Intelligent Invoice Information Extraction and Analysis Method and System Integrating Deep Learning

Abstract

This invention provides an intelligent method and system for extracting and analyzing invoice information using deep learning, relating to the field of deep learning technology. First, based on the field flow and association logic of invoice business scenarios, it generates semantic traceability tracing trajectories for fields, collects scenario-related features of the invoice carrier to generate scenario feature guidance signals, and uses these signals to locate the physical presentation areas of key fields and extract their original information to form a set. Next, it constructs a dynamic field association model, inputting the set of original information for key fields and scenario-related features into the model, and uses deep learning algorithms to model dynamic dependency relationships to generate a dynamic field association graph. Finally, based on this graph, it integrates information to generate scenario-based structured invoice analysis results, including core field content, scenario association logic, and dynamic association descriptions, enabling efficient and accurate processing of invoice information.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for intelligent invoice information extraction and analysis that integrates deep learning. Background Technology

[0002] In the field of bill processing, with the continuous growth of business volume and the increasing complexity of business scenarios, higher demands are placed on the accurate extraction and in-depth analysis of bill information. Traditional bill information processing methods mainly rely on manual identification and simple rule matching. In terms of manual identification, staff need to spend a lot of time and energy to check and record various types of information on the bill one by one, which is not only inefficient but also prone to human error, especially when processing a large number of bills, the error rate will increase significantly.

[0003] Rule-based matching methods typically pre-define fixed rules to extract invoice information, such as identifying key fields based on specific formats or locations. However, invoice formats vary widely, and the format and field layout differ significantly across different business scenarios. Furthermore, invoice formats may change frequently as business evolves. This makes it difficult for pre-defined rules to adapt to various complex invoice situations, accurately extract key field information, or deeply analyze the dynamic relationships between fields and their correlation with business scenarios. Consequently, these methods fail to meet the demands of comprehensive, accurate, and efficient processing of invoice information in real-world business operations. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an intelligent invoice information extraction and analysis method integrating deep learning, the method comprising:

[0005] Based on the field flow association logic of the bill business scenario, the semantic generation source and propagation path of key fields are traced to generate the semantic source tracing trajectory of fields. The key fields include bill code, bill number, amount information, tax information, and buyer and seller information.

[0006] Collect scene association features of the invoice carrier, and generate scene feature guidance signals based on the semantic tracing trajectory of the fields. The scene association features are the layout association information, field function association information and business process association information of the invoice carrier in the business scenario.

[0007] Based on scene feature guidance signals, the physical presentation area of ​​key fields in the ticket carrier is targeted and locked, and the original information of key fields in the physical presentation area of ​​key fields is extracted. The original information includes text representation information and field form association information, forming a set of original information of key fields.

[0008] Construct a field dynamic association model by inputting the original information set of key fields and scene association features into the field dynamic association model, and using deep learning algorithms to model the dynamic dependency relationships between key fields and between key fields and scenes, thereby generating a field dynamic association graph;

[0009] Based on the field dynamic association graph, the complete original information of key fields, scenario association information and dynamic dependency relationships are integrated to generate scenario-based structured invoice analysis results. The scenario-based structured invoice analysis results include the core content of the fields, scenario association logic and dynamic association description.

[0010] In another aspect, embodiments of the present invention also provide an intelligent invoice information extraction and analysis system integrating deep learning, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this invention generates a semantic tracing trajectory of fields by establishing a field flow association logic based on the bill business scenario. This allows for tracing the semantic generation source and propagation path of key fields, effectively avoiding information extraction errors caused by semantic misunderstanding. Collecting scenario association features of the bill carrier and generating scenario feature guidance signals can pinpoint the physical presentation area of ​​key fields, accurately extracting native information and improving the accuracy and relevance of information extraction. Constructing a dynamic field association model utilizes deep learning algorithms to model the dynamic dependencies between key fields and between key fields and scenarios, generating a dynamic field association graph. Finally, based on this dynamic field association graph, a scenario-based structured bill analysis result is generated, which not only includes the core content of the fields but also covers the scenario association logic and dynamic association description, greatly improving the efficiency and quality of bill information processing. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the intelligent invoice information extraction and analysis method integrating deep learning provided in the embodiments of the present invention.

[0013] Figure 2 This is a simplified logical diagram of the intelligent invoice information extraction and analysis method integrating deep learning provided in this embodiment of the invention.

[0014] Figure 3 This is a schematic diagram of the hardware architecture of the intelligent invoice information extraction and analysis system that integrates deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings. See also: Figures 1-2The following section provides a detailed introduction to this intelligent invoice information extraction and analysis method that integrates deep learning.

[0016] Step S110: Based on the field flow association logic of the bill business scenario, trace the semantic generation source and propagation path of key fields to generate the semantic source tracing trajectory of fields. Key fields include bill code, bill number, amount information, tax information, and buyer and seller information.

[0017] In invoice processing, different types of invoices involve the flow of multiple key fields in their respective business scenarios. For example, in the business scenario of value-added tax (VAT) invoices, the aforementioned key fields need to flow through multiple business stages such as issuance, transmission, authentication, and deduction. In order to accurately generate semantic traceability of fields, it is first necessary to clarify the flow of each key field in the entire business process.

[0018] Step S111: Collect the field flow records of the entire bill business process to form a field flow record set. The field flow records cover the complete flow path of key fields from generation to archiving in different business scenarios, including the appearance form of the fields and related fields in each stage of business application, review, confirmation and archiving.

[0019] In the context of VAT invoice transactions, each step—from the invoice application to the tax authority's review, the company's confirmation of receipt, and finally the archiving in the tax system—generates field flow records. For example, in the application stage, the company submits an invoice application, which includes the initial form of key fields such as buyer and seller information and amount information. In the review stage, the tax authority verifies these fields and may generate related fields such as review comments. In the confirmation stage, both the buyer and seller confirm the invoice information, at which point the form of key fields may become more standardized. In the archiving stage, the invoice information is stored in the database according to a specific format, and key fields exist in a specific data form. Collecting the records generated at each of these stages constitutes a set of field flow records.

[0020] Step S112: Perform association logic analysis on the field flow record set, and use the association rule mining algorithm to extract the associated fields and flow conditions of each key field in different business nodes; based on the timestamp of the field appearance and the business node sequence, determine the generation trigger event, propagation data format and receiving system identifier of the key field in the business process, and form a field flow association logic diagram.

[0021] Step S1121: Based on the business scenario identifier in the field flow record, divide the field flow record set into multiple field flow record subsets corresponding to different business scenarios.

[0022] In the context of VAT invoice transactions, field transfer records contain specific business scenario identifiers, such as "VAT invoice issuance" and "VAT invoice authentication." These identifiers allow all collected field transfer records to be categorized into corresponding business scenario subsets for subsequent analysis tailored to different scenarios.

[0023] Step S1122: Parse each field flow record subset one by one, extract the key field names, business nodes, related field names and flow triggering conditions in each field flow record, and form a single record parsing result.

[0024] Taking the "Issuance of Value-Added Tax Special Invoices" business scenario subset as an example, each record within it is analyzed. For instance, a record might be an invoice application submitted by a company. Analyzing this record reveals key field names such as "Buyer Name," "Seller Name," and "Amount." The business node is "Invoice Application," and associated field names might include "Application Number" and "Application Date." The triggering condition for the flow might be "The application information is complete and conforms to the invoice issuance specifications."

[0025] Step S1123: Integrate the parsing results of all individual records and construct a field node association matrix. The rows of the field node association matrix represent key fields, the columns represent business nodes, and the matrix elements represent the associated fields and flow conditions of the key field in the corresponding business node.

[0026] The parsing results of individual records obtained above are then integrated. For example, rows can be key fields such as "Buyer Name," "Seller Name," and "Amount," while columns can be business nodes such as "Invoice Application," "Tax Review," "Company Confirmation," and "System Archive." Matrix elements can be related fields such as "Application Number" and the flow trigger condition "Application information is complete and complies with invoicing specifications" for "Buyer Name" under the "Invoice Application" node.

[0027] Step S1124: Based on the field node association matrix, use the association rule mining algorithm to extract the generation trigger events of each key field in each business node. The generation trigger events are the business operations or information input behaviors that cause the key field to appear for the first time.

[0028] Association rule mining algorithms are used to analyze the association matrix of field nodes. In the "Invoice Application" business node, for the key field "Buyer Name", by analyzing the association relationship between this field and other fields and flow conditions in the matrix, it can be determined that its generation trigger event is the operation of the enterprise entering the buyer name information on the invoice application interface.

[0029] Step S1125: Analyze the data format of each key field during the circulation process, determine the encoding method and structure definition of the data format, including the image format of paper carriers, the message format of electronic transmission, and the database record format of system storage.

[0030] For example, "amount information" may exist in the form of handwritten or printed text on a paper application form during the business application process, i.e., the image format of the paper carrier; during electronic transmission, it may be converted into a specific message format, such as XML format, which defines structural information such as amount tags and data types; in the system storage stage, it will exist in the database record format. For example, in a MySQL database, amount information may be stored as a DECIMAL type field with specific precision and length definitions.

[0031] Step S1126: Determine the receiving system identifier for each key field at each business node, forming a list of receiving system identifiers. The receiving system identifier is used to uniquely identify the business processing module, the review node, and the archiving system.

[0032] At the "Tax Audit" stage, key fields such as "Amount Information" are transmitted to the tax authority's audit system, which has a unique identifier, such as "TaxAuditSys_001". In the "System Archiving" stage, key fields are transmitted to the tax archive management system, whose receiving system identifier might be "TaxArchiveSys_002". Arranging these identifiers in the order of the business stages creates a list of receiving system identifiers.

[0033] Step S1127: Construct the basic framework of key field flow association logic. The basic framework of key field flow association logic takes key fields as the core, arranges each business node according to the business process sequence, and marks the associated fields and flow conditions.

[0034] Centered on the key field of "amount information," the system arranges business nodes such as "invoice application," "tax review," "company confirmation," and "system archiving" in the order of the VAT special invoice business process. Each node is labeled with its associated fields, such as "application number" for the "invoice application" node, as well as the processing conditions, such as "complete application information."

[0035] Step S1128: Add the list of generated trigger events, propagation data formats, and receiving system identifiers to the basic framework of key field flow association logic, and improve the field flow details of each business node.

[0036] In the "Invoice Application" node, the generation trigger event for "Amount Information" is supplemented with "Enterprise Input Amount Data", the data format is "Paper Image Format", and the system identifier "Invoice Application Processing Module Identifier" is received. In the "Tax Audit" node, the data format is supplemented with "Electronic Message Format", and the system identifier "Tax Audit System Identifier" is received, etc., to make the field flow details of each business node more complete.

[0037] Step S1129: Compare the supplemented key field flow association logic basic framework with the preset standard business process rules to verify whether the flow order between each business node is consistent with the standard business process rules, and verify whether the associated fields and conditions are consistent with the rules defined in the standard business process rules.

[0038] The pre-defined standard business process rules stipulate the correct workflow for VAT invoice transactions, such as requiring an invoice application first, followed by tax review, then company confirmation, and finally system archiving. The workflow node order in the supplementary basic framework is compared with this standard to ensure consistency. Simultaneously, the associated fields and conditions of each node are verified to conform to the standard rules, such as whether the associated fields in the review process include the necessary review items and whether the workflow conditions are reasonable.

[0039] Step S11210: Input the improved key field flow association logic basic framework into the graph generation algorithm to generate a field flow association logic graph containing details of the entire key field flow process.

[0040] The basic framework, refined through the above steps, contains detailed flow information of key fields across various business nodes. This information is then input into a graph generation algorithm, which constructs a graphical diagram of field flow relationships. Nodes represent business nodes and key fields, edges represent flow relationships, and the edges are labeled with information such as associated fields, flow conditions, and propagation data formats.

[0041] Step S113: Input the field flow association logic diagram into the deep learning semantic tracing module. This deep learning semantic tracing module is based on the semantic rules of bill business, learns the evolution rules of key field semantics in the flow process, and establishes a semantic evolution model.

[0042] The semantic rules for invoice transactions define the semantic meaning and evolution rules of key fields in different business scenarios. For example, in the VAT invoice process, "amount information" may only represent an estimated amount during the invoice application stage, with relatively vague semantics. During the review stage, after verification by the tax authorities, its semantics become more precise, representing the deductible amount that complies with tax regulations. During the confirmation stage, after confirmation by both the buyer and seller, its semantics are further clarified as the transaction amount agreed upon by both parties. The deep learning semantic tracing module learns these semantic rules and the flow of key fields in the field flow association logic diagram to grasp the evolutionary patterns of semantics, thereby establishing a semantic evolution model.

[0043] Step S114: Based on the semantic evolution model, locate the initial semantic generation node of each key field, extract the business scenario information and initial representation form of the field from the initial semantic generation node. The initial semantic generation node is the business link in which the key field first has business meaning.

[0044] The established semantic evolution model is used to analyze key fields. Taking "tax amount information" as an example, in the VAT special invoice business, tax amount information is calculated based on the amount and tax rate. During the invoice application stage, the company may estimate a tax amount, but this information may not yet be confirmed by the tax authorities and does not have formal business meaning. However, during the tax review stage, the tax amount calculated by the tax system based on the confirmed amount and applicable tax rate is the first time it has formal business meaning. Therefore, the "tax review" stage is the initial semantic generation node for "tax amount information." Extracting the business scenario information of this node, such as "tax review scenario, VAT special invoice tax amount calculation," the initial representation of the field might be "1234.56 yuan (tax rate 13%)".

[0045] Step S115: Trace the propagation path of the key field from the initial semantic generation node to the final archiving node, record the supplementary, adjustment or enhancement behavior of each business node in the propagation path for the semantics of the key field, and form a detailed record of semantic propagation.

[0046] Starting from the initial semantic generation node "Tax Audit" for "Tax Information" and ending at the final "System Archive" node, the process may involve business nodes such as "Enterprise Confirmation." At the "Enterprise Confirmation" node, the enterprise may verify the tax information. If it is found to be correct, this is a semantic reinforcement action; if problems are found and feedback is provided, the tax authorities make adjustments, and the enterprise confirms again—this is a semantic adjustment action. Recording in detail the semantic supplementation, adjustment, or reinforcement actions that occur at each business node creates a detailed semantic propagation record.

[0047] Step S116: Integrate the initial semantic generation node information, propagation path and semantic propagation details of key fields to build a basic framework for semantic tracing.

[0048] The initial semantic generation node information of "tax amount information", such as node name, business scenario information, initial field representation form, propagation path, such as "tax audit → enterprise confirmation → system archiving", and semantic propagation details, such as the verification behavior when the enterprise confirms and the situation of strengthening semantics, are integrated together to build a basic framework to describe the source of the semantics of key fields.

[0049] Step S117: Set the semantic association weight of the fields. Based on the degree of influence of different business nodes on the semantics of key fields, assign corresponding association weights to each propagation link to strengthen the semantic influence representation of the core propagation link.

[0050] In VAT invoice transactions, the "tax audit" stage has the greatest impact on the semantics of "tax amount information" because it determines the official tax amount. The "enterprise confirmation" stage is next, serving to verify and reinforce the information. The "system archiving" stage primarily stores information and has a relatively small impact on semantics. Based on these levels of impact, a higher association weight (e.g., 0.6) is assigned to the "tax audit → enterprise confirmation" propagation stage, while a lower association weight (e.g., 0.3) is assigned to the "enterprise confirmation → system archiving" propagation stage.

[0051] Step S118: Generate the initial semantic tracing trajectory of key fields by integrating the semantic tracing framework with the association weights.

[0052] The semantic tracing framework combines each propagation stage with its corresponding association weight. For example, in the semantic tracing framework for "tax information," each stage in the propagation path carries an association weight, forming the initial semantic tracing trajectory described above: "Tax audit (weight 0.6) → Enterprise confirmation (weight 0.3) → System archiving (weight 0.1)." The total weight is 1 to reflect the comprehensive impact of each stage on the semantics.

[0053] Step S119: Collect semantic traceability cases of fields in historical ticket processing, compare the semantic traceability trajectories in the cases with the preset business rule base, filter out the trajectory cases that meet the business rules, and form a traceability optimization dataset.

[0054] We collect past semantic tracing cases from processing VAT invoices, which contain the semantic tracing trajectories of key fields. A pre-defined business rule base contains rules that should be followed for the semantic tracing of key fields in VAT invoice processing, such as "the initial semantic generation node of tax amount information must be the tax review stage" and "the propagation path must include the enterprise confirmation stage." By comparing the trajectories in the cases with these rules, cases that meet the rules are selected to form the tracing optimization dataset.

[0055] Step S1110: Use the source optimization dataset to adjust the initial semantic source tracing trajectory, correct the deviation records in the propagation path, supplement the missing semantic association information, and finally generate the field semantic source tracing trajectory.

[0056] The cases in the source tracing optimization dataset conform to business rules. By analyzing these cases, potential deviations in the initial semantic source tracing trajectory can be identified. For example, in one initial trajectory, the propagation path of "tax amount information" might omit the "enterprise confirmation" step. By referring to cases in the dataset, this trajectory can be corrected by supplementing the "enterprise confirmation" information, thereby generating an accurate field semantic source tracing trajectory.

[0057] Step S120: Collect the scene association features of the invoice carrier, and generate scene feature guidance signals based on the field semantic tracing trajectory. The scene association features are the layout association information, field function association information and business process association information of the invoice carrier in the business scenario.

[0058] Invoice carriers, such as paper or electronic VAT invoices, exhibit specific scenario-related characteristics in their business contexts. For example, paper invoices have a fixed format, including headers, bodies, and footers, each containing different fields—this is format-related information. Different fields have different functions; for instance, the "amount" and "tax amount" fields together constitute the core value information of the invoice—this is field function-related information. The business processes of invoice issuance, authentication, and deduction are also reflected on the invoice carrier; for example, an authentication mark appears after successful authentication—this is business process-related information. After collecting these characteristics, scenario feature guidance signals are generated by combining them with field semantic tracing trajectories, which are used to subsequently locate the physical presentation areas of key fields.

[0059] Step S121: Input the bill carrier into the scene feature acquisition module. The scene feature acquisition module performs overall business scene positioning on the bill carrier, determines the specific business area and business process to which the bill carrier belongs, and forms a business scene positioning result.

[0060] The scene feature acquisition module analyzes the input invoice carrier. For a VAT special invoice, the module identifies the specific business area to which it belongs, namely "VAT invoice management," by recognizing the identification information on the invoice, such as the words "VAT special invoice" and the specific coding rules of the invoice code. The business process may be "certified and awaiting deduction," thus forming a business scene positioning result.

[0061] Step S122: Based on the business scenario positioning results, extract the layout association information of the bill carrier to form a layout association information set. The layout association information includes the overall layout form of the bill, the field arrangement method, the regional division mode, and the positional relationship between different regions.

[0062] Step S1221: Based on the business scenario positioning result, retrieve the standard invoice format library corresponding to the business scenario. The standard invoice format library contains common invoice overall layout templates, field arrangement templates, and area division templates for this business scenario.

[0063] In the "Value-Added Tax (VAT) Invoice Management" business scenario, the standard invoice template library contains various templates for VAT special invoices. For example, there are different versions of VAT special invoice templates, including templates used in different regions and periods. These templates define the standard format for overall layout, field arrangement, and regional division.

[0064] Step S1222: Calculate the layout similarity between the bill carrier and each template in the standard bill format library, select the template with the highest similarity as a reference, and determine the overall layout form of the bill carrier, including symmetric type coding, regional distribution ratio value, and position coordinate range of core fields.

[0065] The VAT invoices to be analyzed are compared with templates in the standard invoice template library. The layout similarity is calculated by comparing the overall structure of the invoices, the shape and size of each area, etc. The template with the highest similarity is selected, and the overall layout of the invoice is determined based on this template. For example, the symmetrical type code might be "left-right symmetrical," the area distribution ratio might be "20% for the header, 60% for the body, and 20% for the footer," and the position coordinate range of core fields such as "amount" and "tax amount" might be "slightly right of the middle of the body, horizontal coordinate x1-x2, vertical coordinate y1-y2."

[0066] Step S1223: Perform region segmentation and identification on the document carrier. According to the field functional attributes and location distribution, divide the document carrier into multiple functional regions and extract the boundary features and size features of each functional region.

[0067] Based on the functional attributes of the fields, the VAT invoice is divided into functional areas such as "Buyer Information Area", "Seller Information Area", "Product Details Area", "Amount and Tax Amount Area", and "Remarks Area". By identifying boundary features such as edge lines and color differences of each area, the range of the area is determined, and then the size features of each area are extracted, such as "Width a, Height b of Buyer Information Area".

[0068] Step S1224: Analyze the field arrangement in each functional area, record the spacing characteristics, alignment and hierarchical relationship between fields. Field arrangement methods include horizontal arrangement, vertical arrangement and mixed arrangement.

[0069] In the "Product Details Area," fields such as "Product Name," "Specifications," "Quantity," "Unit Price," and "Amount" are typically arranged horizontally with fixed spacing between each field. In the "Buyer Information Area," fields such as "Buyer Name" and "Buyer Taxpayer Identification Number" may be arranged vertically, aligned vertically. Analyze and record the spacing characteristics between these fields, such as "horizontal spacing c, vertical spacing d," alignment methods such as "left alignment" and "center alignment," and hierarchical relationships such as "the Product Name field is a first-level field in the Product Details Area, and the specific product names below it are second-level fields."

[0070] Step S1225: Construct a regional association matrix. The rows and columns of the regional association matrix represent functional areas, and the matrix elements represent the positional relationship between two functional areas, including adjacency, inclusion, and relative orientation.

[0071] Both rows and columns represent the functional areas defined above. For example, a row might be the "Buyer Information Area" and a column might be the "Seller Information Area." Matrix elements could be "adjacent, located in the top left and top right corners." A row might be the "Product Details Area" and a column might be the "Amount and Tax Area." Matrix elements could be "adjacent, located below." A row might be the "Remarks Area" and a column might be the "Table Tail Area." Matrix elements could be "contained, with the Remarks Area located within the Table Tail Area."

[0072] Step S1226: Extract the core field identifier of each functional area, determine the key field that plays a dominant role in the functional area, and establish the correspondence between the core field and the functional area.

[0073] In the "Amount and Tax Amount Area," the core fields are clearly "Amount" and "Tax Amount," which play a dominant role in this area and determine the core value of the invoice. In the "Buyer Information Area," the core field is "Buyer Taxpayer Identification Number," used to uniquely identify the buyer company. Establish a correspondence between these core fields and functional areas, such as "Amount and Tax Amount Area → Amount, Tax Amount" and "Buyer Information Area → Buyer Taxpayer Identification Number."

[0074] Step S1227: Analyze the positional correlation strength between the functional area where the core field is located and other functional areas. Based on the distance and functional correlation between functional areas, determine the representation method of positional correlation strength.

[0075] The "Amount / Tax Amount Area" and the "Product Details Area" are strongly related because the amount is calculated based on the quantity and unit price in the product details. Furthermore, the two areas are close together on the invoice, resulting in a strong positional correlation. Conversely, the "Remarks Area" and the "Buyer Information Area" are less related and relatively far apart, leading to a weaker positional correlation. The strength of this positional correlation can be represented by setting a numerical range, such as 0-1, with larger values ​​indicating a stronger correlation.

[0076] Step S1228: Integrate the overall layout, field arrangement, region division mode, region association matrix, core field region correspondence and position association strength information to form the initial layout association information.

[0077] The overall layout obtained from the above steps, such as symmetric type encoding, regional distribution ratio, etc.; field arrangement methods, such as horizontal arrangement, vertical arrangement and corresponding spacing, alignment, etc.; regional division mode, such as the division results of each functional area; regional association matrix; core field regional correspondence; location association strength information, etc., are integrated together to form the initial layout association information.

[0078] Step S1229: Set the feature importance threshold, filter each element in the initial layout association information, remove elements whose importance is lower than the feature importance threshold, and retain the core layout association elements.

[0079] Based on the importance of the bill business, a feature importance threshold is set, such as 0.7. Each element in the initial layout's associated information is evaluated. For example, the regional distribution ratio in the overall layout is very important for subsequently locating key fields, with an importance value potentially reaching 0.9, which is above the threshold and is therefore retained. Conversely, minor alignment deviations in field arrangement, with an importance value potentially below the threshold (around 0.3), are discarded. Through this screening process, core layout-related elements are retained.

[0080] Step S12210: Arrange the optimized layout association elements in logical order to form a layout association information set containing the core association information of the bill layout.

[0081] The core layout elements that are retained, such as key parameters in the overall layout, important field arrangement methods, and core regional relationship matrix information, are arranged in a logical order from the whole to the part and from the primary to the secondary, thus forming a set of layout relationship information.

[0082] Step S123: Analyze the functional attributes of each field in the invoice carrier, combine the business scenario positioning results, determine the functional positioning of each field in the corresponding business process, extract the functional synergy and functional complementarity relationships between fields, and form a set of field functional association information.

[0083] In the context of VAT invoice transactions, the functional attributes of each field differ. The "Buyer's Taxpayer Identification Number" and "Seller's Taxpayer Identification Number" identify the two parties in the transaction, serving an identity verification function; the "Amount" and "Tax Amount" reflect the transaction value, serving a value measurement function. Based on the business scenario positioning result "Certified and Pending Deduction," the functional positioning of the "Tax Amount" field can be determined as "tax amount to be deducted, used by the enterprise to deduct input tax." There are functional synergies between fields, such as the "Quantity," "Unit Price," and "Amount" fields working together to calculate the amount by multiplying the quantity by the unit price; and complementary relationships, such as the "Amount" and "Tax Amount" fields complementing each other, jointly constituting the total tax burden information related to the invoice. Extracting the above functional positioning, synergistic relationships, and complementary relationships forms a set of field functional association information.

[0084] Step S124: Retrieve the standard business process template of the business domain to which the invoice belongs, calculate the matching degree between the presentation order of the fields in the invoice carrier and the expected order of the fields in the standard business process template, and extract the flow association nodes and association rules of the fields in the business process based on the matching degree results to form a set of business process association information.

[0085] The standard business process template for the business area covered by VAT special invoices specifies the expected presentation order of fields, such as first buyer and seller information, then product details, followed by amount and tax information, and finally remarks and signature information. The actual field presentation order in the invoice carrier is compared with this template to calculate the matching degree. For example, if the field presentation order of an invoice is completely consistent with the template, the matching degree is 100%; if the order of some fields is reversed, the matching degree will decrease. Based on the matching degree results, the flow-related nodes of the fields in the business process are extracted, such as the flow order of "product details field → amount field → tax field," and association rules, such as "the amount field must be presented after the product details field," forming a set of business process related information.

[0086] Step S125: Integrate the set of layout-related information, the set of field function-related information, and the set of business process-related information, and generate scenario-related features through unified processing of feature dimensions.

[0087] The sets of layout-related information, field-functional related information, and business process-related information may have different feature dimensions. For example, layout-related information may be represented by spatial coordinates and dimensions, field-functional related information may be represented by function type and correlation strength, and business process-related information may be represented by time sequence and node relationships. By unifying the processing of feature dimensions, the information from these different dimensions is transformed into a unified feature space. For example, all features are converted into numerical vectors, allowing for comprehensive analysis and processing to generate scene-related features.

[0088] Step S126: Analyze the semantic tracing trajectory of the field, extract the semantic generation nodes, core propagation paths and key related field information of the key fields in the semantic tracing trajectory, and form a set of core tracing information.

[0089] The semantic traceability trajectory of the fields is analyzed. Taking the semantic traceability trajectory of "tax amount information" as an example, the semantic generation node "tax review process" is extracted, the core propagation path "tax review → enterprise confirmation → system archiving" is extracted, and key related field information such as "amount information" and "tax rate information" (the tax amount is calculated based on the amount and tax rate) is extracted. The above information is integrated to form the core traceability information set.

[0090] Step S127: Input the core information set of traceability into the guidance signal generation module. The guidance signal generation module establishes a mapping channel between traceability information and scenario features based on the correspondence between semantic generation nodes and business scenarios.

[0091] The guiding signal generation module stores the correspondence between semantic generation nodes and business scenarios, such as "tax audit process" corresponding to "value-added tax special invoice audit scenario". After the core traceability information set is input into the module, the module establishes a mapping channel between the semantic generation nodes, core propagation paths, and key related fields in the traceability information and the layout association information, field function association information, and business process association information in the scenario association features, so that the traceability information can guide the analysis direction of the scenario features.

[0092] Step S128: Based on the mapping channel, filter the feature information in the scene association features that is directly related to the semantic origin of the key field, strengthen the representation strength of the feature information, and form an association feature enhancement set.

[0093] By mapping channels, information directly related to the semantic origin of key fields in scene-related features is identified. For example, for "tax amount information," its semantic origin is related to key related fields such as "amount information" and "tax rate information." Therefore, the layout-related information (such as the position of these fields on the invoice) and the field function-related information (such as their relationship to tax calculation) related to "amount information" and "tax rate information" in scene-related features are directly relevant features. This information is then enhanced, such as by increasing its weight in the feature vector, to make its representation stronger, forming a set of enhanced related features.

[0094] Step S129: Based on the weights of each feature in the associated feature enhancement set and the node depth of the key field in the semantic tracing trajectory, generate an initial guiding signal containing feature coordinate information, extracted priority weights, and association strength coefficients.

[0095] In the feature enhancement set, each feature has a different weight, with higher-weighted features being more important for locating key fields. The node depth of a key field in the semantic tracing trajectory is also considered; for example, "tax amount information" has 3 nodes in the propagation path, with a node depth of 3. Deeper nodes may require higher extraction priority. Combining these two factors, an initial guiding signal is generated. Feature coordinate information indicates the possible location range of the key field on the invoice, extraction priority weights determine the order in which key fields are extracted, and the association strength coefficient represents the degree of association between each feature and the key field.

[0096] Step S1210: Using an optimization algorithm with the objective function of improving the positioning accuracy of key fields, the priority weights and correlation strength coefficients in the initial guidance signal are iteratively adjusted to generate the final scene feature guidance signal.

[0097] Optimization algorithms such as gradient descent are used to improve the accuracy of key field localization as the objective function. Through multiple iterations, the priority weights and association strength coefficients in the initial guiding signal are adjusted. For example, in a certain iteration, if it is found that the accuracy of locating "tax information" according to the current priority weights and association strength coefficients is not high, the algorithm will adjust these two parameters according to the error situation until the localization accuracy reaches the expected target. The signal obtained at this point is the final scene feature guiding signal.

[0098] Step S130: Based on the scene feature guidance signal, locate the physical presentation area of ​​the key field in the document carrier, extract the original information of the key field within the physical presentation area of ​​the key field, the original information includes text representation information and field form association information, and form a set of original information of the key field.

[0099] The feature coordinate information in the scene feature guidance signal indicates the possible location range of key fields, and the extracted priority weights determine which key field to locate first. For example, if the guidance signal indicates that the physical presentation area of ​​"amount information" may be located in the lower middle part of the document, it has a higher priority, so this area is locked first. After locking, the text representation information within the area is extracted, such as the text content "10000.00", as well as the field form association information, such as the font, font size, and color of the text, and the spacing and alignment with the surrounding "tax amount" field, forming a set of native information for the key fields.

[0100] Step S131: Analyze the feature pointing information in the scene feature guidance signal to determine the layout positioning basis, function association positioning basis, and business process positioning basis of the key fields in the ticket carrier.

[0101] Feature-based positioning information encompasses multiple positioning criteria. Layout-based positioning may be determined by layout association information in scene-related features, such as the fixed location area of ​​a key field within the document layout; function-related positioning is based on field function association information, such as the functional synergy with other key fields; and business process-related positioning is based on business process association information, such as the presentation position of the key field on the document corresponding to the flow sequence within the business process. By analyzing this information, it becomes clear how to locate the key field.

[0102] Step S132: Based on the layout positioning criteria, preliminary screening of candidate physical areas in the document carrier that meet the layout characteristics of key fields is carried out. The layout characteristics include the area size range, the location distribution pattern, and the spacing relationship with other areas.

[0103] Based on the layout and positioning criteria, such as the "Amount Information" field typically being a rectangular area 5 cm wide and 1 cm high in the VAT invoice layout, its location is below the "Product Details Area" and above the "Tax Information" field, with a spacing of 0.5 cm between it and the "Product Details Area" and 0.3 cm between it and the "Tax Information" field. By searching the invoice carrier according to these layout characteristics, preliminary screening is performed to identify candidate physical areas that meet the criteria.

[0104] Step S133: Based on the functional association positioning criteria, analyze the functional synergy between the candidate physical region and the surrounding field region, and filter out the candidate physical regions that are closely related to the key field functions to narrow down the positioning range.

[0105] The "Amount Information" field is closely related to the "Quantity" and "Unit Price" fields in the "Product Details Area" because the amount is obtained by multiplying the quantity by the unit price. Analyze the functional collaboration between the initially selected candidate physical areas and the "Product Details Area," such as whether the area has a logical data calculation relationship with the "Product Details Area," and filter out candidate physical areas with close functional relationships to further narrow down the location range.

[0106] Step S134: Based on the business process positioning criteria, verify whether the presentation order of the candidate physical areas after narrowing down the scope in the document carrier conforms to the field flow order in the business process, and lock the physical presentation area of ​​the final key field.

[0107] In the VAT invoice business process, the field flow order is usually "goods details → amount → tax amount". Verify whether the presentation order of the candidate physical areas after narrowing down the scope conforms to this order, such as whether the area is presented after the "goods details area" and before the "tax amount information" field. If it conforms, then lock that area as the physical presentation area of ​​the final key field.

[0108] Step S135: Perform pixel-level scanning on the physical presentation area of ​​each key field, extract the original pixel distribution information of the text within the physical presentation area of ​​the key field, preserve the original shape and arrangement structure of the text, and form the original pixel information of the text.

[0109] Image scanning technology is used to perform pixel-level scanning of the locked physical display area to obtain information such as the color and brightness of each pixel within the area. For example, within the physical display area of ​​the "Amount Information" field, each stroke of the text "10000.00" is composed of specific pixels. After scanning, the original distribution of these pixels is preserved, including the original shape of the text such as its size, font, and slant, as well as the arrangement structure between the text, such as whether there are spaces and the alignment, forming the original pixel information of the text.

[0110] Step S136: Using deep learning-driven text morphology recognition technology, the original pixel information of the text is analyzed to extract the original features of the strokes, the original features of the character combinations, and the original features of the text arrangement, forming the original information of the text morphology.

[0111] Step S1361: Input the original pixel information of the text into the input layer of the deep learning text morphology recognition model. The deep learning text morphology recognition model includes a feature extraction subnetwork, a morphology parsing subnetwork, and a feature integration subnetwork.

[0112] The original pixel information of the text is input into the input layer of the deep learning text morphology recognition model in the form of image data, and the subsequent morphology analysis process begins.

[0113] Step S1362: Perform multi-scale convolution operations on the original pixel information of the text through the convolutional layer of the feature extraction sub-network to extract pixel distribution features at different scales and form a multi-scale pixel feature set.

[0114] The convolutional layers of the feature extraction subnetwork contain convolutional kernels of different sizes, such as 3x3 and 5x5. When performing convolution operations on the raw pixel information of text, convolutional kernels of different sizes can extract features at different scales. For example, small convolutional kernels extract stroke detail features, while large convolutional kernels extract the overall outline features of the text. Through multi-scale convolution operations, pixel distribution features at different scales are obtained and combined to form a multi-scale pixel feature set.

[0115] Step S1363: Use an edge detection algorithm to process the multi-scale pixel feature set, enhance the response value of the stroke edge features, and use a filter to suppress the response value of the background area to form an enhanced pixel feature set.

[0116] Edge detection algorithms can identify the edge parts of text strokes. By enhancing the response values of these edge parts, the stroke features become more prominent. At the same time, filters such as Gaussian filters are used to process the background area, reducing the response values of the background area and minimizing background interference. After the above processing, an enhanced pixel feature set is obtained, in which the text stroke features are clearer.

[0117] Step S1364: In the morphological analysis sub-network, perform stroke segmentation on the enhanced pixel feature set, separate the pixel regions corresponding to different strokes of individual characters, and extract the length feature, arc feature, and direction feature of each stroke to form the original stroke features.

[0118] The morphological analysis sub-network first performs stroke segmentation on the characters in the enhanced pixel feature set. For example, for the character "人", the pixel regions corresponding to the left-falling stroke and the right-falling stroke can be segmented. Then, analyze the pixel regions of each stroke to extract the length feature, such as the number of pixels in the stroke; the arc feature, such as the degree of curvature of the stroke; and the direction feature, such as whether the stroke is from the upper left to the lower right or from the upper right to the lower left. These features together form the original stroke features.

[0119] Step S1365: Perform character combination analysis on the segmented individual character pixel regions, and extract the connection method, overlap feature, and spacing feature between characters to form the original character combination features.

[0120] In the arrangement of text, there are different combination methods between characters. For example, the two characters "金额" are connected left and right, and the numbers in "123" are arranged continuously. Analyze the segmented individual character pixel regions, extract the connection method between characters, such as left connection, right connection; the overlap feature, such as whether there is an intersection and overlap of character strokes; and the spacing feature, such as the pixel distance between characters, to form the original character combination features.

[0121] Step S1366: Expand the analysis scope to the physical presentation area of the entire keyword field, and extract the arrangement direction, alignment method, line spacing feature, and column spacing feature of all the characters in the physical presentation area of the keyword field to form the original text arrangement features.

[0122] Analyze the overall arrangement of the characters in the physical presentation area of the entire keyword field. The arrangement direction may be horizontal or vertical; the alignment method may be left alignment, right alignment, or center alignment; the line spacing feature refers to the pixel distance between different lines of characters; the column spacing feature refers to the pixel distance between different columns of characters. Extract these features to form the original text arrangement features.

[0123] Step S1367: The original features of strokes, character combinations, and text arrangement are processed to unify dimensions through the fully connected layer of the morphological parsing subnetwork, so that the original features of strokes, character combinations, and text arrangement are in the same dimensional space.

[0124] The original features of strokes, character combinations, and text arrangements may have different dimensions. For example, the original features of strokes may be three-dimensional vectors, the original features of character combinations may be five-dimensional vectors, and the original features of text arrangements may be four-dimensional vectors. The fully connected layers of the morphological parsing subnetwork transform these features of different dimensions into the same dimensional space through matrix transformations and other operations, for example, converting them all into ten-dimensional vectors, so as to facilitate subsequent integration processing.

[0125] Step S1368: Input the original stroke features, character combination features, and text arrangement features after unifying dimensions into the feature integration subnetwork, and perform weighted fusion according to the preset feature weight coefficients to generate comprehensive text morphology features.

[0126] The feature integration subnetwork pre-defines the weight coefficients for each feature. For example, the weight coefficient for the original stroke feature is 0.4, the weight coefficient for the original character combination feature is 0.3, and the weight coefficient for the original text arrangement feature is 0.3. The features after unifying the dimensions are multiplied by their respective weight coefficients and then summed to obtain the comprehensive text morphology features.

[0127] Step S1369: Calculate the similarity between the comprehensive text morphology features and the original pixel information of the text in the feature space. If the similarity is higher than the preset threshold, it is determined that the comprehensive text morphology features meet the originality requirements, and the comprehensive text morphology features that have passed the originality verification are output as the original text morphology information containing the original stroke features, the original character combination features, and the original text arrangement features.

[0128] In the feature space, both the integrated text morphological features and the original pixel information of the text have corresponding feature vector representations. The similarity between these two vectors is measured by methods such as calculating the cosine similarity. A preset similarity threshold, such as 0.85, is used. If the calculated similarity is higher than this threshold, it indicates that the integrated text morphological features have well preserved the morphological features in the original pixel information of the text, meeting the requirement of originality, and are therefore output as the original text morphological information.

[0129] Step S137: Analyze the relationship between the text and surrounding elements within the physical presentation area of ​​the key field, and form the field form relationship information.

[0130] Within the physical presentation area of ​​key fields, text and surrounding elements exhibit various relationship patterns. For example, the positional relationship between text and borders: the "Amount Information" field might be located within a rectangular border, with the text's top, bottom, left, and right margins to the border each being 0.2 cm. The combination of text and symbols: for instance, an amount might be followed by the "¥" symbol. The distribution of text and blank space: there might be blank space around the text to highlight its content. Analyzing these relationship patterns helps to establish field morphological association information.

[0131] Step S138: Associate and bind the original pixel information of the text, the original text shape information, and the field shape association information to form the original information unit of a single key field.

[0132] The original pixel information of text belonging to the same key field, such as the pixel distribution of the "amount information" field; the original text form information, such as the strokes, character combinations, and arrangement features of the text in this field; and the field form association information, such as the association forms with surrounding elements, are associated and bound together to form a complete information unit, that is, the original information unit of a single key field.

[0133] Step S139: Set an information integrity threshold, scan the original information units of all key fields, filter out the original information units whose number of information dimensions is lower than the integrity threshold, and retain the original information units that meet the integrity requirements.

[0134] The information integrity threshold can be set based on the number of information dimensions that the original information unit of the key field should contain. For example, if the original pixel information of the text, the original text shape information, and the field shape association information are all complete, the integrity threshold is set to 3. All original information units are scanned to check whether each unit contains information in these three dimensions. If a unit is missing original text shape information, its information dimension count is 2, which is lower than the threshold, and it is filtered out; if all three dimensions are complete, it is retained.

[0135] Step S1310: Integrate all the native information units of the key fields that meet the requirements, classify and arrange them according to the type of the key fields, and form a key field native information set containing complete native information of all key fields.

[0136] The retained original information units that meet the requirements, such as the original information units of key fields such as "invoice code", "invoice number", "amount information", "tax information" and "buyer and seller information", are classified according to the type of key field. For example, "invoice code" and "invoice number" are classified as identification fields, "amount information" and "tax information" are classified as value fields, and "buyer and seller information" is classified as subject fields. Then they are arranged to form a set of original information of key fields.

[0137] Step S140: Construct a field dynamic association model. Input the original information set of key fields and the scene association features into the field dynamic association model. Model the dynamic dependency relationship between key fields and between key fields and the scene through deep learning algorithms to generate a field dynamic association graph.

[0138] The Field Dynamic Association Model is a deep learning-based model capable of learning the complex relationships between key fields and between key fields and scenarios. After inputting the original information sets of key fields and scenario-related features into the model, it captures the dynamic dependencies between key fields through multi-layer neural network computation and learning. For example, there is a calculation dependency relationship between "amount information" and "tax amount information" (tax amount = amount × tax rate); and dynamic dependencies between key fields and scenarios. For instance, in the "certified and awaiting deduction" scenario, the dependency relationship of "tax amount information" is waiting for the enterprise to perform the deduction operation. By modeling these relationships, a field dynamic association graph is generated, graphically representing these dependencies.

[0139] Step S141: Build the basic architecture of the field dynamic association model. This basic architecture includes a feature input layer, an association modeling layer and a graph generation layer. The feature input layer supports the synchronous input of the original information set of key fields and scene-related features.

[0140] The feature input layer receives the native information set of key fields and scene-related features, and converts them into a data format that the model can process, such as tensors. The association modeling layer is the core of the model, containing multiple neural network layers for learning and modeling dynamic dependencies. The graph generation layer then converts the relationships learned by the association modeling layer into a graphical, dynamic field association graph.

[0141] Step S142: Perform feature transformation processing on the original information set of key fields, converting textual representation information and field morphological association information into numerical feature vectors that can be processed by the field dynamic association model, forming a set of key field feature vectors.

[0142] The textual representations in the original information set of key fields, such as the text content "10000.00", and the field morphological associations, such as the positional relationship between the text and the border, are all non-numerical information. They need to undergo feature transformation processing, such as converting the text content into corresponding numerical codes and the positional relationships into coordinate values. This transforms the above information into numerical feature vectors, with each key field corresponding to a feature vector, thus forming a set of key field feature vectors.

[0143] Step S143: Perform dimension adaptation processing on the scene-related features, adjust the feature dimensions of the scene-related features to make the dimensions of the scene-related features consistent with the key field feature vector set, and form an adapted scene feature vector.

[0144] Scene-related features may have different dimensions than the key field feature vector set. For example, a scene-related feature might be a 100-dimensional vector, while each vector in the key field feature vector set is 50-dimensional. Dimensional adaptation processing, such as using principal component analysis to reduce the dimensionality of the scene-related features or adding zero vectors to increase their dimensionality, can adjust them to have the same dimensions as the key field feature vector set, thus forming an adapted scene feature vector.

[0145] Step S144: Input the key field feature vector set and the adaptive scene feature vector into the feature input layer, and transmit the key field feature vector set and the adaptive scene feature vector to different processing branches of the associated modeling layer through the feature splitting transmission mechanism.

[0146] After receiving the key field feature vector set and the adaptation scene feature vector, the feature input layer transmits the key field feature vector set to the "field association branch" of the association modeling layer through the feature splitting transmission mechanism. This branch is specifically used to learn the dependency relationship between key fields. The adaptation scene feature vector is transmitted to the "scene association branch" of the association modeling layer. This branch is specifically used to learn the dependency relationship between key fields and the scene.

[0147] Step S145: In the key field branch of the association modeling layer, a multilayer perceptron is used to perform deep processing on the key field feature vector set to extract the potential association features between key fields and form intra-field association features.

[0148] A multilayer perceptron contains multiple hidden layers, each composed of multiple neurons. The key field feature vector set is input into the multilayer perceptron. After processing by neurons in the first hidden layer, simple correlation features are extracted, such as linear relationships between fields. The second hidden layer then extracts more complex non-linear correlation features. After deep processing through multiple hidden layers, the latent correlation features between key fields are finally extracted, such as the non-linear calculation relationship between "amount information" and "tax information," forming intra-field correlation features.

[0149] Step S146: In the scene association branch of the association modeling layer, a recurrent neural network is used to perform temporal modeling on the adaptive scene feature vector to capture the dynamic association pattern between scene features and key fields, forming scene field association features.

[0150] For example, step S1461: Arrange the adaptive scenario feature vectors according to the time sequence of the business process to form a time-series scenario feature vector sequence, so that the order of the time-series scenario feature vector sequence is consistent with the order of the business process.

[0151] In the VAT special invoice business, the time sequence of the business process is: invoice application → tax review → enterprise confirmation → system archiving. The adaptation scenario feature vector will contain features related to these business process stages. Arranging the adaptation scenario feature vectors according to this time sequence forms a time-series scenario feature vector sequence.

[0152] Step S1462: Construct the input sequence of the recurrent neural network, and divide the temporal scene feature vector sequence into vector subsequences of fixed length, with each vector subsequence corresponding to a stage of the business process.

[0153] The temporal scene feature vector sequence is divided according to the stages of the business process, with each stage corresponding to a vector subsequence of fixed length. For example, the invoice application stage corresponds to a vector subsequence of length 20, the tax review stage corresponds to a vector subsequence of length 20, and so on, to construct the input sequence of the recurrent neural network.

[0154] Step S1463: Input the vector subsequence into the hidden layer of the recurrent neural network, and capture the correlation between the scene feature vector of each stage and the scene feature vector of the preceding and subsequent stages through the state transmission of the hidden layer neurons.

[0155] The hidden layer neurons of a recurrent neural network possess memory capabilities. After the current stage's vector subsequence is input, the neuron's state depends not only on the current input but also on the states of neurons in previous stages. Through this state propagation, the network can capture the correlation between the current stage's scene feature vector and the scene feature vectors of previous and subsequent stages. For example, the causal relationship between the scene feature vectors of the tax audit stage and the invoice application stage, as well as the continuity relationship with the scene feature vectors of the enterprise confirmation stage.

[0156] Step S1464: Set up the attention mechanism module, calculate the correlation score between the hidden layer output features and the key field feature vectors, and perform a weighted summation of the hidden layer output features based on the correlation score to obtain the weighted hidden layer features.

[0157] The attention mechanism module is used to highlight the hidden layer output features that are relevant to the key field feature vector. It calculates the relevance score between each element in the hidden layer output feature and the key field feature vector; elements with higher relevance scores have a greater impact on the relationship between the key field and the scene. Then, each element in the hidden layer output feature is multiplied by its corresponding relevance score, and the results are summed to obtain a weighted hidden layer feature, making the network pay more attention to important feature information.

[0158] Step S1465: The weighted hidden layer features are processed by the output layer of the recurrent neural network to generate stage association features corresponding to each vector subsequence. These stage association features contain a description of the association between the scene and key fields in that stage.

[0159] The output layer of a recurrent neural network performs linear transformations and activation functions on the weighted hidden layer features, converting them into feature vectors that represent the association between the scene and key fields at that stage—the stage-related features. For example, the stage-related features of the invoice application stage might describe the association between the initial entry of the "buyer / seller information" field and the scene at that stage.

[0160] Step S1466: Perform time-series fusion processing on all stage-related features. Based on the stage weights of the business process, weighted aggregate the related features of different stages to form a set of time-series related features.

[0161] Different stages of the business process have varying importance, so corresponding stage weights are assigned. For example, the tax audit stage has a weight of 0.4, the enterprise confirmation stage has a weight of 0.3, and the invoice application stage and system archiving stage each have a weight of 0.15. The associated features of each stage are multiplied by their corresponding stage weights and then summed to obtain a time-series associated feature set. This time-series associated feature set integrates the association features between various stage scenarios and key fields.

[0162] Step S1467: Use the temporal pattern mining algorithm to analyze the temporal correlation feature set, extract the dynamic change pattern of the relationship between scene features and key fields, and the dynamic change pattern includes the change curve of correlation strength and the transformation sequence of correlation type.

[0163] Temporal pattern mining algorithms analyze temporal association feature sets to identify dynamic change patterns. For example, the change curve of association strength may show that during the tax audit stage, the association strength between the scenario feature and the "tax amount information" field increases sharply, remains stable during the enterprise confirmation stage, and decreases slightly during the system archiving stage; the transformation sequence of association type may be from "preliminary association" in the invoice application stage to "strong association" in the tax audit stage, and then to "confirmed association" in the enterprise confirmation stage.

[0164] Step S1468: Based on the dynamic change pattern, construct a scenario field association model. This scenario field association model can describe the association pattern between scenario features and key fields at different business stages.

[0165] Based on the extracted dynamic change patterns, a scenario field association model is constructed. The model includes association pattern parameters for different business stages, such as association strength parameters and association type parameters. These parameters describe how scenario features and key fields are associated at different business stages.

[0166] Step S1469: Reconstruct the temporal correlation feature set through the scene field correlation model to generate reconstructed features that can accurately represent the dynamic correlation rules. Optimize the dimensions of the reconstructed features, remove redundant dimensions, retain the core correlation dimensions, and form scene field correlation features that contain scene features and the dynamic correlation rules of key fields.

[0167] By using a scene field association model to process the time-series association feature set, features that more accurately represent dynamic association patterns are reconstructed. Then, the reconstructed features are optimized in terms of dimensions. By analyzing the importance of each dimension, redundant dimensions that do not contribute much to representing dynamic association patterns are removed, while core association dimensions, such as association strength and association type, are retained, forming scene field association features.

[0168] Step S147: Construct an association fusion submodule. Input the in-field association features and the scene field association features into the association fusion submodule. Through the feature interaction mechanism, achieve deep fusion of the two types of association features and generate comprehensive association features.

[0169] The feature interaction mechanism in the association fusion submodule can be operations such as element-wise multiplication, addition, and concatenation. For example, it can concatenate the in-field association features and the scene field association features to form a longer feature vector; or it can multiply corresponding elements in two features to highlight their interaction relationship. Deep fusion is achieved through these methods to generate comprehensive association features, which simultaneously contain the association information between key fields and between key fields and the scene.

[0170] Step S148: Based on the comprehensive association features, an initial association graph is constructed using a graph neural network algorithm. The nodes in the initial association graph correspond to key fields and scene association elements, and the lines between nodes correspond to association relationships.

[0171] Graph neural network algorithms can process graph-structured data. By inputting comprehensive relational features into a graph neural network, the algorithm automatically identifies key fields and scenario-related elements as nodes, such as nodes for "amount information," "tax information," and "tax audit scenario." Then, based on the relationships inherent in the comprehensive relational features, connections are established between the nodes. For example, a connection between the "amount information" node and the "tax information" node indicates an intra-field relationship between them; a connection between the "tax information" node and the "tax audit scenario" node indicates a scenario-field relationship between them, thus constructing an initial relational graph.

[0172] Step S149: Query the preset weight adjustment strategy table according to the business scenario identifier, and dynamically calculate the weight value of the connection between nodes in the initial association graph based on the association strength coefficient of the key field to complete the weight adjustment.

[0173] The preset weight adjustment strategy table stores the calculation rules for the connection weights between nodes under different business scenarios. By querying this table based on a business scenario identifier, such as "VAT invoice audit scenario," the corresponding calculation rules are obtained. Then, combined with the association strength coefficient of key fields (e.g., the association strength coefficient between "amount information" and "tax information" is 0.9), the weight value of the connection between these two nodes in the initial association graph is dynamically calculated according to the calculation rules, such as 0.9 × preset base weight 1.0 = 0.9, thus completing the weight adjustment.

[0174] Step S1410: Optimize the structure of the adjusted association graph, remove redundant association lines, supplement missing core association nodes, and generate a field dynamic association graph that can represent the dynamic dependency relationship between key fields and between key fields and scenarios.

[0175] The adjusted association graph may contain some redundant connections, such as connections between two nodes with very weak association strength. These connections do not contribute much to the representation of the overall association relationship and can be removed. Additionally, some core association nodes may be missing. For example, the "tax rate information" node is a core association node when calculating tax amounts; if this node is not present in the initial association graph, it needs to be added. Through the above structural optimization, the final dynamic field association graph is generated.

[0176] Step S150: Based on the field dynamic association diagram, integrate the complete original information of key fields, scenario association information and dynamic dependency relationships to generate scenario-based structured invoice analysis results. The scenario-based structured invoice analysis results include the core content of fields, scenario association logic and dynamic association description.

[0177] The dynamic field relationship diagram clearly illustrates the dynamic dependencies between key fields and between fields and scenarios. It integrates complete native information of key fields, such as their specific content and morphological characteristics; scenario-related information, such as the format of the invoice, field functions, and business processes; and dynamic dependencies, such as calculation relationships between fields and stage-related relationships with scenarios, all organized in a structured manner to generate scenario-based structured invoice analysis results. This scenario-based structured invoice analysis result clearly presents the core information of the invoice, the scenario-related logic, and the dynamic relationship descriptions, facilitating user understanding and use of invoice information.

[0178] Step S151: Analyze the dynamic association graph of the fields, extract the association node information, association connection weight and association scene elements corresponding to each key field, and form the graph analysis result.

[0179] The dynamic relationship graph of the fields is analyzed. Each key field node, such as the "Amount Information" node, is associated with other key field nodes, such as the "Tax Information" node, as well as scenario element nodes, such as the "Tax Audit Scenario" node. These are the relationship node information. The weight value of the connection between the related nodes, such as 0.9, is the relationship connection weight. Scenario nodes such as "Tax Audit Scenario" are the related scenario elements. The above information is extracted to form the graph analysis result.

[0180] Step S152: Extract the dynamic dependency relationship description between key fields from the graph analysis results to form a set of field association relationships. The dynamic dependency relationship description includes association type, association strength and association triggering conditions.

[0181] From the graph analysis results, the relationships between key fields are filtered out. For example, the relationship type between "Amount Information" and "Tax Information" is "Calculation Relationship," with a relationship strength of 0.9, and the triggering condition is "Amount Information is Determined and Tax Rate Information is Known." The relationship type between "Buyer Information" and "Seller Information" is "Transaction Relationship," with a relationship strength of 0.8, and the triggering condition is "A Valid Transaction Contract Exists." These dynamic dependency descriptions are then organized to form a set of field relationship relationships.

[0182] Step S153: Retrieve the original information set of key fields, extract the core content of the text representation of each key field and the core information associated with the field form, and form a core information set of key fields.

[0183] Key information is extracted from the original information set of key fields. This includes textual representations of core content such as "10000.00 yuan" in the "Amount Information" field and "1300.00 yuan" in the "Tax Information" field; and the relationship between field forms and core information, such as the "Amount Information" field being located below the "Product Details Area" and left-aligned with the "Tax Information" field. This information is then integrated to form the core information set of key fields.

[0184] Step S154: Integrate the set of field relationships with the set of core information of key fields, establish a corresponding mapping between the core information of key fields and the relationships, and form a set of field relationship information.

[0185] The relationships in the set of field relationships are mapped to the core information in the set of core information of key fields. For example, the core content of "amount information" "10000.00 yuan" and the core content of "tax information" "1300.00 yuan" are mapped to each other through the "calculation relationship". The core content of "buyer information" and the core content of "seller information" are mapped to each other through the "transaction relationship". This forms a set of field relationship information, which contains both the core information of key fields and the relationships between them.

[0186] Step S155: Extract the core scene elements from the scene association features to form a set of core scene information. The core scene elements include business scene type, layout core features, and core nodes of business process.

[0187] The scene association features contain multiple scene elements, from which the core elements are extracted. Business scene types include "VAT invoice certified and awaiting deduction"; core layout features include "the amount information field is located in the lower middle of the invoice, with a size of 5cm × 1cm"; core business process nodes include "tax review node" and "enterprise confirmation node". These core scene elements are integrated to form a set of core scene information.

[0188] Step S156: Integrate the core information set of the scene with the field association information set to construct a scene field association framework and determine the association and positioning of each key field core information in the corresponding scene.

[0189] This involves integrating the business scenario type, layout features, and business process nodes from the core information set of the scenario with the key field information and their relationships from the field association information set. For example, in the scenario of "VAT special invoices have been certified and are awaiting deduction," the position of the "tax amount information" field is determined based on the layout features, and its association with the "tax review node" and "enterprise confirmation node" is determined based on the business process nodes. This constructs a scenario field association framework, clarifying the association and positioning of each key field's core information within the scenario.

[0190] Step S157: Based on the scenario field association framework, arrange the core information of key fields, core information of scenarios, and field associations in a structured manner according to the business logic order to form a preliminary structured result.

[0191] The business logic sequence typically proceeds from basic invoice information to core transaction information, and then to business process information. Following this order, key fields such as "invoice code," "invoice number," "buyer / seller information," "amount information," and "tax information"; scenario-based core information such as business scenario type, core layout features, and core business process nodes; and field relationships such as the calculation relationship between "amount information" and "tax information" are arranged to form a preliminary structured result.

[0192] Step S158: Supplement the preliminary structured results with scenario-based descriptions, adding corresponding scenario adaptation descriptions to the core information of each key field. The scenario adaptation descriptions include the scenario function positioning and the scenario association value.

[0193] In the scenario of "VAT special invoices have been certified and are awaiting deduction," scenario adaptation instructions are added to the "Tax Amount Information" field. The scenario's functional positioning is "as the basis for enterprises to deduct input tax," and its associated value is "directly affecting the enterprise's tax burden and profit situation." Similarly, scenario adaptation instructions are added to the "Purchaser Information" field. The scenario's functional positioning is "identifying the entity responsible for tax deduction," and its associated value is "ensuring the correct attribution of deduction rights." These scenario-based additions enrich and detail the initial structured results.

[0194] Step S159: Construct a scenario-based structured template. This scenario-based structured template is divided into information display modules according to the presentation requirements of different business scenarios. Each information display module corresponds to a type of core information.

[0195] For different business scenarios, such as "certified and awaiting deduction" and "deducted and archived," corresponding scenario-based structured templates are constructed. The templates are divided into different information display modules, such as the "Basic Invoice Information Module" to display invoice code, number, etc.; the "Transaction Core Information Module" to display amount, tax amount, etc.; the "Business Process Information Module" to display authentication status, deduction status, etc.; and the "Scenario Association Explanation Module" to display scenario adaptation instructions, etc.

[0196] Step S1510: Fill the structured information containing scenario adaptation instructions into the corresponding information display module of the scenario-based structured template, and integrate them to form a scenario-based structured invoice analysis result containing core field content, scenario association logic and dynamic association description.

[0197] The structured information, supplemented with contextualized information, is then populated into the corresponding modules of the contextualized structured template according to information type. For example, "invoice code" and "invoice number" are populated into the "basic invoice information module"; "amount information," "tax amount information," and their related relationships are populated into the "core transaction information module"; business scenario type and core nodes of the business process are populated into the "business process information module"; and scenario adaptation instructions are populated into the "scenario association instructions module." Through integration, the final contextualized structured invoice analysis results are generated.

[0198] Throughout the implementation of this method, sensitive data was collected, such as taxpayer identification numbers and company names in buyer and seller information. To protect the privacy of this data and prevent leakage, data anonymization techniques were employed to process the collected sensitive data. For example, some digits of the taxpayer identification number were replaced with asterisks (*), and key characters in the company name were encrypted. Simultaneously, encrypted transmission protocols, such as SSL, were used during data transmission to ensure that data was not intercepted during transmission. In the data storage stage, encrypted storage methods were used to encrypt the sensitive data in the database, allowing only authorized personnel to access it through decryption, thereby achieving the protection of sensitive data.

[0199] Figure 3 The diagram illustrates the hardware structure of a deep learning-integrated intelligent invoice information extraction and analysis system 100, provided by an embodiment of the present invention, for implementing the aforementioned deep learning-integrated intelligent invoice information extraction and analysis method. Figure 3 As shown, the intelligent ticket information extraction and analysis system 100 integrating deep learning may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0200] In one possible design, the intelligent invoice information extraction and analysis system 100 integrating deep learning can be a single server or a server group. The server group can be centralized or distributed (e.g., the intelligent invoice information extraction and analysis system 100 integrating deep learning can be a distributed system). In some embodiments, the intelligent invoice information extraction and analysis system 100 integrating deep learning can be local or remote. For example, the intelligent invoice information extraction and analysis system 100 integrating deep learning can access information and / or data stored in machine-readable storage medium 120 via a network. As another example, the intelligent invoice information extraction and analysis system 100 integrating deep learning can directly connect to machine-readable storage medium 120 to access stored information and / or data.

[0201] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by the intelligent invoice information extraction and analysis system 100 incorporating deep learning to perform or use in order to accomplish the exemplary methods described in this invention.

[0202] In the specific implementation process, one or more processors 110 execute computer-executable instructions stored in machine-readable storage medium 120, so that processor 110 can execute the intelligent ticket information extraction and analysis method integrating deep learning as described in the above method embodiment. Processor 110, machine-readable storage medium 120 and communication unit 140 are connected through bus 130. Processor 110 can be used to control the sending and receiving actions of communication unit 140.

[0203] The specific implementation process of processor 110 can be found in the various method embodiments executed by the intelligent invoice information extraction and analysis system 100 that integrates deep learning. The implementation principle and technical effect are similar, and will not be repeated here.

[0204] Furthermore, this embodiment of the invention also provides a readable storage medium containing computer-executable instructions. When a processor executes the computer-executable instructions, the above-mentioned intelligent invoice information extraction and analysis method integrating deep learning is implemented.

[0205] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. A method for intelligent bill information extraction and analysis based on deep learning, characterized in that, The method includes: Based on the field flow association logic of the bill business scenario, this study traces the semantic generation source and propagation path of key fields to generate a field semantic source tracing trajectory. The key fields include bill code, bill number, amount information, tax information, and buyer / seller information. The determination method of the field flow association logic includes: collecting field flow records of the entire bill business process to form a field flow record set. The field flow records cover the complete flow path of key fields from generation to archiving under different business scenarios, including the appearance form of the field and related fields in each stage of business application, review, confirmation, and archiving; performing association logic analysis on the field flow record set, using association rule mining algorithms to extract the related fields and flow conditions of each key field at different business nodes; and determining the generation trigger event, propagation data format, and receiving system identifier of the key field in the business process based on the timestamp of the field appearance and the business node sequence to form a field flow association logic diagram. Collect scene association features of the invoice carrier, and generate scene feature guidance signals based on the semantic tracing trajectory of the fields. The scene association features are the layout association information, field function association information and business process association information of the invoice carrier in the business scenario. Based on scene feature guidance signals, the physical presentation area of ​​key fields in the ticket carrier is targeted and locked, and the original information of key fields in the physical presentation area of ​​key fields is extracted. The original information includes text representation information and field form association information, forming a set of original information of key fields. Construct a field dynamic association model by inputting the original information set of key fields and scene association features into the field dynamic association model, and using deep learning algorithms to model the dynamic dependency relationships between key fields and between key fields and scenes, thereby generating a field dynamic association graph; Based on the field dynamic association graph, the complete original information of key fields, scenario association information and dynamic dependency relationships are integrated to generate scenario-based structured invoice analysis results. The scenario-based structured invoice analysis results include the core content of the fields, scenario association logic and dynamic association description. The collection of scene-related features of the invoice carrier, and the generation of scene feature guidance signals based on the semantic tracing trajectory of the fields, include: The bill carrier is input into the scene feature acquisition module, which performs overall business scene positioning on the bill carrier, determines the specific business area and business process to which the bill carrier belongs, and forms a business scene positioning result. Based on the business scenario positioning results, the layout association information of the bill carrier is extracted to form a layout association information set. The layout association information includes the overall layout form of the bill, the field arrangement method, the area division mode, and the positional relationship between different areas. Analyze the functional attributes of each field in the invoice carrier, combine the business scenario positioning results to determine the functional positioning of each field in the corresponding business process, extract the functional synergy and functional complementarity relationships between fields, and form a set of field functional association information; Retrieve the standard business process template of the business domain to which the bill belongs, calculate the matching degree between the presentation order of the fields in the bill carrier and the expected order of the fields in the standard business process template, and extract the flow association nodes and association rules of the fields in the business process based on the matching degree results to form a set of business process association information; The integrated layout-related information set, field function-related information set, and business process-related information set are processed uniformly through feature dimensions to generate the scenario-related features; Parse the semantic source tracing trajectory of the field, extract the semantic generation nodes, core propagation paths and key related field information of the key fields in the semantic source tracing trajectory, and form a set of core source tracing information; The core information set for tracing is input into the guidance signal generation module, which establishes a mapping channel between tracing information and scenario features based on the correspondence between semantic generation nodes and business scenarios. Based on the mapping channel, filter out the feature information in the scene association features that is directly related to the semantic origin of the key fields, strengthen the representation strength of the feature information, and form an association feature enhancement set; Based on the weights of each feature in the associated feature enhancement set and the node depth of the key field in the semantic tracing trajectory, an initial guiding signal containing feature coordinate information, extraction priority weights, and association strength coefficients is generated. An optimization algorithm is adopted, with the objective function of improving the positioning accuracy of key fields. The priority weights and correlation strength coefficients in the initial guidance signal are iteratively adjusted to generate the final scene feature guidance signal.

2. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 1, characterized in that, The field flow association logic based on the bill business scenario traces the semantic generation source and propagation path of key fields, generating a field semantic source tracing trajectory, including: The logic diagram of field flow association is input into the deep learning semantic tracing module. This deep learning semantic tracing module is based on the semantic rules of bill business, learns the evolution rules of key field semantics in the flow process, and establishes a semantic evolution model. Based on the semantic evolution model, the initial semantic generation node of each key field is located, and the business scenario information and initial representation form of the field are extracted from the initial semantic generation node. The initial semantic generation node is the business link in which the key field first has business meaning. Trace the propagation path of key fields from the initial semantic generation node to the final archiving node, and record the supplementary, adjusted or enhanced semantics of key fields by each business node in the propagation path to form a detailed record of semantic propagation. Integrate the initial semantic generation node information, propagation path and semantic propagation details of key fields to build a basic framework for semantic tracing; Set semantic association weights for fields, and assign corresponding association weights to each propagation link based on the degree of influence of different business nodes on the semantics of key fields, thereby strengthening the semantic influence representation of core propagation links; By integrating the semantic tracing framework with association weights, an initial semantic tracing trajectory for key fields is generated. Collect semantic traceability cases of fields in historical invoice processing, compare the semantic traceability trajectories in the cases with the preset business rule base, filter out the trajectory cases that meet the business rules, and form a traceability optimization dataset; The semantic source tracing initial trajectory is adjusted using the source tracing optimization dataset, the deviation records in the propagation path are corrected, the missing semantic association information is supplemented, and finally the semantic source tracing trajectory of the field is generated.

3. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 1, characterized in that, The step of using scene feature guidance signals to target and locate the physical presentation area of ​​key fields in the document carrier, and extracting the original information of the key fields within the physical presentation area of ​​the key fields, includes: Analyze the feature-guided signals in the scene to determine the layout positioning basis, functional association positioning basis, and business process positioning basis of key fields in the invoice carrier; Based on the layout positioning criteria, candidate physical areas that meet the layout characteristics of key fields are initially screened in the document carrier. The layout characteristics include the area size range, the location distribution pattern, and the distance relationship with other areas. Based on the functional association positioning criteria, analyze the functional synergy between candidate physical regions and surrounding field regions, and filter out candidate physical regions that are closely related to the functions of key fields to narrow down the positioning range; Based on the business process positioning criteria, verify whether the presentation order of the narrowed-down candidate physical areas in the invoice carrier conforms to the field flow order in the business process, and finally determine the physical presentation area of ​​the key fields. The physical presentation area of ​​each key field is scanned at the pixel level to extract the original pixel distribution information of the text within the physical presentation area of ​​the key field, preserving the original shape and arrangement structure of the text to form the original pixel information of the text. Using deep learning-driven text morphology recognition technology, the original pixel information of the text is analyzed to extract the original features of the strokes, the original features of the character combinations, and the original features of the text arrangement, thus forming the original information of the text morphology. Analyze the relationship between text and surrounding elements within the physical presentation area of ​​key fields, including the positional relationship between text and borders, the combination relationship between text and symbols, and the distribution relationship between text and blank areas, to form field form association information; The original pixel information of the text, the original text shape information, and the field shape association information are associated and bound together to form the original information unit of a single key field; Set an information integrity threshold, scan all native information units of key fields, filter out native information units with fewer information dimensions than the integrity threshold, and retain native information units that meet the integrity requirements; Integrate all eligible key field native information units, classify and arrange them according to the type of key field, and form a key field native information set containing complete native information of all key fields.

4. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 1, characterized in that, The construction of the field dynamic association model involves inputting the original information set of key fields and scene-related features into the field dynamic association model, and using deep learning algorithms to model the dynamic dependencies between key fields and between key fields and scenes, generating a field dynamic association graph, including: The basic architecture for building a dynamic field association model includes a feature input layer, an association modeling layer, and a graph generation layer. The feature input layer supports the synchronous input of the native information set of key fields and scene-related features. The original information set of key fields is processed by feature transformation, which converts textual representation information and field form association information into numerical feature vectors that can be processed by the field dynamic association model, forming a set of key field feature vectors. The scene-related features are adapted by adjusting their dimensions to match the dimensions of the key field feature vector set, thus forming a scene-adapted feature vector. The key field feature vector set and the adaptive scene feature vector are input into the feature input layer. Through the feature splitting transmission mechanism, the key field feature vector set and the adaptive scene feature vector are respectively transmitted to different processing branches of the association modeling layer. In the key field branch of the association modeling layer, a multilayer perceptron is used to perform deep processing on the key field feature vector set to extract the potential association features between key fields and form intra-field association features. In the scene association branch of the association modeling layer, a recurrent neural network is used to perform temporal modeling on the adaptive scene feature vectors, capture the dynamic association patterns between scene features and key fields, and form scene field association features. Construct an association fusion submodule, input the in-field association features and the scene field association features into the association fusion submodule, and achieve deep fusion of the two types of association features through the feature interaction mechanism to generate comprehensive association features; Based on comprehensive association features, an initial association graph is constructed using a graph neural network algorithm. The nodes in the initial association graph correspond to key fields and scene-related elements, and the lines between nodes correspond to association relationships. The system queries the preset weight adjustment strategy table based on the business scenario identifier, and dynamically calculates the weight value of the connection between nodes in the initial association graph based on the association strength coefficient of the key field, thus completing the weight adjustment. The adjusted association graph is structurally optimized by removing redundant association lines, supplementing missing core association nodes, and generating a dynamic association graph that can represent the dynamic dependencies between key fields and between key fields and scenarios.

5. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 1, characterized in that, The method, based on a dynamic field association graph, integrates the complete original information of key fields, scenario-related information, and dynamic dependencies to generate scenario-based structured invoice analysis results, including: The dynamic association graph of the fields is analyzed to extract the association node information, association connection weight and association scene elements corresponding to each key field, and form the graph analysis result. The dynamic dependency relationship description between key fields is extracted from the graph analysis results to form a set of field association relationships. The dynamic dependency relationship description includes association type, association strength and association triggering conditions. Retrieve the original information set of key fields, extract the core content of the textual representation of each key field and the core information associated with the field form, and form a set of core information of key fields; Integrate the set of field relationships with the set of core information of key fields, establish a corresponding mapping between the core information of key fields and their relationships, and form a set of field relationship information; Extract core scene elements from scene association features to form a set of core scene information. The core scene elements include business scene type, layout core features, and business process core nodes. By integrating the core information set of the scenario with the field association information set, a scenario field association framework is constructed to determine the association and positioning of the core information of each key field in the corresponding scenario; Based on the scenario field association framework, the core information of key fields, core information of scenarios, and field association relationships are arranged in a structured manner according to the business logic order to form a preliminary structured result; The preliminary structured results are supplemented with scenario-based information by adding corresponding scenario adaptation descriptions to the core information of each key field. The scenario adaptation descriptions include scenario function positioning and scenario association value. Construct a scenario-based structured template. This scenario-based structured template is divided into information display modules according to the presentation requirements of different business scenarios. Each information display module corresponds to a type of core information. The structured information, including scenario adaptation instructions, is filled into the corresponding information display module of the scenario-based structured template, and integrated to form a scenario-based structured invoice analysis result that includes core field content, scenario association logic, and dynamic association description.

6. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 1, characterized in that, The association logic analysis of the field flow record set is performed, and the association rule mining algorithm is used to extract the associated fields and flow conditions of each key field in different business nodes; Based on the timestamps of field appearances and the sequence of business nodes, the generation trigger events, propagation data formats, and receiving system identifiers of key fields in the business process are determined, forming a logical diagram of field flow associations, including: Based on the business scenario identifier in the field flow record, the field flow record set is divided into multiple field flow record subsets corresponding to different business scenarios; Each field flow record subset is parsed one by one, extracting the key field names, business nodes, associated field names, and flow triggering conditions from each field flow record to form a single record parsing result; Integrate the parsing results of all individual records to construct a field node association matrix. The rows of the field node association matrix represent key fields, the columns represent business nodes, and the matrix elements represent the associated fields and flow conditions of the key field in the corresponding business node. Based on the field node association matrix, an association rule mining algorithm is used to extract the generation trigger events of each key field in each business node. The generation trigger events are business operations or information input behaviors that cause the key field to appear for the first time. Analyze the data format of each key field during the circulation process, determine the encoding method and structure definition of the data format, including the image format of paper carriers, the message format of electronic transmission, and the database record format stored in the system; Determine the receiving system identifier for each key field at each business node to form a list of receiving system identifiers. The receiving system identifier is used to uniquely identify the business processing module, the review node, and the archiving system. Construct a basic framework for the flow and association logic of key fields. In this framework, key fields are the core, and each business node is arranged according to the business process sequence, with associated fields and flow conditions marked. The generated trigger events, propagation data formats, and receiving system identifier lists are added to the basic framework of key field flow association logic to improve the field flow details of each business node; The supplemented key field flow association logic framework is compared with the preset standard business process rules to verify whether the flow order between each business node is consistent with the standard business process rules, and to verify whether the associated fields and conditions are consistent with the rules defined in the standard business process rules. The improved key field flow association logic framework is input into the graph generation algorithm to generate a field flow association logic graph containing details of the entire key field flow process.

7. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 2, characterized in that, Based on the business scenario positioning results, the format association information of the invoice carrier is extracted to form a format association information set, including: Based on the business scenario positioning results, the standard invoice format library corresponding to the business scenario is retrieved. The standard invoice format library contains common invoice overall layout templates, field arrangement templates and area division templates for this business scenario. Calculate the layout similarity between the bill carrier and each template in the standard bill format library, select the template with the highest similarity as a reference, and determine the overall layout of the bill carrier, including symmetric type coding, regional distribution ratio, and the position coordinate range of core fields. The document carrier is divided into multiple functional regions according to the field functional attributes and location distribution. The boundary features and size features of each functional region are extracted. Analyze the field arrangement in each functional area, and record the spacing characteristics, alignment, and hierarchical relationship between fields. The field arrangement includes horizontal arrangement, vertical arrangement, and mixed arrangement. Construct a regional association matrix, where the rows and columns of the regional association matrix represent functional regions, and the matrix elements represent the positional relationship between two functional regions, including adjacency, inclusion, and relative orientation. Extract the core field identifiers of each functional area, identify the key fields that play a dominant role in that functional area, and establish the correspondence between the core fields and the functional areas. Analyze the positional correlation strength between the functional area where the core field is located and other functional areas, and determine the representation method of positional correlation strength based on the distance and functional correlation between functional areas; Integrate the overall layout, field arrangement, region division mode, region association matrix, core field region correspondence and position association strength information to form the initial layout association information; Set a feature importance threshold to filter each element in the initial layout association information, remove elements with importance lower than the feature importance threshold, and retain the core layout association elements; The optimized layout elements are arranged in a logical order to form a layout association information set containing the core association information of the bill layout.

8. The intelligent invoice information extraction and analysis method integrating deep learning according to claim 3, characterized in that, The deep learning-driven text morphology recognition technology analyzes the original pixel information of the text to extract the original features of the strokes, character combinations, and text arrangement, forming the original text morphology information, including: The original pixel information of the text is input into the input layer of the deep learning text morphology recognition model, which includes a feature extraction subnetwork, a morphology parsing subnetwork, and a feature integration subnetwork. The convolutional layers of the feature extraction subnetwork perform multi-scale convolution operations on the original pixel information of the text to extract pixel distribution features at different scales, forming a multi-scale pixel feature set. An edge detection algorithm is used to process multi-scale pixel feature sets, enhancing the response values ​​of stroke edge features, and a filter is used to suppress the response values ​​of background regions, forming an enhanced pixel feature set. In the morphological parsing subnetwork, stroke segmentation is performed on the enhanced pixel feature set to separate the pixel regions corresponding to different strokes of a single character, and the length features, curvature features and direction features of each stroke are extracted to form the original stroke features. The segmented individual character pixel regions are analyzed to extract the connection methods, overlap features and spacing features between characters, forming the original features of character combinations. Expand the analysis scope to the entire physical presentation area of ​​the key field, and extract the arrangement direction, alignment, line spacing and column spacing features of all text within the physical presentation area of ​​the key field to form the original text arrangement features; The original features of strokes, character combinations, and text arrangement are processed by the fully connected layer of the morphological parsing subnetwork to unify the dimensions, so that the original features of strokes, character combinations, and text arrangement are in the same dimensional space. The original features of strokes, character combinations, and text arrangement after unification are input into the feature integration subnetwork, and weighted and fused according to the preset feature weight coefficients to generate comprehensive text morphology features. Calculate the similarity between the comprehensive text morphological features and the original pixel information of the text in the feature space. If the similarity is higher than the preset threshold, it is determined that the comprehensive text morphological features meet the originality requirements, and the comprehensive text morphological features that have passed the originality verification are output as the original text morphological information containing the original features of strokes, the original features of character combinations, and the original features of text arrangement.

9. A smart ticket information extraction and analysis system integrating deep learning, characterized in that, The intelligent invoice information extraction and analysis system integrating deep learning includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the intelligent invoice information extraction and analysis method integrating deep learning as described in any one of claims 1-8.

Images