content, and generate a missing annotation list;

[0208] The semantic analysis result is generated by integrating the location and type of the illegal expressions, the list of inaccurate descriptions and the list of missing annotations.

[0209] In one embodiment, the intelligent correction module 30 is specifically configured to:

[0210] Classify the illegal expressions into high-risk, medium-risk, and low-risk priority levels based on the severity of the illegal expressions in the semantic analysis results, and assign the highest correction priority to the illegal expressions of the high-risk priority level;

[0211] generating a corrective action execution queue in the order of the high-risk, medium-risk and low-risk priority levels;

[0212] detecting whether there are conflicting tasks in the correction operation execution queue for executing multiple correction operations on the same structured document unit or the same auxiliary commentary text, deduplicating or merging the conflicting tasks, and generating an optimized correction operation execution queue;

[0213] According to the optimized correction operation execution queue, correction operations are performed on the structured document content and the auxiliary explanation text through the intelligent correction module.

[0214] In one embodiment, the intelligent correction module 30 is specifically configured to:

[0215] Extracting a synonym replacement table for illegal expressions in the semantic analysis results from a preset compliance policy library through an intelligent correction module;

[0216] Based on the grammatical structure of the structured document content and the auxiliary explanation text, matching candidate replacement words that meet current grammatical requirements from the synonym replacement table;

[0217] Synchronously replacing the illegal expressions in the structured document content and the auxiliary commentary text with the candidate replacement words to generate replaced structured document content and auxiliary commentary text;

[0218] Extracting contextual semantic vectors of misleading expressions and ambiguous terms in the replaced structured document content and auxiliary commentary text based on the list of inaccurate descriptions in the semantic analysis result;

[0219] Generate multiple candidate rewriting sentences based on the context semantic vector by calling the pre-trained language model through the intelligent correction module;

[0220] Filtering candidate sentences without sensitive words from the rewriting candidate sentences according to the compliance policy library to generate a set of compliant rewriting sentences;

[0221] Replacing misleading expressions and ambiguous terms in the structured document content and auxiliary explanatory text with statements in the set of compliant rewritten statements to generate updated structured document content and auxiliary explanatory text;

[0222] Based on the missing annotations in the semantic analysis results, the intelligent correction module locates the end of the product description paragraph in the updated structured document content and the auxiliary commentary text;

[0223] The intelligent correction module matches the standard prompt template of the current product type from the compliance policy library, and inserts the standard prompt template into the end of the paragraph to generate structured document content and auxiliary explanation text that complete the correction operation.

[0224] In one embodiment, the intelligent correction module 30 is specifically configured to:

[0225] Load the prohibited content feature library in the preset compliance policy library through the prohibited content processing module;

[0226] Performing semantic similarity analysis based on the prohibited content feature library to match prohibited content paragraphs in the structured document content and the auxiliary commentary text;

[0227] Deleting the prohibited content paragraphs in the structured document content and the auxiliary commentary text, recording the deleted position of the prohibited content paragraph, and extracting the adjacent paragraphs of the deleted position as the associated context;

[0228] Invoking a pre-trained language model to generate a coherent rewritten text based on the associated context;

[0229] The coherent rewritten text is inserted into the deleted position recorded in the structured document content and the auxiliary commentary text to generate the structured document content and the auxiliary commentary text that have completed the correction operation.

[0230] In one embodiment, the format reconstruction module 40 is specifically configured to:

[0231] Load the format template set in the preset compliance policy library through the format reconstruction module;

[0232] Parsing the current format elements of the structured document content and the auxiliary commentary text through a format reconstruction module;

[0233] Detecting format differences between the current format element and a target industry template in the format template set by a format reconstruction module;

[0234] Adjusting, by a format reconstruction module, paragraph spacing, title font style, and numbering sequence rules of the structured document content and the auxiliary explanatory text based on the format difference items, so that the structured document content and the auxiliary explanatory text match the format rules of the target industry template;

[0235] The adjusted structured document content and auxiliary explanatory text are output as a compliant format document through the format reconstruction module.

[0236] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external user terminal via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the server side of an intelligent correction method based on semantic analysis.

[0237] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps of a user-side intelligent correction method based on semantic analysis.

[0238] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0239] Extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0240] Performing semantic understanding processing on the structured document content and the auxiliary explanation text through a semantic analysis module to generate a semantic analysis result;

[0241] According to the semantic analysis results, a correction operation is performed on the structured document content and the auxiliary explanation text by an intelligent correction module;

[0242] Input the structured document content and auxiliary explanatory text that have completed the correction operation into the format reconstruction module to generate a compliant format document;

[0243] Output the document in the compliant format.

[0244] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0245] Extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source;

[0246] Performing semantic understanding processing on the structured document content and the auxiliary explanation text through a semantic analysis module to generate a semantic analysis result;

[0247] According to the semantic analysis results, a correction operation is performed on the structured document content and the auxiliary explanation text by an intelligent correction module;

[0248] Input the structured document content and auxiliary explanatory text that have completed the correction operation into the format reconstruction module to generate a compliant format document;

[0249] Output the document in the compliant format.

[0250] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0251] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0252] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0253] It should be noted that if any software tools or components other than those of the Company appear in the embodiments of this application, they are merely for illustration and do not represent actual use. The above embodiments are intended only to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some of the technical features therein with equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. An intelligent correction method based on semantic analysis, characterized in that: The following steps are involved: Extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source; Performing semantic understanding processing on the structured document content and the auxiliary explanation text through a semantic analysis module to generate a semantic analysis result; According to the semantic analysis results, a correction operation is performed on the structured document content and the auxiliary explanation text by an intelligent correction module; Input the structured document content and auxiliary explanatory text that have completed the correction operation into the format reconstruction module to generate a compliant format document; Output the document in the compliant format.

2. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: Extracting structured document content from the solution content, and extracting auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source, including: Extracting structured document content from the solution content, wherein the structured document content includes document page units, chapter units, and table units; parsing the document format of the structured document content according to the identification tags of the document page unit, chapter unit, and table unit; Based on the parsed document format, a mapping relationship table between each unit in the structured document content and the auxiliary commentary text is established in the commentary data source; Extracting, from the explanation data source, page explanation text corresponding to each document page unit, chapter summary text corresponding to each chapter unit, and data annotation text corresponding to each table unit according to the mapping relationship table; The page explanation text, chapter summary text and data annotation text are classified and integrated according to corresponding document page units, chapter units and table units to generate a structured auxiliary explanation text set.

3. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: The semantic analysis module performs semantic understanding processing on the structured document content and the auxiliary explanation text to generate semantic analysis results, including: Loading a preset compliance policy library, which includes a set of illegal statements, an inaccurate description feature library, and a missing annotation policy; Performing word segmentation processing on the structured document content and the auxiliary explanation text to generate semantically related context sentence fragments; According to the set of illegal expressions, matching illegal expressions in the context sentence fragments and marking the position and type of the illegal expressions; Based on the inaccurate description feature library, identifying ambiguous terms and misleading expressions in the context sentence fragments, and generating an inaccurate description list; According to the missing annotation strategy, detect the unannotated product providers, data sources and risk warnings in the structured document content, and generate a missing annotation list; The semantic analysis result is generated by integrating the location and type of the illegal expressions, the list of inaccurate descriptions and the list of missing annotations.

4. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: According to the semantic analysis results, the intelligent correction module performs correction operations on the structured document content and the auxiliary explanation text, including: Classify the illegal expressions into high-risk, medium-risk, and low-risk priority levels based on the severity of the illegal expressions in the semantic analysis results, and assign the highest correction priority to the illegal expressions of the high-risk priority level; generating a corrective action execution queue in the order of the high-risk, medium-risk and low-risk priority levels; detecting whether there are conflicting tasks in the correction operation execution queue for executing multiple correction operations on the same structured document unit or the same auxiliary commentary text, deduplicating or merging the conflicting tasks, and generating an optimized correction operation execution queue; According to the optimized correction operation execution queue, correction operations are performed on the structured document content and the auxiliary explanation text through the intelligent correction module.

5. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: According to the semantic analysis results, the intelligent correction module performs correction operations on the structured document content and the auxiliary explanation text, including: Extracting a synonym replacement table for illegal expressions in the semantic analysis results from a preset compliance policy library through an intelligent correction module; Based on the grammatical structure of the structured document content and the auxiliary explanation text, matching candidate replacement words that meet current grammatical requirements from the synonym replacement table; Synchronously replacing the illegal expressions in the structured document content and the auxiliary commentary text with the candidate replacement words to generate replaced structured document content and auxiliary commentary text; Extracting contextual semantic vectors of misleading expressions and ambiguous terms in the replaced structured document content and auxiliary commentary text based on the list of inaccurate descriptions in the semantic analysis result; Generate multiple candidate rewriting sentences based on the context semantic vector by calling the pre-trained language model through the intelligent correction module; Filtering candidate sentences without sensitive words from the rewriting candidate sentences according to the compliance policy library to generate a set of compliant rewriting sentences; Replacing misleading expressions and ambiguous terms in the structured document content and auxiliary explanatory text with statements in the set of compliant rewritten statements to generate updated structured document content and auxiliary explanatory text; Based on the missing annotations in the semantic analysis results, the intelligent correction module locates the end of the product description paragraph in the updated structured document content and the auxiliary commentary text; The intelligent correction module matches the standard prompt template of the current product type from the compliance policy library, and inserts the standard prompt template into the end of the paragraph to generate structured document content and auxiliary explanation text that complete the correction operation.

6. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: After performing correction operations on the structured document content and the auxiliary explanation text by the intelligent correction module according to the semantic analysis results, the method further includes: Load the prohibited content feature library in the preset compliance policy library through the prohibited content processing module; Performing semantic similarity analysis based on the prohibited content feature library to match prohibited content paragraphs in the structured document content and the auxiliary commentary text; Deleting the prohibited content paragraphs in the structured document content and the auxiliary commentary text, recording the deleted position of the prohibited content paragraph, and extracting the adjacent paragraphs of the deleted position as the associated context; Invoking a pre-trained language model to generate a coherent rewritten text based on the associated context; The coherent rewritten text is inserted into the deleted position recorded in the structured document content and the auxiliary commentary text to generate the structured document content and the auxiliary commentary text that have completed the correction operation.

7. The intelligent correction method based on semantic analysis according to claim 1, characterized in that: The structured document content and auxiliary explanatory text that have completed the correction operation are input into the format reconstruction module to generate a compliant format document, including: Load the format template set in the preset compliance policy library through the format reconstruction module; Parsing the current format elements of the structured document content and the auxiliary commentary text through a format reconstruction module; Detecting format differences between the current format element and a target industry template in the format template set by a format reconstruction module; Adjusting, by a format reconstruction module, paragraph spacing, title font style, and numbering sequence rules of the structured document content and the auxiliary explanatory text based on the format difference items, so that the structured document content and the auxiliary explanatory text match the format rules of the target industry template; The adjusted structured document content and auxiliary explanatory text are output as a compliant format document through the format reconstruction module.

8. An intelligent correction device based on semantic analysis, characterized in that: The intelligent correction device based on semantic analysis includes: A structured extraction module, configured to extract structured document content from the solution content and extract auxiliary explanation text corresponding to each unit in the structured document content from the explanation data source; A semantic analysis module, configured to perform semantic understanding processing on the structured document content and the auxiliary explanation text through the semantic analysis module to generate a semantic analysis result; An intelligent correction module, configured to perform correction operations on the structured document content and the auxiliary explanation text according to the semantic analysis result through the intelligent correction module; A format reconstruction module is used to input the structured document content and auxiliary explanation text that have completed the correction operation into the format reconstruction module to generate a compliant format document; The compliance output module is used to output the compliance format document.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and an intelligent correction program based on semantic analysis stored in the memory and capable of running on the processor. When the intelligent correction program based on semantic analysis is executed by the processor, the steps of the intelligent correction method based on semantic analysis as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores an intelligent correction program based on semantic analysis, which, when executed by a processor, implements the steps of the intelligent correction method based on semantic analysis as described in any one of claims 1 to 7.

Images