Archives management method and system based on registration behavior graph and era perception search
By constructing a cataloging behavior graph and a time-drift knowledge graph, the problems of the propagation of cataloging errors and the low recall rate of historical archive retrieval in the archive management system are solved, and efficient error repair and cross-era retrieval are achieved.
Patent Information
- Application Number
- CN202611045377.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing record management systems suffer from systemic propagation of cataloging errors and low recall rates for historical record retrieval. Current technologies cannot effectively identify the root causes and propagation chains of errors, nor can they effectively handle semantic misalignment caused by time drift.
We construct a record-keeping behavior graph, use graph neural networks to cluster error root causes and block the propagation chain, and simultaneously construct a knowledge graph of archival terminology across historical periods, automatically expanding modern queries to a set of equivalent terms covering all historical periods.
It significantly improved the identification rate and repair efficiency of cataloging errors, reduced the workload of manual review, and improved the retrieval recall and discoverability of historical archives, forming a positive technical closed loop of "repair-feedback-retrieval optimization".
Smart Images

Figure CN123068639A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of archival management technology, and more specifically, to an archival management method and system based on bibliographic behavior mapping and era perception retrieval. Background Technology
[0002] The informatization and intelligent transformation of archival management has placed higher technical demands on cataloging efficiency, quality control, and retrieval services. Utilizing artificial intelligence technologies such as computer vision, natural language processing, and knowledge graphs to assist in the automated extraction, standardized organization, and semantic retrieval of archival information has become an important development direction in this field. However, in the engineering practice of intelligent archival management systems, there are two independent but common technical challenges that seriously affect system effectiveness, and existing technical solutions have not yet provided effective and systematic solutions.
[0003] In terms of cataloging quality control, current mainstream technical solutions mainly rely on field-level rule verification, which essentially involves independent judgment of each record. However, cataloging errors in real-world scenarios are not isolated, randomly distributed phenomena, but rather exhibit significant systematic clustering and chain propagation characteristics. A small number of root causes of errors spread rapidly within the cataloging behavior network through three paths: homology of cataloger behavior, propagation of defects in cataloging templates, and circular referencing of historical samples, ultimately forming a large-scale cluster of errors from the same source. The record-by-record verification technology lacks the architectural design to reverse-locate the root causes of error sets and block the propagation chain, resulting in a linear increase in the workload of manual review and data scale, with systematic biases remaining for a long time.
[0004] In terms of retrieval services, the archival holdings span a wide period of time, and there is a significant "time drift" phenomenon between the terminology used in historical archives and modern general language. This manifests in three types of semantic misalignment: historical changes in the names of institutions and administrative entities, semantic evolution of social and professional terms, and cross-period name distortions of the same entity. Existing vector retrieval models, due to their training corpus primarily consisting of modern text, suffer from severe semantic representation limitations in historical archive content, resulting in a significantly lower recall rate for historical archives compared to modern archives. Furthermore, the accuracy of general optical character recognition engines in recognizing handwritten content, seals, and complex table structures on archival pages is insufficient for practical applications, and general knowledge graphs do not include domain-specific semantic drift relationships related to the diachronic evolution of language in the archival field. These technological shortcomings, combined, constitute a systemic technical bottleneck restricting the improvement of intelligent archival management. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to provide an archive management method and system based on bibliographic behavior mapping and era perception retrieval.
[0006] Firstly, a method for tracing and correcting errors in archival records is provided, including:
[0007] S1. Construct a relationship graph of recording behavior, with the recording behavior unit as the node, and establish weighted connection edges between the nodes based on three conditions: the time homology of the recorder, the homology of the recording template, and the homology of the batch import.
[0008] S2. Perform field-level rule verification on the bibliographic records and mark the bibliographic behavior units containing errors as error node sets; use the error node sets as seeds on the bibliographic behavior association graph, use graph neural networks to perform feature aggregation, and perform community detection to divide the error nodes into K clusters. Perform root cause determination on each cluster to confirm the root cause type of each cluster.
[0009] S3. Based on the root cause type of each cluster confirmed in step S2, the clusters are attributed to the corresponding error source category;
[0010] S4. Starting from each cluster, perform graph traversal on the descriptive behavior association graph to identify high-risk descriptive records generated by root cause propagation corresponding to that cluster but not yet marked, thus forming the propagation influence domain.
[0011] S5. Within the propagation impact domain, using archival records as the restoration objects, a set overlay algorithm is used to generate a restoration priority list;
[0012] S6. Write the repair results back to the file management system, extract the field values before and after the repair record, and output the field difference pairs.
[0013] Preferably, the root cause determination in step S2 includes: calculating the proportion of dominant edge types within a cluster, and determining the type with the highest proportion of dominant edge types as the corresponding root cause type when it exceeds a preset threshold; verifying the consistency of error field types within a cluster; calculating the purity within a cluster of three types of homologous features, and triggering homologous root cause determination when the purity exceeds a threshold; outputting the root cause confidence score by weighted scoring based on the above indicators, and triggering manual verification for clusters with confidence scores below the threshold; the error source categories in step S3 include: attributing the error to the bibliographer's personal bias when the dominant edge is a bibliographer's time homologous edge, attributing the error to the bibliographer's template defect when the dominant edge is a template homologous edge, and attributing the error to batch import error when the dominant edge is a batch homologous edge.
[0014] Preferably, the weight of the weighted connection edge in step S1 is obtained by weighted summation of at least two of the recorder time homology indicator function, template homology indicator function, and batch homology indicator function; in step S2, the optimal number of clusters K is automatically determined by the modularity optimization algorithm.
[0015] Preferably, the graph traversal in step S4 is a weighted breadth-first search, adding neighboring nodes with edge weights greater than a preset threshold to the propagation influence domain; step S6 also includes: establishing a dynamic deviation profile for each logger, calculating a personalized early warning score when a logger submits a new logging record, and triggering a verification prompt when the score exceeds the early warning threshold.
[0016] Secondly, a cross-era semantic retrieval method for archives is provided, the method receiving the field difference pair output from step S6 of the method described in any of the first aspects as input, including the following steps:
[0017] S1. Construct a knowledge graph of the temporal drift of archival terminology, using historical terms, modern equivalent terms, types of drift relationships, and historically active time intervals as elements to express cross-temporal semantic equivalence relationships;
[0018] S2. After the received field differences pass the confidence assessment, they are incrementally written into the era drift knowledge graph as new word pairs; vector embedding is re-performed on the repaired archive records to update the retrieval vector index incrementally;
[0019] S3. Based on time slices of archival corpora, drifting candidate word pairs are automatically mined through word frequency transition statistical tests, and then pushed for confirmation and storage after confidence filtering.
[0020] S4. Perform era-context detection on user queries, and use the era-drift knowledge graph to find equivalent terms for modern terms in various historical periods to construct an extended query set;
[0021] S5. Perform vector semantic retrieval and knowledge graph entity expansion retrieval on the extended query set, and generate comprehensive retrieval results through a fusion ranking algorithm.
[0022] Preferably, the fusion ranking in step S5 includes an era matching reward item, which gives additional ranking reward weight when the formation year of the candidate archive falls within the historical active time range of the query expansion terms; the drift relationship type includes a historical term type without modern equivalent terms, which triggers retrieval expansion through at least one of the following: the intersection of the query time condition and the historical active time range, related event keywords, topic tags, and archive formation time.
[0023] As a preferred option, the word frequency transition statistical test in step S3 is as follows: the archives are divided into multiple time slices according to timestamps, a temporal word frequency matrix is constructed, a statistical significance test is performed on the word frequency changes between adjacent slices, and the candidate transition words that pass the test are paired according to temporal complementarity, semantic relevance and contextual distribution similarity. After comprehensive confidence scoring and mismatch filtering, they are pushed to be confirmed and added to the database.
[0024] Thirdly, an archival management system based on bibliographic behavior mapping and era-aware retrieval is provided, including:
[0025] The error tracing and repair module for the cataloging behavior graph is used to execute steps S1 to S6 of the method described in the first aspect, construct a cataloging behavior association graph, perform error root cause clustering attribution and propagation influence domain expansion, generate repair paths, and extract field difference pairs before and after repair.
[0026] The era-aware semantic retrieval module is used to execute steps S1 to S5 of the method described in the second aspect, receive the field difference before and after repair to update the archival terminology era drift knowledge graph, expand the era context of user queries, and combine vector retrieval and knowledge graph retrieval to realize cross-era archival retrieval.
[0027] A multimodal archive parsing engine is used to identify scanned archive documents and output a structured representation of archive content, providing the recording field input for the error tracing and repair module of the recording behavior map;
[0028] The archival knowledge graph storage module is used to store archival entities and their relationships, providing knowledge graph retrieval support for the era-aware semantic retrieval module;
[0029] The vector retrieval library storage module is used to store the semantic vector representation of archival records, providing vector retrieval support for the era-aware semantic retrieval module.
[0030] Preferably, the multimodal document parsing engine performs semantic block segmentation on the scanned document, dividing the page into a title area, a body text area, a table area, a handwritten annotation area, and a seal area, and then sends them to the corresponding recognition submodules for processing.
[0031] Preferably, the system supports deployment and operation under a dual-graphics card configuration: the multimodal parsing model adopts INT8 quantization, the graph neural network inference stage is executed on the central processing unit, and the time-drift knowledge graph and vector retrieval library are decoupled from the graphics processor and run independently.
[0032] The beneficial effects of this invention are:
[0033] 1. This invention constructs a record-by-record behavior association graph and introduces a graph neural network for error root cause clustering, thereby improving the error identification granularity from record-by-record to the root cause cluster level. This enables batch location of the source of systematic errors and blocking of the propagation chain, significantly improving the systematic record-by-record error identification rate, greatly reducing the workload of manual review, and preventing the repair cost from rising linearly as the data scale continues to grow. This fundamentally overcomes the technical defects of the existing record-by-record verification scheme, which is inefficient.
[0034] 2. This invention constructs a knowledge graph of archival terminology across different historical periods, automatically expanding modern user queries to a set of equivalent terms covering all historical periods. This bridges the diachronic semantic gap without replacing existing vector retrieval models, significantly improving the recall rate of historical archive retrieval and the recall rate of queries containing terms with time-drifted terms. It also significantly improves the discoverability of historical archives without sacrificing the accuracy of modern archive retrieval.
[0035] 3. This invention deeply integrates the process of tracing and repairing bibliographic errors with the time-drift knowledge graph update mechanism. The field differences generated during repair are automatically and incrementally written into the knowledge graph after confidence assessment, so that the retrieval capability is enhanced in sync with the continuous improvement of bibliographic quality, forming a positive technical closed loop of "repair-feedback-retrieval optimization". At the same time, it is supplemented by a multimodal archive parsing engine and a lightweight deployment scheme, which has excellent performance in terms of automated extraction of bibliographic fields, recognition of handwritten annotations and hardware resource adaptation, combining technological advancement and engineering practicality. Attached Figure Description
[0036] Figure 1 This is a diagram of the overall architecture of the AIMS-CBG system.
[0037] Figure 2 A flowchart for the process of archive entry and automatic cataloging;
[0038] Figure 3 Flowchart for constructing the behavior association graph and clustering and attributing error root causes;
[0039] Figure 4a This is a performance comparison chart of an experiment for tracing the source of cataloging errors;
[0040] Figure 4b This is a performance comparison chart for another type of recordation error tracing experiment;
[0041] Figure 5a A comparison chart of the performance of the "Time-Aware Retrieval" experiment;
[0042] Figure 5b A comparison chart of the performance of the Era Perception Retrieval Experiment. Detailed Implementation
[0043] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0044] Example 1:
[0045] Existing bibliographic quality control technologies verify bibliographic errors one by one in isolation, which has a time complexity of O(n) and cannot distinguish between random errors and systematic errors, identify the root cause of errors, or block the error propagation chain.
[0046] This invention is the first to propose modeling archival recording errors as a systematic propagation phenomenon on a recording behavior network: a limited number of "error root causes" spread into a large number of recording errors of the same origin through three propagation paths: the homogeneity of recording behavior, the reusability of recording templates, and the homogeneity of batch imports.
[0047] Specifically, to address the problems of existing technologies, Embodiment 1 of this application provides a method for tracing and repairing errors in archival record keeping, including:
[0048] S1. Construct a relationship graph of recording behavior, with recording behavior units as nodes, and establish weighted connection edges between nodes based on three conditions: time homology of recorders, homology of recording templates, and homology of batch imports.
[0049] Specifically, the cataloging behavior unit is uniquely identified by a triple of (cataloger ID, template ID, batch ID). Multiple cataloging records under the same triple are merged into one behavior node. The node attributes include field semantic embedding vectors and cataloging metadata.
[0050] Furthermore, the weights of the weighted connection edges are obtained by weighted summation of at least two of the following: the bibliographer's temporal homology indicator function, the template homology indicator function, and the batch homology indicator function. The bibliographer's temporal homology indicator function has a value of 1 when two bibliographic records are submitted by the same bibliographer within a predetermined time window, and a value of 0 otherwise. The weight coefficients are obtained through logistic regression with L2 regularization learning from historical labeled data and support online incremental updates.
[0051] S2. Perform field-level rule verification on the bibliographic records, and mark the bibliographic behavior units containing errors as error node sets; use the error node sets as seeds on the bibliographic behavior association graph, use graph neural networks to perform feature aggregation, and perform community detection to divide the error nodes into K clusters. Perform root cause determination on each cluster to confirm the root cause type of each cluster.
[0052] Specifically, the error node set is marked by field-level rule verification on the cataloging records; the error node set is used as a seed on the cataloging behavior association graph, and a graph attention network is used for feature aggregation. Community detection is performed by combining Louvain modularity optimization, and the error nodes are divided into K clusters. The root cause type is confirmed after verification by the root cause determination rule.
[0053] The root cause determination rule includes: calculating the proportion of dominant edge types within a cluster (the proportion of edges with the same origin as the recorder's time, the same origin as the template, and the same origin as the batch), and determining the dominant edge type with the highest proportion exceeding a preset threshold. The system determines the corresponding root cause type; verifies the consistency of error field types within clusters, considering a field-level strong correlation signal when the error rate of the same field exceeds twice the cluster mean; calculates the purity within clusters of three homologous features: bibliographer-time, template ID, and batch ID, and determines purity exceeding a threshold. The determination of the common root cause is triggered at certain times; the confidence level of the root cause is output by weighted scoring based on the above indicators, and the confidence level is lower than the threshold. Clustering triggers manual verification of the closed loop.
[0054] In addition, the Louvain modularity optimization algorithm is used in step S2 to automatically determine the optimal number of clusters K without manual preset, and the time complexity is O(n log n), which supports real-time processing of hundreds of thousands of bibliographic records.
[0055] S3. Based on the root cause type of each cluster confirmed in step S2, the clusters are attributed to the corresponding error source category.
[0056] Specifically, when the dominant edge is a time-related edge of the same bibliophile and the field error pattern is highly consistent with the behavior of a specific bibliophile, it is attributed to the bibliophile's personal bias; when the dominant edge is a template-related edge and multiple bibliophiles under the same template present the same field error, it is attributed to the bibliographic template defect; when the dominant edge is a batch-related edge and the error occurs in batches within a short time window, it is attributed to batch import error; for clusters whose confidence level does not reach the threshold, manual verification is pushed and the attribution is corrected and the threshold parameters are updated based on the verification results.
[0057] S4. Starting from each cluster, perform graph traversal on the descriptive behavior association graph to identify high-risk descriptive records generated by the root cause propagation corresponding to that cluster but not yet marked, thus forming the propagation influence domain.
[0058] S5. Within the propagation impact domain, using archival records as the restoration objects, a set overlay algorithm is used to generate a restoration priority list.
[0059] Specifically, within the propagation impact domain, the archival records are taken as the objects of restoration. The restoration operations under each root cause cluster are regarded as a set (the elements of which are the bibliographic records that can be associated with the restoration by the operation). A greedy approximate maximum weighted set coverage algorithm is used to generate a restoration priority list: in each round, the restoration action that can cover the most unrestored records in a single operation is selected first, and the maximum number of bibliographic records to be restored is covered with the minimum number of manual review operations.
[0060] S6. Write the repair results back to the file management system, extract the field values before and after the repair record, and output the field difference pairs.
[0061] Step S6 also includes: establishing a dynamic deviation profile for each cataloger that includes the historical error rate and main error patterns of each field; calculating a personalized early warning score when the cataloger submits a new catalog record; and triggering a field-level key check prompt when the early warning threshold is exceeded, thereby achieving advance prevention of cataloging errors.
[0062] Example 2:
[0063] Modern vector retrieval models (such as BGE-M3 and E5) primarily use modern Chinese text as their training corpus, lacking effective representations of vocabulary from government archives from 1949 to 1985. When users search for historical archives using modern terminology, a significant semantic gap exists between the modern terminology used and the historical terminology used in the archives, resulting in a large number of relevant historical archives being unretrievable.
[0064] Furthermore, based on Embodiment 1, Embodiment 2 of this application provides a cross-era semantic retrieval method for archives. Instead of replacing the existing vector retrieval model, it constructs a knowledge graph of the era-drift of archival terminology and automatically expands modern query terms into a set of equivalent terms covering various historical periods before retrieval, fundamentally bridging the semantic gap.
[0065] Specifically, the method receives the field difference pair output from step S6 of the method in Embodiment 1 as input, and includes the following steps:
[0066] S1. Construct a knowledge graph of the temporal drift of archival terminology, using historical terms, modern equivalent terms, types of drift relationships, and historically active time intervals as elements to express cross-temporal semantic equivalence relationships.
[0067] Specifically, the types of drift relationships include institutional renaming, semantic evolution, terminology dissolution, institutional mergers and reorganizations, and conceptual extensions. Among them, the modern equivalent term field of terminology dissolution entries is null. This type of historical terminology is triggered by at least one of the following: the intersection of query time conditions and historical active time intervals, related historical event keywords, subject tags, and archive formation time. It does not rely on reverse lookup of modern equivalent terms.
[0068] S2. After the received field differences are evaluated with confidence, they are incrementally written into the era drift knowledge graph as new word pairs; vector embedding is re-executed on the repaired archive records to update the retrieval vector index incrementally.
[0069] Specifically, perform the following incremental updates:
[0070] ① Extract the pre-repair and post-repair field values from the records to form field difference pairs; conduct confidence assessment based on the self-consistency score and cross-record consistency verification of the repair difference pairs. After the assessment is passed, the new word pairs are incrementally written into the archival terminology era-drift knowledge graph without further expert confirmation; ② Re-perform vector embedding on the repaired archival records to update the retrieval vector index incrementally, so that the repaired content can be correctly recalled by era-aware semantic retrieval; ③ After the knowledge graph and retrieval index are updated, subsequent cross-era retrieval requests will automatically use the updated graph and index, forming a closed loop of continuous feedback on retrieval effect from repair quality.
[0071] S3. Based on time slices of archival corpora, drift candidate word pairs are automatically mined through word frequency transition statistical tests. After confidence filtering, they are pushed for confirmation and storage, realizing semi-automatic accumulation of knowledge about time drift.
[0072] Specifically, the automatic mining method for drift candidate word pairs in step S3 is as follows: the archives are divided into multiple slices with a granularity of 10 years according to the timestamp, a temporal word frequency matrix is constructed, and a chi-square test (significance level p<0.01) is performed on the word frequency changes between adjacent slices. The candidate transition words that pass the test are paired according to temporal complementarity, semantic relevance and contextual distribution similarity. After comprehensive confidence score and mismatch filtering, they are pushed to archive experts for confirmation and entry into the database. The comprehensive confidence score is obtained by weighted summation of three factors: normalized chi-square statistic, cosine similarity of word vectors and Jensen-Shannon divergence of context.
[0073] S4. Perform era-context detection on user queries, and use the era-drift knowledge graph to find equivalent terms for modern terms in various historical periods, thus constructing an extended query set covering multiple historical periods.
[0074] S5. Perform vector semantic retrieval and knowledge graph entity expansion retrieval on the expanded query set. By introducing a fusion ranking algorithm with an era matching reward factor, generate comprehensive retrieval results covering different historical periods.
[0075] The fusion ranking in step S5 includes an era matching reward item. When the formation year of a candidate archive falls within the historical active time range of the query expansion terms, it is given additional ranking reward weight to improve the ranking of the era matching historical archives in the search results. The drift relationship type includes a historical term type without modern equivalent terms. This type of historical term triggers search expansion through at least one of the following: the intersection of the query time condition and the historical active time range, related event keywords, topic tags, and the archive formation time.
[0076] Through the above-mentioned incremental update mechanism, each cataloging repair operation in Example 1 substantially expands the era-drift knowledge graph and optimizes the retrieval index, thereby continuously improving the accuracy and recall of cross-era retrieval as the cataloging quality improves. The two methods work together to serve the overall technical goal of improving the quality of archival cataloging and high-quality retrieval.
[0077] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.
[0078] Example 3:
[0079] To address the shortcomings of existing technologies, such as their inability to identify systemic root causes and propagation chains, lack of archival structure awareness, absence of historical semantic mapping capabilities, and lack of archival domain-specific semantic drift relationships, Embodiment 3 of this application provides an Archival Intelligent Management System based on Cataloging Behavior Graph and Era-Aware Retrieval (hereinafter referred to as AIMS-CBG), such as... Figure 1 As shown, it includes:
[0080] The Descriptive Behavior Graph Error Source Tracing and Repair Module (CBGE) is used to execute steps S1 to S6 of the method described in Example 1, construct a descriptive behavior association graph, perform error root cause clustering attribution and propagation influence domain expansion, generate repair paths, and extract field difference pairs before and after repair.
[0081] Specifically, the error tracing and repair module for the bibliographic behavior map includes the following sub-modules:
[0082] Submodule 1: Cataloging Behavior Graph (CBG) Construction Submodule
[0083] Specifically, this submodule is used to define the relationship diagram of recording behavior. ,in:
[0084] node ∈V: A recording behavior unit, defined as a unique combination of the triples (Cataloger ID, Template ID, Batch ID). Multiple recording records under the same triple are merged into one behavior node, whose attribute vector is the weighted average of the attribute vectors of the records in that group. The node weight is the number of records in that group.
[0085]
[0086] The components are: field semantic embedding (encoded by the BGE-M3 model, dimension 768), bibliographer identifier, bibliograph template ID, bibliograph timestamp, and batch import ID.
[0087] Edge establishment rules: Node and A directed edge is established between two points if any of the following conditions are met:
[0088] Behavioral origins: = And | - |≤30 days;
[0089] Same template origin: = ;
[0090] Batch same origin: =
[0091] Edge weight:
[0092]
[0093] Among them, the weighting coefficient , , It is learned on a labeled dataset using logistic regression L2 regularization, with a typical value of =0.30, =0.45, =0.25 (the sum of the three is 1)
[0094] Submodule 2: Error Root Cause Clustering (ERCC) Submodule
[0095] This submodule is used to perform the following steps:
[0096] Step S1 (Error Node Marking): Perform field-level rule validation (file number format, time validity, institutional name standardization, etc.) on all bibliographic records, and mark the bibliographic behavior units containing erroneous records as error node sets. ⊆V;
[0097] Step S2 (Graph Neural Network Clustering): In above Using the seed as a reference, a Graph Attention Network (GAT) is used for feature aggregation:
[0098]
[0099] Attention coefficient This reflects the strength of the erroneous association between node j and node i;
[0100] The technical details of integrating GAT feature aggregation with Louvain are as follows:
[0101] ① Input feature construction: For each error node in CBG ∈ Its input feature vector It is composed of components shown in Table 1:
[0102] Table 1. Types of Feature Vector Components
[0103]
[0104] ②GAT Network Structure: A 2-layer GAT is used, with the first layer having 4 attention heads (each head outputting 128 dimensions) and the second layer having 1 attention head (outputting 64 dimensions). Inter-layer ELU activation + Dropout (p=0.3) is applied. GAT operates only on the subgraph induced by Verr, with the input being all nodes in Verr and their adjacent edges.
[0105] ③ Training Objective: Employ self-supervised reconstruction loss, without relying on manually labeled root cause tags. Specifically, the GAT training objective is to minimize the following loss function:
[0106]
[0107] The reconstruction loss is: ,
[0108] For MLP decoders (64→256→822).
[0109] Embedded in GAT output nodes;
[0110] Sparse regularization ,
[0111] Encourage compact embedding within clusters; λ=0.1. Training epoch=200, learning rate 1e-3, Adam optimizer.
[0112] ④ Attention Output → Louvain Connection: GAT outputs the 64-dimensional embedding of each error node. Then, calculate the similarity matrix of node pairs (i,j). ( (for sigmoid), As the new weights of edges in the Louvain algorithm, replacing the original edge weights. Louvain maximizes modularity. Divide the target into communities and output K communities. ,..., }
[0113] ⑤ Dominant Feature Determination: For each community Calculate the intra-cluster purity of the three types of homology features respectively:
[0114]
[0115]
[0116]
[0117] The homologous type corresponding to the maximum value among the three is taken as the dominant feature. Additional verification conditions: the dominant purity must be ≥0.6 and the normalized entropy of the erroneous field type within this community must be... A score of ≤1.5 (meaning the error pattern is not too scattered) is required to confirm the root cause type; otherwise, it is marked as "awaiting manual verification".
[0118] ⑥ Root cause confidence calculation:
[0119]
[0120] in Primarily based on purity, For the entropy of the error field, (Maximum entropy of 5 error fields) / The cluster percentages are represented by weights α=0.5, β=0.3, and γ=0.2. Clusters with a confidence level below 0.5 trigger manual verification to close the loop.
[0121] ⑦ Reproducible example parameters are shown in Table 2:
[0122] Table 2 Reproducible Example Parameters
[0123]
[0124] Step S3 (Clustering Confirmation and Root Cause Mapping): Based on the output of GAT feature aggregation and Louvain community detection in Step S2, confirm the K incorrect clusters { , ,…, Each cluster corresponds to a root cause of error;
[0125] Step S4 (Root Cause Type Attribution): For each cluster Extract its dominant feature: if the dominant feature is the bibliographer ID, it is attributed to bibliographer personal bias; if it is the bibliograph template ID, it is attributed to template flaw; if it is the batch ID, it is attributed to batch import error.
[0126] Step S5 (Expanding the Propagation Influence Domain): For each root cause cluster... ,exist above Perform a weighted breadth-first search (BFS) starting from the edge weights. > (default Neighboring nodes of (=0.55) are added to the propagation influence domain D( ); D( ) in and Records that are associated but have not yet been marked as error nodes are marked as high-risk warning records.
[0127] Submodule 3: Minimum Cost Repair Planning (MCRP)
[0128] This submodule is used to propagate influence domain D ( Within this section, define the node repair cost:
[0129]
[0130] in For nodes Risk score for transmission to unmarked neighbors.
[0131] A priority list for repairs is generated by using a greedy approximate maximum weighted set coverage: Within the propagation impact domain, repair operations under each root cause cluster are treated as a set (the elements of which are the catalog records that the operation can be associated with for repair). In each round, the repair action that can cover the largest number of unrepaired records in a single operation is selected first, and the maximum number of catalog records to be repaired is covered with the fewest manual review operations. For root causes attributed to TemplateFlaw, the system automatically generates template repair suggestions, and a single repair can cover all erroneous records from that template source.
[0132] Submodule 4: Cataloger Bias Profile (CBP)
[0133] This submodule is used to create a dynamic deviation profile for each cataloger:
[0134]
[0135] Record the historical error rate and main error patterns in each bibliographic field (such as "year is habitually omitted in the document number field" or "full name of the responsible organization is missing in the author field"). When a bibliographer submits a new bibliographic record, the system calculates a personalized warning score:
[0136]
[0137] like (Default 0.40) Triggers field-level key check prompts to prevent recording errors in advance.
[0138] For example, in an experiment on approximately 32,000 bibliographic records of a company, the CBGE algorithm identified 11 error root cause clusters, covering 86.4% of the total error records; the efficiency of batch repair of systematic errors was improved by about 6 times compared with manual review of each record, and the workload of manual review was reduced by 83.7%.
[0139] The era-aware semantic retrieval module is used to execute steps S1 to S5 of the method described in Embodiment 2, receive the field difference pair before and after repair to update the archival terminology era drift knowledge graph, expand the era context of user queries, and combine vector retrieval and knowledge graph retrieval to realize cross-era archival retrieval.
[0140] Specifically, this module includes the following sub-modules:
[0141] Submodule 1: Temporal Terminology Drift Graph (TTDG)
[0142] This submodule is used to define the knowledge triplet for time-drift:
[0143] , , , [ , ]
[0144] Meaning of each field:
[0145] Historical terminology;
[0146] : The corresponding modern equivalent term;
[0147] : Drift relationship type;
[0148] [ , [: The time period during which historical terms were active (e.g., [1966, 1978]).
[0149] Drift Relationship Type Defined into the following five categories:
[0150] Table 3 Classification of Drift Relationships
[0151]
[0152] TTDG is stored in the Neo4j database in the form of a graph. The initial version was constructed by a dozen experts in the field of archives through collaborative annotation (the annotation volume is expected to be about 500-1000 drift triples, covering the main collection period of a certain company). It will be continuously expanded through an automatic mining mechanism.
[0153] TTDG Features: Existing general knowledge graphs (CN-DBpedia, OpenKG, etc.) are all based on modern ontology and do not contain diachronic semantic drift relationships in the archival field. TTDG is an original knowledge structure specifically designed for the evolution of Chinese historical archival language.
[0154] Search triggering mechanisms for terms without modern equivalents:
[0155] For TTDG entries with a drift relationship type of "concept dissolution" (i.e., historical terms no longer have corresponding modern terms), the modern equivalent term field hmod is null. Such terms cannot be retrieved through the reverse lookup link of "modern term → historical equivalent term". This invention employs the following multi-dimensional triggering mechanism to compensate for this deficiency:
[0156] (1) Historical period interval trigger: When the time condition included in the user query matches the historical active time interval of the TTDG entry [ , When there is an overlap, add the historical terms of all "concept dissolution" entries within that time period to the extended query set;
[0157] (2) Event node triggering: Add related historical event nodes to the "concept resolution" category in TTDG. When the user query contains keywords related to the event, the search expansion of the corresponding historical terms is triggered.
[0158] (3) Topic tag trigger: Each "concept resolution" category entry is labeled with a topic tag. When a user queries an entity type identified by NER that matches a certain topic tag, all historical terms in the "concept resolution" category under that topic are triggered.
[0159] (4) File creation time trigger: When the user-defined search time range coincides with the historical terminology activity interval [ , When there is overlap, historical terms related to "concept dissolution" that were active during that time period will be automatically added to the search query as expanded terms.
[0160] The above four triggering methods are based on an "OR" logical relationship. The extension is triggered when any one of the conditions is met, without relying on the reverse lookup of modern equivalence terms.
[0161] Submodule 2: Semi-Auto Drift Mining (SADM) Submodule
[0162] This submodule is used to perform a semi-automatic drift knowledge mining process based on archival corpora:
[0163] Step D1 (Time Slicing): Using the archive timestamp as anchor, divide the archives into time slices in 10-year units. , ,... ;
[0164] Step D2 (Construction of Temporal Frequency Matrix): For high-frequency noun entities in the archives, count their frequency of occurrence in each time slice to construct a temporal frequency matrix. (V represents vocabulary size, and N represents the number of time slices).
[0165] Step D3 (Transition Detection): Perform a chi-square test (significance level p < 0.01) on adjacent slices of the word frequency matrix to detect word pairs with significant frequency transitions, and select word pairs that meet the criteria. , Marked as a drift candidate;
[0166] Drift word pair identification and matching process:
[0167] Word frequency transition detection can only identify "a word showing a significant frequency change in a certain time slice", but it cannot automatically determine the semantic equivalence or drift correspondence between two words. This invention uses the following process to complete the conversion from "transition words" to "drift word pairs":
[0168] (1) Candidate transition word extraction: For each word w in the temporal word frequency matrix M, calculate its adjacent time slices. and The chi-square statistic between the two values is used to mark words that satisfy p < 0.01 as candidate transition words. In this example, approximately 5,200 candidate transition words are extracted from 37,500 file directories through this step.
[0169] (2) Candidate word pairing: For all candidate transition word pairs ( , Candidate drift word pairs are formed when the following conditions are met:
[0170] Table 4 Candidate Drift Word Pairs
[0171]
[0172] (3) Confidence score: For candidate drift word pairs that meet the above three conditions, calculate the comprehensive confidence score:
[0173] Score = 0.3 × (Minimum Chi-square / Maximum Chi-square) + 0.4 × Cosine Similarity + 0.3 × (1 - JS Divergence)
[0174] in:
[0175] Chi-square minimum: the smaller chi-square statistic between the two words.
[0176] Maximum chi-square value: The maximum chi-square statistic of all candidate transition words (used for normalization).
[0177] Cosine similarity: Cosine similarity in the Word2Vec embedding space
[0178] JS divergence: Jensen-Shannon divergence
[0179] (4) Mismatch filtering rules: Candidate word pairs in the following situations are judged as mismatches and filtered:
[0180] The two words belong to different parts of speech (e.g., one is an institution name and the other is a place name), which is determined by the NER annotation results in the archival domain;
[0181] The two words have no overlap in their active time periods and are separated by more than 20 years (excluding accidental cross-era connections).
[0182] The overall confidence score is <0.45.
[0183] (5) Expert confirmation process and data:
[0184] Table 5. Statistical data of each stage of drift word pair identification.
[0185]
[0186] The main reasons for expert rejection included: semantic relevance but not equivalence, specific historical context, and expert judgment in the field of archives deeming it inappropriate to establish a drift relationship. The drift word pairs confirmed by the experts were labeled with drift relationship types (5 categories) and written into the TTDG.
[0187] Step D4 (Ingestion and TTDG Update): After matching, scoring and mismatch filtering in Step D3, the candidate drifting word pairs are finally confirmed by experts in the field of archives, and then automatically written into the TTDG.
[0188] This mechanism transforms the unit cost of expanding TTDG from "manual annotation of each triplet" to "automatic candidate generation + rapid expert confirmation", significantly reducing the cost of knowledge base maintenance.
[0189] Sub-module 3: Era-Aware Query Rewriting & Retrieval (EAQRR)
[0190] This submodule is used to perform the following steps:
[0191] Step E1 (Contemporary Context Detection): Identify the set of modern terms in user query q. }, for each In the TTDG, perform a reverse search for all historical equivalent terms; simultaneously, for historical terms in the TTDG where the modern equivalent term field is null, add the matched historical terms to the extended term set H_null based on at least one of the following matching results: query time condition, event keyword, topic tag, and document creation time.
[0192]
[0193] Step E2 (Constructing the Extended Queryset): Construct the extended queryset Q_exp, which consists of three parts:
[0194]
[0195] Step E3 (Multi-level Hybrid Retrieval): For the expanded query set Each subquery in the query performs two types of retrieval:
[0196] Vector semantic retrieval: The query vector is obtained by calling the BGE-M3 encoded subquery, and the approximate nearest neighbor (ANN) retrieval is performed in the Milvus / FAISS vector library to recall the candidate file set;
[0197] Knowledge graph entity retrieval: The identified entities (including historical equivalent terms) are expanded into one-hop / two-hop neighbors in the archive knowledge graph to supplement related archives;
[0198] Step E4 (Cross-Era Integration Sorting): Sort all candidate files according to the following comprehensive score:
[0199]
[0200] Among them, the era matching reward items are:
[0201]
[0202] If the archive era matches the historical active period of the query extended terms, an additional bonus weight is awarded. (Default 0.15) Improves the ranking of historical records in the results.
[0203] Quantification of technical effects: such as Figure 5a and Figure 5bAs shown, on a test set of approximately 18,000 digitized historical archives (1949-1985) held by a company, EARE improved the retrieval recall from 24.3% (baseline BGE-M3) to 71.8%, an increase of 195.5%; the precision increased from 28.1% to 66.3%; and the average last rank improved from 0.21 to 0.62. The recall of queries containing time-shifted terms increased from 18.7% to 74.2%, an increase of 296.8%. For queries containing time-shifted institutional names (accounting for approximately 42% of historical archive queries), the recall improvement was particularly significant, reaching +218.4%.
[0204] The Multimodal Document Parsing Engine (MAPE) is used to identify scanned documents and output a structured representation of document content, providing the recording field input for the error tracing and repair module of the recording behavior map.
[0205] The multimodal document parsing engine performs semantic block segmentation on the input scanned document, dividing the page into title area, body text area, table area, handwritten annotation area and seal area. After preserving the legal hierarchical structure information of the document, it sends them to the corresponding recognition submodule for processing.
[0206] For example, suppose the multimodal content of a file page is represented as follows:
[0207]
[0208] The components are: printed text, handwritten annotations, graphics and images, table structure, and stamps / signatures.
[0209] This invention proposes a semantic block segmentation strategy for archive pages: before being fed into a multimodal model, the scanned pages of the archives are classified into regions (title area, body text area, table area, annotation area, and seal area) to preserve the legal hierarchical structure information of the archives and avoid misinterpretation of the archive layout structure by the general multimodal model.
[0210] With 2×RTX A6000 (2×48GB) hardware:
[0211] Qwen3-VL-4B-Thinking was used as the multimodal parsing backbone to process four modalities: printed text, handwritten annotations, tables, and images.
[0212] By fine-tuning the LoRA domain (with approximately 5,000 pages of labeled samples and LoRA rank r=16), the accuracy of handwritten annotation recognition was improved from approximately 63% of the general model to 87.2%.
[0213] The seal area uses an independent seal semantic recognition module to extract the seal text content (for institutional registration) and seal type (official seal / private seal / sealed seal), with a confidence level ≥ 0.91.
[0214] The output of this module is a structured archive content description (JSON format), which serves as input for subsequent automatic filling of bibliographic fields. The F1 score of the bibliographic fields is automatically extracted to reach 0.913.
[0215] Hardware feasibility description:
[0216] Qwen3-VL-4B-Thinking (INT8): Approximately 10GB of video memory, can run with a single A6000 card;
[0217] BGE-M3 (Vector Encoding): Approximately 3GB of video memory, can run on the same card as VL models;
[0218] GAT Graph Neural Network (CBGE module): No GPU inference required, CPU operation is sufficient; training consumes approximately 8GB of video memory.
[0219] TTDG Knowledge Graph (EARE Module): Stored in Neo4j, CPU + memory are sufficient;
[0220] The entire system can be fully deployed in a 2×A6000 48GB environment without requiring additional computing power.
[0221] The archival knowledge graph storage module is used to store archival entities and their relationships, providing knowledge graph retrieval support for the era-aware semantic retrieval module.
[0222] The vector retrieval library storage module is used to store the semantic vector representation of archival records, providing vector retrieval support for the era-aware semantic retrieval module.
[0223] In addition, the entire system supports full deployment and operation under a 2×RTX A6000 (48GB×2) hardware configuration without the need for additional large-scale computing power investment: the multimodal parsing model adopts INT8 quantization, the large language model citation inference adopts 4-bit quantization, the graph neural network inference stage is executed on the CPU, and the era-drift knowledge graph and vector retrieval library are decoupled from the GPU and run independently.
[0224] This application also provides a more specific application example, as follows:
[0225] Experimental background: The experimental dataset consists of 12,000 catalog records of the archives of a certain company generated from January to April 2024, spanning 120 days and involving 12 catalogers. The types of archives include administrative archives (8,400 records), scientific and technological archives (2,100 records), and accounting archives (1,500 records).
[0226] Error identification: Field-level rule validation was performed on 12,000 records, and a total of 498 erroneous records and 691 erroneous fields were identified. Among them, 339 were systematic errors (accounting for 68.0% of the erroneous records) and 159 were sporadic errors.
[0227] Root Cause Clustering: 12,000 bibliographic records were merged into behavioral nodes at the granularity of (bibliographer ID, template ID, batch ID) triplets to construct a Bibliographic Behavior Association Graph (CBG), containing 1,243 behavioral nodes and 4,872 behavioral edges. 498 erroneous records were distributed across 387 behavioral nodes. After GAT feature aggregation, Louvain community detection was performed, resulting in 11 error clusters. After manual verification, the errors were attributed to: 4 clusters of Bibliographer Personal Bias (associated with 221 erroneous records); 3 clusters of Template Flaw (associated with 98 erroneous records); and 4 clusters of Batch Errors (associated with 120 erroneous records). An additional 59 scattered errors were partially covered through propagation domain expansion. The overall error coverage rate reached 86.4%.
[0228] Propagation Domain Expansion: A weighted BFS propagation domain expansion (edge weight threshold = 0.55) was performed on the 339 systematic error records, further identifying 118 high-risk warning records. These 118 records were manually reviewed, confirming that 97 of them contained errors originating from the same source, resulting in a warning accuracy rate of 82.2%.
[0229] Repair efficiency: For identified systemic errors, MCRP (Most Cost Repair Path Planning) is performed to generate repair solutions. During the repair process, the number of manual operations required for batch repair of systemic errors is reduced compared to the number of operations required for tracing each error one by one. After implementation, the number of manual operations required for batch repair of systemic errors is reduced by 83.7% compared to the number of operations required for tracing each error one by one; the scope of manual review is reduced from full review of 498 error records to targeted review of only 118 high-risk warning records.
[0230] Furthermore, a time-aware semantic retrieval was conducted, as detailed below:
[0231] Search background: Based on a catalog database of documents and archives held by a certain archive from 1949 to 1990, a total of 37,500 records were recorded.
[0232] Knowledge Graph of Drifting Terms: Five categories of terms were extracted from the archive catalog: archive name, document title, institution name, historical place name, and job title. Based on the chi-square test of word frequency transition (SADM, p<0.01), the drifting relationship of terms across eras was identified, and a knowledge graph of time drifting terms in archives (TTDG) was constructed. It contains 3,842 drifting word pairs and 5 types of drifting relationship, with an average of about 1,280 drifting word pairs per historical period.
[0233] Search performance comparison: 186 query statements containing historical terms (covering political, economic and organizational categories) were selected, and comparative experiments were conducted using BGE-M3 baseline vector retrieval and the EARE era-aware retrieval of this invention.
[0234] In addition, this application also provides an optimization scheme for RTX A6000×2 graphics cards, as shown in Table 6.
[0235] Table 6 shows the optimization scheme for RTX A6000×2 graphics cards.
[0236]
[0237] Total GPU memory budget: GPU0≈12.5GB / GPU1≈17GB. Both cards have sufficient memory to run inference and light training simultaneously.
[0238] Furthermore, the embodiments of this application are compared with the prior art, and the comparison results and the core technical indicators of this application are shown in Tables 7 and 8.
[0239] Table 7 Comparison with Existing Technologies
[0240]
[0241] Table 8 Core Technical Indicators
[0242]
[0243] It should be noted that EARE and MAPE are two different search extension modules, which evaluate the recall and precision metrics respectively; the recall of queries containing time-shifted terms is an independent experimental subset, corresponding to a specific query set containing institution names that have shifted across time (accounting for 42% of the total queries).
[0244] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in Embodiments 1 and 2. Therefore, the parts in this embodiment that are the same as or similar to those in Embodiments 1 and 2 can be referred to each other, and will not be repeated in this application.
Claims
1. A method for tracing and correcting errors in archival record keeping, characterized in that, include: S1. Construct a relationship graph of recording behavior, with the recording behavior unit as the node, and establish weighted connection edges between the nodes based on three conditions: the time homology of the recorder, the homology of the recording template, and the homology of the batch import. S2. Perform field-level rule verification on the bibliographic records and mark the bibliographic behavior units containing errors as error node sets; use the error node sets as seeds on the bibliographic behavior association graph, use graph neural networks to perform feature aggregation, and perform community detection to divide the error nodes into K clusters. Perform root cause determination on each cluster to confirm the root cause type of each cluster. S3. Based on the root cause type of each cluster confirmed in step S2, the clusters are attributed to the corresponding error source category; S4. Starting from each cluster, perform graph traversal on the descriptive behavior association graph to identify high-risk descriptive records generated by root cause propagation corresponding to that cluster but not yet marked, thus forming the propagation influence domain. S5. Within the propagation impact domain, using archival records as the restoration objects, a set overlay algorithm is used to generate a restoration priority list; S6. Write the repair results back to the file management system, extract the field values before and after the repair record, and output the field difference pairs.
2. The method for tracing and correcting archival recording errors according to claim 1, characterized in that, The root cause determination in step S2 includes: calculating the proportion of dominant edge types within a cluster; when the proportion of the dominant edge type exceeds a preset threshold, it is determined as the corresponding root cause type; verifying the consistency of error field types within a cluster; calculating the purity of the three types of homologous features within a cluster; when the purity exceeds a threshold, homologous root cause determination is triggered; outputting the root cause confidence score by weighted scoring based on the above indicators; clusters with confidence scores below the threshold trigger manual verification; the error source categories in step S3 include: when the dominant edge is a time homologous edge, it is attributed to the bibliographer's personal bias; when the dominant edge is a template homologous edge, it is attributed to the bibliographing template defect; when the dominant edge is a batch homologous edge, it is attributed to the batch import error.
3. The method for tracing and correcting archival recording errors according to claim 2, characterized in that, The weight of the weighted connection edge in step S1 is obtained by weighted summation of at least two of the recorder time homology indicator function, template homology indicator function, and batch homology indicator function; in step S2, the optimal number of clusters K is automatically determined by the modularity optimization algorithm.
4. The method for tracing and correcting archival record errors according to claim 3, characterized in that, The graph traversal in step S4 is a weighted breadth-first search, adding neighboring nodes with edge weights greater than a preset threshold to the propagation influence domain; step S6 also includes: establishing a dynamic deviation profile for each logger, calculating a personalized early warning score when a logger submits a new logging record, and triggering a verification prompt when the score exceeds the early warning threshold.
5. A cross-era semantic retrieval method for archives, characterized in that, The method receives the field difference pair output from step S6 of the method according to any one of claims 1 to 4 as input, and includes the following steps: S1. Construct a knowledge graph of the temporal drift of archival terminology, using historical terms, modern equivalent terms, types of drift relationships, and historically active time intervals as elements to express cross-temporal semantic equivalence relationships; S2. After the received field differences pass the confidence assessment, they are incrementally written into the era drift knowledge graph as new word pairs; vector embedding is re-performed on the repaired archive records to update the retrieval vector index incrementally; S3. Based on time slices of archival corpora, drifting candidate word pairs are automatically mined through word frequency transition statistical tests, and then pushed for confirmation and storage after confidence filtering. S4. Perform era-context detection on user queries, and use the era-drift knowledge graph to find equivalent terms for modern terms in various historical periods to construct an extended query set; S5. Perform vector semantic retrieval and knowledge graph entity expansion retrieval on the extended query set, and generate comprehensive retrieval results through a fusion ranking algorithm.
6. The cross-era semantic retrieval method for archives according to claim 5, characterized in that, The fusion ranking in step S5 includes an era matching reward item, which gives additional ranking reward weight when the formation year of the candidate archive falls within the historical active time range of the query expansion terms; the drift relationship type includes a historical term type without modern equivalent terms, which triggers retrieval expansion through at least one of the following: the intersection of the query time condition and the historical active time range, related event keywords, topic tags, and archive formation time.
7. The cross-era semantic retrieval method for archives according to claim 6, characterized in that, The word frequency transition statistical test mentioned in step S3 is as follows: the archives are divided into multiple time slices according to timestamps, a temporal word frequency matrix is constructed, and a statistical significance test is performed on the word frequency changes between adjacent slices. Candidate transition words that pass the test are paired according to temporal complementarity, semantic relevance and contextual distribution similarity. After comprehensive confidence scoring and mismatch filtering, they are pushed to be confirmed and added to the database.
8. An archive management system based on bibliographic behavior mapping and era-aware retrieval, characterized in that, include: The error tracing and repair module for the cataloging behavior graph is used to execute steps S1 to S6 of the method described in claim 1, construct a cataloging behavior association graph, perform error root cause clustering attribution and propagation influence domain expansion, generate a repair path, and extract field difference pairs before and after repair. The era-aware semantic retrieval module is used to execute steps S1 to S5 of the method described in claim 5, receive the field difference pair before and after repair to update the archival terminology era drift knowledge graph, expand the era context of user queries, and combine vector retrieval and knowledge graph retrieval to realize cross-era archival retrieval. A multimodal archive parsing engine is used to identify scanned archive documents and output a structured representation of archive content, providing the recording field input for the error tracing and repair module of the recording behavior map; The archival knowledge graph storage module is used to store archival entities and their relationships, providing knowledge graph retrieval support for the era-aware semantic retrieval module; The vector retrieval library storage module is used to store the semantic vector representation of archival records, providing vector retrieval support for the era-aware semantic retrieval module.
9. The archive management system based on bibliographic behavior mapping and era-aware retrieval according to claim 8, characterized in that, The multimodal document parsing engine performs semantic block segmentation on the scanned document, dividing the page into title area, body text area, table area, handwritten annotation area and seal area, and sends them to the corresponding recognition submodule for processing.
10. The archival management system based on bibliographic behavior mapping and era-aware retrieval according to claim 9, characterized in that, The system supports deployment and operation under dual graphics card configuration: the multimodal parsing model adopts INT8 quantization, the graph neural network inference stage is executed on the central processing unit, and the time-drift knowledge graph and vector retrieval library are decoupled from the graphics processor and run independently.
