Archive classification and identification method and system based on artificial intelligence

Abstract

The invention relates to the technical field of archive classification, and particularly discloses an archive classification and identification method and system based on artificial intelligence. The method comprises the following steps: firstly, obtaining document structured and non-document structured metadata of a to-be-classified file, then obtaining text node data, document initial traces and node associated data according to the document structured metadata, obtaining context association according to the non-document structured metadata, disassembling the node data according to file content, and associating and merging the node data into a node data chain; the method comprises the following steps: constructing a dynamic knowledge graph, generating a to-be-pre-stored archive category, if the to-be-pre-stored archive category is not in a preset category, matching a similar preset category by using a cosine similarity algorithm, supplementing semantics based on the graph to obtain a newly defined label category, and finally classifying and identifying archives according to the new category, and meanwhile, the preset archive category can be dynamically optimized along with the extension of the new category. Therefore, repeated classification errors are avoided, and the problem that the to-be-preset archive category cannot be classified and identified when the to-be-preset archive category is not matched with the preset archive category can also be avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of file classification, and in particular to a file classification and recognition method and system based on artificial intelligence. Background Art

[0002] File classification is a basic task in file management, which refers to the process of systematically classifying files with internal connections according to certain standards and logic to form an ordered system for efficient management and utilization. Existing file classification and recognition usually match the files to be pre-stored with the preset file categories and identify the document categories to be recognized. However, when the files to be preset do not match the preset file categories, there will be a problem that the files to be pre-stored cannot be classified and recognized. Therefore, a file classification and recognition method based on artificial intelligence is needed to solve the above problems. Summary of the Invention

[0003] The purpose of the present invention is to provide a file classification and recognition method and system based on artificial intelligence to solve the technical problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: A file classification and recognition method based on artificial intelligence, including: Obtaining the digital data of the file to be classified, where the digital data includes document structured metadata and non-document structured metadata; Obtaining text node data according to the document structured metadata, and obtaining initial document trace data and node association data according to the text node data; Obtaining context association data and file content disassembled node data according to the non-document structured metadata, and associating and merging the node association data and the file content disassembled node data to obtain a node data chain; Obtaining a dynamic knowledge graph according to the initial document trace data, the context association data and the node data chain, and obtaining the file category to be pre-stored according to the dynamic knowledge graph; Judging whether the file category to be pre-stored is within the preset file categories; If the file category to be pre-stored is not within, performing a similarity match between the file category to be pre-stored and the preset file categories to obtain a similar preset file category, and performing semantic supplementation of the file category on the similar preset file category based on the dynamic knowledge graph to obtain a newly defined label category; Re-classifying and recognizing the file to be classified according to the newly defined label category.

[0005] Preferably, the steps of obtaining text node data according to the document structured metadata and obtaining initial document trace data and node association data according to the text node data include: Obtain text logic units according to the document structured metadata, where the text logic units include start units, segmentation units, and end units; Split the document structured metadata according to the text logic units to obtain text node data; Obtain the transfer information of the file to be classified according to the text node data, and trace the initial document trace data according to the transfer information; Obtain progressive connection data within a time window according to the start unit, the segmentation unit, and the end unit, and obtain corresponding connection context data according to the connection data, and use the connection context data as node association data.

[0006] Preferably, the steps of obtaining context association data and file content decomposition node data according to the non-document structured metadata include: Obtain the storage path hierarchy and email thread structure of the file to be classified according to the non-document structured metadata, and obtain a cross-file organization topology according to the storage path hierarchy and the email thread structure; Obtain a plurality of connection nodes according to the cross-file organization topology; Obtain corresponding node data according to the plurality of connection nodes, perform semantic extraction on the plurality of node data to obtain a plurality of node semantic data; Associate the plurality of node semantic data according to a preset timestamp to obtain context association data; Decompose the non-document structured metadata based on a preset content structure to obtain a preset decomposition path, and extract file content decomposition node data according to the preset decomposition path.

[0007] Preferably, the steps of obtaining a dynamic knowledge graph according to the initial document trace data, the context association data, and the node data chain, and obtaining the categories of files to be pre-stored according to the dynamic knowledge graph include: Obtain basic document data according to the initial document trace data, and map the basic document data into a coding space according to a preset receiving time sequence for coding to obtain a basic document code; Obtain context-associated text data according to the context association data, perform keyword extraction on the associated text data to obtain context label data, and map the context label data into a coding space according to a preset receiving time sequence for coding to obtain a context label code; Obtain the data chain structure features according to the node data chain, where the data chain structure features include category inheritance structure features, behavior-driven structure features, and time limit constraint structure features; Map the category inheritance structure features, the behavior-driven structure features, and the time limit constraint structure features into the encoding space according to the preset receiving time sequence respectively for encoding, to obtain the category inheritance structure feature encoding, the behavior-driven structure feature encoding, and the time limit constraint structure feature encoding; Perform encoding splicing on the basic document encoding, the context label encoding, the category inheritance structure feature encoding, the behavior-driven structure feature encoding, and the time limit constraint structure feature encoding to obtain the document splicing encoding, and use the spliced document splicing encoding as the dynamic knowledge graph.

[0008] Preferably, the step of supplementing the semantics of the archive category for the similar preset archive category based on the dynamic knowledge graph to obtain the newly defined label category includes: Obtain the first high-frequency key behavior data according to the dynamic knowledge graph, and extract keywords according to the first high-frequency key behavior data to obtain the first high-frequency key behavior keywords; Obtain the second high-frequency key behavior data according to the similar preset archive category, and extract keywords according to the second high-frequency key behavior data to obtain the second high-frequency key behavior keywords; Perform semantic splicing on the first high-frequency key behavior keywords and the second high-frequency key behavior keywords to obtain the newly defined label category.

[0009] Preferably, the step of reclassifying and identifying the to-be-classified archive according to the newly defined label category includes: Based on the similar preset archive category, split the newly defined label category into a similar archive category and an amplified archive subcategory; Construct a reference sample of the clustering model according to the similar archive category and the amplified archive subcategory, and expand the clustering model with the reference sample to obtain an expanded clustering model, Input the to-be-classified archive into the expanded clustering model, and perform reclassification and identification on the to-be-classified archive through K-means clustering.

[0010] This application also provides an archive classification and identification system based on artificial intelligence, including: The first acquisition module is used to acquire the digital data of the to-be-classified archive, where the digital data includes document structured metadata and non-document structured metadata; The second acquisition module is used to obtain text node data according to the document structured metadata, and obtain document initial trace data and node association data according to the text node data; A third acquisition module, configured to acquire context-associated data and archive content decomposition node data according to the non-document structured metadata, and perform associated merging on the node-associated data and the archive content decomposition node data to obtain a node data chain; A fourth acquisition module, configured to acquire a dynamic knowledge graph according to the document initial trace data, the context-associated data, and the node data chain, and acquire a to-be-prestored archive category according to the dynamic knowledge graph; A judgment module, configured to judge whether the to-be-prestored archive category is within a preset archive category; If the to-be-prestored archive category is not within, perform similarity matching on the to-be-prestored archive category and the preset archive category to obtain a similar preset archive category, and perform archive category semantic supplementation on the similar preset archive category based on the dynamic knowledge graph to obtain a newly defined label category; An identification module, configured to re-classify and identify the to-be-classified archive according to the newly defined label category.

[0011] Preferably, the second acquisition module includes: A first acquisition unit, configured to acquire a text logic unit according to the document structured metadata, where the text logic unit includes a start unit, a segmentation unit, and an end unit; A splitting unit, configured to split the document structured metadata according to the text logic unit to obtain text node data; A second acquisition unit, configured to acquire the transfer information of the to-be-classified archive according to the text node data, and trace the document initial trace data according to the transfer information; A third acquisition unit, configured to acquire progressive connection data within a time window according to the start unit, the segmentation unit, and the end unit, and acquire corresponding connection context data according to the connection data, and use the connection context data as node-associated data.

[0012] The present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0013] The present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0014] The beneficial effects of this application are as follows: In the present invention, digital data of files to be classified is first obtained, including document-structured and non-document-structured metadata. Then, text node data, initial document trace data, and node association data are obtained based on the document-structured metadata. Context association data and file content disassembled node data are obtained according to the non-document-structured metadata, and they are combined and merged to form a node data chain. Based on this, a dynamic knowledge graph is constructed to generate the categories of files to be pre-stored. If it is not within the preset categories, the cosine similarity algorithm is used to match similar preset categories, and new defined label categories are obtained by supplementing semantics based on the graph. Finally, the files are re-classified and identified according to the new categories, realizing the automatic expansion of the category system and the accurate positioning of files. Furthermore, the preset file categories can be dynamically optimized as the new categories expand, avoiding repeated classification errors, and also being able to avoid the problem that when the categories of files to be pre-stored do not match the preset file categories, the files to be pre-stored cannot be classified and identified. Brief Description of the Drawings

[0015] Figure 1 It is a schematic flowchart of the method according to an embodiment of this application.

[0016] Figure 2 It is a schematic structural diagram of the system according to an embodiment of this application.

[0017] Figure 3 It is a schematic internal structure diagram of a computer device according to an embodiment of this application.

[0018] The realization, functional features, and advantages of the purpose of this application will be further described with reference to the embodiments and the accompanying drawings. Detailed Embodiments

[0019] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0020] As Figures 1-3 shown, this application provides an artificial-intelligence-based file classification and recognition method, including: S1. Obtain the digital data of the files to be classified, where the digital data includes document-structured metadata and non-document-structured metadata; S2. Obtain text node data according to the document-structured metadata, and obtain initial document trace data and node association data according to the text node data; S3. Obtain context association data and file content disassembled node data according to the non-document-structured metadata, and combine and merge the node association data and the file content disassembled node data to obtain a node data chain; S4. Obtain a dynamic knowledge graph according to the initial document trace data, the context association data, and the node data chain, and obtain the categories of files to be pre-stored according to the dynamic knowledge graph; S5. Determine whether the file category to be pre-stored is within the preset file categories; If the file category to be pre-stored is not within the preset file categories, perform a similarity match between the file category to be pre-stored and the preset file categories to obtain a similar preset file category, and perform semantic supplementation of the file category for the similar preset file category based on the dynamic knowledge graph to obtain a newly defined tag category; S6. Re-classify and identify the file to be classified according to the newly defined tag category.

[0021] As described in the above steps S1 - S6, since the existing file classification and identification usually match the file category to be pre-stored with the preset file categories and identify the document category to be identified, but when the preset file category to be pre-stored does not match the preset file categories, there will be a problem that the file category to be pre-stored cannot be classified and identified. The present invention first obtains the digital data of the file to be classified, where the digital data includes document structured metadata and non-document structured metadata, so as to comprehensively collect the structured (such as text content, metadata tags) and unstructured data (such as storage path, email thread) of the file, providing a multi-dimensional information basis for subsequent analysis, and at the same time obtaining structured and unstructured data, breaking through the limitation of traditional methods that only rely on text content. For example, it can capture the department attribution reflected in the storage path of the file or the business process nodes implied in the email correspondence; The specific process of obtaining the digital data of the file to be classified: Adopt the method of docking distributed crawlers with API interfaces to crawl data in real time from multi-source heterogeneous platforms such as OA systems, email servers, and file storage systems, and perform standardization processing through ETL (Extract-Transform-Load) tools to ensure data integrity and consistency; Then, obtain text node data according to the document structured metadata, and obtain document initial trace data and node association data according to the text node data. In this way, through logical unit splitting, a long document is transformed into a computable node network. For example, the "fact statement - legal basis - judgment result" structure in legal documents can be accurately identified, and at the same time, the document initial trace data can record the modification history. For example, the multiple amendment records of patent application documents can be used as classification references (such as distinguishing between "first draft" and "authorized version"); Then, context-related data and archive content decomposition node data are obtained according to the non-document structured metadata, and the node-related data and the archive content decomposition node data are associated and merged to obtain a node data chain. In this way, the association between different archives is identified through the topological structure. For example, the "project establishment report" and the "acceptance report" of the same project can be associated through the storage path and the email process and classified into the same project category. At the same time, non-text data such as the storage path and the email process are converted into computable semantic tags. For example, the "Finance Department" path is directly mapped to the classification feature of the "finance category" archive, and the unstructured data is decomposed into a node chain, and the scattered information is integrated into an ordered data structure to provide association support for the knowledge graph; Secondly, a dynamic knowledge graph is obtained according to the document initial trace data, the context-related data, and the node data chain, and a to-be-prestored archive category is obtained according to the dynamic knowledge graph. In this way, the dynamic knowledge graph can comprehensively represent the archive features, and the knowledge graph can integrate multi-source heterogeneous data. The dynamic characteristics can adapt to the changes of the archive content and provide semantic support for category generation. At the same time, the to-be-prestored category is automatically generated through graph mapping to realize data-driven intelligent classification and avoid the limitations of manual definition; Then, it is judged whether the to-be-prestored archive category is within the preset archive category. If the to-be-prestored archive category is not, the to-be-prestored archive category is matched with the preset archive category for similarity to obtain a similar preset archive category, and the archive category semantics of the similar preset archive category is supplemented based on the dynamic knowledge graph to obtain a newly defined label category. In this way, there is no need for manual maintenance of the category library, and the system can automatically expand the category system according to the new archive features. And through the splicing of high-frequency behavior keywords, it is ensured that the new category not only conforms to the existing classification logic but also contains new features, providing a basic basis for subsequent re-identification. Among them, the above judgment method is implemented based on the cosine similarity algorithm; Finally, the to-be-classified archive is re-classified and identified according to the newly defined label category. In this way, when the newly defined label category is readjusted according to the to-be-classified archive and the to-be-classified archive fails to match the corresponding category, the newly defined label category is added to identify the to-be-classified archive. Furthermore, the preset archive category can be dynamically optimized as the new category expands, avoiding repeated classification errors, ensuring the accurate positioning of the new archive in the existing classification system, and also avoiding the problem that the to-be-prestored archive category cannot be classified and identified when it does not match the preset archive category.

[0022] In one embodiment, step S2 of obtaining text node data according to the document structured metadata and obtaining document initial trace data and node-related data according to the text node data includes: S201. Obtain text logic units according to the document structured metadata, where the text logic units include a starting unit, a segmentation unit, and an ending unit; S202. Split the document structured metadata according to the text logic units to obtain text node data; S203. Obtain the transfer information of the file to be classified according to the text node data, and trace the initial trace data of the document according to the transfer information; S204. Obtain progressive connection data within a time window according to the starting unit, the segmentation unit, and the ending unit, and obtain corresponding connection context data according to the connection data, and use the connection context data as node association data.

[0023] As described in the above steps S201 - S204, the present invention first obtains text logic units according to the document structured metadata, where the text logic units include a starting unit, a segmentation unit, and an ending unit. In this way, the logical framework of the text is extracted from the structured metadata, and the unordered text is divided into units with clear starting, segmentation, and ending attributes, providing a logical anchor for subsequent data splitting. At the same time, it can also achieve the structured parsing of the text, converting the unstructured text content into logical units that can be understood by a computer, laying a foundation for accurately extracting node data; Among them, through the syntactic analysis technology of natural language processing (NLP), the document structured metadata is pre - processed to identify the logical starting identifiers of the document (such as "Table of Contents", "Introduction", "Chapter 1", etc.), segmentation marks (such as chapter titles, first - line indentation of paragraphs, delimiters), and ending features (such as "Conclusion", "References", "Appendix", etc.), and use the Hidden Markov Model (HMM) or Conditional Random Field (CRF) algorithm to divide the document into three types of logical units (starting unit, segmentation unit, and ending unit) based on document format features (such as font size, bold style, paragraph spacing) and semantic features (such as the overview statement of the starting paragraph, the summary statement of the ending paragraph); Then, split the document structured metadata according to the text logic units to obtain text node data. In this way, the complete document can be split into independent text nodes (such as chapters, paragraphs, sub - headings) according to the logical units, each node corresponding to a specific semantic module, and the granularity splitting of the document content can be achieved, making the data processing refined from "full - text analysis" to "node analysis" and improving the accuracy of information extraction; Next, obtain the transfer information of the file to be classified according to the text node data, and trace the initial trace data of the document according to the transfer information. In this way, the transfer records such as the creation, modification, and approval of the document can be extracted from the text nodes, forming a complete version iteration chain. Moreover, by tracing the initial traces (such as the creator, modification time, approver), background attributes (such as department affiliation, business process stage) can be provided for file classification, enhancing the comprehensiveness of the classification basis; Finally, obtain the progressive connection data within the time window according to the starting unit, the segmenting unit, and the ending unit, and obtain the corresponding connection context data according to the connection data. Use the connection context data as node association data. In this way, the progressive relationship of the starting, segmenting, and ending units in the time dimension (such as causal, transitional, and chronological connections between paragraphs) is extracted, forming a logical chain between nodes. At the same time, a chronological association network of the document content is constructed, making the independent node data form an overall with context logic, providing relational data support for the knowledge graph; Among them, set time window parameters (such as the last 7 days, 30 days, etc.), and filter the text node data within this time range according to the creation time or the latest modification time of the document. The time window can be dynamically adjusted. For example, an hourly window is used for real-time updated documents, and an annual window is used for historical documents.

[0024] In one embodiment, the step S3 of obtaining context association data and file content disassembled node data according to the non-document structured metadata includes: S301. Obtain the storage path hierarchy and email thread structure of the file to be classified according to the non-document structured metadata, and obtain the cross-file organization topology according to the storage path hierarchy and the email thread structure; S302. Obtain multiple connection nodes according to the cross-file organization topology; S303. Obtain the corresponding node data according to the multiple connection nodes, perform semantic extraction on the multiple node data, and obtain multiple node semantic data; S304. Associate the multiple node semantic data according to a preset timestamp to obtain context association data; S305. Decompose the non-document structured metadata based on a preset content structure to obtain a preset decomposition path, and extract file content disassembled node data according to the preset decomposition path.

[0025] As described in the above steps S301 - S305, the present invention first obtains the storage path hierarchy and email thread structure of the file to be classified according to the non - document structured metadata. Among them, the storage path of the file to be classified is traversed through the file system interface, and the multi - level path (such as server / company / finance department / 2024 / budget report.docx) is disassembled into hierarchical nodes, stored in a tree - like structure, and the cross - file organization topology is obtained according to the storage path hierarchy and the email thread structure. In this way, the storage path (such as folder hierarchy) and email thread relationship (such as email communication chain) of the file are extracted from the unstructured data, and a cross - file organization association topology graph is constructed to present the location and association network of the file in the enterprise data architecture. Secondly, the storage path reflects the physical organization logic of the file (such as department, time, business type), the email thread reflects the collaboration relationship in the business process, and the topology graph integrates these implicit associations to provide a "context scenario" basis for classification; Then, multiple connection nodes are obtained according to the cross - file organization topology. In this way, key connection nodes (such as departments, personnel, suppliers, project names, etc.) are extracted from the topology graph as the basic units for semantic extraction. Secondly, the connection nodes are the core carriers of the topological relationship. After extraction, the key entities in the data can be focused on, providing target objects for subsequent semantic analysis. At the same time, the essence of the topology graph is a set of nodes and edges. Nodes represent entities and edges represent relationships. Extracting nodes is a basic step in deconstructing the topological relationship, facilitating the subsequent targeted extraction of semantic information related to entities; Secondly, corresponding node data are obtained according to the multiple connection nodes, and semantic extraction is performed on the multiple node data to obtain multiple node semantic data. In this way, for each connection node, the text content related to it is extracted from the file data, and structured semantic data are generated through semantic analysis (such as keyword extraction, entity recognition), and the unstructured node association information is transformed into computable semantic features, providing semantic support for constructing context - associated data, upgrading the topological relationship from "structural association" to "semantic association". At the same time, a simple node name lacks semantic connotation, and key information needs to be extracted in combination with the text content related to the node (such as email body, path description) to reflect the actual role of the node in the file (such as the "Legal Affairs Department" node corresponding to the semantic meaning of "contract clause review"); Next, associate the multiple node semantic data according to a preset timestamp to obtain context-associated data. In this way, the node semantic data is sorted and associated in chronological order (such as the email sending time, path update time) to form a context chain with temporal logic, and a time dimension is given to the semantic data to reflect the development context of the archival business process (such as the chronological order of "first draft → revision → review"), making the context-associated data a "traceable semantic process". At the same time, the chronological order of the business process is a key factor in understanding the archival background. For example, the chronological order of "drafting → review → signing" of a procurement contract directly affects its classification stage (such as "pending review", "effective"), and this logic needs to be solidified through timestamps; Finally, disassemble the non-document structured metadata based on a preset content structure to obtain a preset disassembly path, and extract the archival content disassembly node data according to the preset disassembly path. In this way, according to the preset content disassembly rules (such as "business type → time → participant"), the unstructured metadata (such as storage path, email subject) is disassembled into node data in a standard format, and the unstructured data is converted into structured disassembly nodes, which are merged with the node data chain of the structured metadata to form a unified data structure, facilitating subsequent knowledge graph construction.

[0026] In one embodiment, step S4 of obtaining a dynamic knowledge graph according to the document initial trace data, the context-associated data, and the node data chain, and obtaining the archival categories to be pre-stored according to the dynamic knowledge graph includes: S401. Obtain basic document data according to the document initial trace data, and map the basic document data into a coding space according to a preset receiving chronological order for coding to obtain a basic document code; S402. Obtain context-associated text data according to the context-associated data, extract keywords from the associated text data to obtain context label data, and map the context label data into a coding space according to a preset receiving chronological order for coding to obtain a context label code; S403. Obtain the data chain structure features according to the node data chain, where the data chain structure features include category inheritance structure features, behavior-driven structure features, and time limit constraint structure features; S404. Map the category inheritance structure features, the behavior-driven structure features, and the time limit constraint structure features into a coding space according to a preset receiving chronological order for coding to obtain a category inheritance structure feature code, the behavior-driven structure feature code, and the time limit constraint structure feature code; S405. Encode and splice the encoded basic document, the encoded context label, the encoded category inheritance structure feature, the encoded behavior-driven structure feature, and the encoded time limit constraint structure feature to obtain a document splicing code, and use the obtained document splicing code as the dynamic knowledge graph.

[0027] As described in the above steps S401 - S405, the present invention first obtains basic document data according to the initial document trace data, maps the basic document data to the encoding space according to the preset receiving time sequence for encoding to obtain a basic document code. In this way, the basic content (text body, keywords) is extracted from the initial document trace data (such as creator, modification record, transfer information), mapped to the encoding space according to the preset time sequence to generate a structured basic document code, and the core content of the document is converted into a computer-processable encoding format. While preserving the content semantics, it provides underlying data support for the subsequent construction of the knowledge graph to ensure the computability of content features. At the same time, the basic document data is the core basis for file classification. By encoding, natural language is converted into a numerical representation in the vector space (such as word embedding), which meets the input requirements of the artificial intelligence model and is convenient for subsequent fusion with other features; Next, obtain the associated text data of the context according to the context association data, extract keywords from the associated text data to obtain context label data, and map the context label data to the encoding space according to the preset receiving time sequence for encoding to obtain a context label code. In this way, keywords are extracted from the context association data (such as email approval opinions, storage path levels) as context labels (such as "urgent", "quarterly summary"), encoded as context label codes in chronological order, and the implicit scenario features of the document (such as business urgency, time period) are supplemented, so that the classification is not only based on the content itself, but also combined with the context environment in which it is generated, improving the accuracy of classification. At the same time, the context information of the document (such as the "urgent" flag in the approval process, the "quarterly" label in the storage path) is an important auxiliary basis for classification. By encoding it into a feature vector, it can form a complement with the basic content code; Then, obtain the data chain structure features according to the node data chain, where the data chain structure features include category inheritance structure features, behavior-driven structure features, and time limit constraint structure features. In this way, the logical relationship of the data chain is abstracted into structured features, enabling the classification model to understand the "business logic context" of the file, rather than relying only on the surface content. And the essence of the node data chain is the digital mapping of the business process. By extracting the three types of structure features, the complex relationship network can be converted into structured parameters recognizable by the model to support the semantic reasoning of the knowledge graph; Secondly, map the category inheritance structure feature, the behavior-driven structure feature, and the time-limit constraint structure feature into the coding space respectively according to the preset receiving time sequence for coding, so as to obtain the category inheritance structure feature coding, the behavior-driven structure feature coding, and the time-limit constraint structure feature coding. In this way, the three types of structure features are respectively mapped into the coding space to generate corresponding feature codings (such as category inheritance coding, behavior-driven coding, time-limit constraint coding), preserving the time sequence logic of the features, and transforming the abstract structure features into computable numerical vectors, which is convenient for splicing and fusing with other codings (basic documents, context labels) to form a complete file feature representation. At the same time, the time sequence of the structure features (such as the sequence of processes and the validity of time constraints) is crucial for classification. Through time sequence coding, the logical order of the features can be ensured to be preserved in the vector space, avoiding semantic distortion; Finally, splice the basic document coding, the context label coding, the category inheritance structure feature coding, the behavior-driven structure feature coding, and the time-limit constraint structure feature coding to obtain the document splicing coding, and use the spliced document splicing coding as the dynamic knowledge graph. In this way, the basic document coding, the context label coding, and the three types of structure feature codings are spliced according to the rules to form a multi-dimensional document splicing coding, which is used as the digital representation of the dynamic knowledge graph. The knowledge graph integrates the content, context, structural relationship, and time sequence features of the file, realizing the upgrade from "single content classification" to "full-dimensional semantic understanding", supporting intelligent category generation and matching. At the same time, a single feature cannot comprehensively represent the complexity of the file. The knowledge graph formed by coding splicing can be regarded as the "digital twin" of the file, and its dynamic nature is reflected in its ability to be updated in real time with the input of new data to adapt to the changes in the business scenario; Specific implementation process: Unify the basic document coding, the context label coding, the category inheritance structure coding, the behavior-driven structure coding, and the time-limit constraint structure coding into the same dimension, adjust the vector space dimension of each coding through linear transformation (such as a fully connected layer), and then splice various codings in the time order of data generation. For example, first splice the basic document coding (generated at creation), and then sequentially splice the context label coding (generated at approval) and the structure feature coding (generated at filing) to form a dynamically growing coding sequence. Among them, node generation: Each coding segment serves as a node in the graph, containing the coding vector, time stamp attribute, and edge relationship definition: Adjacent coding segments are connected by time sequence edges, and at the same time, semantically related coding segments (such as "budget" in the basic document coding and "budget approval" in the context label coding) are connected by semantic edges. When there is new data in the document (such as new modification records, supplementary approval opinions), a new coding segment is generated and added to the graph, and the node representation of the entire graph is updated through a graph neural network (GNN) to ensure that the graph evolves dynamically with the document life cycle.

[0028] In one embodiment, step S5 of supplementing the semantic meaning of the file category for similar preset file categories based on the dynamic knowledge graph to obtain a newly defined tag category includes: S501. Obtain the first high-frequency key behavior data according to the dynamic knowledge graph, and extract keywords according to the first high-frequency key behavior data to obtain the first high-frequency key behavior keywords; S502. Obtain the second high-frequency key behavior data according to the similar preset file category, and extract keywords according to the second high-frequency key behavior data to obtain the second high-frequency key behavior keywords; S503. Semantically splice the first high-frequency key behavior keywords and the second high-frequency key behavior keywords to obtain a newly defined tag category.

[0029] As described in steps S501-S503 above, the present invention first obtains the first high-frequency key behavior data according to the dynamic knowledge graph, and extracts keywords according to the first high-frequency key behavior data to obtain the first high-frequency key behavior keywords. In this way, high-frequency behavior patterns related to the file to be classified (such as creation, review, revision, etc.) are identified from the dynamic knowledge graph, and keywords representing these behaviors are extracted to form the first high-frequency key behavior keywords. Moreover, the high-frequency behavior data reflects the core action characteristics of the file in the business process, and keyword extraction can transform the implicit behavior pattern into an explicit semantic label, providing a basis for the "behavior dimension" of the category semantic supplement. At the same time, the dynamic knowledge graph integrates all-dimensional features of the file, and the behavior-driven structure is a direct manifestation of the business process. By extracting high-frequency behavior keywords, it can be ensured that the newly defined tag category can accurately reflect the actual processing scenario of the file (such as actions like "review" and "revision"); Next, obtain the second high-frequency key behavior data according to the similar preset file category, and extract keywords according to the second high-frequency key behavior data to obtain the second high-frequency key behavior keywords. In this way, from the historical data of the similar preset file category, the high-frequency behavior patterns and keywords of the files under this category are extracted to form the second high-frequency key behavior keywords. Moreover, the high-frequency behaviors of the similar categories reflect the general business logic of this type of file. By extracting keywords, it can be ensured that the newly defined tag category maintains logical consistency with the existing classification system, avoiding confusion in the classification system. At the same time, the preset category system is the summary of long-term business practices, and the high-frequency behavior keywords of the similar categories represent the "standard behavior patterns" of this type of file. Combining the behavior characteristics of the file to be classified with the standard pattern can achieve the integration of the "old category logic" and the "new file characteristics"; Finally, semantically splice the first high-frequency key behavior keywords and the second high-frequency key behavior keywords to obtain a newly defined label category. In this way, the first high-frequency keywords of the file to be classified are semantically combined with the second high-frequency keywords of similar categories to generate a newly defined label category that combines the characteristics of the new file and the logic of the existing category. And through semantic splicing, the new label can not only accurately describe the unique behavior characteristics of the file to be classified (such as "document amendment"), but also inherit the classification logic of similar categories (such as "patent application"), realizing the dynamic expansion of the classification system. At the same time, semantic-level splicing of keywords is achieved through vector calculation to avoid ambiguity caused by simple string splicing.

[0030] In one embodiment, step S6 of re-classifying and identifying the file to be classified according to the newly defined label category includes: S601. Based on the similar preset file categories, split the newly defined label category into a similar file category and an amplified file sub-category; S602. Construct a reference sample of the clustering model according to the similar file category and the amplified file sub-category, and expand the clustering model with the reference sample to obtain an expanded clustering model. S603. Input the file to be classified into the expanded clustering model, and re-classify and identify the file to be classified through K-means clustering.

[0031] As described in the above steps S601-S603, the present invention first splits the newly defined label category into a similar file category and an amplified file sub-category based on the similar preset file categories. In this way, the newly defined label category is split into a "similar file category" and an "amplified file sub-category". The former inherits the basic classification logic of the preset category, and the latter supplements the unique attributes of the new file. And through splitting, the balance between "classification system compatibility" and "new file specificity" is achieved, avoiding the impact of the new category on the original classification system, and at the same time ensuring that the unique attributes are accurately represented. At the same time, the newly defined label category is essentially an extension of the preset category. After splitting, the "general classification logic" and the "new features" can be clearly distinguished, providing a clear reference sample structure for the subsequent expansion of the clustering model. Next, construct a reference sample of the clustering model according to the similar file category and the amplified file sub-category, and expand the clustering model with the reference sample to obtain an expanded clustering model. In this way, according to the split similar categories and amplified sub-categories, a reference sample containing new and old features is constructed for training the clustering model to enable it to have the ability to identify new categories. And after the clustering model is expanded, it can process traditional files and new files at the same time, avoiding manual maintenance of the model, realizing the self-evolution of the classification system, and improving the adaptability to changes in business scenarios. Finally, input the file to be classified into the expanded clustering model, and re-classify and identify the file to be classified through K-means clustering. In this way, input the file to be classified into the expanded clustering model, calculate its distance from each clustering center through the K-means algorithm, achieve precise classification based on the newly defined label categories, and utilize the learning ability of the expanded model to automatically map new files to the correct categories, avoiding manual intervention, improving the classification efficiency and accuracy, especially suitable for newly emerging files that appear frequently. At the same time, K-means clustering is an efficient unsupervised learning algorithm that can quickly adapt to new features by updating the clustering centers. Combining it with the expansion of reference samples can achieve dynamic expansion of the classification ability without changing the model architecture, conforming to the self-optimization logic of the artificial intelligence system, ensuring the precise positioning of new files in the existing classification system, and also avoiding the problem that when the preset file category does not match the preset file category, it is impossible to classify and identify the preset file category; Specific steps: The specific steps are to divide the expanded clustering model into K clusters, and then each cluster is represented by a centroid. The clustering model algorithm continuously optimizes the centroid position through iteration, minimizing the distance between the samples in the same cluster and the centroid. For example, for the extraction of files to be classified, regard the similar file categories and the expanded file sub-categories as each cluster being represented by a centroid, and then regard the file to be classified as a sample, making the file to be classified close to the similar file categories and the expanded file sub-categories, so as to gather the same categories and obtain the file category of the file to be classified.

[0032] This application also provides an archive classification and recognition system based on artificial intelligence, including: The first acquisition module is used to acquire the digital data of the file to be classified, where the digital data includes document structured metadata and non-document structured metadata; The second acquisition module is used to acquire text node data according to the document structured metadata, and acquire document initial trace data and node association data according to the text node data; The third acquisition module is used to acquire context association data and archive content disassembled node data according to the non-document structured metadata, and perform associative merging on the node association data and the archive content disassembled node data to obtain a node data chain; The fourth acquisition module is used to acquire a dynamic knowledge graph according to the document initial trace data, the context association data and the node data chain, and acquire the preset file category according to the dynamic knowledge graph; The judgment module is used to judge whether the preset file category is within the preset file categories; If the file category to be pre-stored is not in a certain state, perform a similarity match between the file category to be pre-stored and a preset file category to obtain a similar preset file category, and perform semantic supplementation of the file category on the similar preset file category based on a dynamic knowledge graph to obtain a newly defined label category; An identification module for re-classifying and identifying the file to be classified according to the newly defined label category.

[0033] In one embodiment, the second acquisition module includes: A first acquisition unit for acquiring text logic units according to the document structured metadata, where the text logic units include a start unit, a segmentation unit, and an end unit; A splitting unit for splitting the document structured metadata according to the text logic units to obtain text node data; A second acquisition unit for acquiring the circulation information of the file to be classified according to the text node data, and tracing the initial trace data of the document according to the circulation information; A third acquisition unit for acquiring progressive connection data within a time window according to the start unit, the segmentation unit, and the end unit, and acquiring corresponding connection context data according to the connection data, and using the connection context data as node association data.

[0034] This application also provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0035] This application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0036] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided in this application and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be obtained in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0037] It should be noted that, in this article, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, apparatus, article or method comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, apparatus, article or method. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, apparatus, article or method comprising the element.

[0038] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall equally be included in the patent protection scope of the present invention.

Claims

1. An artificial intelligence-based method for file classification and recognition, characterized in that, Including: Obtain the digital data of the file to be classified, where the digital data includes document structured metadata and non-document structured metadata; Obtain text node data according to the document structured metadata, and obtain initial document trace data and node association data according to the text node data; Obtain context association data and file content disassembled node data according to the non-document structured metadata, and perform correlation and merging on the node association data and the file content disassembled node data to obtain a node data chain; Obtain a dynamic knowledge graph according to the initial document trace data, the context association data and the node data chain, and obtain the file category to be pre-stored according to the dynamic knowledge graph; Judge whether the file category to be pre-stored is within the preset file categories; If the file category to be pre-stored is not within, perform a similarity match between the file category to be pre-stored and the preset file categories to obtain a similar preset file category, and perform semantic supplementation of the file category on the similar preset file category based on the dynamic knowledge graph to obtain a newly defined label category; Re-classify and identify the file to be classified according to the newly defined label category.

2. The method for classifying and identifying archives based on artificial intelligence according to claim 1, wherein The step of obtaining text node data according to the document structured metadata and obtaining initial document trace data and node association data according to the text node data includes: Obtain text logical units according to the document structured metadata, where the text logical units include a start unit, a segmentation unit and an end unit; Split the document structured metadata according to the text logical units to obtain text node data; Obtain the transfer information of the file to be classified according to the text node data, and trace the initial document trace data according to the transfer information; Obtain progressive connection data within a time window according to the start unit, the segmentation unit and the end unit, and obtain corresponding connection context data according to the connection data, and use the connection context data as node association data.

3. The method for classifying and identifying archives based on artificial intelligence according to claim 1, wherein The step of obtaining context association data and file content disassembled node data according to the non-document structured metadata includes: Obtain the storage path hierarchy and email thread structure of the file to be classified according to the non-document structured metadata, and obtain a cross-file organization topology according to the storage path hierarchy and the email thread structure; Obtain a plurality of connection nodes according to the cross-file organization topology; Obtain corresponding node data according to the plurality of connection nodes, perform semantic extraction on the plurality of node data to obtain a plurality of node semantic data; Associate the plurality of node semantic data according to a preset timestamp to obtain context association data; Disassemble the non-document structured metadata based on a preset content structure to obtain a preset disassembly path, and extract file content disassembled node data according to the preset disassembly path.

4. The method for classifying and identifying archives based on artificial intelligence according to claim 1, wherein The step of obtaining a dynamic knowledge graph according to the initial document trace data, the context association data and the node data chain and obtaining the file category to be pre-stored according to the dynamic knowledge graph includes: Obtain the basic document data according to the initial trace data of the document, and map the basic document data into the coding space according to the preset receiving time sequence for coding to obtain the basic document coding; Obtain the associated text data of the context according to the context associated data, extract keywords from the associated text data to obtain the context label data, and map the context label data into the coding space according to the preset receiving time sequence for coding to obtain the context label coding; Obtain the data chain structure features according to the node data chain, where the data chain structure features include category inheritance structure features, behavior-driven structure features, and time limit constraint structure features; Map the category inheritance structure features, the behavior-driven structure features, and the time limit constraint structure features into the coding space according to the preset receiving time sequence for coding to obtain the category inheritance structure feature coding, the behavior-driven structure feature coding, and the time limit constraint structure feature coding; Perform coding splicing on the basic document coding, the context label coding, the category inheritance structure feature coding, the behavior-driven structure feature coding, and the time limit constraint structure feature coding to obtain the document splicing coding, and use the spliced document splicing coding as the dynamic knowledge graph.

5. The method for classifying and identifying archives based on artificial intelligence according to claim 1, wherein The step of supplementing the semantics of the file category for the similar preset file categories based on the dynamic knowledge graph to obtain the newly defined label categories includes: Obtain the first high-frequency key behavior data according to the dynamic knowledge graph, and extract keywords from the first high-frequency key behavior data to obtain the first high-frequency key behavior keywords; Obtain the second high-frequency key behavior data according to the similar preset file categories, and extract keywords from the second high-frequency key behavior data to obtain the second high-frequency key behavior keywords; Perform semantic splicing on the first high-frequency key behavior keywords and the second high-frequency key behavior keywords to obtain the newly defined label categories.

6. The method for classifying and identifying archives based on artificial intelligence according to claim 1, wherein The step of re-classifying and identifying the file to be classified according to the newly defined label categories includes: Based on the similar preset file categories, split the newly defined label categories into similar file categories and amplified file sub-categories; Construct a reference sample of the clustering model according to the similar file categories and the amplified file sub-categories, and expand the clustering model with the reference sample to obtain an expanded clustering model; Input the file to be classified into the expanded clustering model, and perform re-classification and identification on the file to be classified through K-means clustering.

7. An archive classification and recognition system based on artificial intelligence, characterized in that, Include: The first acquisition module is used to acquire the digital data of the file to be classified, where the digital data includes document structured metadata and non-document structured metadata; The second acquisition module is used to obtain the text node data according to the document structured metadata, and obtain the document initial trace data and node association data according to the text node data; The third acquisition module is used to obtain the context associated data and the file content disassembled node data according to the non-document structured metadata, and perform joint merging on the node association data and the file content disassembled node data to obtain the node data chain; A fourth acquisition module, configured to obtain a dynamic knowledge graph according to the initial document trace data, the context association data, and the node data chain, and obtain the file categories to be pre-stored according to the dynamic knowledge graph; A judgment module, configured to judge whether the file categories to be pre-stored are within the preset file categories; If the file categories to be pre-stored are not within the preset file categories, perform a similarity match between the file categories to be pre-stored and the preset file categories to obtain similar preset file categories, and perform semantic supplementation of the file categories on the similar preset file categories based on the dynamic knowledge graph to obtain newly defined label categories; An identification module, configured to re-classify and identify the files to be classified according to the newly defined label categories.

8. An artificial intelligence-based file classification and recognition system according to claim 7, characterized in that, The second acquisition module includes: A first acquisition unit, configured to obtain text logic units according to the document structured metadata, where the text logic units include start units, segmentation units, and end units; A splitting unit, configured to split the document structured metadata according to the text logic units to obtain text node data; A second acquisition unit, configured to obtain the transfer information of the files to be classified according to the text node data, and trace the initial document trace data according to the transfer information; A third acquisition unit, configured to obtain progressive connection data within a time window according to the start unit, the segmentation unit, and the end unit, and obtain corresponding connection context data according to the connection data, and use the connection context data as node association data.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images