A method and system for semantic processing of enterprise content for generative engines

By uniformly parsing and semantically reconstructing multi-source unstructured content of enterprises, AI-readable semantic units are generated, solving the problems of ambiguous semantic boundaries and insufficient quantitative assessment of reference value in existing technologies. This achieves high-quality semantic unit conversion and generative engine adaptation, improving the content understanding and reference accuracy of generative AI.

CN123070122APending Publication Date: 2026-11-10BEIJING YINGDA TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611264432.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to identify semantic boundaries and reconstruct semantics from multi-source unstructured content from enterprises, resulting in insufficient understanding of enterprise content, inaccurate citations, and poor response stability by generative AI platforms.

Method used

By uniformly parsing unstructured content from multiple sources within an enterprise, a set of candidate semantic fragments is constructed, semantic reconstruction is performed, and AI-readable semantic units are generated. Through generative engine adaptation and feedback optimization, accurate identification and high-value screening of semantic units are achieved.

Benefits of technology

It achieves effective extraction of factual fragments and accurate identification of semantic boundaries in unstructured content, improves the independence and self-containment of semantic units, realizes quantitative evaluation of the reference value of AI-readable semantic units and dynamic adaptation of generative engines, and improves the ability of generative engines to accurately identify and understand content.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070122A_ABST
    Figure CN123070122A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of natural language processing technology, specifically a semantic processing method and system for enterprise content for generative engines. The method includes: unified parsing, cleaning, and entity recognition of multi-source unstructured content; analysis of topic, entity, and contextual relationships to identify semantic boundaries and segment the content; construction of candidate semantic segments; content recognition, information density analysis, and semantic integrity assessment of the candidate segments; semantic reconstruction based on the assessment results; calculation of citation scores to filter segments with citation value; construction of an AI-readable semantic unit set according to a unified structure; acquisition of target engine cognitive features for structured expression reorganization; determination of semantic defect locations through question-and-answer verification and feedback optimization until the engine's recognition requirements are met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically a semantic processing method and system for enterprise content oriented towards generative engines. Background Technology

[0002] With the development of generative artificial intelligence and AI search technology, users are increasingly shifting from traditional search engine retrieval to directly querying generative AI platforms for answers. Company websites, product manuals, technical documents, design drawings, case studies, FAQs, and internal knowledge bases have become crucial information sources for generative AI in knowledge retrieval, content understanding, and answer generation. To improve the online visibility of company information, current technologies typically employ methods such as webpage editing, keyword optimization, FAQ creation, press release distribution, and SEO optimization to organize and promote company content, thereby enhancing the search ranking and exposure of company information in traditional search engines.

[0003] However, existing technologies primarily cater to human reading and traditional search engine content organization methods. A large amount of enterprise content remains scattered in unstructured formats such as PDF files, technical documents, design drawings, parameter tables, and images. Furthermore, individual content fragments often contain multiple semantic information simultaneously, including background descriptions, parameter explanations, conclusions, and evidence citations. This results in blurred semantic boundaries and fragmented factual expression, making it difficult to form independent semantic units that can be directly recognized, understood, and referenced by generative AI. Consequently, generative AI platforms suffer from insufficient understanding of enterprise content, inaccurate citations, and poor response stability. Therefore, how to identify semantic boundaries and reconstruct semantic units from multi-source unstructured enterprise content to generate AI-readable content suitable for generative engines to understand and reference has become a pressing technical problem in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for semantic processing of enterprise content for generative engines. This addresses the difficulties in existing technologies in identifying semantic boundaries and reconstructing semantics for multi-source unstructured enterprise content, thereby generating AI-readable semantic units that can be accurately identified, understood, and referenced by generative engines.

[0005] A semantic processing method for enterprise content oriented towards generative engines includes the following steps: uniformly parsing multi-source unstructured content of the enterprise and constructing a candidate semantic fragment set, specifically:

[0006] Collect heterogeneous content data from multiple sources of enterprises. For different types of raw content data, use corresponding parsing methods for unified parsing processing. Standardize the parsed text content and establish a unified content expression set. Perform content cleaning processing, entity recognition processing on the cleaned text content, extract key semantic entities from the text, construct entity sets, perform topic change analysis, entity change analysis, and context association analysis on each text content to identify semantic boundaries, continuously segment the unified text content, and construct a candidate semantic segment set.

[0007] The candidate semantic fragments are semantically reconstructed to generate AI-readable semantic units, specifically:

[0008] The candidate semantic fragments are subjected to content recognition, information density analysis, and semantic integrity assessment. Based on the semantic integrity assessment results, semantic reconstruction processing is performed, and the citation score of the current candidate semantic fragment is calculated. Based on the calculation results, it is determined whether the current candidate semantic fragment has citation value for the generative engine, and AI-readable semantic units are constructed. All constructed AI semantic units are organized in a unified format to construct a set of AI-readable semantic units.

[0009] The AI-readable semantic units are adapted to a generative engine, and iterative optimization is performed based on the feedback results, specifically as follows:

[0010] The content cognitive features of the target generative engine are obtained. Based on the content cognitive features, the AI-readable semantic units are restructured and reorganized into a structured expression. The generative engine is verified by question-and-answer testing using a set of test questions. Based on the answer results, the semantic defect locations of the AI-readable semantic units are determined, and the AI-readable semantic units are optimized by feedback until they meet the recognition requirements of the generative engine.

[0011] Preferably, the unified parsing process is performed as follows:

[0012] Utilize the enterprise content access module to acquire multi-source content from both internal and external sources in real time;

[0013] For different types of raw content data, corresponding parsing methods are used for unified parsing processing, specifically:

[0014] If the current content type is a text document, then perform text extraction processing to generate the corresponding text content;

[0015] If the current content type is image data, then the optical character recognition algorithm is used to perform text recognition processing to generate the corresponding OCR text content;

[0016] If the current content type is video or audio data, then a speech recognition algorithm is used to perform speech-to-text processing to generate the corresponding speech-text content;

[0017] If the current content type is web page data, then the web page structure is parsed using a web page text extraction algorithm to remove the navigation bar, advertising areas, and irrelevant links, and generate the corresponding web page text content.

[0018] Preferably, the content cleaning process is performed as follows:

[0019] Set a threshold for determining duplicate content And calculate the semantic similarity between any two text contents. , Representing text With text Semantic similarity between them;

[0020] The semantic similarity between two pieces of text is compared with a set threshold for determining duplicate content. The comparison result is used to determine whether the text content is duplicated. Specifically:

[0021] If the comparison results satisfy the formula If so, the corresponding text content is determined to be duplicate content, and the duplicate text is deleted;

[0022] If the comparison results satisfy the formula If so, the current text content will be retained;

[0023] Entity recognition processing is performed on the cleaned text content to extract key semantic entities and construct an entity set. ,in, Indicates the first A semantic entity, Indicates the total number of semantic entities.

[0024] Preferably, the identification of semantic boundaries is specifically as follows:

[0025] Calculate the topic feature vector of the current text segment and the degree of thematic difference between it and adjacent text fragments. Simultaneously calculate the degree of entity change between adjacent text segments. Calculate the contextual association between text fragments And set a semantic boundary determination threshold. ;

[0026] The calculated contextual relevance between text segments is compared with a set semantic boundary determination threshold. Based on the comparison result, it is determined whether the current text content belongs to the same semantic unit. Specifically:

[0027] If the comparison results satisfy the formula If a semantic boundary exists at the current text position, the current text content will be segmented.

[0028] If the comparison results satisfy the formula If so, it is determined that the current text content still belongs to the same semantic unit.

[0029] Preferably, the information density analysis is as follows:

[0030] Statistics Number of fact entities in each candidate semantic segment and the total number of words in the text ;

[0031] And calculate the information density value of the candidate semantic fragments. ,in, Indicates the first The information density of each candidate semantic fragment.

[0032] Set minimum information density threshold The system compares the set minimum information density threshold with the calculated information density value, and determines whether the current candidate semantic segment contains invalid information based on the comparison result. Specifically:

[0033] If the comparison results satisfy the formula If so, it is determined that the current candidate semantic fragment contains marketing description information or invalid explanatory information, and a low-value identifier is generated. The current candidate semantic fragments do not participate in the construction of subsequent AI-readable semantic units;

[0034] If the comparison results satisfy the formula If no marketing description information or invalid explanatory information is found in the current candidate semantic fragment, a high-value identifier is generated. The current candidate semantic fragments participate in the construction of subsequent AI-readable semantic units.

[0035] Preferably, the semantic completeness assessment is as follows:

[0036] Count the number of fact entities in the current candidate semantic segments respectively. and the number of context dependencies Calculate the semantic completeness of the current candidate semantic segment. ;

[0037] Set semantic integrity threshold The set semantic completeness threshold is compared with the calculated semantic completeness, specifically:

[0038] If the comparison results satisfy the formula If so, it is determined that the current candidate semantic segment has semantic missing or strong contextual dependency, and a semantic reconstruction identifier is generated. ;

[0039] If the comparison results satisfy the formula If the current candidate semantic segment does not have semantic missing or strong contextual dependency, a semantic reconstruction identifier is generated. .

[0040] Preferably, the step of determining whether the current candidate semantic fragment has reference value for the generative engine based on the calculation results is as follows:

[0041] A mapping relationship between candidate semantic fragments and original content is established based on the constructed content index relationship. ,in, This indicates the constructed content index relationship. This represents the candidate semantic fragments after semantic reconstruction, and simultaneously constructs a semantic evidence chain. The evidence completeness score is calculated based on the source completeness, location traceability, version validity, and relational completeness in the semantic evidence chain. And calculate the citation score of the current candidate semantic fragment. ;

[0042] Set referenceability threshold The set referability threshold is compared with the calculated referability score. Based on the comparison result, it is determined whether the current candidate semantic fragment has referential value for the generative engine. Therefore:

[0043] If the comparison results satisfy the formula If the current candidate semantic fragment is not considered valid, an invalid identifier is generated. ;

[0044] If the comparison results satisfy the formula If so, the current candidate semantic fragment is determined to have reference value for the generative engine, and a valid identifier is generated. .

[0045] Preferably, the construction of the AI-readable semantic unit set is as follows:

[0046] For valid identifiers that satisfy the formula The candidate semantic fragments are organized according to a unified structure to construct AI-readable semantic units. ;

[0047] If all the constructed AI semantic units are organized according to a unified format to form a set of AI-readable semantic units, then:

[0048] ;

[0049] in, This represents the constructed set of AI-readable semantic units;

[0050] Set minimum field completeness threshold Calculate the first Field completeness of each AI-readable semantic unit ,in, This indicates the number of valid fields in the current semantic unit. Valid fields are data fields that are not empty and pass format validation. Indicates the total number of preset fields. Indicates the completeness of the calculated fields;

[0051] The set minimum field completeness threshold is compared with the calculated field completeness. Based on the comparison result, it is determined whether the current AI-readable semantic unit is valid. Specifically:

[0052] If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be invalid, and an invalidation flag is generated. ;

[0053] If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be valid, and a valid identifier is generated. ;

[0054] The set of AI-readable semantic units and their corresponding validity identifiers together form the AI ​​semantic processing data set. ,in, This represents a dataset used for AI semantic processing.

[0055] Preferably, the step of determining the semantic defect location of the AI-readable semantic unit based on the answer result is as follows:

[0056] Calculate the reference matching degree of AI-readable semantic units ,in, This indicates the number of pieces of information in the answer that correctly reference the current AI-readable semantic unit. R represents the expected number of references for the current test question. This indicates the reference matching degree of the current AI-readable semantic unit;

[0057] Set reference matching threshold The set reference matching threshold is compared with the calculated reference matching degree. Based on the comparison result, it is determined whether the current AI-readable semantic unit can be effectively recognized by the generative engine. Specifically:

[0058] If the comparison results satisfy the formula If so, it is determined that the current AI-readable semantic unit has a semantic expression deviation, and an optimization identifier is generated. ;

[0059] If the comparison results satisfy the formula R If so, it is determined that the current AI-readable semantic unit can be effectively recognized by the generative engine and a valid identifier can be generated. ;

[0060] For optimized identifiers, the formula is satisfied. The AI-readable semantic unit performs feedback optimization processing, specifically as follows:

[0061] By analyzing the missing information, misquoted information, and low-confidence information in the generative engine's response results, the location of semantic defects in the current AI-readable semantic units can be determined. ,in, Indicates the first A set of semantic defects corresponding to each AI-readable semantic unit.

[0062] A semantic processing system for enterprise content for generative engines includes: a unified parsing module for multi-source content, a semantic reconstruction module, and a generative engine adaptation module.

[0063] The multi-source content unified parsing module is used to perform unified parsing of multi-source unstructured content of enterprises and construct a set of candidate semantic fragments;

[0064] The semantic reconstruction module is used to perform semantic reconstruction on candidate semantic fragments and generate AI-readable semantic units;

[0065] The generative engine adaptation module is used to adapt AI-readable semantic units to a generative engine and iteratively optimize them based on feedback results.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention performs unified parsing and semantic boundary identification on multi-source unstructured content of enterprises. Based on a comprehensive judgment of topic difference, entity change degree and contextual relevance, it continuously segments unified text content and constructs a set of candidate semantic segments. This solves the technical problem in the prior art that the semantic boundaries of multi-source heterogeneous content are vague and difficult to segment into independent semantic units. It realizes the effective extraction of factual fragments and accurate identification of semantic boundaries in unstructured content.

[0068] 2. This invention analyzes the information density and assesses the semantic completeness of candidate semantic segments. It calculates information density based on the ratio of the number of factual entities to the total number of words in the text to filter low-value segments. At the same time, it constructs a semantic completeness calculation model based on the number of factual entities and the number of contextual dependencies to identify segments with semantic deficiencies or strong contextual dependencies and performs splitting, merging, and recombination processing. This solves the technical problems in the prior art where background descriptions and factual information of a single content segment are mixed, factual expressions are scattered, and it is difficult to form independent factual units. It realizes the accurate identification and semantic reconstruction of effective factual information in candidate semantic segments and improves the independence and self-containment of semantic units.

[0069] 3. This invention calculates the citation value score of candidate semantic segments, constructs a multi-dimensional quantitative evaluation model based on information density value, semantic integrity, evidence integrity score and timeliness score, and compares it with a set threshold to determine whether the candidate semantic segments have citation value for generative engines. This solves the technical problem in the prior art of lacking a means to quantitatively evaluate the citation value of semantic units and the difficulty in screening high-value factual information, and realizes the objective quantitative evaluation and screening of the citation value of AI-readable semantic units.

[0070] 4. This invention constructs a unified AI-readable semantic unit structure and sets up a field integrity detection mechanism. It organizes semantic units according to a unified structure that includes semantic unit number, subject information, topic information, core fact information, parameter information, evidence information, application scenario information, and relationship information. It also calculates the field integrity to verify the validity of the semantic units. This solves the technical problems in the prior art where enterprise content lacks a unified AI-readable organizational form and generative engines have difficulty in directly recognizing and referencing it. It achieves a high-quality conversion of enterprise content into structured and standardized AI-readable semantic units.

[0071] 5. This invention acquires the content cognitive features of generative engines and restructures them into a structured expression. It combines question-and-answer verification and feedback optimization to form a closed-loop iterative mechanism. Based on the engine's content organization preferences, citation preferences, and semantic expression preferences, it generates various adapted expression forms such as question-and-answer content, knowledge cards, or structured web pages. It verifies the citation matching degree through a set of test questions and performs targeted information supplementation, association reconstruction, evidence enhancement, and order adjustment optimization operations based on the location of semantic defects. This solves the technical problem in the prior art that static content is difficult to adapt to different generative engine cognitive methods and cannot be continuously optimized. It realizes the dynamic adaptation and continuous optimization of AI-readable semantic units with different generative engine cognitive methods, and improves the ability of enterprise content to be accurately identified, understood, and cited by generative engines. Attached Figure Description

[0072] Figure 1This is a schematic diagram illustrating the overall method steps of a semantic processing method for enterprise content oriented towards generative engines according to the present invention.

[0073] Figure 2 This is a schematic diagram of the overall structure of an enterprise content semantic processing system for generative engines according to the present invention. Detailed Implementation

[0074] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0075] Example 1

[0076] Reference Figure 1 As an embodiment of the present invention, a method for semantic processing of enterprise content for generative engines is provided, comprising the following steps:

[0077] S1: Perform unified parsing of multi-source unstructured content of enterprises and construct a set of candidate semantic fragments.

[0078] Specifically, constructing a candidate semantic fragment set involves collecting multi-source heterogeneous content data from enterprises, uniformly parsing content from different sources and in different formats, constructing a unified text content set based on the parsed content, and further performing semantic boundary recognition to generate a candidate semantic fragment set. The specific implementation is as follows:

[0079] Utilizing the enterprise content access module, real-time acquisition of multi-source content from both internal and external sources is achieved. Enterprise content includes at least one or more of the following: company website pages, product manuals, PDF documents, technical white papers, images, video subtitles, FAQ documents, case studies, and internal knowledge bases. Specifically:

[0080] Obtain the first The original content data corresponding to the enterprise-type content source at the current collection time. ,in, Indicates the first The original content data corresponding to the enterprise-type content source. , This indicates the total number of content sources for the enterprise;

[0081] Synchronously acquire the content attribute information corresponding to the original content data Content attribute information should include at least one or more of the following: content source, content type, publication time, version information, and the business domain to which it belongs, to construct a content attribute set. ,in, Indicates the source information of the content. Indicates content type information, Indicates the publication time information. Indicates version information, Indicates business domain information, Indicates the first The collection of content attributes corresponding to the class content.

[0082] For different types of raw content data, corresponding parsing methods are used for unified parsing processing, as follows:

[0083] If the current content type is a text document, then perform text extraction processing to generate the corresponding text content. , Indicates the first The text content obtained from parsing a text-like document;

[0084] If the current content type is image data, then the optical character recognition algorithm is used to perform text recognition processing to generate the corresponding OCR text content. , Indicates the first Text content obtained from image content recognition;

[0085] If the current content type is video or audio data, then a speech recognition algorithm is used to perform speech-to-text processing to generate the corresponding speech-text content. , Indicates the first Text content obtained by transcribing audio and video content;

[0086] If the current content type is webpage data, then the webpage structure is parsed using a webpage text extraction algorithm to remove navigation bars, advertising areas, and irrelevant links, generating the corresponding webpage text content. , Indicates the first The main text obtained by extracting content from web pages.

[0087] To eliminate the differences in expression between different content sources and data formats, the parsed text content is standardized to establish a unified content representation set. ,in, Indicates the first A unified text content, This represents the unified collection of text content that has been constructed. Indicates the total amount of uniform text content;

[0088] Meanwhile, to ensure subsequent semantic tracing capabilities, a corresponding content index relationship is established for each unified text content. ,in, Indicates the source of the content. Indicates the content location identifier. Indicates the time of content generation. Indicates the content version identifier. Indicates the first The index relationship corresponding to a unified text content.

[0089] To improve the accuracy of subsequent semantic boundary recognition, content cleaning processing is performed on the unified text content set, as follows:

[0090] Set a threshold for determining duplicate content And calculate the semantic similarity between any two text contents. , Representing text With text Semantic similarity between them;

[0091] The semantic similarity between two pieces of text is compared with a set threshold for determining duplicate content. The comparison result is used to determine whether the text content is duplicated. Specifically:

[0092] If the comparison results satisfy the formula If so, the corresponding text content is determined to be duplicate content, and the duplicate text is deleted;

[0093] If the comparison results satisfy the formula If so, the current text content will be preserved.

[0094] Entity recognition processing is performed on the cleaned text content to extract key semantic entities and construct an entity set. ,in, Indicates the first A semantic entity, which includes at least one or more of the following: enterprise entity entity, product entity, parameter entity, application scenario entity, qualification entity, and case entity.

[0095] To achieve semantic boundary recognition of text content, topic change analysis, entity change analysis, and context association analysis are performed on each piece of text content, as detailed below:

[0096] Calculate the topic feature vector of the current text segment and the degree of thematic difference between it and adjacent text fragments. ;

[0097] Simultaneously calculating the degree of entity change between adjacent text segments, we have:

[0098] ;

[0099] in, It indicates the degree of entity change between adjacent text segments.

[0100] Simultaneously calculate the contextual relationship between text segments. And set a semantic boundary determination threshold. ;

[0101] The calculated contextual relevance between text segments is compared with a set semantic boundary determination threshold. Based on the comparison result, it is determined whether the current text content belongs to the same semantic unit. Specifically:

[0102] If the comparison results satisfy the formula If a semantic boundary exists at the current text position, the current text content will be segmented.

[0103] If the comparison results satisfy the formula If so, it is determined that the current text content still belongs to the same semantic unit.

[0104] Based on the semantic boundary positions, the unified text content is continuously segmented to construct a candidate semantic segment set. ,in, Indicates the first A number of candidate semantic segments, This represents the constructed set of candidate semantic fragments. This represents the total number of candidate semantic segments.

[0105] To ensure the effectiveness of subsequent AI-readable semantic unit construction, integrity checks are performed on candidate semantic segments, as follows:

[0106] Set minimum semantic integrity threshold Simultaneously calculating the semantic completeness of each candidate semantic segment, we have:

[0107] ;

[0108] in, Indicates the first The number of fact entities in each candidate semantic fragment includes: product name, technical parameters, functional description, application scenario, and qualification evidence. Indicates the first The total number of semantic nodes in each candidate semantic fragment, including: entity nodes, attribute nodes, and relation nodes. Indicates the first The semantic completeness of each candidate semantic fragment;

[0109] The calculated semantic completeness is compared with the set minimum semantic completeness threshold. Based on the comparison result, it is determined whether the current candidate semantic segment has semantic missing parts. Specifically:

[0110] If the comparison results satisfy the formula If the current candidate semantic segment is found to have semantic missing information, a corresponding invalid identifier will be generated. Remove the current candidate semantic fragment;

[0111] If the comparison results satisfy the formula If so, the current candidate semantic segment is determined to be valid, and a corresponding valid identifier is generated. .

[0112] The candidate semantic fragment set and its corresponding validity identifier together form the basic data set for semantic processing. ,in, This represents the basic data set for semantic processing. Represents a set of candidate semantic fragments. This indicates the validity of the candidate semantic fragment.

[0113] S2: Semantically reconstruct candidate semantic fragments to generate AI-readable semantic units.

[0114] Specifically, generating AI-readable semantic units involves performing content recognition, information density analysis, semantic integrity assessment, and evidence chain association on candidate semantic fragments based on the constructed semantic processing foundation dataset. Then, based on semantic integrity and citationability, the candidate semantic fragments are split, merged, and recombined to construct a set of AI-readable semantic units. The specific implementation is as follows:

[0115] For semantic processing, the validity identifiers in the basic dataset satisfy the formula The candidate semantic segments are then subjected to semantic decomposition processing, as follows:

[0116] Content recognition processing is performed on candidate semantic segments to extract key semantic information, including subject information. Product Information Parameter information Scene information and supporting information ;

[0117] Based on the key semantic information extracted from the candidate semantic fragments, a corresponding set of content features is constructed. ,in, Indicates the first A set of content features for each candidate semantic fragment.

[0118] To identify valid factual information in candidate semantic segments, information density analysis is performed on each candidate semantic segment, as detailed below:

[0119] Statistics Number of fact entities in each candidate semantic segment and the total number of words in the text ;

[0120] And calculate the information density value of the candidate semantic fragments. ,in, Indicates the first The information density of each candidate semantic fragment.

[0121] Set minimum information density threshold The system compares the set minimum information density threshold with the calculated information density value, and determines whether the current candidate semantic segment contains invalid information based on the comparison result. Specifically:

[0122] If the comparison results satisfy the formula If so, it is determined that the current candidate semantic fragment contains marketing description information or invalid explanatory information, and a low-value identifier is generated. The current candidate semantic fragments do not participate in the construction of subsequent AI-readable semantic units;

[0123] If the comparison results satisfy the formula If no marketing description information or invalid explanatory information is found in the current candidate semantic fragment, a high-value identifier is generated. The current candidate semantic fragments participate in the construction of subsequent AI-readable semantic units.

[0124] To improve the generative engine's ability to understand content, semantic completeness analysis is performed on candidate semantic segments, and the number of fact entities in each candidate semantic segment is counted. and the number of context dependencies Context dependency refers to the dependency relationship in which a semantic element (including entity, attribute, relation or event) in the current semantic segment cannot be understood independently within the current segment, and must rely on other text content outside the segment (including preceding, following or external reference content) to determine its complete semantics.

[0125] Calculating the semantic completeness of the current candidate semantic segment, we have:

[0126] ;

[0127] in, Indicates the first The semantic completeness of each candidate semantic fragment. , This represents the weighting coefficient, the specific value of which is set by the implementer based on the actual application scenario, and must satisfy the formula. ;

[0128] Set semantic integrity threshold The set semantic completeness threshold is compared with the calculated semantic completeness, specifically:

[0129] If the comparison results satisfy the formula If so, it is determined that the current candidate semantic segment has semantic missing or strong contextual dependency, and a semantic reconstruction identifier is generated. ;

[0130] If the comparison results satisfy the formula If the current candidate semantic segment does not have semantic missing or strong contextual dependency, a semantic reconstruction identifier is generated. .

[0131] For semantic reconstruction identifiers that satisfy the formula For the candidate semantic fragments, perform semantic reconstruction processing as follows:

[0132] If the current candidate semantic segment contains multiple independent topic entities, then the current candidate semantic segment is split according to the position of the topic entity change.

[0133] If the current candidate semantic segment has a cross-segment reference relationship or a nomenclature relationship, then its associated candidate semantic segment is obtained, and semantic merging processing is performed;

[0134] If the current candidate semantic segment has semantic missing information, supplementary information is extracted from the associated text content and semantic reorganization is performed.

[0135] Based on semantic reconstruction processing, a new set of candidate semantic segments is constructed. ,in, This represents the candidate semantic fragment after semantic reconstruction.

[0136] Furthermore, evidence associations are established for each candidate semantic segment, as follows:

[0137] A mapping relationship between candidate semantic fragments and original content is established based on the constructed content index relationship. ,in, This indicates the constructed content index relationship. This represents the candidate semantic fragment after semantic reconstruction.

[0138] Simultaneously constructing a semantic evidence chain, we have:

[0139] ;

[0140] in, Indicates source information, Indicates the original position. Indicates the release time. Indicates version information, Indicates a relationship. This represents the constructed semantic evidence chain.

[0141] The completeness score of evidence is calculated based on the source completeness, location traceability, version validity, and relational completeness in the semantic evidence chain. (This is the weighted sum of all parameters in the semantic evidence chain set. The weights of each parameter are set by the implementers according to the actual application, and the sum of all weight coefficients is 1.)

[0142] The citation score of the current candidate semantic fragment is then:

[0143] ;

[0144] in, Indicates the completeness of evidence score. Indicates timeliness score, , , , These represent the weighting coefficients, The calculated citation score is used to determine whether the current candidate semantic fragment has citation value for the generative engine. Specifically:

[0145] Set referenceability threshold The set referability threshold is compared with the calculated referability score. Based on the comparison result, it is determined whether the current candidate semantic fragment has referential value for the generative engine. Therefore:

[0146] If the comparison results satisfy the formula If the current candidate semantic fragment is not considered valid, an invalid identifier is generated. ;

[0147] If the comparison results satisfy the formula If so, the current candidate semantic fragment is determined to have reference value for the generative engine, and a valid identifier is generated. .

[0148] For valid identifiers that satisfy the formula The candidate semantic fragments are processed using AI semantic unit construction, specifically as follows:

[0149] By organizing the current candidate semantic fragments according to a unified structure and constructing AI-readable semantic units, we have:

[0150] ;

[0151] in, Indicates the semantic unit number. Indicates main information, Indicates topic information, Indicates core factual information, Indicates parameter information, Indicates evidence information, Indicates application scenario information. Indicates relationship information. This represents the constructed AI-readable semantic unit.

[0152] If all the constructed AI semantic units are organized according to a unified format to form a set of AI-readable semantic units, then:

[0153] ;

[0154] in, This represents the constructed set of AI-readable semantic units.

[0155] To ensure the effectiveness of subsequent generative engine adaptation, integrity checks are performed on AI-readable semantic units, as follows:

[0156] Set minimum field completeness threshold ;

[0157] Calculate the first Field completeness of each AI-readable semantic unit ,in, This indicates the number of valid fields in the current semantic unit. Valid fields are data fields that are not empty and pass format validation. This indicates the total number of preset fields. Preset fields include at least: main field, topic field, core fact field, parameter field, evidence field, scenario field, and related fields. Indicates the completeness of the calculated fields;

[0158] The set minimum field completeness threshold is compared with the calculated field completeness. Based on the comparison result, it is determined whether the current AI-readable semantic unit is valid. Specifically:

[0159] If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be invalid, and an invalidation flag is generated. ;

[0160] If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be valid, and a valid identifier is generated. .

[0161] The set of AI-readable semantic units and their corresponding validity identifiers together form the AI ​​semantic processing data set. ,in, This represents a dataset used for AI semantic processing.

[0162] S3: Adapt AI-readable semantic units to a generative engine and iteratively optimize based on feedback results.

[0163] Specifically, iterative optimization based on feedback results involves using a constructed AI semantic processing dataset to perform generative engine adaptation and expression processing on AI-readable semantic units, and then using the generative engine's question-answering verification results to further optimize the AI-readable semantic units. The specific implementation is as follows:

[0164] Obtain the content cognition features corresponding to each generative engine, and construct a generative engine feature set. ,in, Indicates the first Content organization preferences of a generative engine Indicates the first Content citation preferences of a generative engine Indicates the first The semantic expression preferences of a generative engine Indicates the first Each generative engine corresponds to a set of engine features, which are obtained through a set of test questions and answers, collecting information such as answer length, number of cited sources, preference for webpage structure, and preference for lists.

[0165] Based on the engine feature sets corresponding to each generative engine, the AI-readable semantic units are reorganized into structured representations, as follows:

[0166] Extracting subject information, core factual information, parameter information, and evidence information from the current AI-readable semantic unit. ;

[0167] Generate corresponding adapted content according to the preset expression template, specifically:

[0168] If the current generative engine focuses on question-and-answer expression, then it generates standard question-and-answer content. ;

[0169] If the current generative engine focuses on knowledge organization and expression, then it generates knowledge card content. ;

[0170] If the current generative engine focuses on webpage knowledge referencing, then it generates structured semantic webpage content. .

[0171] And form a collection of content adapted to the generative engine. ,in, Indicates the first The set of adapted content corresponding to each AI-readable semantic unit.

[0172] To verify whether the current adapted content can be effectively recognized by the generative engine, a set of test questions is constructed. ,in, This refers to the issue of corporate entity identification. This indicates a product information issue. Indicates a technical parameter issue. This indicates an application scenario issue;

[0173] Input the set of test questions into the target generative engine and obtain the corresponding answer results. , where a Indicates the first The answer results for each test question This indicates the obtained answer result.

[0174] Analyze the enterprise content recognition in the response results and calculate the reference matching degree of AI-readable semantic units. ,in, This indicates the number of pieces of information in the answer that correctly reference the current AI-readable semantic unit. R represents the expected number of references for the current test question. This indicates the reference matching degree of the current AI-readable semantic unit;

[0175] In addition, the expected number of referenced information is the total number of key information points from the target AI readable semantic units that are pre-annotated for each test question when constructing the test question set, and that are expected to be correctly referenced in the generative engine's answer results.

[0176] If the answer correctly references the current AI-readable semantic unit, it means that the answer correctly references the current AI-readable semantic unit if it satisfies at least one of the following: subject consistency, parameter consistency, evidence source consistency, and scenario consistency.

[0177] Set reference matching threshold The set reference matching threshold is compared with the calculated reference matching degree. Based on the comparison result, it is determined whether the current AI-readable semantic unit can be effectively recognized by the generative engine. Specifically:

[0178] If the comparison results satisfy the formula If so, it is determined that the current AI-readable semantic unit has a semantic expression deviation, and an optimization identifier is generated. ;

[0179] If the comparison results satisfy the formula R If so, it is determined that the current AI-readable semantic unit can be effectively recognized by the generative engine and a valid identifier can be generated. ;

[0180] For optimized identifiers, the formula is satisfied. The AI-readable semantic unit performs feedback optimization processing, specifically as follows:

[0181] By analyzing the missing information, misquoted information, and low-confidence information in the generative engine's response results, the location of semantic defects in the current AI-readable semantic units can be determined. ,in, Indicates the first A set of semantic defects corresponding to each AI-readable semantic unit.

[0182] Based on the semantic defect set, the current AI-readable semantic units are optimized, including: supplementing missing factual information; adjusting the relationship between subjects and parameters; enhancing the content of the evidence chain description; and reorganizing the expression order of core facts, as follows:

[0183] Content analysis is performed on the answers returned by the generative engine to identify missing information, incorrect citations, low confidence information, and content obfuscation information in the answers, and the corresponding semantic defects are located to construct a set of semantic defects.

[0184] Based on the different types of defects in the semantic defect set, targeted optimization processing is performed on the current AI-readable semantic unit, specifically as follows:

[0185] When the generative engine fails to correctly identify the enterprise entity information, the corresponding entity information is re-extracted from the original content index relationship and semantic evidence chain and added to the current AI-readable semantic unit;

[0186] When the generative engine makes an incorrect reference to product parameters, technical indicators, or application scenarios, it re-validates the relationship between the main information, parameter information, and scenario information, and re-establishes the corresponding semantic relationship.

[0187] When the generative engine's response is lacking in factual basis or the source of the citation is unclear, it traces back to the original content location based on the content index relationship and supplements the source information, publication time information, version information and relationship information to enhance the integrity of the evidence chain of the current AI readable semantic unit;

[0188] When the generative engine can recognize the current content, but the confidence level of the answer is low or the answer result is ranked low, the topic tags, core fact description order and content organization structure of the current AI readable semantic unit are adjusted to improve the recognition priority of the generative engine for the current semantic unit.

[0189] After optimization, the corresponding AI-readable semantic units are regenerated. Then, the generative engine adaptation verification process is performed again; if the re-verified AI readable semantic unit still has semantic defects, the feedback optimization process continues until the current AI readable semantic unit meets the recognition requirements of the generative engine.

[0190] Ultimately, the optimized AI-readable semantic units and their corresponding optimized states are combined to form a generative content output set.

[0191] Based on the output of generative content sets, the system adapts AI-readable semantic units to the cognitive methods of different generative engines, enabling enterprise content to be organized and expressed in a way that better aligns with the understanding habits of generative engines. Furthermore, it continuously optimizes semantic units based on the actual citation results of generative engines, thereby improving the ability of enterprise content to be accurately identified, understood, and cited by generative engines. This solves the problem in existing technologies where enterprise content cannot be effectively identified and utilized by generative artificial intelligence.

[0192] Example 2

[0193] Reference Figure 2 As an embodiment of the present invention, an enterprise content semantic processing system for generative engines is provided, comprising:

[0194] The multi-source content unified parsing module is used to uniformly parse unstructured content from multiple sources of an enterprise and construct a set of candidate semantic fragments; the semantic reconstruction module is used to semantically reconstruct the candidate semantic fragments and generate AI-readable semantic units; the generative engine adaptation module is used to adapt the AI-readable semantic units to a generative engine and iteratively optimize them based on feedback results.

[0195] Furthermore, the multi-source content unified parsing module is used to uniformly parse multi-source unstructured content of an enterprise and construct a set of candidate semantic fragments, specifically:

[0196] The multi-source content unified parsing module includes a content acquisition unit, a content parsing unit, a content cleaning unit, a semantic boundary recognition unit, and a candidate segment selection unit.

[0197] The content acquisition unit is used to acquire multi-source content from both internal and external sources in real time. Enterprise content includes at least one or more of the following: company website pages, product manuals, PDF documents, technical white papers, image materials, video subtitles, FAQ documents, case studies, and internal knowledge bases. The unit acquires the following content in real time: The original content data corresponding to the enterprise-type content source at the current collection time. and its corresponding content attribute information Construct a collection of content attributes ;

[0198] The content parsing unit performs unified parsing processing on different types of raw content data using corresponding parsing methods. It performs text extraction on text documents, optical character recognition (OCR) on image data, speech-to-text transcription on video or audio data, and webpage text extraction on webpage data. Finally, it standardizes the parsed text content to establish a unified content representation set. At the same time, a corresponding content index relationship is established for each unified text content. ;

[0199] The content cleaning unit is used to clean up a uniform set of text content and sets a threshold for determining duplicate content. Calculate the semantic similarity between any two text contents. The semantic similarity is compared with a threshold to identify and remove duplicate content;

[0200] The semantic boundary recognition unit is used to perform entity recognition processing on the cleaned text content to extract key semantic entities and construct an entity set. Furthermore, the text content undergoes topic change analysis, entity change analysis, and contextual association analysis to calculate the topic feature vector of the current text segment. and the degree of thematic difference between adjacent text fragments Degree of change in the entity and contextual relevance Set semantic boundary determination threshold When the conditions are met The system determines if a semantic boundary exists at the current text position and segments it accordingly. It then continuously segments the unified text content based on the semantic boundary positions to construct a candidate semantic segment set. ;

[0201] The candidate segment filtering unit is used to perform integrity checks on candidate semantic segments and sets a minimum semantic integrity threshold. Calculate the semantic completeness of each candidate semantic segment. ,Will Candidate semantic segments are eliminated, and valid candidate semantic segments are retained. The set of candidate semantic segments and their corresponding validity identifiers are combined to form the basic data set for semantic processing.

[0202] The semantic reconstruction module is used to semantically reconstruct candidate semantic fragments and generate AI-readable semantic units, specifically:

[0203] The semantic reconstruction module includes a content feature extraction unit, an information density analysis unit, a semantic integrity evaluation unit, a semantic reconstruction processing unit, an evidence chain association unit, a citation evaluation unit, and a semantic unit construction unit.

[0204] The content feature extraction unit is used to extract valid identifiers from the semantic processing dataset. The candidate semantic segments are subjected to semantic decomposition processing to extract key semantic information and construct a content feature set;

[0205] The information density analysis unit is used to perform information density analysis on each candidate semantic segment, count the number of fact entities and the total number of words in the text, and calculate the information density value. Set minimum information density threshold ,Will The segments are identified as low-value segments;

[0206] The semantic completeness evaluation unit is used to perform semantic completeness analysis on candidate semantic segments, count the number of context dependencies (i.e., the number of dependencies in the current semantic segment that cannot be understood independently within the current segment and must rely on other text content outside the segment to determine its complete semantics), and calculate the semantic completeness. Set semantic integrity threshold ,Will Segment generation semantic reconstruction identifier ;

[0207] The semantic reconstruction processing unit is used to process semantic reconstruction identifiers that satisfy... The candidate semantic fragments are subjected to semantic reconstruction processing: when there are multiple independent topic entities in the candidate semantic fragments, they are split according to the change position of the topic entities; when there are cross-segment reference relationships or referential relationships, the related fragments are obtained and merged; when there are semantic missing items, supplementary information is extracted from the related text content and reorganized to construct a set of reconstructed candidate semantic fragments.

[0208] The evidence chain association unit is used to establish evidence association relationships for each candidate semantic fragment, establish a mapping relationship between candidate semantic fragments and original content based on content index relationships, and construct a semantic evidence chain. The evidence completeness score is calculated based on the source completeness, location traceability, version validity and relationship completeness in the semantic evidence chain.

[0209] The referability evaluation unit is used to calculate the referability score of candidate semantic segments. Set the referenceability threshold ,Will Remove fragments, retain Valid fragments;

[0210] Semantic unit construction units are used to construct units for valid identifiers that satisfy Candidate semantic fragments are constructed into AI-readable semantic units according to a unified structure. All AI semantic units are organized into a set of AI-readable semantic units, and field integrity checks are performed on the AI-readable semantic units to calculate the field integrity score. Set the minimum field completeness threshold ,Will The semantic units are marked as invalid, and the set of AI-readable semantic units and their corresponding validity marks are combined to form the AI ​​semantic processing data set.

[0211] The generative engine adaptation module is used to adapt AI-readable semantic units to a generative engine and iteratively optimize them based on feedback results. Specifically:

[0212] The generative engine adaptation module includes an engine feature acquisition unit, a structured expression recombination unit, a question-answering verification unit, a citation matching analysis unit, and a feedback optimization unit;

[0213] The engine feature acquisition unit is used to acquire the content cognition features corresponding to each generative engine and construct the generative engine feature set. The engine feature set is obtained through question and answer of a set of test questions, collecting information such as answer length, number of cited sources, preference for web page structure, and preference for lists.

[0214] The structured expression reorganization unit is used to reorganize the structured expression of AI readable semantic units according to the engine feature set corresponding to each generative engine. It extracts the subject information, core fact information, parameter information and evidence information in the current AI readable semantic units, and generates corresponding adapted content according to the preset expression template: if the generative engine focuses on question and answer expression, it generates standard question and answer content; if it focuses on knowledge organization expression, it generates knowledge card content; if it focuses on web page knowledge reference, it generates structured semantic web page content, forming a set of adapted content for the generative engine.

[0215] The question-answering verification unit is used to construct a set of test questions (including questions on enterprise identification, product information, technical parameters, and application scenarios), input the set of test questions into the target generative engine, and obtain the corresponding answer results.

[0216] The citation matching analysis unit is used to analyze the identification of enterprise content in the response results and calculate the citation matching degree. Set reference matching threshold ,Will Semantic unit generation optimization identifier ;

[0217] The feedback optimization unit is used to optimize the flags that meet the requirements. The AI-readable semantic unit performs feedback optimization processing, analyzes missing information, incorrect citations, and low-confidence information in the answer results, determines the location of semantic defects, constructs a semantic defect set, and performs targeted optimization operations based on the defect type: when the generative engine fails to correctly identify the enterprise entity information, it re-extracts the corresponding entity information from the original content index relationship and semantic evidence chain and supplements it into the current AI-readable semantic unit; when the generative engine makes incorrect citations of product parameters or application scenarios, it re-verifies and re-establishes the relationship between entity information, parameter information, and scenario information; when the generative engine's answer lacks factual basis or the source of citation is unclear, it traces back the original content position based on the content index relationship and supplements the source information, publication time information, version information, and relationship information to strengthen the evidence chain. Completeness; when the generative engine's response has low confidence or is ranked low, the topic tags, core fact description order, and content organization structure of the current AI-readable semantic unit are adjusted; after optimization, the AI-readable semantic unit is regenerated and the generative engine adaptation verification process is executed again. If semantic defects still exist after re-verification, feedback optimization is continued until the recognition requirements are met; the optimized AI-readable semantic units and their corresponding optimization states are combined to form a generative content output set, realizing the adaptation between AI-readable semantic units and different generative engine cognition methods, enabling enterprise content to be organized and expressed in a way that is more in line with the understanding habits of generative engines, and continuously optimizing semantic units based on the actual citation results of generative engines, thereby improving the ability of enterprise content to be accurately recognized, understood, and cited by generative engines.

[0218] Example 3

[0219] As an embodiment of the present invention, a method for semantic processing of enterprise content based on the CIDR-3 semantic decomposition algorithm is provided, as follows:

[0220] Content recognition processing based on the CIDR-3 algorithm involves obtaining a set of candidate semantic segments after unified parsing, and then performing entity recognition on each candidate semantic segment to extract enterprise entity, product entity, technical parameter entity, application scenario entity, and evidence entity, forming a candidate semantic feature set. ,in, Indicates the number of corporate entities. Indicates the quantity of product entities. Indicates the number of parameter entities. Indicates the number of entities in the scene. Indicates the number of evidentiary entities.

[0221] The information density assessment module based on the CIDR-3 algorithm calculates the information density score, which is for the first... For each candidate semantic segment, the number of valid fact entities and the total number of words in the text are counted, and an information density score is calculated. ,in, Indicates the first Information density score of each candidate semantic segment This indicates the number of valid fact entities in the candidate semantic fragment. This indicates the number of text words in the candidate semantic segment.

[0222] Compare the information density score with the preset information density threshold Comparing them, we have:

[0223] When the formula is satisfied If the current candidate semantic segment contains a lot of low-value descriptive information, a low-density identifier is generated.

[0224] When the formula is satisfied If the current candidate semantic segment meets the factual information density requirement, it will proceed to the next processing step.

[0225] In addition, the CIDR-3 algorithm uses information density scoring to quantify the proportion of effective facts in candidate semantic segments.

[0226] Semantic boundary recognition processing based on the CIDR-3 algorithm, targeting adjacent candidate semantic segments. and The degree of topic difference, the degree of entity change, and the degree of contextual relevance are calculated separately, as follows:

[0227] Constructing semantic boundary evaluation parameters, we have:

[0228] ;

[0229] in, Indicates the degree of thematic difference between adjacent segments. Indicates the degree of change of an entity. Indicates the degree of contextual relevance. , , Indicates the corresponding weight parameters;

[0230] semantic boundary evaluation parameters With semantic boundary determination threshold Comparing them, we have:

[0231] When the formula is satisfied If so, it is determined that there is a semantic boundary at the current position, and semantic segmentation is performed;

[0232] When the formula is satisfied If so, the current semantic continuity relationship is maintained.

[0233] Evidence chain association processing based on the CIDR-3 algorithm involves establishing a mapping relationship between semantic segments and the original content after semantic boundary segmentation. Specifically:

[0234] Establish a mapping relationship between semantic fragments and original content ,in, Indicates the source document, Show content location, Indicates the release time. Indicates version information;

[0235] To calculate the completeness of evidence, we have:

[0236] ;

[0237] in, Indicates the completeness of the source. Indicates location traceability. Indicates version validity. Indicates the completeness of association. , , , This represents the evidence weight parameter.

[0238] The AI ​​citationability scoring based on the CIDR-3 algorithm integrates information density scoring, semantic completeness scoring, evidence completeness scoring, and timeliness scoring. Constructing an AI citationability evaluation model yields the following:

[0239] ;

[0240] in, Indicates the first AI citation score for each candidate semantic fragment Indicates information density score, Indicates semantic completeness score. Indicates the completeness of evidence score. Indicates timeliness score, , , , This represents the weighting parameter for calculating citationability.

[0241] Scoring citationability With referenceability threshold Comparing them, we have:

[0242] When the formula is satisfied Then, the semantic fragments of the generative engine reference conditions are determined, and AI-readable semantic units are generated.

[0243] Based on the question-and-answer feedback results, CIDR-3 parameters are adaptively and iteratively optimized. After completing the construction of AI-readable semantic units, the set of test questions is input into the target generative engine to obtain the answer results, specifically:

[0244] Calculate semantic unit reference matching degree ,in, This indicates the number of pieces of information in the response that correctly match the current semantic unit. Indicates the expected amount of information corresponding to the test question;

[0245] Based on the citation matching results, the internal parameters of the CIDR-3 algorithm are dynamically adjusted, rather than manually modified. Specifically:

[0246] When the reference matching degree is lower than a preset threshold, the expression weight of low-contribution fields in the current semantic unit is reduced, and the weight of core fact fields is increased. ,in, Indicates the first Class semantic field weights This represents the weight adjustment amount calculated based on the reference feedback;

[0247] When the accuracy of the answer decreases, adjust the weight of the evidence chain association. Increase the contribution of evidentiary features such as source completeness and version validity to the citationability score;

[0248] When the semantic boundary misidentification rate increases, adjust the threshold parameter in the semantic boundary determination model. Re-execute candidate semantic segmentation;

[0249] When the generative engine's performance in recognizing the current structured representation deteriorates, adjust the field sorting parameters, label association parameters, and summary generation parameters in the representation recombination model. ,in, It represents a structured representation of the parameter set.

[0250] It should be noted that by limiting the CIDR-3 semantic decomposition algorithm to a continuous execution process including content recognition, information density judgment, semantic boundary recognition, evidence chain association, and AI citationability scoring, and by binding the information density scoring model, citationability scoring model, and weight parameter adjustment mechanism in the feedback optimization process, CIDR-3 is transformed from a conceptual algorithm into a technical processing flow with clear calculation steps and parameter update mechanisms. At the same time, by automatically adjusting the algorithm weights, threshold parameters, and structured expression parameters based on the feedback results of the generative engine, the reliance on manual content operation adjustments is avoided, and dynamic adaptation and continuous optimization of enterprise content semantic units for the generative engine are achieved.

[0251] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of protection claimed by the present invention.

Claims

1. A semantic processing method for enterprise content oriented towards generative engines, characterized in that: Includes the following steps: The system performs unified parsing of multi-source unstructured content from enterprises and constructs a set of candidate semantic fragments, specifically as follows: Collect heterogeneous content data from multiple sources of enterprises. For different types of raw content data, use corresponding parsing methods for unified parsing processing. Standardize the parsed text content and establish a unified content expression set. Perform content cleaning processing, entity recognition processing on the cleaned text content, extract key semantic entities from the text, construct entity sets, perform topic change analysis, entity change analysis, and context association analysis on each text content to identify semantic boundaries, continuously segment the unified text content, and construct a candidate semantic segment set. The candidate semantic fragments are semantically reconstructed to generate AI-readable semantic units, specifically: The candidate semantic fragments are subjected to content recognition, information density analysis, and semantic integrity assessment. Based on the semantic integrity assessment results, semantic reconstruction processing is performed, and the citation score of the current candidate semantic fragment is calculated. Based on the calculation results, it is determined whether the current candidate semantic fragment has citation value for the generative engine, and AI-readable semantic units are constructed. All constructed AI semantic units are organized in a unified format to construct a set of AI-readable semantic units. The AI-readable semantic units are adapted to a generative engine, and iterative optimization is performed based on the feedback results, specifically as follows: The content cognitive features of the target generative engine are obtained. Based on the content cognitive features, the AI-readable semantic units are restructured and reorganized into a structured expression. The generative engine is verified by question-and-answer testing using a set of test questions. Based on the answer results, the semantic defect locations of the AI-readable semantic units are determined, and the AI-readable semantic units are optimized by feedback until they meet the recognition requirements of the generative engine.

2. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The unified parsing process is performed as follows: Utilize the enterprise content access module to acquire multi-source content from both internal and external sources in real time; For different types of raw content data, corresponding parsing methods are used for unified parsing processing, specifically: If the current content type is a text document, then perform text extraction processing to generate the corresponding text content; If the current content type is image data, then the optical character recognition algorithm is used to perform text recognition processing to generate the corresponding OCR text content; If the current content type is video or audio data, then a speech recognition algorithm is used to perform speech-to-text processing to generate the corresponding speech-text content; If the current content type is web page data, then the web page structure is parsed using a web page text extraction algorithm to remove the navigation bar, advertising areas, and irrelevant links, and generate the corresponding web page text content.

3. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The content cleaning process is performed as follows: Set a threshold for determining duplicate content And calculate the semantic similarity between any two text contents. , Representing text With text Semantic similarity between them; The semantic similarity between two pieces of text is compared with a set threshold for determining duplicate content. The comparison result is used to determine whether the text content is duplicated. Specifically: If the comparison results satisfy the formula If so, the corresponding text content is determined to be duplicate content, and the duplicate text is deleted; If the comparison results satisfy the formula If so, the current text content will be retained; Entity recognition processing is performed on the cleaned text content to extract key semantic entities and construct an entity set. ,in, Indicates the first A semantic entity, Indicates the total number of semantic entities.

4. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The identification of semantic boundaries is specifically as follows: Calculate the topic feature vector of the current text segment and the degree of thematic difference between it and adjacent text fragments. Simultaneously calculate the degree of entity change between adjacent text segments. Calculate the contextual association between text fragments And set a semantic boundary determination threshold. ; The calculated contextual relevance between text segments is compared with a set semantic boundary determination threshold. Based on the comparison result, it is determined whether the current text content belongs to the same semantic unit. Specifically: If the comparison results satisfy the formula If a semantic boundary exists at the current text position, the current text content will be segmented. If the comparison results satisfy the formula If so, it is determined that the current text content still belongs to the same semantic unit.

5. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The information density analysis is as follows: Statistics Number of fact entities in each candidate semantic segment and the total number of words in the text ; And calculate the information density value of the candidate semantic fragments. ,in, Indicates the first Information density of each candidate semantic segment; Set minimum information density threshold The system compares the set minimum information density threshold with the calculated information density value, and determines whether the current candidate semantic segment contains invalid information based on the comparison result. Specifically: If the comparison results satisfy the formula If so, it is determined that the current candidate semantic fragment contains marketing description information or invalid explanatory information, and a low-value identifier is generated. The current candidate semantic fragments do not participate in the construction of subsequent AI-readable semantic units; If the comparison results satisfy the formula If no marketing description information or invalid explanatory information is found in the current candidate semantic fragment, a high-value identifier is generated. The current candidate semantic fragments participate in the construction of subsequent AI-readable semantic units.

6. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The semantic completeness assessment is as follows: Count the number of fact entities in the current candidate semantic segments respectively. and the number of context dependencies Calculate the semantic completeness of the current candidate semantic segment. ; Set semantic integrity threshold The set semantic completeness threshold is compared with the calculated semantic completeness, specifically: If the comparison results satisfy the formula If so, it is determined that the current candidate semantic segment has semantic missing or strong contextual dependency, and a semantic reconstruction identifier is generated. ; If the comparison results satisfy the formula If the current candidate semantic segment does not have semantic missing or strong contextual dependency, a semantic reconstruction identifier is generated. .

7. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The process of determining whether the current candidate semantic fragment has reference value for the generative engine based on the calculation results is as follows: A mapping relationship between candidate semantic fragments and original content is established based on the constructed content index relationship. ,in, This indicates the constructed content index relationship. This represents the candidate semantic fragments after semantic reconstruction, and simultaneously constructs a semantic evidence chain. The evidence completeness score is calculated based on the source completeness, location traceability, version validity, and relational completeness in the semantic evidence chain. And calculate the citation score of the current candidate semantic fragment. ; Set referenceability threshold The set referability threshold is compared with the calculated referability score. Based on the comparison result, it is determined whether the current candidate semantic fragment has referential value for the generative engine. Therefore: If the comparison results satisfy the formula If the current candidate semantic fragment is not considered valid, an invalid identifier is generated. ; If the comparison results satisfy the formula If so, the current candidate semantic fragment is determined to have reference value for the generative engine, and a valid identifier is generated. .

8. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The specific steps for constructing the AI-readable semantic unit set are as follows: For valid identifiers that satisfy the formula The candidate semantic fragments are organized according to a unified structure to construct AI-readable semantic units. ; If all the constructed AI semantic units are organized according to a unified format to form a set of AI-readable semantic units, then: ; in, This represents the constructed set of AI-readable semantic units; Set minimum field completeness threshold Calculate the first Field completeness of each AI-readable semantic unit ,in, This indicates the number of valid fields in the current semantic unit. Valid fields are data fields that are not empty and pass format validation. Indicates the total number of preset fields. Indicates the completeness of the calculated fields; The set minimum field completeness threshold is compared with the calculated field completeness. Based on the comparison result, it is determined whether the current AI-readable semantic unit is valid. Specifically: If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be invalid, and an invalidation flag is generated. ; If the comparison results satisfy the formula If so, the current AI-readable semantic unit is determined to be valid, and a valid identifier is generated. ; The set of AI-readable semantic units and their corresponding validity identifiers together form the AI ​​semantic processing data set. ,in, This represents a dataset used for AI semantic processing.

9. The enterprise content semantic processing method for generative engines as described in claim 1, characterized in that: The determination of the semantic defect location of the AI-readable semantic unit based on the answer result is as follows: Calculate the reference matching degree of AI-readable semantic units ,in, This indicates the number of pieces of information in the answer that correctly reference the current AI-readable semantic unit. R represents the expected number of references for the current test question. This indicates the reference matching degree of the current AI-readable semantic unit; Set reference matching threshold The system compares the set reference matching threshold with the calculated reference matching degree, and determines whether the current AI-readable semantic unit can be effectively recognized by the generative engine based on the comparison result. Specifically: If the comparison results satisfy the formula If so, it is determined that the current AI-readable semantic unit has a semantic expression deviation, and an optimization identifier is generated. ; If the comparison results satisfy the formula R If so, it is determined that the current AI-readable semantic unit can be effectively recognized by the generative engine and a valid identifier can be generated. ; For optimized identifiers, the formula is satisfied. The AI-readable semantic unit performs feedback optimization processing, specifically as follows: By analyzing the missing information, misquoted information, and low-confidence information in the generative engine's response results, the location of semantic defects in the current AI-readable semantic units can be determined. ,in, Indicates the first A set of semantic defects corresponding to each AI-readable semantic unit.

10. A semantic markup processing system for enterprise content oriented towards generative engines, employing the semantic markup processing method for enterprise content oriented towards generative engines as described in any one of claims 1 to 9, characterized in that, include: A unified parsing module for multi-source content, a semantic reconstruction module, and a generative engine adaptation module; The multi-source content unified parsing module is used to perform unified parsing of multi-source unstructured content of enterprises and construct a set of candidate semantic fragments; The semantic reconstruction module is used to perform semantic reconstruction on candidate semantic fragments and generate AI-readable semantic units; The generative engine adaptation module is used to adapt AI-readable semantic units to a generative engine and iteratively optimize them based on feedback results.