Automatic archive category judgment and description template matching method based on pattern recognition

Abstract

The invention relates to the field of artificial intelligence and pattern recognition, and discloses an archive category automatic judgment and description template matching method based on pattern recognition. The method comprises the following steps: acquiring original data and metadata of an archive; performing structured analysis and semantic feature extraction to generate a multi-dimensional feature vector; the archive category is judged through the ensemble classification model; searching a description template with the highest matching degree in the dynamic template library, and automatically generating and registering a new template if there is no high matching template; and finally, automatic recording is completed. Through a dynamic template mechanism and online incremental learning, full-automatic, self-adaptive and sustainable evolution of archive description is realized, and description efficiency, consistency and system robustness are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and pattern recognition, specifically relating to a method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition. Background Technology

[0002] With the deepening of digital transformation, archival management plays an increasingly important foundational role in key areas such as government affairs, finance, healthcare, and corporate operations. Archival cataloging, as the core link in the organization and retrieval of archival information, directly affects the efficiency of subsequent classification, storage, sharing, and intelligent analysis due to its degree of standardization and structuring. Currently, mainstream archival management systems generally rely on preset static cataloging templates, manually configuring field rules, format constraints, and metadata mapping relationships to achieve structured parsing of specific types of archives. While this method is effective in handling traditional archives with stable formats and clearly defined categories, its underlying logic is based on the assumption that "archival types are known a priori and remain unchanged over time," making it difficult to adapt to the rapid evolution and increasing diversity of archival forms in real-world scenarios.

[0003] Automatic classification technology for archives based on pattern recognition has become a research hotspot in recent years. It aims to automatically assign category labels by semantically analyzing and extracting features from archive content using optical character recognition (OCR), natural language processing (NLP), and deep learning models. The fundamental goal of this research is to reduce human intervention and improve the automation level of archive classification.

[0004] After classifying the documents, the existing system still relies on manual pre-assignment of corresponding cataloging templates for each type of document, resulting in a serious disconnect between the "identification" and "cataloging" stages. When faced with non-standard documents generated by emerging business or historical documents with frequently changing formats, the system often fails to generate accurate and complete structured metadata due to its lack of dynamic evaluation and adjustment capabilities for template applicability, and may even become completely ineffective.

[0005] Existing bibliographic template matching mechanisms generally suffer from three major drawbacks: The binding relationship between templates and file categories is a hard-coded static mapping, which cannot be flexibly adapted according to the actual semantic features of the file content; The lack of feedback loops regarding the effectiveness of template execution means that even if the cataloging results have missing, misaligned, or semantically biased fields, the system cannot autonomously detect and optimize the matching strategy. In scenarios where multi-source heterogeneous archives continue to flood in, the maintenance of the template library relies heavily on expert experience, resulting in long update cycles and high costs, which severely restricts the adaptive evolution capability of the archive management system.

[0006] The aforementioned problems are particularly prominent in scenarios such as high-frequency business changes, dynamic adjustments to regulations, or emergency archiving of sudden events, leading to a decline in the quality of archival recording and an increase in the risk of information loss. There is an urgent need for a dynamic recording template matching and optimization mechanism that can integrate status awareness and autonomous decision-making capabilities. Summary of the Invention

[0007] This invention provides a method for automatic determination of archival categories and matching of cataloging templates based on pattern recognition, aiming to solve the problem that existing archival management systems have static and fixed cataloging templates that cannot adapt to emerging or format-variant archival types. This method achieves automatic identification, classification, and real-time matching of optimal cataloging templates for unknown or variable archival types by constructing a dynamically evolving archival feature extraction model and template matching mechanism, thereby eliminating the need for manual intervention in configuration.

[0008] The present invention provides a method for automatic determination of archive categories and matching of bibliographic templates based on pattern recognition, including: Obtain the original data content and metadata information of the file to be processed; The original data content is subjected to structured parsing and semantic feature extraction to generate a multi-dimensional feature vector of the archive; The multidimensional feature vector of the archive is input into a pre-trained archive category discrimination model, which outputs the category label to which the archive belongs; Based on the category tags, retrieve the bibliographic template with the highest relevance from the dynamic template library; If there is no bibliographic template in the dynamic template library with a matching degree greater than a preset threshold, the automatic template generation process is triggered to construct a new bibliographic template based on the structural features and semantic elements of the current archive and register it in the dynamic template library. The matching or generated cataloging template is applied to the file to be processed to complete the automatic cataloging.

[0009] Preferably, obtaining the original data content and metadata information of the file to be processed includes: Read binary stream data of archive files from the archive storage medium; Parse the header information of the archive file to obtain the file format identifier, creation time, modification time, file size, encoding method, and author information; If the file is a compound document, then recursively parse its nested objects and extract the type identifier and hierarchical path of each sub-object; Synchronously collect the contextual information of the archives in the business system, including the source institution code, business process number, associated case file identifier, and processing stage mark.

[0010] Preferably, the original data content is subjected to structured parsing and semantic feature extraction to generate a multi-dimensional feature vector of the archive, including: For text-based archives, a dictionary- and rule-based named entity recognition method is used to extract key fields such as personal names, place names, organization names, time, amount, and identification number, and to count their frequency of occurrence, location distribution, and contextual co-occurrence relationships. For tabular files, parse their row and column structure, identify the header field names, and extract the cell data type distribution, value range, proportion of empty values, and cross-column association patterns. For image-based archives, visual text content is extracted using optical character recognition technology, and visual semantic features are generated by combining image texture, layout partitioning, and stamp area detection results. For hybrid archives, the above-mentioned analysis results are integrated to construct a unified heterogeneous feature space, in which each dimension corresponds to a semantic element or structural attribute, and a weight coefficient is assigned to reflect its discriminative power in category determination.

[0011] Preferably, the multidimensional feature vector of the archive is input into a pre-trained archive category discrimination model, and the category label to which the archive belongs is output, including: The standardized multidimensional feature vectors of the archives are then input in parallel into a linear kernel support vector machine model and a gradient boosting tree model. Obtain the class probability distribution vectors of the two model outputs respectively; The two probability distribution vectors are linearly combined according to the preset weighted fusion coefficients to obtain the integrated prediction probability distribution; The category label with the highest probability value is selected as the output, and the category label is the leaf node path in the tree hierarchy.

[0012] Preferably, based on the category tag, the most relevant bibliographic template is retrieved from the dynamic template library, including: Map the category labels to template index tree nodes in the template library; Traverse all candidate templates under this node and calculate the structural fit between each candidate template and the current file's multidimensional feature vector; The method for calculating the structural adaptability is as follows: for each cataloging field defined in the template, determine whether the current file contains the semantic element corresponding to the field. If it does, add the preset weight value of the field; otherwise, do not add it. Sum all the accumulated values ​​and divide by the total template weight to obtain the normalized fitness score; The candidate template with the highest fit score that is greater than or equal to the preset threshold is selected as the matching result.

[0013] Preferably, the automatic template generation process is triggered, and a new bibliographic template is constructed based on the structural features and semantic elements of the current archive, including: Cluster center fitting is performed on the multidimensional feature vectors of the current archive to identify the set of semantic elements that appear frequently and stably. Based on the set of semantic elements, a list of bibliographic fields is automatically generated. Each field includes a field name, data type, required field flags, value constraints, and source location rules. The source location rule describes the extraction path of the field content in the original file. For text files, it is a regular expression or context keyword anchor point; for table files, it is row and column coordinates or table header matching rules; for image files, it is OCR recognition area coordinates and post-processing logic. Assign a unique template identifier to the newly generated cataloging template and establish a two-way index relationship between it and the category label of the current archive; Write the new template to the persistent storage area of ​​the dynamic template library and update the template index tree.

[0014] Preferably, the step of performing cluster center fitting on the multidimensional feature vector of the current archive to identify a set of frequently occurring semantic elements includes: From the multidimensional feature vectors of the current archive, select dimensions that are non-zero and have a confidence level greater than the threshold to form an initial set of elements; Query the historical similar archive database to obtain the feature vectors of several archives that have the same or similar category tags as the current archive; Perform single-center K-means clustering on the feature vectors of the aforementioned files, and fit the cluster center vectors; Dimensions whose median value in the cluster center vector is greater than a preset activity threshold are defined as stable semantic elements.

[0015] Preferably, when constructing the new bibliographic template, a field redundancy detection and merging operation is performed, including: Calculate the co-occurrence correlation coefficient of any two elements in the stable semantic element set in historical similar archives; If the correlation coefficient of two elements is greater than the preset high correlation threshold, they will be merged into a composite field. The source location rules for the composite field define multi-source extraction logic, including the main path and the fallback path.

[0016] Preferably, the matching or generated cataloging template is applied to the file to be processed to complete the automatic cataloging, including: Based on the source location rules of each field in the cataloging template, locate and extract the corresponding content in the original archive data; The extracted content is formatted and standardized, including date format standardization, number precision adjustment, unit conversion, and illegal character filtering. The standardized content is populated into a predefined bibliographic data structure to form a structured bibliographic record; The structured catalog records are linked to the original archival documents and stored in the archival management database; Record the log information for this matching or generation operation, including file identifier, matching template identifier, fit score, operation time, and operation type.

[0017] Preferably, the document category discrimination model is continuously optimized through a combination of offline training and online incremental learning, including: The model is reconstructed periodically using full historical labeled data; After each successful generation of a new template, the current file sample and its category label are added to the training buffer pool; When the number of samples in the buffer pool reaches the preset batch threshold, the model fine-tuning process is triggered, updating only the decision boundary parameters in the model related to the newly added category.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves complete automation of archive category determination and template matching by constructing a multi-dimensional feature vector and dynamically evolving cataloging template library for archives.

[0019] 2. For emerging or format-variant archive types, the system can automatically generate suitable bibliographic templates without manual intervention and incorporate them into the knowledge base for subsequent use, completely solving the fundamental defect of the traditional static template mechanism lacking adaptability.

[0020] 3. This invention employs an integrated classification model and a dual verification mechanism for structural fit, ensuring the accuracy of category determination and the reliability of template matching.

[0021] 4. At the same time, through an online incremental learning strategy, the system has the ability to continuously evolve, and can dynamically adjust its recognition and cataloging capabilities as the ecology of archival types evolves.

[0022] 5. The design of field redundancy detection and multi-source location rules improves the generalization and robustness of newly generated templates, effectively avoiding the problem of overfitting a single file instance.

[0023] 6. The overall solution significantly reduces the labor costs of archive management, improves cataloging efficiency and data consistency, and provides a solid technical foundation for the intelligent processing of large-scale digital archives. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the multidimensional feature vector generation and dynamic template matching mechanism for archives in this invention. Figure 3 This is a logical flowchart of the data acquisition and multimodal structured parsing of the files to be processed in this invention; Figure 4 This is a logical framework diagram of the integrated learning and hierarchical label output of the archive category discrimination model in this invention; Figure 5 This is a flowchart illustrating the logical flow of dynamic template library retrieval, fit calculation, and automatic generation of new templates in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal file processing system and the dynamic template library in this invention. Detailed Implementation

[0025] Please refer to Figures 1 to 6 This invention provides a method for automatic determination of archival categories and matching of cataloging templates based on pattern recognition. Its core objective is to address the problem in traditional archival management systems where cataloging templates are statically fixed and unable to adapt to emerging or format-variant archival types. By constructing a dynamically evolving archival feature extraction model and template matching mechanism, the system can perform structured parsing, semantic feature extraction, and category discrimination on raw data of unknown or variable archival types. It can also achieve real-time matching or automatic generation of the optimal cataloging template from a dynamic template library, thereby completely eliminating the need for manual intervention in configuration.

[0026] The method includes the following steps: acquiring the original data content and metadata information of the archive to be processed; performing structured parsing and semantic feature extraction on the original data content to generate a multi-dimensional feature vector of the archive; inputting the multi-dimensional feature vector of the archive into a pre-trained archive category discrimination model and outputting the category label to which the archive belongs; based on the category label, searching the dynamic template library for the cataloging template with the highest correlation; if there is no cataloging template with a matching degree greater than a preset threshold in the dynamic template library, triggering the automatic template generation process, constructing a new cataloging template based on the structural features and semantic elements of the current archive, and registering it in the dynamic template library; applying the matched or generated cataloging template to the archive to be processed to complete the automatic cataloging.

[0027] First, step S1 is executed to obtain the original data content and metadata information of the archive to be processed. This step specifically includes reading the binary stream data of the archive file from the archive storage medium; parsing the header information of the archive file to obtain the file format identifier, creation time, modification time, file size, encoding method, and author information; if the archive is a composite document, recursively parsing its internal nested objects to extract the type identifier and hierarchical path of each sub-object; and synchronously collecting the archive's context information in the business system, including the source institution code, business process number, associated case file identifier, and processing stage marker. During this step, the system first accesses the archive storage location through the standard file system interface and calls underlying I / O operations to read the complete byte sequence.

[0028] Subsequently, based on internationally recognized file format standards such as PDF, DOCX, XLSX, and JPEG, the MagicNumber field in the file header is parsed to determine its primary type. For compound document types, the system initiates a recursive parser, traversing its internal object tree layer by layer, recording the MIME type, relative path, and nesting depth of each sub-object.

[0029] Simultaneously, the system interfaces with the business management system's API to obtain the file's contextual metadata within the business process, including but not limited to the source institution's unified social credit code, the current approval stage number, the unique identifier of the case file, and processing status markers such as "archived" and "pending review." All of the above information is uniformly organized into a structured metadata container, serving as the contextual basis for subsequent feature extraction.

[0030] Next, step S2 is executed to perform structured parsing and semantic feature extraction on the original data content, generating a multi-dimensional feature vector for the archive. This step specifically includes using a dictionary-based and rule-based named entity recognition method for text-type archives to extract key fields such as personal names, place names, organization names, time, amount, and identification number, and to statistically analyze their frequency of occurrence, location distribution, and contextual co-occurrence relationships. For tabular files, parse their row and column structure, identify the header field names, and extract the cell data type distribution, value range, proportion of empty values, and cross-column association patterns. For image-based archives, visual text content is extracted using optical character recognition technology, and visual semantic features are generated by combining image texture, layout partitioning, and stamp area detection results. For hybrid archives, the above-mentioned analysis results are integrated to construct a unified heterogeneous feature space, in which each dimension corresponds to a semantic element or structural attribute, and a weight coefficient is assigned to reflect its discriminative power in category determination.

[0031] In practice, the system first schedules the corresponding parsing engine based on the file type identified in step S1. For plain text or rich text documents, the system loads a pre-built domain dictionary, including administrative division directories, legal entity directories, historical figures databases, and financial terminology tables, and performs named entity recognition in conjunction with a set of regular expression rules. The recognition results are further processed into feature items, such as the frequency of occurrence of the "ID number" field, the numerical value of the "contract amount" field, and its Boolean flag appearing in the first 20% of the document.

[0032] For tabular documents, the system calls the table structure recognition module to reconstruct the logical table using a row and column segmentation algorithm. It extracts the field names from the header row and infers the data type of each column, including integers, floating-point numbers, dates, and strings. Simultaneously, it calculates the null value rate, extreme value range, and correlation coefficient with other columns for each column.

[0033] For image files, the system first performs image preprocessing, including noise reduction, binarization, and skew correction, and then performs layout analysis to divide the file into text areas, table areas, image areas, and stamp areas.

[0034] High-precision OCR recognition is performed in the text area to output the text content; circular detection and character recognition are performed in the seal area to determine whether it contains a legally valid seal.

[0035] For audio-visual files, the system first splits the file type into pure audio (such as WAV, MP3) or video (such as MP4, AVI).

[0036] For pure audio files, audio preprocessing (noise reduction, volume normalization, and channel separation) is performed. Then, a high-precision speech recognition (ASR) engine is called to convert the speech content into a text transcript and mark the timestamp of each sentence (start millisecond - end millisecond). At the same time, audio acoustic features are extracted, including fundamental frequency, spectral entropy, speech rate (words / second), sentiment coefficient (outputting positive / neutral / negative probabilities through a speech sentiment recognition model), and background noise intensity level.

[0037] For video files, the system uses a video decoding module to separate the video frame sequence (sampled at 10 frames per second) and an independent audio track: the audio track extracts speech text and acoustic features according to pure audio processing logic; the video frame sequence first performs scene segmentation (based on inter-frame difference), and then performs image parsing on key frames (scene start frame, screen change frame) (similar to layout analysis, OCR recognition, and stamp / object detection for image-type files), while extracting subtitle text (through subtitle track parsing or intra-frame text OCR), human face detection results (quantity, location) and scene labels (such as "meeting scene" and "on-site investigation scene").

[0038] All parsing results of audio-visual archives (speech text, acoustic features, image features, subtitle text, and timing information) are mapped to dedicated dimensions of the feature space, including "audio duration", "speech text length", "average sentiment coefficient", "number of video keyframes", and "whether it contains meeting scene tags", etc., and are incorporated into a multi-dimensional feature vector after being assigned corresponding weights.

[0039] The parsing results of all modalities are mapped to a predefined feature dimension space, which contains more than 200 fixed dimensions. Each dimension corresponds to a specific semantic or structural attribute, such as "whether it contains a court seal", "number of contract signatories", "number of table columns greater than 10", etc. The value of each dimension is a real number or a Boolean value, and it is accompanied by an initial weight calculated from historical training data, which is used as input for subsequent classification models.

[0040] Then, step S3 is executed, in which the multidimensional feature vector of the archive is input into the pre-trained archive category discrimination model, and the category label to which the archive belongs is output.

[0041] The document category discrimination model is an ensemble classifier based on support vector machines and gradient boosting trees. Its training process includes collecting a historically labeled document sample set, with each sample containing a complete multi-dimensional feature vector and a manually confirmed category label; stratified sampling of the sample set, dividing it into a training set, a validation set, and a test set; training a linear kernel support vector machine model and a gradient boosting tree model on the training set respectively; evaluating the classification accuracy, recall, and F1 score of the two models on the validation set to determine the weighted fusion coefficient; and linearly combining the output probabilities of the two models according to the weighted fusion coefficient to form the final category prediction result. The category label system adopts a tree-like hierarchical structure, with the top level representing broad categories such as administrative documents, judicial files, engineering drawings, and research reports, and the lower levels progressively refining to specific business subcategories.

[0042] When performing this step, the system normalizes the feature vectors generated in step S2 and inputs them in parallel into two independent sub-models.

[0043] The Support Vector Machine (SVM) sub-model employs a linear kernel function, and its decision hyperplane exhibits strong discriminative power for linearly separable patterns in a high-dimensional feature space. The Gradient Boosting Tree (GRB) sub-model constructs a decision tree forest through multiple iterations, excelling at capturing nonlinear interactions and local patterns between features. Each sub-model outputs a probability distribution vector, the dimension of which is equal to the total number of leaf nodes in the category label system.

[0044] The system performs a weighted summation of the two probability vectors based on the fusion coefficients α and β determined during the offline training phase to obtain the final ensemble prediction probability. Subsequently, the system selects the category label with the highest probability value as the output. This label is a hierarchical path string, such as "judicial case file / civil case / first instance judgment," and its structure strictly follows a predefined category tree.

[0045] Next, step S4 is executed, whereby, based on the category label, the most relevant bibliographic template is retrieved from the dynamic template library. This step specifically includes mapping the category label to a template index tree node in the template library; traversing all candidate templates under that node; and calculating the structural fit between each candidate template and the current archive's multidimensional feature vector. The method for calculating the structural fit is as follows: For each cataloging field defined in the template, determine whether the current archive contains the semantic element corresponding to that field. If it does, add the preset weight value of that field; otherwise, do not add it. Sum all the accumulated values ​​and divide by the total template weight to obtain the normalized fitness score; The candidate template with the highest fit score that is greater than or equal to the preset threshold is selected as the matching result.

[0046] In the actual retrieval process, the system first decomposes the complete category path output in step S3 into a sequence of nodes from root to leaf. The template library internally maintains an index tree isomorphic to the category tree, with each node storing a list of pointers to all documented templates under that category. After locating a leaf node, the system retrieves all its candidate templates. For each candidate template, the system iterates through its list of field definitions. Each field definition includes the field name, data type, required field flags, and preset weight values.

[0047] The system checks if the multidimensional feature vector of the current file contains a dimension that semantically corresponds to the field and whose value is valid. If it does, the weight value of that field is added to the sum. After traversing all fields, the system divides the sum by the sum of the weights of all fields in the template to obtain a normalized fit score, which ranges from 0 to 1. The system records the scores of all candidate templates and selects the one with the highest score. If the highest score is greater than or equal to a preset threshold, such as 0.85, the match is considered successful, and the process proceeds to the next step; otherwise, the automatic template generation process begins.

[0048] If no matching template is found, proceed to step S5 to trigger the automatic template generation process, constructing a new bibliographic template based on the structural features and semantic elements of the current archive. This step specifically includes fitting cluster centers to the multidimensional feature vectors of the current archive to identify a set of frequently occurring and stable semantic elements. Based on the semantic element set, a list of bibliographic fields is automatically generated. Each field includes a field name, data type, required field flags, value constraints, and source location rules. The source location rules describe the extraction path of the field's content in the original file. For text files, these are regular expressions or contextual keyword anchors; for table files, they are row and column coordinates or table header matching rules; and for image files, they are OCR recognition area coordinates and post-processing logic. A unique template identifier is assigned to the newly generated bibliographic template, and a bidirectional index relationship is established between it and the category label of the current file. The new template is written to the persistent storage area of ​​the dynamic template library, and the template index tree is updated.

[0049] In this step, the system first filters out all non-zero dimensions with a confidence level greater than a threshold from the feature vectors, forming an initial feature set. Then, the system queries a historical similar archive database to obtain the feature vectors of the most recent archives with the same or similar category labels as the current archive. These vectors are then subjected to K-means clustering, with K set to 1, meaning a single cluster center is fitted. Dimensions with median values ​​greater than a preset activity threshold in the center vector are identified as stable semantic elements within that category.

[0050] Based on this stable set of elements, the system generates bibliographic fields for each element. The field names are automatically generated from the element semantics, such as "Plaintiff's Name", "Project Number"; the data types are inferred from the instances of the elements in the current file; the mandatory flag is set to true; the value constraints are deduced from the format of the instance values, for example, the date format is YYYY-MM-DD.

[0051] The generation of the source location rules depends on the parsing context in step S2. For text elements, the system records their offset positions in the original text and the surrounding anchor keywords to generate context-based extraction rules; for table elements, it records the serial number, row number, column number or header name of the table where they are located; for image elements, it records the rectangular area coordinates during OCR recognition.

[0052] For audio-visual semantic elements, the source location rules include temporal constraints and multi-source verification logic: the location rules for audio fields specify the timestamp interval of speech recognition (such as "the contract amount field corresponds to the audio interval from 1 minute 23 seconds to 1 minute 35 seconds"), keyword anchor points (such as "based on the content within 50 words after 'total contract amount'"), and emotional feature thresholds (such as "important clauses need to meet an emotional coefficient of neutral or above"); the location rules for video fields include the range of video frame numbers (such as "the seal area in frames 10-20"), subtitle track identifiers (such as "text in the 3rd to 5th of subtitle track 2"), and audio track association rules (such as "take effect when the text 'date of signature' in the video is consistent with the speech content at 1 minute 10 seconds of the audio"). For audio-visual composite fields (such as "meeting resolution content"), the source location rules include the main path (keyword matching of the speech transcript) and the fallback path (supplementation of video subtitle text, extraction of text from key frames), ensuring the integrity of the extraction of field content.

[0053] After all field definitions are completed, the system generates a globally unique 16-digit digital template identifier, serializes the new template into a JSON format object, and writes it into persistent storage through the distributed database interface. At the same time, the system updates the template index tree and adds a reference to this template under the corresponding category node.

[0054] Finally, step S6 is executed to apply the matched or generated bibliographic template to the said to-be-processed file to complete automatic bibliographic recording. This step specifically includes locating and extracting the corresponding content in the original file data according to the source location rules of each field in the bibliographic template; performing format standardization processing on the extracted content, including unifying the date format, adjusting the digital precision, unit conversion, and filtering illegal characters; The standardized content is populated into a predefined cataloging data structure to form a structured cataloging record; the structured cataloging record is associated with the original archival file and stored in the archival management database; the log information of this matching or generation operation is recorded, including the archival identifier, matching template identifier, fit score, operation time and operation type.

[0055] When performing this step, the system iterates through the field list of the cataloging template and, for each field, calls the extractor specified by its source location rule.

[0056] The text extractor performs regular expression matching or keyword anchoring search; the table extractor accesses specified rows, columns, or headers; the image extractor crops specified areas and performs OCR.

[0057] All extracted raw content is fed into a format standardization engine. This engine has multiple built-in conversion rules: The date string is parsed into a standard timestamp and reformatted to ISO8601; Amounts are uniformly retained to two decimal places; length units such as "meter" and "centimeter" are converted to basic units; All content has been Unicode normalized and controlled for character filtering.

[0058] Standardization is implemented for audio-visual content extraction: speech-to-text conversion must be segmented according to the "Chinese Text Segmentation Standard," polyphonic characters must be corrected through contextual semantics, and dialects or professional terms must be uniformly expressed according to national standard expressions; watermark characters and advertising texts are removed from video subtitles, and the font encoding is uniformly UTF-8, organized according to the "paragraph-sentence" structure; audio-visual temporal information is standardized to ISO8601 format, and keyframe numbers are uniformly named according to the "file ID-frame number" format; acoustic feature data is uniformly retained to 3 decimal places, and sentiment coefficients are classified into "neutral (0.3-0.7), positive (>0.7), negative (<0.3)" as labeled data to ensure that the audio-visual recorded content is consistent with the data structure of other types of archives, which facilitates subsequent retrieval and analysis.

[0059] The standardized content is populated into a predefined XML or JSON cataloging structure. This structure is fully compatible with the data model of the records management system. The system then binds this cataloging record to the unique identifier of the original archive and writes it to the central archive via transactional database operations. Simultaneously, a detailed operation log is appended to the audit log file, including the archive hash value, template identifier, fit score (if a match), operation timestamp, and operation type (match or generate).

[0060] Throughout the entire process, the system also incorporates an online incremental learning mechanism. Whenever a new template is successfully generated in step S5, the feature vector of the current file and its new class label (which may be an existing class or a newly extended subclass) are added to the training buffer pool.

[0061] When the number of samples in the buffer pool reaches a preset batch threshold, such as 100, the system triggers a model fine-tuning process. This process uses only new samples from the buffer pool to update the parameters of the ensemble classifier in step S3. The Support Vector Machine model adjusts its support vector set through an incremental learning algorithm; the Gradient Boosting Tree model adds several trees to the existing forest, specifically for distinguishing new categories. This process ensures that the model incorporates new knowledge without compromising its ability to distinguish existing categories.

[0062] In addition, during the template generation process in step S5, the system performs field redundancy detection. This detection calculates the co-occurrence correlation coefficient of any two features in the current stable feature set within similar archives by querying the historical archive feature database.

[0063] If the correlation coefficient is greater than a preset high correlation threshold, such as 0.95, the system will merge these two elements into a composite field. For example, if "Contract Signing Date" and "Effective Date" are highly correlated, they will be merged into "Contract Date," whose source location rules include two extraction paths, corresponding to two locations in the original text. The system prioritizes the first path during cataloging; if it fails, it will fall back to the second path. This design significantly improves the generalization ability of the new template, avoiding the generation of overly specialized fields due to the accidental structure of a single instance.

[0064] In summary, this invention constructs a closed-loop, adaptive, and evolvable intelligent archival cataloging system through the coordinated operation of the aforementioned six core steps. This method not only achieves high-precision automatic cataloging of known archival types but also possesses the ability to autonomously recognize and construct templates for unknown or variant archival types, fundamentally solving the adaptability deficiencies of traditional static template mechanisms.

[0065] At the system level, the present invention also provides an automatic archival category determination and cataloging template matching system based on pattern recognition, which includes an archival data acquisition module, a multi-dimensional feature generation module, a category discrimination module, a template matching module, a template generation module, and an automatic cataloging execution module.

[0066] The archive data acquisition module is responsible for executing step S1. It integrates multiple file protocol parsers and business system adapters to ensure the integrity of metadata collection.

[0067] The multi-dimensional feature generation module is responsible for executing step S2, and it further includes a text parsing unit, a table parsing unit, an image parsing unit, and a feature fusion unit. The text parsing unit is configured to perform named entity recognition and contextual semantic analysis on plain text or document body; the table parsing unit is configured to recognize the logical structure of spreadsheets or scanned tables and extract field mapping relationships; the image parsing unit is configured to perform image preprocessing, layout analysis, OCR recognition, and seal detection; the feature fusion unit is configured to map the output results of each parsing unit to a unified feature space and assign different dimensional weights according to the importance of the elements.

[0068] The category discrimination module is responsible for executing step S3. Its archive category discrimination model is continuously optimized through a combination of offline training and online incremental learning. Offline training periodically reconstructs the model using all historical labeled data. Online incremental learning adds the current archive sample and its new category label to the training buffer after each successful generation of a new template. When the number of samples in the buffer reaches a preset batch threshold, a model fine-tuning process is triggered, updating only the decision boundary parameters related to the newly added category, thus maintaining the stability of the existing category discrimination capability.

[0069] The template matching module is responsible for executing step S4 and is tightly coupled with the dynamic template library. The dynamic template library is stored in a distributed key-value database, where the key is a hash code of the category label and the value is a serialized object of all bibliographic templates under that category. Each bibliographic template object contains a template identifier, a list of field definitions, a total weight value, a creation timestamp, and a usage frequency counter. During retrieval, the template matching module first quickly locates the candidate template set using the category label, and then performs a secondary sorting and filtering based on structural fit.

[0070] The template generation module is responsible for executing step S5. When building a new catalog template, it performs field redundancy detection and merging operations. Field redundancy detection is achieved by calculating the co-occurrence correlation coefficient of different semantic elements in historical similar archives. If the correlation coefficient of two elements is greater than the preset high correlation threshold, they are merged into a composite field, and multi-source extraction logic is defined in the source positioning rules.

[0071] The automatic cataloging execution module is responsible for executing the S6 steps. Its built-in format standardization engine and database transaction controller ensure the accuracy and consistency of the cataloged data.

[0072] The entire system is deployed using a microservice architecture, with modules communicating asynchronously via message queues to ensure high concurrency and fault isolation. The dynamic template library is implemented using a distributed NoSQL database, supporting horizontal scaling and high-availability read / write operations. All operations are logged with complete audit logs to meet compliance requirements for record management.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatic determination of archival categories and matching of bibliographic templates based on pattern recognition, characterized in that, include: Obtain the original data content and metadata information of the file to be processed; The original data content is subjected to structured parsing and semantic feature extraction to generate a multi-dimensional feature vector of the archive; The multidimensional feature vector of the archive is input into a pre-trained archive category discrimination model, which outputs the category label to which the archive belongs; Based on the category tags, retrieve the bibliographic template with the highest relevance from the dynamic template library; If there is no bibliographic template in the dynamic template library with a matching degree greater than a preset threshold, the automatic template generation process is triggered to construct a new bibliographic template based on the structural features and semantic elements of the current archive and register it in the dynamic template library. The matching or generated cataloging template is applied to the file to be processed to complete the automatic cataloging.

2. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 1, characterized in that, The process of obtaining the original data content and metadata information of the file to be processed includes: Read binary stream data of archive files from the archive storage medium; Parse the header information of the archive file to obtain the file format identifier, creation time, modification time, file size, encoding method, and author information; If the file is a compound document, then recursively parse its nested objects and extract the type identifier and hierarchical path of each sub-object; Synchronously collect the contextual information of the archives in the business system, including the source institution code, business process number, associated case file identifier, and processing stage mark.

3. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 2, characterized in that, The original data content is subjected to structured parsing and semantic feature extraction to generate a multi-dimensional feature vector of the archive, including: For tabular files, parse their row and column structure, identify the header field names, and extract the cell data type distribution, value range, proportion of empty values, and cross-column association patterns. For audio-visual archives, speech text content in the audio is extracted using speech recognition technology, and audio semantic features are generated by combining audio features; for video files, the video frame sequence and audio track are split, visual features of the video frames are extracted using image parsing logic, and semantic features of the audio track are extracted using the above audio processing logic, and audio-visual features are fused to generate comprehensive audio-visual semantic features. For image-based archives, visual text content is extracted using optical character recognition technology, and visual semantic features are generated by combining image texture, layout partitioning, and stamp area detection results. For hybrid archives, the above-mentioned analysis results are integrated to construct a unified heterogeneous feature space, in which each dimension corresponds to a semantic element or structural attribute, and a weight coefficient is assigned to reflect its discriminative power in category determination.

4. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 3, characterized in that, The multidimensional feature vector of the archive is input into a pre-trained archive category discrimination model, which outputs the category label to which the archive belongs, including: The standardized multidimensional feature vectors of the archives are then input in parallel into a linear kernel support vector machine model and a gradient boosting tree model. Obtain the class probability distribution vectors of the two model outputs respectively; The two probability distribution vectors are linearly combined according to the preset weighted fusion coefficients to obtain the integrated prediction probability distribution; The category label with the highest probability value is selected as the output, and the category label is the leaf node path in the tree hierarchy.

5. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 4, characterized in that, Based on the category tags, the most relevant bibliographic templates are retrieved from the dynamic template library, including: Map the category labels to template index tree nodes in the template library; Traverse all candidate templates under this node and calculate the structural fit between each candidate template and the current file's multidimensional feature vector; The method for calculating the structural adaptability is as follows: for each cataloging field defined in the template, determine whether the current file contains the semantic element corresponding to the field. If it does, add the preset weight value of the field; otherwise, do not add it. Sum all the accumulated values ​​and divide by the total template weight to obtain the normalized fitness score; The candidate template with the highest fit score that is greater than or equal to the preset threshold is selected as the matching result.

6. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 5, characterized in that, Trigger the automatic template generation process to construct a new bibliographic template based on the structural features and semantic elements of the current archive, including: Cluster center fitting is performed on the multidimensional feature vectors of the current archive to identify the set of semantic elements that appear frequently and stably. Based on the set of semantic elements, a list of bibliographic fields is automatically generated. Each field includes a field name, data type, required field flags, value constraints, and source location rules. The source location rule describes the extraction path of the field content in the original file. For text files, it is a regular expression or context keyword anchor point; for table files, it is row and column coordinates or table header matching rules; for image files, it is OCR recognition area coordinates and post-processing logic. For audio files, the parameters include speech recognition timestamp range, keyword trigger threshold, and emotion feature extraction algorithm parameters; for video files, the parameters include video frame sequence range, subtitle track identifier, audio track timestamp range, and image feature matching rules. Assign a unique template identifier to the newly generated cataloging template and establish a two-way index relationship between it and the category label of the current archive; Write the new template to the persistent storage area of ​​the dynamic template library and update the template index tree.

7. The method for automatic determination of archive categories and matching of cataloging templates based on pattern recognition according to claim 6, characterized in that, The process of performing cluster center fitting on the multidimensional feature vectors of the current archive to identify a set of frequently occurring and stable semantic elements includes: From the multidimensional feature vectors of the current archive, select dimensions that are non-zero and have a confidence level greater than the threshold to form an initial set of elements; Query the historical similar archive database to obtain the feature vectors of several archives that have the same or similar category tags as the current archive; Perform single-center K-means clustering on the feature vectors of the aforementioned files, and fit the cluster center vectors; Dimensions whose median value in the cluster center vector is greater than a preset activity threshold are defined as stable semantic elements.

8. The method for automatic determination of archival categories and matching of cataloging templates based on pattern recognition according to claim 7, characterized in that, When constructing the new bibliographic template, field redundancy detection and merging operations are performed, including: Calculate the co-occurrence correlation coefficient of any two elements in the stable semantic element set in historical similar archives; If the correlation coefficient of two elements is greater than the preset high correlation threshold, they will be merged into a composite field. The source location rules for the composite field define multi-source extraction logic, including the main path and the fallback path.

9. The method for automatic determination of archival categories and matching of cataloging templates based on pattern recognition according to claim 8, characterized in that, Applying the matched or generated cataloging template to the file to be processed completes the automatic cataloging, including: Based on the source location rules of each field in the cataloging template, locate and extract the corresponding content in the original archive data; The extracted content is formatted and standardized, including date format standardization, number precision adjustment, unit conversion, and illegal character filtering. Specialized standardization processing is carried out on the extracted audio and video content, including standardization of sentence segmentation for speech-to-text conversion, correction of polyphonic characters, standardization of dialects / professional terms, format cleaning of video subtitles, and temporal alignment of extracted frame content; The standardized content is populated into a predefined bibliographic data structure to form a structured bibliographic record; The structured catalog records are linked to the original archival documents and stored in the archival management database; Record the log information for this matching or generation operation, including file identifier, matching template identifier, fit score, operation time, and operation type.

10. The method for automatic determination of archival categories and matching of cataloging templates based on pattern recognition according to claim 9, characterized in that, The document category discrimination model is continuously optimized through a combination of offline training and online incremental learning, including: The model is reconstructed periodically using full historical labeled data; After each successful generation of a new template, the current file sample and its category label are added to the training buffer pool; When the number of samples in the buffer pool reaches the preset batch threshold, the model fine-tuning process is triggered, updating only the decision boundary parameters in the model related to the newly added category.

Images