Seal watermark rule analysis method and system based on natural language processing

Abstract

The application discloses a seal watermark rule analysis method and system based on natural language processing, relates to the cross technical field of natural language processing and image processing, and comprises the following steps: acquiring a supply chain financial contract seal image containing a watermark; performing a preprocessing operation on the seal image; and extracting a watermark text from the preprocessed seal image; in the analysis process, the automatic seal identification and trigger logic are synchronously associated, the constructed dependency chain semantic graph contains automatic seal associated attributes, the association mapping between the dependency chain and the automatic seal system is realized, the automatic seal is accurately triggered, automatic seal misoperation is reduced, and the contract stamping automation efficiency is improved; the resolution model contains a fuzzy completion module, which can complete the fuzzy time, fuzzy participants and fuzzy automatic seal information in the change notice and evaluate the completion result confidence, thereby guaranteeing the accuracy of the change information, providing reliable data support for the dynamic update of the dependency chain, and further improving the robustness of the change processing.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of natural language processing and image processing, and in particular to a method and system for parsing seal watermark rules based on natural language processing. Background Technology

[0002] In business operations, seal watermarks are crucial for ensuring the legality and authenticity of documents, and are widely used in important documents such as contracts and supporting documents. Their rules not only include stylistic features but also rely heavily on textual expression to convey compliance logic. Currently, the interpretation of seal watermark rules largely depends on manual processes. Staff must meticulously review the details of the watermark text and industry-specific terminology to determine compliance. This method is inefficient and prone to errors due to misunderstandings of semantic meaning and unfamiliarity with industry terminology. Especially when faced with differentiated textual rules from different partners and industries, manual interpretation struggles to quickly and accurately adapt, significantly increasing the time cost and compliance risks associated with document verification.

[0003] With the widespread use of electronic documents, the amount of watermark data requiring parsing has surged, making traditional manual and simple technical methods increasingly inadequate. Existing technologies primarily focus on style feature extraction, failing to adequately cover the semantic relationships and core rules of logical expression in watermark text. They cannot uncover compliance requirements behind the text through text understanding, struggle to adapt to complex and ever-changing text rules, and lack accuracy and adaptability. Consequently, they cannot provide enterprises with stable text rule parsing support. Therefore, there is an urgent need for technical solutions that can accurately process text expression rules to address the aforementioned core issues. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of low efficiency and large error caused by manual parsing of seal watermark rules, as well as the inability of existing technologies to cover semantic associations and complex logical rules and to adapt to automatic seal systems. Therefore, a new seal watermark rule parsing method and system based on natural language processing is proposed.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a seal watermark rule parsing method based on natural language processing, comprising the following steps:

[0006] Obtain the watermarked seal image of the supply chain finance contract; perform preprocessing operations on the seal image;

[0007] Extract watermark text from the preprocessed seal image;

[0008] Natural language processing technology is used to perform semantic analysis on the extracted watermark text to obtain the identification results of the roles of supply chain participants, seal types, and seal dependencies among participants;

[0009] Based on the identification results, a dependency chain semantic graph is constructed, with supply chain participants as nodes and seal dependency relationships as edges.

[0010] Furthermore, the preprocessing operation performed on the seal image includes the following specific steps:

[0011] A binarization algorithm is used to remove background noise from the seal image;

[0012] Optimize the edge contour of the watermark text using an edge detection algorithm;

[0013] A threshold segmentation algorithm is used to separate the stamp pattern area from the watermark text area.

[0014] Output the image area containing the watermark text.

[0015] Furthermore, the extraction of watermark text is semantically parsed using natural language processing technology, and the specific steps are as follows:

[0016] Enable dependency inference model;

[0017] By relying on the inference model to adjust the semantic encoding weight, priority is given to obtaining the recognition results of exclusive seal terms and corresponding automatic seal identifiers in the supply chain finance field, such as supplier official seal, manufacturer acceptance seal, logistics carrier seal, and bank loan seal.

[0018] The inference model is used to label the roles of supply chain participants, the role of stamping actions, and the role of time constraints. The time constraint roles include time-related expressions such as before, after, simultaneously with, and until completion.

[0019] Furthermore, the dependent reasoning model also includes a term disambiguation module, which constructs a polysemous dictionary of supply chain finance terms to record the mapping relationship between the scenarios and meanings of polysemous terms and the corresponding automatic stamp identification mapping relationship.

[0020] The terminology disambiguation module identifies polysemous terms and assigns them their corresponding automatic stamps based on contextual scene recognition and role association verification.

[0021] Furthermore, the dependency reasoning model also includes a sentence parsing module, which constructs a library of ambiguous sentence templates;

[0022] Among them, the ambiguous sentence template library includes typical ambiguous templates containing automatic seal triggering logic in supply chain finance scenarios;

[0023] After performing ambiguous sentence matching on the watermark text, the sentence parsing module determines the true logic of the ambiguous sentences and the corresponding automatic stamp triggering logic by supplementing the context semantics and associating with the causal reasoning rule base.

[0024] Furthermore, it also includes steps for handling dynamic changes in supply chain participants:

[0025] Enable the resolution model;

[0026] Modal-specific processing is performed on change notifications sent by supply chain participants, which contain text, tables, and attached images, using a resolution model to obtain text modality, table modality, and attached image modality; the change initiator, change description, change effective time, and automatic stamp change elements are extracted;

[0027] We perform weighted fusion of text modality, table modality, and attached image modality, and map the features of text modality, table modality, and attached image modality into a unified feature vector;

[0028] The change type is identified based on the unified feature vector, and the entity boundaries of the old participants, new participants, and the corresponding old and new automatic seal identifiers are located.

[0029] Furthermore, the resolution model also includes a fuzzy completion module;

[0030] The fuzzy completion module addresses the fuzzy time, fuzzy participants, and fuzzy automatic seal information in the change notification by combining the historical change cycle of the associated participants, the business cooperation list, and the automatic seal filing database to complete the fuzzy information and obtain the completion result.

[0031] The fuzzy completion module assesses the confidence level of the completion results and triggers a manual confirmation process when the confidence level is lower than a preset threshold.

[0032] Furthermore, it also includes providing early warnings about the health of dependency chains:

[0033] A health warning model was activated, and historical stamping time data and automatic stamping historical trigger data were collected. The data were divided into two time dimensions: the past week period and the past 1 to 4 weeks period. The data of each dimension were time-series encoded by the branches of the long short-term memory network to obtain time-series encoded features.

[0034] Collect real-time business data and real-time status data of automatic seals from supply chain participants to obtain automatic seal-related features. Convert discrete business features into continuous feature vectors and normalize the continuous business features to obtain business embedded features.

[0035] The time-series coding features, business embedding features, and automatic stamping-related features are fused into a unified health feature vector by a lightweight fully connected layer according to preset weights. Based on the unified health feature vector, the expected stamping time of the participants and the expected triggering time of the automatic stamping are calculated.

[0036] Furthermore, the health warning model also includes a dynamic warning layer. The dynamic warning layer uses a gradient boosting tree algorithm to train a health predictor, outputs the expected stamping time of the participants and the expected triggering time of the automatic stamp, and maps the unified health feature vector to the health score of the participants through the health predictor. The health threshold is dynamically adjusted based on the urgency of the contract, the business load of the participants, and the load of the automatic stamping system.

[0037] Furthermore, it also includes a step to verify the accuracy of the parsing:

[0038] The verification algorithm consists of logical consistency verification, business compliance verification, timing rationality verification, ambiguity resolution result verification, and automatic seal logic verification.

[0039] If a logical anomaly, compliance anomaly, timing anomaly, ambiguity resolution anomaly, or automatic stamping logic anomaly is detected, the dependency inference model iterative optimization process is triggered, the domain adaptation weight and automatic stamping association mapping rules are adjusted, and the parsing process is re-executed until the preset business requirements are met.

[0040] The seal watermark rule parsing system based on natural language processing includes: a seal image module, an image preprocessing module, an extraction module, a semantic parsing module, and a semantic graph construction module;

[0041] The seal image module is used to acquire watermarked images of supply chain finance contract seals.

[0042] The image preprocessing module is used to perform preprocessing operations on the seal image acquired by the seal image module to enhance the clarity of the watermark text;

[0043] The extraction module is used to extract the watermark text from the stamp image processed by the image preprocessing module;

[0044] The semantic parsing module is used to perform semantic parsing on the watermark text extracted by the extraction module using natural language processing technology, and to obtain the identification results of the roles of supply chain participants, seal types and seal dependencies between participants;

[0045] The semantic graph construction module is used to construct a dependency chain semantic graph based on the recognition results of the semantic parsing module. The dependency chain semantic graph uses supply chain participants as nodes and seal dependency relationships as edges.

[0046] In summary, due to the adoption of the above-mentioned method and system for parsing seal watermark rules based on natural language processing, the beneficial effects of this invention are:

[0047] This invention uses a dependency reasoning model to fully identify explicit and implicit dependencies among supply chain participants, avoiding the misinterpretation of rule texts containing sequential dependencies as parallel stamping rules. This effectively prevents the chain of dependencies from breaking, ensuring that the contract stamping process of suppliers, manufacturers, logistics providers, and banks proceeds according to the preset dependency chain rules, reducing supplier financing delays, and maintaining supplier cash flow and production progress.

[0048] This invention automatically processes multimodal change notifications sent by supply chain participants through a resolution model, extracts change rules and updates dependency chains, eliminating the need for manual modification of dependency chain configurations, shortening change response time, avoiding the misjudgment of new participant stamps as invalid, and reducing cooperation disputes between participants.

[0049] This invention uses a health warning model to collect historical and real-time data to assess the health of the dependency chain, providing early warnings of the risk of dependency chain breakage. This prevents subsequent participants from being stuck when one participant fails to complete the stamping on time, thus shortening the contract approval cycle. Through a verification algorithm, it detects logical anomalies, compliance anomalies, timing anomalies, ambiguity resolution anomalies, and automatic stamping logic anomalies, triggering an iterative optimization process of the dependency reasoning model. This improves the accuracy of dependency chain parsing and automatic stamping logic matching, preventing dependency chain parsing vulnerabilities from being maliciously exploited, and reducing the risk of bank fraud and bad debts.

[0050] This invention synchronously associates the automatic seal identifier with the triggering logic during the parsing process. The constructed dependency chain semantic graph contains automatic seal association attributes, realizing the association mapping between the dependency chain and the automatic seal system, ensuring accurate triggering of the automatic seal, reducing automatic seal misoperation, and improving the efficiency of automated contract stamping. The resolution model includes a fuzzy completion module, which can complete the fuzzy time, fuzzy participants, and fuzzy automatic seal information in the change notification and evaluate the confidence of the completion result, ensuring the accuracy of the change information, providing reliable data support for the dynamic update of the dependency chain, and further improving the robustness of change processing. Attached Figure Description

[0051] Figure 1 A flowchart of the present invention is shown;

[0052] Figure 2 A system block diagram of the present invention is shown. Detailed Implementation

[0053] The following will describe, with reference to the accompanying drawings of the embodiments of the present invention, the seal watermark rule parsing method and system based on natural language processing of the present invention clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0054] To more clearly and intuitively demonstrate the practical application effects and advantages of the seal watermark rule parsing method and system based on natural language processing, and to verify its feasibility and effectiveness, the invention is further described below with reference to embodiments. Through specific scenario simulations and data calculations, the method is explained in detail how it plays a role in seal watermark rules, helping readers better understand the technical details and practical value of the invention. The invention is further described below with reference to embodiments;

[0055] Example 1:

[0056] See Figure 1 - Figure 2 To eliminate interference from the seal pattern and enhance the clarity of the watermark text, the following preprocessing operations need to be performed on the acquired watermarked supply chain finance contract seal image: The Otsu adaptive binarization algorithm is used to convert the color / grayscale seal image into a black-and-white binary image. By setting an optimal grayscale threshold, paper texture, slight stains, and background noise are removed, highlighting the grayscale difference between the watermark text and the seal pattern. The Sobel edge detection algorithm is used to perform convolution operations on the binary image to strengthen the edge contours of the watermark text characters, solving the problem of blurred text edges caused by watermark overlap and improving the accuracy of subsequent OCR recognition. A region-based threshold segmentation algorithm is used to separate the seal pattern region and the watermark text region according to the spatial distribution characteristics of the seal pattern and the watermark text. The final output is an image region containing only the watermark text, providing high-quality input for subsequent text processing in the element extraction layer.

[0057] The dependency reasoning model employs a four-layer mechanism—domain-adaptive encoding, hierarchical role labeling, causal reasoning, and multi-dimensional verification—to fully acquire the identification results of explicit and implicit dependencies, thus resolving the fragmentation problem of dependency chains. Building upon the original four-layer mechanism, the dependency reasoning model adds a term disambiguation module and a sentence parsing module, forming a six-layer collaborative mechanism that significantly improves the accuracy of dependency chain parsing. Simultaneously, it integrates with an automatic stamping system to achieve seamless integration of parsing results with stamping execution. The specific steps for constructing the dependency reasoning model are as follows:

[0058] Step 0A involves the element extraction layer, which employs a two-stage, bidirectional Transformer architecture combining basic semantic encoding and domain-adaptive encoding. In the first stage, the element extraction layer pre-trains weights on a general corpus to perform word segmentation and basic semantic encoding on the watermark rule text, capturing the text's general semantic features. In the second stage, the element extraction layer introduces corpus from the supply chain finance domain for domain-adaptive training, adjusting the semantic encoding weights to ensure that the dependent inference model prioritizes recognizing supply chain finance-specific terms such as supplier's official seal, manufacturer's acceptance seal, logistics provider's carrier seal, and bank's loan seal. Simultaneously, it identifies the corresponding automatic seal identifiers for each specific term, with the supplier's official seal corresponding to the supplier's automatic electronic seal and the manufacturer's acceptance seal corresponding to the manufacturer's automatic acceptance seal. After encoding is completed, the element extraction layer uses a hierarchical role labeling algorithm to first label the participating roles and stamping action roles in the text. Participating roles include suppliers, manufacturers, logistics providers, and banks. Stamping action roles include the stamping operation of each participant's exclusive seal and the automatic stamping operation. Then, based on part-of-speech association rules and semantic distance calculation, time constraint roles are labeled. Time constraint roles include expressions of precedence, following, synchronization with, and until completion, ensuring that the labeling of all three types of roles is complete and the associations are accurate. Step 0B involves the relational reasoning layer, which constructs a collaborative reasoning mechanism combining attention mechanisms, a causal reasoning rule base, and dependency link verification. The attention mechanism focuses on the associated text areas of time-constrained roles, participating roles, and stamping action roles, strengthening the weight of key semantics and extracting explicit dependencies. Explicit dependencies include explicit automatic stamping trigger dependencies, such as the supplier's automatic electronic stamping triggering the manufacturer's automatic acceptance stamping. The causal reasoning rule base incorporates core business logic rules of supply chain finance and automatic stamping triggering logic rules. Core business logic rules of supply chain finance include: loan disbursement requires confirmation of shipment, and acceptance requires official stamping. Automatic stamping triggering logic rules include: automatic stamping requires the completion of a previous automatic stamping step. Based on the rules in the rule base... Causal logic is used to uncover implicit dependencies in rule texts. These implicit dependencies include implicit automatic stamp trigger dependencies, such as the implicit prerequisite that the automatic stamping of a bank's automatic loan disbursement stamp is completed by a logistics provider. The dependency link verification module checks the integrity of the extracted explicit dependencies and the uncovered implicit dependencies, while also verifying the integrity of the automatic stamp trigger link. This ensures that each dependency has a complete action trigger source, action receiver, and constraint link, and that each automatic stamp trigger dependency has a complete link of preceding automatic stamps, subsequent automatic stamps, and trigger conditions, avoiding isolated dependencies without a source and isolated automatic stamp trigger logic.

[0059] Step 0C involves constructing a polysemous dictionary of supply chain finance terms using a terminology disambiguation module. This dictionary records the mapping between the scenarios and meanings of polysemous terms, as well as the mapping between the corresponding automatic stamp identifiers. When parsing rule text, the terminology disambiguation module accurately labels polysemous terms and their corresponding automatic stamp identifiers through contextual scenario recognition and role association verification. Contextual scenario recognition includes identifying scenarios such as the manufacturer completing acceptance, which identifies the scenario to which the acceptance stamp belongs. Role association verification ensures that the logistics provider role can only be associated with the logistics provider's acceptance stamp action and the corresponding logistics provider's automatic acceptance stamp identifier. For example, in a scenario where the manufacturer completes acceptance, the acceptance stamp corresponds to the manufacturer's acceptance stamp and the manufacturer's automatic acceptance stamp identifier; in a scenario where the logistics provider accepts goods, the acceptance stamp corresponds to the logistics provider's acceptance stamp and the logistics provider's automatic acceptance stamp identifier, ensuring that the meanings of polysemous terms are not confused with their corresponding automatic stamp identifiers.

[0060] Step 0D involves constructing an ambiguous sentence template library using a sentence parsing module. This library includes typical ambiguous templates from supply chain finance scenarios, as well as templates containing automatic stamp triggering logic. Typical ambiguous templates include: Participant A and Participant B take actions followed by Participant C's action, or Participant A or Participant B take actions followed by Participant C's action. Ambiguous templates containing automatic stamp triggering logic include: Participant A automatically stamps their seal, and Participant B automatically stamps their seal, followed by Participant C automatically stamping their seal. When parsing rule text, the sentence parsing module first matches ambiguous sentences in the ambiguous sentence template library, then supplements the context semantics and associates it with the causal reasoning rule library to determine the true logic of the ambiguous sentences, and simultaneously determines the corresponding automatic stamp triggering logic.

[0061] Contextual semantic supplementation: For example, subsequent statements in the text that require both to be completed before they can be triggered can clearly establish a parallel relationship between them;

[0062] The causal reasoning rule base is associated with the following: For example, the necessity of the rule that bank loans require dual confirmation from both the logistics provider and the manufacturer is clearly defined. For example, when parsing the bank loan after the logistics provider and the manufacturer have stamped the document, the sentence parsing module determines the true logic of the sentence based on the business rule that bank loans require dual confirmation. It determines that the bank loan is triggered after both the logistics provider and the manufacturer have completed stamping. At the same time, it determines the triggering logic of the corresponding automatic seal. It determines that the bank's automatic loan seal is triggered after both the logistics provider's automatic seal and the manufacturer's automatic acceptance seal are affixed. This avoids misjudgment of sentence logic and automatic seal triggering logic.

[0063] Step 0E involves constructing a multi-partner dependency chain semantic graph based on the explicit and implicit dependencies output by the dependency reasoning model, while incorporating automatic stamp association information. The multi-partner dependency chain semantic graph uses participants as nodes and dependencies as edges. In addition to explicit or implicit dependency identifiers and ambiguity resolution markers, edge attributes include an automatic stamp association attribute. This attribute records the automatic stamp identifier and triggering condition for each dependency. For example, an automatic stamp identifier might link a supplier's automatic electronic stamp to a manufacturer's automatic acceptance stamp. An automatic stamp triggering condition might be triggered by the manufacturer's automatic acceptance stamp being affixed after the supplier's automatic electronic stamp is affixed. By adding this automatic stamp association attribute, the dependency chain and the automatic stamp system are mapped together, providing a basis for subsequent automatic stamp triggering and execution. This also enables the visualization of the dependency chain, the traceability of the parsing process, and the verification of the automatic stamp logic.

[0064] Step 0F, the accuracy verification step, employs a verification algorithm consisting of logical consistency verification, business compliance verification, timing rationality verification, ambiguity resolution result verification, and automatic seal logic verification.

[0065] The logical consistency check compares the dependency relationship with the logical rules of no circular dependency and no contradictory dependency. If a circular dependency is detected from bank stamp to logistics provider stamp and from logistics provider stamp to bank stamp, it is immediately marked as a logical anomaly.

[0066] The business compliance verification calls the supply chain finance business compliance library. If the dependency model resolves that the bank's loan stamp was affixed earlier than the logistics provider's carrier stamp, it is judged as a compliance anomaly because it violates the compliance rule that bank loan disbursement must be based on carrier confirmation.

[0067] The timing rationality check is based on the timing logic of first satisfying the constraints and then executing the action. If a dependency relationship is detected where the manufacturer's acceptance stamp is affixed before the supplier's official stamp is affixed, it is judged as a timing abnormality.

[0068] The ambiguity resolution results are verified by comparing the meaning of the terms after ambiguity resolution with the standard of matching the context of the terminology meaning and the business rules of the sentence logic. If there is a mismatch, it is judged as an ambiguity resolution anomaly.

[0069] The automatic seal logic verification compares the automatic seal triggering logic with the rules that automatic seals require the completion of a prior automatic seal and that the automatic seal identifier matches the role of the participating party. If it is detected that the bank's automatic loan seal has no prior automatic seal triggering conditions or that the logistics provider's role is associated with the manufacturer's automatic acceptance seal identifier, it is determined that the automatic seal logic is abnormal.

[0070] If any anomaly exists in the verification algorithm, the system triggers the dependency reasoning model iterative optimization process, adjusts the domain adaptation weight of the element extraction layer, the attention weight of the relation reasoning layer and the automatic seal association mapping rule, and re-executes the parsing process until the dependency chain parsing accuracy and the automatic seal logic matching accuracy both meet the preset business requirements.

[0071] The conflict resolution model achieves robust handling of dynamic changes among supply chain participants through precise multimodal processing, dynamic modality weighting, and a segmented conflict resolution mechanism. It simultaneously handles changes related to the automatic stamping system, ensuring consistency between the dependency chain and the automatic stamping system after changes. Building upon the original three mechanisms, the model adds a fuzzy completion module and a collaborative resolution module, forming a five-mechanism collaborative processing architecture. This significantly enhances the system's responsiveness to dynamic changes, conflict handling capabilities, and automatic stamping change adaptation capabilities. The specific steps for constructing the conflict resolution model are as follows:

[0072] Step 1A involves sending change notifications containing text, tables, and attached images to supply chain participants through a cross-modal information fusion layer. These notifications include changes related to automatic seals, such as changing the logistics provider's automatic seal identifier and adjusting the automatic seal triggering order. This establishes a processing mechanism that enables modal-specific processing, weighted fusion, and joint extraction of elements, while also enhancing the ability to extract elements related to automatic seal changes.

[0073] The text modality employs a domain-adaptive text encoder pre-trained on a supply chain change notification corpus to extract elements such as the change initiator, change description, and change effective time from the change notification. It also extracts automatic stamp change elements, including the old automatic stamp identifier, the new automatic stamp identifier, and the automatic stamp triggering order adjustment requirements.

[0074] The table modality uses a structural parsing network to identify the participant type column, seal type column, and change status column in the table. The seal type column includes an automatic seal type column. For example, the logistics provider type column in the table corresponds to the original logistics provider and the new logistics provider. The automatic seal type column corresponds to the automatic seal carried by the original logistics provider and the automatic seal carried by the new logistics provider. The change status column corresponds to disabled and enabled. It accurately extracts the old and new participant information in the table, and also extracts the old automatic seal identifier and the new automatic seal identifier information.

[0075] The attached image modality employs fine-grained image segmentation and text recognition fusion technology. The attached image modality includes scanned documents containing change instructions and sample images of new automatic seals. First, the text area and seal pattern area in the image are segmented. The seal pattern area contains the automatic seal pattern area. Then, optical character recognition is performed on the text area to extract the changed text. The changed text contains the automatic seal change text. The validity of the old and new seals is confirmed by seal feature comparison of the seal pattern area. At the same time, the validity and uniqueness of the old and new automatic seals are confirmed. The seal feature comparison includes the seal name, pattern style, and automatic seal exclusive identifier.

[0076] A cross-modal information fusion layer constructs an effectiveness evaluation module. This module assesses the effectiveness of each modality based on element completeness, information clarity, and relevance to the change theme. Element completeness refers to whether the change elements included in the modality are complete, including the completeness of automatic seal change elements. Information clarity refers to whether the text and images in the modality are clear and identifiable, including the clarity of the automatic seal image. Relevance to the change theme refers to the degree of relevance between the modal information and the core change requirements, including the degree of relevance to the automatic seal change requirements. For example, if the text in the change notification is blurry but the table data is complete, and the table data contains complete new automatic seal identification data, the text modality has low effectiveness, while the table modality has high effectiveness. The system automatically tilts the weighted fusion towards the table modality. Subsequently, the attention fusion mechanism maps the features extracted from the text modality, table modality, and attachment image modality into a unified feature vector based on the effectiveness weights of each modality. The features extracted from the multimodality include automatic seal change features, avoiding information bias and omission of automatic seal change elements caused by fixed weights.

[0077] The text classification sub-model identifies change types based on a unified feature vector. Change types include changing suppliers, changing manufacturers, changing logistics providers, changing banks, and automatic seal change types, which include changing the automatic seal identifier and adjusting the automatic seal triggering order.

[0078] The named entity recognition sub-model adopts a strategy that combines entity boundary refinement with domain entity dictionary matching to accurately locate the entity boundaries of old participants, new participants, old seals, and new seals. At the same time, it accurately locates the entity boundaries of old automatic seal identifiers and new automatic seal identifiers. For example, it distinguishes the entity differences between automatic seal A carried by the original logistics provider and automatic seal B carried by the new logistics provider, avoiding entity confusion and automatic seal identifier confusion, and ensuring accurate extraction of change elements and automatic seal change elements.

[0079] Step 1B: The fuzzy completion module addresses the fuzzy time, fuzzy participants, and fuzzy auto-stamp information in the change notification. Fuzzy time includes recent and before the end of the month; fuzzy participants include new cooperating logistics providers and alternative suppliers; and fuzzy auto-stamp information includes new logistics providers whose auto-stamps are not clearly identified and adjustments to the auto-trigger order are not clearly defined. A fuzzy information completion rule base is constructed, and the auto-stamp fuzzy information completion rules are also incorporated. The auto-stamp fuzzy information completion rules include that the auto-stamp identifier of new participants must match the participant type, and the auto-stamp trigger order must conform to the core dependency chain order.

[0080] For ambiguous time periods, the fuzzy completion module completes the ambiguous time period into a specific time range by associating the historical change cycle of the participants with the contractually agreed cycle. For example, the historical change cycle of the participants is more likely to be implemented in the latter half of each month for manufacturers, and the contractually agreed cycle is likely to be implemented in the contract terms, which require one week's advance notice for changes. The specific time range is likely to be completed in the latter half of the month, such as the 20th to the 30th of this month.

[0081] For ambiguous participants, the fuzzy completion module completes the ambiguous participants into explicit participants by associating the participants' business cooperation list with clues in the change notice attachments. For example, the business cooperation list of the participants is a list of logistics partners that the manufacturer has registered. Clues in the change notice attachments are such as the qualification numbers of the logistics partners mentioned in the attachments. Explicit participants' names are such as new logistics partners being completed as Logistics Partner B.

[0082] For fuzzy automatic seal information, the fuzzy completion module completes the fuzzy automatic seal information into explicit information by associating the automatic seal registration database of participating parties with the automatic seal triggering logic of the core dependency chain. For example, the automatic seal registration database of participating parties includes a logistics provider's automatic seal type of "carrier automatic seal." The automatic seal triggering logic of the core dependency chain includes a manufacturer's automatic acceptance seal triggering a logistics provider's carrier automatic seal. Explicit information includes, for example, a new logistics provider's automatic seal being completed as "Logistics Provider B's carrier automatic seal," or adjusting the automatic triggering order to complete it as the manufacturer's automatic acceptance seal triggering a logistics provider B's carrier automatic seal. Simultaneously, the fuzzy completion module performs a confidence level judgment on the completion result, including the automatic seal completion result. If the confidence level is higher than a preset threshold, the completion result is automatically confirmed; if the confidence level is lower than the preset threshold, a manual confirmation process is triggered to avoid completion deviations and errors in automatic seal information completion.

[0083] Step 1C involves constructing a conflict resolution layer that includes conflict type subdivision, dedicated resolution algorithms, and resolution result verification, while also incorporating conflicts related to automatic stamping into the conflict type and resolution scope.

[0084] Based on supply chain change scenarios, the conflict resolution layer categorizes conflicts during dependency chain changes into five types: dependency missing conflicts, redundancy conflicts, logical reversal conflicts, overlapping permissions conflicts, and automatic stamp association conflicts. Dependency missing conflicts occur when a new participant fails to establish dependencies with upstream and downstream participants; for example, after changing logistics providers, the new logistics provider fails to establish a dependency with the bank. Redundancy conflicts occur when old participants are not completely removed, resulting in simultaneous dependencies between old and new participants; for example, both the original and new logistics providers have dependencies with the manufacturer. Logical reversal conflicts occur when the new participant's dependency relationship is in the reverse order of core business rules; for example, the new logistics provider's dependency chain is from bank stamping to new logistics provider stamping, violating the core rule of logistics provider stamping to bank stamping. Overlapping permissions conflicts occur when a new participant has dependencies with other participants... Overlapping stamping permissions among parties, for example, both the new logistics provider and the manufacturer have the authority to trigger the bank's stamp; automatic seal association conflicts refer to mismatches between the automatic seal identifier and the participating party's role, and conflicts between the automatic seal triggering logic and the dependency chain. For example, the automatic seal identifier is associated with the automatic acceptance stamp identifier, which violates the automatic seal rules for the corresponding carrier of the logistics provider. Conflicts between the automatic seal triggering logic and the dependency chain, for example, the automatic seal triggering logic for the new logistics provider is triggered after the bank's automatic loan disbursement stamp, which violates the rule that the automatic seal for the logistics provider is triggered after the automatic seal for the carrier is triggered by the bank's automatic loan disbursement stamp.

[0085] To address dependency missing conflicts, the conflict resolution layer invokes a core rule matching and completion algorithm to automatically match core dependency rules from manufacturers to logistics providers to banks, completing the dependencies between new participants and their upstream and downstream partners. For redundant conflicts, the layer employs a change effective time priority algorithm, retaining new participant dependencies within the effective time frame specified in the change notification and deleting old ones. For logical reversal conflicts, the layer initiates a core rule forced correction algorithm to adjust reversed dependencies to conform to the core business rules. For overlapping permission conflicts, the layer invokes a permission attribution dictionary, using permission attribution terms... The system clearly defines the exclusive stamping permissions of each participant, removes overlapping permissions, and ensures that each participant's permissions are unique. In case of automatic seal association conflicts, the conflict resolution layer calls the automatic seal association rule library. The automatic seal association rule library clarifies the correspondence between participant roles, automatic seal types, and triggering logic, and corrects automatic seal identifiers. For example, the automatic acceptance seal identifier of the new logistics provider is changed to the carrier automatic seal identifier. The automatic seal triggering logic is also adjusted. For example, the triggering logic after the bank automatic loan disbursement seal of the new logistics provider's carrier automatic seal is changed to the triggering logic after the manufacturer's automatic acceptance seal, ensuring that the automatic seal association information is consistent with the participant roles and dependency chain logic.

[0086] The conflict resolution layer constructs a conflict resolution result verification library. The conflict resolution result verification library contains verification rules for the following: after resolution, the dependency chain has no logical contradictions, the new dependency chain conforms to the core business rules, there are no unprocessed conflict residues, and the automatic seal association information has no abnormalities. The library verifies the resolved dependency relationship and the automatic seal association information to ensure that the conflict is completely eliminated, the automatic seal association information is accurate, and no new conflict is generated.

[0087] Step 1D involves using the collaborative resolution module to address scenarios where multiple parties initiate changes simultaneously. For example, when a manufacturer changes its logistics provider, and a bank adjusts the carrier stamp type required for loan disbursement, the automatic loan disbursement stamp triggering logic is simultaneously adjusted. A collaborative change priority rule base is constructed, incorporating automatic stamp collaborative change priority rules. These rules prioritize changes to the automatic stamp triggering logic over changes to the automatic stamp configuration of a participating party. The collaborative change priority rule base clearly defines that changes to core participants take precedence over changes to auxiliary participants, business-critical changes take precedence over routine changes, and changes to the core logic of the automatic stamp take precedence over changes to the automatic stamp identifier. Specifically, changes to core participants take precedence over changes to auxiliary participants; for example, a bank as a core participant has its loan disbursement rule changes taken precedence over a logistics provider change. Business-critical changes take precedence over routine changes; for example, changes to loan dependence take precedence over changes to the participant name. Changes to the core logic of the automatic stamp take precedence over changes to the automatic stamp identifier; for example, changes to the automatic loan disbursement stamp triggering logic take precedence over the automatic stamp identifier configuration of the new logistics provider. When handling collaborative changes, the collaborative resolution module first identifies the type of each participant, the changed business attributes, and the automatic seal change attributes. Participant types include core participants or auxiliary participants; changed business attributes include critical business or routine business; and automatic seal change attributes include core logic changes or identifier changes. Then, based on the collaborative change priority rule base, the change order is sorted, and the change requests and automatic seal change requests of each participant are processed sequentially. For example, changes to the carrier seal type and automatic loan disbursement seal triggering logic that bank loan disbursements depend on are processed first, followed by changes to the manufacturer changing logistics providers and the new logistics provider's automatic seal identifier configuration. This avoids dependency chain conflicts and automatic seal logic confusion caused by simultaneous changes. Meanwhile, the collaborative resolution module has added a collaborative change chain impact assessment function. After processing a change of a participant, it automatically assesses whether the change will affect the change requirements and automatic seal change requirements of other participants. For example, if a bank adjusts the automatic loan seal trigger logic to require confirmation of the automatic seal by a new logistics provider, the new logistics provider replaced by the manufacturer needs to configure the corresponding automatic seal identifier. This allows for the early correction of change elements and automatic seal change elements, avoiding chain conflicts and abnormal automatic seal associations.

[0088] Step 1E involves performing a dynamic update operation on the dependency chain. This dynamic update is based on the conflict-free change elements output by the resolution model, the resolved dependency relationships, and the corrected automatic stamp association information. It constructs a dynamic update mechanism that includes incremental map updates, permission synchronization verification, historical version retention, and automatic stamp information synchronization updates. Incremental graph updates employ incremental operations on nodes and edges to avoid a full reconstruction of the dependency chain semantic graph. For example, when changing logistics providers, only the original logistics provider node and its associated edges are deleted, and a new logistics provider node and its corresponding dependent edges are added. Simultaneously, the automatic seal association attribute is incrementally updated, deleting the automatic seal identifier from the original logistics provider's dependent edges and adding it to the new logistics provider's dependent edges, thus improving update efficiency. Permission synchronization verification simultaneously verifies the stamping and automatic seal permissions of new participants while updating the dependency chain. For example, a new logistics provider must pass both qualification verification and seal registration verification to be granted the right to affix its carrier seal. Similarly, a new logistics provider must pass automatic seal identifier registration and trigger logic configuration verification to be granted automatic seal operation permissions, preventing unqualified participants from entering the dependency chain and unauthorized participants from gaining access to the system. Automatic stamp operation permissions are granted; historical version retention: Historical versions are retained in the dependency chain semantic graph after each change, along with the corresponding automatic stamp association information. This records the change time, content, executor, and automatic stamp change details, facilitating subsequent change traceability, error rollback, and automatic stamp change auditing. For example, if a change error is discovered, a quick rollback to the previous version's dependency chain and corresponding automatic stamp association information is possible; automatic stamp information is synchronized and updated: After completing the dependency chain graph update, the system automatically calls the automatic stamp system's information synchronization interface to synchronize the updated automatic stamp association information to the automatic stamp system. The updated automatic stamp association information includes the new logistics provider's automatic stamp identifier and the new logistics provider's automatic stamp trigger logic, ensuring consistency between the dependency chain and the automatic stamp system, providing data support for accurate automatic stamp triggering in the future.

[0089] Step 1F involves performing an accuracy verification operation for change processing. The accuracy verification of change processing is carried out through a multi-dimensional verification method, which includes the effectiveness of fuzzy information completion, the effectiveness of collaborative conflict resolution, the accuracy of automatic stamp information, and the original verification dimensions. The original verification dimensions include logical consistency, business compliance, and time sequence rationality. The effectiveness of ambiguous information completion is assessed by matching the completion result with subsequent explicit notifications. For example, if the completed information (recent period) matches the explicit notification that the change will take effect on the 25th of this month, and the completed new logistics provider's automatic seal (Logistics Provider B's automatic seal) matches the subsequent notification confirming the registration of Logistics Provider B's automatic seal, then the completion is deemed effective. The effectiveness of resolving collaborative conflicts is assessed by evaluating the logical consistency of the dependency chain after the collaborative change. For example, if there are no cross-conflicts or logical contradictions in the dependency chain after the collaborative change between the bank and the manufacturer, then the resolution is deemed effective. The accuracy of automatic seal information is assessed by matching the automatic seal identifier with the participant's role and the automatic seal triggering logic with the dependency chain logic. For example, if the new logistics provider's automatic seal identifier is the carrier's automatic seal and the triggering logic is triggered after the manufacturer's automatic acceptance seal, it matches both the logistics provider's role and the dependency chain logic, thus the automatic seal information is deemed accurate. The original verification dimensions continue to verify the dependency chain logic, compliance, and timing. Through multi-dimensional verification, we ensure that fuzzy information is accurately completed, collaborative change conflicts are thoroughly resolved, automatic seal information is matched without error, and the dependency chain logic is intact and consistent with the information of the automatic seal system after the change.

[0090] The health status early warning model achieves accurate early warning of dependency chain breakage risks through deep fusion of multi-dimensional features, intelligent adjustment of dynamic health status thresholds, and an early warning closed-loop optimization mechanism. It also integrates with an automatic stamping system to enable automatic stamping intervention after an early warning. The health status early warning model focuses on balancing accuracy and efficiency. By simplifying the feature fusion layer, dynamic early warning layer, and early warning execution process, it improves response efficiency while ensuring accurate early warnings and reducing manual review costs and the risk of accidental automatic stamping. The specific steps for constructing the health status early warning model are as follows:

[0091] Step 2A involves performing a feature fusion layer operation. This layer includes key temporal feature encoding and core business feature embedding, while also incorporating features related to automatic seal operation. These features include the historical success rate of automatic seal triggering and the delay duration of automatic seal operation. Key temporal feature encoding divides historical automatic seal time data into two time dimensions: a period of the past week and a period of the past 1 to 4 weeks. The historical automatic seal time data includes historical automatic seal time data. Two independent long short-term memory network branches are used to perform temporal encoding on the data for each dimension. The data for the past week reflects the current seal efficiency and the current trigger efficiency of the automatic seal, while the data for the past 1 to 4 weeks reflects the historical average efficiency and the historical average trigger efficiency of the automatic seal, capturing the seal patterns and automatic seal operation patterns at different time scales. Core business feature embedding collects real-time business data from participants and real-time status data of the automatic seal. Real-time business data includes the urgency of the contract and the busyness of the participant's business. Real-time status data of the automatic seal includes... The system's online status and automatic seal operation queue length are used to construct a business feature dictionary. Discrete business features are converted into continuous feature vectors through an embedding layer. Discrete business features include business activity levels (idle, normal, busy) and the automatic seal system's online status (online, offline). Continuous business features are integrated into the feature space through normalization processing. Continuous business features include quantitative indicators of contract urgency and the automatic seal operation queue length. Finally, a feature fusion layer uses a lightweight fully connected layer to fuse temporal coding features, business embedding features, and automatic seal-related features into a unified health feature vector according to preset weights. This avoids interference from secondary features and improves feature processing efficiency and health assessment accuracy. The health assessment accuracy includes the accuracy of automatic seal operation risk assessment.

[0092] Step 2B involves a dynamic early warning layer. Based on the fused health feature vector, the dynamic early warning layer trains a health predictor using a gradient boosting tree algorithm. The predictor outputs the expected stamping time and the expected triggering time for automatic stamping for participating parties. The expected stamping time is the core early warning indicator, while the expected triggering time for automatic stamping is an auxiliary early warning indicator. Simultaneously, the dynamic early warning layer constructs a dynamic health threshold generation algorithm. Adjustments are made based on contract urgency, participating party workload, and automatic stamping system load to dynamically optimize health thresholds. For example, the health threshold for urgent contracts is higher than for ordinary contracts to avoid delays; the health threshold for busy participating parties is appropriately lowered to avoid false warnings due to workload; and the health threshold related to automatic stamping operations is appropriately increased when the automatic stamping system is under high load to prevent system overload.

[0093] Step 2C involves executing the early warning operation. The early warning execution mechanism includes a closed-loop execution mechanism encompassing tiered early warning response, backup participant linkage, early warning result feedback, and automatic stamping operation intervention.

[0094] The health predictor (trained based on gradient boosting tree algorithm) maps the unified health feature vector to the health score of the participants from 0 to 100, where the higher the health score of the participants, the better their health.

[0095] The tiered early warning response classifies warnings into three levels: mild, moderate, and severe, based on the difference between the participant's health score and the dynamic health threshold.

[0096] A mild warning only sends a reminder notification to the participants, explaining the difference between the current waiting time and the expected time. At the same time, it sends a status reminder to the automatic stamping system. The status reminder may be that the automatic stamping of participant A may be delayed, which does not affect the downstream process and the automatic stamping operation.

[0097] The moderate warning sends a reminder notice to the participants, which includes an explanation of the impact of the delay on the downstream process, such as the logistics provider being unable to carry the goods on time due to your delay in stamping. At the same time, it sends a pending trigger notice to the downstream participants, informing them that they need to wait for the upstream to complete the stamping. It also sends a delay trigger preparation instruction to the automatic stamping system, such as temporarily storing the downstream participants' automatic stamping trigger request.

[0098] In addition to sending urgent reminders, severe warnings will suspend downstream participants' stamping requests and automatic stamping trigger requests to avoid invalid waiting and accidental triggering of automatic stamps.

[0099] When a participant triggers a severe warning and fails to complete the stamping operation within a preset time frame, the system automatically calls upon the backup participant database. It matches alternative participants based on their type, such as a backup logistics provider for the manufacturer. A connection request is sent to the backup participant, simultaneously pushing basic contract information and the stamping requirements. At the same time, the system automatically activates the automatic stamping system module to temporarily activate the automatic stamping permissions corresponding to the pending stamping operation for the backup participant. For example, it activates temporary automatic stamping permissions for a backup logistics provider, ensuring that the backup participant can immediately perform the stamping and automatic stamping operations, preventing a complete break in the dependency chain. The early warning result feedback module records the early warning level, response measures, actual results, and automatic stamp operation status for each early warning. For example, if a participant stamps the stamp on time and the automatic stamp is triggered normally after a mild early warning, it is marked as an effective early warning with no loss and the automatic stamp is normal. If a backup participant is activated after a severe early warning and the backup participant's automatic stamp is successfully triggered, it is marked as an effective early warning that reduces losses and the automatic stamp is successfully activated. False early warnings are marked as invalid early warnings that require threshold adjustment and the automatic stamp is not falsely triggered. The feedback data is periodically used to update the feature weights of the health predictor, the adjustment parameters of the dynamic health threshold generation algorithm, and the automatic stamp operation intervention strategy, thereby achieving iterative optimization of the early warning model and the automatic stamp operation intervention logic and reducing the subsequent false early warning rate and automatic stamp misoperation rate.

[0100] Example 2:

[0101] See Figures 1-2 The seal watermark rule parsing system based on natural language processing includes: a seal image module, an image preprocessing module, an extraction module, a semantic parsing module, a semantic graph construction module, a resolution module, a health status warning module, and a parsing accuracy verification module. These modules work collaboratively to achieve the parsing, change processing, risk warning, and accuracy verification of seal watermark rules in supply chain finance contracts, as detailed below:

[0102] The seal image module's core function is to acquire watermarked images of supply chain finance contract seals. This module interfaces with the supply chain finance contract management platform's image acquisition interface, receiving contract image files containing seal and watermark information uploaded by enterprise clients and bank contract signing clients. The module also performs format validation on the received image files, supporting common formats such as JPG, PNG, and PDF. If an unsupported format is detected, it immediately sends an error message to the uploading client, specifying the supported formats, ensuring the image can proceed smoothly into subsequent processes.

[0103] The image preprocessing module performs preprocessing operations on the seal image acquired by the seal image module. The aim is to enhance the clarity of the watermark text and eliminate interference from the seal pattern on subsequent text extraction. First, the image preprocessing module uses the Otsu adaptive binarization algorithm to convert the color or grayscale seal image into a black and white binary image, removing background noise. Then, it uses the Sobel edge detection algorithm to enhance the edge contours of the watermark text characters. Finally, it uses a region-based thresholding algorithm to separate the seal pattern region from the watermark text region, outputting an image region containing only the watermark text, providing high-quality input for the extraction module.

[0104] The extraction module extracts the watermark text from the stamp image processed by the image preprocessing module. Employing high-precision optical character recognition technology, the extraction module converts the watermark content from image format into editable text data. Simultaneously, it cleans the recognized text, removing redundant spaces, garbled characters, and invalid information, and corrects grammatical errors to ensure the accuracy and completeness of the extracted watermark text, providing a foundation for the semantic analysis module.

[0105] The semantic parsing module employs natural language processing technology to semantically analyze the watermark text extracted by the extraction module, identifying the roles of supply chain participants, seal types, and seal dependencies among participants. The semantic parsing module incorporates a dependency reasoning model, working through a six-layer collaborative mechanism: domain-adaptive coding, hierarchical role labeling, terminology disambiguation, sentence parsing, explicit and implicit dependency extraction, and dependency link verification. It prioritizes identifying supply chain finance-specific seal terminology and corresponding automatic seal identifiers, accurately labeling roles, resolving ambiguities, extracting and verifying dependencies, and avoiding isolated dependencies or isolated automatic seal triggering logic.

[0106] The semantic graph construction module builds a dependency chain semantic graph based on the recognition results of the semantic parsing module. The dependency chain semantic graph uses supply chain participants as nodes and seal dependency relationships as edges. The module first organizes the information output by the semantic parsing module, clarifying the types of graph nodes and the attributes of edges; then it inputs participants as nodes into the graph and completes the node information; next, it connects the corresponding nodes using dependency relationships and records the edge attributes; finally, it completes the graph visualization rendering and stores the graph data in a structured format to provide data support for subsequent modules.

[0107] The conflict resolution module handles dynamic change requests from supply chain participants, such as changing logistics providers or adjusting automatic stamp identifiers, ensuring that the dependency chain logic remains consistent and matches the information in the automatic stamp system after the change. The module has a built-in conflict resolution model that first processes multimodal change notifications (text, tables, and attached images) to extract change information; then, it converts the multimodal features into a unified feature vector; next, it identifies the change type, locates entity boundaries, and uses a fuzzy completion module to complete fuzzy information and assess confidence; finally, it uses a conflict resolution layer to handle various conflicts, and then verifies the resolution results using a conflict resolution result verification library to ensure complete conflict elimination.

[0108] The health alert module assesses the health of the multi-partner seal dependency chain, providing early warnings of the risk of chain breakage, and simultaneously initiating operational interventions in conjunction with the automatic seal system. The module incorporates a built-in health alert model. It first collects historical stamping times, historical automatic seal triggers, real-time business data from participating parties, and real-time automatic seal status data, processing and fusing them into a unified health feature vector. Then, a health predictor is trained through a dynamic alert layer, outputting the expected duration and dynamically adjusting the health threshold. Finally, it executes mild, moderate, and severe graded alert responses and interventions, recording alert results for model optimization to reduce false alarm rates and automatic seal misoperation rates.

[0109] The accuracy verification module verifies the accuracy of the watermark rule parsing results to ensure they meet business requirements. This module employs algorithms for logical consistency, business compliance, temporal rationality, ambiguity resolution, and automatic seal logic verification. If any logical or compliance anomaly is detected, an iterative optimization process based on the dependency inference model is immediately triggered, adjusting relevant weights and rules and controlling the system to re-execute the parsing process. If all dimensions are normal, the parsing results are confirmed as valid and synchronized to the supply chain finance contract management platform and the automatic seal system to support subsequent business processes.

[0110] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the natural language processing-based seal watermark rule parsing method and system and its inventive concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for parsing watermark rules based on natural language processing, characterized in that, include: Obtain the image of the seal on a watermarked supply chain finance contract; Perform preprocessing operations on the seal image; Extract watermark text from the preprocessed seal image; By enabling the dependency reasoning model and using natural language processing technology to perform semantic parsing on the extracted watermark text, the identification results of the roles of supply chain participants, seal types, and seal dependency relationships among participants are obtained. The dependent reasoning model adjusts the semantic encoding weights, prioritizes the identification of exclusive seal terms and corresponding automatic seal identifiers in the supply chain finance field, and marks the roles of supply chain participants, the role of the seal affixing action, and the role of time constraints. Based on the identification results, a dependency chain semantic graph is constructed, wherein the supply chain participants are nodes and the seal dependency relationship is an edge. The dependency chain semantic graph includes an auto-stamp association attribute, which is used to record the auto-stamp identifier and triggering conditions corresponding to each dependency relationship; The dependency reasoning model includes a relational reasoning layer, which constructs a collaborative reasoning mechanism that combines an attention mechanism, a causal reasoning rule base, and dependency link verification. The attention mechanism focuses on the associated text regions of time-constrained roles, participating roles, and stamping action roles, strengthening the weight of key semantics and extracting explicit dependencies, which include explicit automatic stamping trigger dependencies. The causal reasoning rule base incorporates the core business logic rules of supply chain finance and automatic stamping trigger logic rules. Based on the causal logic in the rule base, implicit dependencies in the rule text are mined, which include implicit automatic stamping trigger dependencies. The dependency link verification module performs a link integrity check on the extracted explicit dependencies and the mined implicit dependencies.

2. The seal watermark rule parsing method based on natural language processing according to claim 1, characterized in that, The preprocessing operations performed on the seal image include: A binarization algorithm is used to remove background noise from the seal image; Optimize the edge contour of the watermark text using an edge detection algorithm; A threshold segmentation algorithm is used to separate the stamp pattern area from the watermark text area. Output the image area containing the watermark text.

3. The seal watermark rule parsing method based on natural language processing according to claim 2, characterized in that, The dependent reasoning model also includes a term disambiguation module, which constructs a polysemous dictionary of supply chain finance terms to record the mapping relationship between the scenarios and meanings of polysemous terms and the corresponding automatic stamp identification mapping relationship. The terminology disambiguation module identifies polysemous terms and assigns them their corresponding automatic stamps based on contextual scene recognition and role association verification.

4. The seal watermark rule parsing method based on natural language processing according to claim 3, characterized in that, The dependency reasoning model also includes a sentence parsing module, which constructs a library of ambiguous sentence templates; Among them, the ambiguous sentence template library includes typical ambiguous templates containing automatic seal triggering logic in supply chain finance scenarios; After performing ambiguous sentence matching on the watermark text, the sentence parsing module determines the true logic of the ambiguous sentences and the corresponding automatic stamp triggering logic by supplementing the context semantics and associating with the causal reasoning rule base.

5. The method for parsing seal watermark rules based on natural language processing according to claim 1, characterized in that, It also includes steps for handling dynamic changes in supply chain participants: Enable the resolution model; Modal-specific processing is performed on change notifications sent by supply chain participants, which contain text, tables, and attached images, using a resolution model to obtain text modality, table modality, and attached image modality; the change initiator, change description, change effective time, and automatic stamp change elements are extracted; We perform weighted fusion of text modality, table modality, and attached image modality, and map the features of text modality, table modality, and attached image modality into a unified feature vector; The change type is identified based on the unified feature vector, and the entity boundaries of the old participants, new participants, and the corresponding old and new automatic seal identifiers are located.

6. The method for parsing seal watermark rules based on natural language processing according to claim 5, characterized in that, The resolution model also includes a fuzzy completion module; The fuzzy completion module addresses the fuzzy time, fuzzy participants, and fuzzy automatic seal information in the change notification by combining the historical change cycle of the associated participants, the business cooperation list, and the automatic seal filing database to complete the fuzzy information and obtain the completion result. The fuzzy completion module assesses the confidence level of the completion results and triggers a manual confirmation process when the confidence level is lower than a preset threshold.

7. The method for parsing seal watermark rules based on natural language processing according to claim 1, characterized in that, This also includes providing early warnings about the health of the dependency chain: A health warning model was activated, and historical stamping time data and automatic stamping historical trigger data were collected. The data were divided into two time dimensions: the past week period and the past 1 to 4 weeks period. The data of each dimension were time-series encoded through the branches of the long short-term memory network to obtain time-series encoded features. Collect real-time business data and real-time status data of automatic seals from supply chain participants to obtain automatic seal-related features. Convert discrete business features into continuous feature vectors and normalize the continuous business features to obtain business embedding features. The time-series coding features, business embedding features, and automatic stamping-related features are fused into a unified health feature vector by a lightweight fully connected layer according to preset weights. Based on the unified health feature vector, the expected stamping time of the participants and the expected triggering time of the automatic stamping are calculated.

8. The method for parsing seal watermark rules based on natural language processing according to claim 7, characterized in that, The health warning model also includes a dynamic warning layer. The dynamic warning layer uses a gradient boosting tree algorithm to train a health predictor, outputs the expected stamping time of the participants and the expected triggering time of the automatic seal, and maps the unified health feature vector to the health score of the participants through the health predictor. The health threshold is dynamically adjusted based on the urgency of the contract, the business load of the participants and the load of the automatic seal system.

9. The method for parsing seal watermark rules based on natural language processing according to claim 1, characterized in that, It also includes a step to verify the accuracy of the parsing: The verification algorithm consists of logical consistency verification, business compliance verification, timing rationality verification, ambiguity resolution result verification, and automatic seal logic verification. If a logical anomaly, compliance anomaly, timing anomaly, ambiguity resolution anomaly, or automatic stamping logic anomaly is detected, the dependency inference model iterative optimization process is triggered, the domain adaptation weight and automatic stamping association mapping rules are adjusted, and the parsing process is re-executed until the preset business requirements are met.

10. A seal watermark rule parsing system based on natural language processing, implementing the seal watermark rule parsing method based on natural language processing as described in any one of claims 1-9, characterized in that, include: The system includes a seal image module, an image preprocessing module, an extraction module, a semantic parsing module, and a semantic graph construction module. The seal image module is used to acquire watermarked images of supply chain finance contract seals. The image preprocessing module is used to perform preprocessing operations on the seal image acquired by the seal image module to enhance the clarity of the watermark text; The extraction module is used to extract the watermark text from the stamp image processed by the image preprocessing module; The semantic parsing module is used to perform semantic parsing on the watermark text extracted by the extraction module using natural language processing technology, and to obtain the identification results of the roles of supply chain participants, seal types and seal dependencies between participants; The semantic graph construction module is used to construct a dependency chain semantic graph based on the recognition results of the semantic parsing module. The dependency chain semantic graph uses supply chain participants as nodes and seal dependency relationships as edges.

Images