Multi-modal agent driven enterprise oa review and approval decision method and system

Abstract

The application relates to the technical field of data processing, and discloses a multi-modal Agent-driven enterprise OA examination and approval decision method and system. The method comprises the following steps: inputting each clause of a contract into an A-party interest Agent and a B-party interest Agent to generate opposite interpretation vectors and construct a clause semantic game graph, calculating the cosine distance of each clause node to mark the contradictory nodes, calculating the downstream influence radius along the directed edges to divide the isolated contradictory nodes and the cascading contradictory nodes, and reconfiguring the preset linear approval path into a bifurcated arbitration approval path. The application solves the problems that the existing contract examination scheme cannot detect cross-clause semantic contradictions, cannot quantitatively depict the contradiction propagation range, and the preset linear approval path cannot be dynamically reconfigured according to the contradiction structure.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a multimodal agent-driven enterprise OA review and approval decision-making method and system. Background Technology

[0002] With the deepening of enterprise digital transformation, contract review and approval decisions in office automation systems have become core aspects of enterprise operation and management. Currently, enterprise contract review mainly relies on two methods: manual clause-by-clause reading and rule matching. The former depends on the professional experience of the reviewers, while the latter relies on comparison with a pre-set keyword library and templates. In recent years, single-agent contract review solutions based on large language models have gradually emerged. These solutions input the contract text into a single language model instance, generate review conclusions clause by clause, and combine this with retrieval-enhanced generation technology to match external knowledge bases, pushing the review conclusions to various approval nodes in a linear approval process.

[0003] However, the aforementioned existing technical solutions have fundamental technical flaws in enterprise OA contract review scenarios. Single-Agent linear review schemes process clauses sequentially, limiting their reasoning scope to the semantics of individual clauses. They inherently lack the ability to perceive cross-clause logical relationships and cannot identify the objectively existing semantic contradictions between different clauses in the contract document. While rule-based schemes based on keyword matching and template comparison can detect anomalies in specific fields, they also fail to capture cross-clause semantic contradictions formed by logical dependencies such as obligation triggering, breach triggering, and exemption triggering. Existing confidence-based routing schemes can identify the uncertainty of a single conclusion and trigger manual review, but the triggering basis for manual intervention is the probability distribution within the model, rather than the objectively existing structural contradictions in the contract text. When reviewers intervene, they are still facing AI conclusions rather than original contradictory evidence, posing a risk of anchoring bias. Summary of the Invention

[0004] This application provides a multimodal agent-driven enterprise OA review and approval decision-making method and system, which solves the problems of existing contract review schemes that cannot detect semantic contradictions across clauses, cannot quantitatively characterize the scope of contradiction propagation, and cannot dynamically reconstruct the signing and approval hierarchy based on the contradiction structure of the preset linear approval path. It has achieved significant improvements in both the accuracy of contract review and the rationality of the approval path.

[0005] Firstly, this application provides a multimodal agent-driven enterprise OA review and approval decision-making method, which includes: Step S1: Input the text content of each clause in the contract document into Party A's interest agent and Party B's interest agent respectively. Party A's interest agent and Party B's interest agent independently generate opposing interpretation vectors for each clause. Based on the logical triggering relationship between each clause, construct the opposing interpretation vectors into a clause semantic game graph. Step S2: Calculate the cosine distance between the opposing interpretation vectors of A and B for each node in the semantic game graph of the clauses, and mark the nodes whose cosine distance exceeds a preset threshold as contradictory nodes to obtain the set of contradictory nodes in the semantic game graph of the clauses. Step S3: Starting from each conflict node in the set of conflict nodes, propagate downstream along the directed edges of the semantic game graph of the clause, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. Step S4: Based on the type of the isolated conflict node and the cascaded conflict node and the corresponding downstream influence radius, the preset linear approval path is reconstructed into a bifurcated arbitration approval path with the conflict node as the axis. The conflict solution and approval authority level corresponding to each bifurcation are determined by the bifurcated arbitration approval path.

[0006] Secondly, this application provides a multimodal agent-driven enterprise OA review and approval decision-making system, the multimodal agent-driven enterprise OA review and approval decision-making system comprising: The input module is used to input the text content of each clause in the contract document into the A party's interest agent and the B party's interest agent respectively. The A party's interest agent and the B party's interest agent independently generate opposing interpretation vectors for each clause, and construct the opposing interpretation vectors into a clause semantic game graph based on the logical triggering relationship between each clause. The marking module is used to calculate the cosine distance between the opposing interpretation vectors of A and B for each clause node in the clause semantic game graph, and to mark the nodes whose cosine distance exceeds a preset threshold as contradictory nodes, thereby obtaining the set of contradictory nodes in the clause semantic game graph. The calculation module is used to start from each conflict node in the set of conflict nodes, propagate along the directed edges of the semantic game graph of the clause to the downstream clauses, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. The reconstruction module is used to reconstruct a preset linear approval path into a bifurcated arbitration approval path with the conflict node as the axis, based on the type of the isolated conflict node and the cascaded conflict node and the corresponding downstream influence radius. The bifurcated arbitration approval path determines the conflict solution and approval authority level corresponding to each bifurcation.

[0007] Thirdly, a multimodal agent-driven enterprise OA review and approval decision-making device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the multimodal agent-driven enterprise OA review and approval decision-making device to execute the aforementioned multimodal agent-driven enterprise OA review and approval decision-making method.

[0008] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned multimodal agent-driven enterprise OA review and approval decision-making method.

[0009] The technical solution provided in this application involves inputting the text content of each clause in the contract document into two agents—the Party A's interest agent and the Party B's interest agent—each with an opposing role. These two agents independently generate opposing interpretation vectors for each clause, and based on the logical triggering relationships between the clauses, construct a clause semantic game graph from these opposing interpretation vectors. This fundamentally solves the technical shortcomings of existing single-agent linear review schemes, which process clauses sequentially, limit their reasoning scope to the semantic range of a single clause, and fail to perceive cross-clause logical relationships. The clause semantic game graph uses opposing interpretation vectors as node attributes and obligation triggering relationships, breach of contract triggering relationships, and exemption triggering relationships as directed edge types. It comprehensively depicts the semantic opposition distribution and logical triggering structure among the clauses in the contract document, providing an objective structural basis for identifying cross-clause semantic contradictions, rather than relying on probability distribution estimations within the model. Based on this, by calculating the cosine distance between the opposing interpretation vectors of A and B at each node in the semantic game diagram of the contract terms, and marking nodes whose cosine distance exceeds a preset threshold as contradictory nodes, a set of contradictory nodes is obtained, realizing the quantitative identification of the semantic opposition structure that objectively exists in the contract document. The contribution of the cosine distance algorithm feature in this scheme is to transform the semantic differences of dual-Agent adversarial reasoning into a calculable and comparable numerical quantity, providing a clear starting point for subsequent contradiction propagation analysis.

[0010] Furthermore, this application takes each conflicting node in the conflicting node set as a starting point and propagates it layer by layer downstream along the directed edges of the semantic game graph of the clauses. It calculates the downstream influence radius of each conflicting node and divides the conflicting node set into isolated conflicting nodes and cascading conflicting nodes based on the downstream influence radius. This solves the technical deficiency of existing solutions, which can only output isolated risk markers and cannot quantitatively characterize the scope of conflict impact. The core contribution of the downstream influence radius algorithm output in this solution lies in establishing a quantitative mapping relationship between the severity of the conflict structure and the approval decision-making authority. Based on this, this application reconstructs the preset linear approval path into a bifurcated arbitration approval path centered on the conflicting nodes, according to the types of isolated and cascading conflicting nodes and their corresponding downstream influence radii. The bifurcated arbitration approval path determines the conflict resolution solutions and approval authority levels corresponding to each bifurcation. This fundamentally solves the technical problem that existing preset linear approval paths cannot dynamically adjust the approval level and handling solutions according to the conflict structure, and there is a systemic mismatch between approval authority and conflict severity. It realizes that the approval path topology is generated in real-time driven by the conflict distribution of the contract content, rather than relying on a fixed organizational structure diagram. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an embodiment of the multimodal agent-driven enterprise OA review and approval decision-making method in this application. Figure 2 This is a schematic diagram of the cosine distance distribution of the dual-Agent interpretation vectors in an embodiment of this application; Figure 3 This is a schematic diagram comparing the accuracy of cross-clause contradiction detection in the embodiments of this application. Detailed Implementation

[0013] This application provides a multimodal agent-driven enterprise OA review and approval decision-making method and system. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the multimodal agent-driven enterprise OA review and approval decision-making method in this application includes: Step S1: Input the text content of each clause in the contract document into Party A's interest agent and Party B's interest agent respectively. Party A's interest agent and Party B's interest agent independently generate opposing interpretation vectors for each clause, and construct a semantic game diagram of the clauses based on the logical triggering relationship between each clause. Specifically, the "Agent of Party A's Interests" and the "Agent of Party B's Interests" are two large language model instances with opposing roles. They are constrained by system prompts that maximize Party A's rights and Party B's rights, respectively. Semantic vectors are output independently for the same clause text, and the concatenation of these two semantic vectors constitutes the opposing interpretation vector of that clause. The logical triggering relationships between clauses are obtained through semantic parsing of conditional sentences, result sentences, and cross-references within the clause text. These relationships are categorized into three types: obligation triggering, breach of contract triggering, and exemption triggering. Directed edges connect the corresponding clause nodes, with the direction of the directed edges pointing from the triggering clause to the triggered clause. All clause nodes and directed edges together constitute the clause semantic game graph.

[0015] Step S2: Calculate the cosine distance between the opposing interpretation vectors of A and B for each clause node in the clause semantic game graph, and mark the nodes whose cosine distance exceeds the preset threshold as contradictory nodes to obtain the set of contradictory nodes in the clause semantic game graph. Specifically, the cosine distance is calculated for the interpretation vectors of Party A and Party B at the same clause node. The larger the cosine distance value, the greater the difference in semantic interpretation between the two parties for that clause. The preset threshold is determined by adding one standard deviation to the mean of the cosine distance distribution of the interpretation vectors of undisputed clauses in the contract text. Clause nodes exceeding this threshold are marked as contradictory nodes, and all contradictory nodes constitute a set of contradictory nodes.

[0016] Step S3: Starting from each conflict node in the set of conflict nodes, propagate downstream along the directed edges of the semantic game graph of the clauses, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. Specifically, the downstream influence radius is defined as the ratio of the total number of downstream clause nodes reachable from a given conflicting node along the directed edges of the semantic game graph to the total number of all clause nodes in the graph, with a value ranging from 0 to 1. The propagation process employs a layer-by-layer breadth-first traversal, expanding layer by layer along the directed edges from each conflicting node until no new nodes are reachable. Conflicting nodes with a downstream influence radius below a preset threshold are marked as isolated conflicting nodes, while those above the preset threshold are marked as cascaded conflicting nodes. The preset threshold is set based on empirical values ​​of the proportion of core rights and obligations clauses in the full text of the contract.

[0017] Step S4: Based on the type of isolated conflict nodes and cascading conflict nodes and their corresponding downstream influence radius, the preset linear approval path is reconstructed into a bifurcated arbitration approval path with the conflict node as the axis. The conflict solution and approval authority level corresponding to each bifurcation are determined by the bifurcated arbitration approval path.

[0018] Specifically, the bifurcated arbitration approval path is a non-linear approval structure obtained by bifurcating three types of conflict solutions—intercepting approval nodes for directly interested parties for isolated conflict nodes, generating modification clauses for cascading conflict nodes, and deleting clauses and supplementary agreements—on top of a pre-defined linear approval path. The signing authority level corresponding to each bifurcated node is determined jointly by the type of cascading conflict node and its downstream influence radius. The larger the downstream influence radius, the higher the signing authority level, thus forming a bifurcated arbitration approval path with conflict nodes as the axis and the severity of the conflict matching the approval authority.

[0019] In one specific embodiment, step S1 includes: The text content of each clause in the contract document is input into the A-party interest agent and the B-party interest agent respectively. The A-party interest agent is set to maximize the rights of the A-party and generates an A-party interpretation vector for each clause. The B-party interest agent is set to maximize the rights of the B-party and generates a B-party interpretation vector for each clause. The A-party interpretation vector and the B-party interpretation vector corresponding to each clause are combined into the opposing interpretation vector of each clause. Based on the textual semantics of each clause, the logical triggering relationships between each clause are extracted, and the logical triggering relationships are divided into obligation triggering relationships, breach of contract triggering relationships and exemption triggering relationships. The logical triggering relationships of each clause are then labeled as the edge type of directed edges. By using the opposing interpretation vectors of each clause as node attributes and the directed edges between clauses and their types as graph structures, a semantic game graph of clauses is constructed.

[0020] Specifically, both the A-party's interest agent and the B-party's interest agent use the original text of the contract terms as input. The input format is a single-term text string after sentence segmentation. The output is a fixed-dimensional semantic vector, with the vector dimension consistent with the hidden layer dimension of the selected pre-trained language model, typically 768 or 1024 dimensions, depending on the deployed model. The system prompt word constraint model for the A-party's interest agent assigns higher attention weights to words expressing the A-party's rights and restricting the B-party's obligations during encoding, while the system prompt word constraint model for the B-party's interest agent assigns higher attention weights to words expressing the B-party's rights and restricting the A-party's obligations. Both agents are based on the same underlying model, differing only in system prompt words, thus ensuring that the opposition stems solely from role constraints. The A-party's interpretation vector and the B-party's interpretation vector for each term are concatenated column-wise to obtain an opposing interpretation vector with twice the dimension of the original vector. The concatenation method, rather than weighted fusion, is chosen because it fully preserves the semantic distribution of each party without introducing information loss due to fusion weights.

[0021] The criteria for determining the triggering relationship of obligations are: 1) Conditional clauses in the clause text stating that one party must perform a certain action to trigger the execution of another clause. Typical linguistic features include conditional clauses containing keywords such as "must," "shall," and "may," and explicit numbering references between clauses. 2) Result clauses in the clause text triggering the breach of contract. 3) Result clauses in the clause text triggering compensation, punishment, or termination rights if one party fails to perform its obligations. Typical linguistic features include keywords such as "breach of contract," "compensation," and "termination," and refer to a specific clause number. 4) Result clauses in the clause text triggering exemption from liability. 5) Valid condition clauses in the clause text exempting one or both parties from liability under agreed-upon circumstances or force majeure. Typical linguistic features include keywords such as "force majeure," "exempt from liability," and "deemed performance," and reference to the exemption clause number. These three criteria are applied through keyword matching and semantic template comparison of the clause text. Each clause's relationship is categorized into a unique edge type based on the matching results; no relationship can belong to both types simultaneously, thus ensuring unambiguous edge type labeling. Each clause node is stored as its opposite interpretation vector as a node attribute, and the relationship between clauses is stored as the corresponding edge type as an edge attribute. The nodes and directed edges together constitute the clause semantic game graph. The direction of the directed edges in the graph is uniformly from the triggering clause node to the triggered clause node, and the direction definition is unambiguous.

[0022] In one specific embodiment, step S2 includes: For each clause node in the semantic game graph, the interpretation vectors of Party A and Party B are extracted respectively. Based on the interpretation vectors of Party A and Party B, the cosine distance of each clause node is calculated to obtain the cosine distance of the interpretation vectors of each clause node. The cosine distance of the interpretation vector of each clause node is compared with a preset threshold. Clause nodes whose cosine distance of interpretation vector exceeds the preset threshold are marked as contradictory nodes, thus obtaining a set of contradictory node labels. Based on the edge type of the directed edges corresponding to each conflicting node in the conflicting node label set, each conflicting node is divided into obligation conflicting nodes, breach of contract conflicting nodes, and exemption conflicting nodes, resulting in a set of conflicting nodes classified by edge type.

[0023] Specifically, the cosine distance is calculated by splitting the opposing interpretation vectors stored for the same clause node into the interpretation vectors of Party A and Party B. The splitting method involves taking the first half of the opposing interpretation vectors and the second half, taking Party A's interpretation vector and Party B's interpretation vector. The cosine distance is calculated as follows: the inner product of Party A's and Party B's interpretation vectors is divided by the product of their respective L2 norms to obtain the cosine similarity. Then, 1 is subtracted from the cosine similarity to obtain the cosine distance. The cosine distance ranges from 0 to 2; a larger value indicates a greater difference in semantic interpretation between Party A and Party B regarding the clause. The preset threshold is set to 0.6. This value is determined based on the statistical distribution of the cosine distances of the interpretation vectors of each clause node in an undisputed standard contract text. The mean of the distribution plus twice the standard deviation is taken as the upper bound of the threshold. Clause nodes with a value below 0.6 have similar semantic interpretations by both parties and do not constitute a substantial contradiction. Clause nodes with a value above 0.6 have significant semantic contradictions between the two parties and are marked as contradictory nodes. All contradictory nodes constitute a contradictory node label set.

[0024] The rules for classifying conflicting nodes by edge type are as follows: Extract the edge type of all directed edges corresponding to each conflicting node in the semantic game graph of the clauses from the conflicting node label set. If all directed edges of a conflicting node are of the obligation-triggered type, then the conflicting node is labeled as an obligation-conflict conflicting node; if all directed edges are of the breach-of-contract-triggered type, then it is labeled as a breach-of-contract-conflict conflicting node; if all directed edges are of the exemption-of-liability-triggered type, then it is labeled as an exemption-of-liability-conflict conflicting node. When a conflicting node corresponds to directed edges of multiple edge types, its category is determined based on the edge type with the highest proportion among the directed edges corresponding to the conflicting node. If the number of two edge types is the same, the category is determined by the priority order of obligation-triggered relationship over breach-of-contract-triggered relationship, and breach-of-contract-triggered relationship over exemption-of-liability-triggered relationship. This ensures that each conflicting node uniquely belongs to one category, and all classification results constitute a set of conflicting nodes classified by edge type.

[0025] Figure 2 This is a schematic diagram of the cosine distance distribution of the dual-Agent interpretation vectors in this embodiment. In this embodiment, the horizontal axis represents the cosine distance between the interpretation vectors of Party A and Party B, the vertical axis represents the number of clauses within the corresponding distance interval, the diagonally filled bars represent the cosine distance distribution of normal clauses, the cross-filled bars represent the cosine distance distribution of contradictory clauses, and the vertical dashed line represents the preset threshold of 0.6. The cosine distances of normal clauses are concentrated in the interval between 0.1 and 0.4, while the cosine distances of contradictory clauses are concentrated in the interval between 0.6 and 0.9. The distribution of the two types of clauses shows a clear separation on both sides of the preset threshold, verifying the rationality of using the cosine distance exceeding the preset threshold as the condition for marking contradictory nodes.

[0026] In one specific embodiment, step S3, starting from each conflicting node in the conflicting node set, propagates downstream along the directed edges of the semantic game graph of the clauses, including: Taking each conflict node in the set of conflict nodes as the starting point of propagation, the first layer of propagation is carried out along the directed edges corresponding to each conflict node in the semantic game graph of the clauses, to the downstream clause nodes directly connected to each conflict node, so as to obtain the first layer of downstream clause node set of each conflict node. Based on the directed edges corresponding to each downstream clause node in each first-level downstream clause node set, the process continues to propagate to the secondary downstream clause nodes directly connected to each downstream clause node, iterating layer by layer until there are no new downstream clause nodes reachable in the clause semantic game graph, thus obtaining the full set of downstream clause nodes for each conflict node. The ratio of the number of nodes in the full set of downstream clause nodes of each conflict node to the total number of all clause nodes in the clause semantic game graph is determined as the downstream influence radius of each conflict node, thus obtaining the downstream influence radius corresponding to each conflict node.

[0027] Specifically, the first-level propagation is executed as follows: starting from each conflicting node, propagation proceeds along all directed edges originating from that node in the semantic game graph of the clauses, reaching the clause node pointed to by the endpoint of each directed edge. All endpoint clause nodes are then added to the first-level downstream clause node set of that conflicting node. Propagation occurs only along the direction of the directed edges, without backtracking in the opposite direction, ensuring that the propagation direction is consistent with the logical triggering direction. The termination condition for each layer of iteration is: when all endpoint nodes pointed to by directed edges originating from the current layer's downstream clause node set in the semantic game graph already exist in the visited node record of that conflicting node, i.e., when no new unvisited clause nodes are reachable, the iteration terminates. The full downstream clause node set of each conflicting node is the union of the downstream clause node sets of each layer, and deduplication is performed on the union result. Each clause node is recorded only once in the full downstream clause node set, and is not counted repeatedly due to being reached repeatedly by multiple propagation paths.

[0028] The denominator for calculating the downstream influence radius is the total number of all clause nodes in the clause semantic game graph. The statistical scope includes both conflicting nodes and non-conflicting nodes, excluding any nodes, to ensure consistency and comparability of the denominator across all conflicting nodes. The numerator is the number of nodes in the complete set of downstream clause nodes for each conflicting node, excluding the conflicting node itself. The downstream influence radius ranges from 0 to 1. A value of 0 indicates that the conflicting node has no downstream reachable nodes in the clause semantic game graph, while a value close to 1 indicates that the conflicting influence of the node almost covers all clauses. The downstream influence radius serves as a quantitative basis for subsequently distinguishing between isolated and cascading conflicting nodes.

[0029] In one specific embodiment, step S3, calculating the downstream influence radius of each conflict node, includes: For the full set of downstream clause nodes of each conflict node, extract the edge type of the directed edge corresponding to each downstream clause node, and classify and count each downstream clause node in the full set of downstream clause nodes according to the obligation triggering relationship, the breach triggering relationship and the exemption triggering relationship to obtain the number of downstream clause nodes of each type of conflict node. Based on the number of downstream clause nodes of each conflict node, the ratio of the number of downstream clause nodes corresponding to each edge type to the total number of clause nodes in the clause semantic game graph is calculated to obtain the downstream influence radius of each conflict node. The downstream influence radii of each conflict node are weighted and summed according to the weights of the obligation triggering relationship, the breach of contract triggering relationship, and the exemption triggering relationship, to obtain the weighted downstream influence radius of each conflict node. The weighted downstream influence radius is used as the downstream influence radius corresponding to each conflict node.

[0030] Specifically, the method for extracting the edge type of the directed edges corresponding to each downstream clause node is as follows: extract the edge type of all directed edges reaching the downstream clause node. When a downstream clause node is reached by only one type of directed edge, directly use that edge type as the attribute edge type of the downstream clause node. When a downstream clause node is reached by multiple types of directed edges, take the edge type with the highest proportion among the directed edges reaching the node as the attribute edge type. If the proportions are the same, determine the attribute according to the priority of obligation triggering relationship over breach of contract triggering relationship, and breach of contract triggering relationship over exemption triggering relationship, to ensure that each downstream clause node uniquely belongs to one type of edge. The statistical caliber for the number of downstream clause nodes by type is as follows: regardless of how many directed edges reach the same downstream clause node, it is counted only once in the count of its attribute edge type, and is not counted repeatedly. The sum of the number of downstream clause nodes of the three types is equal to the total number of duplicate nodes in the entire set of downstream clause nodes.

[0031] The downstream influence radius for each category is obtained by dividing the number of downstream clause nodes corresponding to the obligation triggering relationship, breach of contract triggering relationship, and exemption triggering relationship by the total number of clause nodes in the semantic game graph of the clauses. The value range of the downstream influence radius for each of the three categories is 0 to 1. The weighted downstream influence radius is obtained by weighted summation of the downstream influence radii of the three categories. The weight of the obligation triggering relationship is set to 0.5, the weight of the breach of contract triggering relationship is set to 0.3, and the weight of the exemption triggering relationship is set to 0.2, with the sum of the three weights being 1. The weight values ​​are determined based on the degree of contribution of the three triggering relationships to the contract performance risk. The obligation triggering relationship is directly related to the core performance obligation of the contract and has the largest contribution to the overall risk. The breach of contract triggering relationship is related to compensation liability and has a medium contribution. The exemption triggering relationship is related to liability exemption and has the smallest contribution. The value range of the weighted downstream influence radius is 0 to 1. This value is used as the downstream influence radius corresponding to each conflict node and is used for the subsequent division of isolated conflict nodes and cascading conflict nodes.

[0032] In one specific embodiment, step S3, dividing the set of conflicting nodes into isolated conflicting nodes and cascading conflicting nodes based on the downstream influence radius, includes: The weighted downstream influence radius corresponding to each contradiction node in the contradiction node set is compared with the preset division threshold. Each contradiction node whose weighted downstream influence radius is lower than the preset division threshold is marked as an isolated contradiction node, and each contradiction node whose weighted downstream influence radius is not lower than the preset division threshold is marked as a cascaded contradiction node, thus obtaining a subset of isolated contradiction nodes and a subset of cascaded contradiction nodes. For each level of cascading conflict node in the subset of cascading conflict nodes, extract the corresponding set of all downstream clause nodes. Mark the level of cascading conflict nodes that contain conflict nodes of obligation, conflict nodes of breach of contract, or conflict nodes of exemption in the set of all downstream clause nodes as composite cascading conflict nodes, and mark the remaining level of cascading conflict nodes as single cascading conflict nodes, thus obtaining a subset of cascading conflict nodes subdivided by cascading type. By merging the subset of isolated contradictory nodes, the single cascaded contradictory nodes, and the compound cascaded contradictory nodes, we obtain a set of contradictory nodes divided according to isolated contradictory nodes and cascaded contradictory nodes.

[0033] Specifically, the preset threshold for division is set to 0.4. This value is determined based on the empirical value of the proportion of core rights and obligations clauses in the full text of the contract. When the weighted downstream influence radius of a conflict node is not less than 0.4, it indicates that the influence scope of the conflict node covers at least 40% of the clause nodes in the entire diagram, belonging to a conflict that has a systematic impact on the overall rights and obligations structure of the contract, and is marked as a cascading conflict node. When it is less than 0.4, it indicates that the influence scope is limited to a few clauses, and is marked as an isolated conflict node. The criteria for determining a compound cascading conflict node are as follows: if at least one node in the full set of downstream clause nodes of a cascading conflict node belongs to any one of the following categories: obligation conflict node, breach of contract conflict node, or exemption conflict node, that is, if there is a conflict node in the downstream clause nodes, then the cascading conflict node is marked as a compound cascading conflict node, indicating that there are secondary conflict nodes in the propagation path of the conflict node, and the conflict structure is nested; a cascading conflict node that does not contain any conflict nodes in the full set of downstream clause nodes is marked as a single cascading conflict node, indicating that there are no secondary conflicts in its propagation path, and the conflict structure is a single-layer propagation.

[0034] The subdivision of single-cascaded conflict nodes and compound-cascaded conflict nodes directly determines the allocation of approval authority levels in the subsequent step S4. Compound-cascaded conflict nodes, due to their nested conflict structure, correspond to higher approval authority levels, while single-cascaded conflict nodes correspond to lower approval authority levels, providing differentiated input for step S4. The final merged set of conflict nodes consists of three parts: a subset of isolated conflict nodes, a subset of single-cascaded conflict nodes, and a subset of compound-cascaded conflict nodes. These three parts do not overlap, and the total number of nodes after merging is equal to the total number of nodes in the original conflict node label set, ensuring that all conflict nodes are covered without omission.

[0035] In one specific embodiment, step S4 includes: For each isolated conflict node in the subset of isolated conflict nodes, based on the mapping relationship between the edge type of the directed edge corresponding to each isolated conflict node and the direct stakeholder, a stakeholder approval node is inserted at the corresponding position of the preset linear approval path to obtain the isolated conflict approval path. For each level of cascading conflict node in the subset of cascading conflict nodes, taking each level of cascading conflict node as the axis, and based on the edge type combination of the weighted downstream influence radius corresponding to each level of cascading conflict node and the full set of downstream clause nodes, three types of conflict solution forks are generated: modified clauses, deleted clauses, and supplementary agreements, thus obtaining the set of conflict solution forks corresponding to each level of cascading conflict node; Based on the number of complex cascading conflict nodes and the weighted downstream influence radius in each conflict solution fork set, the approval authority levels of each conflict solution fork are divided into the department head level, the legal review level, and the senior management decision-making level, thus obtaining the approval authority levels corresponding to each conflict solution fork. By merging the isolated conflict approval path with the set of conflict solutions and the various approval authority levels, a bifurcated arbitration approval path is obtained.

[0036] Specifically, the mapping relationship of direct stakeholders is determined based on the edge type of the directed edges: isolated conflict nodes corresponding to obligation triggering relationships are inserted into the approval node of the person in charge of the party directly bound by the obligation clause; isolated conflict nodes corresponding to breach of contract triggering relationships are inserted into the approval node of the person in charge of the contract management department; and isolated conflict nodes corresponding to exemption triggering relationships are inserted into the approval node of the legal department. The insertion position is after the approval stage of the clause corresponding to the conflict node in the preset linear approval path and before the next approval stage, thus obtaining an isolated conflict approval path containing stakeholder approval nodes. The generation basis for the three types of conflict solution forks is as follows: when the proportion of obligation triggering relationships in the edge type combination of the entire downstream clause node set is the highest, a modification clause fork is generated, with the direction of adjusting the obligation performance conditions as the handling direction; when the proportion of breach of contract triggering relationships is the highest, a deletion clause fork is generated, with the direction of removing conflicting breach of contract clauses as the handling direction; when the proportion of exemption triggering relationships is the highest, a supplementary agreement fork is generated, with the direction of supplementing the exemption scope agreement as the handling direction; when the weighted downstream influence radius is not less than 0.7, the three types of forks are generated simultaneously, and the approver chooses the handling path.

[0037] The three-tiered boundaries for approval authority levels are determined based on the number of complex cascading conflict nodes and the weighted downstream influence radius: when the number of complex cascading conflict nodes is 0 and the weighted downstream influence radius is less than 0.6, it corresponds to the department head level; when the number of complex cascading conflict nodes is not 0 or the weighted downstream influence radius is between 0.6 and 0.8, it corresponds to the legal review level; when the number of complex cascading conflict nodes is not 0 and the weighted downstream influence radius is not less than 0.8, it corresponds to the senior management decision-making level. The higher level is used when both conditions are met, ensuring a monotonous correspondence between the severity of the conflict and the approval authority level. The merging method for isolated conflict approval paths and conflict solution bifurcation sets is as follows: using a preset linear approval path as the main trunk, the stakeholder approval nodes inserted in the isolated conflict approval path are embedded into the corresponding positions on the main trunk according to the clause order. The conflict solution bifurcation sets corresponding to each level of cascading conflict node are attached to the corresponding positions of the cascading conflict node in the main trunk in a bifurcation structure. Each bifurcation node is labeled with its corresponding approval authority level. The merged result constitutes a bifurcation arbitration approval path centered on the conflict node.

[0038] Figure 3This diagram illustrates the comparison of cross-clause contradiction detection accuracy in the embodiments of this application. In these embodiments, the horizontal axis represents the number of test contract samples, and the vertical axis represents the cross-clause contradiction detection accuracy. The three curves correspond to single-Agent linear review, RAG enhanced review, and the dual-Agent game graph method of this scheme, respectively. As the number of test samples increases from 50 to 500, the detection accuracy of this scheme continuously improves from 0.67 to 0.89. This is higher than the 0.51 to 0.63 of the single-Agent linear review method and the 0.58 to 0.71 of the RAG enhanced review method at all sample sizes. This verifies that the cross-clause contradiction detection effect of dual-Agent adversarial reasoning combined with clause semantic game graph propagation is superior to the existing single-modal linear review scheme.

[0039] The above describes the multimodal agent-driven enterprise OA review and approval decision method in the embodiments of this application. The following describes the multimodal agent-driven enterprise OA review and approval decision system in the embodiments of this application. One embodiment of the multimodal agent-driven enterprise OA review and approval decision system in the embodiments of this application includes: The input module is used to input the text content of each clause in the contract document into the A party's interest agent and the B party's interest agent respectively. The A party's interest agent and the B party's interest agent independently generate opposing interpretation vectors for each clause, and construct the opposing interpretation vectors into a clause semantic game graph based on the logical triggering relationship between each clause. The marking module is used to calculate the cosine distance between the opposing interpretation vectors of A and B for each clause node in the clause semantic game graph, and to mark the nodes whose cosine distance exceeds a preset threshold as contradictory nodes, thereby obtaining the set of contradictory nodes in the clause semantic game graph. The calculation module is used to start from each conflict node in the set of conflict nodes, propagate along the directed edges of the semantic game graph of the clause to the downstream clauses, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. The reconstruction module is used to reconstruct a preset linear approval path into a bifurcated arbitration approval path with the conflict node as the axis, based on the type of the isolated conflict node and the cascaded conflict node and the corresponding downstream influence radius. The bifurcated arbitration approval path determines the conflict solution and approval authority level corresponding to each bifurcation.

[0040] This invention also provides a multimodal agent-driven enterprise OA review and approval decision-making device, which can be a server. The multimodal agent-driven enterprise OA review and approval decision-making device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the multimodal agent-driven enterprise OA review and approval decision-making device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the multimodal agent-driven enterprise OA review and approval decision-making device stores the corresponding data in this embodiment. The network interface of the multimodal agent-driven enterprise OA review and approval decision-making device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0041] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the multimodal agent-driven enterprise OA review and approval decision-making method.

[0042] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0043] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a multimodal agent-driven enterprise OA review and approval decision-making device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal agent-driven enterprise OA review and approval decision-making method, characterized in that, The method includes: Step S1: Input the text content of each clause in the contract document into Party A's interest agent and Party B's interest agent respectively. Party A's interest agent and Party B's interest agent independently generate opposing interpretation vectors for each clause. Based on the logical triggering relationship between each clause, construct the opposing interpretation vectors into a clause semantic game graph. Step S2: Calculate the cosine distance between the opposing interpretation vectors of A and B for each node in the semantic game graph of the clauses, and mark the nodes whose cosine distance exceeds a preset threshold as contradictory nodes to obtain the set of contradictory nodes in the semantic game graph of the clauses. Step S3: Starting from each conflict node in the set of conflict nodes, propagate downstream along the directed edges of the semantic game graph of the clause, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. Step S4: Based on the type of the isolated conflict node and the cascaded conflict node and the corresponding downstream influence radius, the preset linear approval path is reconstructed into a bifurcated arbitration approval path with the conflict node as the axis. The conflict solution and approval authority level corresponding to each bifurcation are determined by the bifurcated arbitration approval path.

2. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 1, characterized in that, Step S1 includes: The text content of each clause in the contract document is input into the A-party interest agent and the B-party interest agent respectively. The A-party interest agent is set to maximize the rights of the A-party and generates an A-party interpretation vector for each clause. The B-party interest agent is set to maximize the rights of the B-party and generates a B-party interpretation vector for each clause. The A-party interpretation vector and the B-party interpretation vector corresponding to each clause are combined to form the opposing interpretation vector of each clause. Based on the textual semantics of each clause, the logical triggering relationships between each clause are extracted, and the logical triggering relationships are divided into obligation triggering relationships, breach of contract triggering relationships and exemption triggering relationships. The logical triggering relationships of each clause are then labeled as the edge types of directed edges. The opposing interpretation vectors of each clause are used as node attributes, and the directed edges between each clause and their edge types are used as graph structures to construct the semantic game graph of the clauses.

3. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 2, characterized in that, Step S2 includes: For each clause node in the semantic game graph, the interpretation vector of Party A and the interpretation vector of Party B are extracted respectively. Based on the interpretation vector of Party A and the interpretation vector of Party B, the cosine distance of each clause node is calculated to obtain the cosine distance of the interpretation vector of each clause node. The cosine distance of the interpretation vector of each clause node is compared with a preset threshold. Clause nodes whose cosine distance of the interpretation vector exceeds the preset threshold are marked as contradictory nodes, thus obtaining a set of contradictory node marks. Based on the edge type of the directed edges corresponding to each conflicting node in the conflicting node label set, each conflicting node is divided into obligation conflicting nodes, breach of contract conflicting nodes, and exemption conflicting nodes, thus obtaining the conflicting node set classified by edge type.

4. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 3, characterized in that, In step S3, starting from each conflicting node in the set of conflicting nodes, the process propagates downstream along the directed edges of the semantic game graph of the clauses, including: Taking each contradictory node in the set of contradictory nodes as the starting point of propagation, the first layer of propagation is carried out along the directed edges corresponding to each contradictory node in the semantic game graph of the clauses, to the downstream clause nodes directly connected to each contradictory node, so as to obtain the first layer of downstream clause node set of each contradictory node. Based on the directed edges corresponding to each downstream clause node in each first-level downstream clause node set, the propagation continues to the secondary downstream clause nodes directly connected to each downstream clause node, iterating layer by layer until there are no new downstream clause nodes reachable in the clause semantic game graph, thus obtaining the full set of downstream clause nodes for each conflict node. The ratio of the number of nodes in the full set of downstream clause nodes of each conflict node to the total number of all clause nodes in the clause semantic game graph is determined as the downstream influence radius of each conflict node, thus obtaining the downstream influence radius corresponding to each conflict node.

5. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 4, characterized in that, In step S3, the downstream influence radius of each conflict node is calculated, including: For the full set of downstream clause nodes of each conflict node, extract the edge type of the directed edge corresponding to each downstream clause node, and classify and count each downstream clause node in the full set of downstream clause nodes according to the obligation triggering relationship, the breach of contract triggering relationship and the exemption triggering relationship to obtain the number of downstream clause nodes of each type of conflict node. Based on the number of downstream clause nodes of each conflict node, the ratio of the number of downstream clause nodes corresponding to each edge type to the total number of clause nodes in the clause semantic game graph is calculated to obtain the downstream influence radius of each conflict node. The downstream influence radii of each conflict node are weighted and summed according to the weights of the obligation triggering relationship, the breach of contract triggering relationship, and the exemption triggering relationship, to obtain the weighted downstream influence radius of each conflict node. The weighted downstream influence radius is used as the downstream influence radius corresponding to each conflict node.

6. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 5, characterized in that, In step S3, the set of conflicting nodes is divided into isolated conflicting nodes and cascading conflicting nodes based on the downstream influence radius, including: The weighted downstream influence radius corresponding to each contradictory node in the set of contradictory nodes is compared with a preset division threshold. Each contradictory node whose weighted downstream influence radius is lower than the preset division threshold is marked as an isolated contradictory node, and each contradictory node whose weighted downstream influence radius is not lower than the preset division threshold is marked as a cascaded contradictory node, thus obtaining a subset of isolated contradictory nodes and a subset of cascaded contradictory nodes. For each level of cascading conflict node in the subset of cascading conflict nodes, extract the corresponding set of all downstream clause nodes. Mark the level of cascading conflict nodes that contain the obligation conflict node, the breach of contract conflict node, or the exemption conflict node in the set of all downstream clause nodes as composite cascading conflict nodes, and mark the remaining level of cascading conflict nodes as single cascading conflict nodes, thus obtaining the subset of cascading conflict nodes subdivided by cascading type. The isolated contradictory node subset, the single cascaded contradictory node, and the composite cascaded contradictory node are merged to obtain the contradictory node set divided according to isolated contradictory nodes and cascaded contradictory nodes.

7. The multimodal agent-driven enterprise OA review and approval decision-making method according to claim 6, characterized in that, Step S4 includes: For each isolated conflict node in the subset of isolated conflict nodes, based on the mapping relationship between the edge type of the directed edge corresponding to each isolated conflict node and the direct stakeholder, a stakeholder approval node is inserted at the corresponding position of the preset linear approval path to obtain the isolated conflict approval path. For each level of cascading conflict node in the subset of cascading conflict nodes, taking each cascading conflict node as the axis, and based on the weighted downstream influence radius corresponding to each cascading conflict node and the edge type combination of the full set of downstream clause nodes, three types of conflict solution forks are generated: modified clause, deleted clause, and supplementary agreement, thus obtaining the conflict solution fork set corresponding to each level of cascading conflict node; Based on the number of composite cascaded conflict nodes in each set of conflict solution forks and the weighted downstream influence radius, the approval authority levels of each conflict solution fork are divided into department head level, legal review level and senior management decision-making level, thus obtaining the approval authority level corresponding to each conflict solution fork. The isolated conflict approval path is merged with the conflict solution fork set and each of the approval authority levels to obtain the forked arbitration approval path.

8. A multimodal agent-driven enterprise OA review and approval decision-making system, characterized in that, For implementing the multimodal agent-driven enterprise OA review and approval decision-making method as described in any one of claims 1-7, the multimodal agent-driven enterprise OA review and approval decision-making system includes: The input module is used to input the text content of each clause in the contract document into the A party's interest agent and the B party's interest agent respectively. The A party's interest agent and the B party's interest agent independently generate opposing interpretation vectors for each clause, and construct the opposing interpretation vectors into a clause semantic game graph based on the logical triggering relationship between each clause. The marking module is used to calculate the cosine distance between the opposing interpretation vectors of A and B for each clause node in the clause semantic game graph, and to mark the nodes whose cosine distance exceeds a preset threshold as contradictory nodes, thereby obtaining the set of contradictory nodes in the clause semantic game graph. The calculation module is used to start from each conflict node in the set of conflict nodes, propagate along the directed edges of the semantic game graph of the clause to the downstream clauses, calculate the downstream influence radius of each conflict node, and divide the set of conflict nodes into isolated conflict nodes and cascaded conflict nodes based on the downstream influence radius. The reconstruction module is used to reconstruct a preset linear approval path into a bifurcated arbitration approval path with the conflict node as the axis, based on the type of the isolated conflict node and the cascaded conflict node and the corresponding downstream influence radius. The bifurcated arbitration approval path determines the conflict solution and approval authority level corresponding to each bifurcation.

9. A multimodal agent-driven enterprise OA review and approval decision-making device, characterized in that, The system includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the multimodal agent-driven enterprise OA review and approval decision-making method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the multimodal agent-driven enterprise OA review and approval decision-making method as described in any one of claims 1 to 7.

Images