Multi-algorithm collaborative intelligent document analysis system

Abstract

The invention relates to the technical field of information processing and artificial intelligence, in particular to a multi-algorithm collaborative intelligent document analysis system, which comprises a graph construction step: extracting text primitives and spatial position coordinates under a pixel coordinate system, and constructing a heterogeneous graph data structure; hypothesis evolution: generating multiple groups of structured mapping hypotheses; a logic verification step: calculating a total logic cost value representing a logic conflict degree; an iterative optimization step: comparing the total logic cost value with a preset convergence threshold; if the total logic cost value is greater than the convergence threshold and the current number of iterations does not reach a preset maximum value, returning to execute the hypothesis evolution step; if the total logic cost numerical value is smaller than or equal to a convergence threshold value, or the current number of iterations reaches a preset maximum value, structured analysis data are output; according to the method, the interpretability of the algorithm and the reliability of engineering delivery are remarkably improved, and the data delivery integrity and the system stability in a complex network environment and a physical storage process are ensured.

Description

Technical Field

[0001] This invention relates to the fields of information processing and artificial intelligence technology, specifically to a multi-algorithm collaborative document intelligent parsing system. Background Technology

[0002] Intelligent document parsing is a key step in converting unstructured documents into structured data. The efficiency and accuracy of parsing directly affect the operational performance of systems such as automated office systems and financial auditing systems. The parsing process mainly includes text primitive extraction, spatial layout modeling, and logical semantic mapping, and achieves page layout restoration through the correlation analysis of character blocks within the document. Existing technologies mainly rely on serial processing flows or fixed heuristic rules, which cannot fully utilize the multi-source logical information such as arithmetic coherence and semantic coherence within the document for closed-loop decision-making, making it difficult to adapt to non-standard documents with complex and ever-changing layouts. Furthermore, the passive response recognition mode struggles to implement dynamic error correction when visual cues are weak, and cannot effectively cope with the impact of environmental factors such as blurry printing and borderless interference. In addition, the spatial topology of document layouts exhibits significant non-linear characteristics and logical coupling effects, making it difficult for existing technologies to capture the dynamic evolution patterns between text primitives, resulting in limited parsing accuracy and robustness, especially when the layout structure undergoes drastic changes or logical conflicts occur. Therefore, a solution is urgently needed to address the problems of poor document parsing logical self-consistency and low layout restoration accuracy in existing technologies. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a multi-algorithm collaborative document intelligent parsing system. Specifically, the technical solution of this invention includes: The memory stores computer programs, and the processor executes the computer programs to perform the following steps: The graph construction steps are as follows: obtain the layout data of the document to be parsed, extract the text primitives and their spatial coordinates in the pixel coordinate system, calculate the Euclidean distance and projection overlap rate between any two text primitives, and construct a heterogeneous graph data structure with text primitives as nodes and distance and overlap rate as edge weights. Hypothesis evolution steps: Based on the topological features of the heterogeneous graph data structure, multiple sets of structured mapping hypotheses are generated. Each set of hypotheses defines the parent-child hierarchical relationship between nodes and data attribute labels. Logical verification steps: Inject the semantic content of the text primitives into the structured mapping hypothesis, and calculate the total logical cost value representing the degree of logical conflict by weighted summation; Iterative optimization steps: Compare the total logic cost with the preset convergence threshold: If the total logical cost is greater than the convergence threshold and the current iteration number has not reached the preset maximum value, then based on the characteristics of the local region that generates the high cost value, perform a topology reconstruction operation of node splitting or node aggregation on the heterogeneous graph data structure, and return to the execution hypothesis evolution step. If the total logical cost is less than or equal to the convergence threshold, or if the current iteration count reaches the preset maximum value, then the structured mapping hypothesis with the minimum total logical cost is locked as the final model, and structured parsed data is output.

[0004] Optionally, the processor executes the following logic when calculating the projection overlap rate: Obtain the projection intervals of two text primitives on the horizontal coordinate axis; Calculate the intersection length and union length of two projection intervals; The ratio obtained by dividing the intersection length by the union length is determined as the projection overlap rate. Only when the projection overlap rate is greater than zero, the corresponding connection edge is established in the heterogeneous graph data structure.

[0005] Optionally, the processor performs the following steps when calculating the total logical cost: Calculate the arithmetic consistency cost component, which represents the validity of numerical calculation logic; the semantic coherence cost component, which represents the fluency of natural language text; and the data format compliance cost component, which represents whether the field content conforms to the predefined regular expression format. Obtain the preset weight coefficient table, which defines the normalized weights of the above three components for different document types. The total logical cost is obtained by multiplying each component by its corresponding normalized weight and summing the results.

[0006] Optionally, the processor executes the following logic when calculating the arithmetic consistency cost component: Identify the first set of nodes with sub-item attribute labels and the second set of nodes with total attribute labels in the structured mapping hypothesis; Extract the text values ​​of all nodes in the first node set and calculate the cumulative sum; Calculate the absolute value of the difference between the accumulated sum and the text value of the second node; The absolute value of the difference is mapped to a value between 0 and 1, which is used as the arithmetic consistency cost component.

[0007] Optionally, the processor executes the following logic when computing the semantic coherence cost components: Extract text sequences marked as multi-line concatenation in the structured mapping hypothesis; Input the text sequence into a pre-trained language probability model to obtain the probability value of the text sequence as natural language generation; The reciprocal of the probability value is used as an index of perplexity. The perplexity index is used as a component of semantic coherence cost, where a higher perplexity index indicates a higher degree of semantic logical conflict.

[0008] Optionally, the processor performs the following steps when executing the topology reconstruction operation: Node aggregation: In response to feedback signals that have low semantic coherence cost but high arithmetic consistency cost, identify multiple independent nodes that are spatially adjacent and semantically continuous, merge them into an aggregate node, and update the boundary coordinates of the aggregate node. Node splitting: In response to feedback signals that data format compliance costs are high, a single node containing multiple data format characteristics is identified and split into multiple independent child nodes based on the position of the whitespace delimiter; Reconstruction Update: Based on the node states after aggregation or splitting, recalculate the edge weights of the involved nodes and update the adjacency matrix of the heterogeneous graph data structure.

[0009] Optionally, it is assumed that during the evolutionary steps, the processor specifically executes: A probabilistic statistical model is used to perform subgraph isomorphism search on heterogeneous graph data structures; Calculate the probability value of matching the identified sub-graph structure with the preset table template, column text template or paragraph template; A preset number of candidate tree structures with the highest matching probability values ​​are selected as the structured mapping hypothesis.

[0010] Optionally, the processor is also configured to perform the following when outputting structured parsed data: Generate confidence metadata containing logical decision paths; Convert structured parsed data into JSON or XML format files; Write the file to physical storage media or send it to downstream business systems via a network interface; If a write failure or network timeout is detected during the output process, a retry operation is performed and an error log is logged.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention establishes a closed-loop iterative graph optimization mechanism, breaking the limitation of continuous error propagation; it transforms document parsing from a traditional unidirectional serial pipeline into a dynamic iterative process based on hypothesis-verification; by constructing a heterogeneous graph data structure and continuously optimizing topological features, the system can use the internal logical rules of the document to reverse correct the initial visual recognition results, significantly improving the parsing accuracy and robustness in scenarios with weak visual cues such as frameless tables and complex columns; 2. This invention achieves deep integration of multimodal constraints, ensuring high self-consistency of the parsing results; comprehensively considers arithmetic consistency, semantic coherence, and data format compliance, transforming business rules into quantifiable logical costs; by simulating the logical judgment of professionals on account balance, contextual fluency, and format standardization, it can automatically identify and intercept illogical parsing assumptions, making the parsing output not just a collection of characters, but rigorous data that conforms to business logic; 3. This invention possesses adaptive topology reconstruction capabilities, flexibly handling complex non-standard layouts; it can dynamically execute node splitting or aggregation operations based on the conflict characteristics of logical verification feedback; this mechanism empowers the system to dynamically adjust the visual focus according to logical understanding requirements, and can automatically repair numerical fragmentation caused by excessive visual segmentation or field adhesion caused by insufficient segmentation; this dynamic correction significantly enhances the system's adaptability to different print qualities, layout formats, and image resolutions. 4. This invention significantly improves the interpretability of the algorithm and the reliability of its engineering delivery; while outputting structured data, the system can generate confidence metadata containing logical decision paths, transforming the parsing process from a black box to a traceable white box, meeting the stringent requirements for data evidence in fields such as financial auditing and legal compliance; coupled with a robust exception handling and retry mechanism, it ensures the integrity of data delivery and system stability in complex network environments and physical storage processes. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1: Please see Figure 1 A multi-algorithm collaborative document intelligent parsing system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: The graph construction steps are as follows: obtain the layout data of the document to be parsed, extract the text primitives and their spatial coordinates in the pixel coordinate system, calculate the Euclidean distance and projection overlap rate between any two text primitives, and construct a heterogeneous graph data structure with text primitives as nodes and distance and overlap rate as edge weights. Hypothesis evolution steps: Based on the topological features of the heterogeneous graph data structure, multiple sets of structured mapping hypotheses are generated. Each set of hypotheses defines the parent-child hierarchical relationship between nodes and data attribute labels. Logical verification steps: Inject the semantic content of the text primitives into the structured mapping hypothesis, and calculate the total logical cost value representing the degree of logical conflict by weighted summation; Iterative optimization steps: Compare the total logic cost with the preset convergence threshold: If the total logical cost is greater than the convergence threshold and the current iteration number has not reached the preset maximum value, then based on the characteristics of the local region that generates the high cost value, perform a topology reconstruction operation of node splitting or node aggregation on the heterogeneous graph data structure, and return to the execution hypothesis evolution step. If the total logical cost is less than or equal to the convergence threshold, or if the current iteration count reaches the preset maximum value, then the structured mapping hypothesis with the minimum total logical cost is locked as the final model, and structured parsed data is output.

[0015] This embodiment discloses a document intelligent parsing system with multi-algorithm collaboration; the system is based on a hypothesis-verification iterative optimization mechanism, which transforms the document parsing task from a traditional serial pipeline into a closed-loop iterative graph optimization problem; The memory and processor form the physical foundation of this system; the processor is configured to perform the following core steps: Atlas construction steps: The processor performs atlas initialization; text primitives refer to the smallest semantic units extracted by OCR, i.e., optical character recognition technology. They are usually independent character blocks or word blocks, which carry text content and bounding box information in the pixel coordinate system of the document page; In this step, the processor obtains the image resolution parameters of the document to be parsed, including the image width. With image height Extract the original bounding box information of text primitives in the page pixel coordinate system; to eliminate the influence of different document resolutions on geometric calculations, the processor performs coordinate normalization: normalizing the original coordinates of all text primitives... Mapped to a range of values relative coordinates between The calculation formula is as follows:

[0016]

[0017] Based on the normalized relative coordinates, the processor calculates the Euclidean distance and projection overlap rate between any two text primitives. Based on the above calculation results, the processor constructs a heterogeneous graph data structure. A heterogeneous graph refers to a graph model containing multiple edge attributes, where text primitives are the nodes of the graph, and the weights of the connecting edges between nodes are defined by the normalized distance and overlap rate. Specifically, the processor uses a weighted fusion formula to calculate the edge weights. The calculation formula is:

[0018] in For nodes With nodes The normalized Euclidean distance between them and The preset scaling factor is used, and it satisfies... Normalized Euclidean distance The calculation method is as follows: compute nodes With nodes raw pixel distance between center points and divide it by the current page diagonal pixel length. The calculation formula is as follows:

[0019] Used to balance the contributions of geometric distance and projection alignment in edge weight calculation, such as This formula fuses the attenuation properties of geometric distance and the alignment properties of projective overlap into a single scalar weight; this step completes the mapping from unstructured images to scale-invariant structured graph data, i.e., establishes a visual adjacency graph; assuming the evolution steps are: Based on the topological features of heterogeneous graph data structures, such as node clustering and orientation, the processor generates multiple sets of structured mapping hypotheses. The specific mapping logic is as follows: the processor uses the maximum spanning tree algorithm to extract the graph's backbone skeleton and assigns edge weights... Greater than the preset connectivity threshold The nodes are identified as logically related parent-child candidate pairs; then the hierarchical depth is determined based on the vertical projection inclusion relationship of the relative coordinates of the nodes, and the related nodes are combined into table rows or paragraph blocks based on the arrangement direction features, thereby generating a variety of tentative tree-like logical structures. Structured mapping assumptions refer to a tentative inference about the logic of a document layout. Each set of assumptions explicitly defines the parent-child hierarchy between nodes, such as the relationship between table titles and table cells, and data attribute labels, such as amount, date, and total. This process is equivalent to the builder in the system proposing multiple possible interpretations based on visual cues, such as whether this is a three-column text or a table containing merged cells. Logical verification steps: The processor performs semantic injection and logical calculation; the system injects the semantic content of the text primitives, i.e. the text symbols recognized by OCR, into the current structured mapping hypothesis to form a temporary semantic stream; The processor calculates the total logical cost, which represents the degree of logical conflict, using a weighted summation method. This value is a quantitative indicator used to measure the degree to which the parsing results under the current assumptions deviate from the preset rules in terms of arithmetic logic, semantic fluency, and format standardization. Iterative optimization steps: The processor calculates the total logical cost value. With preset convergence threshold The comparison is performed, and branch logic is executed based on the comparison result: Before the first execution of the logic verification, the processor automatically initializes the current iteration number. Subsequently, each time the hypothesis evolution step is repeated, the number of iterations is [number missing]. Automatically increment by 1 until the preset maximum value is reached. until; Branch 1, continue optimization: If And the current iteration number The preset maximum value has not been reached. If the result is negative, it indicates that there is a significant logical contradiction in the current hypothesis, such as uneven table summaries or semantic breaks in sentences. At this point, the processor performs a topological reconstruction operation on the heterogeneous graph data structure by splitting or aggregating nodes, based on the characteristics of the local regions that generate high-cost values, i.e., the regions where logical residuals are concentrated. The system returns to the hypothesis evolution step and generates a new hypothesis based on the reconstructed graph. This process constitutes a feedback loop of perception and cognition. Branch 2, Output: If or If the current parsing result meets the logical consistency requirement or has reached the upper limit of computing resources, the processor locks the structured mapping hypothesis with the minimum total logical cost as the final model and outputs the structured parsing data. This embodiment breaks through the limitation of unidirectional transmission in traditional OCR post-processing by introducing a closed-loop mechanism of logical verification and iterative optimization. Its core advantage lies in its ability to use strong logical constraints within the document, such as accounting identities and semantic coherence, to reverse-correct visual recognition errors, such as numerical segmentation errors caused by printing blur. This white-box self-verifying parsing method significantly improves parsing accuracy in complex scenarios with weak visual cues and strong semantic dependencies, such as in borderless tables.

[0020] Example 2: When calculating the projection overlap rate, the processor executes the following logic: Obtain the projection intervals of two text primitives on the horizontal coordinate axis; Calculate the intersection length and union length of two projection intervals; The ratio obtained by dividing the intersection length by the union length is determined as the projection overlap rate. Only when the projection overlap rate is greater than zero, the corresponding connection edge is established in the heterogeneous graph data structure.

[0021] This embodiment provides specific limitations on the edge weight calculation logic in the graph construction step; The processor calculates the projection overlap ratio. At that time, the following precise logic is executed: Obtain the projection interval: For any two text primitives to be analyzed and The processor obtains their projection intervals on the horizontal coordinate axis, i.e., the X-axis; let... The projection range is , The projection range is ; The processor calculates the intersection length of the two projection intervals. Union length The calculation formula is as follows:

[0022]

[0023] The ratio obtained by dividing the intersection length by the union length is determined as the projection overlap rate, and its calculation formula is as follows:

[0024] The range of values ​​for this number is: This is used to quantify the alignment of two graphic elements in vertical layout. Establishing connecting edges: The processor executes a sparsity strategy only when... Only then are the corresponding connection edges established in the heterogeneous graph data structure; By using threshold filtering based on horizontal projection overlap rate, the system can effectively eliminate node connections that cannot logically belong to a column. This not only significantly reduces the edge density of heterogeneous graphs and reduces the computational complexity of subsequent graph algorithms, but also provides strong geometric feature support for the identification of table column structures, avoiding cross-column erroneous connections. Example 3: When calculating the total logical cost, the processor performs the following steps: Calculate the arithmetic consistency cost component, which represents the validity of numerical calculation logic; the semantic coherence cost component, which represents the fluency of natural language text; and the data format compliance cost component, which represents whether the field content conforms to the predefined regular expression format. Obtain the preset weight coefficient table, which defines the normalized weights of the above three components for different document types. The total logical cost is obtained by multiplying each component by its corresponding normalized weight and summing the results.

[0025] This embodiment provides specific limitations on the construction of the cost function in the logic verification step; The processor calculates the total logical cost value. At that time, a multi-dimensional constraint fusion strategy was adopted, and the specific steps are as follows: Component cost calculation: The processor calculates the cost components for the following three dimensions respectively: Arithmetic consistency cost component : To measure whether the calculation logic of numerical data is valid; Semantic coherence cost components : Measures the fluency of natural language text; Data format compliance cost components This measures whether the content of a field conforms to a predefined regular expression format, such as a date format. The quantitative calculation formula is as follows:

[0026] in The length of the text content that conforms to the regular expression constraints. This is the total length of the text in that node; this formula ensures that the formatting validation results are smoothly mapped to... The interval, when the content completely fails to meet the formatting requirements, A value of 1 results in the maximum logical penalty. Obtain weight coefficients: The processor retrieves a preset weight coefficient table; this table defines the normalized weights of the three components for different document types, such as financial statements, contracts, and technical drawings. ,satisfy ; For example, regarding financial statements, The standard setting is usually higher; however, for legal contracts... The setting is relatively high; Weighted summation: The total logical cost is obtained by multiplying each component by its corresponding normalized weight and then summing the results. The calculation formula is as follows:

[0027] By introducing a dynamic weighting mechanism for document types, the system can adaptively handle parsing tasks under different business scenarios. This multimodal constraint fusion method solves the problem that a single model cannot take into account both visual features and business logic, enabling the system to calculate financial data like an accountant and read contract texts like a legal professional. Example 4: When calculating the arithmetic consistency cost component, the processor executes the following logic: Identify the first set of nodes with sub-item attribute labels and the second set of nodes with total attribute labels in the structured mapping hypothesis; Extract the text values ​​of all nodes in the first node set and calculate the cumulative sum; Calculate the absolute value of the difference between the accumulated sum and the text value of the second node; The absolute value of the difference is mapped to a value between 0 and 1, which is used as the arithmetic consistency cost component.

[0028] This embodiment specifically illustrates the arithmetic consistency cost component. The computational logic is the core method of this invention for error correction using strong logical priors. The processor executes the following logic: Identify the set of nodes: Under the current structured mapping assumption, the processor identifies the first set of nodes with sub-item attribute labels. and the second node with the total attribute label This typically corresponds to the detail row and total row in a table. Calculate the cumulative sum: The processor extracts the text values ​​of all nodes in the first node set. And calculate its cumulative sum. The calculation formula is as follows:

[0029] in, This represents the total number of nodes in the first node set; For the first The text values ​​of each child node; during this process, the processor automatically filters out non-numeric characters; Calculate the absolute value of the difference: Calculate the cumulative sum The text value of the second node The absolute value of the difference between The calculation formula is as follows:

[0030] Mapping cost component: The processor uses an adaptively scaled variant of the Sigmoid function to map the absolute value of the difference. Mapped to The dimensionless values ​​between them serve as the arithmetic consistency cost component. ; Specifically, to address the difference in gradient sensitivity caused by different numerical magnitudes, such as amounts and tax rates, the processor calculates an adaptive scaling factor based on the numerical characteristics of the current node set. The calculation formula is as follows:

[0031] in, This is a preset base slope constant, for example, a value of 10.0. The modulus of the summation; this factor is used to eliminate the influence of physical dimensions in the data, and is substituted into the following formula to calculate the cost component:

[0032] in, The preset allowable deviation threshold is used; by introducing This ensures the exponential term It is always a dimensionless pure number, thus ensuring that the model has stable error correction sensitivity when processing financial data of any magnitude; This technical solution transforms the accounting identity in business logic into a differentiable loss function. When the OCR mistakenly identifies 8 as 3, causing an imbalance in the accounts, this cost component increases dramatically, driving the system to reject the erroneous assumption in subsequent iterations and attempt to find the correct recognition result that makes the equation true. This is a logic-level error correction that conventional NLP models cannot achieve. Example 5: When computing the semantic coherence cost components, the processor executes the following logic: Extract text sequences marked as multi-line concatenation in the structured mapping hypothesis; Input the text sequence into a pre-trained language probability model to obtain the probability value of the text sequence as natural language generation; The reciprocal of the probability value is used as an index of perplexity. The perplexity index is used as a component of semantic coherence cost, where a higher perplexity index indicates a higher degree of semantic logical conflict.

[0033] This embodiment specifically illustrates the semantic coherence cost component. The computational logic is mainly used to resolve ambiguity issues in cross-line text concatenation; The processor executes the following logic: Extracting text sequences: The processor extracts text sequences marked as multi-line concatenations in the structured mapping hypothesis. This usually happens in cell wrappers or paragraphs spanning multiple pages. Obtain the generation probability: (This refers to the text sequence) Input a pre-trained language probability model, such as a lightweight model based on BERT or GPT; the model outputs a probability value of the text sequence as generated natural language. Probability values The probability scores are normalized using the Softmax function, and a minimal bias constant is added to the denominator. ,For example make sure The range of values ​​is within This avoids mathematical singularities where the denominator is zero or the logarithm is undefined when calculating the inverse of the perplexity and the logarithmic mapping. Calculating perplexity: The reciprocal of the probability value is used as the perplexity index. :

[0034] Or take the logarithmic form: ; Determine the cost components: due to the perplexity index The range of values ​​is usually 100. Directly participating in weighted calculations can lead to an imbalance in magnitude; therefore, the processor performs a nonlinear normalization mapping, which... Convert to Semantic coherence cost components within the interval The mapping function is in logarithmic form of the hyperbolic tangent function, and its calculation formula is as follows:

[0035] in, Indicates The natural logarithm with base 0. The perplexity value output by the language model. The hyperparameter for adjusting sensitivity has a range of values. The settings are positively correlated with the complexity of the document layout; the more complex the layout, the more... The larger the value, the more preferably 0.5 in this embodiment; When the text is extremely fluent ( )hour, ,but As the level of confusion increases, that is... Increase It gradually approaches 1; this ensures that the semantic cost component and the arithmetic cost component are of the same order of magnitude, thus guaranteeing the subsequent total logical cost value. The mathematical rigor of calculation; This feature leverages the sensitivity of NLP models to text fluency to effectively solve the problem of determining reading order in complex layouts. When visual dividing lines are missing, the system can decide whether to merge two lines of text based on whether the sentence is fluent, thus relying on semantic features to complete the correct layout restoration when visual features fail. Example 6: When the processor performs a topology reconfiguration operation, it specifically executes the following steps: Node aggregation: In response to feedback signals that have low semantic coherence cost but high arithmetic consistency cost, identify multiple independent nodes that are spatially adjacent and semantically continuous, merge them into an aggregate node, and update the boundary coordinates of the aggregate node. Node splitting: In response to feedback signals that data format compliance costs are high, a single node containing multiple data format characteristics is identified and split into multiple independent child nodes based on the position of the whitespace delimiter; Reconstruction Update: Based on the node states after aggregation or splitting, recalculate the edge weights of the involved nodes and update the adjacency matrix of the heterogeneous graph data structure.

[0036] This embodiment describes in detail the topology reconstruction operation performed by the system during the iterative optimization process, which is a concrete manifestation of the system's dynamic correction capability; When the processor performs a topology reconfiguration operation, it specifically executes the following steps: When making refactoring decisions, the processor follows the principle that format logic takes precedence over semantic logic: Node aggregation: In response to feedback signals that have low semantic coherence cost but high arithmetic consistency cost, the processor identifies multiple independent nodes that are spatially adjacent and semantically contiguous. Feedback signal logic: This signal indicates that the current independent node is semantically fluent, but may be fragmented in numerical logic. For example, a long value may be incorrectly split into two parts, resulting in an error in the total calculation. The processor merges these nodes into a single aggregate node and updates the boundary coordinates of the aggregate node, which is the smallest bounding rectangle of all child node coordinates; this operation aims to correct excessive visual segmentation. Node Splitting: In response to feedback signals indicating high data format compliance costs, the processor identifies single nodes containing multiple data format characteristics. During reconstruction, the processor follows a conflict resolution principle prioritizing format over semantics: if both aggregation and split signals are triggered simultaneously in the same local region, the system prioritizes executing the reconstructive method based on format. The node splitting operation ensures the accuracy of atomic-level data; specifically, when the data format compliance cost component of this region... At that time, regardless of the cost of semantic coherence If the value is too low, the processor will force the node splitting process to prevent critical fields, such as dates and amounts, from being concatenated due to over-aggregation. In the splitting calculation, the minimum splitting interval threshold is set to a percentage of the current page's average character width. This is used to filter out mis-segmentation caused by visual noise; Feedback signal logic: This signal indicates that a node contains attributes that should be separate, such as date and amount being combined, causing date format validation to fail; The processor splits the text into multiple independent child nodes based on the position of whitespace separators or abrupt changes in character spacing. The logic for determining abrupt changes in character spacing is as follows: calculate the set of horizontal spacings between adjacent characters within the current text block; if a certain spacing value... A mutation is determined if the following criteria are met:

[0037] in The average spacing between characters in this line. The average global character width of the page; the average global character width of the page The acquisition logic is as follows: the processor traverses all extracted text elements on the current page and calculates the arithmetic mean of the horizontal width of all text elements; this operation aims to correct visual undersegmentation. Reconstruction Update: Based on the node states after aggregation or splitting, the processor recalculates the edge weights of the involved nodes and updates the adjacency matrix of the heterogeneous graph data structure; this ensures that subsequent hypothesis evolution is based on the corrected graph topology. This embodiment solves the problem of the immutability of static OCR results through a specific reconstruction strategy; the system can dynamically cut and merge connected characters or merge broken characters according to logical feedback, which essentially gives the system a human-like cognitive ability to adjust visual focus in order to understand, and significantly improves its adaptability to non-standard layouts. Example 7: Assuming that during the evolutionary process, the processor specifically executes: A probabilistic statistical model is used to perform subgraph isomorphism search on heterogeneous graph data structures; Calculate the probability value of matching the identified sub-graph structure with the preset table template, column text template or paragraph template; A preset number of candidate tree structures with the highest matching probability values ​​are selected as the structured mapping hypothesis.

[0038] This embodiment further elaborates on the specific implementation of the hypothesis evolution steps; The processor specifically executes the following during the hypothetical evolution step: Subgraph isomorphism search: Using probabilistic statistical models, such as Conditional Random Fields (CRF) or Graph Neural Networks (GNN), subgraph isomorphism search is performed on heterogeneous graph data structures; this process aims to find substructures that conform to specific topological rules from a chaotic network of nodes. Matching probability calculation: The processor calculates the matching probability value of the recognized sub-graph structure that matches the preset table template, column text template, or paragraph template. ; Candidate tree selection: Based on the calculation results, the top-K candidate tree structures with the highest matching probability values ​​are selected as the structured mapping hypothesis; here... The value range is set to This range was determined based on experiments balancing computational resources and search space. The system determines the range based on the total number of nodes in the document to be parsed. The dynamic adjustment is calculated using the following formula:

[0039] This adaptive mechanism ensures that sufficient hypothesis space is preserved when processing complex documents to avoid local optima, while optimizing computational efficiency in processing simple documents. By introducing template-based subgraph search, the system can quickly map the underlying graph data to the higher-level layout semantic structure, such as tables and paragraphs. The Top-K selection strategy ensures that the system will not get stuck in local optima, and retains multiple possible structural assumptions for subsequent logical verification modules to filter, thus balancing search efficiency and parsing robustness. Example 8: The processor is also configured to execute the following when outputting structured parsed data: Generate confidence metadata containing logical decision paths; Convert structured parsed data into JSON or XML format files; Write the file to physical storage media or send it to downstream business systems via a network interface; If a write failure or network timeout is detected during the output process, a retry operation is performed and an error log is logged.

[0040] This embodiment focuses on data integrity and engineered interaction during the system output phase; When the processor outputs structured parsed data, it is configured to execute: Generate confidence level metadata: In addition to outputting the parsing results, the system also generates confidence level metadata containing the logical decision path; for example: Field A and Field B are merged, with a confidence level of 0.95, based on: conforming to the accounting equation; Format conversion: Convert structured parsed data into common JSON or XML format files to facilitate downstream system integration; Persistence and Transmission: The generated files are written to physical storage media, such as a hard disk database, or sent to downstream business systems, such as an ERP system, via a network interface; Exception handling closed loop: If a write failure or network timeout is detected during the output process, the processor automatically performs a retry operation and records the error log to ensure the reliability of data delivery; This embodiment realizes the last mile from algorithm model to engineering application; in particular, the output of the logical decision path transforms the traditional black-box AI into an interpretable white-box system, which has decisive application value for fields such as financial auditing and legal compliance that are sensitive to data sources.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-algorithm collaborative intelligent document parsing system, characterized in that, It includes a memory and a processor. The memory stores the computer program, and the processor executes the computer program to perform the following steps: The graph construction steps are as follows: obtain the layout data of the document to be parsed, extract the text primitives and their spatial coordinates in the pixel coordinate system, calculate the Euclidean distance and projection overlap rate between any two text primitives, and construct a heterogeneous graph data structure with text primitives as nodes and distance and overlap rate as edge weights. Hypothesis evolution steps: Based on the topological features of the heterogeneous graph data structure, multiple sets of structured mapping hypotheses are generated. Each set of hypotheses defines the parent-child hierarchical relationship between nodes and data attribute labels. Logical verification steps: Inject the semantic content of the text primitives into the structured mapping hypothesis, and calculate the total logical cost value representing the degree of logical conflict by weighted summation; Iterative optimization steps: Compare the total logic cost with the preset convergence threshold: If the total logical cost is greater than the convergence threshold and the current iteration number has not reached the preset maximum value, then based on the characteristics of the local region that generates the high cost value, perform a topology reconstruction operation of node splitting or node aggregation on the heterogeneous graph data structure, and return to the execution hypothesis evolution step. If the total logical cost is less than or equal to the convergence threshold, or if the current iteration count reaches the preset maximum value, then the structured mapping hypothesis with the minimum total logical cost is locked as the final model, and structured parsed data is output. When calculating the projection overlap rate, the processor executes the following logic: Obtain the projection intervals of two text primitives on the horizontal coordinate axis; Calculate the intersection length and union length of two projection intervals; The ratio obtained by dividing the intersection length by the union length is determined as the projection overlap rate. Only when the projection overlap rate is greater than zero, the corresponding connection edge is established in the heterogeneous graph data structure.

2. The document intelligent parsing system with multi-algorithm collaboration according to claim 1, characterized in that, When calculating the total logical cost, the processor performs the following steps: Calculate the arithmetic consistency cost component, which represents the validity of numerical calculation logic; the semantic coherence cost component, which represents the fluency of natural language text; and the data format compliance cost component, which represents whether the field content conforms to the predefined regular expression format. Obtain the preset weight coefficient table, which defines the normalized weights of the above three components for different document types. The total logical cost is obtained by multiplying each component by its corresponding normalized weight and summing the results.

3. The document intelligent parsing system with multi-algorithm collaboration according to claim 2, characterized in that, When calculating the arithmetic consistency cost component, the processor executes the following logic: Identify the first set of nodes with sub-item attribute labels and the second set of nodes with total attribute labels in the structured mapping hypothesis; Extract the text values ​​of all nodes in the first node set and calculate the cumulative sum; Calculate the absolute value of the difference between the accumulated sum and the text value of the second node; The absolute value of the difference is mapped to a value between 0 and 1, which is used as the arithmetic consistency cost component.

4. The document intelligent parsing system with multi-algorithm collaboration according to claim 2, characterized in that, When computing the semantic coherence cost components, the processor executes the following logic: Extract text sequences marked as multi-line concatenation in the structured mapping hypothesis; Input the text sequence into a pre-trained language probability model to obtain the probability value of the text sequence as natural language generation; The reciprocal of the probability value is used as an index of perplexity. The perplexity index is used as a component of semantic coherence cost, where a higher perplexity index indicates a higher degree of semantic logical conflict.

5. The document intelligent parsing system with multi-algorithm collaboration according to claim 1, characterized in that, When the processor performs a topology reconfiguration operation, it specifically executes the following steps: Node aggregation: In response to feedback signals that have low semantic coherence cost but high arithmetic consistency cost, identify multiple independent nodes that are spatially adjacent and semantically continuous, merge them into an aggregate node, and update the boundary coordinates of the aggregate node. Node splitting: In response to feedback signals that data format compliance costs are high, a single node containing multiple data format characteristics is identified and split into multiple independent child nodes based on the position of the whitespace delimiter; Reconstruction Update: Based on the node states after aggregation or splitting, recalculate the edge weights of the involved nodes and update the adjacency matrix of the heterogeneous graph data structure.

6. The document intelligent parsing system with multi-algorithm collaboration according to claim 1, characterized in that, Assuming that during the evolutionary process, the processor specifically executes: A probabilistic statistical model is used to perform subgraph isomorphism search on heterogeneous graph data structures; Calculate the probability value of matching the identified sub-graph structure with the preset table template, column text template or paragraph template; A preset number of candidate tree structures with the highest matching probability values ​​are selected as the structured mapping hypothesis.

7. The document intelligent parsing system with multi-algorithm collaboration according to claim 1, characterized in that, The processor is also configured to execute the following when outputting structured parsed data: Generate confidence metadata containing logical decision paths; Convert structured parsed data into JSON or XML format files; Write the file to physical storage media or send it to downstream business systems via a network interface; If a write failure or network timeout is detected during the output process, a retry operation is performed and an error log is logged.

Images