File informatization integration method and system based on deep learning

Abstract

The invention discloses an archive informatization integration method and system based on deep learning, and relates to the technical field of computers. The method comprises the following steps: receiving a paper archive digital scanning image, paper archive layout structure information, electronic archive text data, photo archive image data, an audio and video archive and a transfer text thereof, external business system structured business data, archive original metadata fields, user identity information and user role information; and submitting a query request and a cross-modal query request by a user. According to the method, unified format analysis, text standardization, image correction and metadata verification are performed on heterogeneous archive data from different sources and different formats, so that the problems of fragmentation and non-structuring of the archive data are solved, and standardization and high-quality input of the archive data are realized; performing semantic coding on the multi-modal archive data by using a deep learning technology to generate an archive semantic vector set;

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for integrating archival information based on deep learning. Background Technology

[0002] With the rapid development of information technology, the construction of archival information systems has become a key link for various industries to improve management efficiency and realize the value of data. Traditional archival management systems face numerous challenges when dealing with massive, heterogeneous, and multimodal archival data. Archival data comes from diverse sources, including digitized images of paper archives, electronic documents, photographs, audio and video, as well as structured data from external business systems. These data have inconsistent formats and complex content structures, making data integration difficult and hindering unified management and efficient utilization.

[0003] Existing archival systems have limited semantic understanding capabilities, relying primarily on keyword matching for retrieval. This makes it difficult to meet users' deep, multi-dimensional, and cross-modal query needs. Users prefer to find relevant archives through natural language descriptions or image content, which traditional systems struggle to support. Archival data often involves sensitive information, requiring strict access permissions and data security. Existing systems are inadequate in refined access control, sensitive information anonymization, and risk assessment, making it difficult to fully utilize archives while ensuring data security. These problems severely restrict the in-depth development of archival informatization, hindering the full realization of the potential value of archives. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for integrating archival information based on deep learning.

[0005] To achieve the above objectives, this invention adopts the following technical solution: a deep learning-based archival information integration system, comprising the following modules: an archival preprocessing module, which receives digitized scanned images of paper archives, paper archive layout information, electronic archive text data, photographic archive image data, audio and video archives and their transcribed text, structured business data from external business systems, original archival metadata fields, user identity information, user role information, user-submitted query requests, and cross-modal query requests; and performs preprocessing through archival format parsing, character set and text normalization, image denoising and layout correction, and archival metadata normalization and integrity verification to generate a standardized archival preprocessing dataset; a semantic index construction module, which receives the standardized archival preprocessing dataset and generates an archival semantic vector set, inverted index, vector index, graph index, and version time index through archival semantic encoding, archival feature standardization, and multi-modal index construction; and a knowledge graph module, which receives the standardized archival preprocessing dataset. The system comprises four modules: a semantic vector set and a graph index, which generate an archival knowledge graph through entity extraction, relation extraction, and incremental updates; a semantic relation retrieval module, which receives user-submitted query requests, inverted indexes, vector indexes, the archival semantic vector set, graph index, and archival knowledge graph, and generates a set of candidate results for semantic and relational fusion retrieval through query semantic parsing and rewriting, joint semantic retrieval ranking, and relational path retrieval; a cross-modal retrieval module, which receives user-submitted cross-modal query requests, the archival semantic vector set, vector index, graph index, and archival knowledge graph, and generates a set of cross-modal retrieval results through cross-modal representation alignment and cross-modal similarity calculation; and an access control module, which receives user identity information, user role information, a standardized archival metadata set, a set of candidate results for semantic and relational fusion retrieval, and a set of cross-modal retrieval results, and generates a final retrieval result list through access control rule matching, retrieval result filtering and desensitization, and risk assessment.

[0006] As a further description of the above technical solution: The document preprocessing module performs unified parsing on digitized scanned images of paper archives, paper archive layout information, electronic archive text data, photographic archive image data, audio and video archives and their transcribed text, structured business data from external business systems, and original metadata fields of the archives, generating image parsing objects, text parsing objects, structured field objects, and format parsing objects. It then performs character set unification and text standardization processing on the parsed text content, completing character encoding unification, number and date format standardization, character morphology normalization, and abnormal character removal to form standardized text content. Finally, it performs noise reduction and layout correction processing on the parsed image content, completing noise suppression and tilt correction. The system enhances the layout and obtains corrected image content through binarization. It performs standardization and integrity checks on the original archival metadata fields and the structured business data from external business systems, generating a standardized archival metadata set containing the record number, case number, responsible person identifier, creation time, retention period, and security classification identifier, and generating a unique archival record identifier for each record. It then organizes and encapsulates the standardized text content, corrected image content, structured business data, standardized archival metadata, format parsing objects, user identity information, user role information, user-submitted query requests, and user-submitted cross-modal query requests in a unified manner, outputting a standardized archival preprocessing dataset.

[0007] As a further description of the above technical solution: The semantic indexing construction module receives a standardized preprocessed archival dataset and generates multimodal representations of text content, image content, image content, and structured business data. This results in text modality representation, visual modality representation, and field representation within a unified semantic space. Through multimodal co-coding, archival metadata guidance constraints, entity alignment constraints, and noise suppression, a set of archival semantic vectors with cross-modal consistency and business semantic expression capabilities is formed. Feature standardization processing is then performed on this set of archival semantic vectors to ensure a unified and standardized vector representation in terms of scale and distribution. Based on standardized text content... A reverse index is built, maintaining the mapping relationship between terms, location distribution, and field affiliation. A vector index is constructed based on the set of archival semantic vectors to support nearest neighbor retrieval of archival semantic vectors. A graph index is constructed based on the set of entities and the set of relations, maintaining the searchability of entity-level semantic associations by maintaining the association relationship between entity nodes, relation edges, and archival record identifiers. A version time index is constructed based on archival record identifiers and their creation time, archiving time, and version time, giving the version evolution of archives a traversable structure in the time dimension. The output includes the set of archival semantic vectors, the reverse index, the vector index, the graph index, and the version time index.

[0008] As a further description of the above technical solution: The knowledge graph module receives a standardized preprocessed dataset of archives, a set of archival semantic vectors, and a graph index. It uses standardized text content and a set of standardized archival metadata as input for entity and relation recognition, the set of archival semantic vectors as input for cross-modal consistency judgment, and the graph index as the indexing basis for entity nodes and relation edges. It performs entity recognition processing on the standardized text content, identifying names of persons, organizations, places, projects, document types, responsible persons, keywords, and time expressions. It then combines this with the archival metadata set's record number, file number, responsible person identifier, ownership unit identifier, and creation time to generate an entity set. Based on the archival semantic vector set, it performs consistency alignment between text entities and metadata entities, unifying entities from different sources into the entity set. Finally, it generates formation relationships, participation relationships, citation relationships, custody relationships, derivation relationships, and inheritance relationships between entities based on the text's semantic structure and contextual relationships, combining these with the creation time, change time, or archiving time. Generate a set of relationships; perform entity merging, relationship construction, and temporal attribute organization on the entity set and relationship set through archival knowledge graph construction and incremental update algorithms. In the entity merging stage, entities are unified based on textual semantic consistency, metadata consistency, contextual theme consistency, and entity category consistency. In the relationship construction stage, relationship edges are constructed according to the relationship subject entity, relationship object entity, relationship type, and temporal attribute, and relationship instances are organized into a relationship evolution structure in chronological order. For newly added archives, a local candidate graph is constructed, and the creation of new entity nodes or the updating of the temporal chain of existing entity nodes is determined based on entity consistency. At the same time, the generation of new relationship instances or the insertion of relationship instances into the existing relationship list is determined based on the relationship subject, relationship object, and relationship type. In combination with the visual semantic information in the archival semantic vector set, cross-modal consistency verification is performed on the entity candidates in the image content and the text entities. Conflicting relationships are recorded as relationship evolution branches, and the archival knowledge graph is output.

[0009] As a further description of the above technical solution: The semantic relation retrieval module receives user-submitted query requests, inverted indexes, vector indexes, a set of archival semantic vectors, a graph index, and an archival knowledge graph. It performs semantic parsing and structured processing on the user-submitted query requests, identifying target entities, attribute constraints, relational constraints, and time constraints in the query. The parsing results are then transformed into an executable query plan, mapping keyword conditions, semantic similarity conditions, and entity and relational constraints to the inverted index, vector index, graph index, and archival knowledge graph, respectively. Based on the query plan, it generates a set of candidate archives for text matching using the inverted index and a set of candidate archives for semantic similarity using the vector index. Finally, it fuses textual and semantic features. Simultaneously, the candidate archives are weighted and adjusted based on the quality markers of the archive semantic vector set to form a fusion-ranked candidate archive result. Combining the entity adjacency relationship, relationship type, relationship path structure, and relationship time attribute in the archive knowledge graph, the consistency of the relevance between the candidate archives and the query intent at the entity and relationship levels is determined, and multi-evidence fusion ranking is performed on the candidate archives accordingly. After obtaining the ranking result, the relationship path traversal is performed based on the position of the query entity in the archive knowledge graph to identify archive nodes that have a valid association or time chain association with the query entity. The association results and the fusion ranking results are deduplicated and integrated to output a set of semantic and relation fusion retrieval candidate results.

[0010] As a further description of the above technical solution: The cross-modal retrieval module receives cross-modal query requests, a set of archival semantic vectors, a vector index, and an archival knowledge graph submitted by the user. It parses the cross-modal query requests, determines the modality type of the query content, and maps the query content to query vectors in the same representation space as the set of archival semantic vectors. Based on the query vectors, it performs nearest neighbor retrieval in the vector index, identifying archival semantic vectors that are close to the query vector, forming a cross-modal candidate archive set. It supports retrieval methods such as querying image archives by text and querying text archives by image. Combining the archival knowledge graph, it performs consistency verification on the entities corresponding to the cross-modal candidate archives at the levels of entity adjacency, relationship type, and semantic association structure, filtering out candidate results with insufficient association or incomplete supporting information. Based on the vector retrieval results and entity association verification results, it sorts the cross-modal candidate archive set and outputs a set of cross-modal retrieval results.

[0011] As a further description of the above technical solution: The access control module receives user identity information, user role information, a standardized archive metadata set, a semantic and relational fusion search candidate result set, a cross-modal search result set, and business rules and security policy configuration data. It parses the user identity information and user role information into user access attributes, and parses the security classification, retention period, access conditions, affiliated institution, responsible person identifier, and creation time in the standardized archive metadata set into archive security attributes. It associates the archive records in the semantic and relational fusion search candidate result set and the cross-modal search result set with their corresponding archive security attributes, and matches the user access attributes with the archive security attributes based on the business rules and security policy configuration data to determine the access permission category for each archive record. After completing the access permission determination... The system filters out archives deemed prohibited from viewing, and performs field masking, content blurring, or entity anonymization on archives deemed permissible for viewing with desensitization based on field sensitivity and desensitization rules in the standardized archive metadata set. Sensitive areas involving image content are also partially blurred, masked, or have thumbnail replaced, forming an intermediate search result set. After obtaining the intermediate search result set, the system performs risk assessment and order adjustment on the archives in the intermediate search result set based on the security classification, retention period, access conditions, responsible party identification, and creation time in the standardized archive metadata set, combined with the entity and relationship structure in the archive knowledge graph, and the access behavior characteristics reflected by user identity information and user role information, outputting the final search result list.

[0012] As a further description of the above technical solution: Business rules and security policy configuration data include access level rules, business domain authorization rules, field sensitivity level rules, security classification rules, open condition rules, retention period access restriction rules, de-identification policy rules, field masking rules, content obfuscation rules, entity anonymization rules, image sensitive area processing rules, and risk assessment rules. Access level rules define the range of access levels that can correspond to user identity information and user role information. Business domain authorization rules limit the authorization boundaries for different positions, institutions, or business lines in the archival utilization scenario. Field sensitivity level rules classify the sensitivity of fields in the standardized archival metadata set. Security classification rules map the archive's security classification identifier to the access level. The rules define the control level, access conditions, and restrictions on the display of archives. The retention period access restriction rules specify the access visibility of archives under different retention periods. The desensitization strategy rules indicate the desensitization methods used when desensitized browsing is permitted. Field masking, content blurring, and entity anonymization rules describe desensitization methods at the field and content levels. Image sensitive area processing rules define the methods for performing local blurring, masking, or thumbnail replacement on image content containing sensitive areas. The risk assessment rules map the archive's security classification, access conditions, responsible party identification, creation time, archival group number, file number, and the archive's relationship path information in the archive's knowledge graph as the basis for risk scoring.

[0013] As a further description of the above technical solution: A deep learning-based method for integrating archival information includes the following steps: The system receives digitized scanned images of paper archives, layout structure information of paper archives, text data of electronic archives, image data of photographic archives, audio and video archives and their transcribed text, structured business data from external business systems, original metadata fields of archives, user identity information, user role information, user-submitted query requests, and user-submitted cross-modal query requests. It performs unified parsing, text normalization, image denoising and layout correction, and archive metadata normalization and integrity verification on the above data to generate a standardized archive preprocessing dataset. Upon receiving the standardized archive preprocessing dataset, it generates semantic representations of the normalized text content, image content, image content, and structured business data, and performs feature normalization processing on the generated semantic representations to construct an archive semantic vector set. Simultaneously, it constructs a reverse index, vector index, graph index, and version time index based on the data. The system also receives the standardized archive preprocessing dataset and archive language... The system uses semantic vector sets and graph indexes to identify entities and relationships in archives, and generates an archive knowledge graph through construction and incremental updates. It receives user-submitted query requests and, using inverted indexes, vector indexes, archive semantic vector sets, graph indexes, and the archive knowledge graph, parses and retrieves these requests to generate a set of semantic and relational fusion search candidate results. It also receives user-submitted cross-modal query requests and, using the archive semantic vector sets, vector indexes, and the archive knowledge graph, processes these requests to generate a set of cross-modal search results. Finally, it receives user identity information, user role information, standardized archive metadata sets, the set of semantic and relational fusion search candidate results, and the set of cross-modal search results. Combining this with business rules and security policy configuration data, it performs access permission checks, result filtering and anonymization, and result order adjustments on the search results, outputting the final search result list.

[0014] The present invention has the following beneficial effects: 1. In this invention, heterogeneous archival data from different sources and in different formats are subjected to unified format parsing, text normalization, image correction, and metadata verification. This solves the problems of fragmented and unstructured archival data, lays the foundation for subsequent deep semantic analysis, and achieves the standardization and high-quality input of archival data. By using deep learning technology to semantically encode multimodal archival data, a unified set of archival semantic vectors is generated, and inverted indexes, vector indexes, graph indexes, and version time indexes are constructed. This greatly improves the semantic understanding ability and retrieval efficiency of archival content and supports more complex semantic queries.

[0015] 2. In this invention, entities and relationships are automatically extracted from archival data to construct an archival knowledge graph. This organizes discrete archival information into a structured knowledge network, revealing deep connections between archives. This provides strong support for advanced semantic retrieval and knowledge discovery, enabling users to perform semantic queries based on natural language and cross-modal queries (such as text-to-image search and image-to-text search). Combined with the knowledge graph, relational path retrieval is performed, significantly improving retrieval accuracy, recall rate, and user experience. This meets diverse user retrieval needs, achieves refined permission management based on user identity, role, and archival security attributes, and supports the filtering, de-identification, and risk assessment of sensitive information. Under the premise of ensuring archival data security and compliance, the utilization value of archival information is maximized. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0017] 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.

[0018] Reference Figure 1This invention provides an embodiment of a deep learning-based archival information integration system, comprising the following modules: an archival preprocessing module receives digitized scanned images of paper archives, paper archive layout information, electronic archive text data, photographic archive image data, audio and video archives and their transcribed text, structured business data from external business systems, original archival metadata fields, user identity information, user role information, user-submitted query requests, and cross-modal query requests. It performs preprocessing using archival format parsing algorithms, character set and text normalization algorithms, image denoising and layout correction algorithms, and archival metadata normalization and integrity verification algorithms to generate a standardized archival preprocessing dataset; a semantic indexing construction module receives the standardized archival preprocessing dataset and generates an archival semantic vector set, inverted index, vector index, graph index, and version time index using archival semantic encoding algorithms, archival feature normalization algorithms, and multi-modal indexing construction algorithms; a knowledge graph module receives the standardized archival preprocessing dataset and archival semantic data. The system comprises three modules: a vector set and a graph index; an archive knowledge graph generated through algorithms for extracting archive entities, extracting archive relationships, and constructing and incrementally updating the archive knowledge graph; a semantic relationship retrieval module receiving user-submitted query requests, inverted indexes, vector indexes, archive semantic vector sets, graph indexes, and the archive knowledge graph; a query semantic parsing and rewriting algorithm, a joint semantic retrieval ranking algorithm, and a relationship path retrieval algorithm; a cross-modal retrieval module receiving user-submitted cross-modal query requests, archive semantic vector sets, vector and graph indexes, and the archive knowledge graph; and a cross-modal retrieval result set generated through cross-modal representation alignment and cross-modal similarity calculation algorithms; and an access control module receiving user identity information, user role information, a standardized archive metadata set, a semantic and relationship fusion retrieval candidate result set, and a cross-modal retrieval result set. It generates the final retrieval result list through access control rule matching algorithms, retrieval result filtering and desensitization algorithms, and risk assessment algorithms.

[0019] The input data for the archive preprocessing module includes digitized scanned images of paper archives obtained from the data access interface, paper archive layout structure information, electronic archive text data, photo archive image data, audio and video archives and their corresponding transcribed text, structured business data output by external business systems, and original archive metadata fields. It also includes user identity information, user role information, user-submitted query requests, and cross-modal query requests obtained through the user input interface.

[0020] The document preprocessing module sequentially calls the document format parsing algorithm, character set and text normalization algorithm, image denoising and layout correction algorithm, and document metadata normalization and integrity verification algorithm. It then completes the format unification and semantic cleaning of multi-source heterogeneous data through layer-by-layer structured processing. During the implementation of the document format parsing algorithm, the module reads the resolution, color channels, image size, and page number sequence of the digitized scanned image of the paper document based on its file format. It then identifies areas such as the main text area, text area, margins, binding area, header area, and footer area based on the preset layout template in the paper document's page structure information. Based on the identified areas, the module uses geometric positioning rules to extract the corresponding pixel matrix, separating the image content and page structure into independent structured elements. The module performs the same parsing operation for photographic archive images to maintain the consistency of image-based document processing. When processing electronic document text data, the module defines the page hierarchy, object hierarchy, and text block hierarchy based on the internal file structure of the electronic document format. It extracts all text content, heading structure, paragraph organization, table content, and embedded objects into structured text parsing objects, ensuring that subsequent standardization processing can be performed within clear semantic boundaries. When processing audio and video archives, the module separates the video track, audio track, and metadata track through an audio / video decapsulation program, focusing on reading their corresponding transcribed text and timeline information. This ensures that audiovisual archive content also enters the subsequent semantic processing flow in machine-readable text form. For structured business data output from external business systems, the module reads field names, field types, field values, and field constraints according to field mapping rules, extracting key fields describing the archive's business background into structured field objects. The archive format parsing algorithm ultimately outputs structured parsing objects that can simultaneously carry image content, text content, structured field content, and page layout content, forming the input basis for subsequent algorithms. In the implementation of the character set and text normalization algorithm, the module takes the text content in the structured parsing object as input. First, it executes a character encoding detection program to determine the source encoding format of the text and uniformly converts encodings from different sources to Unicode to avoid unrecognizable or garbled characters in subsequent processing stages. Next, the module performs syntactic-level analysis on numerical expressions in the text, splitting numerical values ​​with units into numerical and unit parts and unifying units according to a standardized unit table. For date expressions, it performs pattern recognition, uniformly converting different writing styles such as Chinese dates, numeric dates, and mixed symbol dates into the standard format while maintaining their semantic position in the document. Finally, the module performs full-width and half-width character normalization on the text, converting full-width Chinese numbers, full-width English characters, and full-width punctuation into half-width forms, making the text more standardized and consistent at the character level.In the noise processing stage, the character set and text normalization algorithm identifies garbled fragments, non-printing control characters, repetitive meaningless symbols, and OCR misaligned characters according to preset noise character recognition rules. It then performs replacement or error marking based on the contextual semantics and character morphology of the misaligned characters, allowing the text content to correct errors as much as possible while preserving its original meaning. The normalized text content processed by this algorithm possesses attributes such as clear structure, consistent expression, and semantic recognizability. In the implementation of the image denoising and layout correction algorithm, the module first performs multi-scale smoothing filtering on the original pixel matrix of the scanned image and photograph to simultaneously suppress random noise and local speckle noise at different scales. Subsequently, bilateral filtering is performed to preserve text edges, line structures, and table borders, improving image clarity while avoiding damage to visual details. After denoising, the module performs a tilt detection step, using Hough transform or text line direction histogram to estimate the overall tilt angle of the page, and performs geometric rotation on the image accordingly to align the text lines and page layout with the horizontal axis. After correction, the module performs binarization enhancement by selecting a global threshold or adaptive threshold algorithm based on the image background texture features. This significantly improves the contrast between the text and background areas in the brightness space, enhancing the subsequent performance of OCR or visual feature encoding. Finally, the module re-aligns the corrected image content with the previously recognized layout structure information, outputting image content that is structurally sound, geometrically stable, and highly readable. In the implementation of the archival metadata standardization and integrity verification algorithm, the module first standardizes the original metadata fields according to the standardization rules for metadata fields such as archival group number, file number, responsible party identifier, creation time, retention period, and security classification. This unifies the field names, field value codes, and date formats from different source systems, allowing metadata to be processed in a unified field table. Subsequently, the module performs consistency and integrity verification based on the dependencies and logical constraints between metadata fields. For example, it determines the archival group to which the archive belongs based on the combination of the file number and archival group number; cross-validates the responsible party identifier with the main information in the structured business data of external business systems; and determines whether the archive is still in its retention period based on the combination of the creation time and retention period. For fields whose missing values ​​can be inferred from the context, the module completes them according to preset inference rules; for fields that cannot be completed and affect the determination of archival business attributes, a missing flag is executed for subsequent manual verification or secondary system verification. After completing metadata verification, the module generates a globally unique archival record identifier for each archival record and records the collection time and preprocessing completion time, making the archival lifecycle traceable and auditable. The archival preprocessing module encapsulates standardized text content, corrected image content, structured business data, standardized archival metadata, format parsing objects, user identity information, user role information, user-submitted query requests, and cross-modal query requests according to a unified data structure, ultimately generating a standardized archival preprocessing dataset.

[0021] The semantic indexing construction module takes as input data the standardized preprocessed archive dataset output by the archive preprocessing module. This standardized preprocessed archive dataset includes at least normalized text content, corrected image content, photographic archive images, structured business data, standardized archive metadata, format parsing objects, and consistent user identity information, user role information, user-submitted query requests, and cross-modal query requests. During implementation, the semantic indexing construction module uses the standardized preprocessed archive dataset as its sole input source, extracting the text content, image content, and structured field information involved in semantic encoding. It also references fields related to archive business attributes from the standardized archive metadata, providing a unified data foundation for subsequent archive semantic encoding algorithms, archive feature standardization algorithms, and multi-modal indexing construction algorithms.

[0022] The semantic indexing construction module sequentially calls the archival semantic encoding algorithm, the archival feature standardization algorithm, and the multi-modal indexing construction algorithm to construct a unified semantic space and multi-type index structure from multi-source archival content. The improvement to the archival semantic encoding algorithm in the semantic indexing construction module introduces a multi-modal co-coding mechanism, an archival metadata-guided attention mechanism, consistency entity alignment constraints based on archival knowledge graphs, and a joint optimization mechanism of contrastive learning and noise suppression, based on traditional text feature extraction and image feature extraction. This substantially enhances the generated archival semantic vectors in terms of cross-modal consistency, sensitivity to archival business features, and robustness to noise and incomplete text. In the implementation steps of the archival semantic coding algorithm, the module first extracts standardized text content, corrected image content, photographic archival images, and structured business data from the standardized archival preprocessing dataset. The standardized text content is input into a deep semantic network to generate a text modal representation containing lexical structure, syntactic structure, discourse structure, and semantic role information. The corrected image content and photographic archival images are input into a visual coding network to generate a visual modal representation containing page layout, paragraph block distribution, title area, illustration area, and salient visual targets. The key fields in the structured business data are transformed into field vector representations through an embedding mapping mechanism, thereby obtaining three types of initial modal representations: text modal vectors, visual modal vectors, and structured field vectors. After obtaining the initial modal vectors, the archival semantic coding algorithm enters the multimodal collaborative coding stage. Through a cross-modal attention structure, the textual and visual modal vectors are jointly input. This structure calculates the relevance score between text fragments and image regions. Based on this relevance, the textual semantic units and image regions are weighted and fused, creating a soft alignment between paragraph semantics, keywords, and entity mentions in the text and page layouts, title blocks, and salient regions in the image. This avoids the separation of textual and image semantics in the representation space. After the multimodal collaborative coding completes the basic alignment, the archival semantic coding algorithm introduces an archival metadata-guided attention mechanism. Fields such as record number, file number, responsible party identifier, creation time, retention period, and security classification identifier are read from standardized archival metadata. These metadata fields are transformed into metadata-guided vectors through a field encoding mechanism, and these metadata-guided vectors are then input into the attention network along with the aforementioned fused multimodal representation. In this attention network, the archival semantic encoding algorithm adjusts the attention weights of each unit within the multimodal representation based on the matching degree between the metadata guiding vector and different semantic fragments and visual regions. This gives higher weights to semantic units closely related to archival cataloging, storage status, responsible parties, and confidentiality levels, while significantly suppressing the weights of background descriptions, general explanatory statements, and low-confidence OCR text fragments that are irrelevant to archival business in the final semantic representation. This makes the archival semantic vector more prominent in terms of semantic features that are of concern to archival management business.After completing the metadata-guided semantic enhancement, the archival semantic coding algorithm further utilizes the entity nodes and relational structures output by the archival knowledge graph module. Through entity alignment constraints, it maps entity mentions in the standardized text content to corresponding entity nodes in the archival knowledge graph. To achieve this mapping, the algorithm comprehensively considers the similarity between entity names and aliases, contextual semantic similarity, and fields such as the responsible party identifier, ownership unit identifier, and project number in the standardized archival metadata. It matches text entity mentions with entity nodes in the archival knowledge graph and embeds the semantics of successfully aligned entity nodes into the existing multimodal representation, ensuring that the semantic representation of the same person, institution, project, or event in different archival records tends to be consistent in the vector space. Through this entity alignment constraint, the archival semantic coding algorithm explicitly introduces the structural information and relational rules of the archival knowledge graph at the semantic coding stage, providing a consistent vector foundation for subsequent entity association retrieval and relational path retrieval, reducing the problem of semantic drift of the same entity in different archives. After fusing multimodal content, archival metadata, and archival knowledge graph entity features, the archival semantic coding algorithm enters the comparative learning stage, optimizing the spatial structure of the archival semantic vectors by constructing positive and negative sample pairs. Semantic representations of the same document in both text and image modalities are considered positive sample pairs, while semantic representations of different documents are considered negative sample pairs. Contrastive learning minimizes the distance between positive sample pairs and maximizes the distance between negative sample pairs, ensuring that the semantic vectors of the same document in different modalities are highly similar in space, while maintaining sufficient discriminative power between semantic vectors of different documents. Building upon this, the document semantic encoding algorithm also introduces a noise suppression constraint mechanism. Based on low-confidence OCR fragment markers, noisy field markers, and redundant content markers provided by the document preprocessing module, semantic units corresponding to noisy regions are assigned lower gradient update weights during contrastive learning, thereby suppressing the interference of noisy fragments on the semantic space structure and improving the stability of document semantic vectors in repeated retrieval and incremental updates. Through the aforementioned multimodal co-coding, document metadata-guided attention, document knowledge graph entity alignment constraints, and joint optimization of contrastive learning and noise suppression, the document semantic encoding algorithm ultimately generates a set of document semantic vectors with cross-modal consistency, sensitivity to document business features, and noise robustness. After the document semantic encoding algorithm generates the set of document semantic vectors, the semantic indexing construction module calls the document feature standardization algorithm to perform unified feature standardization processing on the set of document semantic vectors. The archival feature standardization algorithm pre-calculates the mean and standard deviation of each vector dimension based on a large-scale sample of archival semantic vectors during the training phase, forming vector distribution parameters. During the inference phase, it performs mean shifting and variance scaling on each archival semantic vector to ensure that all vectors have a consistent numerical scale across all dimensions.Subsequently, the archival feature standardization algorithm performs length normalization on each archival semantic vector, normalizing the vectors to a unified norm to ensure stable comparative significance when using inner product similarity or cosine similarity for subsequent retrieval. When extreme outliers are detected in certain dimensions of individual archival semantic vectors, the archival feature standardization algorithm limits the maximum deviation of these dimensions through boundary pruning strategies and compresses excessively extreme values ​​to a preset range according to numerical constraint rules, thereby avoiding the destructive impact of a very small number of abnormal samples on the overall semantic index structure. After processing by the archival feature standardization algorithm, the archival semantic vector set is significantly improved in terms of scale uniformity and distribution smoothness, making it more suitable for constructing an efficient nearest neighbor retrieval structure. After the archival feature standardization algorithm completes the processing of the archival semantic vector set, the semantic index construction module calls the multi-modal index construction algorithm to construct multiple types of indexes under a unified semantic space and a unified data structure. The multi-modal indexing algorithm first performs word segmentation and word position information annotation based on the normalized text content in the standardized archive preprocessing dataset. It then associates each term with the archive record identifier, field position, and word frequency statistics, forming an inverted index structure from term to the set of archive record identifiers. The algorithm stores statistical information such as word frequency, field weight, and position distribution in the inverted list, enabling the system to quickly locate target archive records when performing Boolean or weighted keyword searches. In the vector indexing process, the multi-modal indexing algorithm takes the archive semantic vector set processed by the archive feature standardization algorithm as input, dividing the semantic space into multiple subspaces or cluster structures. Within each subspace, a nearest neighbor search structure is established, allowing for the rapid retrieval of semantically similar archive records on a smaller candidate set when a query semantic vector is received. In the graph indexing process, the multi-modal indexing algorithm utilizes the entity set and relation set output by the archive knowledge graph module to associate entity nodes with archive record identifiers. It maintains the adjacency relationships between entity nodes and the mapping between entity nodes and archive records in the graph structure, enabling the system to quickly retrieve archive records participating in specific relations or located on specific relation paths based on entity identifiers or relation types. During the version time index construction process, the multi-modal index construction algorithm organizes the archive records into a time sequence structure that can be traversed in chronological order based on the archive record identifier and time attributes such as formation time, archiving time, and version time. This enables the system to retrieve the archive version evolution trajectory, archive status in different time periods, and time-related semantic changes within a specified time window.

[0023] Multimodal cooperative coding formula: , Archives Multimodal semantic vector representation. : Multimodal cooperative coding function, used to perform nonlinear transformations on concatenated multimodal vectors. : A vector formed by concatenating text modality vectors, visual modality vectors, and structured field vectors according to their dimensions. Text modality vectors obtained by encoding normalized text content using a deep semantic network. : Visual modality vectors obtained by encoding the corrected image content and photo archive images using a visual coding network. : A structured field vector obtained by encoding key fields of structured business data through an embedding mapping mechanism. The index of the current archive record, used to identify the specific archive. Metadata-guided attention formula: , : No. Attention weights for each multimodal semantic unit, guided by metadata. : A metadata guiding vector obtained by encoding standardized archival metadata such as archival group number, case file number, responsible person identifier, creation time, retention period, and security classification identifier. The attention mapping matrix between metadata and multimodal semantic units is used to calculate the matching score. : No. Each multimodal semantic unit vector originates from the internal representation after multimodal collaborative coding and contains fused features of text fragments and image regions. : Index of all multimodal semantic units in the current archive Perform a summation operation. Archival records obtained under the guidance of metadata-driven attention Weighted multimodal semantic vectors, : Multimodal semantic unit index, used to traverse semantic units in the current file. : Summation index, used to represent the multimodal semantic units traversed in the normalized denominator. The index of the current archive record. Knowledge graph entity alignment loss formula: , Knowledge graph entity alignment loss is used to constrain the consistency between text-based entity representations and archival knowledge graph entity representations. The set of successfully aligned entities, containing entities that have been identified in the normalized text content and matched with entity nodes in the archival knowledge graph. Entity set The number of entities, Entity index, used to iterate through each entity in the entity collection. Entities obtained by archival semantic encoding algorithms under the guidance of multimodal representation and metadata. semantic vector representation, : Entity nodes from the archival knowledge graph The embedding vector representation, The L2 norm of a vector, used to measure the Euclidean distance between two vectors. Noise suppression contrastive learning loss formula: , Contrastive learning loss is used to bring the semantic vectors of the same file closer together in different modalities and to separate the semantic vectors of different files. : A collection of archival records used in comparative learning :gather The number of records in the archives : Positive sample archive record index, corresponding to a pair of text modality and visual modality representations, : Negative sample archive record index, used to traverse the comparison candidate set of the current anchor archive. Regarding archival records The constructed set of candidate files for comparison includes Personal and other records, Archives The noise suppression weight is calculated from the low-confidence OCR markers, noise field markers, and redundant content markers provided by the document preprocessing module, and its value ranges from [value range missing]. and between, Text modality vector With visual modality vectors The similarity function value between them is generally the cosine similarity or the inner product similarity. Archives The text modality semantic vector is output by a deep semantic network or by the text branch of a multimodal coding network. Archives The visual modal semantic vectors are output by the visual coding network or by the visual branch in a multimodal coding network. Archives Visual modal semantic vectors, Temperature coefficient, used to adjust the smoothness of the similarity distribution in contrastive learning. : For sets Index of all candidate files Perform a summation operation. The joint optimization objective formula for archive semantic encoding is: , The joint optimization total loss of the archival semantic coding algorithm. Noise-suppressed contrastive learning loss is used to enhance cross-modal consistency and suppress the influence of noisy samples. Knowledge graph entity alignment loss is used to enhance the consistency between entity representations and the structure of the archival knowledge graph. : Balance coefficient, used to adjust the relative weights of contrastive learning loss and entity alignment loss in the overall optimization.

[0024] The output of the semantic indexing construction module includes a set of archival semantic vectors processed by archival semantic encoding and archival feature standardization algorithms, as well as inverted indexes, vector indexes, graph indexes, and version / time indexes generated by the multi-modal indexing construction algorithm. The set of archival semantic vectors, as a vectorized representation of archival content in a unified semantic space, provides basic semantic representation support for the knowledge graph module, semantic relationship retrieval module, and cross-modal retrieval module. The inverted index, vector index, graph index, and version / time index, as multiple index structures generated by the multi-modal indexing construction algorithm within a unified data structure, provide a unified index foundation for keyword retrieval, semantic similarity retrieval, entity relationship retrieval, and time evolution retrieval, ensuring that the system can perform efficient and stable retrieval processes within a consistent semantic space and index structure under various retrieval modes.

[0025] The input data for the knowledge graph module includes a standardized preprocessed archival dataset, a set of archival semantic vectors, and a graph index. The standardized preprocessed archival dataset provides machine-readable archival text, title structure, and paragraph semantics. The standardized archival metadata set provides archival management attributes such as record number, file number, responsible party identifier, ownership unit identifier, creation time, retention period, and security classification. The archival semantic vector set provides a cross-modal consistent semantic space representation to support entity merging and relation semantic strength assessment. The graph index supports index access and structural traversal between entity nodes, relation edges, and archival records. Based on this input data, the knowledge graph module, through the collaborative work of archival entity extraction algorithms, archival relation extraction algorithms, and archival knowledge graph construction and incremental update algorithms, transforms multi-source archival content into a structured, queryable, and evolvable archival knowledge graph.

[0026] The knowledge graph module sequentially executes the archival entity extraction algorithm, the archival relation extraction algorithm, and the archival knowledge graph construction and incremental update algorithm. While ensuring the accuracy of entity recognition, the completeness of relation extraction, and the stability of the graph structure, it organizes complex archival text and metadata content into a graph structure representation that supports various semantic retrieval and association analysis. The archival knowledge graph construction and incremental update algorithm of this invention improves entity merging strategies, relation evolution representation, cross-modal entity consistency, time-sensitive incremental updates, and relation consistency verification. During the implementation of the archival entity extraction algorithm, the knowledge graph module first constructs a text sequence representation for the standardized text content in the standardized archival preprocessing dataset. Sentences are divided according to natural sentence boundaries and paragraph boundaries, and auxiliary features such as part-of-speech tagging, lexical reconstruction, and syntactic dependency annotation are introduced at each marked position. The archival entity extraction algorithm inputs the aforementioned text sequence into a deep sequence labeling model. It utilizes a context encoding layer to capture semantic information across sentence spans, and then labels candidate entities such as names of persons, organizations, places, projects, document types, responsible persons, keywords, and time expressions through a conditional random field layer or an equivalent sequence discriminant layer, outputting a preliminary entity label sequence. Based on the preliminary labeling results, the algorithm combines paragraph-level topic information with semantic vectors within the context window to merge multiple mentions of the same entity and merge entity names that are split across sentences, ensuring consistency in entity recognition results at the document level. Simultaneously, the algorithm performs field parsing on the standardized archival metadata set, mapping fields such as responsible person identifier, ownership unit identifier, archival group number, file number, creation time, and retention period to an entity type space. These fields are considered high-confidence entities or entity attributes and included in the entity set. To ensure consistency between text entities and metadata entities, the archive entity extraction algorithm calls the text semantic vectors corresponding to the archive record from the archive semantic vector set, maps entity mentions in the semantic space, compares the contextual semantic vectors of the text entity mentions with the semantic representations formed by the metadata fields, and aligns text entities with metadata entities through semantic similarity threshold judgment and entity category constraint rules, eliminating duplicate and ambiguous entities and providing unified entity identifiers for subsequent graph construction. In the implementation of the archive relation extraction algorithm, the knowledge graph module takes the entity set output by the archive entity extraction algorithm and normalized text content as input. First, a syntactic dependency tree is generated for each sentence through a syntactic analyzer, and semantic roles such as predicate center, agent, patient, time adverbial, and place adverbial are identified through a semantic role labeling model. Based on this, the archive relation extraction algorithm identifies candidate entity pairs, calculates the co-occurrence patterns of each pair of entities within the same sentence, paragraph, and across paragraphs, and combines this with the semantic representations of the corresponding paragraphs in the archive semantic vector set to determine the semantic association strength between entities.In the relation type determination stage, the archival relation extraction algorithm, based on a predefined set of relation patterns, determines archival business-related relations such as formation relations, participation relations, referencing relations, custody relations, derivation relations, and inheritance relations. Entity pairs that conform to predicate constraints, dependency structure patterns, and semantic role configurations are assigned specific relation labels. For sentences with explicit time modifiers in the text, the algorithm binds time expressions to relation instances, writing time information such as formation time, change time, or archiving time as time attributes of relation edges into the relation instances. This ensures that each relation not only describes the static connection between entities but also relates to a specific point in time or time interval. When multiple sentences jointly express the same relation, the algorithm utilizes cross-sentence topic consistency analysis and semantic vector aggregation mechanisms to merge duplicate relations into a single relation instance and aggregate information from different contexts into the attributes of that relation instance, thereby avoiding redundant relation edges. After completing entity and relation extraction, the archival knowledge graph construction and incremental update algorithm begins to execute the construction and evolution management of the knowledge graph. In the entity merging phase, the archival knowledge graph construction and incremental update algorithm comprehensively utilizes the archival semantic vector set, the standardized archival metadata set, and entity category and contextual topic information to perform multi-dimensional judgment on mentions of entities with the same or similar names from different archival records. The algorithm first constructs a set of candidate entity pairs, calculates textual semantic similarity, structured metadata consistency score, contextual topic consistency score, and entity category consistency score for each candidate entity pair, and combines these scores according to preset weights to form a comprehensive entity merging score. When the comprehensive score exceeds a dynamic threshold, the algorithm merges the two entity mentions into a single entity node, merging their appearance positions and attribute information in different archives into the attribute set of the unified entity node, realizing a shift from string matching-driven merging to a merging driven by both semantics and structure. During the entity merging process, the algorithm also verifies the merging results based on the responsible party identifier and ownership unit identifier in the standardized archival metadata to prevent the erroneous merging of entities that actually belong to different responsible parties. In the relationship construction and evolution expression stage, the archival knowledge graph construction and incremental update algorithm, based on the relationship instances output by the archival relationship extraction algorithm, uses entity nodes in the entity set as source and target nodes of the relationship. It creates relationship edges for each relationship instance and writes the relationship type, relationship source, relationship confidence, and time attribute into the attribute fields of the relationship edges. Unlike the traditional static approach of retaining only one relationship edge for the same entity pair, this algorithm organizes multiple relationship instances formed by the same entity pair at different times and different business stages into a relationship evolution sequence based on attributes such as formation time, change time, archiving time, and archive type. This allows relationship edges to not only express the current state but also record the trajectory of the relationship's evolution over time and with business changes. The system can support version evolution queries and time-series correlation analysis through the relationship evolution sequence, enabling graph-based tracking of the archival history process.During the incremental update phase of the knowledge graph, the knowledge graph construction and incremental update algorithm constructs a local candidate graph for each batch of newly added archives, aligning the entity and relation instances extracted from the new archives with the entity nodes and relation edges in the existing knowledge graph. The algorithm first determines whether a new entity mention should be assigned to an existing entity node or a new entity node should be created, based on the aforementioned entity merging and comprehensive scoring mechanism. If it is determined to belong to an existing entity node, the algorithm updates the time trajectory of that entity node according to the time attribute in the new archive record and merges the new attribute information into the node attribute set. When processing new relation instances, the algorithm determines whether the corresponding subject entity, object entity, and relation type combination already exists in the knowledge graph. If it already exists but the time attribute or semantic attribute is different, the original relation will not be overwritten; instead, the new relation instance will be inserted into the existing relation evolution sequence, ensuring that the original historical relations are not destroyed, while accurately recording the relation changes brought about by the new archives. In terms of cross-modal fusion, the archival knowledge graph construction and incremental update algorithm utilizes visual semantic vectors from the archival semantic vector set to uniformly align and fuse entity candidates identified in image content and layout structure with entity mentions in text. This integrates multi-source features from both textual and visual modalities for the same entity node, thereby improving the consistency and interpretability of the knowledge graph in multimodal scenarios. To ensure that the knowledge graph maintains structural rationality and semantic consistency as the number of archives grows and versions evolve, the archival knowledge graph construction and incremental update algorithm introduces a relation consistency verification mechanism. When adding or updating relations, this mechanism compares relation descriptions of the same entity pair from different archival sources, checking for semantic conflicts, consistency between the relation subject and object, compliance with preset business constraints, and sufficient contextual support. When a clearly contradictory relation description is detected, the algorithm does not simply overwrite the existing relation. Instead, based on temporal attributes and business semantics, it adds it as a branch relation under different viewpoints or business scenarios to the relation evolution structure, enabling the system to retain complete historical information and interpret it through business rules or manual decision-making. Through the relation consistency verification mechanism, the knowledge graph construction and incremental update algorithm avoids the problem of relation loss and information breakage caused by simple overwriting during the incremental update process, thereby improving the stability and reliability of the knowledge graph in the long-term operation.

[0027] The knowledge graph construction and incremental update algorithm includes: entity semantic incremental aggregation formula: , :entity In the semantic vector after this incremental update :entity The semantic vectors that existed before the update, :entity The number of previously associated file records This incremental batch includes newly appearing entities. A collection of archival records, :gather The number of records in the archives Archives The metadata guides the multimodal semantic vector (consistent with the aforementioned archive semantic encoding algorithm). : Archive record index The entity index in the archival knowledge graph uses a simple weighted average method to represent the old semantics. Semantic information brought by the new archives The entities are fused together to obtain an incremental semantic representation. Incremental alignment formula for entity embedding: , :entity New embedding vectors in knowledge graphs :entity Old embedding vectors in knowledge graphs The entity obtained from the previous equation Incremental semantic vector, Entity embedding update step size coefficient, the value of which is within and between, The entity index in the archival knowledge graph uses an exponential sliding method to embed the graph side. With semantic encoding side Alignment ensures that the embedding of knowledge graph entities is continuously updated with incremental archives. Relation triple scoring formula: , Entity pairs In relation types The relationship score below, :entity Embedded vectors in knowledge graphs :entity Embedded vectors in knowledge graphs Relationship type The corresponding mapping matrix, Candidate subject entity index, Candidate object entity index, The relation type index calculates relation triples based on the current entity embedding and the relation matrix. The scoring provides a quantitative basis for determining whether to add, strengthen, or weaken the relationship. The formula for incremental relationship confidence updates is as follows: , Triplet The new confidence level, Triplet The old confidence level, The score for the triplet relation obtained from the previous equation. The relationship confidence update step size coefficient takes a value within... and between, The Sigmoid function maps real-valued scores to... arrive interval, Subject entity index, Object entity index, Relational type index, utilizing new scores For the original confidence level Perform smooth updates so that the relationship strength gradually converges with the support of multiple incremental archives; when When the edge value exceeds the construction threshold, new or strengthened relation edges are added; when the edge value falls below the deletion threshold, relation edges are decayed or removed. Knowledge graph joint optimization objective formula: Relationship structure loss for: , Total loss during archival knowledge graph construction and incremental updates. The previously defined entity alignment loss is used to constrain... and Consistency Relational structure loss is used to constrain the consistency between triple scores and labeled or pseudo-labeled tags. : Balance coefficient, used to adjust the weights of alignment loss and relational structure loss. The set of relation triples used during the training or update phase. :gather The number of triples, Triplet The strength of the normalized relation, Triplet The target label (existing as) There is no such thing as (or soft label) Entity-relationship triples in the knowledge graph. While maintaining alignment with the archival semantic encoding module, the entity alignment and relation structure are optimized to ensure that the archival knowledge graph conforms to the semantic space while maintaining the structural rationality and stability of the relation network during incremental updates.

[0028] The final output of the knowledge graph module is an archival knowledge graph. The archival knowledge graph consists of three core components: an entity set, a relation set, and a temporal attribute structure. The entity set records the merged archival entity nodes and their attribute information; the relation set records the various types of relationships between entity nodes and their evolution sequence; and the temporal attribute structure records the formation, change, and archiving process of entities and relationships over time.

[0029] The input data for the semantic relationship retrieval module includes user-submitted query requests, inverted indexes, vector indexes, a set of archival semantic vectors, and a graph index recorded in the standardized archival preprocessing dataset, as well as the archival knowledge graph output by the knowledge graph module. User-submitted query requests provide natural language or structured expressions of the retrieval semantics; inverted indexes are used to perform keyword matching; vector indexes are used to perform nearest neighbor searches at the semantic space level; the set of archival semantic vectors supports semantic expression, modality discrimination, and vector reliability calculation; and the graph index and archival knowledge graph support entity localization, relationship judgment, and path reasoning. Integrating these inputs within a unified framework allows users' retrieval intent to be supported by semantic evidence, keyword evidence, and relationship evidence, thereby generating search results consistent with archival business logic.

[0030] The semantic relationship retrieval module sequentially executes the query semantic parsing and rewriting algorithm, the joint semantic retrieval ranking algorithm, and the relationship path retrieval algorithm, constructing a complete retrieval chain from query parsing, candidate retrieval, semantic ranking to relationship reasoning. Improvements to the joint semantic retrieval ranking algorithm include: firstly, introducing a modality preference modeling mechanism in cross-modal query understanding, enabling different modal features to obtain differentiated weights based on query modality preferences; secondly, introducing a semantic vector reliability reweighting mechanism, assigning dynamic credibility factors to semantic similarity by evaluating the noise suppression state, standardization quality, and semantic clustering density of semantic vectors; thirdly, implementing a relationship consistency-guided ranking mechanism, introducing knowledge graph structure constraints on top of semantic ranking, allowing business logic relationships and time chains to play a role in the ranking process; and fourthly, introducing a multi-evidence fusion model, achieving final ranking optimization through the fusion of keyword evidence, semantic vector evidence, vector reliability evidence, and relationship consistency evidence. In the implementation of the query semantic parsing and rewriting algorithm, the module first utilizes a deep semantic parsing model to perform lexical analysis, syntactic analysis, and semantic structure decomposition on the user-submitted query request. This identifies object entities, conditional constraints, relational constraints, and time constraints within the query, and constructs a query intent representation based on the structural clues within the query semantics. During the parsing process, the algorithm further invokes a cross-modal semantic alignment mechanism to compare the query content with the set of document semantic vectors. Based on the semantic features of the query text, it determines whether the query is closer to a text modality, a visual modality, or a structured field modality, and records the identified modal tendency as an attribute of the query plan to guide subsequent index selection strategies. After forming the query intent, the algorithm rewrites the natural language query expression into a structured query plan within the system based on the identification results of entity categories, relation types, and semantic roles. This ensures that the query has a clear correspondence in the inverted index, vector index, and graph index, thus laying a unified foundation for subsequent semantic retrieval and relational retrieval. After the query plan is generated, the module invokes a joint semantic retrieval ranking algorithm. In the implementation of the joint semantic retrieval ranking algorithm, the module first extracts a set of candidate files based on the query plan using the inverted index, and then performs a nearest neighbor search in the set of file semantic vectors based on the query semantic vector using the vector index to form a set of semantic candidate files. Subsequently, the two sets of candidate files are merged and the semantic vector reliability reweighting mechanism is invoked. This mechanism calculates a reliability factor for each file's semantic vector by evaluating noise suppression, vector length normalization quality, and cluster density. This reliability factor is then used as a semantic similarity adjustment parameter, strengthening high-quality semantic vectors in the ranking process, while automatically reducing the weight of low-quality semantic vectors affected by OCR errors, image noise, or incomplete text.After the semantic fusion phase, the joint semantic retrieval ranking algorithm accesses the archive knowledge graph output by the knowledge graph module. Based on the query entities, relational constraints, and temporal limitations identified by the query semantic parsing and rewriting algorithms, it locates the corresponding entity nodes from the graph index and calculates the consistency score of candidate archives in the entity relation space along the adjacency relation structure. This consistency score is comprehensively calculated based on the hop distance between entities, relation type similarity, temporal attribute chain of relation edges, and contextual semantic consistency. Archives with direct relationships to the query entity, multi-hop interpretable paths, or matching temporal conditions are ranked higher, while archives with weak relationships or inconsistent semantic logic are downgraded. Finally, the joint semantic retrieval ranking algorithm integrates keyword evidence, semantic similarity evidence, semantic vector reliability evidence, and relational consistency evidence into a unified ranking score through a multi-evidence fusion model, ensuring consistency and interpretability of the ranking results at the semantic, knowledge, and business logic levels. After the joint semantic retrieval ranking algorithm generates the ranking results, the module calls the relational path retrieval algorithm to further expand potentially overlooked business-related archives. The relational path retrieval algorithm, based on the query entity's position in the archival knowledge graph, starts from the entity node and traverses paths according to entity adjacency edges, relation types, and temporal attribute chains. It judges the validity of candidate paths through path semantic constraints, relation consistency rules, and temporal chain constraints, and identifies archival nodes with multi-hop associations, indirect associations, or historical associations, incorporating these nodes into an expanded candidate set. By performing a deep traversal of the knowledge graph structure, the algorithm enables the system to capture archival records that are difficult to identify in the semantic vector space but are strongly associated in the relational structure, thus significantly improving retrieval recall. Finally, the algorithm deduplicatively merges the expanded candidate set and the ranked candidate set, ensuring that the final candidate set possesses both semantic similarity and relational structural relevance.

[0031] Query the multimodal semantic representation formula: , : query Multimodal semantic vectors Multimodal cooperative coding function, same as the archive side. : The concatenation of query text vector, query structured feature vector, and query metadata vector. : The semantic vector of the query text content : Vector representation of the query's structured conditions or filtering conditions. : Query side metadata preference vector, The index of the current search query. Semantic similarity scoring formula: , : query With archives semantic similarity score, : query Multimodal semantic vectors Archives Metadata guides multimodal semantic vectors. Vector similarity functions, using cosine similarity or inner product similarity. : Query index, : Archive record index. Knowledge graph relevance scoring formula: , : query With archives The knowledge graph relevance score. : query The set of entities identified in the process Archives A collection of related entities. :gather The number of entities, :gather The number of entities, :entity Knowledge graph embedding vectors, :entity Knowledge graph embedding vectors, Similarity function between entity embeddings Query-side entity index, : Archive-side entity index, : Query index, Archive record index. Metadata matching scoring formula: , : query With archives Metadata matching score, : query Metadata preference vector, Archives The metadata vector is obtained by encoding the archival group number, case file number, responsible party identifier, creation time, retention period, and security classification identifier. Metadata matching mapping matrix : Query index, : Archive record index. Joint semantic retrieval ranking score formula: , : query Records The final sorting score, Semantic similarity score Knowledge graph relevance score Metadata matching score Semantic similarity weight coefficient Knowledge graph relevance weight coefficient Metadata matching weight coefficient : Query index, : Archive record index. Joint ranking learning loss formula: , Learning loss for joint semantic retrieval ranking : By query Positive sample archives Negative sample archives The set of training triples, :gather The number of triples in the middle : query Positive sample archives The sorting score, : query negative sample archives The sorting score, The sigmoid function is used to map score differences to probabilities. : Query index, : Index of archive records that have been clicked or marked as relevant. Unclicked or marked as irrelevant archive records index.

[0032] After the query semantic parsing and rewriting algorithm, the joint semantic retrieval ranking algorithm, and the relation path retrieval algorithm are executed sequentially, the semantic relation retrieval module finally outputs a set of candidate results for semantic and relation fusion retrieval. This set of candidate results for semantic and relation fusion retrieval maintains a strict correspondence at the record granularity with the set of archive semantic vectors output by the semantic indexing construction module and the set of entities and relations output by the knowledge graph module.

[0033] The input data for the cross-modal retrieval module includes user-submitted cross-modal query requests, a set of archival semantic vectors, vector and graph indexes, and an archival knowledge graph, all recorded in the standardized archival preprocessing dataset. User-submitted cross-modal query requests express cross-modal retrieval intent and can be presented in image or text form. The archival semantic vector set and vector index provide vector representation and nearest neighbor search capabilities within the cross-modal semantic space. The graph index and archival knowledge graph are used to perform entity association verification and relational structure validation after vector retrieval. These input data collectively form the operational foundation of cross-modal retrieval, enabling computation of different modal query content within a unified semantic space and ensuring semantic and knowledge structure consistency in the retrieval results.

[0034] The cross-modal retrieval module sequentially executes the cross-modal representation alignment algorithm and the cross-modal similarity calculation algorithm. It achieves two types of cross-modal retrieval tasks: text-to-image search and image-to-text search, through semantic mapping of the query content and cross-modal vector retrieval. During the implementation of the cross-modal representation alignment algorithm, the module first performs modality judgment on the cross-modal query request submitted by the user, determining whether the input belongs to the text modality or the image modality by parsing the query content. When the query request is in text form, the cross-modal representation alignment algorithm uses a deep coding network to encode the query text, generating a text query vector containing semantic structure, semantic entities, and semantic topics. This vector is then mapped to the cross-modal semantic space through a cross-modal mapping network, enabling it to be matched with image vectors in the archive semantic vector set under the same metric system. When the query request is in image form, the cross-modal representation alignment algorithm uses a visual coding network to process the query image, extracting layout structure, salient regions, and semantic object features to generate an image query vector. This vector is then mapped to the cross-modal semantic space through a cross-modal mapping network, enabling it to be compared for consistency with text vectors in the archive semantic vector set. The cross-modal representation alignment algorithm, during the above mapping process, relies on the semantic space structure constructed by the semantic index construction module to ensure that query vectors and file semantic vectors of different modalities have the same norm constraints, distance measurement methods, and semantic expression structures in the cross-modal semantic space. This eliminates the expression differences between different modalities and ensures that subsequent cross-modal retrieval can be carried out in a consistent semantic space.

[0035] After the cross-modal semantic vector is generated, the module calls the cross-modal similarity calculation algorithm to perform cross-modal retrieval in the vector index. The cross-modal similarity calculation algorithm uses the cross-modal semantic vector as the query vector and performs a nearest neighbor search in the vector index. It identifies the document semantic vector most similar to the query content through distance or similarity metrics in the cross-modal semantic space, and forms a cross-modal candidate document set accordingly. In the text-to-image search task, the cross-modal similarity calculation algorithm retrieves the image vector closest to the query text vector from the vector index and returns the corresponding document record; in the image-to-text search task, the cross-modal similarity calculation algorithm retrieves the text vector closest to the query image vector and returns the corresponding document record. To improve the semantic reliability and entity consistency of cross-modal retrieval, the cross-modal similarity calculation algorithm, after completing vector retrieval, further utilizes graph indexes and archival knowledge graphs to perform entity-level matching verification. By analyzing the adjacency relationships, relationship types, and contextual semantic structures of the entity nodes corresponding to the candidate archives and query vectors in the knowledge graph, candidate results lacking sufficient entity relevance or semantic consistency are eliminated, ensuring that the cross-modal retrieval results are consistent with the query intent in terms of semantic expression and knowledge structure. Finally, the cross-modal similarity calculation algorithm ranks the candidate archive set based on similarity scores in the cross-modal semantic space or a comprehensive score formed by combining knowledge graph association scores, providing a ranking basis for generating the cross-modal retrieval result set.

[0036] The cross-modal retrieval module ultimately outputs a set of cross-modal retrieval results. This set consists of a collection of files that satisfy cross-modal semantic consistency, vector space similarity, and knowledge graph relevance. It maintains the same structural organization as the semantic and relational fusion retrieval candidate result set, enabling the subsequent access control module to use a unified format to perform access permission determination, content desensitization, and risk ranking.

[0037] The input data for the access control module includes user identity information, user role information, a standardized set of archival metadata, a set of archival semantic vectors, an archival knowledge graph, a set of candidate results for semantic and relational fusion retrieval, a set of cross-modal retrieval results, and pre-configured business rules and security policies. User identity information and user role information together constitute the basic input describing the user's subject characteristics and authorization boundaries. The standardized archival metadata set provides management attributes for each candidate archival record, such as security classification, retention period, access conditions, affiliated institution, responsible party identifier, creation time, archival group number, and case file number. The archival semantic vector set provides the basis for field importance assessment and semantic retention assessment. The archival knowledge graph is used to introduce entity relationship structure and path sensitivity information in access control and risk assessment. The set of candidate results for semantic and relational fusion retrieval and the set of cross-modal retrieval results provide the set of candidate archival records for which access control needs to be performed. The business rules and security policies configuration data are used to describe access level rules, field sensitivity levels, de-identification strategy templates, and risk assessment rules. Based on this input data, the access control module performs access control, content de-identification, and risk ranking processing on the retrieval results through the collaborative action of multiple algorithms.

[0038] The access control module sequentially invokes the access rule matching algorithm, the search result filtering and desensitization algorithm, and the risk assessment algorithm. It performs refined access control on candidate files through three stages: access permission determination, content format adjustment, and risk quantification ranking. The access rule matching algorithm has undergone significant improvements in user access attribute vector modeling, file security attribute vector modeling, relationship constraint determination combined with file knowledge graphs, and the introduction of a session context sensitivity mechanism. This allows access decisions to not only rely on static security level rules but also to comprehensively consider business semantic relevance and relationship path sensitivity. In the implementation steps of the access rule matching algorithm, the module first receives user identity information and user role information. It encodes attributes such as the user's organization, job type, job level, business line, historical authorization scope, and whether they bear specific file management responsibilities into user access attribute vectors. In this vector space, each dimension corresponds to a computable access attribute feature, making user permissions no longer just a set of discrete labels but a structured representation that can be used for similarity and matching operations. Subsequently, the permission rule matching algorithm reads the standardized archival metadata set and parses fields such as the security classification, retention period, access conditions, affiliated institution, responsible party, creation time, archival group number, and case file number for each archival record. These fields are then encoded into archival security attribute vectors according to the dimensions of security classification, time, affiliated institution, and responsible party, giving the archives a unified vector representation in terms of security control. Based on these two vector spaces, the permission rule matching algorithm embeds the access levels and rule constraints configured in the business rules and security policy configuration data into the matching logic, pre-forming an authorization mapping rule model between user access attribute vectors and archival security attribute vectors. After modeling the user access attribute vectors and archival security attribute vectors, the permission rule matching algorithm constructs a set of candidate archives requiring permission determination in the current session based on the archival record identifiers in the semantic and relational fusion retrieval candidate result set and the cross-modal retrieval result set. Each candidate archive is then associated with its corresponding archival security attribute vector. To further enhance the business rationality of access decisions, the permission rule matching algorithm invokes the archival knowledge graph. It retrieves the relationship paths between the entities corresponding to candidate archival records, the entities corresponding to the user's affiliated institution, and the entities corresponding to the user's business objects. Taking into account the length of the relationship path, the type of relationship, the direction of the relationship, and the time attribute, the algorithm determines the rationality and necessity of the current access behavior in terms of archival business semantics. For example, when the entity corresponding to the candidate archive has a direct custody or responsibility relationship with the entity of the user's affiliated institution, the algorithm can increase the confidence of access authorization; when the relationship path is too long or only connected through weak associations, the algorithm imposes stricter restrictions on the access results.The permission rule matching algorithm, through the combined action of user access attribute vectors, archive security attribute vectors, and entity relationship structures in the archive knowledge graph, calculates access decisions for each candidate archive record, classifying them into three access permission categories: allowed direct browsing, allowed anonymized browsing, or prohibited browsing. This access permission category is then bound to the candidate archive record to form an intermediate set of candidate results marked with access permission tags. After the permission rule matching algorithm determines the access permission category, the permission control module invokes a search result filtering and anonymization algorithm to filter and anonymize the intermediate set of candidate results marked with access permission tags. The search result filtering and anonymization algorithm first reads the access permission category of each candidate archive record. For records marked as prohibited browsing, the algorithm directly removes them from the candidate set, ensuring they are not displayed or output in subsequent stages. For records marked as allowed direct browsing, the algorithm maintains the field content in their standardized archive metadata set and the content fields returned by the search module unchanged, without any masking or obfuscation. For records marked as allowing anonymized browsing, the search result filtering and anonymization algorithm identifies highly sensitive fields and partially sensitive fields requiring protection based on the security classification, access conditions, and field sensitivity rules defined in the business rules and security policy configuration data within the standardized archive metadata set. Then, according to preset field masking, content obfuscation, and entity anonymization strategies, it performs full character masking or retains only the field structure for highly sensitive fields, and performs partial character replacement or semantic generalization for partially sensitive fields. When performing field anonymization, the search result filtering and anonymization algorithm references the archive semantic vector set output by the semantic indexing module. By analyzing the contribution of each field's content to the archive semantic vector, it retains fields that have a greater impact on the overall semantic expression, thus maintaining the relevance between the search results and the user's query semantics while meeting the anonymization requirements. When candidate records originate from a cross-modal retrieval result set and contain image content or scanned images, the retrieval result filtering and desensitization algorithm invokes the labeling information for sensitive entities in the archive knowledge graph. Image regions highly associated with these sensitive entities are marked as sensitive areas, and these areas undergo local blurring, occlusion, or replacement with low-resolution thumbnails, thus achieving a desensitization effect consistent with the text modality in the image modality. After processing by the retrieval result filtering and desensitization algorithm, the module generates an intermediate retrieval result set, containing only records that are allowed to be directly viewed and records that have undergone desensitization, with each record possessing a secure content format suitable for direct display. After the retrieval result filtering and desensitization algorithm outputs the intermediate retrieval result set, the access control module invokes a risk assessment algorithm to quantify and sort the risk of each archive record in the intermediate retrieval result set.The risk assessment algorithm is based on fields such as security classification, retention period, access conditions, responsible party, creation time, archival group number, and file number in the standardized archival metadata set. It combines these with the degree of the corresponding entity nodes in the archival knowledge graph, the shortest path length between them and highly sensitive entities, and the distribution of relationship types to assess the inherent sensitivity of the archives within the structured semantic space. Simultaneously, the algorithm constructs a user behavior context based on user identity and role information. It incorporates the distribution of entities retrieved multiple times during the current session, the frequency of continuous access to specific sensitive topics, and the overlap of entity and relationship types with the semantic and relational fusion search candidate result set and the cross-modal search result set. This analysis is used to identify potential unauthorized access behaviors or risk patterns of concentrated access to sensitive archives. The risk assessment algorithm generates a risk score for each record after comprehensively considering the sensitivity attributes of the archive, the sensitivity of the knowledge graph structure, and the sensitivity of user access behavior. This risk score, along with the access permission category output by the permission rule matching algorithm and the content format formed by the search result filtering and de-identification algorithm, are used as the sorting criteria. The order of records in the intermediate search result set is adjusted so that records with higher risk scores are placed further down the list. If necessary, additional user confirmation or audit flags can be triggered. Furthermore, the risk assessment algorithm can register high-risk access events as audit logs, providing a basis for subsequent security operations and compliance reviews. User permission profile coding formula: , :user The permission profile vector is used to characterize identity, role, and permission preferences. The permission profile encoding function takes a concatenated vector as input and outputs a low-dimensional vector representation. The concatenation of user identifier vector, role vector, organization vector, and user metadata vector. User identifier encoding vector User role encoding vector : The encoding vector of the user's organization or department User-side metadata vectors, used to represent access levels, service scope, etc. : Index of the currently accessing user. Metadata permission matching score formula: , :user With archives Metadata permission matching score, :user Permission profile vector, : A mapping matrix between user permission profiles and file metadata Archives The metadata vector is encoded by the archival group number, case file number, responsible party identifier, creation time, retention period, and security classification identifier. User index, Archive record index. Permission rule matching score formula: , : No. The permission rules apply to "users" Access Archives The degree of matching of "". : The aggregated rule matching score : No. Vector representation of each permission rule, encoding rule conditions and action information. : A vector formed by concatenating the user permission profile vector and the file metadata vector according to their dimensions. The Sigmoid function maps real numbers to... arrive interval, The set of permission rules that participate in the matching. : No. The weight coefficient of each permission rule is used to distinguish the importance of the rules. User index, : Archive record index Access control rule index. Knowledge graph sensitivity matching formula: First, define the file sensitivity score: Then define the knowledge graph permission matching score: , Archives The overall sensitivity score, Archives A collection of related entities. :gather Number of entities in China :entity The sensitivity coefficient is given by the entity attributes in the archival knowledge graph. :user With archives Permission matching score on the knowledge graph dimension :user The access control scalar represents the clarity or security level of permissions; a higher value indicates a stronger ability to access sensitive files. User index, : Archive record index Knowledge graph entity index. Joint permission score and access determination formula: Joint permission score: Access determination: , :user Records The combined authority score, ranging from arrive , Metadata permission matching score Permission rule matching score Knowledge graph permission matching score Metadata matching score weighting coefficient Rule matching score weighting coefficient Knowledge graph matching score weighting coefficient The sigmoid function is used to map linear combinations to probabilities. Access determination result This indicates that access is permitted. Access is denied. Indicator function, returns when the condition is true. Returns if condition is false , Access threshold: Controls the minimum federated permission score required to grant access. User index, : Index of archival records.

[0039] After the permission rule matching algorithm, search result filtering and desensitization algorithm, and risk assessment algorithm are executed sequentially, the permission control module finally outputs a final search result list after access permission category determination, content desensitization processing, and risk ranking processing. The final search result list consists of the semantic and relational fusion search candidate result set and the archive records retained after permission control processing in the cross-modal search result set. It maintains a one-to-one correspondence with the candidate results of the aforementioned search modules in terms of record granularity, and meets the system's preset business and security requirements in the three dimensions of access permission, content security, and risk control.

[0040] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based archival information integration system, characterized in that: Includes the following modules: The document preprocessing module receives digitized scanned images of paper documents, layout structure information of paper documents, text data of electronic documents, image data of photographic documents, audio and video documents and their transcribed text, structured business data from external business systems, original metadata fields of documents, user identity information, user role information, user-submitted query requests, and cross-modal query requests. It performs preprocessing through document format parsing, character set and text normalization, image denoising and layout correction, and document metadata normalization and integrity verification to generate a standardized document preprocessing dataset. The semantic index building module receives a standardized archive preprocessing dataset and generates archive semantic vector sets, inverted indexes, vector indexes, graph indexes, and version time indexes through archive semantic encoding, archive feature standardization, and multi-modal index construction. The knowledge graph module receives a standardized preprocessed dataset of archives, a set of semantic vectors of archives, and a graph index. It generates an archive knowledge graph through archive entity extraction, archive relationship extraction, and archive knowledge graph construction and incremental updates. The semantic relation retrieval module receives user-submitted query requests, inverted indexes, vector indexes, archive semantic vector sets, graph indexes, and archive knowledge graphs. Through query semantic parsing and rewriting, joint semantic retrieval sorting, and relation path retrieval, it generates a set of candidate results for semantic and relational fusion retrieval. The cross-modal retrieval module receives cross-modal query requests submitted by users, a set of archival semantic vectors, a vector index, a graph index, and an archival knowledge graph. It generates a set of cross-modal retrieval results through cross-modal representation alignment and cross-modal similarity calculation. The access control module receives user identity information, user role information, standardized archive metadata set, semantic and relational fusion search candidate result set, and cross-modal search result set. Through access rule matching, search result filtering and desensitization, and risk assessment, it generates the final search result list.

2. The deep learning-based archival information integration system according to claim 1, characterized in that: The archive preprocessing module performs unified parsing on digitized scanned images of paper archives, page layout information of paper archives, text data of electronic archives, image data of photo archives, audio and video archives and their transcribed text, structured business data of external business systems, and original metadata fields of archives, generating image parsing objects, text parsing objects, structured field objects, and format parsing objects; The parsed text content is processed to unify the character set and standardize the text, thereby unifying character encoding, standardizing number and date formats, normalizing character morphology, and removing abnormal characters to form standardized text content. The obtained image content is subjected to denoising and layout correction processing to complete noise suppression, tilt correction and layout enhancement, and the corrected image content is obtained through binarization processing. Perform standardization and integrity checks on the original metadata fields of the archives and the structured business data of the external business system, generate a standardized set of archive metadata that includes the archival group number, case number, responsible person identifier, creation time, retention period and security classification identifier, and generate a unique archive record identifier for each archive. The standardized text content, corrected image content, structured business data, standardized archive metadata, format parsing objects, user identity information, user role information, user-submitted query requests, and user-submitted cross-modal query requests are uniformly organized and encapsulated to output a standardized archive preprocessing dataset.

3. The archival information integration system based on deep learning according to claim 1, characterized in that: The semantic index building module receives a standardized archive preprocessing dataset and generates multimodal representations of text content, image content, image content, and structured business data. This forms text modality representation, visual modality representation, and field representation under a unified semantic space. Under the combined effect of multimodal co-coding, archive metadata guidance constraints, entity alignment constraints, and noise suppression processing, it forms a set of archive semantic vectors with cross-modal consistency and business semantic expression capabilities. The semantic vector set of archives is subjected to feature standardization processing to form a unified and standardized vector expression in terms of scale and distribution; An inverted index is built based on the normalized text content, and the mapping relationship between terms, position distribution and field affiliation is maintained in the inverted index; A vector index is constructed based on the set of archival semantic vectors to support nearest neighbor retrieval of archival semantic vectors; A graph index is constructed based on entity sets and relation sets. By maintaining the association between entity nodes, relation edges and record identifiers, the semantic association at the entity level remains searchable. A version time index is constructed based on the archival record identifier and its creation time, archiving time, and version time, so that the version evolution of the archives in the time dimension has a traversable structure; Output archive semantic vector set, inverted index, vector index, graph index and version time index.

4. The deep learning-based archival information integration system according to claim 1, characterized in that: The knowledge graph module receives a standardized archive preprocessing dataset, an archive semantic vector set, and a graph index. It uses the standardized text content and the standardized archive metadata set as input for entity and relation recognition, the archive semantic vector set as input for cross-modal consistency judgment, and the graph index as the index basis for entity nodes and relation edges. Entity recognition processing is performed on the standardized text content to identify personal names, organization names, place names, project names, document names, responsible persons' names, subject terms and time expressions. It also generates an entity set by combining the archival metadata set with the archival group number, file number, responsible person identifier, ownership unit identifier and formation time. Based on the archival semantic vector set, the text entities and metadata entities are aligned for consistency, and entities from different sources are uniformly merged into the entity set. Based on the semantic structure of the text and the contextual relationships, the relationships between entities are generated, including formation relationships, participation relationships, reference relationships, custody relationships, derivation relationships, and inheritance relationships. The relationship set is then generated by combining the formation time, change time, or archiving time. The entity set and relation set are merged, relation built and time attribute organized by constructing an archival knowledge graph and using incremental update algorithms. In the entity merging stage, entities are unified according to text semantic consistency, metadata consistency, context theme consistency and entity category consistency. In the relation building stage, relation edges are constructed according to relation subject entity, relation object entity, relation type and time attribute, and relation instances are organized into relation evolution structure in chronological order. For newly added archives, a partial candidate graph is constructed, and the time chain of existing entity nodes is updated based on entity consistency. At the same time, the relationship subject, relationship object and relationship type are used to determine whether to generate new relationship instances or insert relationship instances into the existing relationship list. The visual semantic information in the archive semantic vector set is combined to perform cross-modal consistency verification on the entity candidates and text entities in the image content. Conflicting relationships are recorded as relationship evolution branches, and the archive knowledge graph is output.

5. The archival information integration system based on deep learning according to claim 1, characterized in that: The semantic relationship retrieval module receives user-submitted query requests, inverted indexes, vector indexes, archive semantic vector sets, graph indexes, and archive knowledge graphs. It performs semantic parsing and structured processing on user-submitted query requests, identifies target entities, attribute constraints, relational constraints, and time constraints in the query, and transforms the parsing results into an executable query plan, mapping keyword conditions, semantic similarity conditions, and entity and relational constraints to the inverted index, vector index, graph index, and archive knowledge graph, respectively. Based on the query plan, a set of candidate files for text matching is generated through inverted index, and a set of candidate files for semantic similarity is generated through vector index. The text features and semantic features are then fused together. At the same time, the candidate files are weighted according to the quality label of the file semantic vector set to form a fused and ranked candidate file result. By combining entity adjacency relationships, relationship types, relationship path structures, and relationship time attributes in the archival knowledge graph, the consistency of the relevance between candidate archives and query intent at the entity and relationship levels is determined, and multi-evidence fusion ranking is performed on the candidate archives accordingly. After obtaining the ranking results, a relational path traversal is performed based on the location of the query entity in the archival knowledge graph to identify archival nodes that have a valid association or time chain association with the query entity. The association results are then deduplicated and integrated with the fused ranking results to output a set of semantic and relational fusion retrieval candidate results.

6. The deep learning-based archival information integration system according to claim 1, characterized in that: The cross-modal retrieval module receives cross-modal query requests, a set of archival semantic vectors, a vector index, and an archival knowledge graph submitted by the user. It parses the cross-modal query requests, determines the modality type of the query content, and maps the query content to query vectors that are in the same representation space as the set of archival semantic vectors. Based on the query vector, nearest neighbor retrieval is performed in the vector index to identify the semantic vectors of archives that are close to the query vector, forming a cross-modal candidate archive set, and supporting retrieval methods of querying image archives by text and querying text archives by image. By combining the knowledge graph of archives, consistency verification is performed on the entities corresponding to cross-modal candidate archives at the levels of entity adjacency, relationship type and semantic association structure, and candidate results with insufficient association or incomplete supporting information are screened out. Based on the vector retrieval results and entity association verification results, the cross-modal candidate file set is sorted, and the cross-modal retrieval result set is output.

7. The archival information integration system based on deep learning according to claim 1, characterized in that: The access control module receives user identity information, user role information, standardized archive metadata set, semantic and relational fusion retrieval candidate result set, cross-modal retrieval result set, and business rules and security policy configuration data. It parses user identity information and user role information into user access attributes, and parses the security level identifier, retention period, access conditions, ownership institution, responsible person identifier, and formation time in the standardized archive metadata set into archive security attributes. The document records in the semantic and relational fusion retrieval candidate result set and the cross-modal retrieval result set are associated with the corresponding document security attributes. Based on the business rules and security policy configuration data, the user access attributes and document security attributes are matched to determine the access permission category of each document record. After completing the access permission determination, the file records that are determined to be prohibited from browsing are filtered, and the file records that are determined to be allowed to be de-identified are processed by field masking, content blurring or entity anonymization according to the field sensitivity and de-identification rules in the standardized file metadata set. The sensitive areas involving image content are partially blurred, masked or thumbnail replaced to form an intermediate search result set. After obtaining the intermediate search result set, based on the security classification, retention period, access conditions, responsible person identification and formation time in the standardized archival metadata set, and combined with the entity and relationship association structure in the archival knowledge graph, as well as the access behavior characteristics reflected by user identity information and user role information, the archival records in the intermediate search result set are risk-assessed and their order is adjusted, and the final search result list is output.

8. The deep learning-based archival information integration system according to claim 7, characterized in that: Business rules and security policy configuration data include access level rules, business domain authorization rules, field sensitivity level rules, security level correspondence rules, opening condition rules, retention period access restriction rules, de-identification policy rules, field masking rules, content blurring rules, entity anonymization rules, image sensitive area processing rules, and risk assessment rules; Access level rules define the range of access levels that can correspond to user identity information and user role information; Business area authorization rules define the authorization boundaries for different positions, organizations, or business lines in the scenario of archive utilization; The field sensitivity level rules classify fields in the standardized archive metadata set according to their sensitivity. The classification rule maps the classification identifier of a file to its access control level; Open condition rules define the conditions and restrictions under which files can be displayed; The rules on access restrictions based on retention period define the access visibility of files under different retention periods; The desensitization policy rules specify the desensitization methods to be used when desensitized browsing is permitted; Field masking rules, content obfuscation rules, and entity anonymization rules describe the desensitization methods at the field and content levels; The image sensitive area processing rules define the methods for performing local blurring, occlusion, or thumbnail replacement on image content containing sensitive areas; the risk assessment rules map the archive's security classification, access conditions, responsible party identification, creation time, archival group number, file number, and the relationship path information of the archive in the archive knowledge graph as the basis for risk scoring.

9. A method for integrating archival information based on deep learning as described in any one of claims 1-8, characterized in that: Includes the following steps: The system receives digitized scanned images of paper archives, layout structure information of paper archives, text data of electronic archives, image data of photo archives, audio and video archives and their transcribed text, structured business data from external business systems, original metadata fields of archives, user identity information, user role information, user-submitted query requests, and user-submitted cross-modal query requests. It performs unified parsing, text normalization, image denoising and layout correction, and archive metadata normalization and integrity verification on the above data to generate a standardized archive preprocessing dataset. The system receives the standardized archive preprocessing dataset, generates semantic representations of the standardized text content, image content, image content, and structured business data, performs feature standardization processing on the generated semantic representations, constructs an archive semantic vector set, and simultaneously constructs an inverted index, vector index, graph index, and version time index based on the data. The system receives the standardized archive preprocessing dataset, archive semantic vector set, and graph index, identifies entities and relationships in the archives, and generates an archive knowledge graph through archive knowledge graph construction and incremental update processing. It receives query requests submitted by users and combines inverted indexes, vector indexes, archive semantic vector sets, graph indexes, and archive knowledge graphs to parse and retrieve the query requests, generating a set of semantic and relational fusion retrieval candidate results. It receives cross-modal query requests submitted by users, and processes the cross-modal query requests by combining the archive semantic vector set, vector index and archive knowledge graph to generate a cross-modal retrieval result set; It receives user identity information, user role information, standardized archive metadata set, semantic and relational fusion search candidate result set, and cross-modal search result set. Combined with business rules and security policy configuration data, it performs access permission determination, result filtering and de-identification processing, and result order adjustment on the search results, and outputs the final search result list.

Images