Semantic matching control optimization method and system based on procurement agreement constraints

Abstract

The application discloses a semantic matching control optimization method and system based on procurement agreement constraints, relates to the technical field of data processing, and comprises the following steps: obtaining a semantic matching result between agreement clause texts and response texts, and extracting corresponding initial confidence degrees; parsing the agreement clause texts into logical expression structures, extracting connection types between the agreement clauses, obtaining logical constraint conditions, performing logical relationship mapping on the initial confidence degrees, generating confidence degree correction propagation paths, and correcting and delivering the initial confidence degrees to obtain a confidence degree atlas; mapping the confidence degree atlas from a semantic matching space to a low-dimensional optimization control space to obtain a probability feasible region; and in the probability feasible region, performing hierarchical correction on an optimization solving process according to confidence degree distribution of the confidence degree atlas to generate a procurement control scheme, so that the problem that the optimization control feasible region is narrowed and the feasibility of the procurement control scheme is reduced due to confidence degree accumulation and attenuation of each agreement clause after semantic matching is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a semantic matching control optimization method and system based on procurement agreement constraints. Background Technology

[0002] A procurement agreement is a crucial written document used by enterprises to define the supply content, delivery requirements, quality standards, service conditions, and performance responsibilities during procurement activities. In actual procurement management, it is typically necessary to semantically match the supplier's response with multiple clauses in the procurement agreement to determine whether the supplier's response covers the agreement requirements, and further generate a procurement control plan based on the matching results. Since procurement agreements usually contain multi-level clause structures, and different clauses often have referencing relationships, conditional relationships, and pre- and post-constraint relationships, it is necessary to conduct item-by-item correlation analysis between agreement clauses and response texts using semantic matching methods to provide a basis for subsequent procurement decisions.

[0003] However, in the process of semantic matching of multiple agreement clauses, the matching results of each clause are not completely independent. Existing semantic matching methods typically calculate the matching degree for each individual clause and generate confidence information based on the matching results. When multiple related clauses participate in the matching sequentially, the confidence deviation generated by the preceding clauses will gradually affect the matching results of subsequent clauses, causing the matching confidence to cumulatively decrease as clauses are passed through. Especially when there are conditional constraints, related references, and multi-level logical relationships in the procurement agreement, a decrease in the confidence of a single clause may further reduce the effective matching range of related clauses, continuously narrowing the range of selectable procurement solutions in the subsequent optimization and control process. This makes the final procurement control solution susceptible to the influence of local low-confidence matching results, reducing the feasibility of the procurement control solution satisfying the overall agreement constraints.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a semantic matching control optimization method and system based on procurement agreement constraints. By parsing the logical constraint relationships between agreement clauses and performing correlation correction and transmission on the semantic matching confidence based on the logical constraint conditions, a stable confidence spectrum is formed and mapped to generate a probabilistic feasible region. Then, the optimization solution process is hierarchically controlled according to the confidence distribution to solve the problem that after semantic matching of multiple agreement clauses, the confidence of the matching results accumulates and decays one by one, causing the feasible region of subsequent optimization control links to continuously narrow, thereby reducing the feasibility of the procurement control scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The semantic matching control optimization method based on procurement agreement constraints includes the following steps: obtaining the semantic matching results between the agreement clause text and the response text, and extracting the initial confidence level corresponding to the semantic matching results; parsing the agreement clause text into a logical expression structure, and extracting the connection type between the agreement clauses from the logical expression structure to obtain the logical constraints; mapping the logical constraints to propagation nodes to obtain an acyclic propagation graph, traversing the acyclic propagation graph to obtain a confidence level correction propagation path, and performing hop-by-hop correction propagation and backtracking convergence on the initial confidence level based on the confidence level correction propagation path to generate a confidence level graph; extracting the node confidence level values ​​of the graph vertices in the confidence level graph to form a high-dimensional confidence level vector to obtain the semantic matching space, and projecting it to a low-dimensional optimization control space using the UMAP algorithm to obtain the probabilistic feasible region for optimization solution; within the probabilistic feasible region, dividing the level intervals according to the confidence level distribution, configuring gradient and hop correction strategies to perform hierarchical correction on the optimization solution process, and outputting the procurement control scheme.

[0007] In a preferred embodiment, obtaining the semantic matching result between the agreement clause text and the response text, and extracting the initial confidence level corresponding to the semantic matching result, specifically involves: segmenting the agreement clause text into a clause semantic fragment set according to semantic units, and segmenting the response text into a response semantic fragment set according to sentence boundaries; performing bidirectional alignment on the clause semantic fragment set and the response semantic fragment set to obtain a fragment-level semantic matching matrix; extracting semantic matching pairs along the main diagonal of the semantic matching matrix and calculating the semantic distance between the semantic matching pairs; and mapping the semantic distance between the semantic matching pairs to the initial confidence level.

[0008] In a preferred embodiment, the step of parsing the agreement clause text into a logical expression structure and extracting the connection types between the agreement clauses from the logical expression structure to obtain logical constraints specifically involves: obtaining the semantic reference relationships of the agreement clause text and performing dependency parsing on the agreement clause text to obtain several logical units; obtaining the coreference and temporal relationships between the several logical units and constructing intra-clause logical relationship edges based on the coreference and temporal relationships; concatenating the intra-clause logical relationship edges according to the inter-clause semantic reference relationships to obtain a cross-clause logical expression graph; labeling the nodes in the cross-clause logical expression graph with logical types and identifying the connection types, including conjunction connection, disjunction connection, and conditional connection; converting the labeled logical expression graph into a sequence of logical constraints and arranging the sequence of logical constraints in topological order to obtain the logical constraints.

[0009] In a preferred embodiment, the step of mapping logical constraints to propagation nodes to obtain an acyclic propagation graph and traversing the acyclic propagation graph to obtain a confidence-corrected propagation path specifically involves: mapping each logical node in the logical constraints to a propagation node in the confidence propagation graph, determining the propagation direction between propagation nodes based on the connection type in the logical constraints, and obtaining initial propagation edges; performing redundancy detection on the initial propagation edges and deleting propagation edges that form closed loops to obtain an acyclic propagation graph; traversing all reachable paths along the propagation direction in the acyclic propagation graph to obtain a confidence-corrected propagation path, and calculating the path length and logical operation type on each path in the confidence-corrected propagation path.

[0010] In a preferred embodiment, the step of correcting the initial confidence level based on the confidence level propagation path by performing hop-by-hop correction and backtracking to generate a confidence level graph specifically involves: selecting the shortest path from the set of confidence level correction propagation paths as the baseline propagation path, and propagating the initial confidence level hop-by-hop along the baseline propagation path according to the logical operation type on the path to obtain the propagation confidence level of each propagation node; using the path backtracking method to merge the propagation confidence levels on non-baseline propagation paths with the propagation confidence levels on the baseline propagation path to obtain the merged confidence level of each propagation node; normalizing the merged confidence level of each propagation node to generate a node confidence level value; and constructing a confidence level graph with propagation nodes as graph vertices and propagation direction and merging relationship as graph edges.

[0011] In a preferred embodiment, the extraction of node confidence values ​​from graph vertices in the confidence graph is combined into a high-dimensional confidence vector to obtain a semantic matching space. This space is then projected onto a low-dimensional optimization control space using the UMAP algorithm to obtain a probabilistic feasible region for optimization. Specifically, this involves: extracting node confidence values ​​from each graph vertex in the confidence graph and combining these values ​​into a high-dimensional confidence vector to obtain the semantic matching space; projecting the high-dimensional confidence vector from the semantic matching space to the low-dimensional optimization control space using the UMAP algorithm to obtain an initial projection point set; performing density clustering on the initial projection point set in the low-dimensional optimization control space to obtain confidence clustering regions; extracting the boundary points of the confidence clustering regions and constructing a convex hull using these boundary points as vertices to obtain the geometric boundary of the probabilistic feasible region; and parameterizing the geometric boundary of the probabilistic feasible region to generate a descriptive parameter set for the probabilistic feasible region.

[0012] In a preferred embodiment, the step of density clustering the initial projection point set in the low-dimensional optimization control space to obtain a confidence cluster region specifically involves: performing neighborhood radius adaptive density clustering on the initial projection point set in the low-dimensional optimization control space to obtain initial clusters; calculating inter-cluster similarity for the initial clusters; merging initial clusters with inter-cluster similarity exceeding a preset threshold to obtain merged clusters; extracting the cluster center points of the merged clusters; constructing confidence equipotential circles with the cluster center points as the centers; and marking the overlapping areas of the confidence equipotential circles as confidence cluster regions.

[0013] In a preferred embodiment, the step of dividing the confidence level intervals according to the confidence level distribution within the probabilistic feasible region and configuring gradient and jump correction strategies to perform hierarchical corrections on the optimization solution process, and outputting a procurement control scheme, specifically involves: obtaining the confidence level distribution of the confidence level map, and dividing the confidence level distribution of the confidence level map into several confidence level intervals according to the confidence level value within the probabilistic feasible region; configuring a corresponding correction strategy for each confidence level interval, the correction strategy including a gradient correction strategy and a jump correction strategy; determining the confidence level of candidate solutions generated during the optimization solution process to determine the confidence level interval to which the candidate solution belongs; dynamically adjusting the search direction and search step size of the candidate solution according to the confidence level interval to which the candidate solution belongs by calling the corresponding correction strategy to obtain the dynamically adjusted candidate solution; verifying the feasibility of the dynamically adjusted candidate solution, and outputting the candidate solution that passes the verification as the procurement control scheme.

[0014] In a preferred embodiment, the feasibility verification of the dynamically adjusted candidate solution specifically involves: extracting the control parameter sequence from the dynamically adjusted candidate solution and obtaining the node confidence value of the graph vertex to which each control parameter belongs in the confidence graph; marking control parameters with node confidence values ​​lower than a preset verification threshold as parameters to be verified, and extracting the associated parameter group of the parameter to be verified in the candidate solution, wherein the associated parameter group includes upstream and downstream parameters that have a logical transitive relationship with the parameter to be verified; and performing feasibility verification on the parameter to be verified and its associated parameter group to obtain candidate solutions that pass the feasibility verification.

[0015] The technical effects and advantages of the semantic matching control optimization method and system based on procurement agreement constraints of this invention are as follows: This invention achieves a quantitative representation of the matching degree between the supplier's response content and the procurement agreement requirements by obtaining the semantic matching results between the agreement clause text and the response text and extracting the corresponding initial confidence scores. Furthermore, it parses the agreement clause text into a logical expression structure and extracts the connection types between the clauses to obtain logical constraints reflecting the relationship between the clauses, enabling the effective identification of conditional constraints, combination relationships, and dependencies between different agreement clauses. Subsequently, it maps the initial confidence scores to logical relationships based on the logical constraints, forming a confidence score correction propagation path, and corrects and propagates the initial confidence scores according to the confidence score correction propagation path, resulting in a result that reflects the confidence score impact between clauses. This invention utilizes a confidence graph to avoid the problem of continuously decreasing overall confidence caused by the step-by-step propagation of confidence from single clauses in traditional semantic matching. Based on this, the confidence graph is mapped from the semantic matching space to a low-dimensional optimization control space, forming a probabilistic feasible region for optimization. This allows the search scope of the procurement control scheme to be dynamically adjusted based on the constraints of the agreement clauses, preventing excessive contraction of the feasible region due to local low-confidence matching results. Finally, within the probabilistic feasible region, the optimization process is hierarchically corrected according to the confidence distribution in the confidence graph, allowing different confidence regions to adopt corresponding optimization adjustment methods, improving the stability and constraint adaptability of the procurement control scheme generation process. Therefore, this invention can maintain the effectiveness of confidence propagation during the semantic matching of multiple agreement clauses, reduce the impact of accumulated matching errors on subsequent optimization processes, and solve the problem of decreasing feasibility of the procurement control scheme due to the continuous narrowing of the feasible region in the optimization control stage caused by the cumulative decay of confidence in matching results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the semantic matching control optimization method based on procurement agreement constraints according to the present invention.

[0017] Figure 2 The system architecture diagram provided for the embodiments of the present invention includes a confidence extraction module, a constraint construction module, a path generation module, a graph construction module, a graph mapping module, and a procurement control module, as well as the data flow relationships between the modules.

[0018] Figure 3 The schematic diagram of the protocol terms logic expression diagram and confidence spectrum structure provided in the embodiments of the present invention exemplarily shows 4 logic units (N1 to N4) and connection types (conditional connection, conjunction connection, disjunctive connection), as well as 4 propagation nodes (P1 to P4) and node confidence values ​​after correction and propagation (0.92, 0.78, 0.85, 0.65).

[0019] Figure 4This is a schematic diagram of confidence correction transmission and timing interaction provided in an embodiment of the present invention. It exemplarily illustrates the timing process of confidence hop-by-hop transmission, path backtracking and merging, and normalization processing along the reference transmission path and non-reference transmission path, and marks the confidence calculation and transmission operations at each time point. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, Figure 1 The present invention provides a semantic matching control optimization method based on procurement agreement constraints, comprising the following steps: S1, obtain the semantic matching results between the agreement terms text and the response text, and extract the initial confidence level corresponding to the semantic matching results; In this embodiment, the semantic matching result between the agreement terms text and the response text is obtained, and the initial confidence level corresponding to the semantic matching result is extracted, specifically as follows: The agreement text is segmented into semantic fragment sets according to semantic units, and the response text is segmented into response semantic fragment sets according to sentence boundaries; Bidirectional alignment is performed on the set of semantic fragments of terms and the set of semantic fragments of responses to obtain a fragment-level semantic matching matrix; Extract semantic matching pairs along the main diagonal of the semantic matching matrix and calculate the semantic distance between the semantic matching pairs; Map the semantic distance of the semantic matching pairs to the initial confidence level.

[0022] It should be noted that, firstly, the electronic agreement document corresponding to the procurement agreement to be evaluated is obtained. The electronic agreement document can be sourced from the procurement management platform, the contract management database, or a scanned agreement document after text recognition. When the agreement document is a scanned document, it undergoes layout correction, character recognition, and encoding standardization processing. All text is converted to a unified character encoding format, and then the original hierarchical structure of the agreement is restored according to clause numbers, chapter titles, and body content to form the agreement clause text. Next, the supplier's response documents are obtained. Response documents include tender documents, commercial response documents, technical response documents, or supplementary explanation documents. The response documents also undergo character encoding standardization, special character removal, and paragraph restoration processing. The body content is extracted based on the titles, serial numbers, and paragraph boundaries in the documents to form the response text.

[0023] After obtaining the text of the agreement terms, the agreement is first initially segmented according to the numbering of the first-level, second-level, and third-level terms. Then, semantic units are identified within each term. When a term contains a description of obligations, restrictions, exceptions, and execution conditions, it is segmented using connectors indicating logical relationships. For cases with multiple consecutive connectors, segmentation is prioritized based on the boundary between conditional and result statements, and then further subdivided according to parallel relationships, ensuring that each semantic fragment corresponds to only one complete constraint, ultimately forming a set of semantic fragments for the terms.

[0024] For the response text, the sentence boundaries are first identified by periods, semicolons, colons and newlines, and then the sentence units are further divided by combining bullet points, serial numbers and list structures. When the same sentence contains multiple independent response items, it is also further divided according to logical connectors so that each response semantic fragment corresponds to only one complete response content, forming a set of response semantic fragments.

[0025] Further, after obtaining the set of clause semantic fragments and the set of response semantic fragments, firstly, a clause fragment index sequence and a response fragment index sequence are established according to the order of each semantic fragment. Then, keywords, action words, object words, constraint words, and quantity information are extracted from each semantic fragment and uniformly converted into semantic feature representations. Subsequently, using each clause semantic fragment as a benchmark, it is compared item by item with all response semantic fragments. A first matching sequence is generated based on the degree of overlap of the semantic features between the two. Then, using each response semantic fragment as a benchmark, it is compared item by item with all clause semantic fragments in reverse to generate a second matching sequence. When a clause semantic fragment and a response semantic fragment are simultaneously located at the optimal matching positions in the first matching sequence and the second matching sequence, respectively, a bidirectional matching relationship is determined. After completing all bidirectional matching, the clause semantic fragments are used as matrix rows, and the response semantic fragments are used as matrix columns. Each set of bidirectional matching results is written into the corresponding matrix cell. Positions where no bidirectional matching relationship is formed are recorded as unmatched states, resulting in a fragment-level semantic matching matrix.

[0026] After obtaining the fragment-level semantic matching matrix, since the bidirectional matching results in the matrix are already arranged according to the clause order and response order, the matrix cells on the main diagonal are traversed sequentially from the top left corner to the bottom right corner, and the clause semantic fragments and response semantic fragments that have formed bidirectional matching relationships are extracted. Each pair of correspondences is considered a semantic matching pair. After extracting the semantic matching pairs, the number of consistent keywords, action words, constraint words, and quantity information in each pair are counted. The number of information items that appear only in the clause text but not in the response text, and the number of information items that appear only in the response text, are also counted. The total number of consistent items and the total number of differing items are then combined. As the number of consistent items increases, the semantic distance decreases; as the number of differing items increases, the semantic distance increases. Finally, the semantic distance of each pair of semantic matching pairs is determined according to the proportion of consistent items to all comparison items and the proportion of differing items to all comparison items.

[0027] Finally, all semantic distances are sorted in ascending order and then divided into several continuous intervals. Preferably, five levels are used: Level 1 corresponds to the smallest 20% of semantic distances, Level 2 to 20%-40%, Level 3 to 40%-60%, Level 4 to 60%-80%, and Level 5 to the largest 20%. Initial confidence intervals are then configured for each level: Level 1 to 0.90-1.00, Level 2 to 0.75-0.90, Level 3 to 0.60-0.75, Level 4 to 0.40-0.60, and Level 5 to 0.20-0.40. For semantic matching pairs falling within the same level, linear interpolation is performed based on their relative position within that level to ensure a continuous rather than discrete distribution of the initial confidence values. For completely identical semantic matching pairs, the initial confidence is set to 1.00. For semantic segments that do not form a valid semantic match, the initial confidence is uniformly set to 0.20.

[0028] In this embodiment, the formula for calculating the semantic distance of semantic matching pairs is as follows:

[0029] in, The semantic distance between the semantic matching pairs. This represents the total number of comparison terms in the current semantic matching pair. The number of semantically consistent items in the semantic fragments of the agreement terms and the semantic fragments of the response. The number of differences between the semantic fragments of the agreement terms and the semantic fragments of the response. The weights for the consistency terms range from 0.45 to 0.60. The weight of the difference term is determined, with a value ranging from 0.40 to 0.55, and satisfies the following condition: + =1.

[0030] S2, parses the agreement terms text into a logical expression structure, and extracts the connection types between the agreement terms from the logical expression structure to obtain logical constraints; In this embodiment, the agreement terms text is parsed into a logical expression structure, and the connection types between the agreement terms are extracted from the logical expression structure to obtain logical constraints, specifically: Obtain the semantic reference relationships of the agreement clause text, and perform dependency parsing on the agreement clause text to obtain several logical units; Obtain the coreference and temporal relationships between several logical units, and construct the logical relationship edges within the clauses based on the coreference and temporal relationships; By concatenating the logical relationship edges within the clauses according to the semantic reference relationships between clauses, a cross-clause logical expression diagram is obtained; Logical type labeling is performed on nodes in the cross-clause logical expression diagram, and connection types are identified, including conjunction connection, disjunction connection and conditional connection; The annotated logic representation diagram is transformed into a sequence of logic constraints, and the sequence of logic constraints is arranged in topological order to obtain the logic constraints.

[0031] It should be noted that after obtaining the text of the agreement terms, a clause index is first established according to the original clause numbers of the agreement, and the main text of each clause is read sequentially to identify the clause reference information contained in the text. When a clause number is identified, the reference relationship between the current clause and the referenced clause is directly established; when the referenced content does not directly provide a clause number, at least two paragraphs before and after are read continuously in combination with the context, and the referenced object is determined based on the most recently appearing clause number, forming a semantic reference relationship between clauses. After completing the identification of all reference relationships, dependency parsing is performed on each sentence of the agreement clause text. The predicate in the sentence is taken as the core component, and words that have subject-predicate, verb-object, attributive-head, adverbial-head, prepositional-object, and supplementary explanatory relationships with the predicate are identified respectively. They are then classified and combined according to dependency relationships, so that the same predicate and its related modifiers form a complete semantic expression unit; for cases where there are multiple predicates in the same sentence, each predicate is used as the core to construct a corresponding logical unit.

[0032] After obtaining several logical units, the system first reads the subject object, action object, and pronoun information of each logical unit sequentially. When a pronoun exists in a logical unit, the system searches backwards through the preceding logical units for noun objects, selecting the nearest and semantically consistent object as the corresponding entity for the pronoun, thus establishing a coreference relationship between the two logical units. If both logical units meet the corresponding conditions, the logical unit within the same clause is prioritized as the coreference object. If multiple candidate objects still exist, the consistency of object names is compared, and the logical unit with completely identical names is selected to establish a coreference relationship. Subsequently, the system continues to identify the temporal sequence information in each logical unit and establishes the sequential execution relationship between logical units by combining the numbering order and action execution order in the clauses. When both coreference and temporal relationships exist between two logical units, a directional logical relationship edge is established according to the execution direction; when only coreference exists, a reference relationship edge is established; and when only temporal relationships exist, an execution relationship edge is established.

[0033] Furthermore, the semantic reference relationships of the clauses are read, and the two clauses with reference relationships are respectively designated as the source clause and the target clause. The logical units corresponding to the reference positions in the source clause and the logical units corresponding to the referenced positions in the target clause are extracted. When a source logical unit references a target logical unit, a connection is established between the endpoint of the logical relationship edge corresponding to the source logical unit and the starting point of the logical relationship edge corresponding to the target logical unit, forming a continuous connection structure for logical relationship edges that were originally located in different clauses. Then, the entire semantic reference relationship is traversed, and the above connection process is repeated until all reference relationships are completed. For cases with multi-level references, such as the first clause referencing the second clause, and the second clause further referencing the third clause, the connection is completed layer by layer according to the reference order, allowing the logical relationship to be continuously transmitted. For cases with circular references, the direction of the first occurrence of the reference is recorded, and the extension to already visited clauses is stopped, thus avoiding the formation of infinite circular connections. After completing the concatenation of all logical relationship edges, each logical unit is used as a node, and logical relationship edges and cross-clause connection relationships are used as connecting edges to form a cross-clause logical expression graph covering all agreement clauses.

[0034] Furthermore, after obtaining the cross-clause logic representation diagram, logical connectors are first extracted according to the logical unit content corresponding to each node in the diagram, and the logical type of the nodes is labeled according to the semantic meaning of the logical connectors. After all nodes are labeled, the connection type is determined according to the logical type between the nodes at both ends of the connection edge; if both ends of the connection edge are conjunctive logical nodes, the corresponding connection edge is marked as a conjunctive connection; if at least one end is a disjunctive logical node, the corresponding connection edge is marked as a disjunctive connection; if the starting node is a conditional logical node, the corresponding connection edge is marked as a conditional connection. When the same connection edge satisfies multiple connection conditions at the same time, the labeling is completed in the order of conditional connection, conjunctive connection, and disjunctive connection to ensure that the conditional constraints have the highest priority in the subsequent logic propagation process.

[0035] Finally, all nodes and edges in the cross-clause logic expression graph are read sequentially. Each logical node's corresponding constraint content is treated as a logical constraint item, and pre- and post-dependent relationships are established between these logical constraints according to the direction of the edges. Then, logical constraints without pre-dependent relationships are identified from all logical constraints, added as starting constraints to the logical constraint condition sequence, and their corresponding output edges are deleted. The remaining logical constraints are then traversed, and logical constraints without pre-dependent relationships are added to the logical constraint condition sequence sequentially. This process is repeated until all logical constraints are sorted. When multiple logical constraints simultaneously meet the addition conditions, they are prioritized for sorting by protocol clause number in ascending order. If clause numbers are the same, they are sorted by logical unit number in ascending order. If logical unit numbers are still the same, they are sorted by the order in which connections were established. If logical constraints with pre-dependent relationships that cannot be eliminated are found, it indicates a circular dependency at the corresponding position. In this case, the initially established dependency is retained, subsequent circular connections are deleted, and the sorting process is repeated until a logical constraint condition sequence without circular dependencies is obtained.

[0036] In this embodiment, the sequence of logical constraints is... Figure 3 The protocol terms are logically represented in the diagram, where each node corresponds to an independent logical unit, namely, a condition node, a conjunction node, a disjunctive node, and a conditional node. The directed edges between nodes are labeled with their connection types (conditional connection, conjunction connection, disjunctive connection), clearly reflecting the dependencies and logical combinations among the protocol terms.

[0037] S3: Map logical constraints to propagation nodes to obtain an acyclic propagation graph. Traverse the acyclic propagation graph to obtain a confidence-corrected propagation path. Based on the confidence-corrected propagation path, perform hop-by-hop correction and backtracking to merge the initial confidence and generate a confidence graph. In this embodiment, logical constraints are mapped to propagation nodes to obtain an acyclic propagation graph. The acyclic propagation graph is then traversed to obtain a confidence-corrected propagation path, specifically: Each logical node in the logical constraints is mapped to a propagation node in the confidence propagation graph, and the propagation direction between propagation nodes is determined according to the connection type in the logical constraints to obtain the initial propagation edge; Redundancy detection is performed on the initial propagation edges, and propagation edges that form closed loops are deleted to obtain an acyclic propagation graph; In the acyclic propagation graph, traverse all reachable paths along the propagation direction to obtain confidence-corrected propagation paths, and calculate the path length and logical operation type of each path in the confidence-corrected propagation path.

[0038] It should be noted that after obtaining the logical constraints, each logical constraint item is read sequentially according to the order of the logical constraint sequence, and each logical constraint item is uniquely converted into a propagation node. Each propagation node retains the corresponding clause number, logical unit number, initial confidence level, and logical type identifier, so that the propagation node can maintain a one-to-one correspondence with the aforementioned semantic matching results. Then, the connection relationships between adjacent logical constraints are read, and the propagation direction between propagation nodes is established according to the connection type. When the connection type is a conditional connection, the condition node is used as the starting propagation node, and the node executed after the condition is met is used as the ending propagation node, so that the confidence level is propagated along the condition triggering direction. When the connection type is a conjunctive connection, all preceding propagation nodes are pointed to a common subsequent propagation node, so that the confidence levels of multiple propagation nodes jointly affect the subsequent propagation node. When the connection type is a disjunctive connection, the same starting propagation node points to multiple alternative propagation nodes, so that the propagation path can extend along multiple optional directions. After establishing the propagation direction, further check whether there are duplicate connections between propagation nodes. When two or more connection edges with the same direction are formed between two propagation nodes, only the connection edge established for the first time is retained. When there are two connection edges with opposite directions, the connection relationship between the predecessor node and the successor node is retained according to the order of arrangement in the logical constraint condition sequence, and the reverse connection is deleted, thereby forming an initial propagation edge with a unique propagation direction.

[0039] In this embodiment, the sequence of logical constraints can be represented as follows: Figure 3The confidence graph structure is derived from the nodes of the protocol clause logic representation graph. Each propagation node (P1 to P4) retains its original logic type and is assigned a confidence value (0.92, 0.78, 0.85, and 0.65 in the example) after undergoing the aforementioned confidence correction propagation path correction. The diagram uses dashed arrows to represent the mapping from the logic graph to the confidence graph, while solid directed edges represent the propagation direction of confidence along the logical dependencies. This diagram exemplifies the correspondence between four logic units and four propagation nodes, as well as the final distribution of each node's confidence under the influence of logical constraints, intuitively demonstrating the transformation process from the logic structure to the confidence propagation structure.

[0040] Further, after obtaining the initial propagation edge, all propagation nodes are traversed sequentially according to the propagation direction, and the propagation paths visited by each propagation node are recorded. When traversing along the current propagation direction, if a propagation node already recorded in the current propagation path is visited again, it is determined that the current propagation edge forms a closed loop. Subsequently, the establishment order of all propagation edges within the closed loop is analyzed, and the propagation edge with the earliest establishment order is retained. The propagation edge with the latest establishment order that leads to the formation of a closed loop is marked as a propagation edge to be deleted. If there are multiple candidate propagation edges to be deleted in a closed loop, propagation edges formed across clauses are deleted first. If they all belong to the same clause, the propagation edge with the largest connection span is deleted, where the connection span is determined by the difference between the number of the starting propagation node and the number of the ending propagation node. If the number differences are the same, the propagation edge with the later establishment order is deleted. After completing the processing of a closed loop, the propagation path is traversed again starting from the starting propagation node, and the existence of a new closed loop is checked. The above process is repeated until all propagation paths do not repeatedly visit the same propagation node. Preferably, after each deletion of a propagation edge, the connection relationship between propagation nodes is updated again, and a complete traversal is performed again; if no new closed loop is detected in two consecutive traversals, it is determined that the current propagation graph has met the acyclic requirement, and an acyclic propagation graph is finally obtained.

[0041] Finally, after obtaining the acyclic propagation graph, firstly, identify all propagation nodes without preceding propagation edges and use them as starting propagation nodes. Then, starting from each starting propagation node, sequentially visit its subsequent propagation nodes along the propagation direction. For each propagation edge traversed, record the corresponding propagation node number and propagation order. When propagation reaches the terminating propagation node without subsequent propagation edges, combine all propagation nodes traversed from the starting propagation node to the terminating propagation node into a complete propagation path and save it as a confidence-corrected propagation path. Then, return to the propagation node with the most recently untraversed propagation edge and continue the above traversal process along the new propagation direction until all propagation edges have been traversed at least once, thus obtaining all confidence-corrected propagation paths. After completing the extraction of confidence-corrected propagation paths, count the number of propagation edges traversed by each confidence-corrected propagation path and use the number of propagation edges as the path length of the corresponding path. At the same time, sequentially read the connection type corresponding to each propagation edge on the path and form a logical operation type sequence according to the propagation order. When conjunctive joins, disjunctive joins, and conditional joins exist simultaneously in the path, their original order in the propagation path is maintained and recorded, and the join types are not reordered. If multiple join types of the same type appear consecutively, they are merged and recorded as consecutive logical operation segments, and the number of consecutive occurrences is counted to reduce the number of times the same logical relationship is repeatedly parsed during subsequent confidence correction.

[0042] In this embodiment, the initial confidence level is corrected and propagated hop-by-hop based on the confidence level correction propagation path, and then backtracked and merged to generate a confidence level map, specifically as follows: The shortest path is selected from the set of confidence correction propagation paths as the baseline propagation path, and the initial confidence is propagated hop by hop along the baseline propagation path according to the logical operation type on the path to obtain the propagation confidence of each propagation node. The path backtracking method is used to combine the transmission confidence on the non-reference transmission path with the transmission confidence on the reference transmission path to obtain the combined confidence of each propagation node. The convergence confidence of each propagation node is normalized to generate a node confidence value; A confidence graph is constructed using propagation nodes as vertices and propagation directions and merging relationships as edges.

[0043] It should be noted that after obtaining the set of confidence correction propagation paths, the path length of each confidence correction propagation path is first read and sorted in ascending order of path length. The path with the shortest path length is selected as the baseline propagation path. When multiple paths have the same length, the number of conditional connections in each path is compared, and the confidence correction propagation path with fewer conditional connections is selected as the baseline propagation path. If the number of conditional connections is still the same, the starting propagation node numbers of the paths are further compared, and the confidence correction propagation path with the smaller starting propagation node number is selected as the baseline propagation path.

[0044] After determining the baseline propagation path, the starting propagation node of the path is taken as the starting point of the propagation. The initial confidence level corresponding to the propagation node is directly used as the starting propagation confidence level, and subsequent propagation nodes are visited one by one according to the propagation direction. Upon traversing each propagation edge, the logical operation type corresponding to that edge is read, and a confidence level is passed based on the logical operation type: When the propagation edge corresponds to a conjunctive connection, the propagation confidence level of the current propagation node is used as the basis for the propagation of the next propagation node, and the smaller of the two, combined with the initial confidence level of the next propagation node, is taken as the new propagation confidence level; when the propagation edge corresponds to a disjunctive connection, the propagation confidence level of the current propagation node is compared with the initial confidence level of the next propagation node, and the larger of the two is taken as the new propagation confidence level, to reflect that propagation can continue as long as one of the multiple alternative conditions is met; when the propagation edge corresponds to a conditional connection, it is first determined whether the logical condition corresponding to the current propagation node is true. If the logical condition is true, the propagation confidence level of the current propagation node continues to be passed along the propagation direction. If the logical condition is false, the propagation in the current direction is stopped, and the propagation confidence level of subsequent propagation nodes is kept at its initial confidence level.

[0045] Further, the remaining confidence-corrected propagation paths in the confidence-corrected propagation path set are read sequentially and processed as non-baseline propagation paths. First, the propagation confidence is calculated along the propagation direction of the non-baseline propagation path. When propagation reaches a propagation node intersecting with the baseline propagation path, this propagation node is designated as the merging node, and propagation stops. Then, starting from the merging node, the propagation is backtracked step-by-step along the propagation direction of the non-baseline propagation path back to its initial propagation node. The propagation confidence of each propagation node on the non-baseline propagation path is recorded sequentially. Returning to the merging node, the backtracked propagation confidence is compared node-by-node with the propagation confidence of the corresponding propagation node on the baseline propagation path. When the same propagation node has propagation confidence corresponding to two propagation paths, the difference between the two is calculated. When the difference does not exceed 0.05, the average of the two is used as the merging confidence of the propagation node. When the difference is greater than 0.05, the larger propagation confidence is retained as the merging confidence, and the smaller propagation confidence is recorded as an auxiliary propagation value for subsequent propagation nodes to refer to. If a propagation node is connected to three or more propagation paths at the same time, the above-mentioned merging process is completed in the order in which the propagation paths arrive at the merging node. Each time a merging is completed, the merging confidence of the current propagation node is updated and then used for the merging of the next propagation path, until all non-baseline propagation paths have completed the merging process.

[0046] Furthermore, after obtaining the convergence confidence scores for all propagation nodes, the largest and smallest convergence confidence scores among all propagation nodes are first calculated and used as the upper and lower boundaries for this normalization process. Then, the convergence confidence scores of each propagation node are read sequentially, and their relative positions within the aforementioned upper and lower boundaries are calculated. The convergence confidence scores of all propagation nodes are then uniformly converted to a range of 0 to 1, generating corresponding node confidence values. When the convergence confidence scores of all propagation nodes are the same, the node confidence values ​​of all propagation nodes are directly set to 1.00 to avoid situations where there is no effective distinguishability during the normalization process. After normalization, the nodes are sorted from largest to smallest confidence scores, and the difference in node confidence scores between adjacent propagation nodes is checked. When the difference is less than 0.01, the original sorting order remains unchanged. When multiple propagation nodes have completely identical node confidence scores, they are rearranged from smallest to largest propagation node number to ensure the uniqueness of the node arrangement.

[0047] Finally, each propagation node is treated as a vertex in the graph, and its node number, protocol clause number, logical type, node confidence value, and propagation level information are written into the corresponding vertex attributes. Then, all propagation edges in the acyclic propagation graph are read, and connections with consistent propagation directions are treated as directed graph edges connecting the corresponding vertices, ensuring the propagation direction aligns with the transmission direction in the logical constraints. Next, all merging relationships formed during path backtracking are read. When two or more propagation paths merge at the same propagation node, a merging relationship graph edge is added between the corresponding vertices, and the merging order and the propagation path numbers participating in the merging are recorded. When both transmission direction graph edges and merging relationship graph edges exist between the same pair of vertices, both types of edges are retained separately without merging, ensuring that both logical propagation relationships and path merging relationships are reflected simultaneously.

[0048] In this embodiment, as Figure 4 As shown, taking three propagation nodes P1, P2, and P3 as examples, the time-series process of confidence correction propagation along the confidence correction propagation path is illustrated in detail. The initial confidence of each node is marked on the left side of the figure (C1=0.90, C2=0.70, C3=0.60). Then, conjunctive propagation (taking the smaller value) and disjunctive propagation (taking the larger value) are performed sequentially along the baseline propagation path (path 1), resulting in a propagation confidence of 0.70 for both P2 and P3. At the same time, the non-baseline paths (path 2 and path 3) merge their respective propagation values ​​to the same node through path backtracking, and compare the confidence differences brought by different paths at the merging point (in this example, the difference is greater than 0.05, so the larger value of 0.70 is retained). Finally, after normalization processing, the final confidence of each node is generated (P1=0.92, P2=0.78, P3=0.85). Figure 4 The text clearly distinguishes the transmission direction of the baseline path and non-baseline paths, and marks the logical operation type (conjunction / disjunction) and corresponding confidence calculation process for each step. It fully reflects the complete data flow from the initial confidence input, through multi-path transmission and backtracking convergence, to the normalized output node confidence value.

[0049] S4. Extract the node confidence values ​​of the graph vertices in the confidence graph and combine them into a high-dimensional confidence vector to obtain the semantic matching space. Then, use the UMAP algorithm to project it onto the low-dimensional optimization control space to obtain the probabilistic feasible region for optimization solution. In this embodiment, the node confidence values ​​of the graph vertices in the confidence graph are extracted and combined into a high-dimensional confidence vector to obtain a semantic matching space. This space is then projected onto a low-dimensional optimization control space using the UMAP algorithm to obtain the probabilistic feasible region for optimization. Specifically: Extract the node confidence value of each vertex in the confidence graph, and combine the node confidence values ​​into a high-dimensional confidence vector to obtain the semantic matching space; The UMAP algorithm is used to project the high-dimensional confidence vector from the semantic matching space to the low-dimensional optimization control space to obtain the initial projection point set. Density clustering of the initial projected point set is performed in the low-dimensional optimization control space to obtain confidence clustering regions. Extract the boundary points of the confidence cluster region and construct the convex hull with the boundary points as vertices to obtain the geometric boundary of the probabilistic feasible region; Parameterize the geometric boundary of the probabilistic feasible region to generate a set of descriptive parameters for the probabilistic feasible region.

[0050] It should be noted that after obtaining the confidence graph, all graph vertices are first read sequentially in ascending order of vertex number, and the node confidence value corresponding to each vertex is extracted. Simultaneously, the protocol clause number, logical level, and propagation level information of the graph vertex are read to ensure that the nodes maintain a fixed arrangement order. Then, a one-to-one correspondence between node numbers and vector dimensions is established, so that each graph vertex corresponds to a fixed dimension in the high-dimensional confidence vector. The node confidence values ​​are then written into the corresponding dimensions sequentially according to the graph vertex number, forming the initial high-dimensional confidence vector. When multiple graph vertices belonging to the same protocol clause exist in the confidence graph, each vertex corresponds to a different dimension and is not merged to fully preserve the differences between logical nodes. When a graph vertex does not participate in the confidence transmission during the aforementioned logical propagation process and therefore has no node confidence value, the initial confidence value of the corresponding graph vertex is used as the data source for that dimension. After all graph vertices have been written, check each dimension for missing data. If a dimension is missing, it is filled by the average confidence value of the nodes of the two adjacent known dimensions. If the missing dimension is located at the beginning or end of the vector, it is filled by the confidence value of the node of the nearest known dimension, until all dimensions have corresponding data.

[0051] Furthermore, all high-dimensional confidence vectors are input into the UMAP algorithm to map them from the semantic matching space to the low-dimensional optimization control space. Specifically, firstly, distance calculations are performed on all high-dimensional confidence vectors, and then 15 to 30 neighboring vectors are selected as a neighborhood set for each high-dimensional confidence vector in ascending order of distance, preferably 20. If the total number of vectors is less than 15, all vectors except the current vector are included in the neighborhood set. Subsequently, high-dimensional spatial neighborhood connections are established based on the distances between each high-dimensional confidence vector and the vectors in its neighborhood set. Connections between nearest-neighbor vectors are considered local neighborhood connections, and the connection strength is determined based on the distance between vectors to form a high-dimensional neighborhood structure for spatial embedding. After establishing the high-dimensional neighborhood structure, the low-dimensional output dimension is set to 2-dimensional, and initial low-dimensional coordinates corresponding to each high-dimensional confidence vector are generated using a random initialization method. Subsequently, the low-dimensional coordinates are iteratively adjusted through the neighborhood cross-entropy optimization process of the UMAP algorithm, ensuring that the neighborhood connectivity relationships in the high-dimensional space are preserved in the low-dimensional space. The number of iterations is set to 200–500, and the iteration ends early when the change in the low-dimensional neighborhood structure is less than 0.001 in 20 consecutive iterations. After spatial embedding, each high-dimensional confidence vector corresponds to a two-dimensional projection coordinate point. All projection coordinate points constitute the initial projection point set, and each projection point is associated with the graph vertex number of the corresponding high-dimensional confidence vector for subsequent density clustering and probabilistic feasible region boundary extraction in the low-dimensional optimization control space.

[0052] Furthermore, the spatial distance between each projection point and its surrounding projection points is calculated according to the projection coordinates, and the number of adjacent projection points within the neighborhood of each projection point is counted. Then, based on the spatial distance and the number of neighboring points, the clustering of projection points is identified. When the distance between several projection points is continuously small and the number of projection points in their neighborhood is consistently higher than the surrounding area, these projection points are assigned to the same initial cluster region. This process is repeated for the remaining projection points until all projection points have been clustered. For cases where there are a few transitional projection points between two cluster regions, the distances between the transitional projection points and the centers of the two cluster regions are counted, and the transitional projection points are assigned to the closer cluster region. If the difference between the two distances does not exceed 5% of the average of all projection distances, these points are temporarily retained as boundary projection points, and their affiliation is further determined in subsequent boundary analysis.

[0053] Furthermore, all projection points are traversed sequentially along the periphery of each cluster region, and the spatial distribution of adjacent projection points around each projection point is statistically analyzed. When a projection point has no adjacent projection points on at least one side, or when all its adjacent projection points are located on the same side of the space, the projection point is marked as a boundary point. Subsequently, the boundary points are sorted clockwise according to their spatial positions in the low-dimensional optimization control space, and adjacent boundary points are connected sequentially to form the outer boundary of the region. After all boundary points have been extracted, the sorted boundary points are connected one by one as vertices to construct a convex hull structure that surrounds all cluster regions, so that all cluster regions are located inside the convex hull. When there are local depressions in the cluster region, the local depressions are ignored, and the outermost boundary point is still used to form the overall convex hull. When multiple independent convex hulls are formed between different cluster regions, the corresponding convex hulls are constructed respectively, and the distance between each convex hull is statistically analyzed. If the shortest distance between two adjacent convex hulls is less than 15% of the average distance between all boundary points, the two convex hulls are reconnected to form a unified geometric boundary.

[0054] Finally, the coordinates of all boundary points are read sequentially according to their arrangement, and the boundary length, boundary turning direction, and boundary connection order between adjacent boundary points are calculated. Simultaneously, the outer length, coverage area, and center position of the entire geometric boundary are statistically analyzed. Then, using the center position of the geometric boundary as a reference point, the direction and distance information of each boundary point relative to the center position are recalculated, and all boundary points are arranged sequentially in a clockwise direction, converting the entire geometric boundary into a unified parameter description form. Next, the total number of boundary points, average boundary length, maximum boundary length, minimum boundary length, number of consecutive boundary turns, and boundary closure state are statistically analyzed, and these parameters are unified into a set of description parameters for the probabilistic feasible region. For cases with multiple independent probabilistic feasible regions, corresponding description parameter sets are established and stored in descending order of coverage area; if multiple probabilistic feasible regions have the same coverage area, they are arranged in ascending order of the x-coordinate corresponding to the center position.

[0055] In this embodiment, density clustering is performed on the initial projection point set in the low-dimensional optimization control space to obtain confidence clustering regions, specifically: In the low-dimensional optimization control space, density clustering with adaptive neighborhood radius is performed on the initial projection point set to obtain the initial clusters; The initial clusters are calculated for inter-cluster similarity. The initial clusters with inter-cluster similarity exceeding a preset threshold are merged to obtain merged clusters. Extract the cluster center points of the merged clusters, construct confidence equipotential circles with the cluster center points as the center, and mark the overlapping areas of the confidence equipotential circles as confidence clustering areas.

[0056] It should be noted that, firstly, a spatial index is established based on the coordinates of the projected points in the low-dimensional optimization control space. Then, the Euclidean distance between each projected point and all its surrounding projected points is calculated sequentially, and these distances are arranged in ascending order to form a neighborhood distance sequence. Subsequently, the average distance between the 10th to 20th nearest neighbor projected points for each projected point is used as the initial neighborhood radius of that point, with a preferred number of nearest neighbors being 15. If the total number of projected points is less than 15, all projected points are included in the calculation. After calculating the neighborhood radii of all projected points, the distribution of all initial neighborhood radii is statistically analyzed. The range within ±10% of the median of all neighborhood radii is defined as the standard neighborhood radius interval. For projected points whose neighborhood radius is greater than the upper limit of the standard neighborhood radius interval, their neighborhood radius is gradually reduced, with each adjustment preferably by 5% of the original neighborhood radius. For projected points whose neighborhood radius is less than the lower limit of the standard neighborhood radius interval, their neighborhood radius is gradually increased by the same adjustment amount until the neighborhood radii of all projected points tend to stabilize, forming an adaptive neighborhood radius result. Then, starting from each projection point, all neighboring projection points within its neighborhood radius are searched. When the number of projection points in the neighborhood reaches the preset clustering threshold, the projection point is marked as the core projection point. Subsequently, starting from the core projection point, the interconnected neighboring projection points are gradually merged into the same cluster, and the process continues to expand outward until no new neighboring projection points satisfy the clustering conditions, ultimately forming several initial clusters.

[0057] Furthermore, the cluster center location, number of projected points within the cluster, average node confidence value, spatial coverage area, and boundary range corresponding to each initial cluster are statistically analyzed and used as cluster features. Then, any two initial clusters are compared. First, the spatial distance between the centers of the two clusters is calculated; then, the shortest distance between the boundaries of the two clusters is calculated; simultaneously, the difference between the average node confidence values ​​and the difference between the projection point distribution densities of the two clusters are compared. Based on these comparison results, an inter-cluster similarity is formed. When the distance between the centers of two clusters does not exceed 20% of the average center distance of all clusters, the shortest distance between their boundaries does not exceed 15% of the average boundary length, and the difference in average node confidence values ​​does not exceed 0.08, the two clusters are considered to have a high degree of similarity. Further, the above similarity values ​​are normalized to form an inter-cluster similarity score, which is then compared with a preset similarity threshold.

[0058] It should be noted that the preset similarity threshold is set to 0.80 to 0.90. This threshold can be obtained through statistical analysis of historical procurement agreement data. At least 1,500 procurement agreements are selected as samples to ensure that the clusters can maintain the continuity of the same semantic regions after merging, while avoiding the erroneous merging of different semantic regions.

[0059] When the inter-cluster similarity reaches or exceeds the preset similarity threshold, the two initial clusters are merged, and the center position, coverage area and average node confidence of the merged cluster are recalculated. Then, the similarity is compared with the remaining clusters, and the above merging process is repeated until the inter-cluster similarity between any two clusters is lower than the preset similarity threshold, and finally several merged clusters are obtained.

[0060] Furthermore, after obtaining the merged clusters, the spatial coordinates of all projected points within each merged cluster are first statistically analyzed, and the average values ​​of the x-coordinates and y-coordinates of all projected points are calculated. The resulting average coordinates are used as the cluster center point of the corresponding merged cluster. When the low-dimensional optimization control space is three-dimensional, the average value of the depth coordinates is further calculated to jointly determine the three-dimensional cluster center position. Subsequently, with the cluster center point as the center, the distances between all projected points within the cluster and the cluster center point are statistically analyzed. The maximum distance covered by 90% of the projected points is used as the radius of the corresponding confidence equipotential circle to avoid a small number of discrete projected points having an excessive impact on the equipotential circle range. If the number of projected points within a cluster is less than 20, the maximum distance among all projected points is directly used as the equipotential circle radius. After constructing all equipotential circles, it is determined whether there is an overlapping region between any two confidence equipotential circles. When there is a common coverage area between two equipotential circles, the number of projection points contained in the common coverage area is further counted. When the number of projection points in the common coverage area reaches more than 10% of the total number of projection points of the corresponding two clusters, the common coverage area is marked as a confidence cluster area. If the number of projection points in the common coverage area is less than the above proportion, only the independent coverage areas of the two equipotential circles are retained and not treated as confidence cluster areas. When three or more confidence equipotential circles overlap simultaneously, all common overlapping areas are merged into a continuous confidence cluster area, and the center position, coverage area, and boundary range of the cluster area are recalculated.

[0061] S5 divides the level intervals according to the confidence distribution within the probabilistic feasible region, and configures gradient and jump correction strategies to perform hierarchical correction on the optimization solution process, and outputs the procurement control scheme.

[0062] In this embodiment, within the probabilistic feasible region, the level intervals are divided according to the confidence distribution, and gradient and jump correction strategies are configured to perform hierarchical corrections on the optimization solution process, outputting a procurement control scheme, specifically as follows: Obtain the confidence distribution of the confidence profile and divide the confidence distribution of the confidence profile into several confidence level intervals according to the confidence value within the probability feasible region; A corresponding correction strategy is configured for each confidence level interval, including a gradient correction strategy and a jump correction strategy. The confidence level of the candidate solutions generated during the optimization process is determined, and the confidence level interval to which the candidate solutions belong is determined. Based on the confidence level range to which the candidate solution belongs, the corresponding correction strategy is invoked to dynamically adjust the search direction and search step size of the candidate solution, so as to obtain the dynamically adjusted candidate solution; The dynamically adjusted candidate solutions are then subjected to feasibility verification, and the candidate solutions that pass the verification are output as procurement control schemes.

[0063] It should be noted that, firstly, the node confidence values ​​corresponding to all graph vertices in the confidence graph are read, and then, according to the projection positions of the graph vertices in the low-dimensional optimization control space, the node confidence values ​​are mapped to the corresponding probabilistic feasible regions, forming a confidence distribution in the probabilistic feasible regions. Subsequently, all node confidence values ​​are sorted in ascending order, and the maximum, minimum, median, and quantile positions of each node confidence value are calculated. Finally, based on the distribution of node confidence values, several consecutive confidence level intervals are defined.

[0064] Preferably, the node confidence values ​​are divided into five level intervals: the first level interval corresponds to 0.90–1.00, the second level interval to 0.75–0.90, the third level interval to 0.60–0.75, the fourth level interval to 0.40–0.60, and the fifth level interval to 0.20–0.40. When the actual node confidence values ​​are relatively concentrated, the 20%, 40%, 60%, and 80% quantiles of all node confidence values ​​in the historical procurement agreement sample are used as the boundaries for each level interval to maintain consistency in the level division between different procurement tasks. After completing the level interval division, the number of graph vertices and spatial coverage corresponding to each level interval are counted sequentially. When the number of graph vertices in a certain level interval is less than 5% of the total number of graph vertices, it is merged with the adjacent level interval. When the number of graph vertices exceeds 40% of the total number of graph vertices, it is further subdivided into two consecutive sub-intervals according to the node confidence values ​​to ensure that each level interval has good distinguishability. Ultimately, each probabilistic feasible region corresponds to a series of confidence level intervals.

[0065] Furthermore, after obtaining the confidence level intervals, a correction strategy configuration table is first established according to the node confidence level corresponding to each level interval, and a corresponding search correction method is specified for each level interval. When the candidate solution is located in the first or second level interval, since the node confidence in the corresponding region is high, a gradient correction strategy is preferred, allowing the candidate solution to continuously search near the current region. When the candidate solution is located in the fourth or fifth level interval, since the node confidence in the corresponding region is low, a jump correction strategy is preferred, allowing the candidate solution to quickly leave the current region and re-search other high-confidence regions. When the candidate solution is located in the third level interval, the correction method is determined based on its distance from the boundary of the adjacent level interval. When it is close to the boundary of a higher level interval, a gradient correction strategy is used; when it is close to the boundary of a lower level interval, a jump correction strategy is used. Preferably, the distance determination range can be set to 20% to 30% of the width of the current level interval. After completing the correction strategy configuration, establish corresponding parameter configurations for the gradient correction strategy and the jump correction strategy respectively. The gradient correction strategy corresponds to a smaller search step size and continuous adjustment mode, while the jump correction strategy corresponds to a larger search step size and cross-region adjustment mode. Preferably, the step size adjustment ratio corresponding to the gradient correction strategy is set to 5% to 15% of the current search step size, and the step size adjustment ratio corresponding to the jump correction strategy is set to 30% to 60% of the current search step size.

[0066] It should be noted that during the optimization process, after each new candidate solution is generated, the control parameters corresponding to the candidate solution are first read, and the candidate solution is mapped to the corresponding coordinate position in the low-dimensional optimization control space according to the parameter set describing the probabilistic feasible region. Then, it is determined whether the coordinate position is inside the probabilistic feasible region. If the candidate solution is outside the probabilistic feasible region, it is directly determined to be in the lowest confidence level interval, and the subsequent jump correction process begins. If the candidate solution is inside the probabilistic feasible region, the node confidence values ​​corresponding to all graph vertices within a preset neighborhood around the candidate solution's position are further read, and the average node confidence value within the neighborhood is calculated as the comprehensive confidence of the candidate solution's current position. The neighborhood is preferably the area formed by the candidate solution's current position and the nearest 15 graph vertices; if the number of graph vertices in the neighborhood is less than 15, all graph vertices are included in the statistics. Next, the overall confidence score is compared with each of the pre-established confidence score intervals. When the overall confidence score falls within the corresponding interval, the confidence score interval to which the candidate solution belongs is determined. If the overall confidence score is exactly at the boundary between two intervals, it is preferentially assigned to the higher interval to improve the efficiency of subsequent optimization search. When three consecutive candidate solutions are determined to belong to the same interval, the current level determination result is maintained. When two consecutive candidate solutions cross more than two intervals, the neighborhood range of the candidate solution is recalculated, and the level determination is completed again.

[0067] Furthermore, after determining the confidence level range to which the candidate solution belongs, the correction strategy configured for the corresponding level range is first read, and the search direction and search step size in the next search process of the candidate solution are adjusted according to the correction strategy. When the gradient correction strategy is invoked, the trend of the change in the confidence values ​​of the nodes in the neighborhood around the current position of the candidate solution is first statistically analyzed, and the search direction is adjusted to the direction in which the node confidence gradually increases. At the same time, the search step size is reduced according to a preset ratio to keep the candidate solution in a continuous search state; preferably, the step size adjustment ratio is 10% of the current step size each time, and the number of consecutive adjustments does not exceed 5 times.

[0068] When the jump correction strategy is invoked, the system first searches for the nearest higher-confidence level interval to the current candidate solution. The center of this interval is then used as the target search area. The search direction is adjusted to the target area, and the search step size is increased so that the candidate solution crosses the current low-confidence region in one step. Preferably, the step size adjustment is set to 45% of the current step size each time, and the number of consecutive jumps does not exceed three. If a candidate solution enters the same low-confidence level interval twice consecutively, the current search direction is stopped, and the nearest higher-confidence level interval is selected as the new search target. After adjusting the search direction and step size, the new candidate solution position is recalculated, and the adjusted control parameters are output as the dynamically adjusted candidate solution, providing input data for subsequent feasibility verification.

[0069] After obtaining dynamically adjusted candidate solutions, a feasibility verification method is first used to perform logical consistency verification and confidence constraint verification on all control parameters in the candidate solutions. Then, it is statistically analyzed whether all control parameters meet the corresponding agreement clause constraints, and it is determined whether the candidate solution is still within the probabilistic feasible region. When all control parameters in a candidate solution pass the verification, and the corresponding node confidence is not lower than the preset verification threshold, the candidate solution is determined to meet the procurement agreement constraint requirements and is retained as a valid candidate solution. If any control parameter in a candidate solution fails the verification, the process returns to the candidate solution dynamic adjustment step, and the corresponding correction strategy is called again to continue optimizing the search until a new candidate solution is generated. For multiple valid candidate solutions that simultaneously meet the verification requirements, the comprehensive node confidence of each candidate solution is further compared, and the candidate solution with the highest comprehensive node confidence is selected as the final procurement control scheme. When the comprehensive node confidence is the same, the candidate solution with fewer adjustments is selected. If the number of adjustments is still the same, the candidate solution with the earliest initial generation time is selected as the final output result. After completing the above screening, the final procurement control plan will be output, and the corresponding node confidence distribution, level range and correction process records will be retained for traceability and verification in the subsequent procurement execution process.

[0070] In this embodiment, the feasibility of the dynamically adjusted candidate solutions is verified, specifically as follows: Extract the control parameter sequence from the dynamically adjusted candidate solution, and obtain the node confidence value of the graph vertex to which each control parameter in the confidence spectrum belongs; Control parameters whose node confidence values ​​are lower than a preset verification threshold are marked as parameters to be verified, and the associated parameter groups of the parameters to be verified in the candidate solutions are extracted. The associated parameter groups include upstream and downstream parameters that have a logical transit relationship with the parameters to be verified. Perform feasibility verification on the parameters to be verified and their associated parameter groups to obtain candidate solutions that pass the feasibility verification.

[0071] It should be noted that after obtaining the dynamically adjusted candidate solutions, a control parameter sequence is first established according to the generation order of each control parameter in the candidate solutions. Then, the parameter number, parameter type, and parameter value corresponding to each control parameter are read sequentially. Simultaneously, based on the mapping relationship saved during the candidate solution generation process, the protocol clause number and logical node number corresponding to each control parameter are determined. Subsequently, using the logical node number as an index, the corresponding graph vertex is retrieved one by one in the previously constructed confidence graph, and the node confidence value corresponding to that vertex is read, ensuring that each control parameter is associated with a unique node confidence value. When a control parameter corresponds to multiple graph vertices, the node confidence values ​​of each graph vertex are counted separately and sorted from low to high according to the logical propagation level. The node confidence value corresponding to the graph vertex with the highest propagation level is selected as the node confidence value for that control parameter. If multiple graph vertices have the same propagation level, the node confidence values ​​corresponding to each graph vertex are further compared, and the node with the lower confidence value is selected first as the node confidence value corresponding to that control parameter, ensuring that subsequent feasibility verification can be judged according to the strictest protocol constraints.

[0072] After completing the matching of all control parameters, check in turn whether each control parameter has obtained the corresponding node confidence value. For control parameters that have not established a corresponding relationship, continue searching along the adjacent propagation nodes in the confidence graph. When the search level does not exceed two levels, use the node confidence value corresponding to the nearest propagation node as the substitute value. If no corresponding relationship is established after more than two levels, use the initial confidence value corresponding to the generation of the control parameter as the node confidence value.

[0073] Furthermore, after obtaining the node confidence values ​​corresponding to all control parameters, the nodes are first sorted in ascending order of confidence value, and each node confidence value is compared with a preset verification threshold. When a node confidence value is lower than the preset verification threshold, the corresponding control parameter is marked as a parameter to be verified. After determining the parameter to be verified, the graph vertex corresponding to the parameter to be verified is used as the starting node. In the confidence graph, all upstream graph vertices are searched forward along the transmission direction, and all downstream graph vertices are searched backward along the transmission direction. The control parameters associated with the corresponding graph vertices are extracted and together form an associated parameter group. The upstream search layer is preferably no more than 3 layers, and the downstream search layer is also preferably no more than 3 layers. When multiple branches appear during the search in a certain direction, all of them are included in the associated parameter group. If the search reaches the starting graph vertex or the ending graph vertex, the search in the corresponding direction is stopped. For cases where multiple parameters to be verified have common associated parameters, only one set of common associated parameters is retained, and a corresponding shared identifier is established so that the same control parameter does not need to participate in subsequent verifications repeatedly, thereby reducing the amount of redundant calculation.

[0074] It should be noted that the preset verification threshold is set to 0.70–0.80. Specifically, it can be obtained through statistical analysis of historical procurement agreement execution results. At least 1500 procurement agreements that have been completed and have actual performance results are selected as statistical samples. The pass rate of the procurement plan corresponding to different node confidence values ​​is calculated. When the lowest node confidence value corresponding to the procurement plan pass rate reaches 90% or higher, it is determined as the preset verification threshold.

[0075] Furthermore, based on the sequence of logical constraints, the logical relationships between the parameter to be verified and its associated parameters are checked sequentially according to the logical transmission order between the control parameters. When there is a conditional connection between the parameter to be verified and the upstream associated parameter, the upstream control parameter is first verified to ensure it meets the corresponding condition. If the condition is not met, the parameter to be verified is deemed to have failed the feasibility check, and the verification of the current parameter is terminated. When there is a conjunctive connection between the parameter to be verified and multiple associated parameters, the control parameters corresponding to all associated parameters are checked one by one to ensure they meet the protocol constraints. Only when all associated parameters meet the constraints is the parameter to be verified deemed to meet the feasibility requirements. When there is a disjunctive connection, all associated parameters are checked sequentially. As long as any one associated parameter meets the protocol constraints, the parameter to be verified is deemed to meet the corresponding logical relationship.

[0076] After completing the logical relationship check, the node confidence values ​​corresponding to the parameter to be verified and all related parameters are statistically analyzed, and the lowest node confidence value is calculated. If the lowest node confidence value is still not lower than the preset verification threshold, the entire related parameter group is determined to have passed the feasibility verification. If the lowest node confidence value is lower than the preset verification threshold, the corresponding parameter to be verified is returned to the candidate solution dynamic adjustment step for further optimization. After all parameters to be verified have completed the above processing, the entire control parameter sequence is checked again to see if there are still any parameters that have not passed the feasibility verification. If not, the current candidate solution is determined to have passed the feasibility verification and is output as the final candidate solution. If there are still control parameters that have not passed the verification, the candidate solution correction and re-verification are continued until all control parameters meet the logical constraints and node confidence requirements. Finally, the candidate solution that has passed the overall feasibility verification is output as the procurement control scheme, and the corresponding verification results and related parameter verification records are retained.

[0077] Example 2, Figure 2 This invention presents a semantic matching control optimization system based on procurement agreement constraints, comprising a confidence extraction module, a constraint construction module, a path generation module, a graph construction module, a graph mapping module, and a procurement control module, as well as the data flow relationships between the modules: The confidence extraction module is used to obtain the semantic matching results between the agreement terms text and the response text, and to extract the initial confidence level corresponding to the semantic matching results; The constraint construction module is used to parse the agreement clause text into a logical expression structure, and extract the connection type between the agreement clauses from the logical expression structure to obtain the logical constraint conditions. The path generation module is used to map logical constraints to propagation nodes, obtain an acyclic propagation graph, and traverse the acyclic propagation graph to obtain a confidence-corrected propagation path. The graph construction module performs hop-by-hop correction and backtracking on the initial confidence level based on the confidence level correction propagation path to generate a confidence level graph. The graph mapping module is used to extract the node confidence values ​​of the graph vertices in the confidence graph and combine them into a high-dimensional confidence vector to obtain the semantic matching space. The UMAP algorithm is then used to project the vector onto the low-dimensional optimization control space to obtain the probabilistic feasible region for optimization. The procurement control module is used to divide the level intervals according to the confidence distribution within the probabilistic feasible region, configure gradient and jump correction strategies to perform hierarchical correction on the optimization solution process, and output the procurement control scheme.

[0078] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0079] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0080] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0081] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semantic matching control optimization method based on procurement agreement constraints, characterized in that, Includes the following steps: Obtain the semantic matching results between the agreement terms text and the response text, and extract the initial confidence scores corresponding to the semantic matching results; The agreement terms are parsed into a logical expression structure, and the connection types between the agreement terms are extracted from the logical expression structure to obtain the logical constraints. Logical constraints are mapped to propagation nodes to obtain an acyclic propagation graph. The acyclic propagation graph is then traversed to obtain the confidence-corrected propagation path. Based on the confidence correction propagation path, the initial confidence is corrected and propagated hop by hop and backtracked to generate a confidence graph; The node confidence values ​​of the graph vertices in the confidence graph are extracted and combined into a high-dimensional confidence vector to obtain the semantic matching space. The UMAP algorithm is then used to project the vector onto a low-dimensional optimization control space to obtain the probabilistic feasible region for optimization solution. Within the probabilistic feasible region, the level intervals are divided according to the confidence distribution, and gradient and jump correction strategies are configured to perform hierarchical corrections on the optimization solution process, outputting a procurement control scheme.

2. The semantic matching control optimization method based on procurement agreement constraints according to claim 1, characterized in that, The process of obtaining the semantic matching results between the agreement terms text and the response text, and extracting the initial confidence level corresponding to the semantic matching results, specifically involves: The agreement text is segmented into semantic fragment sets according to semantic units, and the response text is segmented into response semantic fragment sets according to sentence boundaries; Bidirectional alignment is performed on the set of semantic fragments of terms and the set of semantic fragments of responses to obtain a fragment-level semantic matching matrix; Extract semantic matching pairs along the main diagonal of the semantic matching matrix and calculate the semantic distance between the semantic matching pairs; Map the semantic distance of the semantic matching pairs to the initial confidence level.

3. The semantic matching control optimization method based on procurement agreement constraints according to claim 2, characterized in that, The process of parsing the agreement clause text into a logical expression structure and extracting the connection types between the agreement clauses from the logical expression structure to obtain logical constraints is as follows: Obtain the semantic reference relationships of the agreement clause text, and perform dependency parsing on the agreement clause text to obtain several logical units; Obtain the coreference and temporal relationships between several logical units, and construct the logical relationship edges within the clauses based on the coreference and temporal relationships; By concatenating the logical relationship edges within the clauses according to the semantic reference relationships between clauses, a cross-clause logical expression diagram is obtained; Logical type labeling is performed on nodes in the cross-clause logical expression diagram, and connection types are identified, including conjunction connection, disjunction connection and conditional connection; The annotated logic representation diagram is transformed into a sequence of logic constraints, and the sequence of logic constraints is arranged in topological order to obtain the logic constraints.

4. The semantic matching control optimization method based on procurement agreement constraints according to claim 3, characterized in that, The process of mapping logical constraints to propagation nodes to obtain an acyclic propagation graph, and then traversing the acyclic propagation graph to obtain a confidence-corrected propagation path, specifically involves: Each logical node in the logical constraints is mapped to a propagation node in the confidence propagation graph, and the propagation direction between propagation nodes is determined according to the connection type in the logical constraints to obtain the initial propagation edge; Redundancy detection is performed on the initial propagation edges, and propagation edges that form closed loops are deleted to obtain an acyclic propagation graph; In the acyclic propagation graph, traverse all reachable paths along the propagation direction to obtain confidence-corrected propagation paths, and calculate the path length and logical operation type of each path in the confidence-corrected propagation path.

5. The semantic matching control optimization method based on procurement agreement constraints according to claim 4, characterized in that, The process of performing hop-by-hop correction and backtracking on the initial confidence level based on the confidence level correction propagation path to generate a confidence level map is as follows: The shortest path is selected from the set of confidence correction propagation paths as the baseline propagation path, and the initial confidence is propagated hop by hop along the baseline propagation path according to the logical operation type on the path to obtain the propagation confidence of each propagation node. The path backtracking method is used to combine the transmission confidence on the non-reference transmission path with the transmission confidence on the reference transmission path to obtain the combined confidence of each propagation node. The convergence confidence of each propagation node is normalized to generate a node confidence value; A confidence graph is constructed using propagation nodes as vertices and propagation directions and merging relationships as edges.

6. The semantic matching control optimization method based on procurement agreement constraints according to claim 5, characterized in that, The extracted confidence values ​​of the nodes in the confidence graph are combined to form a high-dimensional confidence vector, resulting in a semantic matching space. This space is then projected onto a low-dimensional optimization control space using the UMAP algorithm to obtain the probabilistic feasible region for optimization. Specifically: Extract the node confidence value of each vertex in the confidence graph, and combine the node confidence values ​​into a high-dimensional confidence vector to obtain the semantic matching space; The UMAP algorithm is used to project the high-dimensional confidence vector from the semantic matching space to the low-dimensional optimization control space to obtain the initial projection point set. Density clustering of the initial projected point set is performed in the low-dimensional optimization control space to obtain confidence clustering regions. Extract the boundary points of the confidence cluster region and construct the convex hull with the boundary points as vertices to obtain the geometric boundary of the probabilistic feasible region; Parameterize the geometric boundary of the probabilistic feasible region to generate a set of descriptive parameters for the probabilistic feasible region.

7. The semantic matching control optimization method based on procurement agreement constraints according to claim 6, characterized in that, The step of performing density clustering on the initial projection point set in the low-dimensional optimization control space to obtain confidence clustering regions specifically involves: In the low-dimensional optimization control space, density clustering with adaptive neighborhood radius is performed on the initial projection point set to obtain the initial clusters; The initial clusters are calculated for inter-cluster similarity. The initial clusters with inter-cluster similarity exceeding a preset threshold are merged to obtain merged clusters. Extract the cluster center points of the merged clusters, construct confidence equipotential circles with the cluster center points as the center, and mark the overlapping areas of the confidence equipotential circles as confidence clustering areas.

8. The semantic matching control optimization method based on procurement agreement constraints according to claim 7, characterized in that, Within the probabilistic feasible region, the system divides the optimization process into graded intervals based on confidence distribution and configures gradient and jump correction strategies to perform graded corrections, outputting a procurement control scheme. Specifically: Obtain the confidence distribution of the confidence profile and divide the confidence distribution of the confidence profile into several confidence level intervals according to the confidence value within the probability feasible region; A corresponding correction strategy is configured for each confidence level interval, including a gradient correction strategy and a jump correction strategy. The confidence level of the candidate solutions generated during the optimization process is determined, and the confidence level interval to which the candidate solutions belong is determined. Based on the confidence level range to which the candidate solution belongs, the corresponding correction strategy is invoked to dynamically adjust the search direction and search step size of the candidate solution, so as to obtain the dynamically adjusted candidate solution; The dynamically adjusted candidate solutions are then subjected to feasibility verification, and the candidate solutions that pass the verification are output as procurement control schemes.

9. The semantic matching control optimization method based on procurement agreement constraints according to claim 8, characterized in that, The feasibility verification of the dynamically adjusted candidate solutions is specifically as follows: Extract the control parameter sequence from the dynamically adjusted candidate solution, and obtain the node confidence value of the graph vertex to which each control parameter in the confidence spectrum belongs; Control parameters whose node confidence values ​​are lower than a preset verification threshold are marked as parameters to be verified, and the associated parameter groups of the parameters to be verified in the candidate solutions are extracted. The associated parameter groups include upstream and downstream parameters that have a logical transit relationship with the parameters to be verified. Perform feasibility verification on the parameters to be verified and their associated parameter groups to obtain candidate solutions that pass the feasibility verification.

10. A semantic matching control optimization system based on procurement agreement constraints, applied to the semantic matching control optimization method based on procurement agreement constraints as described in any one of claims 1-9, characterized in that, It includes a confidence extraction module, a constraint construction module, a path generation module, a graph construction module, a graph mapping module, and a procurement control module. The confidence extraction module is used to obtain the semantic matching results between the agreement terms text and the response text, and to extract the initial confidence level corresponding to the semantic matching results; The constraint construction module is used to parse the agreement clause text into a logical expression structure, and extract the connection type between the agreement clauses from the logical expression structure to obtain the logical constraint conditions. The path generation module is used to map logical constraints to propagation nodes, obtain an acyclic propagation graph, and traverse the acyclic propagation graph to obtain a confidence-corrected propagation path. The graph construction module performs hop-by-hop correction and backtracking on the initial confidence level based on the confidence level correction propagation path to generate a confidence level graph. The graph mapping module is used to extract the node confidence values ​​of the graph vertices in the confidence graph and combine them into a high-dimensional confidence vector to obtain the semantic matching space. The UMAP algorithm is then used to project the vector onto the low-dimensional optimization control space to obtain the probabilistic feasible region for optimization. The procurement control module is used to divide the level intervals according to the confidence distribution within the probabilistic feasible region, configure gradient and jump correction strategies to perform hierarchical correction on the optimization solution process, and output the procurement control scheme.

Images