Intelligent extraction and indexing system for file metadata

Abstract

The invention relates to the technical field of archive information management, and discloses an archive metadata intelligent extraction and indexing system, which comprises a multi-modal preprocessing module for obtaining and preprocessing original multi-modal archive data; the context entity recognition module is used for performing entity recognition and standardization according to the context vector; the cross-modal fusion module is used for carrying out confidence weighted multi-modal fusion and logic verification; the archive association module is used for carrying out association identification and consistency detection between archives; the intelligent indexing module is used for carrying out hierarchical intelligent indexing and quality feedback on the consistency constrained file metadata set; the quality evaluation module is used for carrying out metadata quality evaluation and active repair on the standardized indexing result; the knowledge graph module is used for constructing a time sequence knowledge graph and intelligent retrieval service; according to the method, collaborative extraction of cross-modal information is realized by constructing a confidence-weighted multi-modal fusion model and a bidirectional attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of archival information management technology, and more specifically, to an intelligent extraction and indexing system for archival metadata. Background Technology

[0002] With the deepening implementation of the national informatization strategy, archival institutions at all levels have accumulated a massive amount of digital archival resources. Archival metadata, as structured data describing the content, characteristics, and management information of archives, is a crucial foundation for the effective organization, retrieval, and utilization of archival resources. However, traditional archival metadata extraction and indexing work mainly relies on manual work by archival staff. Staff need to examine the content of each archive one by one, filling in more than 20 metadata fields such as file number, title, responsible person, creation time, and keywords according to archival cataloging rules. This manual indexing model suffers from inefficiency, high costs, and inconsistent standards. On average, each archivist can only complete the metadata entry for 50-80 archives per day, which is insufficient to meet the urgent needs of large-scale archival digitization.

[0003] In recent years, artificial intelligence technology has made breakthroughs in fields such as natural language processing, computer vision, and knowledge graphs, providing new technical approaches for the intelligent extraction and indexing of archival metadata. Existing technologies have been used in archival classification, text recognition, and information retrieval, such as archival classification methods based on LDA topic models and deep learning, text understanding technologies based on BERT pre-trained models, and ontology-based archival semantic retrieval systems. While these technologies have achieved certain results in their respective fields, shortcomings remain in the practical application of intelligent extraction and indexing of archival metadata.

[0004] The main shortcomings of existing technologies include: limitations of single-modal processing (current methods are mainly designed for plain text archives and cannot effectively handle real archives containing mixed media such as text, images, handwritten annotations, and seals), and a lack of collaborative extraction and semantic-level fusion mechanisms for cross-modal information; insufficient historical context awareness (existing named entity recognition technologies lack an understanding of the historical context in which the archives were formed, resulting in high error rates and low standardization); lack of consistency assurance in batch processing (when processing a series of archives, existing technologies lack constraints and detection mechanisms for the consistency of metadata in related archives, leading to inconsistent indexing standards within the same series of archives); and limited metadata quality control methods (lacking a systematic quality assessment and intelligent completion mechanism, often requiring extensive manual review of extraction results). These technical problems severely restrict the practical application of intelligent extraction and indexing technologies for archival metadata. Summary of the Invention

[0005] This invention provides an intelligent extraction and indexing system for archival metadata, which solves the technical problems of mixed-media archive processing, historical context recognition, and batch consistency assurance in related technologies.

[0006] This invention provides an intelligent extraction and indexing system for archival metadata, comprising: The multimodal preprocessing module is used to acquire raw multimodal archive data, preprocess the raw multimodal archive data, and obtain a structured multimodal archive data set. The context entity recognition module extracts context vectors from a structured multimodal archive dataset, performs entity recognition and normalization based on the context vectors, and obtains archive text data with labeled and normalized entities. The cross-modal fusion module performs confidence-weighted multimodal fusion and logical verification based on the archive text data with labeled normalized entities and the structured multimodal archive data set to obtain complete archive metadata records; The archive association module, based on complete archive metadata records, performs association identification and consistency detection between archives to obtain a set of archive metadata with consistency constraints; The intelligent indexing module performs hierarchical intelligent indexing and quality feedback on the metadata set of archives with consistency constraints, and obtains standardized indexing results; The quality assessment module performs metadata quality assessment and proactive repair on the standardized indexing results, resulting in an optimized set of metadata records. The knowledge graph module constructs a time-series knowledge graph and intelligent retrieval service based on optimized metadata records, resulting in a dynamically evolving time-series archive knowledge graph and semantic retrieval scheme.

[0007] In a preferred embodiment, the preprocessing of the raw multimodal archive data includes: The original scanned images of the archives are quality inspected. The Laplacian operator is used to calculate the image sharpness. The difference between the images before and after median filtering is used to assess the noise level. The Hough line detection algorithm is used to identify the image edges and calculate the tilt angle. Adaptive enhancement processing is applied to images with quality scores below a preset threshold, including sharpness enhancement, noise reduction, rotation correction, and contrast enhancement. An end-to-end OCR recognition system integrating visual features and language models is used for text extraction. Convolutional neural networks are used to extract visual features of images, bidirectional long short-term memory networks are used to perform sequence modeling of feature maps, connection-time classification algorithms are used for decoding, and an N-gram language model is introduced for post-processing of recognition results. Characters with a confidence level below the first threshold are marked as low-confidence characters. When the proportion of low-confidence characters exceeds the second threshold, the entire text is marked as high-uncertainty text.

[0008] In a preferred embodiment, the step of extracting context vectors based on a structured multimodal archive dataset and performing entity recognition and normalization based on the context vectors includes: Contextual elements are extracted from a structured multimodal archive dataset, and each contextual element is encoded into an embedding vector and concatenated to form a comprehensive contextual vector. The text is encoded, and the context vector is mapped to the same dimension as the token embedding through a linear transformation. The context vector is then added to the embedding of each token to obtain the context-enhanced token representation. Retrieve relevant historical entities from a multi-period archive knowledge graph based on context vectors to form a candidate entity knowledge set; A knowledge-enhanced sequence labeling model is constructed for named entity recognition. The BiLSTM-CRF architecture is adopted, and external knowledge features are introduced to enhance the model's capabilities. The identified entities are normalized across time periods, and the candidate entity with the highest score is selected as the link result by combining the scores of name similarity, semantic similarity and temporal consistency.

[0009] In a preferred embodiment, the confidence-weighted multimodal fusion and logical verification includes: A multimodal fusion model is constructed to process information from both text and image modalities simultaneously, and the fusion weights are dynamically adjusted based on the confidence level of each modal information. A bidirectional attention mechanism is used to identify semantic associations between different modalities, calculate text-to-image attention and image-to-text attention, and perform weighted aggregation of image features based on attention weights; A joint entity relation extraction model is constructed to identify entities and the relations between them. An extended annotation system is used to transform the relation extraction task into a sequence labeling task, and the relation results in the form of triples are output. Based on the predefined archive metadata template, the relation results in the form of triples are mapped to metadata fields, and the mapping rules from triple relation types to metadata fields are established, and field values ​​that do not conform to the specifications are automatically adjusted. A logical consistency inference engine is built to perform multi-dimensional verification of metadata records, including time consistency verification, job title consistency verification, and geographical consistency verification.

[0010] In a preferred embodiment, the process of identifying and verifying the associations and consistency between the files includes: Construct a mechanism for identifying relationships between files, identify explicit and implicit relationships, and calculate the total relationship score by combining explicit and implicit relationships; A heterogeneous graph neural network is constructed to learn the representation of the archive association graph. Different types of relationships are modeled separately. The node representation is updated layer by layer by stacking multiple HGNNs. Each archive node learns an embedding vector that integrates information from multiple types of relationships. For the file pairs connected in the file association graph, check the consistency of their metadata. For each conflict field, calculate the support of each candidate value. The support is the sum of the weighted confidence of the number of times the candidate value appears in all related files. Output a consistency conflict report. Multiple strategies are employed for conflict resolution, including majority voting, confidence ranking, historical accuracy, and knowledge graph verification. The weighted score selection criteria of the four strategies are integrated. The constraint network is constructed using a message passing mechanism for constraint propagation.

[0011] In a preferred embodiment, the hierarchical intelligent indexing and quality feedback of the archive metadata set with consistency constraints includes: A hierarchical classification model is constructed to classify archives by subject, and the classification system is modeled as a tree structure, with hierarchical constraints introduced to ensure its accuracy. A keyword extraction method guided by knowledge graphs is adopted, which combines TF-IDF scores and knowledge graph relevance, while also considering the relevance of keywords to archive categories; The extracted candidate tags are mapped to the normative words in the thesaurus. The matching methods include exact matching, synonym matching, similar word matching, hypernym expansion, and related word supplementation. A reinforcement learning framework is established to continuously optimize the indexing policy through user feedback on the indexing results. The indexing process is modeled as a Markov decision process, and the policy gradient algorithm is used to update the policy parameters. Establish an indexing quality monitoring mechanism to statistically analyze the indexing acceptance rate of each type of archive, the frequency of use of each subject tag, and the accuracy rate.

[0012] In a preferred embodiment, the metadata quality assessment and proactive repair of the normalized indexing results includes: The quality assessment adopts a five-dimensional indicator system, including completeness assessment, accuracy assessment, consistency assessment, standardization assessment, and timeliness assessment. The overall quality score of the metadata record is calculated by combining the evaluation results of the five dimensions, and a field-level quality score is calculated for each field. A proactive repair mechanism is employed to automatically correct problems identified during quality indicator assessments. Metadata records are classified according to their overall quality score, and a quality classification label and processing recommendations are attached to each metadata record.

[0013] In a preferred embodiment, the construction of a time-series knowledge graph and intelligent retrieval service based on optimized metadata records includes: Knowledge graphs are incrementally constructed by extracting entity and relationship information from optimized metadata records, and an incremental update mechanism is used to avoid repeatedly constructing existing parts. To add a time dimension to the knowledge graph, a temporal knowledge graph is constructed. The temporal triple is represented by a quintuple structure that includes a head entity, a relation, a tail entity, the start time of the relation, and the end time of the relation. An embedding-based completion method is used to predict potential relationships and discover implicit file associations, and time coding is introduced to process temporal triples; Graph neural networks are used for embedding learning, and multi-layer GCN stacking enables nodes to embed and fuse information from multi-hop neighbors. Perform query understanding and identification of entities and intents in the query, and perform query expansion including synonym expansion, hyponym expansion, and related word expansion; The graph query employs a hybrid retrieval strategy that balances precise matching and semantic similarity, and calculates a comprehensive score by combining structured matching and semantic similarity. Build a learning ranking model, collect user interaction data, and continuously optimize the retrieval ranking strategy.

[0014] In a preferred embodiment, classifying metadata records based on a comprehensive quality score includes: Metadata records with a comprehensive quality score greater than or equal to the first score threshold are classified as high-quality records and can be directly stored in the database without manual review. Metadata records with a comprehensive quality score between the second and first score thresholds are classified as medium quality levels. After intelligent repair, they are stored in the database and manually sampled. Metadata records with an overall quality score less than or equal to the second score threshold are classified as low-quality records and are manually reviewed before being stored in the database.

[0015] In a preferred embodiment, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of an intelligent extraction and indexing system for archive metadata as described above.

[0016] The beneficial effects of this invention are as follows: By constructing a confidence-weighted multimodal fusion TranSformer model and a bidirectional attention mechanism, it can simultaneously process multiple modalities of information in archives, such as text, images, handwritten annotations, and seals, and identify semantic relationships between different modalities, achieving collaborative extraction of cross-modal information; addressing the problem of metadata fragmentation caused by the independent processing of different modalities in traditional methods, this invention introduces an intermodal semantic relationship recognition mechanism to integrate complementary information scattered in different modalities into a complete metadata record; and ensures high-quality extraction of information from each modality through technologies such as image quality enhancement, confidence-annotated OCR, in-depth layout analysis, and seal-specific recognition. By constructing a knowledge graph of the archival domain that integrates historical context and employing context-aware entity recognition and standardization methods, the system can accurately identify and standardize historical entities in archives, solving the problem of high error rates in the identification of historical institutions, place names, and terminology using traditional methods. Through the establishment of a series of archival association recognition mechanisms using heterogeneous graph neural networks, combined with multi-strategy integrated conflict resolution and constraint propagation algorithms, dynamic consistency assurance is achieved during batch processing, ensuring that the metadata of associated archives has the same values ​​in fields that should maintain consistency, resulting in standardized and unified indexing results. Through knowledge graph-guided intelligent indexing and reinforcement learning strategy optimization, the system can continuously improve indexing quality based on archival characteristics and user feedback. Through multi-dimensional quality assessment and proactive repair mechanisms, the system automatically identifies and corrects metadata quality issues, marking low-quality records as items awaiting review, reducing the workload of manual review. Attached Figure Description

[0017] Figure 1 This is a module diagram of an intelligent extraction and indexing system for archival metadata in this invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses an intelligent extraction and indexing system for archival metadata, such as... Figure 1 As shown, it includes: The multimodal preprocessing module is used to acquire raw multimodal archive data, preprocess the raw multimodal archive data, and obtain a structured multimodal archive data set. S11, Image Quality Assessment. Scanning the original digitized archival documents yields original multimodal archival data. The input scanned images are then subjected to quality checks. The Laplacian operator is used to calculate image sharpness; images with a Laplacian variance less than 100 are considered blurry. The difference between images before and after median filtering is used to assess noise levels; images with a difference greater than 15 are considered high-noise. The Hough line detection algorithm is used to identify image edges, and the angle between the edge and the horizontal line is calculated as the tilt angle; images with a tilt angle greater than 2 degrees are considered tilted. The image quality score is calculated by weighting three indicators: sharpness, noise level, and tilt angle, with weights of 0.5, 0.3, and 0.2 respectively. Finally, the score is mapped to the 0-1 range using the Sigmoid function.

[0020] S12, Adaptive Image Enhancement Processing. For images with a quality score below 0.6, targeted enhancement processing is performed, including for blurred, high-noise, tilted, and low-contrast images. For blurred images, a deep learning-based super-resolution reconstruction network is used to improve sharpness. This network learns the mapping relationship from low-resolution to high-resolution images to reconstruct high-frequency detail information. For high-noise images, a non-local mean denoising algorithm is used, which removes noise by weighted averaging of similar image blocks while preserving edge information. For tilted images, affine transformation is used for rotation correction based on the detected tilt angle, ensuring the corrected image edges are parallel to the horizontal line. For low-contrast images, adaptive histogram equalization is used to improve contrast. The image is divided into blocks, and histogram equalization is performed on each block. Bilinear interpolation is then used to fuse the blocks, avoiding block artifacts.

[0021] S13, Multimodal OCR Text Recognition. An end-to-end OCR system fusing visual features and a language model is used to extract text from the enhanced image. A Convolutional Neural Network (CNN) is used to extract visual features from the image, resulting in a feature map. A Bidirectional Long Short-Time Memory (BiLSTM) network is then used to model the sequence of the feature map, capturing the contextual dependencies of the text. The BiLSTM processes the sequence through forward and backward LSTMs, concatenating the hidden states from both directions to form a bidirectional representation.

[0022] The Connectionist Temporal Classification (CTC) algorithm is used for decoding. CTC allows the output sequence length to be shorter than the input sequence length by introducing blank labels, thus enabling end-to-end training. For each recognized character, the system outputs its recognition confidence score, which is calculated using the Softmax function, and outputs the normalized probability of each character. Simultaneously, an N-gram language model is introduced for post-processing of the recognition results, correcting common OCR errors. The language model reorders the recognition results based on the statistical probabilities of the character sequences.

[0023] S14, Confidence Threshold Filtering. Each character in the recognized text is traversed, and characters with a confidence level below 0.8 are marked as low-confidence characters and represented by a special symbol in the text. For words composed of multiple consecutive low-confidence characters, the entire word is marked as a word to be verified. The proportion of low-confidence characters in the text is counted; when the proportion exceeds 0.3, the entire text is marked as high-uncertainty text, and a more conservative strategy is applied to this text in subsequent processing. This confidence labeling mechanism effectively identifies potentially erroneous parts in the OCR process and prevents errors from propagating to subsequent steps.

[0024] S15, In-depth Layout Analysis. The LayoutLMv3 multimodal document understanding model is used to analyze the layout of the document pages. This model receives both document images and text sequences as input, and learns a joint representation of images, text, and layout positions through the multimodal TranSformer architecture. The model divides the page into multiple regions, performs semantic classification on each region, and identifies functional areas such as the main text area, title area, handwritten annotation area, stamp area, table area, and image area.

[0025] The steps for constructing the input representation of the multimodal document understanding model are as follows: the embedding vector of the text token, the image patch embedding extracted by CNN, the one-dimensional positional encoding, and the two-dimensional positional encoding are concatenated in sequence to form a comprehensive input representation. Among them, the text embedding captures semantic information, the image embedding captures visual features, the one-dimensional positional encoding represents the positional relationship in the sequence, and the two-dimensional positional encoding represents the spatial positional relationship on the page.

[0026] After feature fusion using a multi-layer TranSformer encoder, the multimodal document understanding model outputs the class probability distribution for each pixel. The argmax operation is then used to obtain the class label for each pixel, thus achieving pixel-level page segmentation. For complex layouts such as multi-column layouts and text-image intersections, the multimodal document understanding model can accurately identify the boundaries and categories of each region by learning the layout patterns of large-scale archival samples.

[0027] S16, Seal Detection and Recognition. A dedicated seal processing workflow is employed for the seal areas in the annotation results. A seal detection algorithm based on color and shape features is used to accurately locate the seal's position coordinates and bounding box. Since seals are typically red and have circular or elliptical features, HSV color space conversion and shape matching can effectively detect the seal's position.

[0028] For detected seal images, a seal character recognition model is used to extract the text content from the seal. Due to issues such as wear, unclear images, and text overlapping with the background, the following processing steps are employed: The seal image is binarized to separate the seal text from the background. Morphological operations are used to remove broken strokes and noise, restoring character outlines. A specially trained seal character recognition model is used for character recognition. This model, trained on a large number of seal samples, has good recognition capabilities for various seal script styles, including seal script and regular script. For blurred seals, multi-frame fusion technology is used; if the same seal appears multiple times, the results of multiple recognitions are combined to improve accuracy.

[0029] S17, Multimodal Consistency Verification. The text content, seal area annotations, and seal information obtained in the above steps are integrated and their consistency verified. For the OCR text in the main text area, pixel-level alignment with the original image is performed, and the matching of the recognition result with the image is checked. For the institution name recognized by the seal, it is compared with the institution name mentioned in the main text to determine if they match. For the handwritten annotation area, its spatial relationship with the main text is identified, and the specific paragraph or sentence to which the annotation points is determined.

[0030] A verification strategy combining rules and models is employed. Rules include temporal logical consistency and organizational responsibility consistency, while the model uses a TranSformer-based text matching model to determine the semantic consistency of information across different modalities. Detected inconsistencies are marked as pending confirmation and output to the quality report. The final structured output data includes complete information such as the original image, enhanced image, regionalized OCR text and confidence scores, region type annotations, stamp locations and text, quality scores, and a list of inconsistencies, forming a high-quality multimodal archival dataset.

[0031] In some embodiments, an attention-based scene text recognition model can be used to replace the traditional OCR model, improving the recognition capability for complex text scenes. This model employs an encoder-decoder architecture, where the decoder dynamically focuses on the image region most relevant to the current character when generating each character, thereby improving the accuracy and robustness of recognizing curved text, artistic fonts, and blurred text.

[0032] The context entity recognition module extracts context vectors from a structured multimodal archive dataset, performs entity recognition and normalization based on the context vectors, and obtains archive text data with labeled and normalized entities. S21, Constructing a Historical Context Vector. This involves extracting contextual elements such as the year of formation, archival category, issuing institution, and location from a structured multimodal archival dataset. The year of formation reflects the historical period in which the archives were created; different periods have different political, economic, and social backgrounds, resulting in different terminology and institutional names. Archival categories, such as documentary archives, scientific and technological archives, and accounting archives, exhibit differences in professional terminology and expressions. The issuing institution reflects the source of the archives; a series of archives from the same institution are related. The location reflects the regional characteristics of the archives; different regions use different place names and dialects.

[0033] The aforementioned contextual elements are encoded into embedding vectors. The dates are numerically normalized and mapped through an embedding layer. Categorical features are encoded using a vocabulary to obtain vector representations. All embedding vectors undergo L2 normalization. The embedding vectors for dates, categories, units, regions, and other dimensions are concatenated to form a comprehensive contextual vector. This contextual vector can represent the historical background information of the archives, providing contextual clues for subsequent entity recognition.

[0034] S22, Context-Aware Text Encoding. The ArchiveBERT model, pre-trained on a large-scale archival corpus, is used to encode the text. This model learns archival terminology, historical expressions, and stylistic features. During encoding, the context vector is fused with the embeddings of text tokens. A linear transformation maps the context vector to the same dimension as the token embedding, and this is added to the embedding of each token to obtain a context-enhanced token representation. The enhanced token sequence is then input into the ArchiveBERT model, outputting hidden states that integrate text content and historical context information.

[0035] S23, Dynamic Knowledge Graph Query. Based on the context vector constructed in step S21, especially the chronological information, relevant historical entities are retrieved from the multi-period archival knowledge graph. The archival knowledge graph is organized hierarchically by historical period and includes databases of institutional names, historical place names, professional terms, and biographical information from different periods.

[0036] For archives created in a specific era, the database of entities from that era is queried. The query process employs a time window mechanism, searching not only entities from the year the archives were created but also entities within a five-year timeframe before and after, to cover potential references. Query criteria also include the location of creation, retrieving only relevant local institutions and place names within that region to minimize interference from irrelevant entities.

[0037] The query results form a candidate entity knowledge set, where each candidate entity includes information such as: entity name, entity type, duration, related description, and a list of aliases. This set serves as external knowledge for entity recognition, guiding the model to identify historical entities appearing in the archival text.

[0038] S24, Knowledge-Enhanced Entity Recognition. A knowledge-enhanced sequence labeling model is constructed for named entity recognition. The model employs a BiLSTM-CRF architecture, which can model the contextual dependencies of sequences and the constraints between labels, and is a classic method for sequence labeling tasks.

[0039] Building upon the standard BiLSTM-CRF, external knowledge features are introduced to enhance the model's capabilities. For each token in the text, the system searches the candidate entity knowledge set to see if an entity starting with that token exists. If it does, the knowledge features of that entity are extracted, including entity type and importance in the knowledge graph, and encoded as a knowledge feature vector. If it does not exist, a zero vector is used.

[0040] The context-aware representation output by BERT is concatenated with the knowledge feature vector and used as the input to BiLSTM. The concatenation process involves connecting the context-aware representation vector of each token with the corresponding knowledge feature vector according to the dimension, forming a comprehensive input vector containing semantic and knowledge information.

[0041] BiLSTM encodes the input sequence. The encoding process includes: a forward LSTM processes the sequence from left to right, and the hidden state at each position is calculated based on the current input and the hidden state at the previous position; a backward LSTM processes the sequence from right to left, and the hidden state at each position is calculated based on the current input and the hidden state at the next position; finally, the forward and backward hidden states at the corresponding positions are concatenated to form a bidirectional representation of that position, so that each position contains contextual information from both directions.

[0042] The output of the bidirectional LSTM is mapped to the label space through a linear layer to obtain the label score at each position. The mapping process is as follows: the bidirectional representation vector at each position is mapped to the label space through a linear transformation, which includes matrix multiplication and bias addition. The output weight matrix maps the bidirectional representation to the score of each label category. The bias term adjusts the base score of each label, and finally the score vector of each label corresponding to that position is obtained.

[0043] A CRF layer is used to perform global optimization of the tag sequence. CRF performs global optimization by calculating the conditional probability of the tag sequence, and the sequence score is the sum of the emission score and the transition score at each position.

[0044] During training, the conditional probability of the correct label sequence is maximized, and during inference, the Viterbi algorithm is used to decode and obtain the optimal label sequence. The labels adopt the BIO annotation system, where B indicates the start of an entity, I indicates the interior of an entity, and O indicates a non-entity. Entity types include person names (PER), organization names (ORG), place names (LOC), time (TIME), and event (EVENT). Through knowledge augmentation, the model can more accurately identify historical entities in archival texts, especially entities existing in knowledge graphs, thus improving recognition accuracy.

[0045] S25, Cross-Period Entity Normalization. For identified entities, they need to be linked to standard entities in the knowledge graph to achieve entity normalization. Since the archives span multiple historical periods, the same entity may have different name formats, necessitating cross-period entity linking.

[0046] For each identified entity mention, candidate entities with similar names are retrieved from the candidate entity knowledge set. Similarity is calculated using a combination of edit distance and character n-gram similarity. Semantic similarity is further calculated by semantically matching the entity's contextual representation with the candidate entity's descriptive text. Temporal consistency constraints are also considered; the duration of a candidate entity should match the era in which the archive was created. A total score is calculated for each candidate entity, combining name similarity, semantic similarity, and temporal consistency scores. The total candidate entity score is a weighted average of the three dimensions: name similarity, semantic similarity, and temporal consistency, with weights of 0.4, 0.4, and 0.2, respectively.

[0047] The candidate entity with the highest score is selected as the link result, and the entity mention is linked to the standard entity to obtain the entity's unique identifier and canonical name. For new entities that cannot find a matching candidate entity, they are marked as entities to be added and submitted to experts for review before being added to the knowledge graph. The final output is the archive text data of the labeled canonical entities. Each entity is labeled with information such as entity type, original mention text, canonical name, and knowledge graph ID, realizing the canonical representation of the entity.

[0048] In some embodiments, due to the limited modeling ability of the BiLSTM-CRF model for long texts and its insufficient capture of long-distance dependencies, a TranSformer-based sequence labeling model can be used to replace the BiLSTM-CRF model. The aim is to enhance the modeling ability for long texts and complex semantic structures. The TranSformer model directly models the dependencies between any two positions in a sequence through a self-attention mechanism, without relying on loop structures. TranSformer employs a multi-head self-attention mechanism, calculating attention weights through query, key, and value matrices, and concatenating the output after parallel processing by multiple attention heads. Through multi-layer TranSformer encoder stacking, the model can learn deep semantic representations of text. Based on the TranSformer output, a CRF layer is also used for sequence labeling, preserving the constraints between labels. This replacement scheme performs better in entity recognition of long text files, capturing entity relationships across sentences and paragraphs, improving the accuracy and completeness of recognition.

[0049] The cross-modal fusion module performs confidence-weighted multimodal fusion and logical verification based on the archive text data with labeled normalized entities and the structured multimodal archive data set to obtain complete archive metadata records; S31, Confidence-Weighted Multimodal Fusion. A multimodal fusion model is constructed to simultaneously process information from both text and image modalities. The text modality includes entities identified by the contextual entity recognition module and their contextual information, while the image modality includes seal recognition results, handwritten annotation recognition results, etc. Each modality's information is accompanied by a confidence score. The confidence score for text entities comes from the output probability of the entity recognition model, while the confidence score for image OCR comes from the OCR confidence evaluation of the multimodal preprocessing module.

[0050] The weights of each modality in the fusion process are dynamically adjusted based on its confidence level. Modal information with high confidence levels should receive greater weights, while modal information with low confidence levels should have reduced weights to avoid the negative impact of low-quality information.

[0051] For the text modality, the text sequence and its entity annotations are used as input, and an embedding layer is used to obtain the text embedding vector. For the image modality, characters from stamp recognition, handwritten annotations, etc., are used as input, and an embedding layer is used to obtain the image text embedding vector. Simultaneously, the positional information of each element on the page is encoded to obtain the positional embedding vector. The average confidence score for each modality is calculated, and the element confidence scores for both the text and image modalities are averaged separately. Based on the confidence scores, dynamic weights are calculated using the Softmax function, with the temperature parameter τ set to 2.0 to control the influence of confidence score differences on the weights.

[0052] Each modal embedding is dynamically weighted based on its confidence level, and then concatenated and fused through a multi-layer TranSformer encoder to learn cross-modal interaction relationships. The fused multimodal feature representation integrates text, image, and location information, and dynamically adjusts the contribution based on the reliability of each modality, enabling a more robust representation of archive content.

[0053] S32, Bidirectional Attention Association. A bidirectional attention mechanism is employed to identify semantic relationships between different modalities of information. For example, the document text mentions approval, the handwritten annotations contain specific approval opinions, and the seal image shows the approving authority. These three pieces of information are semantically related and should be collaboratively extracted to form complete metadata for the approval matter.

[0054] The text-to-image attention algorithm measures the correlation between each token in the text and each element in the image. This is achieved by calculating the relevance score between text and image features, which is then normalized using Softmax to obtain the attention weights. Similarly, the image-to-text attention algorithm calculates the relevance score between the feature vectors of image elements and the feature vectors of text tokens. This relevance score is then exponentially calculated, and the exponential value is divided by the sum of the relevance score exponential values ​​of the image element to all text tokens to obtain the normalized attention weights. For each token in the text, the most relevant image element can be found; similarly, for each element in the image, the most relevant text token can be found.

[0055] Based on attention weights, image features are weighted and aggregated to obtain a text-enhanced image representation.

[0056] Through a bidirectional attention mechanism, text and image features enhance each other, providing a foundation for subsequent relation extraction.

[0057] S33, Joint Relation Extraction. A joint entity relation extraction model is constructed to simultaneously identify entities and the relationships between them, avoiding the error propagation problem of traditional pipeline methods (identifying entities first and then extracting relations). The joint entity relation extraction model uses a sequence-to-sequence labeling approach, transforming the relation extraction task into a sequence labeling task.

[0058] For the input text, the joint entity relation extraction model predicts a label for each token, which includes entity boundary and relation type information. An extended labeling system is adopted, such as relation start - entity type indicating the beginning of an entity and that entity is the subject of a relation, relation interior - entity type indicating the interior of an entity, and relation end - entity type indicating the object entity of the relation.

[0059] Based on the cross-modal enhanced features obtained in step S32, the joint entity relationship extraction model performs sequence labeling through CRF. The labeling results are decoded into triple form, and the triple contains three elements: subject, relationship, and object. For example, Li Ming, holds the position of, and section chief of the finance section form a complete triple relationship.

[0060] For cross-modal relationships, the joint entity relationship extraction model can identify the relationships between entities from different modalities. For example, the "approval number" in the text and the unit name identified in the seal can be extracted as a triple, and the file, approval unit, and XX department form a cross-modal triple relationship.

[0061] The joint entity relationship extraction model outputs all the identified triples to form a set of metadata triples. These triples contain the key factual information in the archive and provide a basis for filling the metadata fields.

[0062] S34, Field mapping and rule constraints. According to the predefined archive metadata template, map the extracted triples to the metadata fields. The archive metadata template contains more than 20 fields such as file number, title, responsible person, formation time, formation unit, retention period, confidentiality level, subject terms, content summary, etc.

[0063] Establish mapping rules from triple relationship types to metadata fields. For example: the subject in the employment relationship is mapped to the responsible person field, the object in the time relationship is mapped to the formation time field, the object in the document issuing unit relationship is mapped to the formation unit field, and the object in the subject relationship is mapped to the subject terms field; for some fields, the information of multiple triples needs to be integrated. For example, for the title field, it may be necessary to combine information such as the document theme, document type, and document issuing unit. Use a templated method to generate the title: the name of the formation unit plus "Regarding", then the subject content, then "of", and finally the document type, and concatenate them in this order to form a complete title.

[0064] For the content summary field, an extractive summarization method is used, selecting the most important sentences from the document text to form a summary. Importance is calculated using the TextRank algorithm. The text is constructed as a sentence graph, with edge weights representing the similarity between sentences. The importance scores of sentences are calculated iteratively, and the top 3-5 sentences with the highest scores are selected as the summary. The specific implementation of the TextRank algorithm is as follows: the document text is segmented into a sequence of sentences, and the semantic similarity between sentences is calculated as the edge weights of the graph. The similarity is calculated using the cosine similarity of sentence vectors. An undirected weighted graph is constructed, with each sentence as a node. The importance score of each node is updated iteratively. The update formula is based on a random walk model, where the score of the current sentence is equal to the damping coefficient multiplied by the weighted sum of the scores of all neighboring sentences pointing to that sentence, plus a uniform distribution term. The iteration continues until convergence, and finally, the top 3-5 sentences are selected as the summary based on their scores. The K value is dynamically adjusted according to the document length: 3 sentences are selected for short documents (less than 10 sentences), 4 sentences for medium-sized documents (10-30 sentences), and 5 sentences for long documents (more than 30 sentences).

[0065] During the mapping process, field values ​​that do not conform to the specifications are automatically adjusted. For example, if the identified date is "October 1, 1999", it is converted to the standard format "1999-10-01"; if an organization's abbreviation is identified, its full name is queried through the knowledge graph and then replaced.

[0066] The initial metadata record is obtained, with each field populated with values ​​extracted and transformed from the triples.

[0067] S35, Logical Consistency Verification. A logical consistency inference engine is built to perform multi-dimensional verification of metadata records, ensuring logical consistency between fields and intelligently completing missing fields.

[0068] Perform time consistency checks. The archival lifecycle contains multiple time points, including creation time, receipt time, registration time, and archiving time. These times should meet a logical order: creation time should be earlier than or equal to receipt time, receipt time should be earlier than or equal to registration time, and registration time should be earlier than or equal to archiving time, forming a complete logical time sequence. If an unreasonable time sequence is found, it is marked as a logical error, and an attempt is made to correct it. Correction strategies include checking for year identification errors (such as misidentifying 2019 as 2009) and inferring the correct time through contextual information.

[0069] Perform a job title consistency check. The responsible person's job title should match their employing unit and the time their personnel file was created. Query the person's employment information for a specific period using the knowledge graph to determine if the job title in the metadata is correct. The job title consistency check steps are as follows: Query the person's employment relationship triplet in the knowledge graph state corresponding to the file's creation time, checking if the person, employment relationship, and job title exist in the knowledge graph; if they exist, it is considered consistent and a true value is returned; if they do not exist, it is considered inconsistent and a false value is returned. If the job title does not match, query the person's actual job title at the creation time and correct it.

[0070] Perform a geographic consistency check. The location of the generating unit should match the location mentioned in the archive. For example, if the generating unit is "XX Bureau of City A," then the location field should be City A or a district of City A, and should not contain place names from other provinces or cities. Geographic consistency is verified through the attribution relationship between units and regions in the knowledge graph.

[0071] For the retention period field, it is inferred based on the archive content and document type. Different types of archives have different retention period regulations, and a mapping rule from document type to retention period is established based on legal regulations and archive retention period regulations. If the retention period is not filled in the metadata, a suggested value is automatically filled in based on the document type.

[0072] For missing fields, a completion method based on related archives is used. If the current file belongs to a series of archives, the corresponding field values ​​of other archives in that series are queried. If the field values ​​of most archives are consistent, then the consistent value is used to complete the missing field of the current file. For example, multiple documents in the same case file usually have the same case file number, record group number, and generating unit, and can be mutually completed.

[0073] After multi-dimensional verification and completion, a complete archive metadata record is output, with each field accompanied by a confidence score and a data source identifier. The confidence score is calculated by comprehensively considering factors such as the output probability of the extraction model, the results of logical verification, and the reliability of the completion method. The data source identifier records whether the field was directly extracted, logically inferred, correlated, or a default value, facilitating subsequent review.

[0074] In some embodiments, an event-based metadata extraction method can be used instead of a triplet extraction method to identify complete event information in the archive. The event extraction model analyzes the archive text into a structured event representation, including elements such as event type, trigger words, and event arguments. Each element of the event directly corresponds to a metadata field, which can capture more complete semantic information compared to triplet extraction, making it particularly suitable for archive types such as meeting minutes and decision documents.

[0075] The archive association module, based on complete archive metadata records, performs association identification and consistency detection between archives to obtain a set of archive metadata with consistency constraints; S41, Multi-granularity Association Identification. A mechanism for identifying associations between archives is constructed to identify related archives belonging to the same series. Associations are divided into explicit and implicit relationships. Explicit associations are based on direct matching of metadata fields; archives with the same record number belong to the same source institution, archives with the same file number belong to the same file, and archives with consecutive file numbers may be a series of documents. The explicit association score is calculated by weighting the matching of record number, file number, and file number, with weights of 0.4, 0.4, and 0.2, respectively. Implicit associations are based on the similarity of archive content; content similarity is measured by calculating the cosine similarity of the archive text representation, which uses context-aware BERT encoding; spatiotemporal proximity measures the proximity of the archives' creation time and location; entity co-occurrence measures the proportion of entities that appear together in two archives; the implicit association score is a weighted score combining content similarity, spatiotemporal proximity, and entity co-occurrence, with weights of 0.5, 0.3, and 0.2, respectively.

[0076] The overall association score balances the weight contributions of explicit and implicit associations, with explicit associations having a weight of 0.7 and implicit associations having a weight of 0.3.

[0077] For all pairs of archives, calculate their association score. When the score exceeds 0.6, establish an association edge between the archives. Finally, an archive association graph is formed, which contains a set of archive nodes and a set of association edges. Each edge is accompanied by an association score and an association type label.

[0078] S42, Heterogeneous Graph Neural Network Learning. A Heterogeneous Graph Convolutional Neural Network (HGNN) is constructed to learn representations of the archive association graph. The archive association graph is a heterogeneous graph, containing various types of nodes such as archive nodes, entity nodes, and topic nodes, as well as various types of edges such as those related to the same file, the same event, mentions, and belonging to.

[0079] The process of constructing the node set of a heterogeneous graph is as follows: perform a union operation on the set of archive nodes, the set of entity nodes, and the set of topic nodes to form a complete node set containing all different types of nodes. The process of constructing the edge set of a heterogeneous graph is as follows: perform a union operation on the set of edges relating between archives, the set of edges mentioning entities in archives, and the set of edges belonging to topics in archives to form a complete edge set containing all different types of relationships.

[0080] Heterogeneous graph neural networks model different types of relations separately. For each relation type, an independent message passing function is defined. Nodes aggregate information about their corresponding neighboring nodes according to a specific relation type. The message aggregation process is as follows: for all neighboring nodes of node i under relation type τ, the feature representations of these neighboring nodes in the current layer are collected. The features of all neighbors are then aggregated using a relation type-specific aggregation function to obtain the aggregated message for node i under that relation type.

[0081] The specific implementation of the aggregation function is as follows: perform a relation-specific linear transformation on the feature representation of neighboring nodes, calculate the importance weight of each neighbor through an attention mechanism, the weight calculation is based on the correlation between the features of the current node and the neighboring nodes and the semantic information of the relation type, sum the weighted features of all neighbors as the aggregation result, and use attention-weighted aggregation.

[0082] Heterogeneous Graph Neural Networks (HGNNs) employ a relation-specific attention mechanism, calculating attention scores between nodes and normalizing them into weights. Through multi-layer stacked HGNNs, node representations are updated layer by layer, ultimately resulting in each archive node learning an embedding vector that integrates information from multiple relation types. This embedding vector reflects the archive's position in the relation graph and its similarity to other archives, providing a representational basis for consistency detection. The node representation update calculation steps are as follows: summing the aggregated messages of the node across all relation types, performing a non-linear transformation through an activation function, and obtaining the node's representation in the next layer.

[0083] S43, Consistency Conflict Detection. For connected file pairs in the file association graph, check the consistency of their metadata. For fields that should maintain consistency, such as the file title of the same file or the project name of the same project, check whether their field values ​​are consistent.

[0084] For a pair of files, iterate through each metadata field and determine if the field values ​​are consistent. The conflict detection steps are as follows: compare the values ​​of the same field in the two files to see if they are equal; determine whether the field should be consistent based on the relationship type between the files; if the field values ​​are not equal and the field should be consistent under the current relationship type, then a conflict is identified and a true value is returned; otherwise, a false value is returned. The implementation steps for consistency field judgment are as follows: establish a mapping rule table between relationship types and consistency fields, such as case file relationships corresponding to fields like case file title, file group number, and retention period, and event relationships corresponding to fields like event name, time, and location; query the mapping table based on the input relationship type and check if the field is in the consistency field list for that relationship type; finally, return a boolean value: true if the field needs to be consistent, otherwise, return a false value.

[0085] When inconsistencies are detected, the confidence level of each archive field value is considered. The confidence level is derived from the field-level confidence score of the cross-modal fusion module. Field values ​​with high confidence are given priority, while field values ​​with low confidence are marked as potentially incorrect.

[0086] For each conflicting field, calculate the support for each candidate value. The support is the sum of the weighted confidence scores of the number of times the candidate value appears in all relevant files. The steps for calculating the support are: traverse all relevant files, find the files where the field value is equal to the candidate value, sum the confidence scores of the field in these files, and obtain the support for the candidate value.

[0087] Output a consistency conflict report, which lists all detected conflicts, including conflict fields, involved files, candidate values ​​and their support and confidence levels, providing a basis for conflict resolution.

[0088] S44, Multi-strategy Conflict Resolution. For each conflict field in the consistency conflict report, a multi-strategy approach is used to resolve the conflict and determine the standard value for that field. The multi-strategy approaches include: The strategy employs several methods: **Majority voting:** Selecting the candidate value that appears most frequently in relevant archives as the standard value. This strategy assumes that the values ​​in most archives are correct and is suitable for cases where a minority of archives contain errors. **Confidence ranking:** Selecting the candidate value with the highest confidence level as the standard value. This strategy relies on the confidence assessment output by the trust model and is suitable for cases with significant differences in extraction quality. **Historical accuracy:** Candidate values ​​are selected based on a weighted average of the historical accuracy rates of each archive or extraction source. Sources with high historical accuracy receive higher weights for their candidate values. Historical accuracy is calculated through user feedback and manual review. **Knowledge graph validation:** The rationality of each candidate value is verified using a knowledge graph. For example, in cases of conflicting institution names, the system checks if each candidate institution existed at the time the archive was created and selects the existing candidate value.

[0089] By combining the above strategies, a weighted score is obtained by integrating the scores of four strategies: majority voting, confidence level, historical accuracy, and knowledge graph verification, with weights of 0.3, 0.3, 0.2, and 0.2, respectively.

[0090] The candidate value with the highest overall score is selected as the standard value. Candidate values ​​with a score difference of less than 0.1 are marked as uncertain and submitted for human-machine collaborative review. For each conflicting field, after determining the standard value, all related files are updated to reflect that field to achieve consistency correction. The original value before correction and the basis for correction are also recorded for traceability.

[0091] The final output field standard value set includes the standard value of each conflicting field, a list of related files, and the basis for correction.

[0092] S45, Constraint Propagation Maintains Consistency. When batch processing files, for fields with established standard values, these standard values ​​are automatically applied when processing subsequent related files, forming a dynamic constraint mechanism.

[0093] A constraint network is constructed, where nodes are archive metadata fields and edges represent consistency constraints. When a standard value for a field is determined, that constraint propagates along the association graph to all relevant archives.

[0094] Constraint propagation employs a message passing mechanism. The constraint message is constructed as follows: when the standard value of field f in file j is determined, a constraint message containing the standard value and the confidence score of that standard value is propagated to the associated file i. The message content includes the determined standard value and the corresponding confidence score.

[0095] File i receives constraint messages from multiple related files, and updates field values ​​by combining all messages. The calculation steps for updating field values ​​are as follows: iterate through all incoming constraint messages, and for each candidate value, sum the confidence scores of all messages containing that candidate value, and select the candidate value with the highest weighted confidence score as the updated value for that field.

[0096] The constraint propagation process iterates until the field values ​​of all files converge and no longer change. Through constraint propagation, even if the initial extraction results of newly processed files contain errors, they can be automatically corrected by the established consistency constraints, ensuring the overall consistency of batch processing.

[0097] The human-machine collaborative feedback received during the processing, i.e. the expert's judgment on the uncertainties, is immediately updated to a high-confidence standard value, triggering a new round of constraint propagation and spreading expert knowledge throughout the entire network.

[0098] The final output is a set of archive metadata with consistency constraints. All archives in this set should have the same values ​​on consistent fields, and the output also includes information on the source and propagation path of the constraints. The updated archive relationship graph and constraint network are also output, providing a consistency foundation for subsequent archive processing.

[0099] The intelligent indexing module performs hierarchical intelligent indexing and quality feedback on the metadata set of archives with consistency constraints, and obtains standardized indexing results; S51, Hierarchical Multi-Label Classification. A hierarchical classification model is constructed to classify archives by subject, into multiple levels such as first-level categories, second-level categories, and third-level categories, with parent-child relationships between categories at each level.

[0100] The classification system is modeled as a tree structure, with the root node representing the top-level category and the leaf nodes representing the finest-grained categories. Each file can belong to multiple category paths, thus making it a multi-label classification task.

[0101] The hierarchical classification model takes the text content and metadata of the archives as input and obtains an archive representation vector through BERT encoding. For each category in the classification system, a binary classifier is constructed to determine whether the archive belongs to that category. Category classification calculates the probability of an archive belonging to each category using linear transformation and the Sigmoid activation function.

[0102] To ensure hierarchical consistency, hierarchical constraints are introduced. If a file belongs to a subcategory, it must belong to all its ancestor categories; conversely, if a file does not belong to a parent category, it should not belong to any of its subcategories. Hierarchical constraints are implemented using a loss function. The calculation steps are as follows: traverse all parent-child category pairs, calculate the difference between the probability of the subcategory and the probability of the parent category. If the difference is positive, it is included in the loss; if the difference is negative or zero, the loss is zero. Finally, sum the losses of all parent-child category pairs to obtain the total hierarchical constraint loss. This loss term penalizes cases where the probability of the subcategory is greater than the probability of the parent category.

[0103] The total loss function is a weighted sum of the classification loss and the hierarchical constraint loss. The calculation process is as follows: the classification loss is added to the hierarchical constraint loss, where the weight parameter λ of the hierarchical constraint loss is set to 0.1 to control the importance of hierarchical constraints in the total loss and ensure that the model maintains the consistency of the hierarchical structure while optimizing classification accuracy.

[0104] The hierarchical classification model outputs the predicted probability of each category. Categories with a probability greater than 0.5 are determined to be the category to which the file belongs. Simultaneously, post-processing with hierarchical constraints is applied: if a subcategory is selected, all ancestor categories are forcibly selected; if a parent category is not selected, all subcategories are removed.

[0105] The final result is a multi-level category label for the archives. A single archive may belong to multiple category paths at the same time, reflecting the multifaceted and complex nature of the archives' subject matter.

[0106] S52, Knowledge Graph-Guided Keyword Extraction. Keywords from the archives are extracted as candidate topic tags. Traditional keyword extraction methods, such as TF-IDF, rely solely on statistical information and may extract high-frequency but semantically low-value words. A knowledge graph-guided approach is used to extract keywords semantically relevant to the archives' themes.

[0107] The TF-IDF algorithm is used to calculate the statistical importance of each word. TF-IDF calculates the statistical importance of a word by multiplying its term frequency and inverse document frequency. The term frequency is log-normalized, which means that the logarithmic number of times the word appears in a document is incremented by 1. The inverse document frequency is calculated by dividing the total number of documents by the number of documents containing the word and then taking the logarithm.

[0108] Knowledge graph relevance is introduced to calculate the semantic association strength between candidate keywords and identified entities. The knowledge graph relevance calculates the association strength by combining three dimensions: path distance, common neighbors, and vector similarity, with weights of 0.4, 0.3, and 0.3, respectively.

[0109] The overall keyword score is derived by weighting the TF-IDF score and the knowledge graph relevance, using a fusion of TF-IDF statistical importance and knowledge graph semantic relevance, with weights of 0.6 and 0.4, respectively. The relevance of keywords to archival categories is also considered, and the score of relevant keywords is improved by leveraging category concept similarity.

[0110] Based on the overall scores, the top 10-15 words are selected as candidate topic tags, forming a set of candidate topic tags. These tags not only appear frequently in the document but are also semantically related to the core entities and subject categories of the archive, accurately reflecting the archive's theme.

[0111] S53, Subject heading normalization mapping.

[0112] Standardized subject heading lists are used in records management, and standard terms from these lists should be used during indexing. Extracted candidate tags are mapped to standardized terms in the subject heading list. For each candidate tag w, a matching standardized term is retrieved from the subject heading list. Matching methods include: Exact matching: Candidate tags are exactly the same as the standard words, and are directly mapped. Synonym matching: Candidate tags are synonyms of standard words. The thesaurus records a list of synonyms for standard words. The system checks whether a candidate tag is in the list of synonyms for a certain standard word. If it exists, it is mapped to that standard word. Similar word matching: Candidate tags are semantically similar to, but not exactly the same as, standard words. The semantic similarity between candidate tags and all standard words is calculated using the cosine similarity of word embedding vectors. The standard word with the highest similarity (greater than 0.7) is selected as the mapping result. Hypernym expansion: For mapped standard words, their hypernyms are automatically added as additional tags according to the hierarchical structure of the thesaurus. For example, after "digital archives management" is mapped to a standard word, its hypernym "archives management" is also added as a tag, forming a multi-level index. Related word supplementation: Based on the correlation between standard words in the thesaurus, related thesaurus words are supplemented for archives.

[0113] After the above mapping and expansion, a standard subject tag set is formed. All tags in this set are standardized terms from the thesaurus and have been expanded according to hierarchy and relevance, resulting in standardized and unified indexing results.

[0114] S54, Reinforcement Learning Strategy Optimization. A reinforcement learning framework is established to continuously optimize the indexing strategy based on user feedback on the indexing results. The indexing process is modeled as a Markov Decision Process (MDP), with the agent representing the indexing system, the state representing the feature representation of the archive, the action representing the selection of an indexing label, and the reward representing user feedback.

[0115] The state includes information such as the text representation of the file, class tags, candidate keywords and their scores, which are encoded as vector representations; the action is to select a tag from the candidate tag set as the indexing result, and the action space is all possible tag selections.

[0116] The reward system defines user feedback on indexing results, including three types: acceptance, modification, and rejection. Acceptance indicates correct indexing and receives a positive reward; rejection indicates incorrect indexing and receives a negative reward; modification indicates partially correct indexing and receives a moderate reward. The reward value setting rule is as follows: if the user accepts the indexing result, a reward of +1.0 is given; if the user modifies the indexing result, a moderate reward of +0.5 is given; and if the user rejects the indexing result, a penalty of -1.0 is given. The discount factor γ is set to 0.9, and the learning rate is set to 0.001.

[0117] For each trajectory, calculate the cumulative sum of the discounted rewards at each time step. The discount factor is used to reduce the weight of future rewards. Calculate the expected cumulative reward of all possible trajectories under the policy distribution to obtain the expected cumulative reward of the policy as the optimization objective.

[0118] The policy gradient algorithm is used to update the policy parameters. The steps for calculating the policy gradient are as follows: For each time step, calculate the gradient of the policy log probability with respect to the parameters, multiply it by the cumulative reward starting from that time step, sum it over all time steps, and calculate the expectation under the trajectory distribution to obtain the gradient of the policy parameters. The cumulative reward is calculated by summing the discounted rewards of the current time step and all subsequent time steps.

[0119] The strategy is updated using collected user feedback data, and the objective function is optimized through gradient ascent. As user feedback data accumulates, the strategy gradually learns the indexing patterns preferred by users, and the indexing quality continuously improves. The optimized indexing strategy model is output, which can intelligently select the most suitable indexing tags based on document characteristics and historical feedback.

[0120] S55, Indexing Quality Monitoring. Establish an indexing quality monitoring mechanism to continuously monitor the system's indexing results and promptly identify quality issues.

[0121] The indexing acceptance rate for each type of archive is calculated. The indexing acceptance rate is the ratio of the number of accepted indexing results to the total number of indexing results.

[0122] For archive types with an acceptance rate of less than 70%, the reasons may be insufficient training samples for that type of archive or an inapplicable indexing strategy. Targeted model optimization should be carried out.

[0123] Analyze the usage frequency and accuracy of each topic tag to identify frequently misused tags. For tags that are frequently rejected, remove them from the candidate tag set or reduce their weight to decrease erroneous indexing.

[0124] To ensure consistency in indexing results, archives with similar content should have similar indexing tags. Calculate the content similarity and tag similarity of archive pairs. If the content similarity is greater than 0.8 but the tag similarity is less than 0.5, mark it as an anomaly and manually review it.

[0125] To ensure the completeness of the indexing results, each document should have an appropriate number of indexing tags, with a normal range of 3-8 tags. Fewer than 3 tags may indicate insufficient extraction, while more than 8 tags may indicate an excessive number of noisy tags. The tag quantity distribution should be statistically analyzed, and documents deviating from the normal range should undergo a quality review.

[0126] Based on the aforementioned monitoring indicators, an indexing quality assessment report is generated. This report includes: overall acceptance rate and acceptance rates for each type, a list of frequently erroneous tags, a list of abnormal indexing files, statistical data on tag quantity distribution, and quality trend analysis. The quality trend analysis compares changes in quality indicators over different time periods. The quality assessment report provides data support for system optimization, guides model improvement, and ensures continuous improvement in indexing quality. Finally, standardized indexing results are output, completing the intelligent indexing task.

[0127] The quality assessment module performs metadata quality assessment and proactive repair on the standardized indexing results, resulting in an optimized set of metadata records.

[0128] S61, Integrity Assessment. Assess the integrity of metadata records, checking that required fields are fully entered. Archive metadata includes required and optional fields. Required fields such as file number, title, responsible person, and creation date must be filled in, while optional fields such as security classification and retention period are entered selectively based on the archive type.

[0129] Iterate through all required fields of the metadata record, count the number of fields filled, and calculate the required field completeness rate as the ratio of the number of filled fields to the total number of required fields.

[0130] For optional fields, the optional field fill rate is the ratio of the number of filled optional fields to the total number of optional fields.

[0131] The completeness score combines the completion of required and optional fields, with required fields having a higher weight.

[0132] Identify missing fields. For required fields that are missing, determine if they can be completed using other methods. Fields that can be completed are marked as items to be completed, and fields that cannot be completed are marked as quality defects. These are then output to the integrity assessment results.

[0133] The integrity assessment results include information such as integrity score, required field completeness rate, optional field fill rate, missing field list, and field completion list.

[0134] S62, Accuracy Assessment. Assess the accuracy of metadata field values ​​by detecting potential errors using multiple validation methods.

[0135] Content matching verification is performed. For metadata extracted from the archive content, its degree of matching with the original text is verified. For example, the title field should appear in the body of the archive or can be derived from the main theme of the body. The preprocessing steps for content matching verification are as follows: text preprocessing is performed on the field values ​​and archive content, including removing punctuation marks, converting to lowercase, and word segmentation; a vocabulary set is constructed, and the intersection of the field value vocabulary and the content vocabulary is calculated; for semantic relevance calculation, the field values ​​and content are encoded into vector representations respectively, semantic embeddings are obtained through a pre-trained language model, and cosine similarity is calculated as the relevance score.

[0136] The steps for calculating content matching score are as follows: calculate the proportion of words in the field value that appear in the file content as the coverage, calculate the semantic relevance between the field value and the file content, and finally multiply the coverage and semantic relevance to obtain the content matching score.

[0137] The specific implementation of calculating the semantic relevance between field values ​​and file content is as follows: Field values ​​and file content are encoded into vector representations using a pre-trained BERT model. The cosine similarity between the two vectors is calculated as the semantic relevance score, ranging from -1 to 1, and mapped to the 0-1 interval through a linear transformation. For long text content, a sliding window approach is used to calculate relevance in segments, and the maximum value is taken as the final score. Simultaneously, considering the semantic characteristics of field types, different relevance calculation weights are applied to different types of fields. A high matching degree indicates that the field value and content are consistent, resulting in high accuracy.

[0138] For entity fields such as person names, organization names, and place names, verification is performed within the knowledge graph. Verification includes checking if the entity exists in the knowledge graph, whether the entity's time information matches the document's creation time, and whether the relationships between entities are reasonable, such as whether the responsible person's position matches their organization. The knowledge graph verification score calculation steps are as follows: Query the knowledge graph to see if the entity exists. If the entity exists and matches the document information, the score is 1; if the entity exists but does not match the document information, the score is 0.5; if the entity is not found in the knowledge graph, the score is 0.

[0139] For numeric or formatted fields, check whether they conform to normal distribution and format specifications. For example, year fields should be four digits within a reasonable range; document number fields should conform to specific encoding rules. Calculate anomaly scores, and mark outliers as suspicious items.

[0140] The accuracy score is calculated based on the above verification results. The scores are weighted by the three dimensions of content matching, knowledge graph verification, and anomaly detection, with weights of 0.5, 0.3, and 0.2, respectively.

[0141] Output accuracy assessment results, including accuracy score, validation results for each field, list of suspected error fields, error type analysis, and other information.

[0142] S63, Assess the consistency of metadata. Metadata consistency includes logical consistency between fields within a single file, as well as consistency between related files.

[0143] Internal consistency checks have been performed in the cross-modal fusion module, including time logic, job matching, and geographic matching. Here, statistical consistency is further checked. For numerical fields, it is checked whether they are within the normal range. For example, the page number field should match the actual number of pages scanned in the archive.

[0144] The consistency of associated files has been checked in the file association module; this section evaluates the effectiveness of consistency maintenance. The actual consistency rate of associated files on fields that should maintain consistency is calculated after consistency constraints are applied: The steps for calculating the series file consistency rate are as follows: Iterate through all related file pairs and all fields that should be consistent, count the number of file pairs with the same field values ​​– the number of field combinations, and divide this number by the product of the number of related file pairs and the number of consistent fields to obtain the series file consistency rate.

[0145] The consistency score is a weighted score combining internal consistency and series consistency, with weights of 0.6 and 0.4, respectively. The output consistency assessment results include the consistency score, a list of internal inconsistencies, and the series inconsistency rate.

[0146] S64, Standardization Assessment. Assess whether metadata field values ​​comply with archival management standards and norms. Various fields have specific cataloging standards, including: time fields should use the ISO 8601 standard format; responsible persons should be listed in full, with multiple responsible persons separated by semicolons; subject terms should use standardized terms; and retention periods should be specified as permanent, 30 years, 10 years, etc.

[0147] Iterate through each field and check if its format and value conform to the standards. For fields whose standardization scores do not conform to the standards, mark them as non-standard items and provide standardization suggestions. The steps for calculating the standardization score are as follows: iterate through all fields, check if each field value conforms to the corresponding standard, count the number of fields that conform to the standards, and divide the number of fields that conform to the standards by the total number of fields to obtain the standardization score.

[0148] The specific implementation of the normative judgment function is as follows: determine the corresponding normative standard based on the field type, such as checking whether the time field conforms to the ISO 8601 format, whether the responsible person field is the full name, and whether the subject term field is in the normative thesaurus; perform normative checks through regular expression matching, thesaurus query, format validation, etc.; check the value range for numeric fields, and check the length limit and character set legality for text fields; finally, return a boolean value, returning 1 if it conforms to the standard, and 0 if it does not.

[0149] Output the normativity assessment results, including normativity score, list of non-normative fields, and normativity recommendations.

[0150] S65, Timeliness Assessment. Assess the timeliness of metadata, checking if it is the latest version and if updates are needed. For dynamically changing information, such as changes in organization names, the metadata should be updated promptly.

[0151] The update interval of metadata is calculated as the difference between the current time and the last update time of the metadata. A timeliness score is then calculated based on the update interval using an exponential decay function. This function uses a natural constant as the base and a negative update interval in days divided by 365 as the exponent. The update interval is in days, and the decay coefficient is 1 / 365. More timely updates result in a higher score.

[0152] Output the timeliness assessment results, including timeliness score, update interval, list of outdated fields, and other information.

[0153] S66, Comprehensive Quality Assessment. Based on the assessment results of the above five dimensions, the comprehensive quality score of the metadata record is calculated. The calculation steps are as follows: multiply the completeness score, accuracy score, consistency score, standardization score, and timeliness score by their respective weights, and sum all weighted scores to obtain the comprehensive quality score. The weights for each dimension are 0.3, 0.3, 0.2, 0.15, and 0.05, respectively.

[0154] For each field, a field-level quality score is calculated, which is the product of four dimensions: existence, accuracy, consistency, and normalization. The field-level quality score is used to identify low-quality fields and guide the priority of remediation.

[0155] Generate a comprehensive quality assessment report, which includes: record-level overall quality score, five-dimensional sub-item scores, field-level quality score list, quality level classification, and details of problem fields.

[0156] S67, Proactive Repair Mechanism. For issues identified during quality assessment, a proactive repair mechanism is employed to attempt automatic correction.

[0157] For missing fields, a completion strategy is adopted, including completion based on related archives, where if the field value is consistent with other archives in the same series, the value is used to complete it; completion based on content inference, where missing information is extracted from the archive text through rules or models; and completion based on default values, where for fields with default rules, such as retention period, default values ​​are filled in according to the archive type.

[0158] For inaccurate fields, correction strategies are employed, including knowledge graph-based correction (if the correct entity name exists in the knowledge graph, the incorrect value is replaced); majority voting-based correction (if the field has different values ​​in multiple extraction sources, the value with the highest confidence or the most frequent occurrence is selected); and providing candidate correction values, listing possible correct values ​​and their basis for user selection.

[0159] For inconsistent fields, the consistency constraints from the application document association module are used for correction. For non-normalized fields, format conversion and thesaurus mapping are used for normalization.

[0160] During the repair process, the basis and confidence level of each repair operation are recorded. Repairs with a confidence level higher than 0.8 are applied automatically, while repairs with a confidence level lower than 0.8 are provided to the user as suggestions.

[0161] Output a list of repair suggestions and optimized metadata records. Repair suggestions include: the problem field, problem type, suggested repair value, basis for repair, and confidence level. The optimized metadata records apply high-confidence repair, improving low-quality fields.

[0162] S68, Quality Tier Management. Metadata records are divided into three levels based on their overall quality score: High quality (overall quality score greater than or equal to 0.8, directly entered into the database without manual review); Medium quality (overall quality score between 0.5 and 0.8, entered into the database after applying intelligent repair, random checks are recommended); Low quality (overall quality score less than or equal to 0.5, must be manually reviewed before entering into the database).

[0163] Each metadata record is appended with a quality grading identifier and processing suggestions, resulting in a final optimized metadata record set. All records in this set have undergone quality assessment and intelligent optimization, leading to improved quality and meeting the actual needs of record management.

[0164] The knowledge graph module constructs a time-series knowledge graph and intelligent retrieval service based on optimized metadata records, resulting in a dynamically evolving time-series archive knowledge graph and semantic retrieval scheme. S71, Incremental Knowledge Graph Construction. Optimizes the extraction of entity and relationship information from metadata records to incrementally construct the knowledge graph. Entities include types such as people, organizations, locations, events, and records. Relationships include types such as appointment relationships, affiliation relationships, location relationships, participation relationships, and referencing relationships.

[0165] For each newly processed metadata record, extract the entity and check if the entity already exists in the knowledge graph. If the entity already exists, update its attribute information; if it does not exist, create a new entity node. The entity node contains attributes such as entity name, entity type, entity description, duration, and related attributes.

[0166] Extract relational triples from the metadata, such as the responsible party's employer or a document mentioning a specific event. For each triple, create a corresponding edge in the knowledge graph, with each edge containing the relation type and relation attributes.

[0167] An incremental update mechanism is employed to avoid repeatedly building existing parts, thus obtaining the basic structure of the archival knowledge graph. For existing entities and relationships, only their attributes and statistical information are updated, such as the number of documents in which the entity appears. This incremental approach improves construction efficiency and supports continuous processing of large-scale archives.

[0168] S72, Temporal Triple Modeling. Archives possess explicit temporal attributes; entities and relations have different states at different historical periods. A temporal dimension is added to the knowledge graph, constructing a temporal knowledge graph. Temporal triples are represented using a quintuple structure, containing five elements: head entity, relation, tail entity, relation start time, and relation end time. For relations that are still valid, the end time is set to a future time or marked as a persistent state.

[0169] For changes in organization name, create organization entities at different times in the graph and establish evolutionary relationships, including XX State-owned Factory as the head entity, Restructuring as the relationship, XX Limited Company as the tail entity, May 1, 1998 as the start time, and Continuous status as the end time.

[0170] Through temporal modeling, knowledge graphs can record historical evolution information, supporting historical question answering and temporal reasoning. When querying someone's position in a specific year, triples are filtered based on timestamps to return valid relations for that period.

[0171] S73, Temporal Knowledge Graph Completion. Information gaps exist in the temporal knowledge graph; relationships exist between some entities but are not explicitly recorded in the archives. A knowledge graph completion algorithm is used to predict potential relationships and discover implicit archive connections. An embedding-based completion method (TranSlating Embeddings for Modeling Multi-relational Data, TranSE) is employed. TranSE embeds entities and relationships into a low-dimensional vector space. For triples, the sum of the head entity vector and the relationship vector should approximate the tail entity vector.

[0172] The following loss function calculation steps are used during training: For each correct triplet and its corresponding negative sampled triplet, calculate the distance. Calculate the difference between the distance of the correct triplet and the distance of the negative sampled triplet, and add the margin parameter γ=1.0. If the result is positive, it is included in the loss; if it is negative, the loss is zero. Summate the losses for all triplet pairs to obtain the total loss. The entity and relation embedding dimensions are set to 100, and the learning rate is set to 0.01. The distance function uses the L1 norm or L2 norm to calculate the distance between the head entity plus relation vector and the tail entity vector.

[0173] For time-series triples, time coding is introduced. Time coding represents time features by constructing periodic numerical sequences. Multiple different time period frequencies are set, and the corresponding sine and cosine values ​​are calculated for each frequency. All calculation results are combined into a complete time coding vector, thereby capturing the periodic patterns at different time scales.

[0174] After training, for any entity pair and relation, its triplet score can be calculated; a higher score indicates a greater probability of the relation existing. Predicted relations with scores exceeding 0.8 are added to the knowledge graph as inferred relations, distinguishing them from extracted relations. The score function is calculated as follows: the head entity vector, relation vector, and time-encoded vector are added together, the tail entity vector is subtracted, and finally, the L2 norm of the resulting vector is calculated as the score.

[0175] Knowledge graph completion uncovers implicit connections between archives. For example, if two archives involve the same event but are not explicitly related, the connection can be inferred through entity co-occurrence and spatiotemporal proximity, providing users with more comprehensive information on archive connections and outputting a dynamically evolving time-series archive knowledge graph.

[0176] S74, Graph Embedding Learning. To support efficient semantic retrieval, vector representations need to be learned for entities and relations in the graph. Graph Neural Networks (GNNs) are used for embedding learning. Entity node embeddings are calculated by aggregating neighbor node information. The specific steps are: transforming the weight matrix of the current layer embedding representation for each neighbor node; normalizing the node based on its degree, where the normalization coefficient is the reciprocal of the square root of the product of the current node's degree and the neighbor nodes' degrees; summing the normalized representations of all neighbor nodes; and obtaining the next layer's node embedding representation through an activation function. By stacking multiple GCNs, node embeddings incorporate information from multi-hop neighbors, capturing the global structure of the graph. For relation embeddings, a relation-aware graph attention network is used, learning different attention weights for different types of edges.

[0177] During training, multi-task learning is employed, simultaneously optimizing tasks such as link prediction, node classification, and relationship prediction. This ensures that the learned embeddings perform well in various downstream tasks, resulting in high-quality entity and relation vector representations. These vectors preserve the semantic and structural information of the graph, providing a foundation for semantic retrieval.

[0178] S75, Query Understanding and Expansion. User queries are typically expressed in natural language and need to be converted into structured graph queries. Query understanding involves identifying entities and intents within the query; named entity recognition is used to identify entities such as names of people, organizations, and locations mentioned in the query; intent classification is used to identify query types, such as searching for files, querying relationships, and tracing history. Entity linking is performed on the identified entities, linking entity mentions in the query to standard entities in the knowledge graph, using a comprehensive judgment based on name matching, semantic similarity, and contextual relevance. Based on the query intent and linked entities, a graph query is constructed. For example, querying "projects Zhang San participated in" is converted into a graph schema: establishing the participation relationship between Zhang San and the query variable, and establishing the type relationship between the query variable and project types, forming a complete query schema through the logical AND connection of these two relationships. Query expansion enriches the query semantics; expansion includes synonym expansion, replacing words in the query with synonyms to broaden the recall scope; hyponym expansion, such as expanding to specific meeting types when querying "meeting"; and association word expansion, adding concepts related to the query topic.

[0179] S76, Hybrid Search Strategy. Performs a graph query, employing a hybrid search strategy that balances precise matching and semantic similarity.

[0180] For structured queries, a graph traversal algorithm is used to find matching entities and paths in the knowledge graph. Starting from the initial entity in the query, the graph is traversed along the specified relation type to find the target entity that matches the pattern. Graph traversal supports multi-hop relation queries and can discover indirectly related files. For semantic queries, vector retrieval is used to encode the query text into a vector and calculate the similarity with the entity embedding vector. The specific steps are: calculate the dot product of the query vector and the entity vector, calculate the L2 norm of the two vectors respectively, divide the dot product result by the product of the two norms, and obtain the cosine similarity as the similarity score between the query and the entity. The entity and file with the highest similarity are returned.

[0181] The steps for calculating the comprehensive score of hybrid retrieval, which combines structured matching and semantic similarity, are as follows: calculate the structured matching score between the file and the query, calculate the vector similarity score between the file and the query, set the weight coefficient to 0.6, multiply the structured matching score by 0.6, add the vector similarity score multiplied by 0.4, and obtain the final comprehensive score.

[0182] The search results are sorted based on their overall scores, and a sorted list is returned. The files in the sorted list satisfy both structured constraints and semantic relevance to the query, indicating high search quality.

[0183] S77, Learning Ranking Optimization. Collect user interaction data with the search system, including query logs, click logs, and dwell time, reflecting the degree to which search results match user needs.

[0184] We construct a learning-based ranking model, transforming retrieval ranking into a machine learning problem. Features include: query-document matching features, document quality features, document importance features, and user historical preference features.

[0185] The pairwiSe learning ranking method is used to learn relative ranking relationships. For two files in the same query, if a user clicks on file A but not file B, then A is considered to be ranked higher than B. Training sample pairs are constructed, containing file i, file j, and query q, labeled as file i being superior to file j. The ranking model calculates the ranking score for each file in the following steps: extract the feature vectors of the file and the query, and perform a transpose multiplication operation between the feature vector and the weight vector to obtain the ranking score of the file for the query.

[0186] During training, the pairwiSe loss function is used. The calculation steps are as follows: For each training sample pair, calculate the difference between the score of the clicked file and the score of the unclicked file, take the negative value of the difference and perform an exponential operation, add 1 and take the logarithm, sum the losses of all sample pairs to obtain the total pairwiSe loss, and optimize to make the score of the clicked file higher than that of the unclicked file.

[0187] The model is continuously updated using user feedback data during training, and the ranking strategy is constantly optimized, making the search results increasingly relevant to user needs. The final output is an optimized search ranking model, forming a complete semantic retrieval solution.

[0188] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, enable the execution of an intelligent extraction and indexing system for archival metadata as described above.

[0189] In one embodiment of the present invention, a specific example is provided: This invention is applied to a real-world scenario at a provincial archive, whose collection includes various types of archives, such as government archives, corporate archives, and personal archives, dating from 1950 to 2020. The system's processing performance is demonstrated using the archives from a company's restructuring in the 1990s as an example.

[0190] The following is an example of the input file data: Table 1: Example of input file data;

[0191] After processing steps S11-S77, the system outputs metadata records as shown in Table 2: Table 2: Example of output metadata records;

[0192] The system successfully identified the historical institution name "XX State-owned Factory" and linked it to its restructured name "XX Limited Company" through a knowledge graph. The unit name identified in the seal was collaboratively extracted with the responsible person information in the main text to form complete metadata. When batch processing the enterprise's series of archives, the system automatically maintained the consistency of fields such as case number and archival group number, ensuring standardized indexing.

[0193] Through the intelligent processing of this invention, the time for extracting archival metadata is reduced from an average of 15 minutes per document by manual methods to 2-3 minutes, and the accuracy rate is increased from 85% by manual indexing to over 92%, thereby improving the efficiency and quality of archival management.

[0194] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An archival metadata intelligent extraction and indexing system, characterized by, The method comprises the following steps: A multi-modal preprocessing module is used to obtain original multi-modal archive data, and the original multi-modal archive data is preprocessed to obtain a structured multi-modal archive data set; A context entity recognition module is used to extract a context vector based on the structured multi-modal archive data set, and to perform entity recognition and standardization according to the context vector to obtain archive text data with labeled standardized entities; A cross-modal fusion module is used to perform confidence-weighted multi-modal fusion and logic verification based on the archive text data with labeled standardized entities and the structured multi-modal archive data set to obtain complete archive metadata records; An archive association module is used to perform association recognition and consistency detection between archives based on the complete archive metadata records to obtain a consistent archive metadata set; An intelligent indexing module is used to perform hierarchical intelligent indexing and quality feedback on the consistent archive metadata set to obtain a standardized indexing result; A quality evaluation module is used to perform metadata quality evaluation and active repair on the standardized indexing result to obtain an optimized metadata record set; A knowledge graph module is used to construct a time-series knowledge graph and intelligent retrieval service based on the optimized metadata records to obtain a dynamically evolving time-series archive knowledge graph and a semantic retrieval scheme.

2. The archival metadata intelligent extraction and indexing system of claim 1, wherein, The preprocessing of the original multi-modal archive data comprises the following steps: Quality detection is performed on the original archive scanned images, the image sharpness is calculated using a Laplacian operator, the noise level is evaluated using the difference between images before and after median filtering, and the tilt angle is calculated using a Hough straight line detection algorithm to identify the image edges; Self-adaptive enhancement processing is performed on images with a quality score lower than a preset threshold, including sharpness improvement, noise reduction processing, rotation correction, and contrast improvement; An end-to-end OCR recognition system that fuses visual features and language models is used for character extraction, a convolutional neural network is used to extract image visual features, a bidirectional long short-term memory network is used to model the sequence of feature maps, a connection time series classification algorithm is used for decoding, and an N-gram language model is introduced for post-processing of the recognition results; Characters with a confidence score lower than a first threshold are marked as low-confidence characters, and when the proportion of low-confidence characters exceeds a second threshold, the entire text is marked as high-uncertainty text.

3. The system of claim 1, wherein, The extraction of the context vector based on the structured multi-modal archive data set and the entity recognition and standardization according to the context vector comprise the following steps: Context elements are extracted from the structured multi-modal archive data set, the context elements are respectively encoded into embedding vectors and spliced to form a comprehensive context vector; The text is encoded, the context vector is mapped to the same dimension as the token embedding through linear transformation, and the context-enhanced token representation is obtained by adding the embedding of each token; Related historical entities are retrieved from the multi-period archive knowledge graph according to the context vector to form a candidate entity knowledge set; A knowledge-enhanced sequence labeling model is constructed for named entity recognition, a BiLSTM-CRF architecture is used, and an external knowledge feature enhancement model is introduced to enhance the model's ability; The recognized entities are standardized across periods, and the highest-scoring candidate entity is selected as the linking result based on the comprehensive name similarity, semantic similarity, and time consistency scores.

4. The system of claim 1, wherein, The confidence-weighted multi-modal fusion and logic verification comprises: A multi-modal fusion model is constructed to process information of text and image modalities, and the fusion weight is dynamically adjusted according to the confidence of each modality information; A bidirectional attention mechanism is used to identify the semantic association between different modal information, and the attention of text to image and the attention of image to text are calculated, and the image features are weighted and aggregated based on the attention weight; A joint entity and relation extraction model is constructed to identify entities and relations between entities, and an extended annotation system is used to convert the relation extraction task into a sequence labeling task, and a relation result in the form of a triple is output; The relation result in the form of a triple is mapped to a metadata field according to a predefined archive metadata template, a mapping rule of the triple relation type to the metadata field is established, and the field values that do not conform to the specification are automatically adjusted; A logic consistency reasoning engine is constructed to perform multi-dimensional verification on the metadata records, including time consistency verification, duty consistency verification, and regional consistency verification.

5. The archival metadata intelligent extraction and indexing system of claim 1, wherein, The association identification and consistency detection between archives comprises: An association identification mechanism between archives is constructed to identify explicit and implicit associations, and the total association score is calculated by integrating the explicit and implicit associations; A heterogeneous graph neural network is constructed to perform representation learning on the archive association graph, different types of relations are modeled respectively, and the node representation is updated layer by layer through multi-layer HGNN stacking, and each archive node learns an embedding vector that integrates multi-type relation information; The consistency of the metadata of the connected archive pairs in the archive association graph is detected, the support degree of each candidate value is calculated for each conflict field, the support degree is the sum of the occurrence times of the candidate value in all related archives weighted by the confidence, and a consistency conflict report is output; A multi-strategy is used for conflict resolution, including majority voting, confidence sorting, historical accuracy, and knowledge graph verification, and a standard value is selected by fusing the weighted scores of the four strategies; A constraint network is constructed to perform constraint propagation using a message passing mechanism.

6. The archival metadata intelligent extraction and indexing system of claim 1, wherein, The hierarchical intelligent indexing and quality feedback of the archive metadata set subject to consistency constraints comprises: A hierarchical classification model is constructed to classify archives by topics, the classification system is modeled as a tree structure, and a hierarchical constraint is introduced; A knowledge graph guided keyword extraction method is used to integrate TF-IDF scores and knowledge graph relevance, and the relevance of keywords and archive categories is considered; The extracted candidate labels are mapped to the standard words in the subject thesaurus, and the matching method includes exact matching, synonym matching, similar word matching, upper word expansion, and related word supplement; A reinforcement learning framework is established to continuously optimize the indexing strategy through user feedback on the indexing results, the indexing process is modeled as a Markov decision process, and a policy gradient algorithm is used to update the policy parameters; An indexing quality monitoring mechanism is established to statistically analyze the indexing acceptance rate of each archive type, the usage frequency and accuracy of each topic label.

7. The system of claim 1, wherein, The metadata quality evaluation and active repair of the standardized indexing result comprises: Five-dimensional quality indicators are used for evaluation, including completeness evaluation, accuracy evaluation, consistency evaluation, specification evaluation, and timeliness evaluation; The comprehensive quality score of the metadata record is calculated based on the evaluation results of the five dimensions, and the field-level quality score is calculated for each field; An active repair mechanism is used to automatically correct the problems found in the quality index evaluation; According to the comprehensive quality score, the metadata records are classified, and a quality classification identifier and processing suggestion are added to each metadata record.

8. The archival metadata intelligent extraction and indexing system of claim 1, wherein, The construction of the time-series knowledge graph and intelligent retrieval service based on the optimized metadata record includes: Incremental construction of the knowledge graph by extracting entity and relationship information from the optimized metadata record, and using an incremental update mechanism to avoid repeated construction of existing parts; Adding a time dimension to the knowledge graph to construct a time-series knowledge graph, and representing the time-series triplets in a five-tuple structure containing the head entity, relationship, tail entity, relationship start time, and relationship end time; Using an embedding-based completion method to predict potential relationships and discover implicit archive associations, and introducing time coding to process time-series triplets; Using graph neural networks for embedding learning, and stacking multiple GCNs to make node embeddings integrate information from multiple-hop neighbors; Query understanding, entity and intent recognition in the query, query expansion, including synonym expansion, hyponym expansion, and associated word expansion; Performing graph query, using a hybrid retrieval strategy that considers both precise matching and semantic similarity, and combining structured matching and semantic similarity calculation to obtain a comprehensive score; Building a learning ranking model to collect user interaction data and continuously optimize the retrieval ranking strategy.

9. The archival metadata intelligent extraction and indexing system of claim 7, wherein, The classification of the metadata records according to the comprehensive quality score includes: Metadata records with a comprehensive quality score greater than or equal to a first score threshold are classified as high-quality, and are directly stored for use without manual review; Metadata records with a comprehensive quality score between a second score threshold and the first score threshold are classified as medium-quality, and are stored after intelligent repair for manual inspection; Metadata records with a comprehensive quality score less than or equal to the second score threshold are classified as low-quality, and are stored after manual review.

10. A computer-readable storage medium, characterized in that, The computer-readable instructions are used to store a computer-readable instruction that can run a file metadata intelligent extraction and indexing system as claimed in any one of claims 1-9 when read by a computer.

Images