Intelligent classification and structured storage method and system for real estate electronic archives

Abstract

This invention relates to the field of electronic real estate records management technology, and discloses a method and system for intelligent classification and structured storage of electronic real estate records. The method involves collecting image data from electronic real estate records, extracting and fusing features, performing preliminary classification based on multimodal fusion features, and storing the records in a structured manner when the classification score FL > classification threshold FY. When the classification score FL ≤ classification threshold FY, the method performs a time-series update of the knowledge graph and stores the records according to the reclassification results. This invention sets up a semantic module, a visual module, and a structural module, and uses a fusion unit to use the semantic embedding vector as a query. It independently performs scaling dot product attention calculations on visual and structural features to generate visual alignment vectors and structural alignment vectors. Finally, these vectors are concatenated and fused, and then subjected to linear transformation and layer normalization to output a multimodal fusion feature vector. This fully mines cross-modal correlation information, thereby improving the accuracy of record classification.

Description

Technical Field

[0001] This invention relates to the field of electronic real estate records management technology, and more specifically to a method and system for intelligent classification and structured storage of electronic real estate records. Background Technology

[0002] Electronic real estate archives refer to the general term for various types of information, such as ownership certificates, registration applications, approval documents, survey maps, mortgage contracts, and seizure documents, which are recorded and stored in electronic form by real estate registration agencies, financial institutions, judicial departments and other relevant entities in the process of handling real estate rights establishment, alteration, transfer and cancellation. With the deepening of the unified real estate registration system and the widespread application of "Internet + government services", the scale of electronic real estate archives is growing exponentially, so it is necessary to classify and store electronic real estate archives. Existing technologies for processing electronic real estate archives typically rely on single-modal features, ignoring the complementarity between multimodal information within the archives, leading to low classification accuracy. Furthermore, real estate archive types dynamically increase or change with policy and regulatory adjustments and the introduction of solutions to historical issues. Existing classification models are mostly statically trained, meaning they cannot automatically identify new types once deployed, resulting in high maintenance costs and slow response times. When faced with missing modalities, existing technologies often discard the archive directly or cause classification failure, lacking effective means to compensate for missing information, thus reducing practicality and fault tolerance. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a method and system for intelligent classification and structured storage of electronic real estate archives, so as to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent classification and structured storage of electronic real estate archives, comprising the following steps: Step S1: Collect image data from the electronic real estate archives and extract features; Step S2: Fuse the extracted feature data to output multimodal fusion features; Step S3: Perform preliminary classification based on multimodal fusion features; Step S4: After preliminary classification, when the classification score FL > classification threshold FY, the preliminary classification results are stored in a structured manner. Step S2: When the classification score FL ≤ classification threshold FY, perform a time-series update of the real estate archive classification knowledge graph, reclassify, and store according to the reclassification results.

[0005] The intelligent classification and structured storage system for electronic real estate archives includes a data acquisition unit, a feature unit, a fusion unit, a classification unit, a judgment unit, a map unit, a storage unit, a retrieval unit, a gap-filling unit, and a mandatory unit. The data acquisition unit is used to acquire electronic real estate archive data. The feature unit extracts features from the electronic real estate archive data. The fusion unit performs multimodal fusion of the extracted features. The classification unit performs preliminary classification based on the multimodal fusion feature vector output by the fusion unit. The judgment unit judges the preliminary classification results. The map unit performs reclassification. The storage unit is used to structure and store the pre-classified and reclassified archive data. The retrieval unit performs data retrieval. The gap-filling unit is used to supplement features within the feature unit. The mandatory unit is used to force updates to the map unit. When the feature unit is used to extract features from archival data, the feature unit includes a semantic module, a visual module, and a structural module. The semantic module uses the BERT language model to extract semantic features from the text content in the archival data and generates a text semantic embedding vector. The visual module uses the visual Transformer model to extract visual features from the image layout of the archival document and generates an image layout feature matrix. The result module uses a layout-aware model to extract the structural features of the archival document and generates a layout structure feature matrix.

[0006] In a preferred embodiment, the semantic embedding vector within the feature unit of the fusion unit is used as a query. Scaling dot product attention calculations are performed independently on the image layout feature matrix output by the visual module and the layout structure feature matrix output by the structure module. After obtaining the attention weights of their respective domains, they are weighted and aggregated to generate visual alignment vectors and structure alignment vectors. The original semantic embedding vector is then concatenated with the two alignment vectors. Finally, a linear transformation layer and layer normalization are applied to output the multimodal fusion feature vectors of the three modules working together.

[0007] In a preferred embodiment, the classification unit inputs the multimodal fusion feature vector generated by cross-modal attention fusion into a feedforward neural network composed of multiple fully connected layers. The output of the last layer is the original logical value with the same dimension as the total number of file types. The output is normalized to the probability distribution of each category by the Softmax function. The category index with the highest probability is taken as the preliminary classification result, and the maximum probability value is output as the classification score FL.

[0008] In a preferred embodiment, the determination unit receives a classification score FL and compares it with its internal classification threshold FY. When the classification score FL > the classification threshold FY, the determination unit sends the preliminary classification result to the storage unit. When the classification score FL ≤ the classification threshold FY, the determination unit sends an update instruction to the map unit.

[0009] In a preferred embodiment, when calculating the classification threshold FY, the determination unit uses a dynamic threshold function based on the archive timestamp and the knowledge graph evolution time to calculate the classification threshold FY. The calculation formula for the dynamic threshold function is as follows: In the formula, YS is the system's preset basic threshold, λ is the maximum adjustment amplitude coefficient, γ is the time decay rate, T is the generation timestamp of the file to be classified, stored in a standardized timestamp format, and T0 is the timestamp of the knowledge graph's most recent evolution update.

[0010] In a preferred embodiment, the graph unit is provided with a real estate archive classification knowledge graph G_t=(V_t,E_t), where V_t is the set of archive type nodes at timestamp t, E_t is the set of associated edges between archive types, and each edge carries a type evolution weight w_t. When the graph unit receives an update instruction, it uses a time-varying graph neural network to update the real estate archive classification knowledge graph in a time sequence. It learns the evolutionary relationship between newly added type nodes and existing type nodes through a graph attention mechanism, updates the evolution weights of the associated edges, and generates an evolved classification knowledge graph G_{t+1}. The graph unit uses the evolved classification knowledge graph to reclassify the archive data, and sends the reclassification results to the storage unit.

[0011] In a preferred embodiment, the storage unit receives preliminary classification results and reclassification results, and stores the archive data in a structured manner according to the preliminary classification results and reclassification results. When the retrieval unit performs a retrieval, the retrieval unit displays all the archive data before classification according to the classification method.

[0012] In a preferred embodiment, when any module in the semantic module, visual module, or structural module is unable to extract effective features due to incomplete archival data, the missing module initiates a modal attention masking mechanism. Based on the type of missing modality, it retrieves the modal complementarity prior model of the historical complete archival data, uses the extracted modal features to reconstruct the missing modality, and generates a replacement feature vector.

[0013] In a preferred embodiment, the mandatory unit is used for mandatory updates of the graph unit. When the mandatory unit detects new real estate registration types, policy and regulation changes, and the release of solutions for handling historical issues, the mandatory unit sends a mandatory instruction to the graph unit. When the graph unit receives the mandatory instruction, it no longer compares the classification score FL with the classification threshold FY, but directly performs incremental graph neural network updates, performs evolution calculations on the local subgraphs affected by the change events, and generates the evolved knowledge graph G_{t+1}.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention sets up a semantic module, a visual module, and a structural module, and uses a fusion unit to use the semantic embedding vector as a query. It independently performs scaling dot product attention calculation on visual features and structural features to generate visual alignment vectors and structural alignment vectors. Finally, the vectors are concatenated and fused, and then subjected to linear transformation and layer normalization to output a multimodal fusion feature vector. This fully explores cross-modal correlation information and improves the accuracy of file classification. 2. This invention sets up a knowledge graph for real estate archive classification and uses a time-varying graph neural network for time-series updates. It can automatically trigger the graph unit to learn the evolutionary relationship between new type nodes and existing type nodes using a graph attention mechanism, update the evolution weight of associated edges, generate an evolved classification knowledge graph, and reclassify the archives based on the new graph. The mandatory unit can directly enforce incremental graph neural network updates when major policy changes are detected, and continuously improve classification capabilities without manual intervention. 3. The present invention sets up a modal attention masking mechanism in the missing unit. When any module in the semantic module, visual module or structural module cannot extract effective features due to incomplete archival data, the missing unit automatically retrieves the modal complementary prior model of the historical complete archival data according to the type of missing modality, and uses the extracted modal features to reconstruct the features of the missing modality to generate alternative feature vectors, thereby ensuring the smooth execution of the subsequent multimodal fusion and classification process. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the classification and storage method of the present invention.

[0016] Figure 2 This is a schematic diagram of the overall system composition of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The intelligent classification and structured storage method and system for electronic real estate archives involved in the present invention are not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1 This invention provides a method and system for intelligent classification and structured storage of electronic real estate records. The method for intelligent classification and structured storage of electronic real estate records includes the following steps: Step S1: Collect image data from the electronic real estate archives and extract features; Step S2: Fuse the extracted feature data to output multimodal fusion features; Step S3: Perform preliminary classification based on multimodal fusion features; Step S4: After preliminary classification, when the classification score FL > classification threshold FY, the preliminary classification results are stored in a structured manner. Step S2: When the classification score FL ≤ classification threshold FY, perform a time-series update of the real estate archive classification knowledge graph, reclassify, and store according to the reclassification results.

[0019] In this embodiment, the application achieves an automated closed loop from original files to final storage by sequentially executing image acquisition and feature extraction, multimodal feature fusion, preliminary classification, scoring determination, and dynamic reclassification. When the classification score is lower than the threshold, the application automatically triggers the knowledge graph temporal update and reclassification to avoid the fixed storage of misclassified data, ensuring the accuracy of classification and the system's adaptability while maintaining the uniformity of the storage structure.

[0020] Reference Figure 2 The intelligent classification and structured storage system for electronic real estate archives includes a data acquisition unit, a feature unit, a fusion unit, a classification unit, a judgment unit, a graph unit, a storage unit, a retrieval unit, a gap-filling unit, and a mandatory unit. The data acquisition unit is used to acquire electronic real estate archive data. The feature unit extracts features from the electronic real estate archive data. The fusion unit performs multimodal fusion of the extracted features. The classification unit performs preliminary classification based on the multimodal fusion feature vector output by the fusion unit. The judgment unit judges the preliminary classification results. The graph unit performs reclassification. The storage unit is used to structure and store the pre-classified and reclassified archive data. The retrieval unit performs data retrieval. The gap-filling unit is used to supplement features within the feature unit. The mandatory unit is used to force updates to the graph unit. When the feature unit is used to extract features from archival data, the feature unit includes a semantic module, a visual module, and a structural module. The semantic module uses the BERT language model to extract semantic features from the text content in the archival data and generates a text semantic embedding vector. The visual module uses the visual Transformer model to extract visual features from the image layout of the archival document and generates an image layout feature matrix. The result module uses a layout-aware model to extract the structural features of the archival document and generates a layout structure feature matrix.

[0021] In this embodiment, by employing the BERT language model, the visual Transformer model, and the layout-aware model respectively, this application can extract high-quality feature representations from three independent dimensions: text semantics, image layout, and document layout structure. The combination of these three enables this application to comprehensively capture multimodal information in real estate archives, laying a solid foundation for subsequent cross-modal fusion and accurate classification, and overcoming the problem of insufficient single-modal feature expression capabilities.

[0022] Reference Figure 2 The semantic embedding vector within the feature unit of the fusion unit is used as a query. Scaling dot product attention calculation is performed independently on the image layout feature matrix output by the visual module and the layout structure feature matrix output by the structure module. After obtaining the attention weights of their respective domains, they are weighted and aggregated to generate visual alignment vectors and structure alignment vectors. The original semantic embedding vector is then concatenated with the two alignment vectors. Finally, a linear transformation layer and layer normalization are applied to output the multimodal fusion feature vectors of the three modules working together.

[0023] In this embodiment, the application adopts a cross-modal attention fusion mechanism, guided by semantic features, to dynamically align visual and layout features. This enables the application to focus on the image regions and structural positions most relevant to the text semantics, and after splicing and fusion, linear transformation and layer normalization are performed to output a fusion feature vector of three modules working together. This improves the utilization efficiency of cross-modal complementary information and enhances the representation ability of classification features.

[0024] Reference Figure 2 The classification unit inputs the multimodal fusion feature vector generated by cross-modal attention fusion into a feedforward neural network composed of multiple fully connected layers (interspersed with ReLU activation and Dropout regularization). The output of the last layer is the original logical value with the same dimension as the total number of file types. It is normalized to the probability distribution of each category by the Softmax function. The category index with the highest probability is taken as the preliminary classification result, and the highest probability value is used as the classification score FL output.

[0025] In this embodiment, multimodal fusion features are input into a multilayer feedforward neural network and ReLU activation and Dropout regularization are used to effectively prevent overfitting. Finally, a normalized class probability distribution is output through Softmax, and the maximum probability value is used as the classification score output, which provides a reliable uncertainty measure for subsequent decision-making units and facilitates confidence-driven dynamic decision-making.

[0026] Reference Figure 2The determination unit receives the classification score FL and compares it with its internal classification threshold FY. When the classification score FL > the classification threshold FY, the determination unit sends the preliminary classification result to the storage unit. When the classification score FL ≤ the classification threshold FY, the determination unit sends an update command to the graph unit. When calculating the classification threshold FY, the determination unit uses a dynamic threshold function based on the archive timestamp and the knowledge graph evolution time. The formula for the dynamic threshold function is as follows: In the formula, YS is the system's preset basic threshold, λ is the maximum adjustment amplitude coefficient, γ is the time decay rate, T is the generation timestamp of the file to be classified, stored in a standardized timestamp format, and T0 is the timestamp of the knowledge graph's most recent evolution update.

[0027] In this embodiment, the automatically output classification score FL is compared and analyzed with the classification threshold FY. When the classification score FL > the classification threshold FY, it is automatically stored. When the classification score FL ≤ the classification threshold FY, it enters the knowledge graph for updating and reclassification. This application ensures processing efficiency while avoiding the solidification of misclassifications, thus guaranteeing the accuracy and efficiency of classification. Furthermore, this application introduces a dynamic threshold function based on the archive timestamp and the knowledge graph evolution timestamp. This allows the classification confidence requirement to be dynamically adjusted according to the interval between the archive generation time and the knowledge graph update time. Archives that are further away from the most recent update time of the knowledge graph have lower thresholds and are more likely to pass the initial classification directly, reducing unnecessary reclassification calculations. Conversely, newly generated archives use higher thresholds, encouraging more use of the latest knowledge graph for reclassification.

[0028] Reference Figure 2 The graph unit contains a real estate archive classification knowledge graph G_t=(V_t,E_t), where V_t is the set of archive type nodes at timestamp t, E_t is the set of associated edges between archive types, and each edge carries a type evolution weight w_t. When the graph unit receives an update instruction, it uses a time-varying graph neural network to update the real estate archive classification knowledge graph in a time sequence. It learns the evolutionary relationship between newly added type nodes and existing type nodes through a graph attention mechanism, updates the evolution weights of the associated edges, and generates an evolved classification knowledge graph G_{t+1}. The graph unit uses the evolved classification knowledge graph to reclassify the archive data, and sends the reclassification results to the storage unit.

[0029] In this embodiment, a time-varying graph neural network is used to perform time-series evolution learning on the knowledge graph of real estate archive classification. The evolutionary relationship between newly added type nodes and existing nodes is automatically extracted through the graph attention mechanism, and the evolution weight of the associated edges is dynamically updated. Therefore, this application can adaptively learn the addition and change of archive types online, and complete knowledge transfer without offline retraining, so as to realize the continuous evolution of the classification model and reduce the cost of manual maintenance.

[0030] Reference Figure 2 The storage unit receives the preliminary classification results and the reclassification results, and stores the archive data in a structured manner according to the preliminary classification results and the reclassification results. When storing, the storage unit uses the real estate unit as the top-level node and establishes four hierarchical layers: basic information layer, ownership subject layer, rights information layer, and business process layer. For each archive file in each layer, a multi-level retrieval index is constructed using the minimum hash inverted index algorithm. The minimum hash inverted index maps similar files to nearby index buckets by calculating the minimum hash signature of the contents of each archive file. For the relationship between archive files across layers, a directed weighted edge is used to model and store it in the relationship table. When the retrieval unit performs a retrieval, the retrieval unit displays all the archive data before classification according to the classification method.

[0031] In this embodiment, the classification results are associated with and stored with the archive data, and the entire data is presented according to the classification method during retrieval. This allows users to quickly browse and locate archives by category. The structured storage method facilitates subsequent data statistics, analysis and export, maintains consistency between the original archives and classification tags, and improves the convenience and traceability of archive management.

[0032] Reference Figure 2 When any module in the semantic module, visual module, or structural module is unable to extract effective features due to incomplete archival data, the missing module activates a modal attention masking mechanism. Based on the type of missing modality, it retrieves the modal complementarity prior model of historical complete archival data, uses the extracted modal features to reconstruct the missing modality, and generates a replacement feature vector.

[0033] In this embodiment of the application, when encountering the common problem of incomplete data in real estate archives, the application adopts a modal attention masking mechanism. When any modal feature extraction fails, the application automatically uses a prior model trained on historical complete archive data to reconstruct the missing modality based on the existing modal features and generate alternative feature vectors. This application can avoid classification failure caused by missing data.

[0034] Reference Figure 2The mandatory unit is used for mandatory updates of the graph unit. When the mandatory unit detects new real estate registration types, policy and regulatory changes, or the release of solutions to historical issues, it sends a mandatory command to the graph unit. Upon receiving the mandatory command, the graph unit no longer compares the classification score FL with the classification threshold FY, but directly performs incremental graph neural network updates, performing evolution calculations on the local subgraphs affected by the change events to generate the evolved knowledge graph G_{t+1}. In this embodiment, when the application detects significant policy and regulatory changes, new real estate registration types, or the release of solutions to historical issues, it can directly trigger incremental local evolution updates of the knowledge graph without waiting for the classification score to fall below a threshold. This enables the application to respond instantly to changes in the external environment and keeps the classification model synchronized with the latest business rules. It is particularly suitable for scenarios where policies in the real estate management field are frequently adjusted, ensuring that the application is always in a relatively accurate state.

[0035] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. The units and algorithm steps of the various examples described in the embodiments can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0036] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0037] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0038] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent classification and structured storage of electronic real estate archives, characterized by: Includes the following steps: Step S1: Collect image data from the electronic real estate archives and extract features; Step S2: Fuse the extracted feature data to output multimodal fusion features; Step S3: Perform preliminary classification based on multimodal fusion features; Step S4: After preliminary classification, when the classification score FL > classification threshold FY, the preliminary classification results are stored in a structured manner. Step S2: When the classification score FL ≤ classification threshold FY, perform a time-series update of the real estate archive classification knowledge graph, reclassify, and store according to the reclassification results.

2. A smart classification and structured storage system for electronic real estate archives, characterized in that: The system includes a data acquisition unit, a feature acquisition unit, a fusion unit, a classification unit, a judgment unit, a graphing unit, a storage unit, a retrieval unit, a gap-filling unit, and a mandatory unit. The data acquisition unit is used to acquire electronic real estate archive data. The feature acquisition unit extracts features from the electronic real estate archive data. The fusion unit performs multimodal fusion of the extracted features. The classification unit performs preliminary classification based on the multimodal fusion feature vector output by the fusion unit. The judgment unit judges the preliminary classification results. The graphing unit performs reclassification. The storage unit is used to structurally store the pre-classified and reclassified archive data. The retrieval unit performs data retrieval. The gap-filling unit is used to supplement features within the feature acquisition unit. The mandatory unit is used to force updates to the graphing unit. When the feature unit is used to extract features from archival data, the feature unit includes a semantic module, a visual module, and a structural module. The semantic module uses the BERT language model to extract semantic features from the text content in the archival data and generates a text semantic embedding vector. The visual module uses the visual Transformer model to extract visual features from the image layout of the archival document and generates an image layout feature matrix. The result module uses a layout-aware model to extract the structural features of the archival document and generates a layout structure feature matrix.

3. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: The semantic embedding vector within the feature unit of the fusion unit is used as a query. Scaling dot product attention calculation is performed independently on the image layout feature matrix output by the visual module and the layout structure feature matrix output by the structure module. After obtaining the attention weights of their respective domains, they are weighted and aggregated to generate visual alignment vectors and structure alignment vectors. The original semantic embedding vector is then concatenated with the two alignment vectors. Finally, a linear transformation layer and layer normalization are applied to output the multimodal fusion feature vectors of the three modules working together.

4. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: The classification unit inputs the multimodal fusion feature vector generated by cross-modal attention fusion into a feedforward neural network composed of multiple fully connected layers. The output of the last layer is the original logical value with the same dimension as the total number of file types. The output is normalized to the probability distribution of each category by the Softmax function. The category index with the highest probability is taken as the preliminary classification result, and the maximum probability value is output as the classification score FL.

5. The intelligent classification and structured storage system for electronic real estate archives according to claim 4, characterized in that: The determination unit receives the classification score FL and compares it with its internal classification threshold FY. When the classification score FL > the classification threshold FY, the determination unit sends the preliminary classification result to the storage unit. When the classification score FL ≤ the classification threshold FY, the determination unit sends an update command to the map unit.

6. The intelligent classification and structured storage system for electronic real estate archives according to claim 5, characterized in that: When calculating the classification threshold FY, the determination unit uses a dynamic threshold function based on the archive timestamp and the knowledge graph evolution time. The formula for the dynamic threshold function is as follows: In the formula, YS is the system's preset basic threshold, λ is the maximum adjustment amplitude coefficient, γ is the time decay rate, T is the generation timestamp of the file to be classified, stored in a standardized timestamp format, and T0 is the timestamp of the knowledge graph's most recent evolution update.

7. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: The graph unit contains a real estate archive classification knowledge graph G_t=(V_t,E_t), where V_t is the set of archive type nodes at timestamp t, E_t is the set of associated edges between archive types, and each edge carries a type evolution weight w_t. When the graph unit receives an update instruction, it uses a time-varying graph neural network to update the real estate archive classification knowledge graph in a temporal sequence. It learns the evolutionary relationship between newly added type nodes and existing type nodes through a graph attention mechanism, updates the evolution weights of the associated edges, and generates an evolved classification knowledge graph G_{t+1}. The graph unit uses the evolved classification knowledge graph to reclassify the archive data, and sends the reclassification results to the storage unit.

8. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: The storage unit receives the preliminary classification results and the reclassification results, and stores the archive data in a structured manner according to the preliminary classification results and the reclassification results. When the retrieval unit performs a retrieval, it displays all the archive data before classification according to the classification method.

9. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: When any module in the semantic, visual, or structural modules is unable to extract effective features due to incomplete archival data, the missing feature unit activates a modal attention masking mechanism. Based on the type of missing modality, it retrieves the modal complementarity prior model of the complete historical archival data, uses the extracted modal features to reconstruct the missing modality, and generates a replacement feature vector.

10. The intelligent classification and structured storage system for electronic real estate archives according to claim 2, characterized in that: The mandatory unit is used for mandatory updates of the graph unit. When the mandatory unit detects new real estate registration types, policy and regulation changes, and the release of solutions for handling historical issues, it sends a mandatory instruction to the graph unit. When the graph unit receives the mandatory instruction, it no longer compares the classification score FL with the classification threshold FY, but directly performs incremental graph neural network updates, performs evolution calculations on the local subgraphs affected by the change events, and generates the evolved knowledge graph G_{t+1}.

Images