A file decomposition method based on multi-dimensional file association analysis and automated interpretation
Patent Information
- Application Number
- CN202510773472.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-11
- Publication Date
- 2026-09-18
- Estimated Expiration
- 2045-06-11
AI Technical Summary
[0002]现有文件解读依赖人工分析或现有系统,存在效率低、智能化不足、动态适应性差等问题
[0014]本发明的有益之处是:
Smart Images

Figure CN120631847B_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a file decomposition method based on multi-dimensional file association analysis and automated interpretation, which involves the intersection of natural language processing and big data analysis. Background Technology
[0002] Current document interpretation relies on manual analysis or existing systems, resulting in low efficiency, insufficient intelligence, and poor dynamic adaptability. Manual interpretation is time-consuming and labor-intensive, and prone to overlooking key information; existing systems lack deep semantic understanding and correlation analysis, making it difficult to capture the interrelationship logic of documents; most systems cannot update their knowledge bases in real time, and document iteration can easily lead to execution risks. Although some automated interpretation tools have emerged, they still do not solve core pain points such as the lack of multi-dimensional correlation analysis and insufficient dynamic decision support. Summary of the Invention
[0003] This invention addresses the problems of existing technologies by providing a document decomposition method based on multi-dimensional document association analysis and automated interpretation. It aims to achieve automated identification and rapid decomposition of policy clauses through a BERT-BiLSTM-CRF hybrid model, improving the efficiency of manual interpretation; constructing a three-layer knowledge graph including entities, relationships, and rules to accurately reveal the cross-domain document linkage logic; developing a document iteration comparison engine and combining it with visualization analysis technology to achieve dynamic early warning of document changes; and integrating system dynamics simulation and quantitative decision-making frameworks to support document effect prediction and multi-scenario simulation. Ultimately, it forms a closed loop of document governance—monitoring, analysis, early warning, and optimization—significantly improving the scientific nature of policy formulation and the agility of implementation response.
[0004] The specific solution proposed in this invention is as follows: This invention provides a file decomposition method based on multi-dimensional file association analysis and automated interpretation, including: Step 1: Perform multi-source heterogeneous data acquisition. Step 2: Intelligently clean the collected data and construct a knowledge graph: Use the BERT-BiLSTM-CRF entity recognition model to extract the implementing entities within the documents from the cleaned data, construct a triplet relationship network of implementing subject-document measure-object, and form a knowledge graph stored in the Neo4j graph database. Step 3: Deep Association Analysis: Using a text vectorization engine, files are mapped to 768-dimensional semantic vectors. A file association matrix is constructed based on cosine similarity. The K-core algorithm is used to select core file nodes. Combined with an improved force-directed layout algorithm, dynamic visualization of the community is achieved. Step 4: Automated Decomposition and Document Simulation: Relying on a professional terminology database and a BiLSTM-CRF model, the system automatically identifies reward / restriction / admission clauses in documents, extracts decision tree rules, and recommends the optimal execution path through a three-level conflict detection algorithm. The document simulator is based on a system dynamics model, pre-sets 12 industry benchmark scenarios, and displays resource flow paths through a Sankey diagram. Step 5: Intelligent Application and Feedback: The structured report engine automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles. It supports GIS map overlay analysis, matches applicable clauses in documents based on the enterprise tag system, pre-screens application materials through OCR+rules engine, collects user feedback cases, and incrementally trains model parameters every month to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
[0005] Furthermore, in step 1 of the file decomposition method based on multi-dimensional file association analysis and automated interpretation, an adaptive crawler system is used to crawl the text of files from 3000+ authoritative websites in real time using dynamic rendering technology (Selenium + Headless Chrome). Anti-crawling mechanisms are avoided through IP rotation and request frequency control. A whitelist mechanism combined with institutional credibility scoring is used to prioritize the acquisition of data from high-weight websites.
[0006] Furthermore, in step 2 of the file decomposition method based on multi-dimensional file association analysis and automated interpretation, the collected data is input into an automated cleaning pipeline. First, version control is performed using the SimHash algorithm to remove duplicate data. Then, a rule engine is used to parse time-sensitive keywords. The BERT-BiLSTM model is used to label the time-sensitive status of the files. Finally, a four-level geocoding library is used to standardize the regional description into a GIS-usable format, laying the foundation for subsequent spatial analysis.
[0007] Furthermore, in step 3 of the file decomposition method based on multi-dimensional file association analysis and automated interpretation, Spark MLlib is used to handle machine learning tasks on large-scale datasets, and TensorFlow is used to build deep learning models for tasks such as file entity recognition, relation extraction, and label prediction. Cross-validation and regularization techniques (L2 / Dropout) are combined to prevent overfitting.
[0008] Gephi is used for topology analysis and visualization of text networks, supporting community discovery and node centrality calculation functions.
[0009] This invention also provides a file disassembly device based on multi-dimensional file association analysis and automated interpretation, including a data acquisition module, a cleaning module, a map construction module, an analysis module, a disassembly module, a simulation module, and an application feedback module. The acquisition module performs multi-source heterogeneous data acquisition. The data cleaning module performs intelligent data cleaning, while the knowledge graph construction module builds a knowledge graph: The knowledge graph construction module uses the BERT-BiLSTM-CRF entity recognition model to extract implementation entities from the cleaned data, constructs a triplet relationship network of implementation subject-document measure-object, and stores the resulting knowledge graph in the Neo4j graph database. Step 3: Deep Association Analysis: Using a text vectorization engine, files are mapped to 768-dimensional semantic vectors. A file association matrix is constructed based on cosine similarity. The K-core algorithm is used to select core file nodes. Combined with an improved force-directed layout algorithm, dynamic visualization of the community is achieved. Step 4: Automated Decomposition and Document Simulation: Relying on a professional terminology database and a BiLSTM-CRF model, the system automatically identifies reward / restriction / admission clauses in documents, extracts decision tree rules, and recommends the optimal execution path through a three-level conflict detection algorithm. The document simulator is based on a system dynamics model, pre-sets 12 industry benchmark scenarios, and displays resource flow paths through a Sankey diagram. Step 5: Intelligent Application and Feedback: The structured report engine automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles. It supports GIS map overlay analysis, matches applicable clauses in documents based on the enterprise tag system, pre-screens application materials through OCR+rules engine, collects user feedback cases, and incrementally trains model parameters every month to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
[0010] Furthermore, the acquisition module of the file disassembly device based on multi-dimensional file association analysis and automated interpretation adopts an adaptive crawler system based on dynamic rendering technology (Selenium + Headless Chrome) to crawl the text of files from 3000+ authoritative websites in real time. It avoids anti-crawling mechanisms by rotating IPs and controlling request frequency, and uses a whitelist mechanism combined with institutional credibility scoring to prioritize the acquisition of data from high-weight websites.
[0011] Furthermore, the cleaning module of the file disassembly device based on multi-dimensional file association analysis and automated interpretation inputs the collected data into the automated cleaning pipeline. First, version control is performed through the SimHash algorithm to remove duplicate data. Then, the timeliness keywords are parsed using a rule engine. The timeliness status of the files is labeled using the BERT-BiLSTM model. Finally, the regional descriptions are standardized into a GIS-usable format using a four-level geocoding library, laying the foundation for subsequent spatial analysis.
[0012] Furthermore, the analysis module of the file disassembly device based on multi-dimensional file association analysis and automated interpretation uses Spark MLlib to handle machine learning tasks on large-scale datasets and uses TensorFlow to build deep learning models for tasks such as file entity recognition, relation extraction, and label prediction. It combines cross-validation and regularization techniques (L2 / Dropout) to prevent overfitting.
[0013] Gephi is used for topology analysis and visualization of text networks, supporting community discovery and node centrality calculation functions.
[0014] The advantages of this invention are: This invention significantly improves document analysis efficiency through automated decomposition, accurately depicts policy linkage networks using knowledge graphs, effectively mitigates implementation risks through dynamic early warning mechanisms, and achieves scientific decision-making through simulation prediction, forming an intelligent management closed loop for the entire document lifecycle, thereby significantly improving governance efficiency and decision-making quality. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example 1
[0017] refer to Figure 1 This invention provides a file decomposition method based on multi-dimensional file association analysis and automated interpretation, including: Step 1: Collect multi-source heterogeneous data. An adaptive crawler system based on dynamic rendering technology (Selenium + Headless Chrome) can be used to crawl the text files of 3000+ authoritative websites in real time. Anti-crawling mechanisms are avoided by rotating IPs and controlling the request frequency. A whitelist mechanism combined with the credibility score of the institution is used to prioritize the acquisition of data from high-weight websites.
[0018] Step 2: Intelligent data cleaning and knowledge graph construction: The collected data is input into an automated cleaning pipeline. First, version control and deduplication are performed using the SimHash algorithm. A rule engine is used to parse time-sensitive keywords, and the BERT-BiLSTM model is used to label the time-sensitive status of documents. A four-level geocoding library is used to standardize regional descriptions into a GIS-usable format, laying the foundation for subsequent spatial analysis. The cleaned data is then processed using the BERT-BiLSTM-CRF entity recognition model to extract implementing entities from the documents. A triplet relationship network of implementing entity-document measure-object is constructed, forming a knowledge graph stored in the Neo4j graph database.
[0019] Step 3: Deep Association Analysis: A text vectorization engine is used to map files into 768-dimensional semantic vectors. A file association matrix is constructed based on cosine similarity. The K-core algorithm is used to select core file nodes, and an improved force-directed layout algorithm is combined to achieve dynamic visualization of the community. Spark MLlib is used for machine learning tasks on large-scale datasets, and TensorFlow is used to build deep learning models for tasks such as file entity recognition, relation extraction, and label prediction. Cross-validation and regularization techniques (L2 / Dropout) are used to prevent overfitting.
[0020] Gephi is used for topology analysis and visualization of text networks, supporting community discovery and node centrality calculation functions.
[0021] Step 4: Automated Decomposition and Document Simulation: Relying on a professional terminology database and a BiLSTM-CRF model, the system automatically identifies reward / restriction / admission clauses in documents, extracts decision tree rules, and recommends the optimal execution path through a three-level conflict detection algorithm. The document simulator is based on a system dynamics model, pre-sets 12 industry benchmark scenarios, and displays resource flow paths through a Sankey diagram. Step 5: Intelligent Application and Feedback: The structured report engine automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles. It supports GIS map overlay analysis, matches applicable clauses in documents based on the enterprise tag system, pre-screens application materials through OCR+rules engine, collects user feedback cases, and incrementally trains model parameters every month to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
[0022] In the application of the method of this invention, a four-layer technology system can be built using a microservice architecture, for example, the specific architecture is as follows: (1) Data layer: Neo4j, a distributed graph database, is used to store policy knowledge graphs and supports fast querying and reasoning about complex relationships. It optimizes the performance of relational queries through native graph storage technology, improving response speed by 40%.
[0023] HBase stores raw policy texts and preprocessed structured data, supporting efficient reading and writing of massive amounts of data and horizontal scaling.
[0024] Data cleaning module: Based on the "Data Cleaning - Deduplication" and "Data Cleaning - Fill Missing Values" functions in Excel, it integrates an automated data cleaning pipeline, including modules for data standardization, filling missing values such as mean / median / interpolation, and noise smoothing such as moving average, to ensure that data quality meets preset standards.
[0025] (2) Analysis layer: Spark MLlib: Responsible for machine learning tasks on large-scale datasets, such as policy classification and cluster analysis, and supports parallel computing.
[0026] TensorFlow: Build deep learning models for tasks such as policy entity recognition, relation extraction, and label prediction, combining cross-validation and regularization techniques (L2 / Dropout) to prevent overfitting.
[0027] Gephi: Enables topology analysis and visualization of policy networks, supporting functions such as community discovery and node centrality calculation.
[0028] (3) Service layer: Flask framework: encapsulates RESTful API interfaces, providing high-concurrency service capabilities of up to 100,000 calls per day, with interface response time controlled within 200ms.
[0029] Quality Control API: Integrates with Excel's "Quality Control and Analysis - Data Validation" function, ensuring the legality of input data through business rule validation (such as policy validity period verification) and integrity validation (key field non-empty check).
[0030] (4) Presentation layer: ECharts visualization component: Generates multi-dimensional charts (such as radar charts and heat maps) to dynamically display policy competitiveness and correlation analysis results.
[0031] AntV Policy Graph Rendering Engine: Based on the force-directed layout algorithm, it supports the visualization and dynamic coloring of policy networks, highlighting core policy nodes. Example 2
[0032] This invention also provides a file disassembly device based on multi-dimensional file association analysis and automated interpretation, including a data acquisition module, a cleaning module, a map construction module, an analysis module, a disassembly module, a simulation module, and an application feedback module. The acquisition module performs multi-source heterogeneous data acquisition. The data cleaning module performs intelligent data cleaning, while the knowledge graph construction module builds a knowledge graph: The knowledge graph construction module uses the BERT-BiLSTM-CRF entity recognition model to extract implementation entities from the cleaned data, constructs a triplet relationship network of implementation subject-document measure-object, and stores the resulting knowledge graph in the Neo4j graph database. Step 3: Deep Association Analysis: Using a text vectorization engine, files are mapped to 768-dimensional semantic vectors. A file association matrix is constructed based on cosine similarity. The K-core algorithm is used to select core file nodes. Combined with an improved force-directed layout algorithm, dynamic visualization of the community is achieved. Step 4: Automated Decomposition and Document Simulation: Relying on a professional terminology database and a BiLSTM-CRF model, the system automatically identifies reward / restriction / admission clauses in documents, extracts decision tree rules, and recommends the optimal execution path through a three-level conflict detection algorithm. The document simulator is based on a system dynamics model, pre-sets 12 industry benchmark scenarios, and displays resource flow paths through a Sankey diagram. Step 5: Intelligent Application and Feedback: The structured report engine automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles. It supports GIS map overlay analysis, matches applicable clauses in documents based on the enterprise tag system, pre-screens application materials through OCR+rules engine, collects user feedback cases, and incrementally trains model parameters every month to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
[0033] The information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.
[0034] Similarly, the device of this invention significantly improves the efficiency of document analysis through automated disassembly, accurately depicts the policy linkage network through knowledge graphs, effectively avoids implementation risks through dynamic early warning mechanisms, and achieves scientific decision-making through simulation prediction, forming an intelligent management closed loop for the entire life cycle of documents, and significantly improving governance efficiency and decision-making quality.
[0035] It should be noted that not all steps and modules in the above processes and device structures are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structure described in the above embodiments can be a physical structure or a logical structure. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.
[0036] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A file disassembly method based on multi-dimensional file association analysis and automated interpretation, characterized by include: Step 1: Perform multi-source heterogeneous data acquisition. Step 2: Perform intelligent cleaning on the collected data and construct a knowledge graph: The collected data is input into an automated cleaning pipeline. First, the SimHash algorithm is used for version control and data deduplication. A rule engine is then used to parse time-sensitive keywords, and the BERT-BiLSTM model is used to label the time-sensitive status of the files. Finally, a four-level geocoding library is used to standardize the regional descriptions into a GIS-usable format, laying the foundation for subsequent spatial analysis. The cleaned data was used to extract the implementing entities within the documents using the BERT-BiLSTM-CRF entity recognition model. A triplet relationship network of implementing entity-document measure-object was constructed, forming a knowledge graph that was stored in the Neo4j graph database. Step 3: Deep Association Analysis: A text vectorization engine maps files to 768-dimensional semantic vectors. A file association matrix is constructed based on cosine similarity. The K-core algorithm is used to select core file nodes. An improved force-directed layout algorithm is combined to achieve dynamic visualization of the community. Spark MLlib is used for machine learning tasks on large-scale datasets. TensorFlow is used to build deep learning models for file entity recognition, relation extraction, and label prediction. Cross-validation and regularization techniques are used to prevent overfitting. Gephi is used for topological analysis and visualization of the text network, supporting community detection and node centrality calculation. Step 4: Automated Decomposition and Document Simulation: Relying on a professional terminology database and a BiLSTM-CRF model, the system automatically identifies reward / restriction / admission clauses in documents, extracts decision tree rules, and recommends the optimal execution path through a three-level conflict detection algorithm. The document simulator is based on a system dynamics model, pre-sets 12 industry benchmark scenarios, and displays resource flow paths through a Sankey diagram. Step 5: Intelligent Application and Feedback: The structured report engine automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles. It supports GIS map overlay analysis, matches applicable clauses in documents based on the enterprise tag system, pre-screens application materials through OCR+rules engine, collects user feedback cases, and incrementally trains model parameters every month to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
2. The file decomposition method based on multi-dimensional file association analysis and automated interpretation according to claim 1 is characterized in that in step 1, an adaptive crawler system is used to crawl the file text of 3000+ authoritative sites in real time based on dynamic rendering technology, and anti-crawling mechanisms are avoided by IP rotation and request frequency control. The whitelist mechanism is combined with the credibility score of the institution to prioritize the acquisition of data from high-weight sites.
3. A file disassembling device based on multi-dimensional file association analysis and automatic interpretation, characterized in that It includes a data acquisition module, a cleaning module, a map construction module, an analysis module, a disassembly module, a simulation module, and an application feedback module. The acquisition module performs multi-source heterogeneous data acquisition; The cleaning module performs intelligent cleaning on the collected data, and the graph construction module constructs a knowledge graph. The cleaning module inputs the collected data into the automated cleaning pipeline. First, it performs version control and deduplication using the SimHash algorithm. Then, it uses a rule engine to parse time-sensitive keywords, combines the BERT-BiLSTM model to label the time status of the files, and uses a four-level geocoding library to standardize the regional descriptions into a GIS-usable format, laying the foundation for subsequent spatial analysis. The graph construction module uses the BERT-BiLSTM-CRF entity recognition model to extract the implementation entities in the documents from the cleaned data, constructs a triplet relationship network of implementation subject-document measures-object of action, and forms a knowledge graph stored in the Neo4j graph database. The analysis module performs deep association analysis: it uses a text vectorization engine to map files into 768-dimensional semantic vectors, constructs a file association matrix based on cosine similarity, uses the K-core algorithm to filter core file nodes, and combines an improved force-directed layout algorithm to achieve dynamic visualization of the community. The analysis module uses Spark MLlib to handle machine learning tasks on large-scale datasets and uses TensorFlow to build deep learning models for tasks such as file entity recognition, relation extraction, and label prediction. It also combines cross-validation and regularization techniques to prevent overfitting. The analysis module uses Gephi for topology analysis and visualization of text networks, and supports community detection and node centrality calculation functions. The disassembly module automates disassembly, and the simulation module performs file simulation: The disassembly module relies on a professional terminology library and a BiLSTM-CRF model to automatically identify reward / restriction / access clauses in the file, extract decision tree rules, and recommend the optimal execution path through a three-level conflict detection algorithm. The file simulator in the simulation module is based on a system dynamics model, presets 12 industry benchmark scenarios, and displays the resource flow path through a Sankey diagram. The application feedback module enables intelligent application and feedback: it automatically generates Markdown documents containing document summaries, clause comparisons, and enterprise profiles through a structured report engine, supports GIS map overlay analysis, matches applicable clauses within documents based on the enterprise tag system, pre-screens application materials through an OCR+rules engine, collects user feedback cases, and incrementally trains model parameters monthly to ensure that adaptation is completed within 24 hours after the release of new documents, forming a closed-loop iterative mechanism of collection-analysis-application-optimization.
4. The file disassembling device based on multi-dimensional file association analysis and automatic interpretation according to claim 3, characterized in that The data collection module uses an adaptive crawler system based on dynamic rendering technology to crawl the text files of 3,000+ authoritative websites in real time. It avoids anti-crawling mechanisms by rotating IPs and controlling the request frequency, and uses a whitelist mechanism combined with institutional credibility scoring to prioritize the acquisition of data from high-weight websites.
Citation Information
Patent Citations
Mirror image type industry linkage engine system and method based on knowledge graph
CN119537864A
Automatic standard term recommendation method
CN119938902A
