An automated review method, system, medium, and product for product promotional copy.

Abstract

This invention discloses an automatic review method, system, medium, and product for product promotional copy, belonging to the field of semantic analysis. The method comprises: converting product promotional copy into structured text data and identifying each target entity and its identifier; based on each entity and identifier, obtaining structured factual data from a preset data source and constructing a factual knowledge graph containing logical relationships and a time dimension; parsing natural language statements in the text data, generating structured data to be verified that matches the graph format, and performing semantic consistency verification, outputting a first verification result; using the graph as a driving force to reverse-search and verify the text data, outputting a second verification result; combining static rules and contextual dynamic rules constructed based on the graph to perform multi-dimensional compliance checks on the text data, outputting a third verification result; and generating a review report by combining the three verification results. By implementing this invention, the accuracy of semantic verification of promotional copy can be improved.

Description

Technical Field

[0001] This invention belongs to the field of semantic analysis and relates to an automatic review method, system, medium, and product for product promotional copy. Background Technology

[0002] In industries such as finance, e-commerce, and advertising, product promotional copy (such as fund posters and product detail pages) often contains complex information such as performance data, relationships between entities, and evaluative statements. The authenticity and compliance of this content are directly related to user decisions and regulatory compliance, thus requiring an efficient and accurate automated review mechanism.

[0003] Currently, the review of promotional copy mainly relies on manual comparison or rule-based systems based on keyword matching: manual review requires checking each item across multiple data sources, which is inefficient and inconsistent in standards; rule-based systems can only recognize fixed text patterns and cannot understand complex language contexts and semantic differences. In addition, traditional automated scripts have great limitations in processing unstructured information (such as images and charts) that are common in modern marketing materials. When faced with diverse natural language and multimodal inputs, they generally suffer from problems such as missed reports, high false positive rates, and incomplete coverage. Summary of the Invention

[0004] This application provides an automatic review method, system, medium, and product for product promotional copy, which can improve the accuracy of semantic verification of promotional copy.

[0005] To achieve the above objectives, in a first aspect, the present invention provides an automatic review method for product promotional copy, comprising: The product promotional text is acquired and converted into structured text data. Each target entity is identified from the structured text data, and the identifier corresponding to each target entity is determined. Based on each target entity and its corresponding identifier, the corresponding structured fact data is obtained from a preset data source, and a fact knowledge graph containing logical relationships between entities and the time dimension is constructed. The natural language statements in the structured text data are parsed to generate structured data to be verified with a format matching the fact knowledge graph. Semantic consistency verification is performed in the fact knowledge graph, and the first verification result is output. Driven by the aforementioned factual knowledge graph, the structured text data is retrieved and verified in reverse, and a second verification result is output. Based on preset rules, multi-dimensional compliance checks are performed on the structured text data, and a third verification result is output; wherein, the rules include static rules and context-dynamic rules constructed based on the fact knowledge graph; An audit report is generated based on the first verification result, the second verification result, and the third verification result.

[0006] Compared to existing technologies, the embodiments of this application have the following beneficial effects: By converting product promotional text into structured text data, the problem that unstructured multimodal inputs (such as images and PDFs) cannot be directly understood by machines is solved, providing a unified and parsable input foundation for subsequent automated processing; furthermore, by identifying target entities from structured text and assigning identifiers to them, the core objects in the text are accurately anchored, establishing an index for cross-data source fact queries; simultaneously, by obtaining structured fact data from preset data sources based on target entities and their identifiers and constructing a fact knowledge graph containing logical relationships between entities and the time dimension, a dynamic, authoritative, and context-complete "fact benchmark" is formed, which can avoid the inconsistency problem caused by scattered facts and heterogeneous sources in traditional review; on this basis, by parsing natural language statements to generate format-matched structured data to be verified and performing semantic consistency verification in the fact knowledge graph, deep semantic understanding and accurate fact comparison of the text statements are achieved. This technology improves the accuracy of verifying complex expressions. Furthermore, it uses a fact knowledge graph to drive reverse retrieval and verification of structured text data, constructing a comprehensive "fact → text" coverage mechanism. This effectively identifies omissions or errors in information extraction, reducing the rate of missed detections during audits. Moreover, by executing multi-dimensional compliance checks that include static rules and dynamic contextual rules built on the fact knowledge graph, compliance judgments are no longer limited to keyword matching. Instead, they can dynamically trigger conditional rules based on factors such as product attributes and contextual semantics in the promotional text, achieving intelligent and refined compliance logic. The synergistic effect of these features forms a closed-loop audit process: "structured input—entity anchoring—dynamic fact graph—positive statement verification + reverse fact verification + contextual compliance." This solves the core problem of low semantic verification accuracy caused by insufficient semantic understanding, fragmented fact sources, and static, rigid rules in existing technologies, achieving high-precision, high-coverage automated auditing of the authenticity, completeness, and compliance of promotional texts.

[0007] In some embodiments of the first aspect of this application, the acquisition and conversion of product promotional copy into structured text data includes: The product promotional text is obtained and the file type is determined. If it is a document format, its document structure is parsed and the text and table content are extracted to obtain temporary document data. If it is an image format, layout structure recognition and character content recognition are performed in parallel, and the recognition results are fused to obtain temporary image data. Based on the temporary document data or image data, structured text data that retains the original semantic structure and has a uniform format is generated.

[0008] Compared with existing technologies, the above embodiments have the following beneficial effects: by acquiring product promotional text and determining its file type, adaptive recognition of heterogeneous input sources is achieved; furthermore, the document structure is analyzed for document formats to extract text and table content, and layout structure recognition and character content recognition are performed in parallel for image formats and the results are fused, effectively balancing the semantic integrity of the layout and the accuracy of the character content, overcoming the one-sidedness of a single recognition model when dealing with complex layouts; on this basis, structured text data with preserved original semantic structure and uniform format is generated based on the processing results, so that machine-readable text with consistent structure and clear hierarchy can be output for different modal inputs, providing a stable and reliable data foundation for subsequent entity recognition and semantic parsing, and significantly improving the robustness and consistency of information extraction in multimodal scenarios.

[0009] In some embodiments of the first aspect of this application, the step of identifying each target entity from the structured text data and determining the identifier corresponding to each target entity includes: Based on the preset named entity recognition model, each target entity is identified from the structured text data, and the identifier explicitly associated with each target entity is extracted simultaneously. For a target entity with no explicit association with an identifier, a query request is constructed based on its name, a preset search interface is called to perform the query, a list of candidate identifiers is obtained, and the string similarity and semantic similarity between each candidate identifier in the list and the name of the target entity are calculated. The candidate identifier with the highest comprehensive score is selected as the identifier of the target entity. The comprehensive score is obtained by weighted fusion of string similarity and semantic similarity.

[0010] Compared with existing technologies, the above embodiments have the following beneficial effects: by extracting explicitly associated identifiers, efficient extraction of labeled entities is achieved; for entities without explicit identifiers, a query request is constructed based on their names and a preset search interface is called to obtain a list of candidate identifiers, solving the identification difficulties caused by ambiguous entity references or abbreviated expressions; furthermore, by calculating the string similarity and semantic similarity between each candidate identifier and the target entity name, and selecting the optimal identifier based on the comprehensive score of weighted fusion, the string similarity ensures the accuracy of literal matching, the semantic similarity understands the semantic association of names through a language model, and the weighted fusion mechanism achieves a balance between precise matching and semantic generalization, so that even in scenarios where entity expressions are not standardized, each target entity can still be accurately assigned a unique identifier that can be used for cross-source queries, providing a reliable anchor point for the construction of fact knowledge graphs.

[0011] In some embodiments of the first aspect of this application, the step of obtaining corresponding structured fact data from a preset data source based on each target entity and its corresponding identifier, and constructing a fact knowledge graph containing logical relationships between entities and a time dimension, includes: Based on each target entity and its corresponding identifier, multiple preset external data source interfaces are called in parallel to obtain the raw response data returned by each external data source; For each of the original response data, the attribute type, attribute value and time information associated with each target entity are extracted to obtain the corresponding structured fact data; For multiple structured fact data with the same target entity and the same attribute type, the structured fact data with the highest priority is retained according to the preset priority of their corresponding external data source interfaces. Based on all the retained structured fact data, a fact knowledge graph containing logical relationships between entities and the time dimension is constructed, with each target entity as the subject, attribute type as the relationship, attribute value as the object of the relationship, and corresponding time information.

[0012] Compared to existing technologies, the above embodiments have the following beneficial effects: By obtaining raw response data through parallel calls to multiple preset external data source interfaces based on each target entity and its identifier, efficient aggregation of multi-source fact data is achieved, avoiding the latency accumulation of serial queries; further, attribute type, attribute value, and time information are extracted from the raw response data to obtain structured fact data, and the heterogeneous interface returns are uniformly transformed into standardized fields; for multiple structured fact data with the same target entity and the same attribute type, the highest priority one is retained according to the preset priority of its corresponding external data source interface, effectively solving the problem of multi-source data conflict and ensuring the consistency and authority of the fact benchmark; on this basis, based on all retained structured fact data, a fact knowledge graph is constructed with the target entity as the subject, attribute type as the relation, attribute value as the relation pointing to the object, and combined with time information, forming a dynamic semantic network containing logical associations between entities and time validity, providing a structured and queryable factual basis for subsequent two-way verification and context rule judgment.

[0013] In some embodiments of the first aspect of this application, the step of performing semantic consistency verification in the fact knowledge graph and outputting a first verification result includes: For each structured data to be verified, extract its declared target entity, attribute type, attribute value, and time information. Then, query the fact knowledge graph to see if there is any structured fact data with the same declared target entity and attribute type. If so, based on the preset language model, perform semantic consistency judgment between the declared attribute value and the attribute value in the corresponding structured fact data, and output the first verification result. Otherwise, construct a query request based on the declared target entity, attribute type, and time information, match and call the corresponding external verification interface from the preset interface metadata dictionary, obtain the returned attribute value data, and call the language model to perform semantic consistency judgment between the returned attribute value data and the declared attribute value, and output the first verification result.

[0014] Compared with existing technologies, the above embodiments have the following beneficial effects: By extracting the target entity, attribute type, attribute value, and time information of each structured data to be verified, and querying whether there is structured fact data with the same target entity and attribute type in the fact knowledge graph, accurate positioning and preliminary matching of the declaration are achieved; if a matching item exists, the semantic consistency judgment between the declared attribute value and the graph attribute value is performed based on the preset language model, so that the system can understand semantically equivalent but different expressions, overcoming the limitations of strict string matching; if there is no matching item in the graph, a query request is constructed according to the target entity, attribute type, and time information of the declaration, and the corresponding external verification interface is matched and called from the preset interface meta-information dictionary to obtain attribute value data, and the language model is called again to perform semantic consistency judgment, thereby extending the verification capability from static graph to dynamic indicators, significantly improving the coverage breadth and depth of fact verification.

[0015] In some embodiments of the first aspect of this application, when performing semantic consistency verification and outputting a verification failure result, the method further includes: Relocate the original declaration text fragment corresponding to the structured data to be verified within the structured text data; The attribute value text content in the original declaration text fragment is compared with the attribute value in the structure data to be verified; if they match, the verification failure result is retained; otherwise, the attribute value in the structure data to be verified is replaced with the attribute value text content in the original declaration text fragment to obtain the corrected structure data to be verified, and the semantic consistency verification is re-executed.

[0016] Compared with existing technologies, the above embodiments have the following beneficial effects: When the semantic consistency verification outputs a verification failure result, the source of the error is traced by relocating the original declaration text fragment corresponding to the structured data to be verified in the structured text data; furthermore, the attribute value text content in the original declaration text fragment is compared with the attribute value in the structured data to be verified. If the two are consistent, the verification failure result is retained, indicating that the error originates from the text content itself, ensuring the authenticity of the review result; if the two are inconsistent, the attribute value in the structured data to be verified is replaced with the attribute value text content in the original declaration text fragment, obtaining the corrected structured data to be verified and re-executing the semantic consistency verification, thereby correcting the misjudgment caused by the information extraction deviation in the previous steps, effectively distinguishing between the two types of situations of "content error" and "system extraction error", and greatly improving the reliability and traceability of the review result.

[0017] In some embodiments of the first aspect of this application, the step of using the fact knowledge graph as a driving force to reverse retrieve and verify the structured text data, and output a second verification result, includes: Traverse each structured fact data in the fact knowledge graph, and for each structured fact data, construct a corresponding query instruction based on its target entity and attribute type as the query conditions; Based on each query instruction and combined with a preset language model, the system searches the structured text data for statements that are semantically consistent with the structured fact data corresponding to each query instruction. If no statement is found or the found statement does not match the corresponding attribute value, it is marked as information omission or factual error.

[0018] Compared with existing technologies, the above embodiments have the following beneficial effects: By traversing the structured fact data in the fact knowledge graph and constructing corresponding query instructions based on the target entity and attribute type as query conditions, a systematic coverage of the entire graph of facts is achieved; furthermore, based on each query instruction and a preset language model, the system searches for semantically consistent declarations in the structured text data, enabling the system to proactively verify whether the text has omitted key facts or made errors in expression. This reverse-driven mechanism does not depend on whether the text proactively declares a fact, but rather uses the fact graph as a benchmark to reverse-check the integrity of the original text; if no declaration is found or the found declaration is inconsistent with the attribute value, it is marked as information omission or factual error, thereby effectively capturing the underreporting problem caused by information extraction failure or writing negligence, complementing the forward verification, and together forming a two-way verification system with high recall.

[0019] Secondly, the present invention also provides an automatic review system for product promotional copy, comprising: a data acquisition module, a graph construction module, a first verification module, a second verification module, a third verification module, and a result output module; The data acquisition module is used to acquire and convert product promotional text into structured text data, identify each target entity from the structured text data, and determine the identifier corresponding to each target entity. The graph construction module is used to obtain corresponding structured fact data from a preset data source based on each target entity and its corresponding identifier, and to construct a fact knowledge graph containing logical relationships between entities and the time dimension. The first verification module is used to parse each natural language statement in the structured text data, generate each structured data to be verified whose format matches the fact knowledge graph, perform semantic consistency verification in the fact knowledge graph, and output the first verification result; The second verification module is used to reverse retrieve and verify the structured text data driven by the fact knowledge graph, and output the second verification result; The third verification module is used to perform multi-dimensional compliance checks on the structured text data based on preset rules and output the third verification result; wherein, the rules include static rules and context-dynamic rules constructed based on the fact knowledge graph; The result output module is used to generate an audit report based on the first verification result, the second verification result, and the third verification result.

[0020] Compared to existing technologies, the above embodiments of this application have the following beneficial effects: By converting product promotional text into structured text data, the problem that unstructured multimodal inputs (such as images and PDFs) cannot be directly understood by machines is solved, providing a unified and parsable input foundation for subsequent automated processing; furthermore, by identifying target entities from structured text and assigning identifiers to them, the core objects in the text are accurately anchored, establishing an index for cross-data source fact queries; simultaneously, by obtaining structured fact data from preset data sources based on target entities and their identifiers and constructing a fact knowledge graph containing logical relationships between entities and the time dimension, a dynamic, authoritative, and context-complete "fact benchmark" is formed, which can avoid the inconsistency problem caused by scattered facts and heterogeneous sources in traditional review; on this basis, by parsing natural language statements to generate format-matched structured data to be verified and performing semantic consistency verification in the fact knowledge graph, deep semantic understanding and accurate fact comparison of the text statements are achieved. This improves the accuracy of verifying complex expressions. Furthermore, by using a fact knowledge graph as a driving force to reverse-search and verify structured text data, a fallback mechanism from "facts → text" is constructed, effectively identifying omissions or errors in information extraction and reducing the rate of missed detections during audits. Moreover, by executing multi-dimensional compliance checks that include static rules and contextual dynamic rules built on the fact knowledge graph, compliance judgment is no longer limited to keyword matching but can dynamically trigger conditional rules based on factors such as product attributes and contextual semantics in the promotional text, achieving intelligent and refined compliance logic. The synergistic effect of these features forms a closed-loop audit process of "structured input—entity anchoring—dynamic fact graph—positive statement verification + reverse fact verification + contextual compliance," solving the core problem of low semantic verification accuracy caused by insufficient semantic understanding, fragmented fact sources, and static rigid rules in existing technologies. This achieves high-precision, high-coverage automated auditing of the authenticity, completeness, and compliance of promotional texts.

[0021] Thirdly, the present invention also provides a computer program product, including a computer program or instructions, characterized in that, when the computer program or instructions are executed, they implement the automatic review method for any product promotional copy of the present invention.

[0022] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the automatic review method for any product promotional copy of the present invention. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an automatic review method for product promotional copy provided in some embodiments of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of an automatic review system for product promotional copy provided in some embodiments of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: Please refer to Figure 1 To improve the accuracy of semantic verification of promotional copy, an embodiment of the present invention provides an automatic review method for product promotional copy, including steps S1 to S6: Step S1: Obtain and convert the product promotional text into structured text data, identify each target entity from the structured text data, and determine the identifier corresponding to each target entity.

[0027] Furthermore, step S1 can be implemented through the following preferred embodiments, including steps S11-S14, as follows: S11: Obtain the product promotional text and determine the file type. If it is a document format, parse its document structure and extract the text and table content to obtain temporary document data. If it is an image format, perform layout structure recognition and character content recognition in parallel, and fuse the recognition results to obtain temporary image data. S12: Based on the temporary document data or temporary image data, generate structured text data that retains the original semantic structure and has a uniform format.

[0028] In this preferred embodiment, adaptive recognition of heterogeneous input sources is achieved by acquiring product promotional text and determining its file type. Furthermore, the document structure is analyzed to extract text and table content for document formats, and layout structure recognition and character content recognition are performed in parallel for image formats, with the results fused. This effectively balances the semantic integrity of the layout with the accuracy of the character content, overcoming the limitations of a single recognition model in handling complex layouts. Based on this, structured text data that retains the original semantic structure and has a unified format is generated from the processing results. This ensures that machine-readable text with consistent structure and clear hierarchy can be output from different modal inputs, providing a stable and reliable data foundation for subsequent entity recognition and semantic analysis, significantly improving the robustness and consistency of information extraction in multimodal scenarios.

[0029] S13: Based on the preset named entity recognition model, identify each target entity from the structured text data, and simultaneously extract the identifiers that are explicitly associated with each target entity. S14: For a target entity with no explicit association with an identifier, construct a query request based on its name, call a preset search interface to perform the query, obtain a list of candidate identifiers, and calculate the string similarity and semantic similarity between each candidate identifier in the list and the name of the target entity. Select the candidate identifier with the highest comprehensive score as the identifier of the target entity; wherein, the comprehensive score is obtained by weighted fusion of string similarity and semantic similarity.

[0030] In this preferred embodiment, the efficient extraction of labeled entities is achieved by extracting explicitly associated identifiers. For entities without explicit identifiers, a query request is constructed based on their names, and a preset search interface is called to obtain a list of candidate identifiers, thus solving the identification difficulties caused by ambiguous entity references or abbreviated expressions. Furthermore, the string similarity and semantic similarity between each candidate identifier and the target entity name are calculated, and the optimal identifier is selected based on the comprehensive score of weighted fusion. String similarity ensures the accuracy of literal matching, semantic similarity understands the semantic association of names through a language model, and the weighted fusion mechanism achieves a balance between precise matching and semantic generalization. Thus, even in scenarios where entity expressions are not standardized, each target entity can still be accurately assigned a unique identifier that can be used for cross-source queries, providing a reliable anchor for the construction of fact knowledge graphs.

[0031] Step S2: Based on each target entity and its corresponding identifier, obtain the corresponding structured fact data from the preset data source, and construct a fact knowledge graph containing logical relationships between entities and the time dimension.

[0032] Furthermore, step S2 can be implemented through the following preferred embodiments, including steps S21-S24, as follows: S21: Based on each target entity and its corresponding identifier, call multiple preset external data source interfaces in parallel to obtain the original response data returned by each external data source; S22: For each of the original response data, extract the attribute type, attribute value and time information associated with each target entity to obtain the corresponding structured fact data; S23: For multiple structured fact data with the same target entity and the same attribute type, retain the structured fact data with the highest priority according to the preset priority of their corresponding external data source interface; S24: Based on all the retained structured fact data, with each target entity as the subject, attribute type as the relation, attribute value as the relation pointing to the object, and combined with the corresponding time information, construct a fact knowledge graph that includes logical relationships between entities and the time dimension.

[0033] In this preferred embodiment, by parallelly calling multiple preset external data source interfaces based on each target entity and its identifier to obtain raw response data, efficient aggregation of multi-source fact data is achieved, avoiding the latency accumulation of serial queries. Furthermore, attribute type, attribute value, and time information are extracted from the raw response data to obtain structured fact data, and the heterogeneous interface returns are uniformly transformed into standardized fields. For multiple structured fact data with the same target entity and attribute type, the highest priority data is retained according to the preset priority of its corresponding external data source interface, effectively solving the multi-source data conflict problem and ensuring the consistency and authority of the factual benchmark. Based on this, a factual knowledge graph is constructed based on all retained structured fact data, with the target entity as the subject, attribute type as the relation, attribute value as the relational object, and time information, forming a dynamic semantic network containing logical relationships between entities and time validity. This provides a structured and queryable factual basis for subsequent two-way verification and contextual rule judgment.

[0034] Step S3: Parse each natural language statement in the structured text data to generate each structured data to be verified whose format matches the fact knowledge graph, and perform semantic consistency verification in the fact knowledge graph to output the first verification result.

[0035] Furthermore, the semantic consistency verification can be implemented through the following preferred embodiment, including step S31, as follows: S31: For each structured data to be verified, extract its declared target entity, attribute type, attribute value, and time information, and query in the fact knowledge graph whether there is structured fact data with the same declared target entity and attribute type; if there is, based on the preset language model, perform semantic consistency judgment between the declared attribute value and the attribute value in the corresponding structured fact data, and output the first verification result; otherwise, construct a query request based on the declared target entity, attribute type, and time information, match and call the corresponding external verification interface from the preset interface meta information dictionary, obtain the returned attribute value data, and call the language model to perform semantic consistency judgment between the returned attribute value data and the declared attribute value, and output the first verification result.

[0036] In this preferred embodiment, by extracting the target entity, attribute type, attribute value, and time information of each structured data to be verified, and querying the fact knowledge graph to see if there is structured fact data with the same target entity and attribute type, accurate positioning and preliminary matching of the declaration are achieved. If a match exists, the semantic consistency of the declared attribute value and the graph attribute value is judged based on the preset language model, so that the system can understand semantically equivalent but different content, overcoming the limitations of strict string matching. If there is no match in the graph, a query request is constructed based on the target entity, attribute type, and time information of the declaration, and the corresponding external verification interface is called from the preset interface meta-information dictionary to obtain the attribute value data. The language model is called again to judge semantic consistency, thereby extending the verification capability from static graph to dynamic indicators, significantly improving the coverage breadth and depth of fact verification.

[0037] Preferably, when performing semantic consistency verification and outputting the verification failure result in step S31, steps S32-S33 are also included, as follows: S32: Locate the original declaration text fragment corresponding to the structured data to be verified in the structured text data; S33: Compare the attribute value text content in the original declaration text fragment with the attribute value in the structure data to be verified; if they match, retain the verification failure result; otherwise, replace the attribute value in the structure data to be verified with the attribute value text content in the original declaration text fragment to obtain the corrected structure data to be verified, and re-execute the semantic consistency verification.

[0038] In this preferred embodiment, when the semantic consistency check outputs a failure result, the source of the error is traced by relocating the original declaration text fragment corresponding to the structured data to be checked in the structured text data. Furthermore, the attribute value text content in the original declaration text fragment is compared with the attribute values ​​in the structured data to be checked. If they match, the failure result is retained, indicating that the error originates from the text content itself, ensuring the authenticity of the review result. If they do not match, the attribute values ​​in the structured data to be checked are replaced with the attribute value text content in the original declaration text fragment, resulting in the corrected structured data to be checked, and the semantic consistency check is re-executed. This corrects the misjudgment caused by the information extraction deviation in the preceding steps, effectively distinguishing between "content error" and "system extraction error," significantly improving the reliability and traceability of the review result.

[0039] Step S4: Driven by the fact knowledge graph, reverse the retrieval and verify the structured text data, and output the second verification result.

[0040] Furthermore, step S4 can be implemented through the following preferred embodiments, including steps S41-S42, as follows: S41: Traverse each structured fact data in the fact knowledge graph, and for each structured fact data, construct a corresponding query instruction based on its target entity and attribute type as the query conditions; S42: Based on each query instruction and combined with a preset language model, search in the structured text data for a statement that is semantically consistent with the structured fact data corresponding to each query instruction. If no statement is found or the found statement is inconsistent with the corresponding attribute value, it is marked as information omission or fact error.

[0041] In this preferred embodiment, by traversing the structured fact data in the fact knowledge graph and constructing corresponding query instructions based on the target entity and attribute type, a systematic coverage of all facts in the graph is achieved. Furthermore, based on each query instruction and a preset language model, the system searches for semantically consistent declarations in the structured text data, enabling the system to proactively verify whether the text has omitted key facts or contained errors. This reverse-driven mechanism does not depend on whether the text proactively declares a fact, but rather uses the fact graph as a benchmark to reverse-check the completeness of the original text. If no declaration is found or the found declaration is inconsistent with the attribute value, it is marked as an information omission or factual error, thereby effectively capturing the underreporting problem caused by information extraction failure or writing negligence. This complements the forward verification and together constitutes a two-way verification system with high recall.

[0042] Step S5: Based on preset rules, perform multi-dimensional compliance checks on the structured text data and output a third verification result; wherein, the rules include static rules and context-dynamic rules constructed based on the fact knowledge graph.

[0043] Step S6: Generate an audit report based on the first verification result, the second verification result, and the third verification result.

[0044] In practice, to more clearly illustrate the solution of this invention, we will use the promotion of fund products in the financial field as a specific preferred implementation scenario for explanation: Step 1: Multimodal text input and structured preprocessing: To address the challenge of traditional review methods in handling unstructured and semi-structured information, this step transforms diverse promotional materials (such as images, PDFs, and Word documents) into unified, structured text that retains the original layout.

[0045] (1) Unified access and material acquisition: The system provides a flexible access interface that can accept various input formats, including image files (PNG, JPG), document files (Word, PDF), plain text strings, and even cloud storage paths. When the input is a cloud storage path, the system will first automatically download the file from the specified path to the local machine for processing, achieving comprehensive coverage and seamless integration of various promotional materials.

[0046] (2) Classification Processing Engine: After receiving input, the system will automatically distribute it to the optimal processing pipeline according to the file type: A. Deep parsing of document files (Word, PDF): For Word documents, the system directly parses their Document Object Model (DOM), extracting meta-information such as text, tables, and heading levels, preserving the original structure to the greatest extent possible. For PDF files, it prioritizes extracting embedded text streams and table data. For "scanned" PDFs generated from images, it automatically identifies and seamlessly transfers them into the image processing pipeline, ensuring no information loss.

[0047] B. Hybrid Recognition and Structuring of Image Files (PNG, JPG): To overcome the limitations of single technologies, we adopt a hybrid strategy combining Visual Language Modeling (VLM) and high-precision Optical Character Recognition (OCR). First, parallel recognition: the system simultaneously utilizes both the VLM and OCR engines. VLM excels at understanding overall layout and can accurately recognize macro-structures such as headings, paragraphs, tables, and charts; while OCR focuses on high-fidelity character-level recognition, ensuring the accuracy of the text content.

[0048] Secondly, complementary advantages and information fusion: using the Large Language Model (LLM) as a "fusion device", the "structural framework" provided by VLM and the "precise text" provided by OCR are intelligently aligned and filled, which can make up for the minor text errors that VLM may produce and the shortcomings of OCR in understanding the layout logic.

[0049] Finally, after the above processing, all inputs are converted into a unified Markdown text format rich in structured information. This format not only preserves the original text's headings, lists, tables, and other layout information, but also provides a clean, standardized, and machine-readable input for subsequent entity recognition and semantic parsing.

[0050] Step 2: Comprehensive Entity Recognition and Verification Context Construction: Based on standardized text, all financial entities mentioned in the promotional materials and their unique identifiers are comprehensively and accurately identified.

[0051] (1) Comprehensive Financial Entity Recognition (NER): The system invokes a Named Entity Recognition (NER) model deeply optimized for the financial field. This model can identify and extract various types of financial entities from text, such as: Financial products: Funds (such as Yi XX Hengjiu 1-year fixed-term bonds), indices (such as CSI 300), stocks (such as Ningx Times), and fund performance benchmarks (such as 80% × China Dividend Index Return + China Total Index Return × 20%), etc.

[0052] Related parties: Fund company (Yi XX Fund), fund manager (Zhang San), custodian (ICBC), etc.

[0053] Key attributes: Securities code (000265), fund share type (Class A shares), fund abbreviation, fund alias, etc.

[0054] (2) Unique Identifier Association and Context Establishment: While identifying entities, the system will extract their unique identifiers (such as fund codes and stock codes), which are crucial for accurately retrieving information from authoritative data sources. All identified core entities and their identifiers together constitute an "entity set," which fully defines the context of this audit task. Unlike traditional methods that focus on a single protagonist, this invention can handle complex scenarios involving comparisons of multiple products and references to multiple market indices, ensuring the comprehensiveness of the verification. For the example poster, the context constructed by the system will include all the above key information, rather than just locking onto a single fund code.

[0055] (3) Implicit Entity Code Completion: For entities that are identified but whose unique identifiers are not explicitly provided in the text (e.g., only mentioning "SSE 50" without providing the code), the system will call an online code search service. This service uses fuzzy matching and semantic analysis to find and associate the corresponding standard code (e.g., 000016.SH). The retrieved code will be stored in a separate supplementary information field for unified use in subsequent steps.

[0056] Step 3: Construction of a dynamic fact knowledge graph: (1) Data source aggregation: All entity codes identified in the previous step are "anchor nodes". The system initiates API calls to multiple open and trusted information data sources such as Wind and Morningstar through parallel processing to obtain comprehensive information about these entities.

[0057] (2) Information Extraction and Fusion: The system extracts structured information from the JSON-formatted data returned by the data source and transforms it into a unified knowledge graph four-tuple (Subject, Predicate, Object, Time) format. The Time dimension records the time validity of the data; this field is mandatory for time-sensitive indicators (such as performance or years of service); it can be empty for basic static information. For example: (000265, Stock Abbreviation, Yi xx Hengjiu 1-Year Fixed Bond A, null); (000265, Fund Manager Name, E Fund Management Co., Ltd., null); (000265, Custodian Name, ICBC Bank Co., Ltd., null); (000265, Fund Manager's Name, Zang XX, null); (000265, Fund Investment Type Description, Bond Fund, null); (000265, Risk Level, R2-Low to Medium Risk, null); (000265, Investment Objective, This fund aims to create stable returns exceeding the benchmark for fund unit holders while pursuing long-term and steady appreciation of fund assets.) (Zang XX, 20 years of investment research experience as a fund manager, 2025-06-30); (3) Conflict resolution and fact-finding: During the information fusion process, if there is a conflict between information from different data sources, the system resolves the conflict according to the preset priority weights of the data sources to ensure the accuracy of the knowledge graph. Finally, a knowledge graph covering all relevant entities is constructed as the "ground truth" for this review.

[0058] Step 4: Forward retrieval verification: (1) Semantic parsing, quadruple generation, and source tracing: The semantic parsing module utilizes the understanding capabilities of the Large Language Model (LLM) to transform various claims in the text into standardized "quadruples to be verified" (Subject, Predicate, Object, Time) that are completely consistent with the knowledge graph format. The LLM also extracts time adverbs from the claims as the Time dimension. For example: Statement: "Managed by Mr. Zang XX, a fund manager with 15 years of investment experience"... "The above data is as of June 30, 2025"; Generate quadruple 1: (000265, Fund Manager Name, Zang XX, 2025-06-30); Generate Quadruple 2: (Tibet XX, Years of Investment Research Experience of Fund Managers, 15 years, 2025-06-30); Copywriting statement: "The one-year yield of 000265 is 5.0%"; Generate Quadruple 3: (000265, One-year Yield, 5.0%, 2025-06-30); Copywriting statement: "The risk level is medium-low risk"; Generate Quadruple 4: (000265, Risk Level, Medium-low Risk, null); When generating each quadruple, the system will accurately record the specific location of the statement in the structured text (Markdown) generated in Step 1 (such as line number, cell coordinates in a certain table, etc.). This traceability information will be bound to the quadruple to provide a direct context link for subsequent review.

[0059] (2) Preliminary Knowledge Graph Verification: The system uses the "quadruples to be verified" generated in the previous step as a query to perform exact matching or semantic matching in the factual knowledge graph constructed in Step 3. All four dimensions of information will be compared simultaneously during the query to ensure that the data and its corresponding timestamp are accurate.

[0060] For example, for the quadruple (000265, Risk Level, Medium-low Risk, null), the system will query the risk level attribute of the subject 000265 in the graph. The return value of the knowledge graph is R2 - Medium-low Risk, and the LLM will determine that the semantics are consistent, and this verification item passes.

[0061] For example, for the quadruple (Tibet XX, Years of Investment Research Experience of Fund Managers, 15 years, 2025-06-30), the system will query the years of investment research experience attribute of the subject Tibet XX in the graph. The return value of the knowledge graph is 20 years, and the LLM determines that this information is incorrect. This verification item fails, and the output result will record the correct information from the factual knowledge graph.

[0062] If the corresponding information does not exist in the graph, it will enter the next step of in-depth verification.

[0063] (3) Dynamic Interface Matching and In-depth Verification: For information not covered by the knowledge graph (usually dynamically changing performance indicators), the system will initiate an external verification process composed of multiple microservices. The specific steps are as follows: A. Verification Trigger: When a quadruple to be verified cannot be verified in the internal knowledge graph, this in-depth verification process is automatically triggered.

[0064] B. Interface Routing: The system maintains an interface metadata dictionary, which acts like a "routing table," mapping various financial indicators (such as yield, peer ranking, Sharpe ratio, etc.) to the dedicated API interfaces that provide this data. The system will precisely match the microservice interface responsible for verifying the indicator from the dictionary based on the predicate (e.g., "Sharpe ratio") in the quadruple to be verified.

[0065] C. Parameter Construction and API Call: The system extracts the subject (e.g., fund code), predicate, and time from the four-tuple to be verified as parameters to construct a standardized API request. Subsequently, the system calls the interface matched in the previous step in real time and passes in these parameters.

[0066] D. Result Comparison and Adjudication: After obtaining the authoritative data returned by the API, the system compares it with the object in the quadruple to be validated. The Large Language Model (LLM) performs the final semantic adjudication, determining whether the two are consistent in value and expression, thus completing the deep validation of this dynamic metric. If the semantic adjudication result is "incorrect," the output will record the correct information returned by the API.

[0067] (4) Error tracing and characterization: For statements that fail to be verified, the system will initiate a secondary tracing process to distinguish the source of the error. For example, if the text is mistakenly written as "the fund manager is Zhang Lei", the statement will be judged as "error" during the initial verification.

[0068] A. Error Review: The system will use LLM to search again in the original, structured text to confirm that the original text does indeed state "the fund manager is Zhang Lei".

[0069] B. Error Characterization: Through review, the system can confirm that the error is due to an "incorrect statement in the text content itself," rather than an error in the image recognition in step one or the semantic parsing in step four, and will display an error message. This step is crucial for accurately recording the cause of the error and avoiding attribution of errors to upstream modules.

[0070] C. Information Extraction and Correction: If the error review finds that the error was in the previous extraction step, but the text content itself is correct, update the object of the previously extracted quadruple.

[0071] In addition, to maximize efficiency, the above three-step verification process for all statements in the document will be carried out in parallel to shorten the overall review time.

[0072] Step 5: Reverse fact-checking, from fact map to original text: To address potential information omissions or extraction errors in preceding processes (such as text-to-image conversion and information extraction), and to achieve a comprehensive, "covert" review of the text content, the process is as follows: Based on the constructed authoritative factual knowledge graph, a reverse-engineered Large Language Model (LLM) is used to verify the original text of the promotional materials point by point. For example, the knowledge graph clearly states that the custodian of the fund (000265) is ICBC Bank Co., Ltd. The LLM will use this "fact" to search for descriptions of "custodian bank" or "custodian" in the original poster text.

[0073] If the original text describes it as "the custodian bank is ICBC", the LLM will determine that the two are consistent and the verification will pass.

[0074] If the original text is described as "custodian behavior establishes xx bank", the LLM will immediately detect the conflict, mark it as a factual error, and provide the correct information from the fact knowledge graph.

[0075] In this way, even if the OCR recognition in step one misses the line of text "custodian bank", the reverse verification mechanism can still find the inconsistency between the text content and authoritative facts, thereby reducing the false negative rate.

[0076] Step Six: Multi-dimensional Compliance Check: In addition to factual cross-validation, the system also activates a multi-dimensional text compliance check engine. This engine operates independently of factual verification and focuses on identifying compliance risks in promotional materials. It primarily includes the following three levels: (1) Basic Information Completeness Check: The system will check whether the promotional materials contain the necessary basic product information (such as fund code and fund name) and whether the legal compliance information required by regulations is missing, based on the preset template. For example, the "contract effective date" of the fund must be included in the performance display section. If missing, the system will mark it as a "basic information omission" error.

[0077] (2) Text standardization and static rule checking, this level is responsible for handling deterministic text errors: A. Standardized word usage: Through language models, the system can identify and suggest corrections for typos or non-standard terms. For example, it requires the use of “component” instead of “component” in all index-related expressions, and the use of “experience” instead of “experience”.

[0078] B. Banned and Sensitive Words: The system will strictly screen for banned or highly sensitive words based on a dynamically updated local rules file. For example, it will directly prohibit the use of expressions such as "positive annual returns" or "investment alternatives" and issue a sensitive word warning. For superlative statements such as "most" or "unique," the system will mark them and prompt that they need to be manually reviewed to confirm whether they have clear and authoritative data to support them.

[0079] (3) Context-dependent dynamic compliance checks: Unlike traditional rule engines, the system can dynamically judge complex compliance rules by combining contextual information obtained from the fact knowledge graph. A. Conditional Statement Validation: When evaluative terms such as "stable performance" or "top-ranked" appear in the text, the system will query the knowledge graph to check for supporting objective data (such as long-term performance or quartile ranking). If no such data is found, it will be marked as "lacking objective evidence." If the text contains the phrase "good flexibility," the system will check whether it is followed by the necessary annotation, "Investors are advised to pay attention to the risks of index fluctuations."

[0080] B. Product Type Matching Verification: The system will determine the compliance of promotional selling points based on the fund's "fund investment type" and "investment scope" in the knowledge graph. For example, for "pure passive ETFs," the system will prohibit the promotion of "excess returns"; for funds whose investment scope includes convertible bonds, the system will prohibit them from using the selling point of "short-term pure bond fund index outperforming money market fund index."

[0081] C. Performance and Holdings Consistency Verification: When a fund manager's investment philosophy involves a specific industry, the system can further link the latest quarterly report holdings data to verify whether the description of the "circle of competence" matches the actual holdings.

[0082] D. Verification of Annotations and Risk Warnings: The system can identify and verify scenarios requiring special annotations. For example, when mentioning "similar" funds, it checks whether the specific classification criteria for similar funds are annotated; when a fund manager manages products that include institutional customization, it checks whether the performance annotations include the statement that "the sales target scope of some products is also different from that of this fund."

[0083] (4) Compliance modification suggestions: While the system identifies compliance prompts, it will also generate clear modification suggestions (for example, for the statement "good flexibility", it will directly suggest "the note 'Investors should pay attention to the risk of index fluctuations' must be added"), providing operations and compliance personnel with one-stop, actionable modification guidance.

[0084] Step 7: Generate Audit Report Results Summary and Classification: The audit report generation module will summarize the verification results of all previous steps and integrate all the problems found.

[0085] Report Generation: The final result is a comprehensive, structured audit report. The report clearly lists every confirmed error, whether it is a factual error, an omission of information, or a compliance risk.

[0086] As a supplement, when implementing and building the system corresponding to this solution, an interactive review, result presentation, and asynchronous processing architecture can be integrated: To implement the automatic review capability of this invention into a complete, efficient, and user-friendly system, this invention also includes a system architecture that supports front-end interaction, multi-dimensional result display, and high-concurrency task processing.

[0087] Front-end Interaction Interface and Task Submission: The system provides a web front-end interface where reviewers can easily upload promotional materials to be reviewed, supporting single or batch file uploads in various formats such as images, PDFs, and Word documents. After uploading, the task is submitted to the back-end system and enters the preprocessing flow described in step one.

[0088] Multi-dimensional audit result output and visualization: After the report is generated in step seven, the audit results will be presented to the user in multiple formats: A. Original text annotation and traceability of results: For image-based text, the system can directly outline and annotate the error location on the original image; for all text, each issue in the review report provides a precise location reference in the original text, making it easy for users to quickly locate and verify the information. At the same time, a reference information document is generated, which lists all referenced data sources and factual basis in detail.

[0089] B. Clear classification display: The system will classify errors into first-level, second-level, and third-level categories based on their nature. First-level categories include text verification (corresponding to the basic information review results in step four and the review results in step five), indicator verification (corresponding to the data indicator review results in step four), and compliance rules (corresponding to the detection results in step six).

[0090] C. Comprehensive Audit Report: The system generates a detailed audit report in HTML or PDF format, which not only includes a categorized list of all issues, but also provides detailed explanations, compliance evidence, and specific modification suggestions for each issue.

[0091] D. Structured Data and Visualization: All audit results will be output in a structured data format (such as JSON) for easy archiving, analysis, or integration with other systems. Simultaneously, the front-end interface will visualize the audit results through dashboards and other means, helping managers gain a macro-level understanding of the compliance risks associated with the documents.

[0092] Highly reliable asynchronous task processing architecture: To handle a large number of concurrent audit requests, the system backend adopts an asynchronous processing architecture based on message queues. A. Message Queue Mechanism: RabbitMQ is used as the message middleware. User-submitted review tasks are encapsulated as messages and sent to the task queue, decoupling task submission from backend processing. Reviewers can submit new tasks without waiting for processing to complete, improving the user experience.

[0093] B. Message Reliability and Fault Tolerance: Through message persistence, producer confirmation, and consumer acknowledgment mechanisms, the system ensures that audit tasks are not lost or duplicated during transmission and processing. If a processing node fails, the task can be reassigned to another healthy node, guaranteeing high system availability.

[0094] C. Buffer and Backpressure Control: Message queues are used as buffers to support batch reception and operation of massive tasks. When backend processing capacity reaches a bottleneck, a backpressure mechanism can be used to slow down task consumption and prevent system overload caused by excessive instantaneous traffic.

[0095] D. Dynamic Batch Processing and Resource Management: The system dynamically adjusts the batch size of each processing node based on the current queue length and system load. Simultaneously, through refined thread pool management, computing resources are allocated rationally, ensuring processing efficiency while avoiding resource waste.

[0096] In addition, each question in the report is accompanied by its explicit source in the original text. This is achieved by utilizing the original text location information recorded and bound along with the quadruple in step four. When users view the report, they can directly see the original text snippet associated with the error message, or quickly jump to the exact location in the original text (or its structured version) by clicking a link.

[0097] In summary, compared with existing technologies, the above embodiments of this application have the following beneficial effects: By converting product promotional text into structured text data, the problem that unstructured multimodal inputs (such as images and PDFs) cannot be directly understood by machines is solved, providing a unified and parsable input foundation for subsequent automated processing; furthermore, by identifying target entities from structured text and assigning identifiers to them, the core objects in the text are accurately anchored, establishing an index for cross-data source fact queries; simultaneously, by obtaining structured fact data from preset data sources based on target entities and their identifiers and constructing a fact knowledge graph containing logical relationships between entities and the time dimension, a dynamic, authoritative, and context-complete "fact benchmark" is formed, which can avoid the inconsistency problem caused by scattered facts and heterogeneous sources in traditional review; on this basis, by parsing natural language statements to generate format-matched structured data to be verified and performing semantic consistency verification in the fact knowledge graph, deep semantic understanding and accurate fact verification of the text statements are achieved. The comparison improves the accuracy of verification for complex expressions. Furthermore, driven by a fact knowledge graph, it reverse-searches and verifies structured text data, constructing a comprehensive "fact → text" coverage mechanism. This effectively identifies omissions or errors in information extraction, reducing the rate of missed detections during audits. Moreover, by executing multi-dimensional compliance checks that include static rules and context-based dynamic rules built on the fact knowledge graph, compliance judgment is no longer limited to keyword matching. Instead, it dynamically triggers conditional rules based on factors such as product attributes and contextual semantics in the promotional text, achieving intelligent and refined compliance logic. The synergistic effect of these features forms a closed-loop audit process: "structured input—entity anchoring—dynamic fact graph—positive statement verification + reverse fact verification + contextual compliance." This solves the core problem of low semantic verification accuracy caused by insufficient semantic understanding, fragmented fact sources, and static, rigid rules in existing technologies, achieving high-precision, high-coverage automated audits of the authenticity, completeness, and compliance of promotional texts.

[0098] Example 2: Please refer to Figure 2 Based on the same inventive concept, the present invention discloses an automatic review system for product promotional copy, comprising: a data acquisition module M1, a graph construction module M2, a first verification module M3, a second verification module M4, a third verification module M5, and a result output module M6; The data acquisition module M1 is used to acquire and convert product promotional text into structured text data, identify each target entity from the structured text data, and determine the identifier corresponding to each target entity.

[0099] Furthermore, the data acquisition module M1 includes: a parsing unit and a conversion unit; The parsing unit is used to acquire the product promotional text and determine the file type. If it is a document format, it parses the document structure and extracts the text and table content to obtain temporary document data. If it is an image format, it performs layout structure recognition and character content recognition in parallel and merges the recognition results to obtain temporary image data. The conversion unit is used to generate structured text data that retains the original semantic structure and has a uniform format based on the temporary document data or temporary image data.

[0100] In this preferred embodiment, adaptive recognition of heterogeneous input sources is achieved by acquiring product promotional text and determining its file type. Furthermore, the document structure is analyzed to extract text and table content for document formats, and layout structure recognition and character content recognition are performed in parallel for image formats, with the results fused. This effectively balances the semantic integrity of the layout with the accuracy of the character content, overcoming the limitations of a single recognition model in handling complex layouts. Based on this, structured text data that retains the original semantic structure and has a unified format is generated from the processing results. This ensures that machine-readable text with consistent structure and clear hierarchy can be output from different modal inputs, providing a stable and reliable data foundation for subsequent entity recognition and semantic analysis, significantly improving the robustness and consistency of information extraction in multimodal scenarios.

[0101] Furthermore, the data acquisition module M1 also includes: an extraction unit and a missing data filling unit; The extraction unit is used to identify each target entity from the structured text data according to a preset named entity recognition model, and simultaneously extract the identifiers that are explicitly associated with each target entity. The missing identifier unit is used to construct a query request based on the name of a target entity with no explicit association with the identifier, call a preset search interface to perform the query, obtain a list of candidate identifiers, and calculate the string similarity and semantic similarity between each candidate identifier in the list and the name of the target entity, and select the candidate identifier with the highest comprehensive score as the identifier of the target entity; wherein, the comprehensive score is obtained by weighted fusion of string similarity and semantic similarity.

[0102] In this preferred embodiment, the efficient extraction of labeled entities is achieved by extracting explicitly associated identifiers. For entities without explicit identifiers, a query request is constructed based on their names, and a preset search interface is called to obtain a list of candidate identifiers, thus solving the identification difficulties caused by ambiguous entity references or abbreviated expressions. Furthermore, the string similarity and semantic similarity between each candidate identifier and the target entity name are calculated, and the optimal identifier is selected based on the comprehensive score of weighted fusion. String similarity ensures the accuracy of literal matching, semantic similarity understands the semantic association of names through a language model, and the weighted fusion mechanism achieves a balance between precise matching and semantic generalization. Thus, even in scenarios where entity expressions are not standardized, each target entity can still be accurately assigned a unique identifier that can be used for cross-source queries, providing a reliable anchor for the construction of fact knowledge graphs.

[0103] The graph construction module M2 is used to obtain corresponding structured fact data from a preset data source based on each target entity and its corresponding identifier, and to construct a fact knowledge graph that includes logical relationships between entities and the time dimension.

[0104] Furthermore, the map construction module M2 includes: a response acquisition unit, a response parsing unit, a filtering unit, and a construction unit; The response acquisition unit is used to call multiple preset external data source interfaces in parallel based on each target entity and its corresponding identifier to obtain the original response data returned by each external data source. The response parsing unit is used to extract the attribute type, attribute value and time information associated with each target entity from each of the original response data to obtain the corresponding structured fact data. The filtering unit is used to retain the structured fact data with the highest priority among multiple structured fact data with the same target entity and the same attribute type, according to the preset priority of their corresponding external data source interfaces. The construction unit is used to construct a fact knowledge graph containing logical relationships between entities and the time dimension based on all retained structured fact data, with each target entity as the subject, attribute type as the relationship, attribute value as the object of the relationship, and corresponding time information.

[0105] In this preferred embodiment, by parallelly calling multiple preset external data source interfaces based on each target entity and its identifier to obtain raw response data, efficient aggregation of multi-source fact data is achieved, avoiding the latency accumulation of serial queries. Furthermore, attribute type, attribute value, and time information are extracted from the raw response data to obtain structured fact data, and the heterogeneous interface returns are uniformly transformed into standardized fields. For multiple structured fact data with the same target entity and attribute type, the highest priority data is retained according to the preset priority of its corresponding external data source interface, effectively solving the multi-source data conflict problem and ensuring the consistency and authority of the factual benchmark. Based on this, a factual knowledge graph is constructed based on all retained structured fact data, with the target entity as the subject, attribute type as the relation, attribute value as the relational object, and time information, forming a dynamic semantic network containing logical relationships between entities and time validity. This provides a structured and queryable factual basis for subsequent two-way verification and contextual rule judgment.

[0106] The first verification module M3 is used to parse each natural language statement in the structured text data, generate each structured data to be verified whose format matches the fact knowledge graph, perform semantic consistency verification in the fact knowledge graph, and output the first verification result.

[0107] Furthermore, the first verification module M3 includes: a forward verification unit; The forward verification unit is used to extract the declared target entity, attribute type, attribute value, and time information for each structured data to be verified, and to query whether there is structured fact data with the same declared target entity and attribute type in the fact knowledge graph. If there is, based on a preset language model, the declared attribute value is semantically consistent with the attribute value in the corresponding structured fact data, and a first verification result is output. Otherwise, based on the declared target entity, attribute type, and time information, a query request is constructed, and the corresponding external verification interface is called from the preset interface metadata dictionary to obtain the returned attribute value data. The language model is then called to perform a semantic consistency judgment between the returned attribute value data and the declared attribute value, and a first verification result is output.

[0108] In this preferred embodiment, by extracting the target entity, attribute type, attribute value, and time information of each structured data to be verified, and querying the fact knowledge graph to see if there is structured fact data with the same target entity and attribute type, accurate positioning and preliminary matching of the declaration are achieved. If a match exists, the semantic consistency of the declared attribute value and the graph attribute value is judged based on the preset language model, so that the system can understand semantically equivalent but different content, overcoming the limitations of strict string matching. If there is no match in the graph, a query request is constructed based on the target entity, attribute type, and time information of the declaration, and the corresponding external verification interface is called from the preset interface meta-information dictionary to obtain the attribute value data. The language model is called again to judge semantic consistency, thereby extending the verification capability from static graph to dynamic indicators, significantly improving the coverage breadth and depth of fact verification.

[0109] Furthermore, the first verification module M3 also includes: a positioning unit and a verification unit; The tracing unit is used to relocate the original declaration text fragment corresponding to the structured text data to be verified in the structured text data when performing semantic consistency verification and outputting the verification failure result. The verification unit is used to compare the attribute value text content in the original declaration text fragment with the attribute value in the structure data to be verified; if they match, the verification failure result is retained; otherwise, the attribute value in the structure data to be verified is replaced with the attribute value text content in the original declaration text fragment to obtain the corrected structure data to be verified, and the semantic consistency verification is re-executed.

[0110] In this preferred embodiment, when the semantic consistency check outputs a failure result, the source of the error is traced by relocating the original declaration text fragment corresponding to the structured data to be checked in the structured text data. Furthermore, the attribute value text content in the original declaration text fragment is compared with the attribute values ​​in the structured data to be checked. If they match, the failure result is retained, indicating that the error originates from the text content itself, ensuring the authenticity of the review result. If they do not match, the attribute values ​​in the structured data to be checked are replaced with the attribute value text content in the original declaration text fragment, resulting in the corrected structured data to be checked, and the semantic consistency check is re-executed. This corrects the misjudgment caused by the information extraction deviation in the preceding steps, effectively distinguishing between "content error" and "system extraction error," significantly improving the reliability and traceability of the review result.

[0111] The second verification module M4 is used to reverse retrieve and verify the structured text data driven by the fact knowledge graph, and output the second verification result.

[0112] Furthermore, the second verification module M4 includes: a condition construction unit and a reverse verification unit; The condition construction unit is used to traverse each structured fact data in the fact knowledge graph, and for each structured fact data, construct a corresponding query instruction based on its target entity and attribute type as query conditions. The reverse verification unit is used to search in the structured text data for whether there is a statement content that is semantically consistent with the structured fact data corresponding to each query instruction, based on each query instruction and combined with a preset language model. If no statement content is found or the found statement content is inconsistent with the corresponding attribute value, it is marked as information omission or fact error.

[0113] In this preferred embodiment, by traversing the structured fact data in the fact knowledge graph and constructing corresponding query instructions based on the target entity and attribute type, a systematic coverage of all facts in the graph is achieved. Furthermore, based on each query instruction and a preset language model, the system searches for semantically consistent declarations in the structured text data, enabling the system to proactively verify whether the text has omitted key facts or contained errors. This reverse-driven mechanism does not depend on whether the text proactively declares a fact, but rather uses the fact graph as a benchmark to reverse-check the completeness of the original text. If no declaration is found or the found declaration is inconsistent with the attribute value, it is marked as an information omission or factual error, thereby effectively capturing the underreporting problem caused by information extraction failure or writing negligence. This complements the forward verification and together constitutes a two-way verification system with high recall.

[0114] The third verification module M5 is used to perform multi-dimensional compliance checks on the structured text data based on preset rules and output the third verification result; wherein, the rules include static rules and context dynamic rules constructed based on the fact knowledge graph.

[0115] The result output module M6 is used to generate an audit report based on the first verification result, the second verification result, and the third verification result.

[0116] In summary, compared with existing technologies, the embodiments of this application have the following beneficial effects: By converting product promotional text into structured text data, the problem that unstructured multimodal inputs (such as images and PDFs) cannot be directly understood by machines is solved, providing a unified and parsable input foundation for subsequent automated processing; furthermore, by identifying target entities from structured text and assigning identifiers to them, the core objects in the text are accurately anchored, establishing an index for cross-data source fact queries; simultaneously, by obtaining structured fact data from preset data sources based on target entities and their identifiers and constructing a fact knowledge graph containing logical relationships between entities and the time dimension, a dynamic, authoritative, and context-complete "fact benchmark" is formed, which can avoid the inconsistency problem caused by scattered facts and heterogeneous sources in traditional review; on this basis, by parsing natural language statements to generate format-matched structured data to be verified and performing semantic consistency verification in the fact knowledge graph, a deep semantic understanding and accurate fact comparison of the text statements are achieved. Yes, it improves the accuracy of verifying complex expressions. Furthermore, by using a fact knowledge graph as a driving force to reverse-search and verify structured text data, a fallback mechanism from "facts → text" is constructed, effectively identifying omissions or errors in information extraction and reducing the rate of missed audits. Moreover, by executing multi-dimensional compliance checks that include static rules and contextual dynamic rules built on the fact knowledge graph, compliance judgment is no longer limited to keyword matching. Instead, it can dynamically trigger conditional rules based on factors such as product attributes and contextual semantics in the promotional text, achieving intelligent and refined compliance logic. The synergistic effect of these features forms a closed-loop audit process of "structured input—entity anchoring—dynamic fact graph—positive statement verification + reverse fact verification + contextual compliance," solving the core problem of low semantic verification accuracy caused by insufficient semantic understanding, fragmented fact sources, and static rigid rules in existing technologies. This achieves high-precision, high-coverage automated auditing of the authenticity, completeness, and compliance of promotional texts.

[0117] Example 3: This invention also provides a computer program product, including a computer program or instructions, capable of running on a computing device or stored in any available medium. When the computer program product is run on at least one computing device, it causes the at least one computing device to execute any of the product promotional materials automatic review methods of this invention.

[0118] Example 4: This invention also provides a computer-readable storage medium storing at least one executable instruction. When the executable instruction is executed on an automatic review system for product promotional materials, the automatic review system for product promotional materials performs an automatic review method for product promotional materials according to any of the above method embodiments.

[0119] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. Similarly, for the purpose of simplification and aiding understanding of one or more aspects of the invention, in the above description of exemplary embodiments of this application, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0120] Those skilled in the art will understand that the modules in the system of the embodiments can be adaptively changed and placed in one or more systems different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

Claims

1. An automatic auditing method of a product promotion script, characterized by, The method comprises the following steps: obtaining and converting product promotion texts into structured text data, identifying target entities from the structured text data, and determining identifiers corresponding to the target entities; based on each of the target entities and the identifier corresponding thereto, obtaining corresponding structured fact data from a preset data source, and constructing a fact knowledge graph containing logical relationships between entities and a time dimension; parsing each natural language statement in the structured text data to generate each to-be-verified structured data matching the format of the fact knowledge graph, and performing semantic consistency verification in the fact knowledge graph, and outputting a first verification result; driven by the fact knowledge graph, reverse searching and checking the structured text data, and outputting a second verification result; based on a preset rule, performing multi-dimensional compliance checking on the structured text data, and outputting a third verification result; wherein the rule includes a static rule and a context dynamic rule constructed based on the fact knowledge graph; based on the first verification result, the second verification result and the third verification result, generating an audit report.

2. The automatic review method of a product promotion script according to claim 1, wherein, The method comprises the following steps: obtaining the product promotion texts and determining the file type, if it is a document format, parsing the document structure and extracting the text and table content to obtain document temporary data, if it is an image format, performing parallel layout structure recognition and character content recognition, and fusing the recognition results to obtain image temporary data; based on the document temporary data or the image temporary data, generating structured text data that retains the original semantic structure and has uniform format.

3. The method of claim 1, wherein the product promotion script is automatically audited by the system. The method comprises the following steps: according to a preset named entity recognition model, identifying each target entity from the structured text data and synchronously extracting identifiers explicitly associated with each target entity; for a target entity without an explicitly associated identifier, constructing a query request based on its name, calling a preset search interface to query, obtaining a list of candidate identifiers, and respectively calculating the string similarity and semantic similarity between each candidate identifier in the list of candidate identifiers and the name of the target entity, selecting the candidate identifier with the highest comprehensive score as the identifier of the target entity; wherein the comprehensive score is obtained by weighted fusion based on the string similarity and the semantic similarity.

4. The method of claim 1, wherein the product promotion script is automatically audited by the system. The method comprises the following steps: based on each of the target entities and the identifier corresponding thereto, obtaining corresponding structured fact data from a preset data source, and constructing a fact knowledge graph containing logical relationships between entities and a time dimension; based on each of the target entities and the identifier corresponding thereto, calling multiple preset external data source interfaces in parallel to obtain original response data returned by each external data source; for each of the original response data, extracting attribute types, attribute values and time information associated with each target entity to obtain corresponding structured fact data; For a plurality of structured fact data with the same target entity and the same attribute type, the structured fact data with the highest priority is retained according to the preset priority of the corresponding external data source interface; Based on all the retained structured fact data, a fact knowledge graph containing logical relationships between entities and time dimensions is constructed, taking each target entity as the subject, the attribute type as the relationship, and the attribute value as the relationship object, and combining the corresponding time information.

5. The automatic review method for product promotional copy as described in claim 4, characterized in that, The semantic consistency verification in the fact knowledge graph outputs a first verification result, including: For each to-be-verified structured data, the declared target entity, attribute type, attribute value and time information are extracted, and it is queried in the fact knowledge graph whether there is structured fact data with the same declared target entity and attribute type; if there is, the declared attribute value and the attribute value in the corresponding structured fact data are judged for semantic consistency based on a preset language model, and a first verification result is output; otherwise, a query request is constructed according to the declared target entity, attribute type and time information, a corresponding external verification interface is matched and called from a preset interface meta information dictionary, attribute value data returned is obtained, and the language model is called to judge the semantic consistency between the returned attribute value data and the declared attribute value, and a first verification result is output.

6. The automatic review method of an advertising copy according to claim 5, wherein When performing semantic consistency verification and outputting a verification failure result, it further includes: Repositioning the original declaration text segment corresponding to the to-be-verified structured data in the structured text data; Comparing the attribute value text content in the original declaration text segment with the attribute value in the to-be-verified structured data; if they are consistent, the verification failure result is retained; otherwise, the attribute value in the to-be-verified structured data is replaced with the attribute value text content in the original declaration text segment to obtain modified to-be-verified structured data, and semantic consistency verification is performed again.

7. The automatic review method for product promotional copy as described in claim 4, characterized in that, The fact knowledge graph is driven to reverse search and check the structured text data, and a second verification result is output, including: Traversing each structured fact data in the fact knowledge graph, taking the target entity and attribute type of each structured fact data as query conditions to construct a corresponding query instruction; Based on each query instruction, combining a preset language model, it is queried in the structured text data whether there is structured fact data with semantic consistency corresponding to each query instruction; if not found or the found declaration content is inconsistent with the corresponding attribute value, it is marked as information missing or fact error.

8. An automatic auditing system for product promotion copy, characterized by, It includes: A data acquisition module, a graph construction module, a first verification module, a second verification module, a third verification module and a result output module; The data acquisition module is configured to acquire and convert product promotional copy into structured text data, identify each target entity from the structured text data, and determine the identifier corresponding to each target entity. The graph construction module is used to obtain corresponding structured fact data from a preset data source based on each target entity and its corresponding identifier, and to construct a fact knowledge graph containing logical relationships between entities and the time dimension. The first verification module is used to parse each natural language statement in the structured text data, generate each structured data to be verified whose format matches the fact knowledge graph, perform semantic consistency verification in the fact knowledge graph, and output the first verification result; The second verification module is used to reverse retrieve and verify the structured text data driven by the fact knowledge graph, and output the second verification result; The third verification module is used to perform multi-dimensional compliance checks on the structured text data based on preset rules and output the third verification result; wherein, the rules include static rules and context-dynamic rules constructed based on the fact knowledge graph; The result output module is used to generate an audit report based on the first verification result, the second verification result, and the third verification result.

9. A computer program product comprising computer programs or instructions, characterized in that, When the computer program or instructions are executed, they implement an automatic review method for product promotional copy as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements an automatic review method for product promotional copy as described in any one of claims 1-7.

Images