A government affair reminding and interfacing system based on communication interaction

Abstract

The application relates to the technical field of electronic government affairs, and discloses an enterprise government affair reminding and government affair department interfacing system based on communication interaction, which comprises an enterprise monitoring module that collects multi-source time sequence data and constructs an enterprise knowledge graph integrating the evolution of industrial and commercial affairs, the fluctuation of tax affairs and the accumulation of intellectual property rights; a government affair policy module that analyzes policy texts, extracts conditional elements, constructs a policy knowledge graph, and performs matching analysis on an enterprise state subgraph and a policy condition tree through a constraint graph neural network matcher to obtain personalized government affair reminding; and a coordination module that analyzes a process dependency graph and dynamically deploys an independent smart contract as a state corridor for an affair, sets a state conversion function and requires joint digital signature verification of the current and previous links, and realizes the integration of real-time state dynamic perception of enterprises, intelligent matching of policy conditions and safe and reliable circulation between multiple government affair departments, thereby reducing the cost of artificial integration of government affairs.

Description

Technical Field

[0001] This application relates to the technical field of e-government and discloses an enterprise government affairs reminder and commission docking system based on communication interaction. Background Technology

[0002] In existing enterprise government affairs reminder and commission-department docking systems, enterprise profile construction largely relies on static data snapshots. This makes it difficult to uniformly express the historical evolution of business attributes, the temporal fluctuations in tax status, and the cumulative decay of intellectual property rights into a computable graph structure, thus failing to dynamically reflect the real-time status of enterprises. Regarding policy matching, existing solutions often use simple keyword matching of applicable subjects, conditions, and timeliness elements in unstructured policy texts. They lack a matching mechanism that transforms natural language conditions into structured condition nodes containing condition types, target attribute paths, and constraint expressions, and uses graph neural network reasoning based on policy condition logic trees and enterprise temporal status subgraphs. This results in insufficient accuracy in determining the applicability of policies with complex logical combinations and timeliness windows. In terms of cross-commission collaboration, existing workflow systems typically rely on centralized servers, requiring manual configuration of process dependencies. They cannot automatically extract sequence, branching, and convergence relationships from service guides and use historical logs for probability correction. Status updates only verify unilateral signatures, lacking mandatory on-chain verification of joint digital signatures between the current and previous departments, making it difficult to prevent unauthorized jumps. Furthermore, a mechanism for dynamically deploying independent state machine instances for individual matters to achieve matter isolation and timeout self-driving has not yet been disclosed. In addition, reminder strategies often use fixed channels for batch pushes, which do not form a closed loop with collaborative process status monitoring. They cannot be optimized online based on enterprise feedback and automatically trigger the reminder status change when there is no response after the timeout. Summary of the Invention

[0003] The purpose of this section is to outline some aspects of the embodiments of this application and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents, and such simplifications or omissions should not be construed as limiting the scope of this application.

[0004] To address the aforementioned technical issues, this application provides a system for enterprise government affairs reminders and inter-departmental communication based on communication interaction.

[0005] On the one hand, this application provides a communication-interactive enterprise government affairs reminder and commission docking system, including an enterprise monitoring module that collects multi-source time-series data of enterprises through data interfaces and constructs an enterprise knowledge graph;

[0006] The government policy module parses government policy texts, extracts applicable subjects, conditions and timeliness elements to construct a policy knowledge graph, and performs a matching analysis to solve for constraint satisfaction, thereby obtaining the personalized government reminders;

[0007] When the personalized government affairs reminder involves multiple commissioning departments, the collaboration module parses the process dependency graph of the government affairs and deploys an independent smart contract instance as a state corridor instance for the current matter. The contract instance sets a state transition function. Any commissioning department must call the state transition function to update the status of the matter. Before the call, it must be verified by the joint digital signature of the current node and the previous node operator. Any state change record generates an immutable on-chain record.

[0008] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0009] The enterprise monitoring module includes a data acquisition unit and a graph unit;

[0010] The data acquisition unit is used to collect the multi-source time-series data, including business registration changes, tax declarations, and intellectual property information.

[0011] The enterprise knowledge graph constructed by the graph unit using the multi-source time-series data is used to reflect the real-time status of the enterprise.

[0012] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0013] The method for constructing the enterprise knowledge graph using the graph unit is as follows:

[0014] The business registration change record is transformed into a quadruple of (enterprise, attribute name, old value, new value, timestamp). A change event node is generated for each quadruple, and the version evolution is represented by directed edges. Adjacent change nodes with the same attribute are connected in series. The vector representation of the change event node is generated by weighting the difference vector between the old and new values ​​and the time decay factor.

[0015] Using the month as the smallest granularity, the time series values ​​of tax declarations are aggregated by a sliding window. Each window generates a tax status node. The attributes of the tax status node include the average tax credit rating within the window, the number of overdue declarations, and the ratio of tax payable to actual tax payable. Tax status nodes of adjacent windows are connected by time-adjacent edges, and the weight of the edge is the Jaccard similarity between the two windows.

[0016] Cumulative counter nodes are established for patents, trademarks, and software copyrights respectively. The counter node value increases with each authorization announcement event. Each authorization event generates an event node. The event node carries the authorization number, authorization date, and classification number. The event node points to the cumulative counter node through a contribution edge. The weight of the edge decreases linearly with the time elapsed since the authorization date.

[0017] Centered on the enterprise entity node identified by the unified social credit code, change event nodes, tax status nodes, authorization event nodes, and cumulative counter nodes are connected through relationship edges.

[0018] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0019] The government policy module collects unstructured or semi-structured original policy texts;

[0020] The original policy texts include announcements, measures, implementation rules, and application guidelines;

[0021] The collected original policy texts are cleaned and segmented into chapters and levels, and policy discourse units are output by title, chapter, and paragraph levels.

[0022] A semantic role labeling model is constructed by a self-attention encoding layer, a dual-path feature extraction layer, and a conditional random field decoding layer to process each policy discourse unit, identify and extract the applicable subject, qualification conditions, and timeliness elements.

[0023] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0024] The matching analysis for satisfying the constraints includes: expanding the enterprise entity nodes and their adjacent attribute nodes, event nodes, and temporal state nodes in the enterprise knowledge graph into an enterprise state subgraph.

[0025] Based on the target path field of the conditional expression structure encapsulated inside the conditional entity node in the policy knowledge graph, locate the corresponding attribute node or time sequence node in the enterprise state subgraph, and extract the current snapshot value and historical sequence.

[0026] Construct a two-stage matcher implemented using a constraint graph neural network.

[0027] As a preferred embodiment of the enterprise government affairs reminder and commission docking system based on communication interaction proposed in this application, the two-stage matcher includes a first stage and a second stage;

[0028] In the first stage, each condition node in the policy condition tree is encoded as a query vector, and the nodes located in the enterprise state subgraph and their neighbors are encoded as key vectors. The local matching score between the condition node and the enterprise node is calculated through cross attention.

[0029] The second stage injects the logical structure of the policy condition tree into the graph attention network, aggregates local matching scores from bottom to top, uses minimum pooling for logical AND nodes and maximum pooling for logical OR nodes, and calculates the time series matching degree by modeling the historical sequence of the time window node through a recurrent neural network and fusing it with the time window parameters, and outputs the overall matching score.

[0030] The overall matching score is compared with a preset threshold to classify enterprises as fully compatible, partially compatible, or hard-conflicting.

[0031] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0032] When the matching analysis result indicates that the enterprise is fully or partially matched, the reminder strategy decision-maker of the context multi-armed gambling machine model is triggered. The reminder strategy decision-maker concatenates the enterprise activity features, historical message interaction feedback features, and urgency and complexity features of government affairs extracted from the enterprise knowledge graph into a context vector, estimates the expected reward for each candidate reminder channel and reminder timing combination, and samples to select the optimal reminder action.

[0033] After generating the reminder message body, the message body hash, enterprise identifier, and reminder action metadata are assembled into a reminder event. The reminder event is recorded on the blockchain through an independent smart contract instance of the collaboration module, and the reminder event is bound to the status corridor instance of the associated government affairs, and a status monitoring timer is started.

[0034] If no response operation from any commissioning department is detected within the timer window, the state transition function of the state corridor instance is triggered, updating the status of the matter from "reminded" to "pending reminder", and notifying the reminder strategy decision-maker to perform secondary reminder optimization.

[0035] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0036] From standardized service guidelines or internal procedures for government affairs, the text describing each step is extracted using a text structure analysis model. Then, through dependency parsing and semantic role labeling, the sequential dependencies, parallel branches, conditional branches, and synchronous convergence relationships between steps are extracted to generate the intermediate process structure.

[0037] The intermediate process structure is transformed into a directed graph representation, where nodes represent the responsibilities of the commissioning departments, edges represent the flow direction of matters, and each edge is attached with a logical expression of the preconditions, which references the condition node identifiers in the policy knowledge graph.

[0038] The directed graph is aligned with process instances extracted from historical case logs. The graph structure is then corrected and probabilistically labeled using a process mining algorithm to obtain the final process dependency graph. Each edge in the process dependency graph also carries a probability distribution of the average processing time estimated based on historical data.

[0039] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0040] Each link in the process dependency graph is mapped to a state enumeration value in a smart contract, and the flow edge between each link is mapped to a state transition function. The transition conditions of the state transition function are encoded as on-chain verifiable assertions of the preconditions.

[0041] During the creation of a state corridor instance, based on the probability distribution of the average processing time, a maximum dwell time limit is preset for each state. If the current state dwell time exceeds the maximum dwell time limit, the state corridor instance automatically triggers the built-in timeout state transition, transfers the matter to the exception processing state, and sends an on-chain event to the associated commissioning department.

[0042] When the state corridor instance is deployed, the contract address is bound to the unique identifier of the government affairs matter and recorded in the contract registry of the collaboration module. At the same time, a matter creation event is published to the enterprise monitoring module, triggering the pre-population of relevant information on the enterprise side.

[0043] As a preferred embodiment of the enterprise government affairs reminder and commission-department docking system based on communication interaction proposed in this application, wherein:

[0044] When the state transition function is called, the caller is required to submit the current state hash signed by the private key of the current processing department node, and the proof that the previous state has been completed signed by the private key of the previous department node in the process dependency graph.

[0045] The verification process first involves performing public key recovery and identity verification on the chain for the private key signature of the current processing node and the private key signature of the previous node.

[0046] After successful verification, the state transition function updates the state variables and stores the composite state change record, which includes the current state, proof of the previous state completion, signatures of both parties, timestamp, and operation hash, as an immutable on-chain log.

[0047] If any party's signature is missing or invalid, the state transition function rolls back and records a violation attempt event, while simultaneously triggering a security notification to other commissioning departments.

[0048] The beneficial effects of this application are as follows:

[0049] This application collects multi-source time-series data such as business registration changes, tax declarations, and intellectual property through an enterprise monitoring module, and constructs an enterprise knowledge graph that integrates attribute version evolution, tax time-series fluctuations, and intellectual property accumulation and decay, thereby achieving a dynamic depiction of the real-time status of the enterprise.

[0050] This application segments policy texts into chapters and labels them with semantic roles through the government policy module, extracts applicable subjects, conditions and timeliness elements and encapsulates them into an executable condition structure to construct a policy knowledge graph. It uses a constraint graph neural network two-stage matcher to perform cross-attention calculation and logical pooling reasoning on the enterprise state subgraph and the policy condition tree, and outputs adaptive classification results, which improves the accuracy of policy matching with complex logical combinations and timeliness windows.

[0051] This application uses a collaborative module to automatically parse the process dependency graph from the service guide and performs probability correction using historical logs. It dynamically deploys independent smart contract instances for each matter as state corridors to achieve matter isolation and timeout self-driving. When the state is updated, it mandates joint digital signature verification between the current node and the previous node. The composite record containing both parties' signatures and state hashes is immutably stored on the blockchain, ensuring the mandatory order of cross-departmental flow and full auditability. The reminder strategy adopts a contextual multi-armed gambling machine model to optimize channels and timing online. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained through these drawings without creative effort. Wherein:

[0053] Figure 1 This application provides an overall business process diagram of an enterprise government affairs reminder and commission docking system based on communication interaction;

[0054] Figure 2 This application provides a flowchart of the enterprise monitoring module of an enterprise government affairs reminder and commission docking system based on communication interaction;

[0055] Figure 3 This application provides a flowchart of the core matching and reminder generation process for the government policy module of a communication-interactive enterprise government affairs reminder and commission-department docking system;

[0056] Figure 4 This application provides a flowchart of the collaborative module status corridor and joint signature verification process for an enterprise government affairs reminder and commission docking system based on communication interaction;

[0057] Figure 5This is a schematic diagram of the policy matching result display interface of a communication-interactive enterprise government affairs reminder and commission docking system provided for this application. Detailed Implementation

[0058] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0060] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0061] Example 1

[0062] like Figure 1 As shown, a communication-interactive enterprise government affairs reminder and commission docking system includes:

[0063] The enterprise monitoring module collects multi-source time-series data of enterprises through data interfaces and constructs an enterprise knowledge graph;

[0064] The enterprise monitoring module includes a data acquisition unit and a graph unit;

[0065] The data acquisition unit is used to collect the multi-source time-series data, including business registration changes, tax declarations, and intellectual property information.

[0066] The data acquisition unit can also collect multi-source time-series data from enterprises on a regular or incremental basis by connecting to the data interfaces opened by market supervision departments, tax departments and intellectual property management departments.

[0067] Specifically, this includes: business registration change records, such as details of changes to registered capital, business scope, shareholder structure, and other attributes; tax declaration records, including tax credit rating, declaration status of each type of tax, overdue records, and actual amount paid; and intellectual property information, covering legal status events such as authorization announcements for patents, trademarks, and software copyrights, changes in bibliographic data, and termination of rights.

[0068] The enterprise knowledge graph constructed by the graph unit using the multi-source time-series data is used to reflect the real-time status of the enterprise.

[0069] The method for constructing the enterprise knowledge graph using the graph unit is as follows:

[0070] The business registration change record is transformed into a quadruple of (enterprise, attribute name, old value, new value, timestamp). A change event node is generated for each quadruple, and the version evolution is represented by directed edges. Adjacent change nodes with the same attribute are connected in series. The vector representation of the change event node is generated by weighting the difference vector between the old and new values ​​and the time decay factor.

[0071] Using the month as the smallest granularity, the time series values ​​of tax declarations are aggregated by a sliding window. Each window generates a tax status node. The attributes of the tax status node include the average tax credit rating within the window, the number of overdue declarations, and the ratio of tax payable to actual tax payable. Tax status nodes of adjacent windows are connected by time-adjacent edges, and the weight of the edge is the Jaccard similarity between the two windows.

[0072] Cumulative counter nodes are established for patents, trademarks, and software copyrights respectively. The counter node value increases with each authorization announcement event. Each authorization event generates an event node. The event node carries the authorization number, authorization date, and classification number. The event node points to the cumulative counter node through a contribution edge. The weight of the edge decreases linearly with the time elapsed since the authorization date.

[0073] Centered on the enterprise entity node identified by the unified social credit code, change event nodes, tax status nodes, authorization event nodes, and cumulative counter nodes are connected through relationship edges.

[0074] In this application, a preferred implementation method for a graph unit includes: graph modeling in the dimension of business registration change, graph modeling in the dimension of taxation, graph modeling in the dimension of intellectual property, and multi-dimensional modeling of enterprise entity nodes;

[0075] The graph modeling of the business registration change dimension transforms each business registration change record into a quadruple of (enterprise, attribute name, old value, new value, timestamp).

[0076] For example, if a company's registered capital changes from 5 million yuan to 10 million yuan, a quadruple (XX Company, Registered Capital, 5 million, 10 million, 2024-03-15) is generated. The graph unit creates a change event node for each quadruple in the graph. The change event node stores the attribute values ​​and time information before and after the change. For the same attribute of the same company, adjacent change event nodes are connected end to end by the evolving directed edges to form an attribute evolution chain, so that the attribute state at any historical point in time can be traced along the chain.

[0077] The vector of the change event node is generated by weighting the difference vector between the old and new values ​​with a time decay factor. When the attribute value is numerical, the difference vector is normalized to the [0,1] interval by the minimum and maximum values ​​of (new value minus old value), and then mapped to a 128-dimensional embedding vector through a fully connected layer. When the attribute value is textual, the difference vector is obtained by subtracting the semantic encoding vector of the new value text from the semantic encoding vector of the old value text element by element. The semantic encoding vector can be extracted by a pre-trained language model. The difference vector is used to encode the magnitude and direction of the attribute change. The physical meaning of the time decay factor is that the more distant the change, the weaker its impact on the current state. Mathematically, it can be expressed as an exponential decay function e. Where λ is the decay coefficient, which can be 0.5 / year, and t is the time interval from the occurrence of the change event to the present, in years. The node vector embedding is obtained by multiplying the difference vector by this time decay factor, i.e. This results in a larger vector representation of recent changes and a suppressed magnitude of changes in the future. The vector embedding of the current node is obtained after weighted summation.

[0078] The graph modeling of the tax dimension uses the natural month as the smallest granularity and performs sliding window aggregation on the time series data of tax declarations. Both the window step size and the window width can be configured. A preferred example can be set to use the month as the step size and the quarter as the window width. Within each window, the average tax credit rating, the cumulative number of overdue declarations, and the ratio of tax payable to tax paid are calculated for each month in the current window. A corresponding tax status node is generated for each window.

[0079] A preferred method for calculating the average tax credit rating, cumulative number of overdue tax returns, and ratio of tax payable to tax paid for each month within the current tax window includes:

[0080] Average Tax Credit Rating: Let the window contain n months, and the tax credit rating scores for each month be as follows: , ,…, Where A corresponds to a score of 90, B to 70, C to 50, and D to 30, the average tax credit rating for this window is the arithmetic mean of the scores for each month, i.e., AvgScore = Then, the AvgScore is mapped back to the rating range to determine the corresponding credit rating label.

[0081] Cumulative number of late applications: This is the sum of the number of late applications that occurred in each month within the specified window. ,in This represents the number of overdue declarations in the i-th month.

[0082] The ratio of tax payable to tax paid: sum the tax payable and tax paid for each month in the window, then divide by the sum. ,in This represents the actual tax paid in the i-th month. Let this be the tax payable for the i-th month; if If the ratio is 0, then the ratio is 1.0.

[0083] The three calculated values ​​mentioned above together constitute the attribute vector of the tax status node corresponding to the current window.

[0084] Tax status nodes in adjacent windows are connected sequentially by directed edges, with the edge weights calculated using the Jaccard similarity of the attribute value sets of the two windows. For vectors composed of continuous attribute values, the Jaccard similarity is calculated using the generalized Jaccard coefficient. ,in and The values ​​of the i-th attribute dimension in the two windows are respectively summed and iterated through the three dimensions: average tax credit rating, cumulative number of overdue declarations, and actual tax payment ratio.

[0085] In this application, by setting Jaccard similarity, adjacent windows with drastic fluctuations in tax status are given lower edge weights, while windows with stable periods are given higher weights, thus ensuring the stability of corporate tax behavior encoded in the graph structure.

[0086] The graph modeling of intellectual property rights establishes cumulative counter nodes for three types of intellectual property rights: patents, trademarks, and software copyrights. The numerical attribute of the counter node increases with each authorization announcement event, reflecting the total cumulative amount of the enterprise in various types of intellectual property rights.

[0087] Furthermore, each specific authorization event generates an independent event node in the graph, whose attributes include authorization number, authorization announcement date, and category number. The event node points to the corresponding cumulative counter node via a contribution edge. The weight of the contribution edge decreases linearly with the time elapsed since the authorization date, i.e., edge weight w = max(0, ...). ), where t is the time elapsed since then, and T is a preset decay period constant. For example, in this embodiment, T is 5 years, that is, the contribution weight of intellectual property rights that have been authorized for 5 years to the current cumulative counter node decays to 0. Since the value of intellectual property rights is time-sensitive, the representation effect of rights obtained earlier on the current innovation capability of enterprises is relatively weakened. This application uses edge weight decay to enable the graph neural network to give higher attention weight to recently authorized events when aggregating neighbor information, so as to more accurately reflect the current intellectual property strength of enterprises.

[0088] The multi-dimensional aggregation of enterprise entity nodes uses the unified social credit code as the unique identifier of the enterprise. An enterprise entity node is created for each enterprise, serving as the central anchor point of the graph. Through semantic relationship edges such as occurrence, generation, and holding, these nodes are connected to the change event node, the tax status node, the authorization event node, and the cumulative counter node, forming a heterogeneous graph structure centered on the enterprise and integrating three types of time-series information: business history, tax fluctuations, and intellectual property accumulation. This heterogeneous graph structure reflects the current snapshot state of the enterprise and preserves a multi-dimensional time-series evolution trajectory, such as... Figure 2 The diagram shown is a flowchart of the enterprise monitoring module.

[0089] For example: when a new business registration change record is detected, its quadruple is extracted and a new change event node is created; the current version evolution chain tail node of the enterprise for this attribute in the graph is retrieved, and a new version evolution directed edge from the tail node to the new node is added, and the new node is attached as the new chain tail; at the same time, the relationship edge between the enterprise entity node and the new event node is updated.

[0090] After each natural month, once a new tax return data is generated, the sliding window moves forward one step, calculates the attribute values ​​of the new window, and generates a new tax status node; establishes temporal adjacent edges from the previous tax status node to the new node, and calculates Jaccard similarity as edge weights.

[0091] When a new patent, trademark, or software copyright authorization announcement event is detected, create the event node in the graph, retrieve the corresponding type of cumulative counter node, and increment its numerical attribute by 1; establish a contribution edge from the new event node to the counter node, and calculate the edge weight;

[0092] The government policy module parses government policy texts, extracts applicable subjects, conditions and timeliness elements to construct a policy knowledge graph, and performs a matching analysis to solve for constraint satisfaction, thereby obtaining the personalized government reminders;

[0093] The government policy module collects unstructured and semi-structured original policy texts;

[0094] The government policy module proactively collects original policy texts publicly released by government departments at all levels through distributed networks and open government data interfaces on their official portals, policy release platforms, and government new media channels. The collected original policy texts cover common document types such as announcements, regulations, implementation rules, and application guidelines, and are in unstructured or semi-structured formats including HTML web pages, PDF attachments, and Word documents. Each collected policy document records its source URL, collection timestamp, and document summary hash value, forming a collection traceability certificate and providing a basis for verifying the authenticity and version consistency of subsequent policy texts.

[0095] The original policy texts include announcements, measures, implementation rules, and application guidelines;

[0096] The collected original policy texts are cleaned and segmented into chapters and levels, and policy discourse units are output by title, chapter, and paragraph levels.

[0097] Because the original policy texts have differences in formatting, redundant characters, and cross-paragraph citations, this application first performs uniform text cleaning processing on the collected texts, including removing HTML tags, standardizing line break encoding, merging line breaks across pages, and removing irrelevant content such as headers and footers.

[0098] After cleaning, the policy text enters the chapter-level segmentation stage. This segmentation stage uses the unique formatting features of policy documents as segmentation clues, including chapter and clause identifiers starting with Chapter X, Article X, hierarchical numbers appearing in the form of I, (I), etc., and layout signals such as font size changes or first-line indentation between titles and body text. For PDF policy texts, document structure parsing tools are used to extract the font name, font size, and X-coordinate offset of each text block, identifying text blocks with significantly increased font size or X-coordinate offset. The segmentation process identifies four levels from top to bottom: title, chapter, clause, and paragraph. Formatting identifiers are identified using regular expression matching. Matching targets include: [1 2 3 4 5 6 7 8 9 10 100 1000] + chapter, [0-9] + clause, and other level identifier patterns. Layout signals such as first-line indentation and font size abrupt changes are used to generate corresponding policy text units for each level. A tree of inclusion and sequential relationships between units is established. This tree structure preserves the original logical organization of the policy text and avoids erroneous confusion of information across chapters.

[0099] A semantic role labeling model is constructed by a self-attention encoding layer, a dual-path feature extraction layer, and a conditional random field decoding layer to process each policy discourse unit, identify and extract the applicable subject, qualification conditions, and timeliness elements.

[0100] Each policy discourse unit is processed using a semantic role labeling model constructed from a self-attention encoding layer, a dual-path feature extraction layer, and a conditional random field decoding layer. This model automatically identifies and extracts three key semantic elements of the policy information in the discourse unit: applicable subject, qualification conditions, and timeliness.

[0101] The automatic identification and extraction refers to transforming semantic role labeling into a sequence labeling task, specifically through the following method:

[0102] The input policy discourse units are encoded at the character level by a self-attention encoding layer, and the contextual semantic vector of each character is output.

[0103] The dual-path feature extraction layer fuses the sequence features and local features of the encoding result to generate a fused feature matrix;

[0104] The conditional random field decoding layer performs character-by-character label prediction on the fused feature matrix. The label system adopts the BIO annotation mode, where B represents the starting character of the feature mention, I represents the internal character of the feature mention, and O represents the non-feature character.

[0105] For the tag sequence obtained by annotation, the character fragments corresponding to consecutive and similar B and I tags are concatenated into a complete semantic element mention text;

[0106] The applicable subject mentions, qualification condition mentions, and timeliness mentions extracted from the same discourse unit are combined into a triple. If a certain type of element does not appear in the discourse unit, the corresponding field is set to null.

[0107] The triple is mounted as an attribute field of the discourse unit node.

[0108] The self-attention encoding layer employs the BERT architecture language model to perform deep semantic encoding on the input policy discourse units. This layer uses a multi-head self-attention mechanism to capture long-distance dependencies between characters within the discourse unit, ensuring that the vector representation of each character incorporates the global contextual information of the sentence. For example, the semantic association between the applicant company and the requirements can be enhanced through self-attention weights, even if they are separated by several characters in the text.

[0109] In this application, a preferred example of the BERT architecture language model can initialize the model parameters using BERT-base-Chinese pre-trained weights. The BERT architecture language model contains a 12-layer Transformer encoder, with 12 attention heads per layer, a hidden layer dimension of 768, and a maximum input sequence length of 512 characters. The input sequence begins with the special symbol [CLS] and ends with [SEP]. The final hidden vector at the [CLS] position serves as the global semantic representation of the discourse unit. During the fine-tuning stage, end-to-end training is performed on a semantic role labeling task, with AdamW selected as the optimizer and a learning rate of 2. The batch size is 16, the training rounds are 10 epochs, and early stopping is selected on the validation set based on the F1 score. The self-attention encoding layer uses a multi-head self-attention mechanism to perform parallel calculation of the semantic association strength between any two characters in the input sequence.

[0110] The dual-path feature extraction layer is located above the encoding layer and consists of parallel bidirectional long short-term memory (LSTM) and dilated convolutional network feature extraction paths. The bidirectional LSM is used to capture sequence dependency features along the text's line order and identify grammatically coherent conditional statements. The dilated convolutional network performs multi-scale semantic convolution on the local context using convolutional kernels with different dilation rates. For example, the dilated convolutional network consists of three parallel dilated convolutional layers with a kernel size of 3 and dilation rates of 1, 2, and 3, respectively, to capture local semantic collocations with a spacing of 1-3 characters. Each layer has 256 output channels, which are concatenated to form a multi-scale feature matrix.

[0111] The outputs of the bidirectional long short-term memory network and the dilated convolutional network are weighted and merged position by position through a gating fusion mechanism.

[0112] In this application, a preferred calculation method includes:

[0113] Let the output feature matrix of the bidirectional long short-term memory network be... The output feature matrix of the dilated convolutional network is set as Calculate the gate vector fusion features Where [;] represents concatenation along the feature dimension, ⊙ represents element-wise multiplication, and σ is the sigmoid activation function. and These are learnable parameters.

[0114] It obtains a fusion feature representation that includes sequential continuity and local saliency, effectively taking into account the stylistic features of policy texts where long conditional statements and short trigger expressions coexist.

[0115] The Conditional Random Field (CRF) decoding layer performs sequence labeling decoding on the fused feature matrix. The fused feature matrix first passes through a fully connected layer, which maps the fused vector at each position to a fractional vector of the same length as the number of label categories, i.e., the emission fractional matrix. The emission fractional matrix and the state transition matrix are input into the CRF layer, and the Viterbi algorithm is used to solve for the optimal label sequence. The labeling system includes subject-type labels, such as the applicant and the supported object; condition-type labels, such as numerical conditions, qualification conditions, and existence conditions; and time-sensitive labels, such as effective time, deadline, and validity period.

[0116] Furthermore, the Conditional Random Field model outputs the globally optimal label sequence through transition constraints between adjacent labels. The decoding process can use the Viterbi algorithm to solve for the optimal label path under the constraints of the emission matrix and the state transition matrix, thus avoiding illegal label jumps that may occur during isolated bit-by-bit classification.

[0117] Furthermore, the Conditional Random Field (CRF) model extracts structured semantic element triples from each policy discourse unit: (applicable subject mention, qualification condition mention, and timeliness mention). These triples are attached to the corresponding nodes in the discourse hierarchy tree along with their respective policy discourse units. The attachment treats the applicable subject mention, qualification condition mention, and timeliness mention in the triple as the subject attribute field, condition attribute field, and timeliness attribute field of the corresponding policy discourse unit node, respectively, and stores them in the node's data structure as key-value pairs. This allows the semantic elements directly associated with a node to be indexed from any node in the discourse hierarchy tree, and child nodes inherit the subject attribute field of the parent node by default, unless the child node explicitly specifies a new applicable subject.

[0118] The matching analysis for satisfying the constraints includes: expanding the enterprise entity nodes and their adjacent attribute nodes, event nodes, and temporal state nodes in the enterprise knowledge graph into an enterprise state subgraph.

[0119] Specifically, starting from the enterprise entity node, the system traverses outwards along the relationship edges connecting the change event node, the tax status node, the authorization event node, and the cumulative counter node. The traversal depth can be preset to two hops. The first hop is from the enterprise entity node to the event node and status node of each dimension. The second hop is from the event node to its associated version evolution node or from the tax status node to the adjacent window node along the temporal adjacent edge.

[0120] Extract all nodes and edges traversed during the traversal to form the enterprise state subgraph.

[0121] Furthermore, each condition entity node in the policy knowledge graph encapsulates a condition expression structure, whose target path field is a string describing the path from the enterprise entity node to the target attribute. A preferred format example is: / tax status[current] / average tax credit rating, which points to the average tax credit rating attribute of the tax status node in the current window.

[0122] / Intellectual Property / Patents / Cumulative Counter, pointing to the current count value of the patent cumulative counter node;

[0123] / Tax Status[*] points to all historical tax status nodes and their time series.

[0124] When the matching engine parses the target path field, it starts from the enterprise entity node in the enterprise status subgraph and performs matching traversal according to the relationship name and node type of each segment in the path to locate the target attribute node or time sequence node.

[0125] The extraction method varies depending on the target node type:

[0126] For attribute nodes, such as cumulative counter nodes or the level attribute of a single tax status node, the current snapshot value of the corresponding attribute field stored in the node can be read directly. For example, for path / intellectual property / patent / cumulative counter, the count value attribute of the cumulative counter node can be read.

[0127] For time-series nodes, such as tax status node sequences, when the target path is marked with [*], it indicates that the time series of all nodes of this type needs to be extracted. The system traces back from the current window node to a preset backtracking depth of L windows (such as 12 months) along the temporal adjacent edges between tax status nodes, extracts the attribute vectors of each window node (average tax credit rating, number of overdue declarations, and ratio of tax payable to actual payment), and arranges them in chronological order to form a historical sequence for subsequent time-series matching degree calculation.

[0128] Based on the target path field of the conditional expression structure encapsulated inside the conditional entity node in the policy knowledge graph, locate the corresponding attribute node or time sequence node in the enterprise state subgraph, and extract the current snapshot value and historical sequence.

[0129] The conditional expression structure is a data structure generated by the semantic role labeling model through structured transformation of qualification condition elements when the government policy module constructs the policy knowledge graph, and is encapsulated inside the conditional entity node.

[0130] A preferred example of a conditional expression structure includes:

[0131] The condition type identifier is an enumeration value, including the numerical threshold type NUM_THRESHOLD, the existence type EXISTENCE, the logical combination type LOGIC_COMB, the geographic affiliation type REGION, the qualification list type QUAL_LIST, and the time-limited window type TIME_WINDOW.

[0132] The target path field is a string type and stores a path expression pointing to an attribute node or time sequence node in the enterprise knowledge graph, which is used to locate the target attribute to be matched in the enterprise state subgraph.

[0133] The constraint operator field is a string type and stores comparison operators, including ≥, ≤, =, ∈, etc.

[0134] The constraint parameter field stores parameters such as the threshold for comparison, list identifier, or time window definition, and the data type depends on the condition type.

[0135] For example, numerical threshold-type conditions store numerical constants, qualification list-type conditions store list name strings and hash values, and time-sensitive window-type conditions store window length, aggregation operators, and target thresholds.

[0136] When generating the conditional expression structure, the qualification condition mention text identified by the semantic role labeling model is processed by the template parser. For example, the policy text fragment "The tax credit rating for the past three years is not lower than level B" is parsed to generate:

[0137] The condition type identifier is TIME_WINDOW;

[0138] The target path field is / tax status[*] / average tax credit rating;

[0139] The constraint operator field is ≥;

[0140] The constraint parameter fields are {window length: 3 years, aggregation operator: minimum value, threshold: score corresponding to level B 70};

[0141] After the structure is serialized in JSON format or key-value pair form, it is stored as the attribute fields of the condition entity node.

[0142] When the matching analysis is started, the structure is read from the condition entity node and deserialized into an in-memory object, and the target path field is parsed to perform the location operation.

[0143] Construct a two-stage matcher implemented using a constraint graph neural network.

[0144] The two-stage matcher includes a first stage and a second stage;

[0145] In the first stage, each condition node in the policy condition tree is encoded as a query vector, and the nodes located in the enterprise state subgraph and their neighbors are encoded as key vectors. The local matching score between the condition node and the enterprise node is calculated through cross attention.

[0146] In this application, a preferred implementation method for two-stage matching includes:

[0147] The first stage is set up as a local matching score calculation module;

[0148] The local matching score calculation module calculates the local matching score of each leaf node in the conditional logic tree for the current state of the enterprise. The specific construction method is as follows:

[0149] S101 encodes the conditional expression structure of the conditional entity node into a query vector, maps the conditional type identifier into a d-dimensional vector through an embedding layer; maps constraint operators into d-dimensional vectors through independent embedding layers; normalizes numerical or window parameters and maps them into d-dimensional vectors through a fully connected layer; and encodes the target path field into a d-dimensional vector through text convolution or a Transformer. These four vectors are concatenated and then fused through a fully connected layer. For example, with a 4d-dimensional input, the output is d-dimensional, yielding the query vector of the current conditional leaf node. .

[0150] S102 uses the target node located by the target path in the enterprise state subgraph and its one-hop neighbor nodes as the context node set. The attribute vector of each node is aggregated through a graph attention network layer with shared weights. The neighbor node information is weighted according to the edge weights between the node and the target node and converged to the target node to obtain the context representation of the target node. The context representation is then mapped to a key vector through a linear transformation. and value vector Let the number of neighboring nodes be m, and j = 1…m.

[0151] S103 calculates the query vector With each key vector Scaling dot product attention weights: The enterprise-side context vector is obtained by summing the value vectors according to their attention weights. .

[0152] S104 will After concatenation, the local matching score of the conditional leaf node is output through two fully connected layers and a sigmoid output layer. ∈[0,1].

[0153] The second stage injects the logical structure of the policy condition tree into the graph attention network, aggregates local matching scores from bottom to top, uses minimum pooling for logical AND nodes and maximum pooling for logical OR nodes, and calculates the time series matching degree by modeling the historical sequence of the time window node through a recurrent neural network and fusing it with the time window parameters, and outputs the overall matching score.

[0154] The second phase is set up as a logic injection and timing module;

[0155] The logic injection and timing module uses a conditional logic tree structure as its framework to construct an isomorphic graph neural network. It calculates the comprehensive matching score of nodes layer by layer from bottom to top, and finally outputs the overall matching score at the root node. The specific construction method is as follows:

[0156] For non-time-sensitive window-type leaf nodes, S201 directly inherits the output of the first stage. As the initial matching value for the current node, for time-window type leaf nodes, the time-sorted attribute vectors of each window extracted from the enterprise state subgraph in the first stage are input into a single-layer GRU network. The GRU hidden state dimension is set to 128, and the hidden state of the last time step is used as the fusion representation of the historical sequence. The time window parameters (window length L and aggregation operator type) in the constraint parameters are mapped through the embedding layer and concatenated with the hidden state. The time-series matching degree is then output through a fully connected layer and a sigmoid function, replacing the previous matching value. As the matching value for that leaf node.

[0157] S202 input is the set of matching scores for all child nodes { The output is the minimum value of the fraction, i.e. =min( The semantics of which all conditions must be satisfied are defined. The logical OR node takes the set of matching scores for child nodes as input and outputs the maximum value. =max( "This corresponds to any condition being met." If there are logical nodes with priorities in the condition tree, then before aggregation, the scores of each child node are multiplied by the learnable weight coefficient before pooling is performed.

[0158] S203 performs the above aggregation layer by layer from bottom to top until the root node, and outputs the overall matching score S∈[0,1] for the enterprise to the policy.

[0159] The two-stage matcher is trained end-to-end, specifically as follows: training samples consist of labeled enterprise policy matching pairs, with labels representing manually reviewed suitability classification results, such as perfect fit, partial fit, hard conflict, etc. The loss function uses mean squared error loss or binary cross-entropy loss, the optimizer uses Adam, and the initial learning rate is set to 1. The batch size can be set to 32. During training, the cross-attention weights in the first stage and the GRU parameters in the second stage are jointly updated, enabling the network to automatically learn deep matching patterns between conditional semantics and enterprise attributes.

[0160] Furthermore, the overall matching score S is compared with a preset threshold: if S≥0.8, it is classified as a complete fit; if 0.4≤S<0.8, it is classified as a partial fit; if S<0.4, it is classified as a hard conflict. For partially fit enterprises, the matcher can reverse the conditional leaf node with the lowest contribution along the conditional logic tree to generate specific qualification correction suggestions.

[0161] like Figure 5 The diagram shows the policy matching results interface. The left side of the interface displays a tree-like menu of the policy knowledge base, organized by policy type (announcements, regulations, implementation rules, application guidelines). The upper right panel displays the current enterprise's suitability classification label (e.g., "fully compatible") and the overall matching score (0.85). The lower right panel displays the nodes of the policy condition tree in the form of a conditional logic tree. The AND nodes aggregate all sub-conditions, and each sub-condition node (e.g., "Tax credit rating ≥ B in the past 3 years", "Registered capital ≥ 1 million yuan", "Owns ≥ 3 invention patents") is marked with its local matching score (0.92, 0.88, 0.75 respectively). This interface intuitively reflects the output results of the overall and local matching scores.

[0162] The overall matching score is compared with a preset threshold to classify enterprises as fully compatible, partially compatible, or hard-conflicting.

[0163] When the matching analysis result indicates that the enterprise is fully or partially matched, the reminder strategy decision-maker of the contextual multi-armed gambling machine model is triggered. This decision-maker concatenates the enterprise activity features, historical message interaction feedback features, and urgency and complexity features of government affairs extracted from the enterprise knowledge graph into a context vector. It then estimates the expected reward for each candidate reminder channel and timing combination and samples to select the optimal reminder action, such as... Figure 3 The diagram shown is a flowchart of the core matching and reminder generation process for the government policy module.

[0164] After generating the reminder message body, the message body hash, enterprise identifier, and reminder action metadata are assembled into a reminder event. The reminder event is recorded on the blockchain through an independent smart contract instance of the collaboration module, and the reminder event is bound to the status corridor instance of the associated government affairs, and a status monitoring timer is started.

[0165] If no response operation from any commissioning department is detected within the timer window, the state transition function of the state corridor instance is triggered, updating the status of the matter from "reminded" to "pending reminder", and notifying the reminder strategy decision-maker to perform secondary reminder optimization.

[0166] The reminder strategy decision-maker employs a contextual multi-armed gambling machine model based on Thompson sampling to achieve online adaptive optimization. Its input consists of contextual feature vectors extracted from enterprise knowledge graphs and government information, and its output is the selected reminder action.

[0167] In this application, a preferred method for feature extraction and context vector concatenation includes:

[0168] Extract the following characteristics from the enterprise entity nodes and their associated behavior logs in the enterprise knowledge graph: enterprise activity characteristics, historical message interaction feedback characteristics, and urgency and complexity characteristics of government affairs.

[0169] The enterprise activity characteristics include the average daily number of times the enterprise logs into the government affairs platform in the past 30 days, the frequency of enterprise business registration change events in the past 90 days, the on-time rate of tax declaration in the past year, and the number of new intellectual property authorization events in the past six months. The on-time rate is obtained by dividing the number of on-time declarations by the number of declarations that should be made. After normalization, these values ​​are mapped to fixed-dimensional values ​​and concatenated to form the enterprise activity feature vector.

[0170] The historical message interaction feedback features are used to query the records of government reminders received by enterprises in the system, and to statistically analyze the historical click-through rate, historical application conversion rate, and number of days since the most recent interaction from historical channels such as SMS, government APP push notifications, WeChat official account template messages, and emails. The statistical values ​​of each channel are arranged in channel order to form a feedback feature vector.

[0171] The urgency and complexity characteristics of the aforementioned government affairs matters are derived by obtaining the application deadline from the timeliness nodes of the policy knowledge graph, calculating the remaining days, and normalizing them into an urgency index; and by extracting the number of commissioning departments and process steps involved in the matter from the process dependency graph, and normalizing them into a complexity index. The two indicators are concatenated to form a feature vector for the matter.

[0172] The feature vectors of enterprise activity features, historical message interaction feedback features, and urgency and complexity features of government affairs are concatenated end to end to obtain a context vector, which is used as the input of the multi-armed gambling machine model.

[0173] Furthermore, the candidate reminder actions consist of a combination of reminder channels and reminder timings. The reminder channel set includes: SMS, government APP push, WeChat official account template messages, and email. The reminder timing set is based on the current time and divided into hourly granularities: immediate (0h), weekday morning (9:00 the next day), weekday afternoon (14:00 the next day), and next weekday (all day the next day). Each action 'a' corresponds to a channel-timing combination, and the size of the action space A is the number of channels × the number of timings.

[0174] Furthermore, a reward prediction is maintained for each action 'a'. The reward is defined as: whether the enterprise initiates a declaration operation within the observation window after the reminder is issued. If it does, the reward is r=1; otherwise, r=0.

[0175] Using the linear contextual gambling machine assumption, the expected reward for each action a is E[r|x,a]= in For the parameter vector of action a, the model for each Maintain a Gaussian posterior distribution Initially, =0, I stands for the identity matrix. Each time a new batch of feedback data is received, {( Update the corresponding action according to the Bayesian update rule. The posterior parameters, x∈ It is a context feature vector, where r∈{0,1} represents the reward, such as whether to apply;

[0176] During decision-making, Thompson sampling draws samples from the current posterior distribution of each action a. Calculate the predicted reward for each action. Choose the action a* that maximizes the predicted reward as the output. .

[0177] New feedback data received ( After that, only the corresponding action is updated. Posterior parameters:

[0178]

[0179]

[0180] After selecting a reminder action, a personalized reminder message body is generated. The message body includes: policy name, summary of the application item, list of departments responsible for handling the application, compatibility classification, and personalized application prompts. The message body can be formatted as plain text or a template message JSON structure according to the requirements of the selected channel. To observe the noise variance, we can represent the random error in reward prediction. It is the covariance matrix before the update; It is the updated covariance matrix.

[0181] After generation, calculate the digest hash value of the message body content, and assemble the following fields into an alert event structure:

[0182] Message body hash: H;

[0183] Corporate identifier: Unified Social Credit Code;

[0184] Reminder Channel: Select Channel Identifier;

[0185] Reminder Timing: Select the timestamp of the selected time;

[0186] Unique identifier for associated government affairs items: Item ID;

[0187] Generate timestamps;

[0188] The reminder event structure is stored on the blockchain as transaction data by calling the reminder evidence storage interface of the independent smart contract instance of the collaboration module.

[0189] The collaboration module queries whether a state corridor instance, i.e. an independent smart contract instance, has been deployed based on the government affairs item ID. If it has been deployed, it calls the instance's binding interface to associate the reminder event transaction hash with the current item status, sets the item status to "already reminded", and presets a maximum dwell time limit for the "already reminded" status in the state corridor instance. The maximum dwell time limit is dynamically set based on the historical average response time or the urgency of the item.

[0190] The state corridor instance internally maintains a logical timer based on block height. When the state enters the alerted state, it records the current block number. Each time there is an on-chain state query or an external trigger call, it checks whether the current block number satisfies (current block number - alerted current block number) > Tblock, where Tblock is the number of blocks converted to the preset maximum dwell time limit. If the timeout condition is met, the state transition function of the state corridor instance is automatically triggered and executed:

[0191] Update the status of the matter from "reminded" to "pending follow-up";

[0192] Generate an event log containing the hash of the previous state, the new state, and the trigger reason as timeout;

[0193] Send on-chain event notifications. The event body contains the event ID, enterprise identifier, and timeout status information.

[0194] After the off-chain listening component of the collaboration module captures the above-mentioned timeout event, it transmits the event information back to the reminder strategy decision-maker. The decision-maker obtains the enterprise identifier and event ID in the event, re-extracts the current context features, introduces the special context features of the timeout reminder, performs Thompson sampling again, selects the optimal channel and timing for the secondary reminder, and adds reminder prompts and deadline warnings to the content of the first message in the secondary reminder message body.

[0195] The second reminder also generates a new reminder event and stores it on the blockchain. It is then bound to the state corridor instance again, and a new timeout limit is set until the enterprise initiates a declaration operation in the state corridor instance. The status then moves to the next department's processing stage, and the timeout monitoring is automatically lifted.

[0196] This application deeply integrates personalized reminder strategy optimization, on-chain evidence storage, and cross-departmental collaborative process status monitoring, improving the timeliness and accuracy of reminders, ensuring the auditability of reminder behavior in an immutable manner, and eliminating the problem of no follow-up after reminders through a timeout self-driven mechanism.

[0197] When the personalized government affairs reminder involves multiple commissioning departments, the collaboration module parses the process dependency graph of the government affairs and deploys an independent smart contract instance as a state corridor instance for the current matter. The contract instance sets a state transition function. Any commissioning department must call the state transition function to update the status of the matter. Before the call, it must be verified by the joint digital signature of the current node and the previous node operator. Any state change record generates an immutable on-chain record.

[0198] From standardized service guidelines or internal procedures for government affairs, the text describing each step is extracted using a text structure analysis model. Then, through dependency parsing and semantic role labeling, the sequential dependencies, parallel branches, conditional branches, and synchronous convergence relationships between steps are extracted to generate the intermediate process structure.

[0199] The intermediate process structure is transformed into a directed graph representation, where nodes represent the responsibilities of the commissioning departments, edges represent the flow direction of matters, and each edge is attached with a logical expression of the preconditions, which references the condition node identifiers in the policy knowledge graph.

[0200] The directed graph is aligned with process instances extracted from historical case logs. The graph structure is then corrected and probabilistically labeled using a process mining algorithm to obtain the final process dependency graph. Each edge in the process dependency graph also carries a probability distribution of the average processing time estimated based on historical data.

[0201] When government affairs involve multiple departments handling matters sequentially or in parallel, the collaboration module must first obtain the process dependency graph of the matter to clarify the flow logic and preconditions of each step, providing accurate structured input for the subsequent deployment of independent smart contract instances (state corridors). This module automatically extracts the process structure from standardized service guidelines or internal procedures for government affairs and uses historical case logs for data-driven correction and probability labeling.

[0202] In this application, a preferred method for extracting process description text includes:

[0203] Standardized service guides or internal procedures describe the processing flow of matters in natural language. For example, an applicant submits materials to the Municipal Market Supervision Bureau, which reviews and approves them before transferring them to the Municipal Tax Bureau for tax verification. After verification, the Municipal Finance Bureau disburses funds. The system uses a pre-trained text structure parsing model to process the documents. This model can identify specific chapter titles such as processing flow and step descriptions in the document, and extract the text paragraphs under them as a sequence of step description texts, preserving the original text's sequential logic and paragraph hierarchy. Each extracted step description text corresponds to a natural language fragment of a processing step, marked as the original step entity text.

[0204] Specifically, by using two natural language processing techniques, dependency parsing and semantic role labeling, the flow relationships between entities in each stage are extracted, resulting in a sequence of descriptive texts for each stage.

[0205] Furthermore, by using named entity recognition, the names of the commissioning departments appearing in the description text of each stage, such as the Market Supervision Bureau and the Tax Bureau, and key action verbs, such as submit, review, transfer, and disburse, are used to label each department's action as a stage entity mention.

[0206] In this application, an example of a preferred method for extracting dependencies between departmental action pairs is as follows:

[0207] By analyzing the syntactic parse tree and semantic role labeling results, the dependency types between entities in the process are extracted, including: sequential dependency, parallel branching, conditional branching, and synchronous merging.

[0208] The sequential dependency refers to temporal sequence conjunctions, such as then, afterwards, in sequence, or the temporal semantics of verbs, such as first...then...Identification, indicating that step B can only be started after step A is completed;

[0209] The parallel branches are parallel conjunctions, such as the recognition of multiple downstream verbs that simultaneously, separately, in parallel or sharing the same upstream link, indicating that link A can be triggered simultaneously by link B and link C after link A is completed.

[0210] The conditional branches are branch structures guided by conditional conjunctions, such as "if...then...", "if not met, then...", "supplementary materials required", etc., which express that under certain conditions, the process moves to a specific branch.

[0211] In the synchronous convergence, multiple upstream links point to the same convergence link, and the initiation condition of the convergence link is jointly determined by multiple upstream links. This can be identified by expressions such as "after all are completed" or "after all have passed", which express the convergence logic of parallel branches.

[0212] Furthermore, the identified process entities are used as nodes, and the aforementioned relationships are used as directed edges, i.e., conditional branch edges, to which conditional description text is attached, generating a preliminary intermediate process structure that is not associated with an external knowledge base. The intermediate process structure is a directed graph, but the semantics of the nodes and edges are not yet aligned with the conditional nodes in the policy knowledge graph, nor does it contain probabilistic time information.

[0213] A preferred method for converting an intermediate process structure into a directed graph representation includes:

[0214] Each entity in the process is mapped to a process node identified by the responsibilities of the commissioning department. The node attributes include: the identifier of the commissioning department, a brief description of the responsibilities of the process, and the node type, such as acceptance node, review node, notification node, and archiving node.

[0215] Each directed edge in the flowchart represents the flow direction of an item. For edges marked as conditional branches in the intermediate process structure, their conditional statements are extracted from the original policy text. Based on the conditional node identification system already constructed in the policy knowledge graph, the conditional statements are transformed into a logical expression. The logical expression uses the conditional node identification in the policy knowledge graph as an atomic predicate, supporting AND (∧), OR (∨), and NOT (¬) combinations.

[0216] For example, a conditional branch edge where the enterprise's tax credit rating has been A for the past three years is transformed into a logical expression containing the condition node identifier COND_A_TAX_3Y_A. If this precondition is not met, the flow edge is blocked.

[0217] The final directed graph is denoted as G=(V,E), where V is the set of nodes in the commissioning department process, E is the set of flow edges, and each edge e∈E is associated with a preconditional logical expression cond(e) and a processing time probability distribution parameter with a default null value.

[0218] The directed graphs described above are mainly generated based on normative documents, which may deviate from the actual situation. For example, there may be common exception paths not mentioned in the documents, or the probability of certain edges occurring in the actual process. Therefore, a process mining algorithm is introduced to correct the graph structure and label the probability using historical case logs.

[0219] Specifically, historical processing instances of the administrative matter are extracted from the case database. Each instance includes: matter ID, the actual handling department of each step, the start time stamp of the step, the end time stamp of the step, and approval opinions, such as return, approval, etc. Event logs. These are grouped by instance ID and sorted by timestamp to form an event sequence.

[0220] Furthermore, the event sequence of each instance is aligned with the current directed graph G, allowing node skipping and insertion. Using an inductive process mining algorithm, such as InductiveMiner, a frequency-weighted process model G' is automatically discovered from the event log. By comparing G and G', the following deviations are identified and corrected:

[0221] If an edge that does not exist in G appears frequently in the log, then add the edge and its instance's conditional context to G and mark it as a log completion edge.

[0222] If a node or edge in G never appears in the log or appears less frequently than a preset threshold, it will be marked as inactive but not deleted, for reference during process execution.

[0223] For branches defined as parallel in G, if the log shows that they are actually mutually exclusive selections, the parallel relationship is corrected to conditional branches based on the log frequency, and the probability of occurrence of each branch is marked.

[0224] In this embodiment, the inductive process mining algorithm automatically discovers process models from event logs according to the following steps:

[0225] Sort the events of each item instance in the historical case log by timestamp, extract the start and end events of each stage, and form an event sequence;

[0226] Construct a direct follow-up relationship graph, where nodes represent links and edges represent the frequency of two links appearing successively.

[0227] Based on the direct following relationship graph, four cutting rules are recursively applied: exclusive cutting, sequential cutting, concurrent cutting, and cyclic cutting, to split the log into sub-logs until it can no longer be cut, thereby generating a process tree;

[0228] The process tree is converted into a Petri net or BPMN model as a frequency-weighted process model G′. After obtaining G′, G is compared with G′, the following deviations are identified and corrections are performed:

[0229] For the corrected process dependency graph G*, an average processing time probability distribution is attached to each edge e. The system counts the time intervals of all instances of the current edge in the historical log, i.e., the sample set of the start time of the next stage minus the completion time of the previous stage. A log-normal distribution is fitted using maximum likelihood estimation to obtain the processing time probability density function parameters of the current edge. The processing time probability density function parameters include the logarithm of the mean μ and the standard deviation σ. The normal distribution is used to preset the maximum dwell time limit for the stage in the state corridor instance and to calculate the urgency of the reminder strategy. Specifically, for each edge e, the time sample set {d1,d2,…,dn} from the completion of the previous stage to the start of the current stage in all historical instances is collected. The natural logarithm is taken and the sample mean and sample standard deviation are calculated as parameters of the log-normal distribution.

[0230] Each link in the process dependency graph is mapped to a state enumeration value in a smart contract, and the flow edge between each link is mapped to a state transition function. The transition conditions of the state transition function are encoded as on-chain verifiable assertions of the preconditions.

[0231] During the creation of a state corridor instance, based on the probability distribution of the average processing time, a maximum dwell time limit is preset for each state. If the current state dwell time exceeds the maximum dwell time limit, the state corridor instance automatically triggers the built-in timeout state transition, transfers the matter to the exception processing state, and sends an on-chain event to the associated commissioning department.

[0232] When the state corridor instance is deployed, the contract address is bound to the unique identifier of the government affairs matter and recorded in the contract registry of the collaboration module. At the same time, a matter creation event is published to the enterprise monitoring module, triggering the pre-population of relevant information on the enterprise side.

[0233] Specifically, the process of compiling the process dependency graph G*=(V,E) into a smart contract state machine is as follows:

[0234] Traverse the node set V in the process dependency graph G*, assign a unique state enumeration identifier to each node in the responsibility process of the commissioning department, and automatically append the built-in states of initial state, completion state, and exception handling state in addition to the business process:

[0235] The initial state is that the event has been created but has not yet entered any stage;

[0236] The completion status means that all mandatory steps have been completed and the matter has been successfully concluded.

[0237] The exception handling status is triggered by timeout or abnormal conditions, and the matter enters the manual intervention channel.

[0238] All state identifiers are defined as enumeration types within the smart contract. After compilation, they are stored as integer values ​​in the contract's storage variable `current_state`. The mapping relationship between state enumeration values ​​and stage nodes is recorded in the contract's read-only mapping table `state_mapping`, which returns readable stage names when queried.

[0239] Furthermore, the traversal process depends on the set of edges E in the graph G*, and for each edge e∈E, a state transition function is defined.

[0240] In this application, a preferred implementation of the transfer function includes:

[0241] Each state transition function accepts the previous node's signature, the current node's signature, and a conditional proof;

[0242] The previous node signature is a digital signature made by the authorized private key of the commissioning department corresponding to the starting point of the current edge in the process dependency graph, for the declaration that the previous state has been completed. The digital signature is used to prove that the commissioning department of the previous stage has confirmed that the handling work within its scope of responsibility has been officially completed and that the current matter is qualified to be transferred to the next stage.

[0243] The current node signature is a digital signature made by the authorized private key of the department corresponding to the end point of the current edge in the process dependency graph, for the declaration of receiving and initiating the processing of this step. The digital signature is used to prove that the department in charge of the current step has officially received the matter and promised to start the processing, rather than being initiated by the system or an unrelated party.

[0244] The conditional proof is an optional parameter. When the transition edge is accompanied by a preconditional logical expression, the field provides proof data for the on-chain data referenced by the conditional expression, such as enterprise qualification attributes and policy condition node status, such as a structured body containing data source, data value and verifiable path. If the transition edge has no additional preconditions, the field can be empty.

[0245] The transfer function also includes a three-stage verification logic, including precondition verification, joint signature verification, and state change and on-chain recording.

[0246] Specifically, the precondition verification reads the precondition logic expression cond(e) attached to edge e. The cond(e) expression is stored in the contract in the form of a verifiable assertion, that is, it is parsed into a combination of calls to a set of atomic condition check functions. Each atomic condition check function corresponds to a condition node identifier in the policy knowledge graph. For example, the contract reads the current snapshot value of the corresponding attribute in the enterprise state subgraph through the oracle interface or the on-chain state query interface, compares it with the condition threshold, and returns a boolean value. If the boolean operation result of cond(e) is false, the function rolls back and returns an error code that the precondition is not met.

[0247] Joint signature verification includes: when the state transition function is invoked, it requires the caller to submit the current state hash signed by the private key of the current processing department node, and a proof that the previous state has been completed signed by the private key of the previous department node in the process dependency graph, such as... Figure 4 The diagram shown is a flowchart of the collaborative module's state corridor and joint signature verification process.

[0248] The verification process first performs public key recovery and identity verification on the chain for the private key signature of the current processing department node and the private key signature of the previous department node, respectively, to ensure that the previous signer is indeed the authorized department in the previous stage of the current matter process, and the current signer is the authorized department in the current stage.

[0249] After successful verification, the state transition function updates the state variables and stores the composite state change record, which includes the current state, proof of the previous state completion, signatures of both parties, timestamp, and operation hash, as an immutable on-chain log.

[0250] If any party's signature is missing or invalid, the state transition function rolls back and records a violation attempt event, while simultaneously triggering a security notification to other commissioning departments.

[0251] In this application, a preferred method for joint signature verification includes:

[0252] After the state transition function performs and passes the precondition verification, it enters the joint signature verification stage. This stage is the core security mechanism for enforcing the cross-departmental workflow sequence and preventing skipping steps in operations. Its technical principle is that each legitimate transition of the matter's state is bound to the joint result of two cryptographic actions: the previous commissioning department's confirmation of delivery and the current commissioning department's "acceptance confirmation." If either one is missing, the workflow cannot be established.

[0253] When the state transition function is called, it requires the caller to submit two signature parameters in the transaction data:

[0254] The signature of the current state hash is a digital signature made by the authorized private key of the current processing department node, which is a hash value obtained by concatenating the current state identifier, the unique identifier of the matter, and the timestamp of the current block. The digital signature proves that the call is made by the authorized department in the current stage, and that the call is initiated in real time to prevent replay attacks.

[0255] The signature proving the completion of the previous state is digitally signed by the authorized private key of the previous department node in the process dependency diagram, using the hash value of the concatenation of the previous state identifier, the unique identifier of the matter, and the block timestamp when the previous state was completed. The signature proves that the previous commissioning department has officially confirmed the completion of the work within its scope of responsibility, and solidifies the completion statement in a non-repudiable cryptographic manner. The block timestamp when the previous state was completed can be read from the state change record in the contract storage.

[0256] Both signatures were generated using the elliptic curve digital signature algorithm, and the signature curve is consistent with the signature curve used by the blockchain network, ensuring that the on-chain native verification command can directly execute the signature verification operation.

[0257] The verification process is executed within the smart contract through a pre-compiled contract or built-in cryptographic functions. The specific steps are as follows:

[0258] First, the contract reads the list of authorized public key addresses of the previous department and the list of authorized public key addresses of the current department from its own storage. When the state corridor instance is deployed, the public key addresses are pre-written into the contract's permission mapping table according to the information of the commissioning department bound to each node in the process dependency diagram, and cannot be changed during the handling of the matter.

[0259] Secondly, the contract invokes elliptic curve signature recovery, using the hash of the concatenated hash value of the current state as the message digest and the signature of the current state hash as the signature input to recover the public key address. This public key address is then compared with the authorized public key address corresponding to the current stage in the permission mapping table. If the addresses match, the current node's signature verification passes; otherwise, the signature verification fails.

[0260] Similarly, the contract uses the hash of the proof completed in the previous state as the message digest and the signature of the proof completed in the previous state as the signature input to recover the public key address. The public key address is then compared with the authorized public key address corresponding to the previous step in the permission mapping table. If the addresses match, the signature verification of the previous node is successful.

[0261] The cryptography used in the verification process ensures that the private key is held only by the corresponding authorized department, and that no third party can forge a valid signature that matches the authorized public key address. Therefore, the legitimacy and authorization of the signer's identity can be confirmed from a technical perspective by comparing the addresses.

[0262] After all signatures have been verified, the state transition function performs a state update operation. It updates the contract's state variables to the target state enumeration value pointed to by the transition edge, and simultaneously records the current block timestamp as the entry time of the new state.

[0263] After the state update is complete, the contract constructs a composite state change record, which contains the following fields:

[0264] Previous state identifier;

[0265] Current state identifier;

[0266] The hash digest of the proof of the previous state completion;

[0267] The signature of the current processing department node;

[0268] The signature of the previous department node;

[0269] The block timestamp in which the state change occurred;

[0270] Caller address;

[0271] The transaction hash for this call;

[0272] After being serialized, the above fields are written into the event log storage area of ​​the contract as an immutable on-chain log. The log is a structured event and can be indexed and queried back through the event listening interface of the blockchain node. Due to the chain hash structure and consensus mechanism of the blockchain, any state change record that has been put on the chain cannot be modified or deleted afterward, providing a complete and reliable audit trail for business collaboration between multiple commissions and departments.

[0273] If either the current node's signature or the previous node's signature is missing (i.e., an empty byte is passed in), or signature recovery fails, or the recovered public key address does not match the authorized address in the permission mapping table, the state transition function executes the following security processing flow:

[0274] The current transaction is rolled back immediately, all modifications to state variables are stopped, and the contract state remains consistent with that before the call.

[0275] Record a violation attempt event, including: the reason for failure, such as invalid previous signature, missing current signature, caller address, target status of the attempted call, timestamp and transaction hash;

[0276] Security notification events are sent through the contract's event mechanism. After the monitoring nodes of the relevant departments capture the events, they can initiate internal security audit processes to investigate whether there are any unauthorized operations or system anomalies.

[0277] After all the above checks pass, the state update will update the contract's state variables to the target state identifier, record the current block timestamp, and issue a state change event.

[0278] The above mapping relationship is compiled into Solidity contract code. After successful compilation, contract bytecode and ABI interface description are generated. Since the process dependency graphs of different government affairs are different, each matter needs to be compiled independently to generate exclusive contract bytecode.

[0279] During the deployment initialization phase, the state corridor instance reads the average processing time probability distribution parameters (μ and σ) of each edge e in the process dependency graph G*, which are log-normal distributions, and presets a maximum dwell time Tmax(i) for each business state, where i is the index of the link node corresponding to the current state.

[0280] A preferred method for calculating the maximum dwell time is as follows: Take the P90 quantile of the processing time for this step in the historical case log (i.e., 90% of historical instances completed within this timeframe) as the baseline value, and multiply it by a configurable safety factor α (default α=1.5) to tolerate certain fluctuations.

[0281]

[0282] in The inverse cumulative distribution function (quantile function) for the processing time distribution of the current stage is determined by the log-normal distribution parameters μi and σi. The parameter values ​​are stored in the contract in the form of block number or Unix timestamp difference.

[0283] The contract maintains a dwell start timestamp (tenter) for each state. Each time a state change occurs and a new state is entered, the current block timestamp is recorded as the tenter. The contract provides a public state check function (checkTimeout()), which can be called by any external account or timer contract.

[0284] Specifically, read the current state current_state and its corresponding maximum dwell time Tmax;

[0285] Calculate the dwell time Δt = block.timestamp - tenter;

[0286] If Δt > T_max, the built-in timeout state transition is triggered, the current_state is updated to the exception handling state, and a timeout event is emitted.

[0287] The on-chain timeout event carries fields including: a unique identifier for the event, a status identifier for the timeout stage, the block number and timestamp of the timeout occurrence, the maximum dwell time limit, and the actual dwell time, etc.

[0288] Once the monitoring nodes of the relevant departments capture the event, they can initiate manual intervention or escalate the processing procedures accordingly.

[0289] When the system creates a state corridor instance for a specific government matter, it calls the contract deployment service of the collaboration module. The deployment service sends the compiled exclusive contract bytecode to the node through the deployment transaction of the blockchain network. After the transaction is confirmed, the on-chain contract address is returned.

[0290] The contract constructor accepts a unique identifier (item ID) of the government affairs matter as a parameter. After deployment, the contract address and the item ID are bound one-to-one. The physical meaning of the binding relationship is that the smart contract instance only serves the current processing flow of the specific matter, and contract instances between different matters are completely isolated.

[0291] The collaboration module maintains an on-chain contract registry smart contract. After a newly deployed state corridor instance is successfully deployed, it calls the register interface of the registry contract to register the following information on the chain: unique identifier of the government matter, state corridor contract address, deployment timestamp, process dependency graph version number. The registry provides a reverse query interface, allowing external systems to query the corresponding contract address through the matter ID or obtain matter information through the contract address, thus achieving bidirectional traceability between matters and contracts.

[0292] Furthermore, after the contract is deployed and registered, the collaboration module publishes a matter creation event to the enterprise monitoring module. The event is an asynchronous message that includes the matter ID, matter name, a list of departments involved, and the status corridor contract address.

[0293] Upon receiving the event, the enterprise monitoring module triggers a pre-population operation for relevant information on the enterprise side. Specifically, based on the policy requirements associated with the event ID, it retrieves the enterprise's qualification information, historical application data, and other necessary information from the enterprise's knowledge graph in advance, which may be required in subsequent processing stages, and caches them in the temporary data area for event processing. This pre-population operation reduces the time spent by various departments repeatedly querying enterprise information during the processing stage, improves collaborative processing efficiency, and provides the enterprise side with a data foundation for event progress awareness.

[0294] It is important to note that the constructions and arrangements of this application shown in several different exemplary embodiments are merely illustrative. Although only two embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible without substantially departing from the novel teachings and advantages of the subject matter described in this application. These modifications may include, for example, changes in the size, dimensions, structure, shape, and proportions of various elements, as well as parameter values ​​(e.g., temperature, pressure, etc.), installation arrangements, the use of materials, colors, orientations, etc. For example, an element shown as integrally formed may be composed of multiple parts or elements, the position of elements may be inverted or otherwise altered, and the nature or number or position of discrete elements may be changed or altered. Therefore, all such modifications are intended to be included within the scope of this application. The order or sequence of any process or method steps may be changed or rearranged by alternative embodiments. Any "apparatus plus function" clause is intended to cover, and not only structurally equivalent but also equivalent structures, the structures performing the functions described herein. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of this application. Therefore, this application is not limited to a particular embodiment, but extends to various modifications that still fall within the scope of the appended claims.

[0295] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of performing this application as currently considered, or those features that are not relevant to implementing this application) may be omitted.

[0296] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those of ordinary skill in the art who benefit from this disclosure, the development effort will be a routine task in design, manufacturing, and production without requiring extensive experimentation.

[0297] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application, and all such modifications and substitutions should be covered within the scope of the claims of this application.

Claims

1. A government affair reminding and handling interface system based on communication interaction of an enterprise, characterized in that, include: The enterprise monitoring module collects multi-source time-series data of enterprises through data interfaces and constructs an enterprise knowledge graph; The government policy module parses government policy texts, extracts applicable subjects, conditions and timeliness elements to construct a policy knowledge graph, and performs matching analysis to solve constraint satisfaction problems, thereby obtaining personalized government reminders; The enterprise monitoring module includes a data acquisition unit and a graph unit; The method for constructing the enterprise knowledge graph using the graph unit is as follows: The business registration change record is transformed into a quadruple of (enterprise, attribute name, old value, new value, timestamp). A change event node is generated for each quadruple, and the version evolution is represented by a directed edge. Adjacent change nodes with the same attribute are connected in series. The vector representation of the changed event node is generated by weighting the difference vector between the old and new values ​​and the time decay factor; Using the month as the smallest granularity, the time series values ​​of tax declarations are aggregated by a sliding window. Each window generates a tax status node. The attributes of the tax status node include the average tax credit rating within the window, the number of overdue declarations, and the ratio of tax payable to actual tax payable. Tax status nodes of adjacent windows are connected by time-adjacent edges, and the weight of the edge is the Jaccard similarity between the two windows. Cumulative counter nodes are established for patents, trademarks, and software copyrights respectively. The counter node value increases with each authorization announcement event. Each authorization event generates an event node. The event node carries the authorization number, authorization date, and classification number. The event node points to the cumulative counter node through a contribution edge. The weight of the edge decreases linearly with the time elapsed since the authorization date. Centered on the enterprise entity node identified by the unified social credit code, change event nodes, tax status nodes, authorization event nodes, and cumulative counter nodes are connected through relationship edges; When the personalized government affairs reminder involves multiple commissioning departments, the collaboration module parses the process dependency graph of the government affairs and deploys an independent smart contract instance as a state corridor instance for the current matter. The contract instance sets a state transition function. Any commissioning department must call the state transition function to update the status of the matter. Before the call, it must be verified by the joint digital signature of the current node and the previous node operator. Any state change record generates an immutable on-chain record. A two-stage matcher is constructed using a constraint graph neural network; The two-stage matcher includes a first stage and a second stage; In the first stage, each condition node in the policy condition tree is encoded as a query vector, and the nodes located in the enterprise state subgraph and their neighbors are encoded as key vectors. The local matching score between the condition node and the enterprise node is calculated through cross attention. The second stage injects the logical structure of the policy condition tree into the graph attention network, aggregates local matching scores from bottom to top, uses minimum pooling for logical AND nodes and maximum pooling for logical OR nodes, and calculates the time series matching degree by modeling the historical sequence of the time window node through a recurrent neural network and fusing it with the time window parameters, and outputs the overall matching score. The overall matching score is compared with a preset threshold to classify enterprises as fully compatible, partially compatible, or hard-conflicting. When the matching analysis result indicates that the enterprise is fully or partially matched, the reminder strategy decision-maker of the context multi-armed gambling machine model is triggered. The reminder strategy decision-maker concatenates the enterprise activity features, historical message interaction feedback features, and urgency and complexity features of government affairs extracted from the enterprise knowledge graph into a context vector, estimates the expected reward for each candidate reminder channel and reminder timing combination, and samples to select the optimal reminder action. After generating the reminder message body, the message body hash, enterprise identifier, and reminder action metadata are assembled into a reminder event. The reminder event is recorded on the blockchain through an independent smart contract instance of the collaboration module, and the reminder event is bound to the status corridor instance of the associated government affairs, and a status monitoring timer is started. If no response operation from any commissioning department is detected within the timer window, the state transition function of the state corridor instance is triggered, updating the status of the matter from "reminded" to "pending reminder", and notifying the reminder strategy decision-maker to perform secondary reminder optimization.

2. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 1, characterized in that: The data acquisition unit is used to collect the multi-source time-series data, including business registration changes, tax declarations, and intellectual property information. The enterprise knowledge graph constructed by the graph unit using the multi-source time-series data is used to reflect the real-time status of the enterprise.

3. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 1, characterized in that: The government policy module collects unstructured and semi-structured original policy texts; The original policy texts include announcements, measures, implementation rules, and application guidelines; The collected original policy texts are cleaned and segmented into chapters and levels, and policy discourse units are output by title, chapter, and paragraph levels. A semantic role labeling model is constructed by a self-attention encoding layer, a dual-path feature extraction layer, and a conditional random field decoding layer to process each policy discourse unit, identify and extract the applicable subject, qualification conditions, and timeliness elements.

4. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 1, characterized in that: The matching analysis for satisfying the constraints includes: expanding the enterprise entity nodes and their adjacent attribute nodes, event nodes, and temporal state nodes in the enterprise knowledge graph into an enterprise state subgraph. Based on the target path field of the conditional expression structure encapsulated within the conditional entity node in the policy knowledge graph, the corresponding attribute node or time sequence node is located in the enterprise state subgraph, and the current snapshot value and historical sequence are extracted.

5. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 1, characterized in that: From standardized service guidelines or internal procedures for government affairs, the text describing each step is extracted using a text structure analysis model. Then, through dependency parsing and semantic role labeling, the sequential dependencies, parallel branches, conditional branches, and synchronous convergence relationships between steps are extracted to generate the intermediate process structure. The intermediate process structure is transformed into a directed graph representation, where nodes represent the responsibilities of the commissioning departments, edges represent the flow direction of matters, and each edge is attached with a logical expression of the preconditions, which references the condition node identifiers in the policy knowledge graph. The directed graph is aligned with process instances extracted from historical case logs. The graph structure is then corrected and probabilistically labeled using a process mining algorithm to obtain the final process dependency graph. Each edge in the process dependency graph also carries a probability distribution of the average processing time estimated based on historical data.

6. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 5, characterized in that: Each link in the process dependency graph is mapped to a state enumeration value in a smart contract, and the flow edge between each link is mapped to a state transition function. The transition conditions of the state transition function are encoded as on-chain verifiable assertions of the preconditions. During the creation of a state corridor instance, based on the probability distribution of the average processing time, a maximum dwell time limit is preset for each state. If the current state dwell time exceeds the maximum dwell time limit, the state corridor instance automatically triggers the built-in timeout state transition, transfers the matter to the exception processing state, and sends an on-chain event to the associated commissioning department. When the state corridor instance is deployed, the contract address is bound to the unique identifier of the government affairs matter and recorded in the contract registry of the collaboration module. At the same time, a matter creation event is published to the enterprise monitoring module, triggering the pre-population of relevant information on the enterprise side.

7. The enterprise government affairs reminder and commission docking system based on communication interaction as described in claim 6, characterized in that: When the state transition function is called, the caller is required to submit the current state hash signed by the private key of the current processing department node, and the proof that the previous state has been completed signed by the private key of the previous department node in the process dependency graph. The verification process first involves performing public key recovery and identity verification on the chain for the private key signature of the current processing node and the private key signature of the previous node. After successful verification, the state transition function updates the state variables and stores the composite state change record, which includes the current state, proof of the previous state completion, signatures of both parties, timestamp, and operation hash, as an immutable on-chain log. If any party's signature is missing or invalid, the state transition function rolls back and records a violation attempt event, while simultaneously triggering a security notification to other commissioning departments.

Images