Purchase process penetration supervision method and system based on document engine

Abstract

This invention discloses a method and system for penetrating supervision of procurement processes based on a document engine. The method includes: first, obtaining contextual representation data of target procurement documents selected from the procurement entities to be supervised from the document engine. This data includes the entity identifier, relevant procurement document features, and the target document's own features. Then, in an edge inference node, a procurement document inference model is invoked to infer the correlation between the supervised entity and the target document based on the contextual data, obtaining a correlation confidence level. This model is trained and optimized in the cloud and then migrated and deployed to the edge node to achieve latency significantly lower than cloud inference. Finally, based on the correlation confidence level, a penetrating supervision instruction is executed on the target procurement document. This invention effectively reduces the response latency of intelligent supervision and improves the ability to perform real-time, accurate, and penetrating analysis of complex procurement processes.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for penetrating and monitoring the procurement process based on a document engine. Background Technology

[0002] In corporate procurement process oversight, traditional methods primarily rely on pre-defined rule engines for compliance checks or on post-event audits to identify issues. These approaches struggle to provide real-time, dynamic, and comprehensive analysis of complex procurement processes that span multiple departments and stages. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for penetrating and supervising the procurement process based on a document engine.

[0004] In a first aspect, embodiments of the present invention provide a method for penetrating and monitoring the procurement process based on a document engine, including:

[0005] Obtain the contextual representation data of the target purchase order in the document engine. The target purchase order is selected for the purchase entity to be supervised. The contextual representation data of the target purchase order includes the entity identifier data of the purchase entity to be supervised, the feature representation of at least one related purchase order that is related to the purchase entity to be supervised, and the feature representation of the target purchase order.

[0006] In the edge inference node, the purchase order inference model is invoked based on the contextual representation data of the target purchase order to infer the correlation between the procurement entity to be supervised and the target purchase order, thereby obtaining the correlation confidence between the procurement entity to be supervised and the target purchase order. The purchase order inference model is pre-trained and optimized in the cloud training sub-cluster and then deployed and migrated to the edge inference node. The inference latency of the purchase order inference model in the edge inference node is lower than that of the purchase order inference model in the cloud training sub-cluster.

[0007] Based on the confidence level between the procurement entity to be regulated and the target procurement document, a penetrating supervision instruction is executed on the target procurement document.

[0008] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0009] Compared to existing technologies, the beneficial effects of this invention include: Employing a document engine-based procurement process penetration supervision method and system disclosed in this invention, contextual representation data of target procurement documents selected for the procurement entity to be supervised is obtained from the document engine. This data includes the entity identifier, relevant procurement document features, and the target document's own features. Subsequently, a procurement document inference model is invoked at the edge inference node to infer the correlation between the supervised entity and the target document based on this contextual data, obtaining a correlation confidence level. This model is trained and optimized in the cloud and then migrated and deployed to the edge node to achieve latency significantly lower than cloud inference. Finally, penetration supervision instructions are executed on the target procurement document based on the correlation confidence level. This invention effectively reduces the response latency of intelligent supervision and enhances the ability to perform real-time, accurate penetration analysis of complex procurement processes. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating the steps of the procurement process penetration supervision method based on a document engine provided in this embodiment of the invention;

[0012] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0014] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0015] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the document engine-based procurement process penetration supervision method provided in this embodiment. The following is a detailed description of the document engine-based procurement process penetration supervision method.

[0016] Step S201: Obtain the contextual representation data of the target procurement document in the document engine. The target procurement document is selected for the procurement entity to be supervised. The contextual representation data of the target procurement document includes the entity identification data of the procurement entity to be supervised, the feature representation of at least one related procurement document that is related to the procurement entity to be supervised, and the feature representation of the target procurement document.

[0017] Step S202: In the edge inference node, the purchase order inference model is invoked to infer the correlation between the procurement entity to be supervised and the target purchase order based on the contextual representation data of the target purchase order, thereby obtaining the correlation confidence between the procurement entity to be supervised and the target purchase order; the purchase order inference model is pre-trained and optimized in the cloud training sub-cluster and then deployed and migrated to the edge inference node; the inference latency of the purchase order inference model in the edge inference node is lower than the inference latency of the purchase order inference model in the cloud training sub-cluster;

[0018] Step S203: Based on the correlation confidence between the procurement entity to be supervised and the target procurement document, execute a penetration supervision instruction on the target procurement document.

[0019] In an embodiment of the invention, for example, the server receives a penetrating oversight task instruction, which specifies the procurement entity to be overseen. The procurement entity to be overseen is an entity in the procurement process, such as a supplier, a procurement department, or a project team. The server parses the instruction and extracts a unique entity identifier for the procurement entity to be overseen, such as a supplier code or department ID.

[0020] Subsequently, the server interacts with the document engine. The document engine is the core business system managing all document data and process status in the enterprise's procurement process. The server submits a query request to the document engine, requesting all pending procurement documents related to the subject entity based on the aforementioned entity identifier. These relationships include direct and indirect associations. A direct association means the document's business fields explicitly contain the entity identifier, such as the "supplier" field in a purchase order. An indirect association refers to a business connection established with the entity through paths such as the document's applying department, related projects, or historical approval chains. The document engine retrieves information from its database based on the query criteria and returns a list of identifiers for pending procurement documents. These pending procurement documents span multiple procurement categories and are distributed across different process stages, including procurement requirements, inquiries, contracts, and payments.

[0021] The server selects a document from the list of pending purchase orders as the target purchase order to be processed. The server constructs contextual representation data for this target purchase order. Contextual representation data is a structured collection of data used to represent the target purchase order itself and its association with the entity to be regulated, specifically including the following three parts:

[0022] First, the entity identification data of the procurement entities to be regulated. The server does not directly use their business codes, but rather converts them into a machine-readable deep feature representation. The server accesses a pre-built entity feature library, which encodes the historical behavioral data, qualification information, risk ratings, and industry attributes of various entities through a pre-trained model, generating a fixed-length high-dimensional feature vector for each entity. Therefore, the entity identification data is specifically represented as a feature vector representing the comprehensive profile of that entity.

[0023] Second, the server retrieves the feature representation of at least one relevant procurement document that is related to the procurement entity under supervision. Based on pre-defined association rules or graph queries, the server obtains other documents from the document engine that have historically been related to the procurement entity under supervision and may have a logical connection with the current target procurement document. For example, if the target procurement document is a contract, relevant procurement documents may include previous purchase requests, technical agreements, and other related quotations. For each relevant procurement document, the server obtains its feature representation. The feature representation of each document is a vector generated through a feature extraction process. This vector integrates multi-dimensional information such as the document type, key numerical fields (e.g., amount, quantity), time information, semantic embedding vector of the text summary, and process status. These feature representations of relevant documents are organized into an ordered list or a feature matrix, constituting the historical association context of the target procurement document.

[0024] Third, the feature representation of the target purchase order itself. The server extracts complete data of the target purchase order from the document engine, including structured fields and unstructured text content. Structured fields include, but are not limited to, document number, purchase category code, total amount, creation time, approval node status, etc. Unstructured text content includes document description, contract terms, remarks, etc. The server standardizes and normalizes the structured fields and maps them into numerical vectors through an embedding layer. For unstructured text, the server calls the text feature extraction module, which performs word segmentation and semantic encoding on the text, outputting a dense vector representing the core semantics of the text. Subsequently, the server uses a feature fusion network to fuse and reduce the dimensionality of the vectors from the structured fields, the text semantic vector, and other possible features, generating a unified, fixed-dimensional feature representation vector for the target purchase order.

[0025] After extracting and generating the three parts of data mentioned above, the server assembles them according to a predefined data structure to form the final contextual representation data of the target purchase order. This data structure typically includes the following components: a feature vector representing the entity to be monitored; a sequence of multiple feature vectors representing a series of related historical documents; and a feature vector representing the current target document. Furthermore, the server can append metadata information such as timestamps, data version numbers, or association strength weights to this data structure for use by subsequent inference models.

[0026] For example, in a specific audit scenario, the entity to be regulated is supplier "X". The target purchase order is a high-value "advance payment application" for this supplier. The contextual representation data constructed by the server includes: the entity feature vector of supplier "X" (which may encode its registration information, historical cooperation records, and past risk events); the feature representation of a set of related purchase orders, such as the feature vector of the "purchase contract" with lenient terms corresponding to this advance payment, and the feature vector of the "delivery acceptance form" showing in the history that the supplier has delayed delivery; and the feature representation vector of the current "advance payment application" itself, which encodes features such as an abnormally high advance payment ratio and an ambiguous application reason.

[0027] Through the steps described above, the server completes the deep feature generation and structured encapsulation of the target purchase order's contextual information. The output of this step provides a complete and high-quality input data foundation for subsequent low-latency, high-precision correlation inference in edge inference nodes.

[0028] In this embodiment of the invention, the edge inference node is pre-configured with model acceleration configuration parameters for the purchase order inference model to perform correlation inference; the step of calling the purchase order inference model in the edge inference node to perform correlation inference between the procurement entity to be supervised and the target purchase order based on the contextual representation data of the target purchase order, and obtaining the correlation confidence between the procurement entity to be supervised and the target purchase order, can be implemented through the following example.

[0029] The purchase order inference model is accelerated and executed according to the model acceleration configuration parameters.

[0030] The inference model for the purchase order after the accelerated configuration execution is invoked. Based on the contextual representation data of the target purchase order, the model infers the correlation between the procurement entity to be supervised and the target purchase order, and obtains the correlation confidence between the procurement entity to be supervised and the target purchase order.

[0031] In this embodiment of the invention, for example, the server sends the encapsulated target purchase order context representation data to a designated edge inference node. This node has pre-deployed a purchase order inference model migrated from a cloud-based training sub-cluster and loaded a model acceleration configuration parameter file optimized for the node's hardware characteristics (such as a specific GPU or AI accelerator card). The server issues a model invocation command to the edge inference node. The node first performs real-time configuration and optimization of the loaded purchase order inference model according to the acceleration configuration parameters. For example, the parameters specify operator fusion of consecutive convolutional layers and activation layers in the model's computation graph to reduce memory access overhead. Based on this rule, the node rewrites and compiles the model's network structure in memory, generating a highly optimized computation graph instance that can be efficiently executed on local hardware. Subsequently, the server instructs the node to invoke this optimized model instance. The model receives the incoming context representation data and performs forward inference computation. Specifically, the feature extraction component in the model processes the features of the relevant document sequence and the target document, calculating the association weight between them and the feature vector of the subject to be regulated through an attention mechanism. The feature generation component synthesizes a comprehensive context-aware representation based on these weighted features. Finally, the model's inference output sub-component calculates a scalar value between 0 and 1 based on this representation, representing the correlation confidence between the regulated purchasing entity and the target purchase order. For example, when processing a high-amount prepayment application from supplier "X," the model might output a high confidence score of 0.92, indicating a strong correlation between the order and supplier "X's" historical behavior patterns, suggesting a strong association requiring close scrutiny. The server then receives this confidence score result for subsequent regulatory decisions.

[0032] In this embodiment of the invention, the model acceleration configuration parameters include the topology fusion rules between the various computing unit nodes in the purchase order inference model;

[0033] The accelerated configuration and execution processing of the purchase order inference model according to the model acceleration configuration parameters can be implemented through the following example.

[0034] The topology of each computing unit node of the purchase order reasoning model is rewritten and compiled according to the topology fusion rules.

[0035] The purchase order inference model, after the accelerated configuration execution process, infers the correlation between the procurement entity to be supervised and the target purchase order based on the contextual representation data of the target purchase order, and obtains the correlation confidence between the procurement entity to be supervised and the target purchase order. This can be implemented through the following example.

[0036] The inference model for the procurement document after topology rewriting and compilation is invoked to infer the correlation between the procurement entity to be supervised and the target procurement document based on the contextual representation data of the target procurement document, thereby obtaining the correlation confidence between the procurement entity to be supervised and the target procurement document.

[0037] In an embodiment of the invention, for example, the server issues an acceleration configuration instruction containing topology fusion rules to the edge inference node. This rule specifically defines the sequence of merging computational unit nodes in the purchase order inference model. For instance, the rule specifies merging the three consecutive computational unit nodes that frequently appear in the model—"LayerNorm - Linear - ReLU activation layer"—into a single composite computational unit node.

[0038] According to this rule, the edge inference node performs topology rewriting on the loaded model computation graph. The node traverses the model's computation graph in memory, identifies all subgraph structures that conform to the "LayerNorm-Linear-ReLU" sequence pattern, and replaces them with a single computation node corresponding to a predefined, highly optimized fused kernel. This node is marked as "FusedLayerNormLinearReLU". After replacing all matching subgraphs, the node calls its local runtime compiler to compile the rewritten computation graph into machine code that can be directly executed on the edge device hardware.

[0039] The server then instructs the node to invoke this compiled optimized model. When processing the contextual representation data of the target purchase order, the model previously required a sequence of three independent calculations and two intermediate result caches for the "LayerNorm-Linear-ReLU" operation; now, this is completed in one go by a single "FusedLayerNormLinearReLU" fused kernel. This significantly reduces memory accesses and kernel startup overhead. The model executes all optimized computational units sequentially via forward propagation, ultimately generating a quantified relevance confidence value through the output layer. For example, when processing the contextual data of supplier X's prepayment application, the optimized model, executed through an efficient kernel, outputs a relevance confidence of 0.92. The server receives this result for further processing.

[0040] In this embodiment of the invention, the model acceleration configuration parameters include the tensor calculation acceleration configuration parameters of the tensor calculation unit node in the purchase order inference model;

[0041] The accelerated configuration and execution processing of the purchase order inference model according to the model acceleration configuration parameters can be implemented through the following example.

[0042] According to the tensor computation acceleration configuration parameters, the tensor computation execution mechanism of the tensor computation unit node is configured and executed for acceleration.

[0043] In this embodiment of the invention, for example, the server issues instructions containing tensor computation acceleration configuration parameters based on the hardware capabilities of the edge inference node. These parameters specify the numerical precision and hardware instruction set to be used during model inference. For instance, the parameters explicitly require that tensor computations such as floating-point matrix multiplication in the model be converted from standard FP32 precision to FP16 mixed-precision computation, and that TensorCores for the edge device's GPU be enabled for computation. After loading these configuration parameters, the edge inference node reconfigures its runtime tensor computation execution mechanism. The node converts the weight parameters and activation values ​​of all tensor computation unit nodes (such as linear transformation layers and attention scoring layers) in the model from FP32 cache to FP16 format. Simultaneously, the node invokes its hardware driver to allocate appropriate TensorCore resources supporting fast FP16 computation for subsequent core operations such as matrix multiplication. When the server instructs the node to perform inference, the model will invoke the configured FP16 computation kernels and TensorCore hardware units when processing feature vectors in the context representation data for large-scale matrix multiplication operations (e.g., calculating attention weights). This significantly improves computational throughput and reduces memory bandwidth usage. The model efficiently outputs relevant confidence scores through a series of accelerated tensor calculations.

[0044] In this embodiment of the invention, the purchase order inference model includes a feature extraction component and a feature generation component. The feature generation component includes a representation generation sub-component and an inference output sub-component. The contextual representation data of the target purchase order is sequentially segmented into multiple sequence feature units after being input into the purchase order inference model. The step of calling the purchase order inference model in the edge inference node to infer the correlation between the procurement entity to be supervised and the target purchase order based on the contextual representation data of the target purchase order, and obtaining the correlation confidence between the procurement entity to be supervised and the target purchase order, can be implemented through the following example.

[0045] Obtain the feature representation vector of each sequence feature unit in the context representation data of the target purchase order;

[0046] In the edge inference node, the feature extraction component is invoked to perform feature relationship modeling on each sequence feature unit based on the feature correlation between the feature representation vectors of each sequence feature unit, so as to obtain the associated feature representation of each sequence feature unit;

[0047] In the edge inference node, the representation generation sub-component is invoked to generate the generated feature representation of the generated input feature based on the feature correlation between the associated feature representation of each sequence feature unit and the associated feature representation of the generated input feature.

[0048] In the edge inference node, the inference output sub-component is invoked to generate feature representations based on the generated input features to infer the correlation between the procurement entity to be regulated and the target procurement document, thereby obtaining the correlation confidence between the procurement entity to be regulated and the target procurement document.

[0049] In this embodiment of the invention, for example, the server sends the context representation data of the target purchase order to the edge inference node and instructs the node to call the purchase order inference model for processing. After receiving the data, the model first parses and sequentially segments it into multiple discrete sequence feature units. Specifically, the feature vector of the subject to be regulated in the context representation data is treated as an independent unit, the feature representation vector of each related purchase order is treated as an independent unit, and the feature representation vector of the target purchase order itself is also treated as an independent unit. These units are arranged in a preset order (such as chronological or logical order) to form a feature vector sequence. The server instructs the edge inference node to first execute the step of "obtaining the feature representation vector of each sequence feature unit in the context representation data of the target purchase order". The node directly reads the parsed sequence data, and each unit itself is a high-dimensional feature vector without additional transformation. For example, in the case of the prepayment order of supplier X, the sequence may include: unit 1 (supplier X subject vector), unit 2 (related framework contract vector), unit 3 (historical acceptance order vector), and unit 4 (current prepayment application order vector). Next, the server instructs the node to call the feature extraction component of the model. This component is usually composed of a multi-layer Transformer encoder. The component receives the aforementioned sequence feature vectors and, through its internal self-attention mechanism, calculates the feature correlation score between any two sequence feature units. Based on this score, it performs weighted aggregation and transformation on the vector information of each unit. This process is equivalent to "modeling the feature relationships of each sequence feature unit based on the feature correlation between their feature representation vectors." After multiple layers of processing, each original sequence feature unit is transformed into an associated feature representation that contains global contextual relationships. For example, the associated feature representation of the current prepayment application (unit 4) now includes its associated information encoding with the supplier's historical contracts (unit 2) and bad records (unit 3). Subsequently, the server instruction node calls the representation generation subcomponent in the feature generation component. This subcomponent adopts a decoder-like structure. It uses the associated feature representation of the subject to be regulated (i.e., the processed result of unit 1) as the initial generation input feature. The subcomponent calculates the correlation between this generated input feature and the associated feature representations of all other units in the sequence (such as units 2, 3, and 4) through a cross-attention mechanism. Based on these relevance weights, the sub-component extracts and integrates key information from the associated feature representations of other units, iteratively updates and enhances the generated input features, and finally outputs a highly refined generated feature representation that incorporates all relevant contexts. This representation can be understood as a deep behavioral representation of the subject to be regulated within the current specific document context. Finally, the server instruction node calls the inference output sub-component in the feature generation component. This sub-component is typically composed of a multilayer perceptron. It takes the generated feature representation obtained in the previous step as input and performs nonlinear transformation and mapping through a fully connected layer.The output layer of the sub-component uses a sigmoid activation function to compress the final mapping result into a scalar value between 0 and 1. This scalar represents the correlation confidence between the regulated purchasing entity and the target purchase order. For example, for a prepayment order from supplier X, the sub-component outputs 0.92, indicating that the model infers a very high correlation between this order and supplier X. The server receives this confidence score to initiate subsequent penetrating regulatory instructions.

[0050] In this embodiment of the invention, the inference output subcomponent is generated by updating the structure of the category determination subcomponent using the confidence calculation subcomponent; the step of calling the inference output subcomponent in the edge inference node to perform correlation inference between the procurement entity to be supervised and the target procurement document based on the generated feature representation of the generated input features, and obtaining the correlation confidence between the procurement entity to be supervised and the target procurement document, can be implemented through the following example.

[0051] In the edge inference node, the category determination subcomponent is invoked to determine the association category between the procurement entity to be supervised and the target procurement document based on the generated feature representation of the generated input features, thereby obtaining the output association category identifier between the procurement entity to be supervised and the target procurement document;

[0052] In the edge inference node, the confidence calculation subcomponent is invoked to perform association confidence calculation on the procurement entity to be supervised and the target procurement document based on the generated feature representation of the generated input features, and to obtain multiple undetermined confidence values ​​under the output association category identifier;

[0053] Select the correlation confidence level between the procurement entity to be supervised and the target procurement document from among the multiple undetermined confidence levels.

[0054] In this embodiment of the invention, exemplarily, the server instructs the edge inference node to call the inference output subcomponent to process the generated feature representation. This subcomponent consists of two cooperating parts: a category determination subcomponent and a confidence calculation subcomponent. The confidence calculation subcomponent is added to the original category determination subcomponent during the model optimization phase through structural updates. First, the server instructs the node to call the category determination subcomponent. This subcomponent is a pre-trained multi-classifier that receives the generated feature representation as input and calculates the probability distribution of the feature representation belonging to various preset associated categories through a fully connected layer and a Softmax function. These associated categories include, for example, "strong association - high risk," "strong association - normal," "weak association," and "suspected irrelevant." The subcomponent outputs the category identifier with the highest probability as the output associated category identifier. For example, for the feature representation generated from supplier X's prepayment invoice, this subcomponent determines that it belongs to the "strong association - high risk" category. Subsequently, the server instructs the node to call the confidence calculation subcomponent. This subcomponent is a parallel regression network branch that also receives the same generated feature representation as input. It calculates confidence estimates under different dimensions or algorithms for the specific category of "strong correlation - high risk" determined in the previous step, using different fully connected layers and activation functions, and outputs multiple undetermined confidence values. For example, it may output three values ​​simultaneously: a value based on the model's predicted probability calibration (0.91), a metric based on the distance from the generated feature representation to the prototype of this class (0.93), and an evaluation value based on the concentration of attention weights during model inference (0.89). Finally, the server instruction node selects one of these undetermined confidence values ​​as the final correlation confidence value according to a preset aggregation strategy. The strategy can be to take the maximum value, the average value, or a weighted average. For example, using the maximum value strategy, the node selects 0.93 from {0.91, 0.93, 0.89} as the correlation confidence between the procurement entity to be monitored and the target procurement document, and returns it to the server. This design improves the robustness and reliability of a single confidence value through multi-angle evaluation.

[0055] In this embodiment of the invention, the step of obtaining the feature representation vector of each sequence feature unit in the context representation data of the target purchase order can be implemented through the following example.

[0056] Obtain a business field feature mapping table, which includes feature representation vectors of multiple standardized business fields;

[0057] Retrieve the feature representation vector of each sequence feature unit in the context representation data of the target purchase order from the business field feature mapping table;

[0058] If a standardized business field that matches the index of the first sequence feature unit in the context representation data of the target purchase order is found in the business field feature mapping table, then the feature representation vector of the standardized business field that matches the index is determined as the feature representation vector of the first sequence feature unit.

[0059] The second sequence feature unit, excluding the first sequence feature unit, in the context representation data of the target purchase order is subjected to feature encoding processing to obtain the feature representation vector of the second sequence feature unit.

[0060] In this embodiment of the invention, for example, when the server executes the step of obtaining the feature representation vector of a sequence feature unit, it first loads a pre-set business field feature mapping table from the storage system. This mapping table is a key-value database, where the key is the code or name of a standardized business field, and the value is a fixed-dimensional feature vector pre-trained from historical data. For example, the table contains an entry with the key "Supplier Code_SUP_XYZ", whose corresponding value is a 256-dimensional floating-point vector; and an entry with the key "Purchase Category_Server", corresponding to another 256-dimensional vector. Subsequently, the server parses the contextual representation data of the target purchase order. For each sequence feature unit that needs to be processed, the server attempts to match it with the key in the mapping table. For example, a sequence feature unit carries the business field identifier "Supplier Code", whose specific value is "SUP_XYZ". The server searches the mapping table for the key "Supplier Code_SUP_XYZ". If the search is successful, i.e., the unit is the first sequence feature unit, the server directly retrieves the corresponding pre-calculated feature vector from the mapping table and determines it as the feature representation vector of that unit. This process enables rapid feature generation of standardized, high-frequency business concepts. For other sequence feature units in the contextual representation data that cannot be directly retrieved from the mapping table—i.e., second sequence feature units—the server performs real-time feature encoding. These units typically contain non-standardized text descriptions, custom fields, or composite information. For example, a unit might contain the text of a contract's "special clause": "Warranty period extended to five years." The server calls a lightweight text encoding model (such as a simplified version of SentenceTransformer) to semantically encode the text, dynamically generating a feature representation vector. This vector has the same dimension as the vector obtained from the mapping table, ensuring compatibility with subsequent model processing. Through this hybrid mechanism, the server efficiently transforms all sequence feature units into a unified format feature representation vector, providing input for subsequent feature relationship modeling.

[0061] In this embodiment of the invention, the document engine filters multiple pending purchase documents for the procurement entity to be supervised, and the target purchase document is any one of the multiple pending purchase documents; the multiple pending purchase documents belong to at least two different procurement process stages of the document engine; at least two purchase document inference models are configured in the edge inference node, and the at least two purchase document inference models correspond one-to-one with at least two procurement process stages;

[0062] The present invention also provides the following embodiments.

[0063] In the edge inference node, the at least two purchase order inference models are invoked to perform correlation inference between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage based on the contextual representation data of each pending purchase order in the corresponding procurement process stage, so as to obtain the correlation confidence between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage.

[0064] Among them, the correlation inference of the at least two purchase order reasoning models is processed concurrently.

[0065] In an embodiment of the invention, for example, after receiving a penetrating supervision instruction for a procurement entity to be supervised (e.g., supplier "X"), the server filters out multiple pending procurement documents related to that entity from the document engine. These pending procurement documents are distributed across at least two different procurement process stages. For example, documents A and B belong to the "contract approval" stage, while documents C and D belong to the "payment execution" stage.

[0066] The server pre-deploys at least two purchase order inference models, each corresponding to a specific process stage, within designated edge inference nodes. For example, a "Contract Stage Inference Model" optimized for analyzing contract text and terms, and a "Payment Stage Inference Model" specifically designed for analyzing payment behavior and invoice matching, are deployed. The server performs the following operations in parallel: constructing contextual representation data for documents A and B in the "Contract Approval" stage, and simultaneously constructing contextual representation data for documents C and D in the "Payment Execution" stage. The server then issues concurrent processing instructions to the edge inference nodes. Upon receiving the instructions and all batch data, the edge inference nodes utilize their multi-core or parallel computing resources to simultaneously load and invoke both the "Contract Stage Inference Model" and the "Payment Stage Inference Model." The Contract Stage Model processes the contextual data for documents A and B, while the Payment Stage Model processes the contextual data for documents C and D. The inference processes of both models are executed concurrently within the nodes. Each model independently performs relevance inference based on its received contextual representation data. For example, the contract phase model analyzes the correlation between supplier "X" and the draft contract (document A) in terms of historical cooperation patterns, risk clauses, etc., and outputs a confidence level (e.g., 0.88). The payment phase model analyzes the correlation between supplier "X" and a payment application (document C) in terms of payment frequency, amount anomalies, etc., and outputs another confidence level (e.g., 0.95). The server synchronously collects relevant confidence level results from different models for each pending purchase document at different stages, thereby obtaining a comprehensive correlation profile of the regulated procurement entity across multiple process stages, providing a phased and fine-grained basis for judgment in implementing penetrating supervision.

[0067] In this embodiment of the invention, the following implementation methods are also provided.

[0068] Obtain the original inference model and training data set, wherein the training data set includes a training subset and a validation subset;

[0069] The original inference model is trained on the training subset in the cloud training sub-cluster to obtain the purchase order inference model;

[0070] In the cloud training sub-cluster, the model inference performance of the purchase order inference model is verified based on the verification subset to obtain the first performance verification index;

[0071] In the edge inference node, the model inference performance of the purchase order inference model is verified based on the verification subset to obtain a second performance verification index;

[0072] If the deviation between the first performance verification indicator and the second performance verification indicator is less than the allowable deviation threshold, then the purchase order inference model is called in the edge inference node to perform correlation inference between the procurement entity to be supervised and each pending purchase order selected for the procurement entity to be supervised. The target purchase order is any pending purchase order selected for the procurement entity to be supervised.

[0073] In this embodiment of the invention, exemplarily, the server executes the training and deployment verification process of the purchase order inference model. The server first retrieves a pre-defined original inference model from the model repository, typically an initial network structure based on Transformer. Simultaneously, the server obtains a training dataset from a historical business database and annotation system. This dataset contains a large number of historical purchase order samples and their corresponding "purchase entity-document" relevance tags annotated by experts. The server divides the dataset into a training subset and a validation subset. Subsequently, the server starts a distributed training task in a cloud-based training sub-cluster. The training task, based on the training subset, uses a backpropagation algorithm and an optimizer (such as AdamW) to iteratively update the parameters of the original inference model to minimize prediction error. After training, an optimized purchase order inference model is obtained. In the cloud cluster, the server uses the validation subset to perform batch inference on the trained model, calculating its accuracy, recall, F1 score, and other metrics as the first performance verification indicators. For example, the accuracy of the cloud model is 95.2%. Next, the server packages and deploys the purchase order inference model and its required acceleration configuration parameters to the target edge inference node. After deployment, the server instructs the edge node to perform localized inference testing on the deployed model using the same validation subset data (already transferred to the node's local machine). After executing the test, the node returns the calculated performance metric (such as accuracy) as a second performance validation metric to the server. For example, the edge node returns an accuracy of 95.0%. The server compares the first and second performance validation metrics. It calculates the deviation, for example, the absolute value of the accuracy deviation is 0.2%. The server compares this deviation with a preset allowable deviation threshold (e.g., 2%). Since 0.2% is less than 2%, the deviation is within the allowable range, and the server determines that the deployment of the model on the edge node meets the performance requirements. Subsequently, the server formally instructs the edge inference node to call this validated purchase order inference model to perform relevance inference on all selected pending purchase orders for the target procurement entity (such as supplier "X"). Any pending purchase order processed becomes the target purchase order.

[0074] In this embodiment of the invention, the verification samples in the verification subset include a first verification sample and a second verification sample; the first verification sample is a verification purchase order that the sample purchaser determines is associatable, and the second verification sample is a verification purchase order that the sample purchaser determines is not associatable; the performance verification of the model inference performance of the purchase order inference model based on the verification subset in the cloud training sub-cluster to obtain a first performance verification index can be implemented through the following example.

[0075] In the cloud training sub-cluster, the purchase order reasoning model is invoked to infer the correlation between the verification samples in the verification subset and the corresponding sample purchasing entities, and the output association category identifier is obtained;

[0076] The percentage of the first sample that was determined to be correlated in the first verification sample was statistically analyzed.

[0077] The percentage of second samples that were misclassified as related in the second verification sample was statistically analyzed.

[0078] The first performance verification index is determined based on the ratio between the first sample proportion parameter and the second sample proportion parameter.

[0079] In this embodiment of the invention, for example, the server performs cloud-based performance verification of the purchase order inference model. The server first prepares a verification subset, where each verification sample contains a sample purchase entity (e.g., supplier "Y") and a verification purchase order along with its contextual data. The verification samples are divided into two categories: first verification samples, whose sample purchase entity and the verification purchase order are determined to be "associative" by expert annotation; and second verification samples, which are determined to be "unassociative". For example, the verification subset contains 100 samples, of which 50 are first verification samples and 50 are second verification samples. The server, in a cloud training sub-cluster, calls the pre-trained purchase order inference model to perform batch inference on all samples in the verification subset. For each sample, the model outputs an output association category identifier based on its contextual representation data; for example, "strong association" or "weak association" is considered "associative," and "suspected irrelevant" is considered "unassociative." After inference, the server performs statistical analysis. First, the server calculates the percentage of first verification samples determined to be associative. Specifically, the server examines the model output of all 50 first validation samples and counts the number of samples that the model correctly identifies as "associative". Assuming 48 samples are correctly identified as associative, the first sample proportion parameter (i.e., recall) is 48 / 50 = 0.96. Next, the server counts the second sample proportion parameter of the second validation samples that are incorrectly identified as associative. The server examines the model output of all 50 second validation samples and counts the number of samples that the model incorrectly identifies as "associative". Assuming 5 samples are incorrectly identified, the second sample proportion parameter (i.e., false positive rate) is 5 / 50 = 0.10. Finally, the server determines the first performance validation metric based on the ratio between the first and second sample proportion parameters. The server does not directly use a single ratio, but rather combines the recall (0.96) with the precision (48 correctly identified as associative / 48 + 5 = 53, approximately 0.906) to calculate the F1 score as the first performance validation metric. In this example, the F1 score is approximately 20.96 / (0.906 + 0.906) = 0.932. The server records this F1 score (0.932) as the first performance validation metric for the model in the cloud environment.

[0080] In this embodiment of the invention, the at least two purchase order inference models share a multi-stage collaborative inference network structure; the step of calling the at least two purchase order inference models in the edge inference node to infer the correlation between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage based on the contextual representation data of each pending purchase order in the corresponding procurement process stage can be implemented through the following example.

[0081] Obtain the contextual representation data of the target purchase order; the target purchase order is any one of the pending purchase orders in the corresponding procurement process stage;

[0082] The target cross-stage shared feature representation is determined by the prior learning backbone network in the multi-stage collaborative reasoning network based on the contextual representation data of the target purchase order; the multi-stage collaborative reasoning network is used to execute multiple process reasoning stages based on the contextual information of the input document.

[0083] The stage mapping component in the multi-stage collaborative reasoning network determines the target stage-specific representation vectors for each of the multiple process reasoning stages based on the contextual representation data of the target procurement document or the baseline feature representation generated when the prior learning backbone network processes the contextual representation data of the target procurement document. The stage mapping component includes a common mapping component and stage information extraction components matched to each of the multiple process reasoning stages. The common mapping component is used to extract the baseline cross-stage shared feature representation, and the stage information extraction component is used to extract the baseline stage-specific representation vectors for the corresponding multiple process reasoning stages based on the baseline cross-stage shared feature representation. The target stage-specific representation vectors for the multiple process reasoning stages are determined based on the baseline stage-specific representation vectors for the multiple process reasoning stages.

[0084] Through each inference output component in the multi-stage collaborative inference network, the inference output information of multiple process inference stages corresponding to the inference output component is determined based on the target stage-specific representation vector of multiple process inference stages corresponding to the inference output component and the target cross-stage shared feature representation; the multi-stage collaborative inference network includes inference output components matched to the multiple process inference stages respectively.

[0085] In this embodiment of the invention, for example, the server performs a collaborative reasoning task for multiple procurement process stages. The server obtains the contextual representation data of a pending procurement document (e.g., a draft server procurement contract) for a procurement entity to be supervised (e.g., supplier "X") in the "contract approval" stage, and uses it as the target procurement document data for current processing. The server instructs the edge inference nodes to call the deployed multi-stage collaborative reasoning network. This network is a unified deep learning model whose internal structure is designed to simultaneously output the reasoning results of multiple process stages. The network first processes the input target procurement document contextual representation data through its prior learning backbone network. This backbone network typically consists of multiple layers of Transformer encoders, which extract deep, abstract features from the input data. A comprehensive feature representation output by the backbone network is defined as the target cross-stage shared feature representation. This representation contains general pattern information in the association between the document and supplier "X" that is independent of specific process stages, such as the basic cooperation closeness between the two parties, historical interaction frequency, etc. Subsequently, the network calls its stage mapping component. This component receives an intermediate layer output (as a baseline feature representation) generated by the backbone network during processing as input. The stage mapping component consists of two parts: a common mapping component and multiple stage information extraction components corresponding to different process stages. The common mapping component (e.g., a multilayer perceptron) performs a nonlinear transformation on the baseline feature representation, outputting a more refined baseline cross-stage shared feature representation. Next, for each process stage requiring inference (e.g., the "contract approval stage" and the "payment execution stage"), its corresponding stage information extraction component begins operation. Each stage information extraction component contains a set of learnable modulation parameters (e.g., gain and bias parameters). The component takes the baseline cross-stage shared feature representation as input and modulates and transforms it using its unique parameters to generate a baseline stage-specific representation vector. For example, the "contract stage" component outputs a vector focusing on encoding clause risks and legal binding force, while the "payment stage" component outputs a vector focusing on encoding payment rhythm and monetary compliance. These baseline vectors undergo further normalization to ultimately form the target stage-specific representation vector for each process stage. Finally, the network invokes its multiple inference output components, each strictly corresponding to a process stage. Each inference output component receives two inputs simultaneously: first, the aforementioned target cross-stage shared feature representation (providing a general context); and second, the target stage-specific representation vector corresponding to its process stage (providing stage-specific focus). Internally, the component combines these two inputs through a feature fusion layer (e.g., concatenating elements followed by a fully connected layer), and then passes them through a final classification or regression head to output the inference output information for that stage. In this method, this information represents the correlation confidence between the regulated procurement entity and the target procurement document from the perspective of that stage.For example, the output component in the contract approval stage might output a confidence level of 0.88, while the output component in the payment execution stage (even if the current document is a contract) might output a predictive confidence level of 0.75 based on information such as payment terms in the contract. The server collects these concurrent, multi-stage outputs to form a multi-dimensional correlation assessment of a single document.

[0086] In this embodiment of the invention, the step of determining the target stage-specific representation vector for each of the plurality of process reasoning stages by the stage mapping component in the multi-stage collaborative reasoning network based on the contextual representation data of the target purchase order or the baseline feature representation generated when the prior learning backbone network processes the contextual representation data of the target purchase order can be implemented through the following example.

[0087] The benchmark cross-stage shared feature representation is determined by the common mapping component in the stage mapping component, based on the contextual representation data of the target purchase document or the benchmark feature representation.

[0088] Using the attention mechanism in the stage mapping component, the first sub-benchmark cross-stage shared feature representation and the second sub-benchmark cross-stage shared feature representation are determined based on the benchmark cross-stage shared feature representation.

[0089] By using each of the stage information extraction components in the stage mapping component, the baseline stage-specific representation vectors of multiple process reasoning stages corresponding to the stage information extraction component are determined based on the second sub-benchmark cross-stage shared feature representation.

[0090] For each of the multiple process inference stages, the target stage-specific representation vector of the multiple process inference stages is determined based on the cross-stage shared feature representation of the first sub-benchmark and the benchmark stage-specific representation vector of the multiple process inference stages.

[0091] In this embodiment of the invention, exemplarily, the server executes specific operations of the stage mapping component in a multi-stage collaborative inference network. The server instructs the edge inference node to call the stage mapping component to process the baseline feature representation (a 512-dimensional vector) of the target procurement document (such as a draft contract for supplier X) transmitted from the prior learning backbone network. First, the common mapping component in the component processes this baseline feature representation. The common mapping component is a feedforward network containing linear layers and activation functions. It maps the input 512-dimensional baseline feature representation into a 256-dimensional baseline cross-stage shared feature representation. This representation is a high-density encoding of common information valuable to all process stages. Subsequently, the stage mapping component calls its built-in attention mechanism. This mechanism is a lightweight self-attention module that takes the 256-dimensional baseline cross-stage shared feature representation as input and decomposes it into two 128-dimensional sub-vectors through the calculation of the attention head: a first sub-baseline cross-stage shared feature representation and a second sub-baseline cross-stage shared feature representation. The first sub-vector mainly retains information related to the general relevance judgment of each stage; the second sub-vector serves as the basis for subsequently generating stage-differentiated features. Next, the components invoke the information extraction components for each stage in parallel. Each component (e.g., the components corresponding to the "Contract Approval Stage" and the "Payment Execution Stage") receives the same 128-dimensional second sub-benchmark cross-stage shared feature representation as input. Each component contains a set of stage-specific learnable parameters: a gain modulation parameter matrix and a bias modulation parameter vector. The component first performs matrix multiplication, scaling and rotating the input second sub-vector with the gain parameters to obtain modulated intermediate features. Then, it adds the bias parameter vector for translation. After this gain and bias modulation processing, each component outputs a 128-dimensional benchmark stage-specific representation vector. For example, the vector for the "Contract Stage" focuses more on encoding contract type and clause pattern, while the vector for the "Payment Stage" focuses more on encoding payment conditions and risk signals. Finally, for each process stage, the server instruction node performs a vector synthesis operation. This operation concatenates the 128-dimensional first sub-benchmark cross-stage shared feature representation (common part) with the corresponding stage-specific benchmark stage-specific representation vector (characteristic part) to obtain a 256-dimensional composite vector. This composite vector is then processed through a stage-specific normalization layer, ultimately outputting a target-stage-specific representation vector for that stage of the process. This vector contains both shared context across stages and specific information relevant to this stage, providing accurate feature inputs for the independent inference outputs of subsequent stages.

[0092] In this embodiment of the invention, the step of determining the benchmark cross-stage shared feature representation based on the contextual representation data of the target purchase order or the benchmark feature representation through the common mapping component in the stage mapping component can be implemented through the following example.

[0093] The compression mapping layer in the public mapping component performs compression mapping processing on the context representation data of the target purchase order or the baseline feature representation to obtain the baseline compressed mapping feature.

[0094] The baseline compressed mapping features are obtained by performing feature activation processing on the baseline compressed mapping features through the feature activation layer in the common mapping component;

[0095] The benchmark mapping features are reconstructed through the reconstruction mapping layer in the public mapping component to obtain the benchmark cross-stage shared feature representation.

[0096] In this embodiment of the invention, exemplarily, the server executes specific operations of the common mapping component. The server instructs the edge inference node to call the common mapping component to process the baseline feature representation (a 512-dimensional vector) of the target procurement document (such as a draft contract for supplier X) transmitted from the prior learning backbone network. First, the compression mapping layer in the component processes this baseline feature representation. The compression mapping layer is a fully connected linear layer with a weight matrix dimension of [256, 512]. This layer performs a linear transformation, projecting the input 512-dimensional baseline feature representation into a lower-dimensional 256-dimensional vector space, obtaining the baseline compressed mapping features. This process achieves dimensionality reduction and preliminary information condensation of the features. Subsequently, the feature activation layer processes the baseline compressed mapping features. This layer applies the ReLU (Rectified Linear Unit) activation function, setting all negative elements in the 256-dimensional baseline compressed mapping feature vector to zero, while retaining positive elements unchanged. This nonlinear transformation introduces the nonlinear expressive power of the model, resulting in the baseline mapping features. Finally, the reconstruction mapping layer processes the baseline mapping features. The reconstruction mapping layer is another fully connected linear layer with a weight matrix of dimension [256, 256]. This layer performs further linear transformations and information integration on the activated 256-dimensional baseline mapping features, ultimately outputting a new 256-dimensional vector, which is the baseline cross-stage shared feature representation. This representation is a refined common feature formed after compression, activation, and reconstruction of the original baseline features, and will be fed into the subsequent attention mechanism for decomposition.

[0097] In this embodiment of the invention, the step of determining the benchmark stage-specific representation vector of multiple process reasoning stages corresponding to the stage information extraction component based on the second sub-benchmark cross-stage shared feature representation through each of the stage information extraction components in the stage mapping component can be implemented through the following example.

[0098] By extracting the gain modulation parameters from the stage information extraction component, the second sub-reference cross-stage shared feature representation is subjected to gain modulation processing to obtain the reference gain modulation feature representation.

[0099] By using the bias modulation parameters in the stage information extraction component, the reference gain modulation feature representation is subjected to bias modulation processing to obtain the reference stage-specific representation vector for the multiple process inference stages.

[0100] In this embodiment of the invention, exemplarily, the server executes the specific operations of the stage information extraction component. The server instructs the edge inference node to call the stage information extraction component corresponding to the "contract approval stage". This component receives a second sub-benchmark cross-stage shared feature representation from the attention mechanism, which is a 128-dimensional vector. First, the component modulates the input using its internal gain modulation parameters. The gain modulation parameters are trainable 128x128-dimensional matrices. The component performs matrix multiplication, multiplying the input 128-dimensional vector by the matrix, selectively amplifying, reducing, or rotating the input features in each dimension to obtain a transformed 128-dimensional benchmark gain modulation feature representation. This process adapts the feature representation to the specific analytical needs of the "contract approval" stage. Subsequently, the component further processes the benchmark gain modulation feature representation using its internal bias modulation parameters. The bias modulation parameters are trainable 128-dimensional vectors. The component performs vector addition, adding the benchmark gain modulation feature representation element-wise to the bias parameter vector, achieving an overall translation of the feature representation. After this bias modulation processing, a 128-dimensional vector is finally output, which is the baseline stage-specific representation vector corresponding to the reasoning stage of the "contract approval" process. This vector already contains specific feature information for contract review. For other stages such as the "payment execution stage," the server instruction node calls its corresponding stage information extraction component in parallel. Each component uses its own unique gain and bias parameters to independently modulate the same second sub-baseline cross-stage shared feature representation, generating the baseline stage-specific representation vector for its respective stage.

[0101] In this embodiment of the invention, the prior learning backbone network includes multiple sub-feature extraction units, and the multi-stage collaborative reasoning network includes stage mapping components matched by the multiple sub-feature extraction units respectively.

[0102] The method of determining the target cross-stage shared feature representation based on the contextual representation data of the target procurement document through the prior learning backbone network in the multi-stage collaborative reasoning network can be implemented through the following example.

[0103] Through the prior learning backbone network, each of the sub-feature extraction units determines the target cross-stage shared feature representation generated by the sub-feature extraction unit based on the sub-encoding structure processing information of the sub-feature extraction unit; the sub-encoding structure processing information of the sub-feature extraction unit is the context representation data of the target purchase order, or other target cross-stage shared feature representations generated by the sub-feature extraction unit.

[0104] The step of determining the target stage-specific representation vector for each of the multiple process reasoning stages through the stage mapping component in the multi-stage collaborative reasoning network, based on the contextual representation data of the target purchase document or the baseline feature representation generated when the prior learning backbone network processes the contextual representation data of the target purchase document, includes:

[0105] Through each of the stage mapping components, based on the sub-encoding structure processing information of the stage mapping component, the target stage-specific representation vectors of each of the plurality of process inference stages generated by the stage mapping component are determined; the sub-encoding structure processing information of the stage mapping component is the sub-encoding structure processing information of the sub-feature extraction unit corresponding to the stage mapping component, or the baseline feature representation generated by the sub-feature extraction unit corresponding to the stage mapping component.

[0106] In this embodiment of the invention, exemplarily, the server executes a specific process involving the collaborative work of the prior learning backbone network and the stage mapping components in a multi-stage collaborative inference network. The prior learning backbone network consists of multiple sequentially connected sub-feature extraction units, for example, three units: sub-unit L1, sub-unit L2, and sub-unit L3. Each sub-feature extraction unit has a specific network layer structure (such as a Transformer layer). Simultaneously, the multi-stage collaborative inference network includes stage mapping components that match these sub-feature extraction units, such as mapping component M1 matching L1, component M2 matching L2, and component M3 matching L3. The server instructs the edge inference nodes to invoke the prior learning backbone network to process the contextual representation data of the target purchase order (such as a draft contract for supplier X). The processing is performed sequentially by unit: First, the sub-feature extraction unit L1 processes the input information according to its sub-encoding structure (such as a self-attention layer plus a feedforward network), which is the contextual representation data of the target purchase order. After L1 processing, a feature representation is generated, which is defined as the target cross-stage shared feature representation (denoted as F1) generated by the L1 unit. Next, the sub-feature extraction unit L2 processes its input information according to its sub-coding structure; this input information is the target cross-stage shared feature representation F1 generated by other sub-feature extraction units (i.e., L1). After L2 processing, it generates its own target cross-stage shared feature representation F2. Then, the sub-feature extraction unit L3 processes its input information according to its sub-coding structure, i.e., the target cross-stage shared feature representation F2 generated by L2. After L3 processing, it generates its own target cross-stage shared feature representation F3. Typically, the F3 output by the deepest sub-unit L3 is used as the final target cross-stage shared feature representation determined for the entire backbone network, and is used for subsequent inference output components. Simultaneously, the server instruction node calls the mapping components of each stage in parallel. Each stage mapping component operates independently, but its input information comes from its matched specific sub-feature extraction unit: Stage mapping component M1 processes its input information according to its sub-encoding structure, which is the sub-encoding structure processing information of its corresponding sub-feature extraction unit L1. Specifically, M1 receives the attention weight matrix (a baseline feature representation) calculated by L1's internal self-attention layer when processing the original context data. Based on this information, M1 determines the target stage-specific representation vectors (denoted as V1_contract, V1_payment) for the multiple process reasoning stages it is responsible for generating (such as the contract stage, payment stage). Stage mapping component M2 processes its input information according to its sub-encoding structure, which is the baseline feature representation generated by its corresponding sub-feature extraction unit L2. Specifically, M2 receives the intermediate tensor generated by L2 after processing F1 and before its feedforward network activation function.Based on this information, M2 determines the target stage-specific representation vectors (V2_contract, V2_payment) for each of the multiple process reasoning stages it is responsible for generating. The stage mapping component M3 processes its input information according to its sub-encoding structure; this input information is the baseline feature representation generated by its corresponding sub-feature extraction unit L3 (e.g., the output of the LayerNorm layer in L3). Based on this information, M3 determines the target stage-specific representation vectors (V3_contract, V3_payment) for each of the multiple process reasoning stages it is responsible for generating. Finally, for each process reasoning stage (e.g., the "contract approval stage"), its complete target stage-specific representation vector is generated by the server instruction node aggregating (e.g., weighted summation or concatenation) the stage vectors from all mapping components, i.e., combining V1_contract, V2_contract, and V3_contract to form the final contract stage representation vector. This design allows feature information from different levels of abstraction to contribute to stage-specific reasoning.

[0107] In this embodiment of the invention, the step of determining the inference output information of multiple process inference stages corresponding to the inference output component through each inference output component in the multi-stage collaborative inference network, based on the target stage-specific representation vector of multiple process inference stages corresponding to the inference output component and the target cross-stage shared feature representation, can be implemented through the following example.

[0108] By using each of the inference output components, the inference output information of the multiple process inference stages corresponding to the inference output component is determined based on the target cross-stage shared feature representation generated by each of the sub-feature extraction units and the target stage-specific representation vector of the multiple process inference stages corresponding to the inference output component generated by each of the stage mapping components.

[0109] In this embodiment of the invention, for example, the server executes the final information synthesis step of the inference output component in the multi-stage collaborative inference network. The server instructs the edge inference node to call the inference output component corresponding to the "contract approval stage". This component receives two types of input information: the first type is the target cross-stage shared feature representation generated by each sub-feature extraction unit (L1, L2, L3), namely F1, F2, and F3. The second type is the target stage-specific representation vector generated by each stage mapping component (M1, M2, M3) corresponding to the "contract approval stage", namely V1_contract, V2_contract, and V3_contract. The feature fusion layer inside the component first performs alignment and concatenation operations on these two sets of six input vectors. For example, F1 and V1_contract are fused element-wise, F2 and V2_contract are concatenated, and F3 and V3_contract are weighted and summed to form three fused intermediate features. Subsequently, these three intermediate features are further aggregated (e.g., through an attention pooling layer) to generate a unified, comprehensive feature vector representing the "contract approval stage" perspective. Finally, this comprehensive feature vector is fed into the output layer at the end of the component (typically a multilayer perceptron). The output layer processes this vector and maps the result to a scalar value between 0 and 1 using a sigmoid activation function. This scalar value represents the inference output information for the "contract approval stage" corresponding to this inference output component; in this method, it is specifically represented as the correlation confidence (e.g., 0.88) between the regulated procurement entity (e.g., supplier X) and the target procurement document (e.g., draft contract) at this stage. The server uses this method to acquire the inference output information for all process stages in parallel.

[0110] In this embodiment of the invention, the multi-stage collaborative reasoning network is trained by the following process and can be implemented through the following examples.

[0111] Obtain sample data matched at each of the multiple process inference stages; the sample data includes input samples with configured target values;

[0112] For each of the multiple process inference stages, the untrained multi-stage collaborative inference network determines the training inference output information corresponding to the sample data based on the input samples in the sample data corresponding to the multiple process inference stages; and updates the model parameters of the common mapping component and the stage information extraction component corresponding to the multiple process inference stages, as well as the model parameters of the inference output component corresponding to the multiple process inference stages, in the multi-stage collaborative inference network based on the training inference output information and the target value in the sample data.

[0113] In this embodiment of the invention, exemplarily, the server executes a training process for a multi-stage collaborative reasoning network. The server first acquires a training sample dataset, which is organized for each of the multiple process reasoning stages (e.g., the "contract approval stage" and the "payment execution stage"). For the "contract approval stage," the server acquires a batch of historical purchase contract documents and their contextual data as input samples, and configures a target value labeled by a business expert for each sample. This target value is a binary label (e.g., 1 indicates a strong correlation with the subject of this stage requiring review, 0 indicates a weak correlation or no correlation). For the "payment execution stage," the server similarly acquires a batch of historical payment document data and corresponding labeled target values. The server initializes an untrained multi-stage collaborative reasoning network, which includes a prior learning backbone network, a stage mapping component, and inference output components for each stage. For each training iteration, the server performs the following operations: For an input sample from the "contract approval stage," the server inputs it into the network. The network processes the sample through its internal structure, and finally, the inference output component corresponding to the "contract approval stage" outputs a training inference output, such as a probability value of 0.7 indicating relevance. The server compares this output value with the target value of the sample (e.g., 1) and calculates the loss (e.g., binary cross-entropy loss). Similarly, for the same input sample, the network generates inference outputs for the "payment execution stage" in parallel and calculates the loss against the target value of that sample in the payment stage. Based on the calculated losses for each stage, the server initiates a backpropagation algorithm. This algorithm calculates the loss gradient and uses an optimizer (e.g., Adam) to update the model parameters of specific components in the multi-stage collaborative inference network. Specifically, the updated parameters include: the weights and bias parameters of the common mapping components included in the stage mapping component; the gain modulation and bias modulation parameters of the stage information extraction components corresponding to each process stage (e.g., contract stage, payment stage) in the stage mapping component; and the weights and bias parameters of the inference output components corresponding to each process stage. This targeted update strategy enables the network to simultaneously learn to extract cross-stage shared features and generate stage-specific features, while optimizing the independent inference capabilities of each stage. The server repeats this process until the model converges.

[0114] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned document engine-based procurement process penetration monitoring method. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113.

[0115] To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other, either directly or indirectly. For example, these components can be electrically connected to each other via one or more communication buses or signal lines.

[0116] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for purchase process penetration supervision based on a document engine, characterized in that, include: Obtain the contextual representation data of the target purchase order in the document engine, wherein the target purchase order is selected for the purchase entity to be supervised; The contextual representation data of the target purchase order includes the entity identification data of the procurement entity to be supervised, the feature representation of at least one related purchase order that is related to the procurement entity to be supervised, and the feature representation of the target purchase order. In the edge inference node, the purchase order inference model is invoked to infer the correlation between the procurement entity to be supervised and the target purchase order based on the contextual representation data of the target purchase order, so as to obtain the correlation confidence between the procurement entity to be supervised and the target purchase order. The purchase order reasoning model is pre-trained and optimized in the cloud training sub-cluster and then deployed and migrated to the edge reasoning node. The inference latency of the purchase order inference model in the edge inference node is lower than the inference latency of the purchase order inference model in the cloud training sub-cluster. Based on the confidence level between the procurement entity to be regulated and the target procurement document, a penetrating supervision instruction is executed on the target procurement document.

2. The method of claim 1, wherein, The edge inference node is pre-configured with model acceleration parameters for the purchase order inference model to perform correlation inference; the step of calling the purchase order inference model in the edge inference node based on the contextual representation data of the target purchase order to perform correlation inference between the procurement entity to be supervised and the target purchase order, and obtaining the correlation confidence between the procurement entity to be supervised and the target purchase order, includes: The purchase order inference model is accelerated and executed according to the model acceleration configuration parameters. The inference model for the purchase order after the accelerated configuration execution is invoked. Based on the contextual representation data of the target purchase order, the model infers the correlation between the procurement entity to be supervised and the target purchase order, and obtains the correlation confidence between the procurement entity to be supervised and the target purchase order.

3. The method of claim 2, wherein, The model acceleration configuration parameters include the topology fusion rules between the nodes of each computing unit in the purchase order reasoning model. The accelerated configuration and execution process for the purchase order inference model according to the model acceleration configuration parameters includes: The topology of each computing unit node of the purchase order reasoning model is rewritten and compiled according to the topology fusion rules. The purchase order inference model, after the accelerated configuration execution processing, performs correlation inference between the procurement entity to be monitored and the target purchase order based on the contextual representation data of the target purchase order, and obtains the correlation confidence between the procurement entity to be monitored and the target purchase order, including: The inference model for the procurement document after topology rewriting and compilation is invoked to infer the correlation between the procurement entity to be supervised and the target procurement document based on the contextual representation data of the target procurement document, thereby obtaining the correlation confidence between the procurement entity to be supervised and the target procurement document.

4. The method of claim 1, wherein, The purchase order reasoning model includes a feature extraction component and a feature generation component. The feature generation component includes a representation generation sub-component and an inference output sub-component. The context representation data of the target purchase order is sequentially segmented into multiple sequence feature units after being input into the purchase order reasoning model. The step of invoking the purchase order inference model in the edge inference node, based on the contextual representation data of the target purchase order, to infer the correlation between the procurement entity to be monitored and the target purchase order, and to obtain the correlation confidence between the procurement entity to be monitored and the target purchase order, includes: Obtain the feature representation vector of each sequence feature unit in the context representation data of the target purchase order; In the edge inference node, the feature extraction component is invoked to perform feature relationship modeling on each sequence feature unit based on the feature correlation between the feature representation vectors of each sequence feature unit, so as to obtain the associated feature representation of each sequence feature unit; In the edge inference node, the representation generation sub-component is invoked to generate the generated feature representation of the generated input feature based on the feature correlation between the associated feature representation of each sequence feature unit and the associated feature representation of the generated input feature. In the edge inference node, the inference output sub-component is invoked to generate feature representations based on the generated input features to infer the correlation between the procurement entity to be regulated and the target procurement document, thereby obtaining the correlation confidence between the procurement entity to be regulated and the target procurement document.

5. The method according to claim 4, characterized in that, The inference output subcomponent is generated by updating the structure of the category determination subcomponent using the confidence calculation subcomponent; the step of calling the inference output subcomponent in the edge inference node based on the generated feature representation of the generated input features to infer the correlation between the procurement entity to be supervised and the target procurement document, and obtaining the correlation confidence between the procurement entity to be supervised and the target procurement document, includes: In the edge inference node, the category determination subcomponent is invoked to determine the association category between the procurement entity to be supervised and the target procurement document based on the generated feature representation of the generated input features, thereby obtaining the output association category identifier between the procurement entity to be supervised and the target procurement document; In the edge inference node, the confidence calculation subcomponent is invoked to perform association confidence calculation on the procurement entity to be supervised and the target procurement document based on the generated feature representation of the generated input features, and to obtain multiple undetermined confidence values ​​under the output association category identifier; Select the correlation confidence level between the procurement entity to be supervised and the target procurement document from among the multiple undetermined confidence levels.

6. The method according to claim 4, characterized in that, The step of obtaining the feature representation vectors of each sequence feature unit in the context representation data of the target purchase order includes: Obtain a business field feature mapping table, which includes feature representation vectors of multiple standardized business fields; Retrieve the feature representation vector of each sequence feature unit in the context representation data of the target purchase order from the business field feature mapping table; If a standardized business field that matches the index of the first sequence feature unit in the context representation data of the target purchase order is found in the business field feature mapping table, then the feature representation vector of the standardized business field that matches the index is determined as the feature representation vector of the first sequence feature unit. The second sequence feature unit, excluding the first sequence feature unit, in the context representation data of the target purchase order is subjected to feature encoding processing to obtain the feature representation vector of the second sequence feature unit.

7. The method according to claim 1, characterized in that, The document engine filters multiple pending purchase documents for the procurement entity to be supervised, and the target purchase document is any one of the multiple pending purchase documents; the multiple pending purchase documents belong to at least two different procurement process stages of the document engine. At least two purchase order reasoning models are configured in the edge reasoning node, and the at least two purchase order reasoning models correspond one-to-one with at least two procurement process stages; The method further includes: In the edge inference node, the at least two purchase order inference models are invoked to perform correlation inference between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage based on the contextual representation data of each pending purchase order in the corresponding procurement process stage, so as to obtain the correlation confidence between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage. Among them, the correlation inference of the at least two purchase order reasoning models is processed concurrently.

8. The method according to claim 7, characterized in that, The at least two purchase order inference models share a multi-stage collaborative inference network structure; the step of calling the at least two purchase order inference models at the edge inference node to infer the correlation between the procurement entity to be supervised and the pending purchase orders in the corresponding procurement process stage based on the contextual representation data of each pending purchase order in the corresponding procurement process stage includes: Obtain the contextual representation data of the target purchase order; the target purchase order is any one of the pending purchase orders in the corresponding procurement process stage; The target cross-stage shared feature representation is determined by the prior learning backbone network in the multi-stage collaborative reasoning network based on the contextual representation data of the target purchase order; the multi-stage collaborative reasoning network is used to execute multiple process reasoning stages based on the contextual information of the input document. The stage mapping component in the multi-stage collaborative reasoning network determines the target stage-specific representation vectors for each of the multiple process reasoning stages based on the contextual representation data of the target procurement document or the baseline feature representation generated when the prior learning backbone network processes the contextual representation data of the target procurement document. The stage mapping component includes a common mapping component and stage information extraction components matched to each of the multiple process reasoning stages. The common mapping component is used to extract the baseline cross-stage shared feature representation, and the stage information extraction component is used to extract the baseline stage-specific representation vectors for the corresponding multiple process reasoning stages based on the baseline cross-stage shared feature representation. The target stage-specific representation vectors for the multiple process reasoning stages are determined based on the baseline stage-specific representation vectors for the multiple process reasoning stages. Through each inference output component in the multi-stage collaborative inference network, the inference output information of multiple process inference stages corresponding to the inference output component is determined based on the target stage-specific representation vector of multiple process inference stages corresponding to the inference output component and the target cross-stage shared feature representation; the multi-stage collaborative inference network includes inference output components matched to the multiple process inference stages respectively.

9. The method according to claim 8, characterized in that, The step of determining the target stage-specific representation vector for each of the multiple process reasoning stages through the stage mapping component in the multi-stage collaborative reasoning network, based on the contextual representation data of the target purchase document or the baseline feature representation generated when the prior learning backbone network processes the contextual representation data of the target purchase document, includes: The compression mapping layer in the public mapping component performs compression mapping processing on the context representation data of the target purchase order or the baseline feature representation to obtain the baseline compressed mapping feature. The baseline compressed mapping features are obtained by performing feature activation processing on the baseline compressed mapping features through the feature activation layer in the common mapping component; The benchmark mapping features are reconstructed through the reconstruction mapping layer in the public mapping component to obtain the benchmark cross-stage shared feature representation. Using the attention mechanism in the stage mapping component, the first sub-benchmark cross-stage shared feature representation and the second sub-benchmark cross-stage shared feature representation are determined based on the benchmark cross-stage shared feature representation. By extracting the gain modulation parameters from the stage information extraction component, the second sub-reference cross-stage shared feature representation is subjected to gain modulation processing to obtain the reference gain modulation feature representation. By using the bias modulation parameters in the stage information extraction component, the reference gain modulation feature representation is subjected to bias modulation processing to obtain the reference stage-specific representation vector for the multiple process inference stages. For each of the multiple process inference stages, the target stage-specific representation vector of the multiple process inference stages is determined based on the cross-stage shared feature representation of the first sub-benchmark and the benchmark stage-specific representation vector of the multiple process inference stages.

10. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-9.

Images