Deep learning based archival metadata intelligent extraction and quality control method

Abstract

This invention relates to the field of archival digitization technology, and more particularly to a method for intelligent extraction and quality control of archival metadata based on deep learning. This method uses a three-branch heterogeneous fusion network to perform layout analysis and text recognition on archival images, obtaining structured extraction results; it then performs named entity recognition on the extracted results to obtain candidate metadata; utilizing an archival domain knowledge graph, it performs graph reasoning and constraint propagation on the candidate metadata to generate verification metadata with confidence scores; finally, it refines the metadata based on the scores to obtain the final metadata. This invention achieves high-precision automated extraction of archival metadata and continuously optimizes the extraction and verification model through a feedback mechanism, effectively improving data quality and processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of archival digitization technology, and in particular to a method for intelligent extraction and quality control of archival metadata based on deep learning. Background Technology

[0002] Extracting archival metadata is a crucial step in the digital management of archives, and its accuracy directly impacts subsequent retrieval, analysis, and long-term preservation. Traditional methods for extracting archival metadata primarily rely on manual data entry or rule-based automated processing.

[0003] To improve efficiency, the industry commonly employs rule-based or template-based automated extraction techniques. These methods typically pre-define fixed layout areas or keyword rules, using optical character recognition (OCR) technology to extract text from specified locations as metadata. However, archival materials are diverse and have complex and varied formats, including tables of different formats, stamps covering text, handwritten annotations, and other interference. Fixed rule templates struggle to adapt to this diversity and lack robustness to layout shifts, stamp obstructions, and other issues, resulting in fragmented, incomplete, or noisy extraction results.

[0004] After extracting preliminary textual information, conventional methods lack effective quality control mechanisms. They typically only perform simple format checks or dictionary matching, failing to verify consistency from the overall logical and domain-specific perspectives of the archival content. For example, they cannot determine whether the extracted "date of creation" is later than the "date of issuance," nor can they identify whether the "institution name" matches the "archival classification number." This isolated data verification method makes it easy to miss erroneous or contradictory metadata, thereby affecting the quality and reliability of the entire archival database.

[0005] In summary, existing technologies lack sufficient extraction accuracy when processing complex format archives and lack the ability to perform internal logical verification based on domain knowledge. This makes it difficult to balance the automation and accuracy of metadata extraction, becoming a bottleneck restricting the progress of archive digitization. Summary of the Invention

[0006] This invention provides a method for intelligent extraction and quality control of archival metadata based on deep learning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for intelligent extraction and quality control of archival metadata based on deep learning, comprising:

[0008] The obtained archive images to be processed are analyzed for layout and text recognition by a three-branch heterogeneous fusion network, which includes a table recognition branch, a seal recognition branch and a text recognition branch. The fusion weight is dynamically calculated based on the regional confidence of each branch and the outputs of each branch are weighted and fused to obtain the structured extraction result.

[0009] Named entity recognition is performed on the structured extraction results to obtain a candidate metadata set;

[0010] A knowledge graph for the archival domain is constructed. The knowledge graph includes metadata field nodes and constraint relationship edges between representation fields, which are temporal constraints, classification constraints, and numerical constraints. The field values ​​in the candidate metadata set are constrained along the constraint relationship edges using a graph reasoning algorithm, and the satisfaction degree of the field value combination is calculated. A confidence score is generated based on the satisfaction degree to obtain a verification metadata set.

[0011] The final metadata set is obtained by correcting the verification metadata set based on the confidence score.

[0012] The final metadata set is output, and the relation edge weights of the knowledge graph are updated based on the feedback. At the same time, the fusion weights are adjusted based on the feedback of the identified erroneous samples.

[0013] Construct a knowledge graph for the archival domain. The knowledge graph includes metadata field nodes and constraint edges between representation fields, encompassing temporal, classification, and numerical constraints.

[0014] Mandatory constraint rules are parsed from the archives management standards, and these mandatory constraint rules are transformed into normative constraint edges and assigned fixed weights.

[0015] Statistical correlation patterns between fields are mined from historical archive data, and these patterns are transformed into statistical constraint edges and assigned adjustable weights.

[0016] Based on the dependency type between the fields involved in the normative constraint edge and the statistical constraint edge, the constraint relationship edge is classified into temporal constraint edge, classification constraint edge and numerical constraint edge; the temporal constraint edge represents the temporal dependency relationship between fields, the classification constraint edge represents the category affiliation or mutual exclusion relationship between fields, and the numerical constraint edge represents the numerical range or computational dependency relationship between fields.

[0017] The normative constraint edges and the statistical constraint edges are connected together to the metadata field nodes. Each metadata field node is connected to other field nodes through multiple constraint relationship edges of different types, thus forming a knowledge graph of the archival domain.

[0018] The field values ​​in the candidate metadata set are constrained along the constraint relationship edges using a graph reasoning algorithm, and the satisfaction degree of the field value combinations is calculated. A confidence score is generated based on the satisfaction degree to obtain the verification metadata set, including:

[0019] The graph reasoning algorithm starts from any field value in the candidate metadata set and propagates constraint information to adjacent field nodes along the constraint relationship edges connected to that field node. When propagating to a normal constraint edge, it determines whether the combination of field values ​​satisfies the mandatory constraint conditions corresponding to that edge to obtain a hard constraint judgment value. When propagating to a statistical constraint edge, it quantifies the degree of compliance of the combination of field values ​​with the constraint conditions corresponding to that edge based on the adjustable weight of that edge to obtain a weighted compliance degree.

[0020] After multi-hop constraint propagation, the satisfaction degree is obtained by combining the hard constraint judgment values ​​of all normative constraint edges with the weighted compliance calculation field values ​​of all statistical constraint edges. When the hard constraint judgment value of any normative constraint edge is not satisfied, the satisfaction degree is set to zero. When the hard constraint judgment values ​​of all normative constraint edges are satisfied, the weighted compliance values ​​of each statistical constraint edge are summed to obtain the satisfaction degree value.

[0021] The satisfaction score is normalized and mapped to a confidence score. Each combination of field values ​​in the candidate metadata set is labeled with its confidence score. The combination of field values ​​with confidence scores higher than a preset score threshold is selected to obtain the verification metadata set.

[0022] The final metadata set is obtained by correcting the verification metadata set based on the confidence score, including:

[0023] Identify target fields in the verification metadata set whose confidence scores are lower than a correction threshold, and locate the field nodes corresponding to the target fields from the knowledge graph;

[0024] Tracing backward from the target field node along the constraint relationship edge to the associated field node whose confidence score is higher than the correction threshold, the target field is deduced in reverse according to the constraint conditions of the connected constraint relationship edge. When the constraint relationship edge is a time series constraint edge, the time range of the target field is deduced based on the time value of the associated field. When the constraint relationship edge is a classification constraint edge, the category of the target field is deduced based on the category value of the associated field. When the constraint relationship edge is a numerical constraint edge, the numerical range of the target field is calculated based on the numerical value of the associated field and the constraint conditions.

[0025] When multiple constraint edges provide different correction candidate values, conflict detection and priority evaluation are performed on the correction candidate values, and the final correction value is selected as the correction value of the target field.

[0026] After updating the corrected value to the target field, the process propagates along the constraint relationship edges to check the associated fields. When the associated field conflicts with the corrected target field, the associated field is corrected accordingly to obtain the final metadata set.

[0027] When multiple constraint edges provide different candidate correction values, conflict detection and priority evaluation are performed on the candidate correction values, and the final correction value is selected as the correction value for the target field, including:

[0028] Conflict detection is performed on the correction candidate values ​​given for multiple constraint relationship edges; when the difference between different correction candidate values ​​exceeds the preset conflict threshold, it is determined that there is a conflict.

[0029] When conflicts exist, the overall confidence of each corrective candidate value is calculated by weighting the weight of each constraint edge and the confidence score of the corresponding associated field.

[0030] If the overall confidence difference of multiple corrected candidate values ​​is less than the preset difference threshold, then the corrected candidate value derived from the normative constraint edge is selected; if they are all derived from the statistical constraint edge, then the corrected candidate value with the highest overall confidence is selected.

[0031] When there is logical mutual exclusion between the candidate values ​​for correction, trace along the constraint relationship edge to the field node with a confidence score higher than the correction threshold to verify the rationality of each candidate value, and select the candidate value for correction that maintains constraint consistency with all fields with confidence scores higher than the correction threshold on the tracing path as the final correction value.

[0032] The weights of the relation edges in the knowledge graph will be corrected based on feedback, and the fusion weights will be adjusted based on feedback from incorrectly identified samples, including:

[0033] Record the field correction records of the final metadata set relative to the verification metadata set. The field correction records include the target field to be corrected, the associated fields involved in the reverse derivation, and the constraint relationship edges connecting the two.

[0034] The effect of each statistical constraint edge on field correction is statistically analyzed. When the combination of field values ​​connected by a statistical constraint edge is verified to be correct after correction, the adjustable weight of the statistical constraint edge is increased. When the verification is incorrect, the adjustable weight of the statistical constraint edge is decreased, thus completing the update of the relation edge weights of the knowledge graph.

[0035] By comparing the structured extraction results with the final metadata set, the image regions and their types corresponding to the corrected fields are identified, and the image regions are marked as misidentified samples and their corresponding branch types are recorded.

[0036] The incorrectly identified samples are input into the three-branch heterogeneous fusion network. The basic weight bias of the corresponding branch is reduced according to the type of region where the incorrect identification occurs. The fusion weight is adjusted by accumulating the statistical information of the incorrectly identified samples.

[0037] A second aspect of the present invention provides an electronic device, comprising:

[0038] processor;

[0039] Memory used to store processor-executable instructions;

[0040] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0041] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0042] This method employs a three-branch heterogeneous fusion network to simultaneously process tables, seals, and text regions in archival images. A dynamic weighted fusion mechanism adaptively integrates multimodal recognition results, significantly improving the accuracy of structured extraction from archives with complex layouts. The named entity recognition module extracts candidate metadata fields and values ​​from the structured results, laying the foundation for subsequent quality control. This method effectively overcomes the high false negative and false positive rates of traditional single recognition models in mixed layouts, achieving efficient and accurate digitization of archival image information.

[0043] The core advantage of this method lies in its constraint reasoning mechanism based on a knowledge graph in the archival domain. The knowledge graph defines the temporal, categorical, and numerical constraint relationships between metadata fields. The graph reasoning algorithm propagates and verifies candidate metadata along the constraint edges, calculating the overall satisfaction of field value combinations and generating a confidence score. This process can automatically discover and identify untrustworthy data with logical contradictions, format abnormalities, or violations of business rules, upgrading quality control from isolated field verification to global correlation verification, significantly improving the inherent consistency and business compliance of metadata.

[0044] By using confidence scoring to drive intelligent correction of the verification metadata set, this method can automatically correct some inferable errors or fill in reasonable missing values, outputting a high-quality final metadata set. The entire process forms a closed loop of "identification-verification-correction," with correction data fed back to the knowledge graph to optimize constraint weights, and error samples fed back to the front-end network to adjust fusion weights, achieving continuous self-iteration and enhancement of performance. This method not only ensures the extraction quality of metadata from individual archives but also continuously improves its adaptability and robustness to complex archive types through a continuous learning mechanism, providing a reliable, fully automated solution for large-scale archive digitization projects. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a deep learning-based intelligent extraction and quality control method for archival metadata.

[0046] Figure 2 Flowchart for metadata confidence correction and constraint propagation. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0049] Figure 1 This is a flowchart illustrating the intelligent extraction and quality control method for archival metadata based on deep learning, as described in an embodiment of the present invention. Figure 1 As shown, the intelligent extraction and quality control methods for archival metadata based on deep learning include:

[0050] The obtained archive images to be processed are analyzed for layout and text recognition by a three-branch heterogeneous fusion network, which includes a table recognition branch, a seal recognition branch and a text recognition branch. The fusion weight is dynamically calculated based on the regional confidence of each branch and the outputs of each branch are weighted and fused to obtain the structured extraction result.

[0051] Named entity recognition is performed on the structured extraction results to obtain a candidate metadata set;

[0052] Construct a knowledge graph for the archival domain. The knowledge graph includes metadata field nodes and constraint relationship edges between representation fields, including temporal constraints, classification constraints, and numerical constraints.

[0053] The field values ​​in the candidate metadata set are constrained along the edge of the constraint relationship by using a graph reasoning algorithm, and the satisfaction degree of the field value combination is calculated. A confidence score is generated based on the satisfaction degree to obtain the verification metadata set.

[0054] The final metadata set is obtained by correcting the verification metadata set based on the confidence score.

[0055] The final metadata set is output, and the relation edge weights of the knowledge graph are updated based on the feedback. At the same time, the fusion weights are adjusted based on the feedback of the identified erroneous samples.

[0056] The three-branch heterogeneous fusion network includes:

[0057] A shared feature extraction layer is used to extract the low-level feature map of the archive image using a multi-scale convolutional encoder;

[0058] The table recognition branch performs line segment detection and cell segmentation based on the underlying feature map, and outputs the table area, table structure and text content within the table.

[0059] The seal recognition branch performs circular region detection and seal text recognition based on the underlying feature map, and outputs the seal area and seal content.

[0060] The text recognition branch performs text line detection and character recognition based on the underlying feature map, and outputs the text region and text content.

[0061] The weighted fusion is implemented through an adaptive fusion module. The adaptive fusion module calculates the fusion weight based on the regional confidence of each branch, adaptively adjusts the weight allocation strategy according to the complexity of different image regions, makes fusion decisions for regions with overlaps or conflicts, and determines the final region type classification.

[0062] The adaptive fusion module also includes a cross-branch attention mechanism, which uses the intermediate layer features of the stamp recognition branch to guide the text recognition branch to enhance the recognition ability of text covered by the stamp or located in the seam.

[0063] The structured extraction results are composed of the regions, their type identifiers, spatial location information, and identification content after fusion decision-making.

[0064] For example, after the archival image enters the processing flow, it first undergoes preliminary feature encoding through a shared feature extraction layer. This shared feature extraction layer employs a multi-scale convolutional encoder architecture, containing four convolutional branches with different receptive fields, and setting the convolutional kernel sizes to 3×3, 5×5, 7×7, and 9×9, respectively. Each branch independently extracts features from the input archival image. Each convolutional branch contains three convolutional operations with channel numbers of 64, 128, and 256, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The feature maps extracted by the four branches are concatenated by channel dimension and then fused using a 1×1 convolution, compressing the concatenated 1024-dimensional features to 512 dimensions, forming a low-level feature map with rich multi-scale information. This feature map simultaneously contains fine-grained local texture information and overall structural information from a larger receptive field.

[0065] The table recognition branch performs a dedicated table parsing task based on this underlying feature map. First, line segment detection is performed using the Hough transform algorithm to detect horizontal and vertical lines in the image, setting a line segment length threshold of 30 pixels and an angle tolerance of 2 degrees to detect the table's border lines. Then, cell segmentation is performed based on the detected line segments, determining the coordinates of the four vertices of each cell through the intersection points of the line segments, dividing the table region into several independent cells. For each cell region, the corresponding feature map region is cropped and fed into the text recognition submodule. This submodule uses a CRNN network structure, containing five convolutional layers and two bidirectional LSTM layers, to output the text content within the cell. The output of the table recognition branch includes the spatial coordinates of the table, the number of rows and columns, the bounding box coordinates of each cell, and the text content within the cell. Simultaneously, a confidence score is calculated for each table region, based on a comprehensive evaluation of the response strength of line segment detection and the character-level probability of text recognition, with a value ranging from 0 to 1.

[0066] The seal recognition branch focuses on the localization and recognition of circular or elliptical seals. It uses the Hough circle detection algorithm to detect circular outlines on the underlying feature map, setting the center positioning accuracy to 5 pixels and the radius search range to 20 to 80 pixels to extract candidate seal regions. For the detected candidate regions, a color space conversion is performed, transforming the RGB image to HSV space, extracting the red channel component, and setting the hue threshold range to 0 to 15 and 165 to 180, and the saturation threshold to be greater than 100, to filter out red seal regions. Considering the special characteristics of the seal region, polar coordinate transformation is used to unfold the circular seal into a rectangular strip. The width of the unfolded image is set to 360 pixels corresponding to 360 degrees, and the height is the seal radius. The unfolded image is fed into a dedicated seal OCR model, which uses an attention-based sequence recognition network capable of handling arc-shaped text. The seal recognition branch outputs the seal's center coordinates, radius, seal text content, and seal type classification label. Confidence calculation combines the fitting error of the circular detection, color matching degree, and text recognition probability.

[0067] The text recognition branch processes regular text content in archival images. First, text line detection is performed, using the PixelLink algorithm to predict the probability of each pixel belonging to text and its connectivity with neighboring pixels. Text instance segmentation results are generated using a probability threshold of 0.6 and a connectivity threshold of 0.5, extracting independent text line regions. Each text line region is corrected using perspective transformation to adjust tilted or curved text lines to a horizontal orientation. The corrected text line image is normalized to a size of 32 pixels in height and variable width, and then input into the character recognition network. The character recognition network uses ResNet-34 as a feature extractor, followed by an attention decoder, outputting the character sequence within the text line and the recognition probability of each character. The text recognition branch output includes the minimum bounding rectangle coordinates of each text line, the text content, an estimated font size, and a line-level confidence score, calculated from the average of the character-level probabilities.

[0068] After completing their respective recognition tasks, the three branches send their results to the adaptive fusion module. This module first collects the region outputs from each branch, including table regions, stamp regions, and text regions, as well as the confidence scores for each region. For spatially overlapping regions, the degree of conflict is determined based on the overlap area ratio; a significant conflict is considered to exist when the overlap area exceeds 30% of the smaller region's area. The fusion weight calculation employs a dynamic normalization strategy. For overlapping regions, the fusion weight wi of each branch is calculated using the formula... Calculate, where c i c j For branch confidence, a i a j For the area of ​​the region, branches with high confidence and large area are given higher weights. For the common scenario of stamps covering text in archival images, the fusion decision rule is set as follows: when the stamp area overlaps with the text area, if the stamp confidence is higher than 0.8, the area is classified as a stamp, but the covered text content is retained as additional information; if the stamp confidence is lower than 0.5, the area is classified as a text.

[0069] Cross-branch attention mechanisms play a synergistic enhancement role in the fusion process. An intermediate feature map, containing structural features of the seal, is extracted from the third convolutional layer output of the seal recognition branch. This feature map is then adjusted to the same size as the feature map from the text recognition branch using bilinear interpolation. A spatial attention map is then calculated, with attention weights obtained by maximizing the seal feature map in the channel dimension and activating it with a sigmoid function. This attention map is then multiplied element-wise with the feature map from the text recognition branch, allowing the text recognition network to acquire additional contextual information in the seal-covered area, improving the accuracy of recognizing partially occluded characters. For seals spanning multiple lines of text, the attention mechanism can also identify seals that cross multiple text lines, associating the separated seal portions at the feature level to assist in the complete recognition of the seal content.

[0070] After the fusion decision is completed, a structured extraction result is output. This structured extraction result is organized in JSON format and includes the type identifier of each region (table, seal, or text), spatial location information (coordinates of the top left and bottom right corners of the smallest bounding rectangle), and recognized content (the cell text matrix of the table, the text sequence of the seal, or the string of the text line).

[0071] Named entity recognition is performed on the structured extraction results, using a BERT-based sequence labeling model to identify entities such as file number, date, organization name, and security classification. The identified content from tables, seals, and text is fed into the model, which outputs entity type labels. Entities and their text values ​​are mapped to field-value pairs, such as mapping "May 15, 2023" to "Document Date: 2023-05-15". All mapping results are then aggregated to form a candidate metadata set.

[0072] Construct a knowledge graph for the archival domain. The knowledge graph includes metadata field nodes and constraint edges between representation fields, encompassing temporal, classification, and numerical constraints.

[0073] Mandatory constraint rules are parsed from the archives management standards, and these mandatory constraint rules are transformed into normative constraint edges and assigned fixed weights.

[0074] Statistical correlation patterns between fields are mined from historical archive data, and these patterns are transformed into statistical constraint edges and assigned adjustable weights.

[0075] Based on the dependency type between the fields involved in the normative constraint edge and the statistical constraint edge, the constraint relationship edge is classified into temporal constraint edge, classification constraint edge and numerical constraint edge.

[0076] The temporal constraint edge represents the temporal dependency between fields, the classification constraint edge represents the category affiliation or mutual exclusion between fields, and the numerical constraint edge represents the numerical range or computational dependency between fields.

[0077] The normative constraint edges and the statistical constraint edges are connected together to the metadata field nodes. Each metadata field node is connected to other field nodes through multiple constraint relationship edges of different types, thus forming a knowledge graph of the archival domain.

[0078] For example, in constructing a knowledge graph for the archival field, the first step is to parse mandatory constraint rules from archival management standards. Archival management standards include, but are not limited to, the "Rules for Archival Description," the "Standards for Electronic Document Archiving and Electronic Archives Management," and industry-specific archival standards. A method combining rule template matching and semantic analysis is used to automatically parse the standard text. Specifically, the standard clauses are input into a natural language processing module, which uses dependency parsing to identify the subject, predicate, and constraints in the clauses, extracting mandatory constraint rules in the form of "Field A must be earlier than Field B" or "The value of Field C must belong to set D." For instance, from the standard clause "The document creation time should be earlier than the archiving time," a temporal constraint rule is extracted, where "document creation time" is the source field node, "archiving time" is the target field node, and the constraint type is identified as a temporal sequence. Each extracted mandatory constraint rule is transformed into a standard constraint edge in the knowledge graph. This type of edge connects two or more related metadata field nodes and is assigned a fixed weight between 0.95 and 1.0, reflecting its enforceable attribute. The fixed weights are set based on the mandatory level of the specification. The "must" type constraints are assigned a weight of 1.0, the "should" type constraints are assigned a weight of 0.98, and the "appropriate" type constraints are assigned a weight of 0.95.

[0079] Unlike mandatory constraint rules, historical archival data contains a large number of implicit statistical association patterns between fields. A historical dataset containing at least 100,000 complete archival records was selected as the mining object, and frequent pattern mining and association rule learning algorithms were used to extract co-occurrence patterns between fields. Specifically, for each pair of fields, their joint probability distribution and conditional probability distribution were calculated. For numerical fields, Pearson correlation coefficient or Spearman rank correlation coefficient were used to measure the correlation strength; for categorical fields, mutual information or chi-square test was used to measure the degree of association; for time-series fields, the distribution characteristics of time differences were statistically analyzed, and significant time-series dependency patterns were identified. Taking the association between archival retention period and document security level as an example, statistical analysis revealed that the permanent retention rate of "top secret" level documents reached 87%, while the permanent retention rate of "ordinary" level documents was only 12%. This statistical significance supports the establishment of statistical constraint edges between the two fields. The mined statistical association patterns were transformed into statistical constraint edges, and their weights were dynamically assigned based on statistical significance, with values ​​ranging from 0.5 to 0.9. Specifically, the association strength is mapped to a weight range using a sigmoid function. Strong associations (with an absolute correlation coefficient greater than 0.7 or mutual information greater than 0.5) are assigned a weight of 0.85 to 0.9, moderate associations are assigned a weight of 0.7 to 0.85, and weak associations are assigned a weight of 0.5 to 0.7. This adjustable weight mechanism allows for fine-tuning and optimization based on actual validation feedback.

[0080] After obtaining the canonical and statistical constraint edges, the constraint edges are classified according to the dependency type between the fields involved. The dependency type refers to the semantic relationship type expressed by the constraint, including three types: temporal dependency, category dependency, and numerical computation dependency. Based on the semantic content of the constraint conditions, constraint edges are classified into temporal constraint edges, categorical constraint edges, and numerical constraint edges. Temporal constraint edges represent the temporal dependency relationship between fields, mainly involving date or timestamp fields. Typical temporal constraints include chained temporal dependencies such as document drafting time earlier than issuance time, issuance time earlier than archiving time, and archiving time earlier than open access time. During graph construction, a time difference range attribute is added to the temporal constraint edges. For example, the constraint "issuance time should be within 7 to 90 days after drafting time" is encoded as a temporal constraint edge, with edge attributes including a minimum time difference of 7 days and a maximum time difference of 90 days. Categorical constraint edges represent the category affiliation or mutual exclusion relationship between fields, mainly involving enumerated or categorical fields. Category attribution relationships, such as "When the document category is 'Contract Agreement,' the retention period should be permanent or long-term," establish a classification constraint edge from the document category node to the retention period node. The edge attribute records the condition category value "Contract Agreement" and the set of allowed result category values. Mutually exclusive relationships, such as "Public and Top Secret classifications are mutually exclusive," establish a self-loop classification constraint edge from the classification level field node. The edge attribute marks the mutually exclusive category pair. Numerical constraint edges represent the numerical range or computational dependency between fields, primarily involving numeric fields. Numerical range constraints, such as "The number of pages should be greater than 0 and less than or equal to 10000," establish a self-loop numerical constraint edge from the page number node. The edge attribute records the lower bound 0 and the upper bound 10000. Computational dependency relationships, such as "The total number of items equals the sum of the number of items in each category," establish a numerical constraint edge from the set of category item count nodes to the total item count node. The edge attribute marks the computational relationship type as summation.

[0081] All normative and statistical constraint edges are connected to the metadata field nodes to form a complete knowledge graph topology. Each metadata field node is connected to other field nodes through multiple constraint edges of different types. The same field node can be connected to temporal constraint edges, classification constraint edges, and numerical constraint edges simultaneously, forming a multi-dimensional constraint network. Each metadata field node, as a vertex in the graph, contains attribute information such as field name, data type, and value range. For example, the "Archived Time" node has a date data type attribute and a value range attribute from January 1, 1900 to the current date. Each metadata field node is connected to other field nodes through multiple constraint edges of different types, forming a multi-dimensional constraint network. Taking the "Document Classification" node as an example, this node is connected to the "Decryption Time" node through a temporal constraint edge, with the constraint "Decryption time should be later than archive time"; it is connected to the "Retention Period" node through a classification constraint edge, with the constraint "Top secret documents should ideally be retained permanently"; and it is connected to the "Usage Frequency" node through a statistical constraint edge, with the constraint "The higher the classification level, the lower the usage frequency". The entire knowledge graph forms a directed weighted graph structure containing 50 to 200 metadata field nodes and 200 to 1000 constraint relationship edges. The directionality of the edges reflects the direction of constraint propagation, and the weight of the edges reflects the credibility or importance of the constraints. To improve the efficiency of graph reasoning, the knowledge graph undergoes topological sorting preprocessing, strongly connected components are identified as cyclic constraint groups, and a hierarchical index structure of field nodes is established. The knowledge graph is stored in a graph database, supporting efficient adjacency queries and path search operations, providing the infrastructure for subsequent constraint propagation and satisfaction calculation.

[0082] The field values ​​in the candidate metadata set are constrained along the constraint relationship edges using a graph reasoning algorithm, and the satisfaction degree of the field value combinations is calculated. A confidence score is generated based on the satisfaction degree to obtain the verification metadata set, including:

[0083] The graph reasoning algorithm starts from any field value in the candidate metadata set and propagates constraint information to adjacent field nodes along the constraint relationship edges connected to that field node. When propagating to a normal constraint edge, it determines whether the combination of field values ​​satisfies the mandatory constraint conditions corresponding to that edge to obtain a hard constraint judgment value. When propagating to a statistical constraint edge, it quantifies the degree of compliance of the combination of field values ​​with the constraint conditions corresponding to that edge based on the adjustable weight of that edge to obtain a weighted compliance degree.

[0084] After multi-hop constraint propagation, the satisfaction degree is obtained by combining the hard constraint judgment values ​​of all normative constraint edges with the weighted compliance calculation field values ​​of all statistical constraint edges. When the hard constraint judgment value of any normative constraint edge is not satisfied, the satisfaction degree is set to zero. When the hard constraint judgment values ​​of all normative constraint edges are satisfied, the weighted compliance values ​​of each statistical constraint edge are summed to obtain the satisfaction degree value.

[0085] The satisfaction score is normalized and mapped to a confidence score. Each combination of field values ​​in the candidate metadata set is labeled with its confidence score. The combination of field values ​​with confidence scores higher than a preset score threshold is selected to obtain the verification metadata set.

[0086] For example, after obtaining the candidate metadata set, it is necessary to perform constraint validation on the field values ​​to ensure that the extraction results conform to the business rules of the archival domain. The core of the graph reasoning algorithm lies in using the constraint relationship edges already constructed in the knowledge graph to transform the archival business rules into a computable constraint propagation process.

[0087] A field value, such as the "document date" field, is selected as the starting node from the candidate metadata set. This field corresponds to a specific node in the knowledge graph, which is connected to other field nodes through multiple constraint edges. Constraint edges are divided into two categories: normative constraint edges and statistical constraint edges. Normative constraint edges represent mandatory rules in archival management, such as the temporal constraint that "the document date must not be later than the archiving date," which does not allow any violations. Statistical constraint edges represent empirical patterns mined from historical archival data, such as "the archiving cycle for a certain type of archive is usually thirty to ninety days," which allows for a certain degree of deviation.

[0088] The constraint propagation process unfolds along the edges directly connected to the current field node. When a canonical constraint edge is encountered, the constraint condition represented by that edge is extracted. Taking a time-series constraint as an example, if the current field "File Date" has a value of "May 15, 2023", and the canonical constraint edge connected to it points to the "Archived Date" field, with the constraint condition "File Date should be earlier than Archived Date", then the value of the "Archived Date" field is searched from the candidate metadata set. Assuming it is "May 10, 2023", this combination of field values ​​is determined to violate the time-series constraint, and the hard constraint judgment value is recorded as unsatisfied. If the "Archived Date" value is "June 20, 2023", then this combination satisfies the time-series constraint, and the hard constraint judgment value is recorded as satisfied.

[0089] To handle statistical constraint edges, an edge weighting mechanism needs to be introduced. Each statistical constraint edge is assigned an adjustable weight during knowledge graph initialization based on historical data analysis results. This weight reflects the importance of the constraint in actual archival operations. For example, the weight value for the statistical constraint edge "archiving cycle" is w. cycle This characterizes the influence of the constraint in the satisfaction calculation. It calculates the degree to which the combination of field values ​​meets the constraint condition when the constraint propagates to that edge. Assume that the statistical constraint specifies a reasonable range for the archiving period as [d]. min d max The interval between the actual extracted file date and the archive date is d. actual If d actualIf the edge falls within a reasonable range, its conformity is 1; if it deviates from the range, the conformity after attenuation is calculated based on the deviation distance. The attenuation function can be either Gaussian or linear; here, we take linear attenuation as an example. When d... actual When the upper limit is exceeded, the degree of conformity is calculated as follows: , where d tolerance This is the allowable deviation tolerance. Multiply this compliance level by the edge weight w. cycle We obtain the weighted compliance degree of the statistical constraint edge.

[0090] The constraint propagation process is not limited to a single hop but extends along the graph structure in multiple hops. Starting from the initial field node, it visits directly adjacent nodes in sequence, and then continues propagating outward from these adjacent nodes. For example, after propagating from "Document Date" to "Archived Date," it can continue propagating from "Archived Date" to the "Retention Period" field to verify whether the classification constraint "the retention period of a certain type of archive should be permanent" is satisfied. The depth of multi-hop propagation is usually set to 2 to 3 hops to avoid excessive computational complexity. During the propagation process, the hard constraint judgment value is recorded for each traversed normative constraint edge, and the weighted compliance degree is accumulated for each traversed statistical constraint edge.

[0091] After completing multi-hop constraint propagation, the satisfaction level is calculated by integrating all collected constraint decision information. First, the hard constraint decision values ​​of all normative constraint edges are checked. If any edge's decision value is unsatisfied, it indicates that the combination of field values ​​violates mandatory business rules; in this case, the satisfaction level is directly set to zero, indicating that the combination is completely unacceptable. If all normative constraint edge decision values ​​are satisfied, the contribution of statistical constraints is further calculated. The weighted compliance levels of all statistical constraint edges are summed to obtain a value S. stat This value reflects the compliance of the field value combination with empirical rules. The satisfaction score is S. stat .

[0092] To facilitate subsequent screening decisions, the satisfaction scores are normalized and mapped to confidence scores. Maximum value normalization is used; firstly, the maximum satisfaction score S of all field value combinations in the candidate metadata set is calculated. max Then divide the satisfaction value of each combination by S. max This yields a confidence score ranging from 0 to 1. If a Sigmoid mapping is used, the satisfaction score is mapped to the 0-1 interval using a function, preserving the continuity of the numerical distribution.

[0093] For each combination of field values ​​in the candidate metadata set, the constraint propagation and satisfaction calculation process described above is repeated to assign a confidence score to all combinations. For example, for a candidate combination containing multiple fields such as "file date," "archive date," and "retention period," there may be multiple permutations of values. Each permutation is evaluated using a graph reasoning algorithm to obtain a corresponding confidence score. To filter out reliable metadata, a preset scoring threshold θ is set. conf This threshold is typically determined based on the archive quality requirements, for example, set to 0.75. Iterate through all field value combinations, retaining those with confidence scores higher than θ. conf The combinations of scores below a threshold are discarded. The resulting combinations of field values ​​constitute a validation metadata set. This set of metadata has high credibility at the business rule level, providing reliable input for subsequent quality corrections.

[0094] The final metadata set is obtained by correcting the verification metadata set based on the confidence score, including:

[0095] Identify target fields in the verification metadata set whose confidence scores are lower than a correction threshold, and locate the field nodes corresponding to the target fields from the knowledge graph;

[0096] Tracing backward from the target field node along the constraint relationship edge to the associated field node whose confidence score is higher than the correction threshold, the target field is deduced in reverse according to the constraint conditions of the connected constraint relationship edge. When the constraint relationship edge is a time series constraint edge, the time range of the target field is deduced based on the time value of the associated field. When the constraint relationship edge is a classification constraint edge, the category of the target field is deduced based on the category value of the associated field. When the constraint relationship edge is a numerical constraint edge, the numerical range of the target field is calculated based on the numerical value of the associated field and the constraint conditions.

[0097] When multiple constraint edges provide different correction candidate values, conflict detection and priority evaluation are performed on the correction candidate values, and the final correction value is selected as the correction value of the target field.

[0098] After updating the corrected value to the target field, the process propagates along the constraint relationship edges to check the associated fields. When the associated field conflicts with the corrected target field, the associated field is corrected accordingly to obtain the final metadata set.

[0099] Combination Figure 2Metadata Confidence Correction and Constraint Propagation Flowchart: When correcting the verification metadata set, the first step is to establish a confidence score correction threshold system. This threshold is set differently based on different file types and field importance. For example, for key fields such as file number and creation date, the correction threshold can be set to 0.85, while for auxiliary fields such as remarks and notes, the correction threshold can be set to 0.70. The process involves iterating through each field in the verification metadata set and its corresponding confidence score, and using numerical comparison to identify all target fields with scores below the correction threshold. These target fields may have insufficient confidence due to poor original image quality, limited text recognition accuracy, or misjudgment by the entity recognition model.

[0100] Once the target field requiring correction is identified, the node corresponding to that field is quickly located using the knowledge graph's node indexing mechanism. The knowledge graph uses field names as node identifiers, and each node stores attribute information such as the field's data type, value range, and regular expression pattern. Starting from the target field node, a reverse tracing is performed along the incoming edges, where incoming edges indicate that the current field is constrained by other fields. During the tracing process, related field nodes with confidence scores higher than the correction threshold are selected; the field values ​​of these high-confidence nodes serve as reliable evidence for the reverse derivation.

[0101] For time-constrained edges, the edge attributes store the time difference range and the chronological relationship. Assuming the target field is "archiving date" and the associated field is "creation date," the time-constrained edge between them defines that the archiving date must be later than the creation date, with the time difference typically between 3 and 24 months. When the confidence level of the creation date is 0.92 and the value is "March 15, 2022," it can be deduced that the archiving date should fall within the time range of June 15, 2022 to March 15, 2024. If the original identified archiving date is "December 10, 2021," it clearly violates the time-constraint, and a reasonable value needs to be reselected within the derived range. During the reselection, other identification candidates for that region are prioritized from the structured extraction results (such as multiple candidate results from OCR output). If a date value falls within the derived range among the candidates, the one with the highest confidence level is selected; otherwise, a recommended value is generated within the derived range based on the median archiving period of similar historical archives.

[0102] For classification constraint edges, the edge attributes include the category mapping relationship and the classification hierarchy. For example, the classification constraint edge between "archive category" and "retention period" stipulates that when the archive category is "permanent," the retention period must be "permanent," and when the archive category is "long-term," the retention period can be either "30 years" or "50 years." When the confidence level of the archive category is 0.89 and its value is "permanent," while the confidence level of the retention period is only 0.62 and its value is "10 years," the mapping rule of the classification constraint edge directly dictates that the retention period should be corrected to "permanent." Classification constraints can also handle multi-level classification cases, such as "institutional classification" determining "department code," and the department code further determining the "archive number prefix."

[0103] For numerical constraint edges, the edge attributes define the numerical calculation relationship and value range constraints between fields. A typical scenario includes the relationship between "page number" and "number of scanned images." Normally, these should be equal, or the number of scanned images should be slightly greater than the number of pages. When the confidence level of the page number field is 0.80 and its value is 125, while the confidence level of the number of scanned images is 0.95 and its value is 128, it can be considered that the page number field may have an identification error. It should be corrected to 128 or kept at 125, and the images should be labeled to include the cover and back cover. Numerical constraint edges can also express summation relationships, such as "total amount" should be equal to the sum of all detailed amounts. By summing the detailed amounts with high confidence, the total amount field can be verified or corrected.

[0104] When a target field yields different candidate values ​​for correction through multiple constraint relationships, conflict detection and priority evaluation are required. Conflict detection first determines whether there are logical contradictions among the candidate values. For example, if the date derived from a time-series constraint is "May 2023," while the date range derived from a classification constraint is "the entire year of 2022," there is a clear conflict. Priority evaluation comprehensively considers the weight of the constraint edges, the confidence level of the associated fields, and the reliability of the constraint type. Generally, numerical constraints have higher priority than classification constraints, and classification constraints have higher priority than time-series constraints, because numerical relationships are the most deterministic, while time-series relationships have greater flexibility. A comprehensive score S is calculated for each candidate correction value. i S i By the constraint edge weight w e Confidence level of related fields c f and constraint type coefficient k t Multiply the values ​​and select the candidate with the highest score as the final correction value; w e The weight is determined by the number of nodes connected by the constraint edge; the more nodes connected, the higher the weight, with a value ranging from 0.3 to 1.0. f The confidence score of the corresponding identification branch is used directly, with a value range of 0-1; k tThe coefficients are preset according to the importance of the constraints, with mandatory constraints (such as required fields and format validation) having a coefficient of 1.0, advisory constraints (such as logical consistency checks) having a coefficient of 0.7, and weak suggestion constraints (such as common value suggestions) having a coefficient of 0.5.

[0105] After correcting the target field, it's necessary to check the impact of the correction on related fields. Starting from the node of the corrected field, traverse all related fields constrained by that field along the outgoing edges. For each related field, recalculate the constraint satisfaction between it and the corrected target field. If a related field originally had high confidence but now conflicts with the corrected target field (e.g., the corrected archiving date is earlier than an approval date), then a linked correction is required for that related field. Linked corrections use an iterative propagation mechanism; each correction may trigger new constraint conflicts, continuing until there are no constraint violations among all fields or the maximum number of iterations is reached. During the iteration process, a correction history is maintained to prevent deadlocks caused by cyclic corrections.

[0106] After the corrections are completed, a final metadata set is constructed. Each field in this set is labeled with information such as the final confidence score, whether it has been corrected, and the constraint path on which the correction was based. This additional information provides a traceable basis for decision-making during subsequent manual review, allowing reviewers to quickly locate the corrected fields and check the rationality of the correction logic.

[0107] When multiple constraint edges provide different candidate correction values, conflict detection and priority evaluation are performed on the candidate correction values, and the final correction value is selected as the correction value for the target field, including:

[0108] Conflict detection is performed on the corrected candidate values ​​given for multiple constraint edges;

[0109] When the difference between different correction candidate values ​​exceeds a preset conflict threshold, a conflict is determined to exist;

[0110] When conflicts exist, the overall confidence of each corrective candidate value is calculated by weighting the weight of each constraint edge and the confidence score of the corresponding associated field.

[0111] If the combined confidence difference of multiple correction candidate values ​​is less than the preset difference threshold, then the correction candidate value derived from the normative constraint edge is selected.

[0112] If all are derived from statistical constraints, then the corrected candidate value with the highest overall confidence is selected.

[0113] When there is logical mutual exclusion between the candidate values ​​for correction, trace along the constraint relationship edge to the field node with a confidence score higher than the correction threshold to verify the rationality of each candidate value, and select the candidate value for correction that maintains constraint consistency with all fields with confidence scores higher than the correction threshold on the tracing path as the final correction value.

[0114] For example, in the process of intelligent extraction of archival metadata, when different constraint relationship edges in the knowledge graph provide different correction suggestions for the same field, a systematic conflict resolution mechanism needs to be established to ensure the accuracy of the correction value.

[0115] For a target field to be corrected, assume there are multiple association constraints from temporal, categorical, and numerical constraints. Temporal constraints may provide correction suggestions based on date order, categorical constraints provide correction candidates based on domain standards, and numerical constraints deduce possible values ​​based on statistical patterns. When discrepancies arise between these candidate correction values ​​from different sources, a conflict detection procedure is first executed. For numerical fields, the relative deviation between different candidate correction values ​​is calculated; a conflict is determined when the relative deviation between any two candidate values ​​exceeds a preset conflict threshold. For textual fields, the difference between candidate values ​​is measured using edit distance or semantic similarity; a conflict is also determined when the difference exceeds a threshold. This conflict threshold is set differently for different field types; for example, a relative deviation of 5% is set for year fields, and an edit distance of 3 is set for document classification number fields.

[0116] Once a conflict is confirmed, a priority evaluation process is initiated. Each constraint edge stores a weight parameter in the knowledge graph, reflecting the reliability of the constraint in historical validation. Simultaneously, each source field associated with the target field has a confidence score, calculated during constraint propagation in the graph reasoning phase. The calculation of the overall confidence score considers both the constraint edge weights and the source field confidence scores. Specifically, for a given corrected candidate value, its overall confidence score equals the product of the constraint edge weights used to derive that value and the confidence score of the field of the starting node of that edge. When multiple constraint edges jointly support the same corrected candidate value, the maximum overall confidence score of all supporting paths is taken as the final overall confidence score of that candidate value. In this way, a quantified confidence index is assigned to each corrected candidate value.

[0117] After obtaining the overall confidence level of each candidate correction value, their values ​​are compared to determine the final correction value. In most cases, the candidate value with the highest overall confidence level is directly selected as the correction result. However, in certain specific scenarios, additional judgment rules need to be introduced. When the difference in the overall confidence level of the top two candidate correction values ​​is less than a preset difference threshold, it indicates that there is no statistically significant difference in their credibility. In this case, the type attribute of the constraint needs to be considered. The constraint relationship edges in the knowledge graph are divided into two categories: normative constraint edges and statistical constraint edges. Normative constraint edges represent mandatory constraints established based on archival industry standards or management regulations, while statistical constraint edges are based on probabilistic patterns obtained from historical data mining. When the overall confidence levels are similar, the candidate correction value derived from the normative constraint edge is preferred because normative constraints have stronger certainty and authority. If both candidate values ​​are derived from the statistical constraint edge, the one with the higher overall confidence level is still selected. This hierarchical decision-making mechanism ensures that domain norms are followed first when there is greater uncertainty.

[0118] In some complex scenarios, different candidate correction values ​​may have logically mutually exclusive relationships. For example, for the classification field of a file, one candidate correction value might be "Secret" and another "Public," and these two attributes cannot logically be true simultaneously. In this case, simply relying on the overall confidence level may not lead to a correct judgment, and further verification of the rationality of each candidate value is required. The verification process is achieved by tracing the constraint relationship edges. Starting from the target field node, the process traces back along the constraint edges related to each candidate correction value to the source field node, checking the confidence scores of all field nodes on the tracing path. For field nodes with confidence scores higher than the correction threshold, their values ​​are considered relatively reliable and should maintain constraint consistency with the finally selected correction value. Specifically, for each candidate correction value, it is checked whether it satisfies the corresponding constraint relationship with all high-confidence fields on the tracing path. If the constraint relationship between a candidate value and a high-confidence field on the path is not valid, the rationality of the candidate value is questioned. In this way, each candidate correction value is verified one by one, and finally, the candidate value that maintains constraint consistency with all high-confidence associated fields is selected as the correction result.

[0119] During the traceability verification process, the directionality of constraint propagation needs to be considered. For time-series constraints, if the target field is the completion date, when tracing back to the start date field, it is necessary to verify whether the corrected completion date is later than the start date. For classification constraints, if the target field is a sub-category of the archive, when tracing back to the major category field, it is necessary to confirm that the sub-category correction value belongs to a sub-category of that major category. For numerical constraints, when tracing back to the associated numerical field, it is necessary to verify whether the corrected numerical relationship conforms to statistical laws or calculation relationships.

[0120] When retrospective verification reveals that all corrected candidate values ​​conflict with constraints of certain high-confidence fields, it indicates a potentially deeper error. In this case, instead of automatically correcting the target field, the field is marked as requiring manual review, and conflict information is appended to the output for reviewers' reference. This conflict case is also recorded for subsequent knowledge graph updates and constraint edge weight adjustments, enabling learning and improvement from error cases.

[0121] Through the aforementioned multi-layered conflict detection and priority assessment mechanism, when providing conflict correction suggestions in the face of different constraints, the mechanism can comprehensively consider the type and weight of the constraints, the credibility of the source fields, and logical consistency, thereby making reasonable correction decisions and ensuring the accuracy and reliability of archival metadata quality control.

[0122] The weights of the relation edges in the knowledge graph will be corrected based on feedback, and the fusion weights will be adjusted based on feedback from incorrectly identified samples, including:

[0123] Record the field correction records of the final metadata set relative to the verification metadata set. The field correction records include the target field to be corrected, the associated fields involved in the reverse derivation, and the constraint relationship edges connecting the two.

[0124] The effect of each statistical constraint edge on field correction is statistically analyzed. When the combination of field values ​​connected by a statistical constraint edge is verified to be correct after correction, the adjustable weight of the statistical constraint edge is increased. When the verification is incorrect, the adjustable weight of the statistical constraint edge is decreased, thus completing the update of the relation edge weights of the knowledge graph.

[0125] By comparing the structured extraction results with the final metadata set, the image regions and their types corresponding to the corrected fields are identified, and the image regions are marked as misidentified samples and their corresponding branch types are recorded.

[0126] The incorrectly identified samples are input into the three-branch heterogeneous fusion network. The basic weight bias of the corresponding branch is reduced according to the type of region where the incorrect identification occurs. The fusion weight is adjusted by accumulating the statistical information of the incorrectly identified samples.

[0127] For example, accurate extraction of archival metadata relies on the collaborative optimization of the recognition network and knowledge constraints. In practical applications, the system continuously accumulates correction data after operation. This correction data can reflect the effectiveness of the knowledge graph constraint rules and reveal the recognition bias of the three-branch heterogeneous fusion network in specific region types. To achieve continuous improvement, a two-way feedback mechanism needs to be established, simultaneously optimizing the constraint edge weights at the knowledge level and the fusion weights at the recognition level.

[0128] During the generation of the final metadata set, field values ​​with confidence scores below a threshold in the verification metadata set are manually reviewed and corrected, or automatically corrected based on historical archive data. Each correction operation generates a detailed field correction record, which stores the changes in field values ​​before and after the correction in a structured format. Specifically, the field correction record contains three key elements: the identifier of the target field being corrected and its original and corrected values, the associated fields involved in the constraint propagation calculation of the target field and their values, and the type and current weight of the constraint relationship edges connecting the target field and the associated fields. For example, when the "Archived Date" field of an archive is originally identified as "2023-15-08" and corrected to "2023-08-15", the record will indicate that the correction involves the temporal constraint relationship edge with the "Document Creation Date" field, and the classification constraint relationship edge with the "Archived Year" field.

[0129] Based on accumulated field correction records, statistical analysis is performed on constraint edges in the knowledge graph. The performance of each constraint edge across all correction records is categorized into three cases: the edge's constraint calculation correctly identifies the target field as an anomaly and successfully corrects it; the edge fails to identify an actual field error; and the edge mistakenly marks a correct field as an anomaly. To quantify the effect of constraint edges, statistical indicators are defined, including the number of successful corrections (Nsuccess), the number of missed corrections (Nmiss), and the number of false corrections (Nfalse). When the combination of field values ​​connected by a constraint edge is verified to be correct after manual correction (i.e., the corrected field values ​​are verified by actual archival data or expert review and confirmed to be correct), the adjustable weight of the edge is updated according to the step size Δw. positive (e.g., 0.1) is increased, and the update rule is to increase the current weight w of the edge. current Adjusted to w current +Δw positive ×(1−w current ), where (1−w current The first item ensures that the weight will not exceed the upper limit of 1. Conversely, if the corrected field value is still incorrect after verification, it indicates that the constraint logic of the constraint edge is flawed or not applicable to this type of file scenario. In this case, a penalty step size Δw is applied. negative (e.g., 0.15) Decrease the adjustable weight of this edge, and update the rule as w. current −Δw negative ×w current This ensures that weights do not fall into negative values. Through multiple iterative updates, the weights of edges with significant constraint effects gradually increase, thus playing a stronger role in subsequent graph reasoning constraint propagation, while the weights of edges with poor constraint effects gradually decrease, reducing their impact on field value satisfaction calculations. This adaptive weight update mechanism enables the knowledge graph to continuously optimize constraint rules based on feedback from actual archival data, improving the accuracy of metadata quality control.

[0130] While updating the knowledge graph weights, adjustments need to be made to the recognition capabilities of the three-branch heterogeneous fusion network. The differences between the structured extraction results and the final metadata set are checked field by field. When a field's value is found to be inconsistent with the final corrected value during the structured extraction stage, the source region of that field value in the original archival image is traced. Based on the field value's location coordinates in the image and its position within the document structure, the type of the region is determined. If the region is located within a clearly defined table cell and contains border features, it is classified as a table region; if the region contains pixel distribution features resembling a red stamp, it is classified as a stamp region; other regions containing regular text arrangements are classified as text regions. The incorrectly identified image regions are cropped and extracted from the original archival image, and their correct field values ​​are labeled as true labels, forming incorrect recognition samples. Simultaneously, the branch type corresponding to each incorrect recognition sample is recorded, specifically one of the table recognition branch, stamp recognition branch, or text recognition branch.

[0131] After accumulating a certain number of misidentified samples, these samples are re-input into the three-branch heterogeneous fusion network for analysis. The proportion of misidentified samples corresponding to different branch types is statistically analyzed, and the error rate index of each branch is calculated. For branch types with high error rates, the calculation strategy for fusion weights is adjusted. Specifically, in the dynamic fusion weight calculation, the final weight of each branch is jointly determined by the region confidence and the basic weight bias. For branches with high misidentification rates, their basic weight bias parameter is reduced, so that even if the region confidence output by this branch is high, its actual weight in the fusion stage will be suppressed. The reduction magnitude is determined based on the accumulated statistical information of misidentified samples; the higher the error rate of a branch, the greater the reduction in its basic weight bias. By continuously accumulating statistical information of misidentified samples, the fusion weight adjustment strategy can adaptively balance the contributions of the three branches, preventing any one branch from dominating the fusion result due to overconfidence, thereby improving the overall robustness and accuracy of the recognition.

[0132] This two-way feedback mechanism enables collaborative optimization between the identification layer and the knowledge layer, allowing the intelligent extraction and quality control methods for archival metadata to continuously improve themselves in practical applications and adapt to the characteristics and changes of different types of archives.

[0133] A second aspect of the present invention provides an electronic device, comprising:

[0134] processor;

[0135] Memory used to store processor-executable instructions;

[0136] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0137] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0138] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent extraction and quality control of archival metadata based on deep learning, characterized in that, include: The obtained archive images to be processed are analyzed for layout and text recognition by a three-branch heterogeneous fusion network, which includes a table recognition branch, a seal recognition branch and a text recognition branch. The fusion weight is dynamically calculated based on the regional confidence of each branch and the outputs of each branch are weighted and fused to obtain the structured extraction result. Named entity recognition is performed on the structured extraction results to obtain a candidate metadata set; A knowledge graph for the archival domain is constructed. The knowledge graph includes metadata field nodes and constraint relationship edges between representation fields, which are temporal constraints, classification constraints, and numerical constraints. The field values ​​in the candidate metadata set are constrained along the constraint relationship edges using a graph reasoning algorithm, and the satisfaction degree of the field value combination is calculated. A confidence score is generated based on the satisfaction degree to obtain a verification metadata set. The final metadata set is obtained by correcting the verification metadata set based on the confidence score. The final metadata set is output, and the relation edge weights of the knowledge graph are updated based on the feedback. At the same time, the fusion weights are adjusted based on the feedback of the identified erroneous samples.

2. The method according to claim 1, characterized in that, The three-branch heterogeneous fusion network includes: A shared feature extraction layer is used to extract the low-level feature map of the archive image using a multi-scale convolutional encoder; The table recognition branch performs line segment detection and cell segmentation based on the underlying feature map, and outputs the table area, table structure, and text content within the table; the seal recognition branch performs circular area detection and seal character recognition based on the underlying feature map, and outputs the seal area and seal content; the text recognition branch performs text line detection and character recognition based on the underlying feature map, and outputs the text area and text content. The weighted fusion is implemented through an adaptive fusion module. The adaptive fusion module calculates the fusion weight based on the regional confidence of each branch, adaptively adjusts the weight allocation strategy according to the complexity of different image regions, makes fusion decisions for regions with overlaps or conflicts, and determines the final region type classification. The adaptive fusion module also includes a cross-branch attention mechanism, which uses the intermediate layer features of the stamp recognition branch to guide the text recognition branch to enhance the recognition ability of text covered by the stamp or located in the seam. The structured extraction results are composed of the regions, their type identifiers, spatial location information, and identification content after fusion decision-making.

3. The method according to claim 1, characterized in that, Construct a knowledge graph for the archival domain. The knowledge graph includes metadata field nodes and constraint edges between representation fields, encompassing temporal, classification, and numerical constraints. Mandatory constraint rules are parsed from the archives management standards, and these mandatory constraint rules are transformed into normative constraint edges and assigned fixed weights. Statistical correlation patterns between fields are mined from historical archive data, and these patterns are transformed into statistical constraint edges and assigned adjustable weights. Based on the dependency type between the fields involved in the normative constraint edge and the statistical constraint edge, the constraint relationship edge is classified into temporal constraint edge, classification constraint edge and numerical constraint edge; the temporal constraint edge represents the temporal dependency relationship between fields, the classification constraint edge represents the category affiliation or mutual exclusion relationship between fields, and the numerical constraint edge represents the numerical range or computational dependency relationship between fields. The normative constraint edges and the statistical constraint edges are connected together to the metadata field nodes. Each metadata field node is connected to other field nodes through multiple constraint relationship edges of different types, thus forming a knowledge graph of the archival domain.

4. The method according to claim 3, characterized in that, The field values ​​in the candidate metadata set are constrained along the constraint relationship edges using a graph reasoning algorithm, and the satisfaction degree of the field value combinations is calculated. A confidence score is generated based on the satisfaction degree to obtain the verification metadata set, including: The graph reasoning algorithm starts from any field value in the candidate metadata set and propagates constraint information to adjacent field nodes along the constraint relationship edges connected to that field node. When propagating to a normal constraint edge, it determines whether the combination of field values ​​satisfies the mandatory constraint conditions corresponding to that edge to obtain a hard constraint judgment value. When propagating to a statistical constraint edge, it quantifies the degree of compliance of the combination of field values ​​with the constraint conditions corresponding to that edge based on the adjustable weight of that edge to obtain a weighted compliance degree. After multi-hop constraint propagation, the satisfaction degree is obtained by combining the hard constraint judgment values ​​of all normative constraint edges with the weighted compliance calculation field values ​​of all statistical constraint edges. When the hard constraint judgment value of any normative constraint edge is not satisfied, the satisfaction degree is set to zero. When the hard constraint judgment values ​​of all normative constraint edges are satisfied, the weighted compliance values ​​of each statistical constraint edge are summed to obtain the satisfaction degree value. The satisfaction score is normalized and mapped to a confidence score. Each combination of field values ​​in the candidate metadata set is labeled with its confidence score. The combination of field values ​​with confidence scores higher than a preset score threshold is selected to obtain the verification metadata set.

5. The method according to claim 1, characterized in that, The final metadata set is obtained by correcting the verification metadata set based on the confidence score, including: Identify target fields in the verification metadata set whose confidence scores are lower than a correction threshold, and locate the field nodes corresponding to the target fields from the knowledge graph; Tracing backward from the target field node along the constraint relationship edge to the associated field node whose confidence score is higher than the correction threshold, the target field is deduced in reverse according to the constraint conditions of the connected constraint relationship edge. When the constraint relationship edge is a time series constraint edge, the time range of the target field is deduced based on the time value of the associated field. When the constraint relationship edge is a classification constraint edge, the category of the target field is deduced based on the category value of the associated field. When the constraint relationship edge is a numerical constraint edge, the numerical range of the target field is calculated based on the numerical value of the associated field and the constraint conditions. When multiple constraint edges provide different correction candidate values, conflict detection and priority evaluation are performed on the correction candidate values, and the final correction value is selected as the correction value of the target field. After updating the corrected value to the target field, the process propagates along the constraint relationship edges to check the associated fields. When the associated field conflicts with the corrected target field, the associated field is corrected accordingly to obtain the final metadata set.

6. The method according to claim 5, characterized in that, When multiple constraint edges provide different candidate correction values, conflict detection and priority evaluation are performed on the candidate correction values, and the final correction value is selected as the correction value for the target field, including: Conflict detection is performed on the correction candidate values ​​given for multiple constraint relationship edges; when the difference between different correction candidate values ​​exceeds the preset conflict threshold, it is determined that there is a conflict. When conflicts exist, the overall confidence of each corrective candidate value is calculated by weighting the weight of each constraint edge and the confidence score of the corresponding associated field. If the overall confidence difference of multiple corrected candidate values ​​is less than the preset difference threshold, then the corrected candidate value derived from the normative constraint edge is selected; if they are all derived from the statistical constraint edge, then the corrected candidate value with the highest overall confidence is selected. When there is logical mutual exclusion between the candidate values ​​for correction, trace along the constraint relationship edge to the field node with a confidence score higher than the correction threshold to verify the rationality of each candidate value, and select the candidate value for correction that maintains constraint consistency with all fields with confidence scores higher than the correction threshold on the tracing path as the final correction value.

7. The method according to claim 1, characterized in that, The weights of the relation edges in the knowledge graph will be corrected based on feedback, and the fusion weights will be adjusted based on feedback from incorrectly identified samples, including: Record the field correction records of the final metadata set relative to the verification metadata set. The field correction records include the target field to be corrected, the associated fields involved in the reverse derivation, and the constraint relationship edges connecting the two. The effect of each statistical constraint edge on field correction is statistically analyzed. When the combination of field values ​​connected by a statistical constraint edge is verified to be correct after correction, the adjustable weight of the statistical constraint edge is increased. When the verification is incorrect, the adjustable weight of the statistical constraint edge is decreased, thus completing the update of the relation edge weights of the knowledge graph. By comparing the structured extraction results with the final metadata set, the image regions and their types corresponding to the corrected fields are identified, and the image regions are marked as misidentified samples and their corresponding branch types are recorded. The incorrectly identified samples are input into the three-branch heterogeneous fusion network. The basic weight bias of the corresponding branch is reduced according to the type of region where the incorrect identification occurs. The fusion weight is adjusted by accumulating the statistical information of the incorrectly identified samples.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images