Agile application construction method and system supporting data fusion and process intelligent auditing

Abstract

The application provides an agile application construction method and system supporting data fusion and process intelligent auditing, and relates to the technical field of intelligent document auditing, wherein the method comprises the following steps: determining a digital base version corresponding to a business document, the digital base version being determined according to an infrared transmission image inside a paper and visible page content; then performing deep semantic analysis on the digital base version, constructing a logical reasoning path for key business elements analyzed to deduce intermediate conclusions; taking the intermediate conclusions as reasoning nodes and adding link identifiers according to logical correlations to construct a logical context map; finally inputting the logical context map into an auditing process instance, generating structured auditing conclusions and natural language auditing conclusions at a current auditing node, encapsulating the two types of conclusions as node auditing opinions to drive the process into a next node until an executable application instance is output. The application improves the accuracy and automation degree of unstructured document auditing.

Description

Technical Field

[0001] This application relates to the field of intelligent document review technology, and in particular to an agile application construction method and system that supports data fusion and intelligent process review. Background Technology

[0002] Agile application building methods, as an application development paradigm that can quickly respond to changes in business needs, have gained widespread attention in business process automation scenarios in fields such as finance, government affairs, and healthcare. By using methods such as visual orchestration and component reuse, they lower the threshold for application building, enabling business personnel to directly participate in the application development process, and have broad application prospects.

[0003] In existing application building technologies, some solutions achieve unified access to business data by integrating multiple data source interfaces, and use a preset rule engine to perform logical validation on form data to generate review results. At the same time, a workflow engine drives the flow of review nodes, and encapsulates the approved business data into executable application instances for output. Other solutions introduce optical character recognition technology to extract text from document images, and build structured data based on the extracted text content, and then combine it with a workflow engine to complete business approval.

[0004] However, the aforementioned methods struggle to accurately extract deep business semantic information from unstructured documents containing multi-source information such as images, handwritten annotations, and seals. Rule engines can only handle pre-defined deterministic logic and cannot address the complex causal and dependency relationships implicit in the document, leading to information omissions or judgment biases in the review results. Therefore, existing technologies suffer from the technical challenge of effectively extracting deep semantic information from unstructured documents and using it for intelligent review. Summary of the Invention

[0005] This application provides an agile application construction method and system that supports data fusion and intelligent process review, in order to solve the problems of low accuracy and low automation in the review of unstructured documents in the prior art.

[0006] To address the aforementioned technical problems, firstly, this application provides a method for building agile applications that supports data fusion and intelligent process review, comprising: Determine the digital base plate corresponding to the business document. The digital base plate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document. The digital base plate is subjected to deep semantic analysis through the thinking chain reasoning mechanism to extract multiple key business elements, and a corresponding logical reasoning path is constructed for each key business element. Each logical reasoning path is used to deduce intermediate conclusions based on the key business elements. The intermediate conclusions output by each logical reasoning path are used as reasoning nodes. Logical link identifiers are added to the reasoning nodes according to the logical connections between them to construct a logical network map. The logical context graph is input into a pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion. The large model is then called to perform semantic understanding on the logical context graph to generate a natural language review conclusion. The structured review conclusion and the natural language review conclusion are encapsulated as node review comments into the review comment field of the current review node, and the review process instance is driven to enter the next review node until all review nodes are completed. Then, an executable application instance is output, which encapsulates complete business data and the node review comments of each review node.

[0007] Optionally, the step of using the intermediate conclusions output by each logical reasoning path as reasoning nodes, adding logical link identifiers to the reasoning nodes according to the logical connections between them, and constructing a logical network graph includes: Collect the intermediate conclusions output by each logical reasoning path, define each intermediate conclusion as a reasoning node, and each reasoning node carries the path identifier of the logical reasoning path to which it belongs; Traverse all inference nodes and identify whether there is a semantic relationship between any two inference nodes. The semantic relationship includes causal relationship, temporal relationship or dependency relationship. When the identification result indicates that there is a semantic association between two inference nodes, a logical link is established between the two inference nodes, a logical link identifier is assigned to the logical link, and the association type is marked on the logical link; A directed graph structure is constructed using each inference node as a graph vertex and each logical link as a graph edge. The directed graph structure is hierarchically divided according to the dependencies between each inference node. Inference nodes without preceding dependent nodes are assigned to the initial level, and inference nodes with the same dependency depth are assigned to the same level. Inference nodes and logical links are combined into a logical network graph according to the hierarchical order.

[0008] Optionally, the step of hierarchically dividing the directed graph structure according to the dependencies between inference nodes, assigning inference nodes without preceding dependent nodes to the initial level, assigning inference nodes with the same dependency depth to the same level, and combining each inference node and each logical link into a logical network graph according to the hierarchical order includes: Based on the dependencies represented by the logical links between each inference node, and combined with the semantic association strength between key business elements in the business document, all inference nodes in the directed graph structure are hierarchically divided. Inference nodes without any preceding dependent nodes are assigned to the initial level, and inference nodes that depend on the inference nodes in the initial level and have no other preceding dependent nodes are assigned to the next level. The hierarchical division is repeated until all inference nodes are assigned to the corresponding level. Inference nodes assigned to the same level are organized into the same level group, and the level groups are arranged in order from low to high level. The hierarchical structure is formed by combining the processing order of the business elements corresponding to the inference nodes in each level group in the document review process. Based on the logical links between the hierarchical groups in the hierarchical structure, the hierarchical structure and the logical links are combined into a logical network diagram. The logical network diagram is used by the current audit node in the audit process instance to read the reasoning nodes in each level group in order from low to high level to perform audit judgment.

[0009] Secondly, this application provides an agile application building system that supports data fusion and intelligent process review, including: The determination module is used to determine the digital base plate corresponding to the business document. The digital base plate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document. The parsing module is used to perform deep semantic parsing on the digital base plate through the thinking chain reasoning mechanism, parsing out multiple key business elements, and constructing a corresponding logical reasoning path for each key business element. Each logical reasoning path is used to derive intermediate conclusions based on the key business elements. The module is used to take the intermediate conclusions output by each logical reasoning path as reasoning nodes, add logical link identifiers to the reasoning nodes according to the logical relationship between each reasoning node, and construct a logical network map. The matching module is used to input the logical context graph into a pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion. It also calls the large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion. The output module is used to encapsulate the structured review conclusion and the natural language review conclusion as node review opinions into the review opinion field of the current review node, and drive the review process instance to the next review node until all review nodes are completed, and then output an executable application instance. The executable application instance encapsulates complete business data and the node review opinions of each review node.

[0010] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, used to execute the computer program to implement the steps of the agile application building method supporting data fusion and intelligent process auditing as described in the first aspect above.

[0011] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the agile application construction method supporting data fusion and intelligent process review as described in the first aspect above.

[0012] The technical solution provided in this application has the following beneficial effects: This application first determines the digital base of a business document by fusing the internal infrared transmission image of the paper with the visible content of the page, thereby fully preserving the explicit and implicit features of the document; then, it performs deep semantic analysis on the digital base through a thought chain reasoning mechanism, constructs logical reasoning paths for the key business elements extracted to derive intermediate conclusions, and realizes the structured extraction of deep semantic information of the document; then, it uses the intermediate conclusions output by each reasoning path as reasoning nodes and adds link identifiers according to the logical connections between nodes to construct a logical network map, organizing the scattered reasoning information into a unified representation with an associated structure; Next, the logical context map is input into the review process instance. At the current review node, the rule engine and the large model are called simultaneously to generate structured review conclusions and natural language review conclusions respectively. The review results are complemented by parallel processing of rule and semantic paths. Finally, the two types of conclusions are encapsulated as node review opinions to drive the process flow until an executable application instance containing complete business data and review opinions of each node is output, ensuring that the comprehensiveness of the review decision is fully preserved in the final application instance.

[0013] Furthermore, this application collects the intermediate conclusions output by each logical reasoning path and defines them as reasoning nodes. It traverses all reasoning nodes to identify causal, temporal, or dependency relationships between nodes. Logical links are established between nodes with semantic associations, and the association type is marked. A directed graph structure is constructed with each reasoning node as a vertex and each logical link as an edge. The directed graph is then hierarchically divided according to the dependency relationships between nodes. Reasoning nodes without preceding dependent nodes are assigned to the initial level, and reasoning nodes with the same dependency depth are assigned to the same level. Each reasoning node and each logical link are combined into a logical network graph according to the hierarchical order.

[0014] Furthermore, by organizing the planar reasoning conclusions into a graph structure with clear hierarchical relationships and dependency order through the above methods, the complex causal and dependency relationships implicit in the document can be explicitly expressed, providing a structured semantic input that can be parsed layer by layer for subsequent review nodes.

[0015] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an agile application construction method supporting data fusion and intelligent process review, provided in an embodiment of this application; Figure 2 A schematic diagram illustrating a specific implementation of an agile application construction method supporting data fusion and intelligent process review, provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an agile application building system that supports data fusion and intelligent process review, provided as an embodiment of this application. Detailed Implementation

[0018] In existing agile application building techniques, for unstructured business documents containing multi-source information such as images, handwritten annotations, stamps, and page layouts, conventional methods typically rely solely on optical character recognition to extract text content, followed by logical verification through a pre-defined rule engine. This approach has several drawbacks when processing documents. First, it is difficult to integrate and utilize the implicit features within the document paper with the visible page content, resulting in compromised document information integrity. Second, because the implicit causal and dependency relationships within the document cannot be effectively identified by the rule engine, the review results are prone to information omissions or judgment biases, affecting the sufficiency of data fusion and the accuracy of review decisions during application building.

[0019] To address the aforementioned issues, this application proposes an agile application construction method that supports data fusion and intelligent process review. Its core lies in first determining the digital template of a business document by fusing the internal infrared transmission image of the paper with the visible content, thus fully preserving both explicit and implicit features of the document. Then, a thought chain reasoning mechanism is used to perform deep semantic analysis on the digital template, constructing logical reasoning paths for the extracted key business elements to derive intermediate conclusions. The intermediate conclusions of each path are organized into a network of reasoning nodes with logical link identifiers, forming a logical network graph, transforming the complex causal relationships implicit in the document into a structured representation. Finally, in a review process instance, a rule engine and a large model are simultaneously invoked to perform rule matching and semantic understanding on the logical network graph. The fusion of these two types of conclusions serves as the node review opinions driving the process flow.

[0020] This method, through the organic combination of digital template fusion, thought chain analysis, and dual-engine collaborative review, enables the effective extraction of deep semantic information from documents and its use in review decisions. It fundamentally solves the problems of insufficient fusion of unstructured document information and inadequate review accuracy in existing technologies, and improves the integrity of data fusion and the level of intelligence in process review in agile application building scenarios.

[0021] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The core of this application is to provide an agile application construction method that supports data fusion and intelligent process review. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: Step 101: Determine the digital base plate corresponding to the business document. The digital base plate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document.

[0023] Among them, business documents refer to unstructured documents containing various information carriers such as text, tables, seals, handwritten annotations, and page layout. These business documents are formed during the business process and carry core business information. Digital templates refer to a complete digital expression that integrates the implicit features of the document paper with the visible information on the surface. These digital templates are used for subsequent semantic analysis and intelligent review. Infrared images are internal images captured by infrared light penetrating the paper. These infrared images contain watermark information, anti-counterfeiting fiber distribution, and traces of alteration overlays inside the paper. Visible content refers to the visual information on the surface of a document that can be recognized by the naked eye or conventional scanning equipment. This visible content includes text information in text areas, cell data in table areas, stamp patterns in signature areas, and handwritten notes in handwritten annotation areas.

[0024] In this embodiment, step 101 includes the following process: Step 1011: In response to the request to build unstructured business documents, multi-source document data generated during the business process is collected through the document digitization all-in-one machine, and infrared transmission scanning is started through the document digitization all-in-one machine to capture infrared transmission images inside the paper.

[0025] In step 1011, the document digitizing all-in-one machine is an integrated hardware device with double-sided scanning and infrared transmission functions. The document digitizing all-in-one machine can start the infrared transmission scanning mode before performing conventional optical character recognition. The multi-source document data includes scanned images, electronic document layout, handwritten signature marks, and document metadata.

[0026] In this embodiment, a construction request for an unstructured business document is first received, which triggers the document digitization all-in-one machine to start working. While collecting multi-source document data generated during the business process, the document digitization all-in-one machine starts an infrared transmission scanning mode, using the ability of infrared light to penetrate paper to capture infrared transmission images inside the paper. These infrared transmission images can reveal watermark information, anti-counterfeiting fiber distribution, and traces of tampering layers hidden inside the paper, thereby obtaining the implicit feature data of the document.

[0027] It should be noted that step 1011 above is not an essential operation for implementing this application. For routine business documents that already have standard electronic documents and do not involve detecting implicit features inside the paper, the digital base plate can be determined directly based on the visible page content corresponding to the document data in the business document, without needing to perform the infrared transmission acquisition process in step 1011.

[0028] Step 1012: Perform pixel-level layout parsing on the multi-source document data to separate text areas, table areas, signature areas, and handwritten annotation areas to form a set of document constituent units with position labels.

[0029] In step 1012, the text area refers to the page area that carries text information, the table area refers to the page area that carries row and column structure data, the signature area refers to the page area that carries a seal or signature image, and the handwritten annotation area refers to the page area that carries handwritten handwriting. The document composition unit set refers to the set of various regional units and their spatial location information after the document page is divided according to content type. Each document composition unit includes a regional type identifier and the coordinate position of the region on the page.

[0030] In this embodiment, after obtaining multi-source document data, a pixel-level layout parsing operation is performed on the multi-source document data. This parsing process traverses every pixel in the document page, identifying different types of layout regions based on the color features, edge features, and texture features of the pixels. Specifically, the document content is separated into text regions carrying text information, table regions carrying structured data, signature regions carrying seal images, and handwritten annotation regions carrying handwritten handwriting. After completing the region separation, the position coordinates of each region on the page are marked, forming a set of document constituent units containing region type and position information. This set transforms the document's layout structure into structured data that can be processed in subsequent steps.

[0031] Step 1013: Overlay and register the infrared transmission image with the visible page content reconstructed based on the document constituent unit set to generate a digital base plate.

[0032] In this embodiment, firstly, based on the position coordinates and region type of each region in the document constituent unit set, the text region, table region, signature region, and handwritten annotation region are reorganized according to their positions in the original document to reconstruct the visible page content of the document; then, the infrared transmission image captured in step 1011 is overlaid and registered with the visible page content. During the registration process, the correspondence between each latent feature in the infrared transmission image and each region in the visible page content is determined based on the position information in the document constituent unit set. The watermark information, anti-counterfeiting fiber distribution, and traces of alteration covering layer inside the paper are spatially aligned and pixel-level fused with the visible text, tables, signatures, and annotations to finally generate a digital base plate that simultaneously contains explicit visual features and implicit anti-counterfeiting features.

[0033] Through the steps described above, this application integrates multi-source heterogeneous data of documents into a unified digital base, providing a complete set of basic data for subsequent semantic analysis.

[0034] Step 102: Perform deep semantic analysis on the digital base plate through the thinking chain reasoning mechanism to extract multiple key business elements, and construct a corresponding logical reasoning path for each key business element. Each logical reasoning path is used to derive intermediate conclusions based on the key business elements.

[0035] Among them, the thought chain reasoning mechanism refers to a reasoning method that decomposes complex reasoning tasks into multiple consecutive intermediate reasoning steps and gradually derives the final conclusion through the logical progression between the steps; key business elements refer to the core information units with business meaning extracted from the document content, including semantically valuable information such as the subject name, date value, amount value, and item identification in the document; logical reasoning path refers to a chain structure composed of multiple reasoning steps linked together according to causal relationships, which is used to simulate the thinking process of deriving unknown conclusions from known information; intermediate conclusion refers to the stage result output by a certain reasoning step in the logical reasoning path, which serves as the input information for subsequent reasoning steps.

[0036] In this embodiment, step 102 includes the following process: Step 1021: Traverse each layout area in the digital base plate, identify the business information carried in each layout area, and extract multiple key business elements from the business information.

[0037] In step 1021, the layout area refers to an independent block in the digital base plate divided according to the content type. The layout area includes text area, table area, signature area and handwritten annotation area; business information refers to text content, numerical data or graphic symbols with semantic meaning identified from the layout area.

[0038] In this embodiment, the process first traverses each layout area in the digital base, and calls the corresponding recognition strategy for each layout area based on its area type. For text areas, semantic analysis technology is used to identify noun entities and numerical descriptions. For table areas, structural parsing technology is used to extract the numerical values ​​and corresponding relationships in the cells. For signature areas, image recognition technology is used to determine the text information and pattern features in the seal. For handwritten annotation areas, handwriting recognition technology is used to extract the annotation content and the corresponding associated position. All identified business-related information is integrated to extract multiple key business elements. Each key business element records its element name, element value, and source position in the document.

[0039] In practical applications, a digital copy of an equipment procurement document contains text areas, table areas, and signature / seal areas. After traversing the text areas, the system identifies the purchaser as Party A, the supplier as Party B, the total equipment price as 5 million yuan, and the delivery date as March 1, 2025. After traversing the table areas, the system extracts an equipment list including 10 units of equipment C and 5 units of equipment D. After traversing the signature / seal areas, the system identifies the official seals of Party A and Party B. From the above business information, five key business elements are extracted: the purchaser as Party A, the supplier as Party B, the total equipment price as 5 million yuan, the delivery date as March 1, 2025, and the equipment list including equipment C and D.

[0040] Step 1022: Create a logical reasoning path for each key business element, and configure a start point and an end point in the logical reasoning path. The start point is the key business element itself, and the end point is the intermediate conclusion to be derived.

[0041] In step 1022, the starting point refers to the entry node of the logical reasoning path, which carries the original information content of the key business elements; the ending point refers to the exit node of the logical reasoning path, which corresponds to the intermediate conclusion that needs to be derived.

[0042] In this embodiment of the application, an independent logical reasoning path is created for each key business element extracted in step 1021. In each logical reasoning path, the key business element itself is configured as the starting point, and the intermediate conclusion to be derived is determined as the termination point based on the data type and semantic attributes of the key business element. Taking the key business element of total equipment price as an example, the key business element is used as the starting point, and the conclusion to be derived, "whether the total equipment price exceeds the procurement budget threshold," is used as the termination point. Taking the key business element of delivery date as an example, the key business element is used as the starting point, and the conclusion to be derived, "whether the delivery date is within a reasonable construction period," is used as the termination point.

[0043] In practical applications, a logical reasoning path is created for the key business element of the purchasing party (Party A), with Party A as the starting point and the intermediate conclusion to be derived, "whether the purchasing party is a registered entity," as the ending point. Similarly, a logical reasoning path is created for the key business element of the total equipment price of 5 million yuan, with the total equipment price as the starting point and the intermediate conclusion to be derived, "whether the total equipment price exceeds a preset threshold," as the ending point.

[0044] Step 1023: Construct at least one intermediate node between the starting point and the ending point. Each intermediate node corresponds to a derivation step. A pointing identifier is set between adjacent nodes. The pointing identifier is used to represent the derivation direction from the earlier node to the later node.

[0045] In step 1023, an intermediate node refers to a reasoning link located between the starting point and the ending point in a logical reasoning path. This intermediate node carries the derivation logic that converts input information into output information. The pointing marker refers to a directed arrow connecting two adjacent nodes. This pointing marker clearly defines the direction of information transmission and derivation.

[0046] In this embodiment, for each logical reasoning path, at least one intermediate node is constructed between the starting point and the ending point, and each intermediate node corresponds to a specific derivation step. When constructing intermediate nodes, the complete derivation process is decomposed into multiple consecutive derivation steps according to the reasoning chain required to derive from the starting point to the ending point, and each derivation step corresponds to an intermediate node. A pointing identifier is set between adjacent nodes, which points from the earlier node to the later node to clarify the information flow and the order of derivation. Taking the reasoning path of whether the total price of the equipment exceeds a preset amount threshold as an example, a first intermediate node is first constructed to determine whether the total price of the equipment exceeds 2 million yuan, and then a second intermediate node is constructed to determine whether a secondary confirmation process needs to be executed after the price exceeds 2 million yuan.

[0047] In practical applications, for a reasoning path with a total equipment price of 5 million yuan, two intermediate nodes are constructed between the starting point (equipment price) and the ending point (whether the total equipment price exceeds a preset threshold). The first intermediate node is for "determining whether the total equipment price exceeds 2 million yuan", and the second intermediate node is for "determining whether a secondary confirmation process is required after exceeding 2 million yuan". A pointer is set between the starting point and the first intermediate node, pointing from the starting point to the first intermediate node; a pointer is set between the first intermediate node and the second intermediate node, pointing from the first intermediate node to the second intermediate node; and a pointer is set between the second intermediate node and the ending point, pointing from the second intermediate node to the ending point.

[0048] Step 1024: Use the data content of the key business elements as the input data of the starting point, activate each intermediate node in sequence according to the pointing identifier, and after each intermediate node is activated, perform a transformation operation on the input data according to the derivation logic carried by its own node to generate intermediate data to be passed to the next node, until the termination point is activated and the intermediate conclusion is output.

[0049] In step 1024, data content refers to the specific numerical or textual information carried by key business elements, and activation refers to triggering the intermediate node to execute the derivation logic it carries; derivation logic refers to the judgment rules or calculation rules built into the intermediate node, which is used to convert input data into output data; intermediate data refers to the stage results generated after the intermediate node executes the derivation logic, which serves as the input information for the next node.

[0050] In this embodiment, the data content of key business elements is first used as the input data of the starting point, and then each intermediate node is activated sequentially according to the direction of the pointing marker. When an intermediate node is activated, it receives the input data passed from the previous node, performs a transformation operation on the input data according to the deduction logic carried by its own node, generates intermediate data, and passes it to the next node. The intermediate data generated by the previous intermediate node is used as the input data of the next intermediate node, and so on until the termination point is activated. After the termination point is activated, the intermediate conclusion is output according to the input data passed from the last intermediate node. This intermediate conclusion is the final deduction result of the logical reasoning path.

[0051] In practical applications, the data content of a total equipment price of 5 million yuan is used as the starting point input data. According to the pointer, the first intermediate node "determines whether the total equipment price exceeds 2 million yuan" is activated first. This intermediate node receives 5 million yuan as input data, performs a numerical comparison logic to determine whether it is greater than 2 million yuan, generates intermediate data "exceeds 2 million yuan" and passes it to the second intermediate node. The second intermediate node "determines whether a secondary confirmation process is required after exceeding 2 million yuan" is activated. This intermediate node receives the intermediate data "exceeds 2 million yuan", performs a judgment logic to determine whether a secondary confirmation is required, generates intermediate data "a secondary confirmation process is required" and passes it to the termination point. The termination point "determines whether the total equipment price exceeds a preset amount threshold" is activated. This termination point receives the intermediate data "a secondary confirmation process is required" and outputs the intermediate conclusion "the total equipment price exceeds the preset amount threshold and a secondary confirmation is required".

[0052] Through the above steps, this application transforms the key business elements in the document into an executable reasoning chain structure, realizing the structured extraction and recursive derivation of the deep semantic information of the document, and providing structured semantic input for subsequent intelligent review.

[0053] Step 103: Take the intermediate conclusions output by each logical reasoning path as reasoning nodes, add logical link identifiers to the reasoning nodes according to the logical connections between them, and construct a logical network map.

[0054] Among them, inference nodes refer to the vertices of the graph structure used to represent intermediate conclusions. Each inference node corresponds to an intermediate conclusion and records the semantic content of the intermediate conclusion. Logical link identifiers refer to the unique identification codes assigned to each logical link. These logical link identifiers are used to distinguish different logical relationships. Logical network graphs refer to the directed graph structure composed of inference nodes and logical links. These logical network graphs are used to describe the logical relationship network between intermediate conclusions in the document.

[0055] In this embodiment, step 103 includes the following process, such as... Figure 2 As shown: Step 1031: Collect the intermediate conclusions output by each logical reasoning path, define each intermediate conclusion as a reasoning node, and each reasoning node carries the path identifier of the logical reasoning path to which it belongs.

[0056] In step 1031, the path identifier is a unique code used to distinguish different logical reasoning paths. The path identifier records the source path information of the intermediate conclusion.

[0057] In this embodiment of the application, the intermediate conclusions output by all logical reasoning paths in step 102 are first collected, and each collected intermediate conclusion is defined as a reasoning node. When defining a reasoning node, the path identifier of the logical reasoning path to which the intermediate conclusion belongs is associated with the reasoning node and stored, so that each reasoning node carries the path identifier of the logical reasoning path to which it belongs. Through the path identifier, it is possible to trace which logical reasoning path each reasoning node was derived from.

[0058] In practical application, step 102 constructs five logical reasoning paths for five key business elements: the purchaser (Party A), the supplier (Party B), the total equipment price of 5 million yuan, the delivery date of March 1, 2025, and the equipment list including equipment C and D. Each of these five logical reasoning paths outputs a corresponding intermediate conclusion. The intermediate conclusion "The total equipment price exceeds the preset amount threshold and requires secondary confirmation" output by the logical reasoning path corresponding to the total equipment price is defined as a reasoning node, and the path identifier of the logical reasoning path for the total equipment price is associated with this reasoning node. Similarly, the intermediate conclusion "The purchaser is a registered entity" output by the logical reasoning path corresponding to the purchaser is defined as another reasoning node, and the path identifier of the logical reasoning path for the purchaser is associated with this node.

[0059] Step 1032: Traverse all inference nodes and identify whether there is a semantic relationship between any two inference nodes. The semantic relationship includes causal relationship, temporal relationship or dependency relationship.

[0060] In step 1032, semantic association refers to the logical connection between the intermediate conclusions represented by two inference nodes. The semantic association includes the causal relationship that one conclusion leads to another conclusion, the temporal relationship that one conclusion occurs before another conclusion, and the dependency relationship that the validity of one conclusion depends on the validity of another conclusion.

[0061] In this embodiment of the application, all inference nodes generated in step 1031 are traversed. For any two inference nodes, the semantic content of the intermediate conclusion carried by each inference node is extracted. The semantic content of the two intermediate conclusions is compared and analyzed to determine whether there is a causal relationship, a temporal relationship or a dependency relationship between them. If any of the above relationships exist, it is determined that there is a semantic association between the two inference nodes. If no such relationship exists, it is determined that there is no semantic association between the two inference nodes.

[0062] Among them, causal relationships are identified by determining whether the intermediate conclusion of the previous inference node serves as the direct cause of the subsequent inference node. For example, if the equipment model and quantity included in the equipment list inference node are equal to the amount in the total equipment price inference node after calculation, then it is determined to be a causal relationship. Temporal relationships are identified by comparing the time attribute values ​​carried by the inference nodes. For example, if the date value of the delivery date inference node is earlier than the date value of the production cycle start date inference node, then it is determined to be a temporal relationship. Dependency relationships are identified by analyzing whether the inference logic of the subsequent inference node uses the intermediate conclusion of the previous inference node as a necessary input. For example, if the judgment logic of the purchaser qualification review inference node explicitly references the filing status of the purchaser's main information inference node as a judgment basis, then it is determined to be a dependency relationship.

[0063] In practical applications, semantic association identification is performed between the inference node corresponding to the total equipment price and the inference node corresponding to the purchaser. The logical relationship between the intermediate conclusion corresponding to the total equipment price ("The total equipment price exceeds the preset threshold and requires secondary confirmation") and the intermediate conclusion corresponding to the purchaser ("The purchaser is a registered entity") is analyzed. It is found that there is no causal, temporal, or dependency relationship between the two, therefore it is determined that there is no semantic association between these two inference nodes. Semantic association identification is also performed between the inference node corresponding to the total equipment price and the inference node corresponding to the delivery date. The logical relationship between the intermediate conclusion corresponding to the total equipment price and the intermediate conclusion corresponding to the delivery date ("The delivery date is within a reasonable construction period") is analyzed. It is found that there is no direct logical connection between the total equipment price and the delivery date, therefore it is determined that there is no semantic association between these two inference nodes.

[0064] Step 1033: When the identification result indicates that there is a semantic association between the two inference nodes, establish a logical link between the two inference nodes, assign a logical link identifier to the logical link, and mark the association type on the logical link.

[0065] In step 1033, a logical link refers to a directed edge connecting two inference nodes, which is used to characterize the logical relationship between two intermediate conclusions; the association type refers to the specific category of semantic association, which includes causal relationship, temporal relationship and dependency relationship.

[0066] In this embodiment of the application, when the identification result of step 1032 is that there is a semantic association between two inference nodes, a logical link is established between the two inference nodes; a unique logical link identifier is assigned to the logical link for subsequent identification and reference; according to the specific category of the semantic association identified in step 1032, the corresponding association type is marked on the logical link to clarify whether the logical link represents a causal relationship, a temporal relationship or a dependency relationship.

[0067] In practical applications, semantic association recognition is performed between the inference node corresponding to the equipment list and the inference node corresponding to the total equipment price. The logical relationship between the intermediate conclusion "the equipment list contains equipment C and D" corresponding to the equipment list and the intermediate conclusion "the total equipment price exceeds the preset amount threshold and requires secondary confirmation" corresponding to the total equipment price is analyzed. It is found that the configuration of equipment C and D in the equipment list directly causes the total equipment price to exceed the preset amount threshold. Therefore, it is determined that there is a causal relationship between the two inference nodes. A logical link is established between the inference node corresponding to the equipment list and the inference node corresponding to the total equipment price. A logical link identifier L1 is assigned to this logical link, and the association type is marked as causal relationship on this logical link.

[0068] Step 1034: Construct a directed graph structure using each inference node as a graph vertex and each logical link as a graph edge.

[0069] In this embodiment of the application, all inference nodes defined in step 1031 are used as the vertex set of the directed graph structure, and all logical links established in step 1033 are used as the edge set of the directed graph structure. The direction of the edge is determined by the pointing relationship between the inference nodes, and a complete directed graph structure is constructed. In this directed graph structure, each vertex corresponds to an intermediate conclusion, each directed edge corresponds to a logical association relationship, and the direction of the edge is determined by the logical pointing between the inference nodes.

[0070] In practical applications, the inference nodes corresponding to the equipment list, the total equipment price, the purchaser, and the delivery date are respectively used as the four vertices of the directed graph structure. The logical link L1 between the inference node corresponding to the equipment list and the inference node corresponding to the total equipment price is used as an edge of the directed graph structure. The direction of this edge is from the inference node corresponding to the equipment list to the inference node corresponding to the total equipment price, representing the causal relationship between the equipment list and the total equipment price. Other inference nodes are not added as edges because there is no logical link between them.

[0071] Step 1035: Divide the directed graph structure into levels according to the dependencies between each inference node. Inference nodes without preceding dependent nodes are assigned to the starting level, and inference nodes with the same dependency depth are assigned to the same level. Combine each inference node and each logical link into a logical network graph according to the hierarchical order.

[0072] Among them, the preceding dependent node refers to other reasoning nodes in the directed graph structure that have logical links pointing to the current node, that is, the prior conclusions on which the reasoning conclusion of the current node depends; the dependency depth refers to the number of logical links traversed from the starting level node to the current node. This dependency depth is used to measure the position level of the node in the reasoning chain.

[0073] Step 1035 may specifically include the following steps: A1: Based on the dependencies represented by the logical links between each inference node, and combined with the semantic association strength between key business elements in the business document, all inference nodes in the directed graph structure are hierarchically divided. Inference nodes without any preceding dependent nodes are assigned to the initial level, and inference nodes that depend on inference nodes in the initial level and have no other preceding dependent nodes are assigned to the next level. The hierarchical division is repeated until all inference nodes are assigned to the corresponding level.

[0074] In step A1, semantic association strength refers to the degree of logical connection between two key business elements. This semantic association strength is quantified by analyzing the co-occurrence frequency, positional proximity, and semantic similarity of the key business elements in the document. The higher the co-occurrence frequency, the closer the position, and the greater the semantic similarity, the higher the semantic association strength between the inference nodes corresponding to the two key business elements. When dividing the hierarchy, inference nodes with high semantic association strength are preferentially assigned to adjacent levels, so that inference nodes with close logical connections are closer in position in the hierarchical structure.

[0075] In this embodiment, firstly, all semantic associations and their association types identified in step 1032 are obtained. For inference node pairs with semantic associations, the semantic association strength value between the inference node pairs is calculated based on the co-occurrence frequency of the key business elements corresponding to the inference nodes in the document, their positional distance on the page, and the semantic similarity of the intermediate conclusion text. Then, all inference nodes in the directed graph structure are traversed, and the number of preceding dependent nodes for each inference node is counted. Inference nodes with zero preceding dependent nodes are assigned to the initial level. For inference nodes with preceding dependent nodes, the priority of hierarchical division is determined based on the semantic association strength between the node and its preceding dependent nodes. Inference nodes with high semantic association strength are preferentially assigned to the next level below the level of the immediate preceding dependent node. The inference nodes that have been divided into levels and their associated logical links are removed from the directed graph structure, the number of preceding dependent nodes of the remaining inference nodes is updated, and the above division operation is repeated until all inference nodes are assigned to the corresponding level.

[0076] In practical applications, there is a causal relationship between the inference node corresponding to the equipment list and the inference node corresponding to the total equipment price. The key business elements corresponding to these two inference nodes, the equipment list and the total contract price, are located in adjacent paragraphs in the document and co-occur 3 times, resulting in a semantic association strength of 0.85. There is no semantic association between the inference node corresponding to the purchaser and the inference node corresponding to the delivery date, resulting in a semantic association strength of 0. The inference nodes corresponding to the equipment list, the purchaser, and the delivery date have no preceding dependent nodes, so these three inference nodes are assigned to the initial level. The inference node corresponding to the total equipment price depends on the inference node corresponding to the equipment list, and the semantic association strength between them (0.85) is higher than the preset threshold of 0.6, so the inference node corresponding to the total equipment price is assigned to the second level, which is adjacent to the initial level.

[0077] A2: Organize the inference nodes assigned to the same level into the same level group, and arrange the level groups in order from low to high level. Combine the processing order of the business elements corresponding to the inference nodes in each level group in the document review process to form a hierarchical structure.

[0078] In step A2, the hierarchical structure refers to the structured organization formed by vertically layering inference nodes according to the depth of their dependencies and horizontally arranging them according to the order of business logic within the same layer. In this hierarchical structure, each inference node has a unique level affiliation and a definite arrangement position. The vertical layering reflects the progressive dependency relationship between inference conclusions, with inference nodes at lower levels providing the reasoning basis for inference nodes at higher levels. The horizontal arrangement reflects the logical order of inference conclusions at the same dependency depth or the natural order of their appearance in the document.

[0079] In this embodiment, inference nodes classified to the same level are organized into a hierarchical group, with each hierarchical group corresponding to a hierarchical number, which increases sequentially from the initial level. The hierarchical groups are arranged in ascending order of hierarchical number to form a vertical hierarchical sequence. Within each hierarchical group, the key business elements corresponding to the inference nodes in that hierarchical group are horizontally sorted according to the order of appearance of the key business elements in the document or the order of processing logic, ultimately forming a hierarchical structure that includes vertical hierarchical relationships and horizontal arrangement order.

[0080] In practical applications, the three inference nodes—equipment list, purchaser, and delivery date—in the initial level are organized into an initial level group, and the equipment total price inference node in the second level is organized into a second level group. The initial level group and the second level group are arranged in ascending order of level. Within the initial level group, the purchaser, delivery date, and equipment list are arranged in the order of appearance of key business elements in the document. Within the second level group, there is only one inference node, so there is no need for horizontal sorting, forming a hierarchical structure.

[0081] A3: Based on the logical links between the hierarchical groups in the hierarchical structure, the hierarchical structure and the logical links are combined into a logical network diagram. The logical network diagram is used by the current audit node in the audit process instance to read the reasoning nodes in each level group in order from low to high level to perform audit judgment.

[0082] In this embodiment, the hierarchical structure formed in step A2 is fused and combined with the logical links established in step 1033. The hierarchical affiliation and horizontal arrangement order of each inference node in the hierarchical structure are retained, as well as the logical links and directions between each inference node. In the logical network map formed after fusion, the inference nodes are vertically distributed in order from low to high level, and the inference nodes of the same level are arranged horizontally. The logical links are connected between inference nodes with dependencies. This logical network map is used by the current review node in the review process instance to read the inference nodes in each level group in order from low to high level and to perform review and judgment layer by layer.

[0083] In practical applications, the three inference nodes (purchaser, delivery date, and equipment list) in the initial hierarchical group and the equipment total price inference node in the second hierarchical group are arranged in hierarchical order. The logical link L1 between the equipment list inference node and the equipment total price inference node is retained, forming a logical network diagram. In this logical network diagram, the initial level contains the three inference nodes (purchaser, delivery date, and equipment list), and the second level contains the equipment total price inference node. The logical link points from the equipment list inference node to the equipment total price inference node. The current audit node reads the three inference nodes in the initial hierarchical group first, and then reads the equipment total price inference node in the second hierarchical group to perform audit judgment, according to the hierarchical order from low to high.

[0084] Through the steps described above, this application organizes the scattered intermediate conclusions into a logical network diagram with clear hierarchical relationships and dependency order, providing structured semantic input that can be parsed layer by layer for subsequent review nodes.

[0085] Step 104: Input the logical context graph into the pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion. It also calls the large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion.

[0086] Among them, the review process instance refers to the specific execution object generated at runtime by the review workflow pre-arranged according to business needs. The review process instance contains multiple review nodes executed in sequence; the rule engine refers to the logic processing component used to execute the preset judgment rules. The rule engine matches the input information with the judgment conditions and outputs structured results; the big model refers to the deep learning model pre-trained on a large-scale corpus. The big model can understand the semantic content of the input text and generate text output that conforms to the expression habits of natural language.

[0087] This application does not impose specific limitations on the model type, internal structure design, parameter design, training process, etc. of large models, and corresponding settings can be made according to the actual situation.

[0088] In this embodiment, step 104 includes the following process: Step 1041: Obtain a pre-built audit process instance, which contains multiple audit nodes, each of which is configured with a node identifier and node execution order.

[0089] In step 1041, the node identifier refers to the identification code used to uniquely distinguish different review nodes; the node execution order refers to the order in which each review node is arranged in the review process instance, and the node execution order determines the flow path of the review process instance.

[0090] In this embodiment of the application, an audit process instance pre-arranged according to business needs is first obtained from the process engine. The audit process instance contains multiple audit nodes. Each audit node is assigned a unique node identifier to distinguish different nodes. At the same time, each audit node is configured with a node execution order to clarify the position of the node in the overall process. The obtained audit process instance records the complete flow path from the starting audit node to the ending audit node.

[0091] Step 1042: When the review process instance is transferred to the current review node, read the rule configuration information attached to the current review node. The rule configuration information contains at least one judgment rule, and each judgment rule consists of a judgment condition and a conclusion template.

[0092] In step 1042, the judgment condition refers to the logical expression used to determine whether the input information meets specific requirements; the conclusion template refers to the preset text frame used to generate a structured conclusion when the judgment condition is met, and the conclusion template contains placeholders that can be instantiated.

[0093] In this embodiment of the application, when the audit process instance flows to a certain audit node according to the node execution order, the audit node is taken as the current audit node, and the rule configuration information pre-attached to the current audit node is read. The rule configuration information contains at least one judgment rule, and each judgment rule is composed of a judgment condition and a conclusion template. The judgment condition is used to make logical judgments on the input information, and the conclusion template is used to generate the corresponding structured conclusion when the judgment condition is true.

[0094] Step 1043: Traverse all nodes and all links in the logical network diagram, and use the node information carried by each node and the link information carried by each link as the matching information.

[0095] In step 1043, node information refers to the semantic content of the intermediate conclusion carried by the reasoning node, and link information refers to the association type and connection relationship represented by the logical link.

[0096] In this embodiment of the application, the logical network map constructed in step 103 is obtained, and all inference nodes and all logical links in the logical network map are traversed. For each inference node, the node information carried by the inference node, i.e., the semantic content of the intermediate conclusion, is extracted. For each logical link, the link information carried by the logical link, i.e., the association type and the identifiers of the two inference nodes connected, is extracted. All extracted node information and all link information are used together as the information to be matched.

[0097] Step 1044: Compare the information to be matched with the judgment conditions in the judgment rule. When the information to be matched meets the judgment conditions, instantiate the conclusion template corresponding to the judgment rule into a structured conclusion.

[0098] In this embodiment of the application, the information to be matched obtained in step 1043 is compared one by one with the judgment conditions in the judgment rules read in step 1042. During the comparison process, the node information and link information in the information to be matched are matched with the preset matching mode in the judgment conditions. When a certain piece of information in the information to be matched meets the matching rule defined in the judgment conditions, it is determined that the information to be matched satisfies the judgment condition. At this time, the conclusion template corresponding to the judgment rule to which the judgment condition belongs is obtained, and the corresponding node information or link information in the information to be matched is filled into the placeholder position in the conclusion template to complete the instantiation of the conclusion template and generate a structured conclusion.

[0099] Step 1045: Read the model call parameters attached to the current review node. The model call parameters include the input template, model identifier, and output format.

[0100] In step 1045, the input template refers to a preset text framework used to organize the information to be matched into a format that the large model can understand; the model identifier refers to the identification code used to uniquely identify the large model to be called; and the output format refers to the data structure specifications that the large model must follow to return the results.

[0101] In this embodiment of the application, the model calling parameters attached to the current review node are read at the same time or after the generation of the structured conclusion. The model calling parameters include an input template, a model identifier, and an output format. The input template specifies how the information to be matched is organized into a text form suitable for input to a large model. The model identifier indicates the specific instance of the large model to be called. The output format specifies the data structure that the results returned by the large model should present.

[0102] Step 1046: Convert the information to be matched into a text sequence according to the connection relationship between nodes and links in the logical network diagram, and fill the text sequence into the input template to generate model input data.

[0103] In step 1046, the text sequence refers to a string sequence in which the node information and link information in the logical network map are organized according to the dependency relationship and connection order between the nodes. This text sequence is used as the input content of the large model. The node information is arranged in order from low to high according to its level. Connecting words for describing the association type are inserted between two nodes with logical link connections, so that the entire text sequence presents a coherent narrative that conforms to the expression habits of natural language.

[0104] In this embodiment, the matching information obtained in step 1043 is acquired, and the node information and link information are converted into a text sequence with a logical order according to the connection relationship between inference nodes and logical links in the logical network diagram. During the conversion process, the node information of each inference node is arranged in order from low to high level, and descriptive text representing the association type is inserted between two inference nodes with logical link connections to form a complete text sequence. The text sequence is then filled into the specified position in the input template read in step 1045 to generate model input data that meets the requirements of large model input.

[0105] Step 1047: Call the corresponding large model interface according to the model identifier, send the model input data to the large model, receive the natural language text returned by the large model, and organize the natural language text into a natural language conclusion according to the output format.

[0106] In this embodiment, the address of the large model interface to be called is determined according to the model identifier read in step 1045, and the model input data generated in step 1046 is sent to the large model through this interface; after receiving the model input data, the large model performs semantic analysis on the input content based on its pre-trained language understanding ability, generates natural language text and returns it; after receiving the natural language text returned by the large model, the natural language text is formatted and organized according to the output format read in step 1045, and converted into a natural language conclusion that conforms to the preset data structure specification.

[0107] Through the above steps, this application utilizes a rule engine and a large model to perform deterministic matching and semantic understanding of the logical context graph, respectively, and generates two types of review conclusions in parallel, thus achieving synergistic complementarity between rule logic and semantic understanding.

[0108] Step 105: Encapsulate the structured review conclusion and the natural language review conclusion as node review comments into the review comment field of the current review node, and drive the review process instance to the next review node until all review nodes are completed, and output an executable application instance. The executable application instance encapsulates complete business data and node review comments of each review node.

[0109] The "Review Comments" field refers to the data storage location within the review node used to store node review comments. This field is created during the initialization of the review process instance. The "Executable Application Instance" refers to the independently running application entity generated after processing by all review nodes. This instance contains all the data and review records required for business processing. The "Complete Business Data" refers to all business-related information collected during the flow of the review process instance. This data includes original document data, key business elements, and intermediate conclusions. The "Node Review Comments" refers to the review judgment result generated by each review node after processing. These comments are stored in association with the corresponding review node.

[0110] In this embodiment, step 105 includes the following process: Step 1051: Read the node identifier of the current audit node in the audit process instance, determine the node identifier of the next audit node according to the preset node flow order, and call the process engine to change the current node pointer of the audit process instance from the node identifier of the current audit node to the node identifier of the next audit node.

[0111] In step 1051, the node flow order refers to the order in which each audit node is pre-configured in the audit process instance. This node flow order determines the flow path of the audit process instance. The process engine refers to the execution component used to control the state transition of the audit process instance. The process engine is responsible for maintaining the current node and executing the jump operation between nodes.

[0112] In this embodiment, the node identifier of the current audit node in the audit process instance is first read. According to the pre-set node flow order, the node identifier of the subsequent node is found at the position corresponding to the node identifier of the current audit node, and the subsequent node is determined as the next audit node. Then, the process engine is called to perform a node jump operation. The process engine changes the current node pointer recorded in the audit process instance from the node identifier of the current audit node to the node identifier of the next audit node, thus completing the flow of the audit process instance from one audit node to the next audit node.

[0113] Step 1052: Repeat the processing operations for each review node until the current node of the review process instance points to the process end marker, and collect the complete business data generated by the review process instance during the process and the node review opinions corresponding to each review node.

[0114] In step 1052, the process end identifier is a special identifier used to mark that the audit process instance has completed all audit nodes. When the current node of the audit process instance points to this identifier, it means that all audit nodes have been processed.

[0115] In this embodiment of the application, after completing the processing of the current review node and encapsulating the node review comments into the review comments field, the operations of steps 104 to 1051 are repeated. That is, for each review node to which the process flows, the logical context graph input, the call of the rule engine and the large model, the generation and encapsulation of the review conclusion, and the node jump are performed. When the current node of the review process instance points to the process end marker, the repeated execution stops, and the complete business data generated by the review process instance in the entire process and the node review comments stored for each review node are collected.

[0116] Step 1053: Combine the complete business data and the node review comments corresponding to each review node into an executable application instance.

[0117] In this embodiment of the application, the complete business data collected in step 1052 and the node review opinions corresponding to each review node are integrated. The complete business data is filled into the business data field according to the data structure specification of the executable application instance, and the node review opinions of each review node are filled into the review record field according to the node execution order. A complete executable application instance is generated by combining them. The executable application instance can run independently and support subsequent business operations.

[0118] Through the above steps, this application encapsulates the audit results and complete business data of each audit node into an executable application instance, realizing the complete recording of the audit process and one-click output of business applications.

[0119] Based on the above embodiments, the construction process of the executable application instance specifically includes: parsing each inference node and its hierarchical relationship in the logical network graph into a call sequence of functional components, wherein each inference node corresponds to a functional component, and the logical links between nodes correspond to the data flow paths between components; parsing each audit node and its flow order in the audit process instance into task nodes and transfer edges of the workflow, wherein each audit node corresponds to a task node, and the node execution order corresponds to the transfer edges of the workflow; and merging the parsed functional component call sequence with the workflow to generate an executable application instance that can be independently deployed and run. This executable application instance can automatically perform data fusion, logical inference, and process auditing based on the input business document, and output the audit results.

[0120] In this way, the present application directly maps the structured semantic information in the logical network diagram to functional components and workflow definitions, realizing agile construction from document semantic parsing to executable applications, without the need for manual coding or process orchestration, which significantly improves the efficiency and flexibility of application construction.

[0121] Figure 3This application provides a schematic diagram of the structure of an agile application building system that supports data fusion and intelligent process review, as shown in the embodiments of this application. Figure 3 As shown, the system includes: The determination module 31 is used to determine the digital base plate corresponding to the business document. The digital base plate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document.

[0122] The parsing module 32 is used to perform deep semantic parsing on the digital base plate through the thinking chain reasoning mechanism, parsing out multiple key business elements, and constructing a corresponding logical reasoning path for each key business element. Each logical reasoning path is used to derive intermediate conclusions based on the key business elements.

[0123] The construction module 33 is used to take the intermediate conclusions output by each logical reasoning path as reasoning nodes, add logical link identifiers to the reasoning nodes according to the logical relationship between each reasoning node, and construct a logical network map.

[0124] The matching module 34 is used to input the logical context graph into a pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion, and calls the large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion.

[0125] The output module 35 is used to encapsulate the structured review conclusion and the natural language review conclusion as node review opinions into the review opinion field of the current review node, and drive the review process instance to enter the next review node until all review nodes are completed and an executable application instance is output. The executable application instance encapsulates complete business data and node review opinions of each review node.

[0126] The agile application building system supporting data fusion and intelligent process review in this application embodiment is used to implement the aforementioned agile application building method supporting data fusion and intelligent process review. Therefore, the specific implementation of the agile application building system supporting data fusion and intelligent process review can be found in the embodiment section of the agile application building method supporting data fusion and intelligent process review above. The specific implementation can be referred to the description of the corresponding embodiment, which will not be repeated here.

[0127] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the agile application construction method supporting data fusion and intelligent process review described above.

[0128] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described agile application construction methods supporting data fusion and intelligent process auditing.

[0129] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0130] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the agile application construction method supporting data fusion and intelligent process review.

[0131] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0133] The foregoing has provided a detailed description of an agile application construction method and system supporting data fusion and intelligent process review, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for building agile applications that supports data fusion and intelligent process review, characterized in that, include: Determining the digital baseplate corresponding to a business document, wherein the digital baseplate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document; determining the digital baseplate corresponding to the business document includes: responding to a request to build an unstructured business document, collecting multi-source document data generated during the business process through a document digitization all-in-one machine, and initiating infrared transmission scanning through the document digitization all-in-one machine to capture the infrared transmission image inside the paper; performing pixel-level page layout analysis on the multi-source document data, separating text areas, table areas, signature areas, and handwritten annotation areas to form a set of document constituent units with position labels; and overlaying and registering the infrared transmission image with the visible page content reconstructed based on the set of document constituent units to generate the digital baseplate. The digital base plate is subjected to deep semantic analysis through the thinking chain reasoning mechanism to extract multiple key business elements, and a corresponding logical reasoning path is constructed for each key business element. Each logical reasoning path is used to deduce intermediate conclusions based on the key business elements. The intermediate conclusions output by each logical reasoning path are used as reasoning nodes. Logical link identifiers are added to the reasoning nodes according to the logical connections between them to construct a logical network map. Based on the dependencies represented by the logical links between each inference node, and combined with the semantic association strength between key business elements in the business document, all inference nodes are hierarchically divided; inference nodes assigned to the same level are organized into the same level group, and the level groups are arranged in order from low to high level. The logical context graph is input into a pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion. It also calls the large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion. The current review node in the review process instance reads the inference nodes in each level group in order from low to high level to perform the review judgment. The structured review conclusion and the natural language review conclusion are encapsulated as node review opinions in the review opinion field of the current review node, and the review process instance is driven to the next review node. After all review nodes are completed, the inference nodes and hierarchical relationships in the logical network graph are parsed into a call sequence of functional components, wherein each inference node corresponds to a functional component, and the logical links between nodes correspond to the data flow paths between components. The review nodes and flow order in the review process instance are parsed into task nodes and transition edges of the workflow, wherein each review node corresponds to a task node, and the node execution order corresponds to the transition edges of the workflow. The parsed functional component call sequence is merged with the workflow to generate an executable application instance that can be deployed and run independently. The executable application instance encapsulates complete business data and the node review opinions of each review node.

2. The method according to claim 1, characterized in that, The process of using intermediate conclusions output from each logical reasoning path as reasoning nodes, adding logical link identifiers to the reasoning nodes based on the logical connections between them, and constructing a logical network graph includes: Collect the intermediate conclusions output by each logical reasoning path, define each intermediate conclusion as a reasoning node, and each reasoning node carries the path identifier of the logical reasoning path to which it belongs; Traverse all inference nodes and identify whether there is a semantic relationship between any two inference nodes. The semantic relationship includes causal relationship, temporal relationship or dependency relationship. When the identification result indicates that there is a semantic association between two inference nodes, a logical link is established between the two inference nodes, a logical link identifier is assigned to the logical link, and the association type is marked on the logical link; A directed graph structure is constructed using each inference node as a graph vertex and each logical link as a graph edge. The directed graph structure is hierarchically divided according to the dependencies between each inference node. Inference nodes without preceding dependent nodes are assigned to the initial level, and inference nodes with the same dependency depth are assigned to the same level. Inference nodes and logical links are combined into a logical network graph according to the hierarchical order.

3. The method according to claim 2, characterized in that, The process involves hierarchically dividing the directed graph structure based on the dependencies between inference nodes, assigning inference nodes without preceding dependent nodes to the initial level, grouping inference nodes with the same dependency depth into the same level, and combining each inference node and each logical link into a logical network graph according to the hierarchical order, including: Based on the dependencies represented by the logical links between each inference node, and combined with the semantic association strength between key business elements in the business document, all inference nodes in the directed graph structure are hierarchically divided. Inference nodes without any preceding dependent nodes are assigned to the initial level, and inference nodes that depend on the inference nodes in the initial level and have no other preceding dependent nodes are assigned to the next level. The hierarchical division is repeated until all inference nodes are assigned to the corresponding level. Inference nodes assigned to the same level are organized into the same level group, and the level groups are arranged in order from low to high level. The hierarchical structure is formed by combining the processing order of the business elements corresponding to the inference nodes in each level group in the document review process. Based on the logical links between the hierarchical groups in the hierarchical structure, the hierarchical structure and the logical links are combined into a logical network diagram. The logical network diagram is used by the current audit node in the audit process instance to read the reasoning nodes in each level group in order from low to high level to perform audit judgment.

4. The method according to claim 1, characterized in that, The process involves deep semantic analysis of the digital baseboard using a thought chain reasoning mechanism to extract multiple key business elements. A corresponding logical reasoning path is then constructed for each key business element. Each logical reasoning path is used to deduce intermediate conclusions based on the key business elements, including: Traverse each layout area in the digital base plate, identify the business information carried in each layout area, and extract multiple key business elements from the business information; Create a logical reasoning path for each key business element, and configure a start point and an end point in the logical reasoning path. The start point is the key business element itself, and the end point is the intermediate conclusion to be derived. At least one intermediate node is constructed between the starting point and the ending point. Each intermediate node corresponds to a derivation step. A pointing marker is set between adjacent nodes. The pointing marker is used to represent the derivation direction from the earlier node to the later node. The data content of the key business elements is used as the input data of the starting point. Each intermediate node is activated in sequence according to the pointing identifier. After each intermediate node is activated, it performs a transformation operation on the input data according to the derivation logic carried by its own node, and generates intermediate data to be passed to the next node, until the termination point is activated and the intermediate conclusion is output.

5. The method according to claim 1, characterized in that, The process involves inputting the logical context graph into a pre-built review process instance, where the review process instance, at the current review node, calls a rule engine to perform rule matching on the logical context graph to generate a structured review conclusion, and then calls a large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion. This includes: Obtain a pre-built audit process instance, which contains multiple audit nodes, each configured with a node identifier and node execution order; When the review process instance is transferred to the current review node, the rule configuration information attached to the current review node is read. The rule configuration information contains at least one judgment rule, and each judgment rule consists of a judgment condition and a conclusion template. Traverse all nodes and all links in the logical network diagram, and use the node information carried by each node and the link information carried by each link as the matching information; The information to be matched is compared with the judgment conditions in the judgment rule. When the information to be matched meets the judgment conditions, the conclusion template corresponding to the judgment rule is instantiated into a structured conclusion. Read the model call parameters attached to the current review node, the model call parameters include input template, model identifier and output format; The information to be matched is converted into a text sequence according to the connection relationship between nodes and links in the logical network diagram, and the text sequence is filled into the input template to generate model input data; The corresponding large model interface is called according to the model identifier, the model input data is sent to the large model, the natural language text returned by the large model is received, and the natural language text is organized into a natural language conclusion according to the output format.

6. The method according to claim 1, characterized in that, The process instance is driven to proceed to the next review node, and after all review nodes are completed, an executable application instance is output. This executable application instance encapsulates complete business data and the review comments from each review node, including: Read the node identifier of the current review node in the review process instance, determine the node identifier of the next review node according to the preset node flow order, and call the process engine to change the current node pointer of the review process instance from the node identifier of the current review node to the node identifier of the next review node. Repeatedly execute the processing operations for each review node until the current node of the review process instance points to the process end marker, and collect the complete business data generated by the review process instance during the process and the node review opinions corresponding to each review node. The complete business data and the node review comments corresponding to each review node are combined into an executable application instance.

7. An agile application building system that supports data fusion and intelligent process review, characterized in that, include: A determination module is used to determine the digital baseplate corresponding to a business document. The digital baseplate is determined based on the infrared transmission image inside the paper corresponding to the business document and the visible page content corresponding to the document data in the business document. Determining the digital baseplate corresponding to the business document includes: responding to a construction request for an unstructured business document, collecting multi-source document data generated during the business workflow through a document digitization all-in-one machine, and initiating infrared transmission scanning through the document digitization all-in-one machine to capture the infrared transmission image inside the paper; performing pixel-level page layout analysis on the multi-source document data to separate text areas, table areas, signature areas, and handwritten annotation areas to form a set of document constituent units with positional annotations; and overlaying and registering the infrared transmission image with the visible page content reconstructed based on the document constituent unit set to generate the digital baseplate. The parsing module is used to perform deep semantic parsing on the digital base plate through the thinking chain reasoning mechanism, parsing out multiple key business elements, and constructing a corresponding logical reasoning path for each key business element. Each logical reasoning path is used to derive intermediate conclusions based on the key business elements. The module is used to take the intermediate conclusions output by each logical reasoning path as reasoning nodes, add logical link identifiers to the reasoning nodes according to the logical relationship between each reasoning node, and construct a logical network map. Based on the dependencies represented by the logical links between each inference node, and combined with the semantic association strength between key business elements in the business document, all inference nodes are hierarchically divided; inference nodes assigned to the same level are organized into the same level group, and the level groups are arranged in order from low to high level. The matching module is used to input the logical context graph into a pre-built review process instance. The review process instance calls the rule engine at the current review node to perform rule matching on the logical context graph to generate a structured review conclusion. It also calls the large model to perform semantic understanding on the logical context graph to generate a natural language review conclusion. The current review node in the review process instance reads the inference nodes in each level group in order from low to high level to perform review judgment. The output module is used to encapsulate the structured review conclusion and the natural language review conclusion as node review opinions into the review opinion field of the current review node, and drive the review process instance to the next review node. After completing all review nodes, it parses each inference node and hierarchical relationship in the logical network graph into a call sequence of functional components, wherein each inference node corresponds to a functional component, and the logical link between nodes corresponds to the data flow path between components; it parses each review node and flow order in the review process instance into task nodes and transition edges of the workflow, wherein each review node corresponds to a task node, and the node execution order corresponds to the transition edge of the workflow; it merges the parsed functional component call sequence with the workflow to generate an executable application instance that can be deployed and run independently. The executable application instance encapsulates complete business data and the node review opinions of each review node.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the agile application construction method supporting data fusion and intelligent process auditing as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the agile application construction method supporting data fusion and intelligent process review as described in any one of claims 1 to 6.

Images