An electronic signature generation method and system for an examination and approval flow
By performing clause-level semantic vector comparison and legal impact assessment on electronically signed documents, the problem of existing systems being unable to detect unstructured text tampering has been solved. This has enabled effective interception of semantic tampering and differentiation of reasonable modifications, thereby improving the security and accuracy of signatures.
Patent Information
- Application Number
- CN202611243070.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-10
AI Technical Summary
Existing electronic signature systems cannot effectively detect semantic changes in unstructured text content, which means that tampered documents can still be signed, posing a security risk.
By acquiring structured and unstructured data of documents to be signed, a pre-trained text semantic encoding model is used to segment and vectorize the semantic units of clauses, calculate cosine similarity, establish a change baseline by combining historical approval records, quantify the level of legal impact, generate a comprehensive score for semantic tampering risk, and trigger the manual review process.
It effectively detects and blocks semantic tampering of unstructured text, avoids false interception of reasonable modifications, and improves the security and accuracy of signatures.
Smart Images

Figure CN123069451A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic signature technology, specifically to a method and system for generating electronic signatures for approval workflows. Background Technology
[0002] In corporate approval processes, approval documents must pass through multiple nodes sequentially, culminating in an electronic signature at the final node to confirm document validity. To prevent unauthorized modification during this process, existing electronic signature systems typically employ anomaly detection methods based on statistical deviations of structured fields. These methods monitor changes in quantifiable fields such as amount, date, and quantity, using this as a pre-signature verification basis. However, these methods only cover the structured field layer and are invisible to semantic changes in unstructured text. For approval documents containing unstructured text, such as contract approvals and policy change approvals, attackers can modify only restrictive clauses, performance conditions, or disclaimers within the text while keeping the structured field values unchanged. This substantially alters the document's actual legal meaning, and such semantic tampering cannot be detected by existing mechanisms. Consequently, documents with altered semantics can still pass detection and obtain electronic signatures, creating a security risk where the signature's validity does not match the document's actual content. Summary of the Invention
[0003] This invention provides a method and system for generating electronic signatures in approval processes, which solves the technical problem in related technologies that the generation of electronic signatures in approval processes lacks the ability to automatically detect and quantify the risks of semantic changes in the main text of documents, resulting in tampered approval documents being incorrectly stamped.
[0004] This invention discloses a method for generating electronic signatures for approval workflows, comprising the following steps: Obtain the complete content data of the document to be signed and approved, separate the set of structured field values from the set of unstructured text paragraphs, and obtain historical version snapshot data; The unstructured text paragraph set is segmented into clause semantic units according to the legal clause segmentation rules, and a pre-trained text semantic encoding model is called to calculate the semantic vector representation of each clause semantic unit; Perform the same segmentation and encoding process on the text content of the most recent preceding node version in the historical version snapshot data, perform clause-by-clause cosine similarity calculation on the clause semantic vector sequence of the current version and the preceding version, mark clauses with cosine similarity lower than the preset semantic change threshold as semantic change clauses, and generate a list of semantic change clause markers. For each changed clause in the semantic change clause tagging list, four key semantic elements are extracted: the obligor, the content of the rights, the restrictions, and the numerical constraints. The clauses are then compared and analyzed in a structured manner with the corresponding clauses in the previous version to determine the change type. The legal impact level is then calculated based on the basic impact weight and the normalized score of the change magnitude corresponding to the change type. Establish a normal change baseline by obtaining historical approval and modification records of similar business documents, compare the change type and legal impact level of each change clause with the normal change baseline, and accumulate the legal impact level of change clauses that exceed the normal change baseline to generate a comprehensive score for semantic tampering risk. The comprehensive score of semantic tampering risk is compared with the preset signature constraint threshold. When the comprehensive score of semantic tampering risk is lower than the preset signature constraint threshold, the electronic signature data is output. When the comprehensive score of semantic tampering risk reaches or exceeds the preset signature constraint threshold, the signature generation is stopped, and the audit prompt data containing the location mark of the changed clause, the change type and the quantitative value of the legal impact level is output, triggering the manual review process.
[0005] Furthermore, the clause-by-clause cosine similarity calculation includes: assuming the current version is... The semantic vector of each clause is The semantic vector of the corresponding clause in the previous version is The cosine similarity is calculated as follows: ; in, Let be the inner product of two vectors. For vectors Norm, ; will satisfy The terms are marked as semantic change terms, in which A preset semantic change threshold is set; in cases where there are clause additions or deletions between the current version and the previous version, resulting in inconsistent sequence lengths, the newly added clauses will be directly included in the semantic change clause marker list.
[0006] Furthermore, before performing clause-by-clause cosine similarity calculation, the process includes: based on the global similarity matrix between the semantic vector sequences of clauses in the current version and the previous version, using the dynamic programming longest common subsequence matching method to pair the clauses in the current version with the clauses in the previous version that are semantically closest, and then performing cosine similarity calculation on the pairing results.
[0007] Furthermore, the calculation of the legal impact level quantification value includes: setting the basic impact weight corresponding to the change type as follows: The normalized score for the change range is Then the first The legal impact level of each amended clause is quantified as follows: ; in, It is calculated by the weighted average of the proportion of the number of elements that change among the four categories of elements: subject of obligation, content of rights, restrictive conditions, and numerical constraints, relative to the total number of elements. The weight of each element category is pre-set according to its degree of influence on the legal meaning.
[0008] Furthermore, the calculation of the semantic tampering risk comprehensive score includes: ; in, This is a collection of change clauses that exceed the normal change baseline. For set The Middle The legal impact level quantification value of each amended clause; the normal amendment baseline is based on the distribution of types of historical clause amendments for similar business documents at the same approval node, and... The statistical range of values indicates that when the type of change of the change clause or If the scope exceeds this statistical range, the clause is deemed to exceed the normal change baseline.
[0009] Furthermore, the normal change baseline is statistically analyzed separately for each approval node, targeting the document from the node. Flow to node In such cases, maintain the corresponding normal distribution range parameters separately.
[0010] Furthermore, the step of comparing the comprehensive score of semantic tampering risk with the preset signature constraint threshold includes: setting a preset low-risk threshold. With high risk threshold ,in When the overall score for semantic tampering risk is lower than When the semantic tampering risk score is not lower than a certain threshold, the electronic signature data is output. And lower than When the risk level is not lower than a certain threshold, an electronic signature with a risk warning is generated, and the review warning data is synchronized to the current approval node for retention; when the comprehensive score of semantic tampering risk is not lower than a certain threshold... If this occurs, the signature generation process is suspended and a manual review process is triggered.
[0011] This invention discloses an electronic signature generation system for approval workflows, comprising: The data acquisition module is used to acquire the complete content data of the document to be signed and approved, separate the set of structured field values from the set of unstructured text paragraphs, and acquire historical version snapshot data; The semantic encoding module is used to segment the unstructured text paragraphs according to the legal clause segmentation rules, and to call the pre-trained text semantic encoding model to calculate the semantic vector representation of each clause semantic unit; The similarity comparison module is used to perform the same segmentation and encoding processing on the text content of the most recent preceding node version in the historical version snapshot data, and to perform clause-by-clause cosine similarity calculation on the clause semantic vector sequence of the current version and the preceding version to generate a list of semantic change clause tags; The impact assessment module is used to extract four key semantic elements—obligation subject, rights content, restrictions, and numerical constraints—from each modified clause in the semantic modification clause tag list, perform structured comparative analysis to determine the type of modification, and calculate the quantitative value of the legal impact level. The risk scoring module is used to obtain historical approval and modification records of similar business documents to establish a normal change baseline. It compares the change type and legal impact level of each change clause with the normal change baseline to generate a comprehensive score for semantic tampering risk. The signature decision module compares the comprehensive score of semantic tampering risk with the preset signature constraint threshold. When the comprehensive score of semantic tampering risk is lower than the preset signature constraint threshold, it outputs electronic signature data. When the comprehensive score of semantic tampering risk reaches or exceeds the preset signature constraint threshold, it stops signature generation and outputs audit prompt data containing the location mark of the changed clause, the change type, and the quantitative value of the legal impact level, triggering the manual review process.
[0012] The beneficial effects of this invention are as follows: This invention introduces clause-level semantic vector comparison before signature generation, incorporating the semantic consistency of unstructured text into the pre-verification conditions for signature generation. This solves the technical problem that existing detection mechanisms based on statistical deviations of structured fields cannot capture semantic-level tampering of the text, achieving the technical effect of identifying and intercepting semantic-level tampering attacks at the signature triggering stage. By assessing the substantive impact of changes through a legal impact level quantification value and establishing a normal change baseline based on historical approval modification records, the judgment process can distinguish between normal approval modifications and abnormal tampering, achieving the technical effect of reducing false interception of reasonable clause revisions. Attached Figure Description
[0013] Figure 1 This is a flowchart of the electronic signature generation method for approval workflow provided in an embodiment of the present invention. Detailed Implementation
[0014] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.
[0015] At least one embodiment of the present invention discloses a method for generating electronic signatures for approval workflows, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Obtain the content data and historical version snapshot data of the document to be signed and approved.
[0016] Retrieve the complete content data of the document currently awaiting signature and approval, and separate the set of structured field values and the set of unstructured text paragraphs from it; at the same time, retrieve the historical version snapshot data of the document saved when it was submitted at each preceding node in the approval flow.
[0017] The structured field value set includes values for quantifiable business fields such as amount, date, and quantity; the unstructured text paragraph set includes natural language text paragraphs such as contract terms, performance conditions, and restrictive statements. Historical version snapshot data records the complete content status of the document at each preceding approval node.
[0018] Step 2: Perform clause semantic unit segmentation and semantic vector encoding on the unstructured text paragraph set.
[0019] According to the legal clause segmentation rules, the unstructured text paragraph set is processed by semantic unit segmentation to generate an independent clause semantic unit sequence; for each clause semantic unit, a pre-trained text semantic encoding model is called to calculate and output the corresponding semantic vector representation.
[0020] It should be noted that legal clause segmentation rules refer to a set of rules for dividing the main text based on clause numbering, paragraph boundary marking, and semantic integrity constraints, so that each segmented unit corresponds to a clause expression that is independent in legal semantics. The pre-trained text semantic encoding model is a conventional pre-trained language model (such as a BERT-like model). The input of the pre-trained text semantic encoding model is a text sequence of a single clause semantic unit, and the output is a fixed-dimensional... The semantic vector representation is used to capture the semantic features of the clause text. After encoding the input text sequence, the pre-trained text semantic encoding model takes the hidden state corresponding to the special marker at the beginning of the sequence and projects it through a linear mapping layer to... The dimensional space serves as the semantic vector output of the semantic unit of this clause.
[0021] Step 3: Perform clause-by-clause cosine similarity calculation on the semantic vector sequences of the current version and the previous version to generate a list of semantic change clause tags.
[0022] Obtain the text content of the most recent preceding node version from the historical version snapshot data, perform the same clause semantic unit segmentation and semantic vector encoding processing as in step 2, and generate a preceding version clause semantic vector sequence.
[0023] Perform clause-by-clause cosine similarity calculations on the semantic vector sequence of the current version's clauses and the semantic vector sequences of the previous versions' clauses. Let the current version be the... The semantic vector of each clause is The semantic vector of the corresponding clause in the previous version is , and All For a 3D vector, the cosine similarity is calculated as follows: ; in, Let be the inner product of two vectors. For vectors Norm, For all terms and conditions Compared with the preset semantic change threshold Performing a comparison will satisfy The clauses are marked as semantic change clauses, and a list of semantic change clause tags is generated.
[0024] It should be noted that the semantic vectors of the terms in both the current and previous versions are generated by the same pre-trained text semantic encoding model, and the output dimensions are both [missing information]. Therefore, the two vectors involved in the cosine similarity calculation have the same dimension. For cases where there are clause additions or deletions between the current version and previous versions, resulting in inconsistent sequence lengths, the newly added clauses are marked as change clauses without corresponding previous clauses and directly included in the semantic change clause marker list.
[0025] In this embodiment, to improve the accuracy of clause alignment, clause alignment processing can be performed on the semantic vector sequences of clauses in the current version and the previous version before performing clause-by-clause cosine similarity calculation. Specifically, based on the global similarity matrix between the semantic vector sequences of clauses, a dynamic programming longest common subsequence matching method is used to pair the clauses in the current version with the clauses in the previous version that are semantically closest, and then cosine similarity calculation is performed on the pairing results, thereby avoiding incorrect pairing caused by clause order adjustment.
[0026] Step 4: Extract key semantic elements and determine the type of change for the semantic change clause, and calculate the quantitative value of the level of legal impact.
[0027] For each amended clause in the semantically amended clause marker list, the named entity recognition and semantic role labeling capabilities of the pre-trained text semantic encoding model are invoked to extract the key semantic elements of the clause, including the obligor, the content of the right, the restrictive conditions, and the numerical constraints. The obligor, the content of the right, the restrictive conditions, and the numerical constraints each correspond to different semantic role categories. A structured comparative analysis is then performed on the key semantic elements of the corresponding clause in the previous version to determine the type of amendment. Amendment types include: expanding the scope of rights, reducing the scope of obligations, deleting restrictive conditions, and modifying numerical constraints.
[0028] Based on the type and extent of the change, calculate the quantitative value of the legal impact level of each changed clause. Let the basic impact weight corresponding to the change type be... The normalized score for the change range is ,but: ; in, The legal impact of the type of change is pre-set, with higher weights for expanding the scope of rights and deleting restrictions, medium weights for modifying numerical constraints, and corresponding weights for reducing the scope of obligations. The calculation is based on the proportion of differences between the current amended terms and their corresponding terms in the previous version across key semantic elements. Specifically, for each of the four categories of elements—obligated parties, rights, restrictions, and numerical constraints—the proportion of elements that have changed between the current and previous versions is calculated out of the total number of elements. The weighted average of these four proportions is then taken as the statistical average. The weights are pre-set according to the degree of influence of each element category on the legal meaning.
[0029] Step 5: Establish a normal change baseline by combining historical approval and modification record data, and generate a comprehensive score for semantic tampering risk.
[0030] Obtain historical approval and modification records for similar business documents, and statistically analyze the normal distribution range of allowed modification types and modification magnitudes across approval nodes. Use this as a baseline for normal changes. Quantify the change type and legal impact level of each modified clause in the current document. The calculation is performed by comparing the changes with the normal distribution range, counting the changes that exceed the normal distribution range, and accumulating their legal impact level quantification values to generate a comprehensive score for semantic tampering risk. : ; in, A collection of change clauses that exceed the normal distribution range. For the change of terms in the collection Indices in the set, summation and traversal All elements in For the first The legal impact level quantification value of a change clause that exceeds the normal distribution range.
[0031] It should be noted that the normal distribution range refers to the distribution of the types of clause changes that occurred in similar business documents between the same approval nodes in historical approval modification records. The statistical range of values is expressed as the mean plus or minus a certain number of standard deviations. When the type of change in the terms of the agreement is changed... If the scope is exceeded, the clause is deemed to have exceeded the normal change baseline.
[0032] In this embodiment of the application, in order to further distinguish the normal change differences between different approval nodes, the normal distribution range can be counted separately according to the approval node pairs, that is, for the document from the node Flow to node In such cases, maintain the corresponding normal distribution range parameters separately to make the comparison calculation more in line with the change mode of the specific approval process.
[0033] Step 6: Based on the comprehensive score of semantic tampering risk, make a decision to generate or postpone the signature, and output the signature data or audit prompt data.
[0034] A comprehensive score for semantic tampering risk With preset signature constraint threshold Execution comparison: when At that time, the signature image generation process is executed, the electronic signature data is output, and the electronic signature is attached to the currently approved document; when When the signature generation process is suspended, a signature delay response is output, and review prompt data is generated. The review prompt data includes the location markers of each changed clause in the document, the corresponding change type, and the quantitative value of the legal impact level. The review prompt data is sent to the superior approval node to trigger the manual review process.
[0035] In this embodiment of the application, to support graded responses to different risk levels, a multi-level threshold determination can be introduced in step 6. Specifically, a low-risk threshold is preset. With high risk threshold ( ),when Electronic signatures are generated directly at the time; when When a risk warning is generated, an electronic signature is created and the review warning data is synchronized to the current approval node for storage; when At this time, the generation of signatures is suspended and the manual review process is triggered, thereby achieving a differentiated balance between security control and approval efficiency.
[0036] On the other hand, the present invention also proposes an electronic signature generation system for approval workflows, comprising: The data acquisition module is used to acquire the complete content data of the document to be signed and approved, separate the set of structured field values from the set of unstructured text paragraphs, and acquire historical version snapshot data; The semantic encoding module is used to perform clause semantic unit segmentation on the unstructured text paragraph set and call the pre-trained text semantic encoding model to calculate the semantic vector representation of each clause semantic unit; The similarity comparison module is used to perform the same segmentation and encoding processing on the text content of the most recent preceding node version in the historical version snapshot data, and to perform clause-by-clause cosine similarity calculation on the clause semantic vector sequence of the current version and the preceding version to generate a list of semantic change clause tags; The impact assessment module is used to extract key semantic elements from each change clause in the semantic change clause tag list, perform structured comparative analysis to determine the change type, and calculate the legal impact level quantification value. The risk scoring module is used to obtain historical approval and modification records of similar business documents, establish a normal change baseline, compare and calculate the quantitative values of the change type and legal impact level of each change clause with the normal distribution range, and generate a comprehensive score for semantic tampering risk. The signature decision module compares the comprehensive score of semantic tampering risk with the preset signature constraint threshold. When the comprehensive score of semantic tampering risk is lower than the preset signature constraint threshold, it outputs electronic signature data. When the comprehensive score of semantic tampering risk reaches or exceeds the preset signature constraint threshold, it stops signature generation and outputs audit prompt data containing the location mark of the changed clause, the change type, and the quantitative value of the legal impact level, triggering the manual review process.
[0037] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.
[0038] Existing signature systems only cover the structured field layer for anomaly detection, failing to detect semantic changes in unstructured text. Attackers can modify text clauses while keeping structured field values unchanged, allowing semantically altered documents to pass detection and obtain electronic signatures. This method introduces clause-level semantic vector comparison before signature generation. It calculates the cosine similarity of the semantic vectors of the current version and previous versions, making semantic consistency of the text a prerequisite for signature generation. This allows semantic-level tampering, invisible to structured field statistical deviations, to be captured during the signature triggering stage. The method assesses the substantive impact of changes using a legal impact level quantification value and establishes a baseline for normal changes based on historical approval behavior distribution. This enables the judgment process to distinguish between normal approval modifications and anomalous tampering, avoiding false interception of reasonable clause revisions. Finally, electronic signatures are only attached to approval documents with verified semantic integrity, effectively intercepting semantic tampering attacks during signature generation.
[0039] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.
Claims
1. A method for generating electronic signatures in an approval workflow, characterized in that, Includes the following steps: Obtain the complete content data of the document to be signed and approved, separate the set of structured field values from the set of unstructured text paragraphs, and obtain historical version snapshot data; The unstructured text paragraph set is segmented into clause semantic units according to the legal clause segmentation rules, and a pre-trained text semantic encoding model is called to calculate the semantic vector representation of each clause semantic unit; Perform the same segmentation and encoding process on the text content of the most recent preceding node version in the historical version snapshot data, perform clause-by-clause cosine similarity calculation on the clause semantic vector sequence of the current version and the preceding version, mark clauses with cosine similarity lower than the preset semantic change threshold as semantic change clauses, and generate a list of semantic change clause markers. For each changed clause in the semantic change clause tagging list, four key semantic elements are extracted: the obligor, the content of the rights, the restrictions, and the numerical constraints. The clauses are then compared and analyzed in a structured manner with the corresponding clauses in the previous version to determine the change type. The legal impact level is then calculated based on the basic impact weight and the normalized score of the change magnitude corresponding to the change type. Establish a normal change baseline by obtaining historical approval and modification records of similar business documents, compare the change type and legal impact level of each change clause with the normal change baseline, and accumulate the legal impact level of change clauses that exceed the normal change baseline to generate a comprehensive score for semantic tampering risk. The comprehensive score of semantic tampering risk is compared with the preset signature constraint threshold. When the comprehensive score of semantic tampering risk is lower than the preset signature constraint threshold, electronic signature data is output. When the comprehensive score of semantic tampering risk reaches or exceeds the preset signature constraint threshold, signature generation is stopped, and audit prompt data containing the location mark of the changed clause, the change type, and the quantitative value of the legal impact level is output, triggering the manual review process.
2. The electronic signature generation method for approval workflow according to claim 1, wherein, The clause-by-clause cosine similarity calculation includes: Let the current version be number The semantic vector of each clause is The semantic vector of the corresponding clause in the previous version is The cosine similarity is calculated as follows: ; in, Let be the dot product of two vectors. For vectors Norm, ; will satisfy The terms are marked as semantic change terms, in which A preset semantic change threshold is set; in cases where there are clause additions or deletions between the current version and the previous version, resulting in inconsistent sequence lengths, the newly added clauses will be directly included in the semantic change clause marker list.
3. The electronic signature generation method for approval workflow according to claim 2, wherein, Before performing the clause-by-clause cosine similarity calculation, the following is also included: Based on the global similarity matrix between the semantic vector sequences of the current version and the previous version, the longest common subsequence matching method of dynamic programming is used to pair the semantically closest clauses in the current version and the previous version, and then the cosine similarity is calculated on the pairing results.
4. The electronic signature generation method for approval workflow according to claim 1, wherein, The calculation of the legal impact level quantification value includes: Let the basic impact weight corresponding to the change type be . The normalized score for the change range is Then the first The legal impact level of each amended clause is quantified as follows: ; in, It is calculated by the weighted average of the proportion of the number of elements that change among the four categories of elements: subject of obligation, content of rights, restrictive conditions, and numerical constraints, relative to the total number of elements. The weight of each element category is pre-set according to its degree of influence on the legal meaning.
5. The electronic signature generation method for approval workflow according to claim 4, wherein, The calculation of the semantic tampering risk comprehensive score includes: ; in, This is a collection of change clauses that exceed the normal change baseline. For set Middle The legal impact level quantification value of each changed clause; the normal change baseline is based on the distribution of the types of historical clause changes in similar business documents at the same approval node, and... The statistical range of values indicates that when the type of change of the change clause or If the scope exceeds this statistical range, the clause is deemed to exceed the normal change baseline.
6. The electronic signature generation method for approval workflow according to claim 5, wherein, The normal change baseline is statistically analyzed separately according to the approval node, targeting the document from the node. Flow to node In such cases, maintain the corresponding normal distribution range parameters separately.
7. The electronic signature generation method for approval workflow according to claim 1, wherein, The steps for comparing the comprehensive score of semantic tampering risk with the preset signature constraint threshold include: Preset low-risk threshold With high risk threshold ,in When the overall score for semantic tampering risk is lower than When the semantic tampering risk score is not lower than a certain threshold, electronic signature data is output. And lower than When the risk level is not lower than a certain threshold, an electronic signature with a risk warning is generated, and the review warning data is synchronized to the current approval node for retention; when the comprehensive score of semantic tampering risk is not lower than a certain threshold... If this occurs, the signature generation process is suspended and a manual review process is triggered.
8. An electronic signature generation system for approval workflows, characterized in that, include: The data acquisition module is used to acquire the complete content data of the document to be signed and approved, separate the set of structured field values from the set of unstructured text paragraphs, and acquire historical version snapshot data; The semantic encoding module is used to segment the unstructured text paragraphs according to the legal clause segmentation rules, and to call the pre-trained text semantic encoding model to calculate the semantic vector representation of each clause semantic unit; The similarity comparison module is used to perform the same segmentation and encoding processing on the text content of the most recent preceding node version in the historical version snapshot data, and to perform clause-by-clause cosine similarity calculation on the clause semantic vector sequence of the current version and the preceding version to generate a list of semantic change clause tags; The impact assessment module is used to extract four key semantic elements—obligation subject, rights content, restrictions, and numerical constraints—from each modified clause in the semantic modification clause tag list, perform structured comparative analysis to determine the type of modification, and calculate the quantitative value of the legal impact level. The risk scoring module is used to obtain historical approval and modification records of similar business documents to establish a normal change baseline. It compares the change type and legal impact level of each change clause with the normal change baseline to generate a comprehensive score for semantic tampering risk. The signature decision module compares the comprehensive score of semantic tampering risk with the preset signature constraint threshold. When the comprehensive score of semantic tampering risk is lower than the preset signature constraint threshold, it outputs electronic signature data. When the comprehensive score of semantic tampering risk reaches or exceeds the preset signature constraint threshold, it stops signature generation and outputs audit prompt data containing the location mark of the changed clause, the change type, and the quantitative value of the legal impact level, triggering the manual review process.
