A method for dynamically evaluating procurement document risks based on semantic understanding

Abstract

The application discloses a kind of based on semantic understanding's procurement document risk dynamic assessment method, comprising the following steps: generating structured procurement document data;Clause semantic boundary identification processing is carried out, and clause fragment set is generated;Perform deep semantic analysis processing, generate procurement semantic element set in combination with procurement domain ontology model;Carry out context-aware vectorization representation processing, generate semantic vector representation set, and construct semantic index set;Build procurement risk knowledge base, carry out semantic bridging and difference quantification, generate risk feature association set;Risk assessment result set is generated using improved BOCPD model, introducing semantic vortex capture mechanism;The updated risk assessment result set is generated by obtaining the update data or new clause data of procurement document data, and the application improves the identification accuracy of implicit risk and cross-clause risk in procurement document.

Description

Technical Field

[0001] This invention relates to the field of procurement risk assessment, and in particular to a method for dynamic risk assessment of procurement documents based on semantic understanding. Background Technology

[0002] As procurement processes become increasingly digital, procurement documents have gradually shifted from traditional paper documents to electronic documents, and are compiled, circulated, reviewed, and archived within procurement management systems, e-procurement platforms, and business collaboration systems. Procurement documents typically include eligibility criteria, technical requirements, commercial terms, contract terms, performance responsibilities, review rules, and risk warnings. These documents exhibit clear contextual dependencies and cross-clause connections. Risks in procurement documents are not always explicitly stated in risk terms; some risks may be hidden within the scope of conditions, responsible parties, negative statements, exceptions, time requirements, and cross-clause references. Therefore, relying solely on manual, line-by-line review of procurement documents is susceptible to the influence of reviewers' experience, business familiarity, and review time constraints. This makes it difficult to ensure consistency in review standards across different procurement documents and to meet the dynamic risk assessment needs arising from frequent document updates.

[0003] Existing risk review techniques for procurement documents primarily employ a combination of keyword matching, rule comparison, template verification, and manual review. Keyword matching typically involves pre-setting risk terms, prohibited terms, or sensitive terms, and then searching the procurement document for text content that matches or is similar to these preset terms. Rule comparison usually compares the amount, time, proportion, and eligibility requirements in the procurement document against pre-set review rules item by item. Template verification typically compares the procurement document under review against a standard procurement document template at the paragraph or field level. While these methods can identify some explicit risks, the same risk type may be expressed in different sentence structures and with different words in the procurement document. Matching only keywords or fixed rules can easily miss clauses with different semantic expressions but the same underlying risk. Furthermore, existing techniques often segment text by paragraph, sentence, or fixed length, failing to adequately consider heading levels, clause numbers, semantic continuity, and contextual relationships. This can easily lead to incorrectly splitting complete clauses or merging multiple independent clauses, resulting in a lack of accurate clause fragments for subsequent risk identification. Summary of the Invention

[0004] One objective of this invention is to propose a dynamic risk assessment method for procurement documents based on semantic understanding. This invention fully utilizes clause semantic boundary identification, context-aware vectorized representation, procurement risk knowledge base, and improved BOCPD model, and has the advantages of accurate identification of implicit risks, strong cross-clause risk analysis capabilities, traceable risk results, and dynamic recalculation after procurement documents are updated.

[0005] A method for dynamic risk assessment of procurement documents based on semantic understanding according to an embodiment of the present invention includes the following steps: Acquire procurement document data and preprocess it to generate structured procurement document data; Perform semantic boundary recognition processing on structured procurement document data to generate a set of clause fragments; Deep semantic parsing is performed on the set of clause fragments, and implicit risk triggering elements in the clause fragments are extracted by combining the procurement domain ontology model to generate a set of procurement semantic elements. The semantic element set of procurement is processed with context-aware vectorization to generate a set of semantic vector representations and to construct a set of semantic indexes. Construct a procurement risk knowledge base, and based on the semantic index set, perform semantic bridging and difference quantification on the procurement semantic element set and the procurement risk knowledge base to generate a risk feature association set; The risk feature association set is input into the improved BOCPD model, a semantic vortex capture mechanism is introduced, the risk rule deviation point is identified and the risk evidence chain is constructed to generate a risk assessment result set. The system acquires updated or newly added clause data from procurement documents, performs incremental semantic parsing on the updated or newly added clause data, updates the procurement semantic element set and risk feature association set, triggers the multidimensional risk assessment model to perform incremental recalculation, and generates an updated risk assessment result set.

[0006] Optionally, procurement document data, including text files, scanned files, image files, and table files, is obtained from the procurement management system. The procurement document data is then processed to identify the file type. Text files are extracted for text content. Scanned files and image files are sequentially processed for image enhancement, noise removal, skew correction, and optical character recognition. Table files are processed for table boundary recognition, cell positioning, and row and column relationship restoration. According to the page position, heading level, paragraph order, and table correspondence in the procurement document data, the identified text and table content are parsed for layout. Abnormal character removal, character encoding standardization, punctuation standardization, duplicate content removal, and field format standardization are then performed. The processed text and table content are then structured and organized according to the original page order and content relationships to generate structured procurement document data.

[0007] Optionally, the generation of the set of clause fragments specifically includes: The structured procurement document data is divided into text units, and a sequence of text units is generated; Semantic boundary features are extracted from adjacent text units in a text unit sequence to generate a semantic boundary feature sequence. Based on the semantic boundary feature sequence, the semantic continuity between adjacent text units is determined, and the position where the semantic continuity is interrupted is determined as the semantic boundary of the candidate clause, generating a set of semantic boundaries of the candidate clause. The semantic boundary set of candidate clauses is subjected to duplicate boundary elimination, adjacent boundary merging and nested boundary correction to generate a semantic boundary set of clauses. The text unit sequence is segmented based on the semantic boundary set of the clauses to generate a set of clause fragments.

[0008] Optionally, the generation of the procurement semantic element set specifically includes: A set of clause fragments is subjected to joint syntactic and semantic parsing to generate a set of clause semantic relations. Map the set of semantic relationships of terms to the procurement domain ontology model to generate an ontology semantic relationship set; Perform contextual semantic completion and cross-clause relationship association on the ontology semantic relationship set to generate a complete ontology semantic relationship set; Based on the risk triggering path in the procurement domain ontology model, risk triggering reasoning is performed on the complete ontology semantic relationship set to extract implicit risk triggering elements and generate a set of implicit risk triggering elements. The set of implicit risk triggering factors is recorded in a structured manner to generate a set of procurement semantic elements.

[0009] Optionally, the construction of the semantic index set specifically includes: Based on the set of procurement semantic elements, determine the contextual relationships between each procurement semantic element and generate a set of contextual relationships; Based on the set of contextual relationships, the interrelated procurement semantic elements are arranged to generate a sequence of contextual semantic elements; Dual-path vector encoding is performed on the sequence of contextual semantic elements to generate a set of clause context vectors and a set of risk path context vectors. The association weight calibration and gating fusion of the clause context vector set and the risk path context vector set are performed to generate a semantic vector representation set. A vector index is created based on the set of semantic vector representations, and the procurement semantic elements and cross-term relationships corresponding to each semantic vector representation are written into the vector index to generate a set of semantic indexes.

[0010] Optionally, the generation of the risk feature association set specifically includes: Acquire procurement risk rules data, procurement standard terms data, and historical procurement risk data, and extract semantic elements to generate a set of risk knowledge units; Based on the risk knowledge unit set, establish the relationship between each risk knowledge unit to generate a procurement risk knowledge base; Based on the semantic vector representations in the semantic index set, risk knowledge units associated with the corresponding procurement semantic elements are retrieved from the procurement risk knowledge base to generate a candidate risk knowledge unit set. Semantically bridge the procurement semantic elements with the candidate risk knowledge unit set to generate a set of semantic bridging relationships; Based on the set of semantic bridging relationships, the differences between procurement semantic elements and corresponding risk knowledge units are quantified to generate a set of risk feature associations.

[0011] Optionally, the generation of the risk assessment result set specifically includes: The improved BOCPD model inputs a risk feature association set into the model. The improved BOCPD model includes a cross-clause risk sequence construction module, a semantic run state inference module, a risk rule deviation change point detection module, and a risk evidence chain aggregation module. The improvement of the improved BOCPD model is as follows: The traditional BOCPD model recursively infers the posterior probability of run length according to the time sequence of the observed data and identifies the data distribution change point based on the reset state of the run length posterior probability. The improved BOCPD model folds and arranges the risk feature association data according to the clause sequence axis and the risk path axis, uses the semantic vortex capture mechanism to identify the round-trip semantic path of the same risk type in the risk trigger path, corrects the semantic run state posterior probability according to the semantic vortex unit, and identifies the risk rule deviation change point with the corrected reset state, generating a risk evidence chain and a risk assessment result set. In the cross-clause risk sequence construction module, the risk feature association data in the risk feature association set is folded and arranged in a dual axis to generate risk observations. The risk observations corresponding to the same risk type and target are aligned to the same risk path position to generate a cross-clause risk observation sequence. In the semantic run state inference module, a semantic vortex capture mechanism is introduced to identify the closed trajectory of risk observations of the same risk type that recur along the risk triggering path in the cross-clause risk observation sequence. The first risk observation in the closed trajectory is determined as the vortex start point, and the last risk observation in the closed trajectory is determined as the vortex end point, generating a set of semantic vortex units. In the risk rule deviation change point detection module, the semantic run state posterior probability corresponding to each clause position is recursively deduced based on the cross-clause risk observation sequence, and the semantic run state posterior probability corresponding to each clause position is corrected by combining the semantic vortex unit set. The clause position where the corrected semantic run state posterior probability switches from the continuation state to the reset state is determined as the risk rule deviation change point, and a risk rule deviation change point set is generated. In the risk evidence chain aggregation module, a risk evidence chain is generated based on the set of risk rule deviation change points. The location and risk level of the risk clause are determined according to the risk type, clause fragment identifier, risk rule deviation change points and semantic vortex units in the risk evidence chain, and a risk assessment result set is generated.

[0012] Optionally, version comparison is performed on the procurement document data. Clause fragments with changed content are identified as updated data, and newly added clause fragments are identified as new clause data. Deep semantic parsing is performed on the updated or new clause data to extract the corresponding implicit risk triggering elements and generate incremental procurement semantic elements. The incremental procurement semantic elements corresponding to the new clause data are written into the procurement semantic element set. Context-aware vectorization representation processing is performed on the incremental procurement semantic elements to update the semantic vector representation set and semantic index set. Based on the updated semantic index set, semantic bridging and difference measurement are re-executed to update the corresponding clause fragment identifier and risk feature association data within the risk triggering path range, resulting in an updated risk feature association set. The updated risk feature association set is input into the improved BOCPD model. Using the clause fragment identifier corresponding to the updated or new clause data as the starting point for incremental recalculation, the cross-clause risk observation sequence, semantic vortex unit set, risk rule deviation change point set, and risk evidence chain within the corresponding risk triggering path range are regenerated. The risk assessment results corresponding to the clause fragment identifier in the risk assessment result set are replaced with the regenerated risk assessment results to generate an updated risk assessment result set.

[0013] The beneficial effects of this invention are: This invention, through structured processing of procurement document data and semantic boundary identification of clauses, can form a set of clause fragments according to the clause hierarchy, semantic continuity, and contextual relationships in the procurement documents, reducing the semantic fragmentation problem caused by fixed-length segmentation or simple paragraph segmentation. Combined with a procurement domain ontology model, deep semantic analysis of the clause fragments can extract implicit risk triggering elements corresponding to risk types, triggering conditions, target objects, cross-clause relationships, and risk triggering paths. Furthermore, through context-aware vectorized representation processing, a set of semantic vector representations and a set of semantic indexes are constructed, enabling procurement semantic elements scattered across different clause fragments to be expressed in relation to the clause context and risk triggering paths. This improves the completeness and accuracy of identifying implicit risks, cross-clause risks, and semantically variant risks in procurement documents.

[0014] This invention constructs a procurement risk knowledge base, semantically bridging and quantifying the differences between procurement semantic elements and risk knowledge units. This clearly records differences in risk types, triggering conditions, affected objects, cross-clause relationships, and risk triggering paths, providing clear data sources and evidence for risk assessment results. Furthermore, it utilizes an improved BOCPD model to perform dual-axis folding arrangement, semantic vortex capture, and risk rule deviation point detection on the risk feature association set. Based on these deviation points, it generates a risk evidence chain and a risk assessment result set, determining the location of risk clauses, risk severity values, and risk levels, thus improving the interpretability and traceability of risk assessment results. Simultaneously, when procurement document data is updated or new clauses are added, incremental recalculation is performed only on the affected procurement semantic elements, risk feature association data, and risk assessment results, reducing redundant processing and improving the processing efficiency and update response capability of dynamic risk assessment in procurement documents. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a dynamic risk assessment method for procurement documents based on semantic understanding proposed in this invention; Figure 2 This is a schematic diagram of the improved BOCPD model, which is a dynamic risk assessment method for procurement documents based on semantic understanding proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 A dynamic risk assessment method for procurement documents based on semantic understanding includes the following steps: Acquire procurement document data and preprocess it to generate structured procurement document data; Perform semantic boundary recognition processing on structured procurement document data to generate a set of clause fragments; Deep semantic parsing is performed on the set of clause fragments, and implicit risk triggering elements in the clause fragments are extracted by combining the procurement domain ontology model to generate a set of procurement semantic elements. The semantic element set of procurement is processed with context-aware vectorization to generate a set of semantic vector representations and to construct a set of semantic indexes. Construct a procurement risk knowledge base, and based on the semantic index set, perform semantic bridging and difference quantification on the procurement semantic element set and the procurement risk knowledge base to generate a risk feature association set; The risk feature association set is input into the improved BOCPD model, a semantic vortex capture mechanism is introduced, the risk rule deviation point is identified and the risk evidence chain is constructed to generate a risk assessment result set. The system acquires updated or newly added clause data from procurement documents, performs incremental semantic parsing on the updated or newly added clause data, updates the procurement semantic element set and risk feature association set, triggers the multidimensional risk assessment model to perform incremental recalculation, and generates an updated risk assessment result set.

[0018] In this embodiment, procurement document data, including text files, scanned files, image files, and table files, is obtained from the procurement management system. The procurement document data undergoes file type identification, text content extraction for text files, image enhancement, noise removal, skew correction, and optical character recognition for scanned and image files, and table boundary identification, cell positioning, and row / column relationship restoration for table files. Based on the page position, heading level, paragraph order, and table correspondence in the procurement document data, the identified text and table content are parsed for layout, and abnormal character removal, character encoding standardization, punctuation standardization, duplicate content removal, and field format standardization are performed. The processed text and table content are then structured according to the original page order and content relationships to generate structured procurement document data.

[0019] In this embodiment, the generation of the clause fragment set specifically includes: The structured procurement document data is divided into text units, and a sequence of text units is generated; The process of generating a text unit sequence is as follows: Read the page order, heading level, clause number, paragraph mark, and table position from the structured procurement document data; determine the start and end positions for the heading content, clause content, paragraph content, and table content in the structured procurement document data; divide the continuous content between the start and end positions into a text unit; and record the corresponding page number, heading level, clause number, paragraph number, and table position for each text unit; sort the text units according to the following order: page number from smallest to largest, vertical position within the same page from top to bottom, and horizontal position within the same vertical position from left to right; and assign consecutive text unit numbers to each text unit based on the sorting results to generate a text unit sequence. Semantic boundary features are extracted from adjacent text units in a text unit sequence to generate a semantic boundary feature sequence. The process of generating semantic boundary feature sequences is as follows: Adjacent text units in the text unit sequence are read sequentially, and the following features are extracted from each adjacent text unit: title level, clause number, paragraph marker, first sentence content, last sentence content, keyword distribution, syntactic dependency relationship, and referent. Title level change features are determined based on whether the title level changes; clause number change features are determined based on whether the clause numbers are consecutive; thematic semantic change features are determined based on the difference in keyword distribution between adjacent text units; syntactic closure features are determined based on whether the syntactic structure of the last sentence of the preceding text unit is complete; and referential connection features are determined based on whether pronouns, abbreviations, and omitted components in the following text unit can be identified as corresponding referents in the preceding text unit. Following the adjacent position order in the text unit sequence, the title level change features, clause number change features, thematic semantic change features, syntactic closure features, and referential connection features corresponding to each group of adjacent text units are recorded as a semantic boundary feature item, generating a semantic boundary feature sequence. Based on the semantic boundary feature sequence, the semantic continuity between adjacent text units is determined, and the position where the semantic continuity is interrupted is determined as the semantic boundary of the candidate clause, generating a set of semantic boundaries of the candidate clause. The semantic boundary set of candidate clauses is subjected to duplicate boundary elimination, adjacent boundary merging and nested boundary correction to generate a semantic boundary set of clauses. The specific process for generating the semantic boundary set of clauses is as follows: Obtain the semantic boundaries of each candidate clause in the candidate clause semantic boundary set, and sort them according to the order in the text unit sequence; compare the sorted candidate clause semantic boundaries one by one. When adjacent candidate clause semantic boundaries correspond to the same text unit position, or the difference between their positions in the text unit sequence does not exceed one text unit, it is determined that there is a duplicate boundary position, and the duplicate candidate clause semantic boundaries are merged into one merged boundary; calculate the text unit span between adjacent boundaries in each merged boundary. When the text unit span between adjacent boundaries is less than 3 text units, and the text unit span between adjacent boundaries is less than 3 text units, the text unit span between adjacent boundaries is determined. When the semantic similarity of the clauses is greater than or equal to 0.85, the adjacent boundaries are merged to generate a merged clause semantic boundary. The nested boundaries in the merged clause semantic boundary set are included. When any clause semantic boundary covers all text units corresponding to another clause semantic boundary, the boundary integrity score of the text content corresponding to the two clause semantic boundaries is calculated. When the boundary integrity score of the included boundary is less than 0.70, the included boundary is deleted. When the boundary integrity score of the included boundary is greater than or equal to 0.70 and higher than the boundary integrity score of the included boundary, the included boundary is retained and the included boundary is deleted to generate a clause semantic boundary set. The text unit sequence is segmented based on the semantic boundary set of the clauses to generate a set of clause fragments; The process of generating the clause fragment set is as follows: based on the position information of each clause semantic boundary in the clause semantic boundary set in the text unit sequence, the text unit sequence is traversed and scanned, and the start and end positions between adjacent clause semantic boundaries are used as segmentation intervals; within each segmentation interval, the corresponding text units are extracted according to the order of the text unit sequence, while maintaining the original arrangement relationship and semantic context relationship of the text units, and the text units in the same segmentation interval are sequentially spliced ​​to form a complete clause content unit; the clause content units generated by each segmentation interval are organized according to the order in the clause semantic boundary set to generate the clause fragment set.

[0020] In this embodiment, the generation of the procurement semantic element set specifically includes: A set of clause fragments is subjected to joint syntactic and semantic parsing to generate a set of clause semantic relations. The specific process for generating the semantic relationship set of clauses is as follows: Each clause fragment in the clause fragment set is read sequentially; syntactic and semantic joint parsing is performed on each clause fragment to identify semantic entities within the clause fragments; numerical, temporal, range, state, condition, responsibility, and negation information with syntactic modification or semantic limitation relationships with the semantic entities is extracted, and the above information is determined as the semantic attributes of the corresponding semantic entities based on the dependency direction and semantic scope of the clause fragments; semantic relationships between semantic entities are determined according to the predicate structure, subject-predicate relationship, verb-object relationship, conditional relationship, negation relationship, and referential relationship in the clause fragments; according to the order of each clause fragment in the clause fragment set, each semantic entity, each semantic attribute, each semantic relationship, and the corresponding clause fragment identifier are recorded to generate the semantic relationship set of clauses. Map the set of semantic relationships of terms to the procurement domain ontology model to generate an ontology semantic relationship set; The generation process of the ontology semantic relationship set is as follows: The procurement domain ontology model includes an ontology concept layer, an ontology attribute layer, an ontology relationship layer, and a risk triggering path layer. The ontology concept layer records the procurement entity, procurement object, qualification conditions, technical requirements, business conditions, performance responsibilities, and risk types. The ontology attribute layer records the numerical, time, scope, status, condition, and responsibility attributes corresponding to each ontology concept. The ontology relationship layer records the subject behavior relationships, condition constraint relationships, rights and obligations relationships, negation relationships, and exception relationships between each ontology concept. The risk triggering path layer records the risk triggering paths composed of ontology concepts, ontology attributes, and ontology relationships. The semantic entities, semantic attributes, and semantic relationships in the clause semantic relationship set are read. Each semantic entity is semantically matched with the ontology concept layer, each semantic attribute is field-matched with the ontology attribute layer, and each semantic relationship is relation-matched with the ontology relationship layer. The matched ontology concepts, ontology attributes, ontology relationships, clause fragment identifiers, and corresponding relationships are recorded to generate the ontology semantic relationship set. Perform contextual semantic completion and cross-clause relationship association on the ontology semantic relationship set to generate a complete ontology semantic relationship set; The process of generating a complete ontology semantic relationship set is as follows: The ontology semantic relationship set is read according to the order of clause fragments in the clause fragment set. Taking the clause fragment corresponding to the current ontology semantic relationship as the center, adjacent clause fragments before and after the current clause fragment are retrieved. For content lacking ontology concepts, ontology attributes, or ontology relationships in the current ontology semantic relationship, the missing ontology concepts, ontology attributes, or ontology relationships are filled in based on ontology semantic relationships in adjacent clause fragments that have the same ontology concepts, the same referents, or continuous semantic relationships as the current ontology semantic relationship, thus completing the contextual semantic completion. For ontology semantic relationships distributed across different clause fragments that have the same ontology concepts, conditional succession relationships, rights and obligations correspondence relationships, negation relationships, or exception relationships, cross-clause relationships are established according to the order of clause fragments, and the starting clause fragment identifier, ending clause fragment identifier, and relationship type corresponding to the cross-clause relationship are recorded. The completed ontology semantic relationships and the established cross-clause relationships are written into the ontology semantic relationship set to generate a complete ontology semantic relationship set. Based on the risk triggering path in the procurement domain ontology model, risk triggering reasoning is performed on the complete ontology semantic relationship set to extract implicit risk triggering elements and generate a set of implicit risk triggering elements. The specific process for generating the implicit risk triggering element set is as follows: Based on each risk triggering path in the procurement domain ontology model, and according to the ontology concepts, ontology attributes, ontology relations, and path order contained in each risk triggering path, the corresponding content is sequentially retrieved in the complete ontology semantic relation set. Each risk triggering path contains 2 to 6 ontology concepts, 1 to 4 ontology attributes, and 1 to 5 ontology relations. The path matching rate between the complete ontology semantic relation set and the corresponding risk triggering path is calculated. When the ontology concept matching rate, ontology attribute matching rate, and ontology relation matching rate are all greater than or equal to 80%, and the path order matching rate is greater than or equal to 90%, it is determined that the ontology concepts, ontology attributes, and ontology relations in the complete ontology semantic relation set satisfy the path start point, path connection relationship, and path end point of the corresponding risk triggering path, and the corresponding risk triggering path is determined to be valid. The risk type identifier corresponding to the valid risk triggering path is read, and the risk type identifier is determined as the risk type. The ontology attributes and conditional constraints used to limit the validity of the risk in the valid risk triggering path are read. When the corresponding attribute value, comparison relationship, and condition range contain at least two valid items, the corresponding attribute value, comparison relationship, and condition range are determined as triggering conditions. The ontology concept corresponding to the endpoint of the established risk triggering path is read, and the procurement entity, procurement object, qualification conditions, technical requirements, commercial conditions, or performance responsibilities corresponding to the ontology concept are determined as the target. Based on the ontology concept, ontology attribute, and ontology relationship matched by the established risk triggering path, the corresponding clause fragment identifier in the complete ontology semantic relationship set is read. When the established risk triggering path corresponds to more than two clause fragment identifiers, it is determined that the established risk triggering path spans multiple clause fragments, and the cross-clause relationship is determined based on the corresponding starting clause fragment identifier, ending clause fragment identifier, and relationship type in the complete ontology semantic relationship set. The risk type, triggering conditions, target, clause fragment identifier, and cross-clause relationship are recorded as implicit risk triggering elements, and each implicit risk triggering element is sorted in ascending order of risk triggering path identifier to generate an implicit risk triggering element set. The set of implicit risk triggering factors is recorded in a structured manner to generate a set of procurement semantic elements; The specific process for generating the procurement semantic element set is as follows: Each implicit risk triggering element in the implicit risk triggering element set is read sequentially, and the risk type, triggering condition, target, clause fragment identifier, and cross-clause relationship are written into the corresponding data fields, with the risk triggering path identifier used as the association identifier between the data fields; the source position of each implicit risk triggering element in the clause fragment set is determined based on the clause fragment identifier, and the starting clause fragment identifier, ending clause fragment identifier, and relationship type corresponding to each implicit risk triggering element are recorded based on the cross-clause relationship; each implicit risk triggering element is sorted according to the clause fragment identifier and risk triggering path identifier, and the sorted implicit risk triggering elements are recorded as procurement semantic elements, thus generating the procurement semantic element set.

[0021] In this embodiment, the construction of the semantic index set specifically includes: Based on the set of procurement semantic elements, determine the contextual relationships between each procurement semantic element and generate a set of contextual relationships; The process of generating the context association set is as follows: Each procurement semantic element in the procurement semantic element set is read sequentially, and the corresponding clause fragment identifier, risk trigger path identifier, and cross-clause relationship are read for each procurement semantic element. When the clause fragment identifiers corresponding to two procurement semantic elements are consecutive, and there is a cross-clause relationship between the two procurement semantic elements, the relationship between the two procurement semantic elements is determined as a clause context association. When the risk trigger path identifiers corresponding to two procurement semantic elements are the same, the relationship between the two procurement semantic elements is determined as a risk path context association. The relationship type of the two mutually related procurement semantic elements and the context association, as well as the corresponding clause fragment identifier and risk trigger path identifier, are recorded to generate the context association set. Based on the set of contextual relationships, the interrelated procurement semantic elements are arranged to generate a sequence of contextual semantic elements; The specific process for generating the context semantic element sequence is as follows: Read each context association relationship in the context association relationship set; take each procurement semantic element as the starting procurement semantic element; arrange the procurement semantic elements associated with the starting procurement semantic element from front to back according to the clause fragment identifiers corresponding to the clause context association relationship; arrange the procurement semantic elements with the same risk trigger path identifier or a sequential connection relationship according to the risk trigger path order corresponding to the risk path context association relationship; when the same procurement semantic element corresponds to both a clause context association relationship and a risk path context association relationship, determine the arrangement position according to the clause fragment identifier and record the corresponding risk trigger path identifier; reserve an arrangement position for repeatedly occurring procurement semantic elements; and write the arranged procurement semantic elements and their corresponding clause fragment identifiers, risk trigger path identifiers, and context association relationships sequentially into the sequence position to generate the context semantic element sequence. Dual-path vector encoding is performed on the sequence of contextual semantic elements to generate a set of clause context vectors and a set of risk path context vectors. The specific process for generating the clause context vector set and the risk path context vector set is as follows: Each procurement semantic element in the context semantic element sequence is read sequentially, and the clause context encoding branch and the risk path context encoding branch are input respectively. The clause context encoding branch reads the risk type, triggering condition, target, and cross-clause relationship corresponding to each procurement semantic element according to the arrangement order of the clause fragment identifiers. Position encoding and attention calculation are performed on adjacent procurement semantic elements to obtain the clause context vector of each procurement semantic element in the preceding and following clause content. The clause context vector set is then generated according to the arrangement order in the context semantic element sequence. The risk path context encoding branch reads the risk type, triggering condition, target, and cross-clause relationship corresponding to each procurement semantic element according to the risk triggering path sequence corresponding to the risk triggering path identifiers. Path position encoding and relationship attention calculation are performed on the procurement semantic elements in the same risk triggering path to obtain the risk path context vector of each procurement semantic element in the risk triggering path. The risk path context vector set is then generated according to the arrangement order in the context semantic element sequence. The association weight calibration and gating fusion of the clause context vector set and the risk path context vector set are performed to generate a semantic vector representation set. The generation process of the semantic vector representation set is as follows: Following the order of the context semantic element sequence, read the clause context vectors in the clause context vector set and the risk path context vectors in the risk path context vector set one by one. The dimension of both clause context vectors and risk path context vectors is set to 256. Calculate the cosine similarity between the clause context vector and the risk path context vector corresponding to the same procurement semantic element, and limit the number of clause context relationships and the number of risk path context relationships to between 0 and 10, respectively. When the corresponding number is greater than 10, it is counted as 10. Determine the clause association weight and the risk path association weight based on the cosine similarity, the number of clause context relationships, and the number of risk path context relationships. Normalize the clause association weight and the risk path association weight so that the sum of the clause association weight and the risk path association weight is 1. The item association weight ranges from 0.20 to 0.80. The 256 vector dimensions of the clause context vector are weighted using the clause association weight, and the 256 vector dimensions of the risk path context vector are weighted using the risk path association weight. The two weighted 256-dimensional vectors are then input into the gating unit. The gating unit calculates the difference between the two weighted vectors in each corresponding dimension and generates 256 gating coefficients using a gating function with values ​​ranging from 0 to 1. The clause semantic components corresponding to each dimension of the clause context vector are retained according to each gating coefficient, and the risk path semantic components corresponding to each dimension of the risk path context vector are retained according to the difference between 1 and each gating coefficient. The retained clause semantic components and risk path semantic components are added dimension by dimension to obtain a 256-dimensional semantic vector representation of the corresponding procurement semantic element. Each semantic vector representation is recorded according to the order in the context semantic element sequence, generating a semantic vector representation set. A vector index is built based on the set of semantic vector representations, and the procurement semantic elements and cross-term relationships corresponding to each semantic vector representation are written into the vector index to generate a set of semantic indexes. The specific process of generating the semantic index set is as follows: Each semantic vector representation in the semantic vector representation set is read sequentially, a unique vector identifier is assigned to each semantic vector representation, and a hierarchical adjacency relationship is established based on the vector distance between each semantic vector representation; each semantic vector representation is used as an index node, and the vector identifier, vector dimension, and hierarchical adjacency relationship are written into the index field of the corresponding index node to form a vector index; based on the correspondence between each semantic vector representation and the procurement semantic element, the procurement semantic element is written into the semantic element field of the corresponding index node, and the cross-terminal relationship corresponding to the procurement semantic element is read, and the starting clause fragment identifier, ending clause fragment identifier, and relationship type in the cross-terminal relationship are written into the cross-terminal relationship field of the corresponding index node; each index node is connected according to the hierarchical adjacency relationship, and each index node is recorded according to the vector identifier to generate the semantic index set.

[0022] In this embodiment, the generation of the risk feature association set specifically includes: Acquire procurement risk rules data, procurement standard terms data, and historical procurement risk data, and extract semantic elements to generate a set of risk knowledge units; The process of generating a risk knowledge unit set is as follows: First, current procurement management systems, compliance review rules, and risk warning rules are read through the data interface of the procurement management system to form procurement risk rule data. Second, standard procurement document templates, standard qualification conditions, standard technical requirements, standard commercial conditions, and standard performance terms are read through the data interface of the procurement template management library to form procurement standard clause data. Third, historical procurement risk data is generated by reading risk review records, review conclusions, objection handling records, and performance anomaly records corresponding to completed procurement projects from the historical project database of the procurement management system. Fourth, semantic elements are extracted from the procurement risk rule data, procurement standard clause data, and historical procurement risk data to extract risk types, triggering conditions, target objects, cross-clause relationships, and risk triggering paths. The data source identifiers and clause locations corresponding to each item are recorded. Risk types, triggering conditions, target objects, and cross-clause relationships with the same risk triggering path are recorded as a risk knowledge unit, generating a risk knowledge unit set. Based on the risk knowledge unit set, establish the relationship between each risk knowledge unit to generate a procurement risk knowledge base; The specific process for generating the procurement risk knowledge base is as follows: Each risk knowledge unit in the risk knowledge unit set is read sequentially. Risk type is used as the classification identifier, the target object as the object identifier, the triggering condition as the condition identifier, cross-clause relationships as the relationship identifier, and the risk triggering path as the path identifier. When two risk knowledge units have the same risk type and the same target object, a risk-like association is established. When the triggering condition of one risk knowledge unit corresponds to the target object of another risk knowledge unit, and the risk triggering paths of the two risk knowledge units are consecutive, a risk transmission association is established. When the cross-clause relationships corresponding to two risk knowledge units have the same starting clause type, ending clause type, and relationship type, a cross-clause association is established. Each risk knowledge unit is treated as a knowledge node, and the risk-like association, risk transmission association, and cross-clause association are treated as knowledge edges. The risk type, triggering condition, target object, cross-clause relationship, risk triggering path, and data source identifier are written into the corresponding knowledge nodes to generate the procurement risk knowledge base. Based on the semantic vector representations in the semantic index set, risk knowledge units associated with the corresponding procurement semantic elements are retrieved from the procurement risk knowledge base to generate a candidate risk knowledge unit set. The specific process for generating the candidate risk knowledge unit set is as follows: Each 256-dimensional semantic vector representation and its corresponding procurement semantic element are sequentially read from the semantic index set. Each 256-dimensional semantic vector representation is used as a retrieval vector. Cosine similarity is calculated between the retrieval vector and the corresponding 256-dimensional knowledge vector in the procurement risk knowledge base. The cosine similarity value ranges from 0 to 1. Risk knowledge units are then arranged in descending order of cosine similarity. From the arrangement results, the top 20 risk knowledge units with the highest cosine similarity are initially selected. When the top 20 risk knowledge units with the highest cosine similarity are selected... When a risk knowledge unit with a cosine similarity lower than 0.72 exists within the knowledge unit, that risk knowledge unit is deleted. When the cosine similarity is greater than or equal to 0.85, the corresponding risk knowledge unit is identified as a highly matched risk knowledge unit. When the cosine similarity is greater than or equal to 0.72 and less than 0.85, the corresponding risk knowledge unit is identified as a risk knowledge unit to be verified. A consistency check is performed on the highly matched risk knowledge units and the risk knowledge units to be verified. When the procurement semantic element and the risk type identifier corresponding to the risk knowledge unit are completely identical, the risk type is considered consistent. The value is set to 1 otherwise. When the target object of the procurement semantic element and the risk knowledge unit belongs to the same ontology concept, the target object consistency value is set to 1. When the two belong to adjacent ontology levels and the ontology level distance is 1, the target object consistency value is set to 0.5, otherwise it is set to 0. Read the risk triggering path corresponding to the procurement semantic element and the risk knowledge unit. When the starting point and ending point of the path are the same and the number of matching path connection relationships accounts for more than or equal to 80% of the total number of path connection relationships in the corresponding risk triggering path, it is determined that the risk triggering path has a front-to-back connection relationship. Retain the risk knowledge units with a risk type consistency value of 1, a target object consistency value greater than or equal to 0.5 and a risk triggering path with a front-to-back connection relationship, and retain the first 8 risk knowledge units in descending order of cosine similarity. When there are fewer than 8 risk knowledge units that meet the conditions, retain all the risk knowledge units that meet the conditions. Record the retained risk knowledge unit identifier, corresponding procurement semantic element, 256-dimensional semantic vector representation, cosine similarity, risk type consistency value, target object consistency value and risk triggering path to generate a candidate risk knowledge unit set. Semantically bridge the procurement semantic elements with the candidate risk knowledge unit set to generate a set of semantic bridging relationships; The specific process of generating the semantic bridging relationship set is as follows: Based on each procurement semantic element and its corresponding candidate risk knowledge unit set, the risk type, triggering condition, target, cross-clause relationship, and risk triggering path in the procurement semantic element are respectively matched with the risk type, triggering condition, target, cross-clause relationship, and risk triggering path in each candidate risk knowledge unit; when the risk type and target are the same, a risk subject bridging relationship is established; when the attribute value, comparison relationship, or condition range in the triggering condition has a corresponding relationship, a triggering condition bridging relationship is established; when the starting clause fragment identifier, ending clause fragment identifier, and relationship type in the cross-clause relationship match the cross-clause relationship in the candidate risk knowledge unit, a cross-clause bridging relationship is established; when the path start point, path connection relationship, and path end point in the risk triggering path have a continuous corresponding relationship with the risk triggering path in the candidate risk knowledge unit, a risk path bridging relationship is established; the procurement semantic element, candidate risk knowledge unit, risk subject bridging relationship, triggering condition bridging relationship, cross-clause bridging relationship, and risk path bridging relationship are associated and recorded to generate the semantic bridging relationship set; Based on the set of semantic bridging relationships, the differences between procurement semantic elements and corresponding risk knowledge units are quantified to generate a set of risk feature associations. The specific process for generating the risk feature association set is as follows: Based on each semantic bridging relationship in the semantic bridging relationship set, obtain the risk type, triggering condition, target, cross-clause relationship, and risk triggering path in the corresponding procurement semantic elements and risk knowledge units; when the risk types are the same, set the risk type difference value to 0, and when the risk types are different, set the risk type difference value to 1; normalize the attribute values ​​in the triggering conditions, calculate the difference between the normalized attribute values, and determine the triggering condition difference value based on whether the comparison relationship is the same and whether the condition range overlaps; determine the target object difference value based on whether the ontology concepts corresponding to the target objects are the same and the hierarchical distance between the ontology concepts; determine the cross-clause relationship difference value based on the number of matching start clause fragment types, end clause fragment types, and relationship types in the cross-clause relationship; determine the risk triggering path difference value based on the number of matching path start point, path connection relationship, and path end point in the risk triggering path and the length of the risk triggering path; associate the risk type difference value, triggering condition difference value, target object difference value, cross-clause relationship difference value, and risk triggering path difference value with the corresponding procurement semantic elements, risk knowledge units, and semantic bridging relationships to generate the risk feature association set.

[0023] In this embodiment, the generation of the risk assessment result set specifically includes: The improved BOCPD model inputs a risk feature association set into the model. The improved BOCPD model includes a cross-clause risk sequence construction module, a semantic run state inference module, a risk rule deviation change point detection module, and a risk evidence chain aggregation module. The improvement of the improved BOCPD model is as follows: The traditional BOCPD model recursively infers the posterior probability of run length according to the time sequence of the observed data and identifies the data distribution change point based on the reset state of the run length posterior probability. The improved BOCPD model folds and arranges the risk feature association data according to the clause sequence axis and the risk path axis, uses the semantic vortex capture mechanism to identify the round-trip semantic path of the same risk type in the risk trigger path, corrects the semantic run state posterior probability according to the semantic vortex unit, and identifies the risk rule deviation change point with the corrected reset state, generating a risk evidence chain and a risk assessment result set. The improved BOCPD model adopts a serial structure. The risk feature association set is first input into the cross-clause risk sequence construction module, which generates five-dimensional risk observations based on the five differences in each risk feature association data. These observations are then arranged into a two-dimensional cross-clause risk observation sequence according to the clause fragment identifier and risk triggering path. The two-dimensional cross-clause risk observation sequence is input into the semantic run state inference module, which outputs the continuation state probability, reset state probability, and semantic vortex unit set corresponding to each clause position. The continuation state probability, reset state probability, and semantic vortex unit set are input into the risk rule deviation change point detection module, which outputs a risk rule deviation change point set composed of clause fragment identifier, risk type, and reset state probability. The risk rule deviation change point set is input into the risk evidence chain aggregation module, which aggregates it with the corresponding risk observations and semantic vortex units to output a risk assessment result set containing risk type, risk clause position, risk degree value, risk level, and risk evidence chain. The training data for the improved BOCPD model comes from historical procurement documents, corresponding risk review records, review conclusions, objection handling records, and performance anomaly records. Historical procurement documents are converted into cross-clause risk observation sequences according to clause fragment identifiers. Procurement reviewers mark each clause location as a risk rule deviation point, the corresponding risk type, risk level, vortex start point, and vortex end point. These are then divided into training, validation, and test sets at 70%, 20%, and 10% respectively. A joint loss function is used during training. This joint loss function is obtained by weighted summation of risk rule deviation point detection loss, semantic run state loss, semantic vortex consistency loss, and risk level classification loss. The weight of the risk rule deviation point detection loss is set to 0.40; the semantic run state loss uses negative log-likelihood loss with a weight of 0.25; the semantic vortex consistency loss uses the mean square error loss of vortex start point error, vortex end point error, and closed trajectory continuity error with a weight of 0.20; and the risk level classification loss uses multi-class cross-entropy loss with a weight of 0.15. The sum of the weights of the four losses is 1. The Lion optimization algorithm is used to update and improve the trainable parameters in the BOCPD model. The initial learning rate is set to 0.03, the batch size is set to 16, the weight decay coefficient is set to 0.001, and the maximum number of training epochs is set to 120. When the joint loss on the validation set does not decrease for 5 consecutive epochs, the learning rate is decayed to 50% of the current learning rate, the lower bound of the learning rate is set to 0.001, and the upper bound of the gradient norm is set to 1. When the joint loss on the validation set does not decrease for 12 consecutive epochs, the improved BOCPD model is considered to have reached convergence, and the model parameters corresponding to the epoch with the minimum joint loss on the validation set are saved. In the cross-clause risk sequence construction module, the risk feature association data in the risk feature association set is folded and arranged in a dual axis to generate risk observations. The risk observations corresponding to the same risk type and target are aligned to the same risk path position to generate a cross-clause risk observation sequence. The process of generating a cross-clause risk observation sequence is as follows: Based on the risk feature association data in the risk feature association set, the corresponding clause fragment identifiers are arranged according to the clause order in the procurement document to form a clause order axis; the corresponding risk trigger paths are arranged according to the path start point, path connection relationship, and path end point to form a risk path axis; the risk type difference value, trigger condition difference value, target difference value, cross-clause relationship difference value, and risk trigger path difference value in each risk feature association data are written into the risk observation value, and the risk observation value is mapped to the intersection of the clause order axis and the risk path axis according to the corresponding clause fragment identifier and risk trigger path; the path position alignment is performed on the risk observation values ​​with the same risk type and target, the risk observation values ​​belonging to the same risk trigger path are arranged to the same risk path position, and the aligned risk observation values ​​are recorded in sequence according to the arrangement order of the clause fragment identifiers to generate a cross-clause risk observation sequence; In the semantic run state inference module, a semantic vortex capture mechanism is introduced to identify the closed trajectory of risk observations of the same risk type that recur along the risk triggering path in the cross-clause risk observation sequence. The first risk observation in the closed trajectory is determined as the vortex start point, and the last risk observation in the closed trajectory is determined as the vortex end point, generating a set of semantic vortex units. The process of generating a semantic vortex unit set is as follows: Each risk observation value in the cross-clause risk observation sequence is read sequentially, and the risk type, target, and triggering path corresponding to each risk observation value are tracked according to the arrangement order of the clause fragment identifiers. The semantic vortex capture mechanism identifies the round-trip semantic path formed between the first occurrence of the risk type and the second occurrence of the same target as a closed trajectory when the same risk type is transmitted from the previous clause fragment to the subsequent clause fragment along the risk triggering path, and the subsequent clause fragment points to the same target again. The first risk observation value appearing in the closed trajectory is determined as the vortex start point, and the risk observation value pointing to the same target again in the closed trajectory is determined as the vortex end point. The vortex start point, vortex end point, and the risk observation values ​​between them are recorded according to the arrangement order of the clause fragment identifiers to form a semantic vortex unit. The above processing is performed on all closed trajectories in the cross-clause risk observation sequence to generate a semantic vortex unit set. In the risk rule deviation change point detection module, the semantic run state posterior probability corresponding to each clause position is recursively deduced based on the cross-clause risk observation sequence, and the semantic run state posterior probability corresponding to each clause position is corrected by combining the semantic vortex unit set. The clause position where the corrected semantic run state posterior probability switches from the continuation state to the reset state is determined as the risk rule deviation change point, and a risk rule deviation change point set is generated. The process of generating the risk rule deviation change point set is as follows: Risk observation values ​​in the cross-clause risk observation sequence are read sequentially according to the arrangement order of the clause fragment identifiers. The current risk observation value is compared with the risk observation value and risk knowledge unit corresponding to the previous clause position. Based on the difference values ​​of risk type, triggering condition, target, cross-clause relationship, and risk triggering path, the continuation state probability and reset state probability corresponding to the current clause position are calculated. The continuation state probability and reset state probability are then normalized to obtain the semantic run-length state posterior probability corresponding to the current clause position. Semantic vortex units are then read. When the current clause position is located between the vortex start point and the vortex end point, based on the number of risk observations within the corresponding semantic vortex unit and the difference in risk triggering paths, the reset state probability corresponding to the current clause position is increased, and the continuation state probability and reset state probability are renormalized to obtain the corrected semantic run state posterior probability; when the corrected reset state probability is greater than the continuation state probability, and the continuation state probability corresponding to the previous clause position is greater than the reset state probability, the current clause position is determined as a risk rule deviation point, and each risk rule deviation point is recorded according to the arrangement order of the clause fragment identifiers to generate a risk rule deviation point set; In the risk evidence chain aggregation module, a risk evidence chain is generated based on the set of risk rule deviation change points. The location and risk level of the risk clause are determined according to the risk type, clause fragment identifier, risk rule deviation change points and semantic vortex unit in the risk evidence chain, and a risk assessment result set is generated. The process of generating the risk assessment result set is as follows: Each risk rule deviation point in the risk rule deviation point set is read sequentially to obtain the corresponding risk observation value, risk type, clause fragment identifier, and semantic vortex unit. Using the risk rule deviation point as the aggregation starting point, risk observation values ​​belonging to the same risk type and corresponding to the same target are traced forward and backward along the corresponding risk triggering path. The risk observation values ​​are connected according to the order of the clause fragment identifiers and written into the corresponding risk rule deviation point and semantic vortex unit to generate a risk evidence chain. The risk type difference value, triggering condition difference value, target difference value, cross-clause relationship difference value, and risk triggering path difference value contained in each risk observation value in the risk evidence chain are added together, and the sum is divided by the total number of difference values ​​to obtain the average risk deviation value. The number of risk rule deviation points is divided by the number of risk observation values ​​in the risk evidence chain to obtain the risk change point density. The number of risk observation values ​​contained in each semantic vortex unit is added together, and the sum is divided by the risk... The number of risk observations in the risk evidence chain is used to obtain the semantic vortex coverage. Corresponding weights are assigned to the average risk deviation, risk variable point density, and semantic vortex coverage. The risk level value is obtained by summing the products of the average risk deviation and its corresponding weight, the risk variable point density and its corresponding weight, and the semantic vortex coverage and its corresponding weight. Risk evidence chains with a risk level value less than 0.25 are classified as low risk; those with a risk level value greater than or equal to 0.25 and less than 0.50 are classified as moderate risk; those with a risk level value greater than or equal to 0.50 and less than 0.75 are classified as relatively high risk; and those with a risk level value greater than or equal to 0.75 are classified as high risk. The clause fragment identifier corresponding to the risk rule deviation variable point is identified as the risk clause location. The risk type, risk clause location, risk level value, risk level, and risk evidence chain are recorded as the risk assessment results, generating a risk assessment result set.

[0024] In this implementation, the procurement document data is compared with previous versions. Clause fragments with changed content are identified as updated data, and newly added clause fragments are identified as new clause data. Deep semantic parsing is performed on the updated or new clause data to extract corresponding implicit risk triggering elements, generating incremental procurement semantic elements. These incremental procurement semantic elements are then written into the procurement semantic element set. Context-aware vectorization is applied to the incremental procurement semantic elements, updating the semantic vector representation set and semantic index set. Based on the updated semantic index set, semantic bridging and difference quantification are re-executed to update the corresponding clause fragment identifiers and risk feature association data within the risk triggering path range, resulting in an updated risk feature association set. This updated risk feature association set is input into the improved BOCPD model. Using the clause fragment identifiers corresponding to the updated or new clause data as the starting point for incremental recalculation, the cross-clause risk observation sequence, semantic vortex unit set, risk rule deviation change point set, and risk evidence chain within the corresponding risk triggering path range are regenerated. Finally, the regenerated risk assessment results replace the risk assessment results corresponding to the clause fragment identifiers in the risk assessment result set, generating an updated risk assessment result set.

[0025] Example 1: A production equipment spare parts procurement project undertaken by a large manufacturing enterprise is used as an application scenario. The procurement documents for this project total 86 pages, including the procurement announcement, supplier qualification requirements, technical requirements, commercial terms, review rules, main contract terms, and performance requirements, totaling approximately 58,000 words. The procurement documents underwent three revisions during preparation. The second revision added a delivery deadline clause, and the third revision adjusted the supplier performance requirements and liability for breach of contract clauses. Some risks in the procurement documents are not directly expressed through explicit risk terms but are scattered throughout different clauses. For example, the qualification requirements clause requires suppliers to have three similar project performance records, the review rules clause uses five similar project performance records as the scoring basis, and the contract clause stipulates that the purchaser can unilaterally adjust the delivery time based on the site conditions, but does not simultaneously specify the corresponding compensation method for the supplier's work period. When using keyword matching, none of the above clauses contain obvious risk terms such as "violation," "exclusion," or "unfairness," making it difficult to directly identify their potential risks.

[0026] The procurement document data is input into the procurement document risk dynamic assessment system. The system first reads the main text, titles, clause numbers, table content, and page positions of the procurement documents, and processes garbled characters, duplicate headers, duplicate footers, and abnormal line breaks to generate structured procurement document data. Through clause semantic boundary recognition processing, the structured procurement document data is divided into 312 clause segments. After manual review, the boundaries of 306 clause segments are consistent with the manual division results, achieving a clause semantic boundary recognition accuracy rate of 98.08%. For content where qualification conditions and supporting document requirements are written consecutively, the system retains them in the same clause segment based on clause number changes, semantic continuity, and referential relationships, avoiding splitting complete qualification requirements into multiple independent sentences.

[0027] The system performs deep semantic analysis on 312 clause fragments and extracts implicit risk triggering elements by combining them with a procurement domain ontology model, generating a procurement semantic element set. This embodiment extracts a total of 684 procurement semantic elements, including 126 qualification restriction semantic elements, 173 technical requirement semantic elements, 142 commercial condition semantic elements, 154 performance responsibility semantic elements, and 89 review rule semantic elements. Based on the clause context and risk triggering path, the system performs context-aware vectorization representation processing on the procurement semantic elements, generating a semantic vector representation set and a semantic index set. For procurement semantic elements of the same risk type located in different clause fragments, the system associates them according to risk type, triggering condition, target, cross-clause relationship, and risk triggering path, enabling risk content scattered across qualification conditions, review rules, and contract clauses to be analyzed together.

[0028] The procurement risk knowledge base is constructed from the company's current procurement management system, standard procurement document templates, internal risk review rules, and historical procurement risk data. In this embodiment, the procurement risk knowledge base contains 1486 risk knowledge units, covering risk types such as unreasonable qualification requirements, misaligned technical parameters, inconsistent review standards, unbalanced payment terms, unequal performance responsibilities, and incomplete delivery deadlines. Based on the semantic index set, the system retrieves risk knowledge units corresponding to procurement semantic elements in the procurement risk knowledge base, and performs semantic bridging and difference quantification on the procurement semantic elements and risk knowledge units to generate a risk feature association set. After processing, a total of 217 risk feature association data items are formed, of which 62 risk feature association data items have a risk level of at least general risk.

[0029] Subsequently, the risk feature association set is input into the improved BOCPD model. The system performs dual-axis folding and arrangement of the risk feature association data according to the clause sequence axis and the risk path axis, generating a cross-clause risk observation sequence. It then uses a semantic vortex capture mechanism to identify closed trajectories where the same risk type repeatedly appears along the risk trigger path and points back to the same target. This embodiment identifies a total of 14 semantic vortex units, one of which consists of three clause fragments: supplier performance requirements, review bonus conditions, and supporting document requirements. This semantic vortex unit indicates that the same supplier performance requirement appears repeatedly in different clause positions with different numbers of conditions, ultimately pointing back to the supplier qualification review result. The improved BOCPD model, based on the cross-clause risk observation sequence, recursively derives the posterior probability of the semantic run state, and combines this with the semantic vortex unit set to correct and reset the state probability, identifying a total of 27 risk rule deviation points and generating corresponding risk evidence chains.

[0030] The system ultimately generated 23 risk assessment results, including 4 high-risk, 7 high-risk, and 12 moderate-risk results. High-risk results included inconsistencies in supplier performance requirements, clearly directional combinations of technical parameters, unilateral adjustments to delivery times by the purchaser without compensation, and unequal scope of liability for breach of contract. Taking the inconsistency in supplier performance requirements as an example, the system connected the qualification clauses, review rules clauses, and supporting documentation clauses into a risk evidence chain, identifying the qualification clauses and review rules clauses as risk clauses, calculating a risk level of 0.81, and classifying them as high-risk. After reviewing the relevant clauses based on the risk evidence chain, the procurement review personnel uniformly adjusted the supplier performance requirements to three items and removed additional performance bonus conditions that were inconsistent with the scale of the procurement project.

[0031] After the third revision of the procurement documents, the system identified 9 updated clause fragments and 2 newly added clause fragments using the procurement document version identifier and clause fragment identifier. Incremental semantic parsing and incremental recalculation were performed only on these 11 clause fragments and their corresponding risk trigger paths. The incremental recalculation took 18.6 seconds, while re-performing the entire procurement document took 126.4 seconds, representing an 85.28% reduction in processing time. The updated risk assessment results showed that 3 of the original 4 high-risk items were downgraded to low-risk, while 1 remained high-risk because the breach of contract clause had not yet been revised. These results demonstrate that the present invention can identify hidden risks scattered across multiple clause fragments and can quickly update risk assessment results after procurement document revisions.

[0032] Table 1 Comparison of the Application Effects of Dynamic Risk Assessment for Procurement Documents

[0033] As shown in Table 1, the 30 procurement documents contained a total of 8,942 clause fragments, of which 186 risks were confirmed by procurement reviewers. Manual review correctly identified 162 risks, keyword and fixed rule methods correctly identified 131 risks, and the method of this invention correctly identified 176 risks, corresponding to a risk identification accuracy rate of 94.62%. The method of this invention correctly identified 68 implicit risks and 49 cross-clause risks, demonstrating that the procurement domain ontology model, semantic bridging, difference quantification, and semantic vortex capture mechanism can improve the ability to identify distributed risk content.

[0034] In terms of processing efficiency, the method of this invention processes a single procurement document in an average of 4.6 minutes, which is lower than the 96.5 minutes of manual review and the 14.8 minutes of keyword and fixed rule methods. After the procurement document is modified, the method of this invention can perform incremental recalculation based on the updated and newly added clause fragments, with an average time of 18.6 seconds, eliminating the need to reprocess the entire procurement document. Furthermore, the method of this invention can incorporate risk type, risk clause location, risk level, risk grade, and risk evidence chain into the risk assessment results, achieving a 100% traceability rate for risk clause locations and a 93.55% completeness rate for risk evidence chains. This provides procurement reviewers with clear sources of risk and a basis for clause modification.

[0035] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic risk assessment of procurement documents based on semantic understanding, characterized in that, Includes the following steps: Acquire procurement document data and preprocess it to generate structured procurement document data; Perform semantic boundary recognition processing on structured procurement document data to generate a set of clause fragments; Deep semantic parsing is performed on the set of clause fragments, and implicit risk triggering elements in the clause fragments are extracted by combining the procurement domain ontology model to generate a set of procurement semantic elements. The semantic element set of procurement is processed with context-aware vectorization to generate a set of semantic vector representations and to construct a set of semantic indexes. Construct a procurement risk knowledge base, and based on the semantic index set, perform semantic bridging and difference quantification on the procurement semantic element set and the procurement risk knowledge base to generate a risk feature association set; The risk feature association set is input into the improved BOCPD model, a semantic vortex capture mechanism is introduced, the risk rule deviation point is identified and the risk evidence chain is constructed to generate a risk assessment result set. The system acquires updated or newly added clause data from procurement documents, performs incremental semantic parsing on the updated or newly added clause data, updates the procurement semantic element set and risk feature association set, triggers the multidimensional risk assessment model to perform incremental recalculation, and generates an updated risk assessment result set.

2. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The procurement document data, including text files, scanned files, image files, and table files, is obtained from the procurement management system. The data undergoes file type identification, text content extraction for text files, image enhancement, noise removal, skew correction, and optical character recognition for scanned and image files, and table boundary identification, cell positioning, and row / column relationship restoration for table files. Based on page position, heading level, paragraph order, and table correspondence in the procurement document data, the identified text and table content are parsed for layout, and abnormal character removal, character encoding standardization, punctuation standardization, duplicate content removal, and field format standardization are performed. The processed text and table content are then structured according to the original page order and content relationships to generate structured procurement document data.

3. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The generation of the set of clause fragments specifically includes: The structured procurement document data is divided into text units, and a sequence of text units is generated; Semantic boundary features are extracted from adjacent text units in a text unit sequence to generate a semantic boundary feature sequence. Based on the semantic boundary feature sequence, the semantic continuity between adjacent text units is determined, and the position where the semantic continuity is interrupted is determined as the semantic boundary of the candidate clause, generating a set of semantic boundaries of the candidate clause. The semantic boundary set of candidate clauses is subjected to duplicate boundary elimination, adjacent boundary merging and nested boundary correction to generate a semantic boundary set of clauses. The text unit sequence is segmented based on the semantic boundary set of the clauses to generate a set of clause fragments.

4. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The generation of the procurement semantic element set specifically includes: A set of clause fragments is subjected to joint syntactic and semantic parsing to generate a set of clause semantic relations. Map the set of semantic relationships of terms to the procurement domain ontology model to generate an ontology semantic relationship set; Perform contextual semantic completion and cross-clause relationship association on the ontology semantic relationship set to generate a complete ontology semantic relationship set; Based on the risk triggering path in the procurement domain ontology model, risk triggering reasoning is performed on the complete ontology semantic relationship set to extract implicit risk triggering elements and generate a set of implicit risk triggering elements. The set of implicit risk triggering factors is recorded in a structured manner to generate a set of procurement semantic elements.

5. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The construction of the semantic index set specifically includes: Based on the set of procurement semantic elements, determine the contextual relationships between each procurement semantic element and generate a set of contextual relationships; Based on the set of contextual relationships, the interrelated procurement semantic elements are arranged to generate a sequence of contextual semantic elements; Dual-path vector encoding is performed on the sequence of contextual semantic elements to generate a set of clause context vectors and a set of risk path context vectors. The association weight calibration and gating fusion of the clause context vector set and the risk path context vector set are performed to generate a semantic vector representation set. A vector index is created based on the set of semantic vector representations, and the procurement semantic elements and cross-term relationships corresponding to each semantic vector representation are written into the vector index to generate a set of semantic indexes.

6. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The generation of the risk feature association set specifically includes: Acquire procurement risk rules data, procurement standard terms data, and historical procurement risk data, and extract semantic elements to generate a set of risk knowledge units; Based on the risk knowledge unit set, establish the relationship between each risk knowledge unit to generate a procurement risk knowledge base; Based on the semantic vector representations in the semantic index set, risk knowledge units associated with the corresponding procurement semantic elements are retrieved from the procurement risk knowledge base to generate a candidate risk knowledge unit set. Semantically bridge the procurement semantic elements with the candidate risk knowledge unit set to generate a set of semantic bridging relationships; Based on the set of semantic bridging relationships, the differences between procurement semantic elements and corresponding risk knowledge units are quantified to generate a set of risk feature associations.

7. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The generation of the risk assessment result set specifically includes: The improved BOCPD model inputs a risk feature association set into the model. The improved BOCPD model includes a cross-clause risk sequence construction module, a semantic run state inference module, a risk rule deviation change point detection module, and a risk evidence chain aggregation module. The improvement of the improved BOCPD model is as follows: The traditional BOCPD model recursively infers the posterior probability of run length according to the time sequence of the observed data and identifies the data distribution change point based on the reset state of the run length posterior probability. The improved BOCPD model folds and arranges the risk feature association data according to the clause sequence axis and the risk path axis, uses the semantic vortex capture mechanism to identify the round-trip semantic path of the same risk type in the risk trigger path, corrects the semantic run state posterior probability according to the semantic vortex unit, and identifies the risk rule deviation change point with the corrected reset state, generating a risk evidence chain and a risk assessment result set. In the cross-clause risk sequence construction module, the risk feature association data in the risk feature association set is folded and arranged in a dual axis to generate risk observations. The risk observations corresponding to the same risk type and target are aligned to the same risk path position to generate a cross-clause risk observation sequence. In the semantic run state inference module, a semantic vortex capture mechanism is introduced to identify the closed trajectory of risk observations of the same risk type that recur along the risk triggering path in the cross-clause risk observation sequence. The first risk observation in the closed trajectory is determined as the vortex start point, and the last risk observation in the closed trajectory is determined as the vortex end point, generating a set of semantic vortex units. In the risk rule deviation change point detection module, the semantic run state posterior probability corresponding to each clause position is recursively deduced based on the cross-clause risk observation sequence, and the semantic run state posterior probability corresponding to each clause position is corrected by combining the semantic vortex unit set. The clause position where the corrected semantic run state posterior probability switches from the continuation state to the reset state is determined as the risk rule deviation change point, and a risk rule deviation change point set is generated. In the risk evidence chain aggregation module, a risk evidence chain is generated based on the set of risk rule deviation change points. The location and risk level of the risk clause are determined according to the risk type, clause fragment identifier, risk rule deviation change points and semantic vortex units in the risk evidence chain, and a risk assessment result set is generated.

8. The method for dynamic risk assessment of procurement documents based on semantic understanding according to claim 1, characterized in that, The procurement document data is compared with previous versions. Clause fragments with changed content are identified as updated data, and newly added clause fragments are identified as new clause data. Deep semantic parsing is performed on the updated or new clause data to extract corresponding implicit risk triggering elements, generating incremental procurement semantic elements. These incremental procurement semantic elements are then written into the procurement semantic element set. Context-aware vectorization is applied to the incremental procurement semantic elements, updating the semantic vector representation set and semantic index set. Based on the updated semantic index set, semantic bridging and difference quantification are re-executed to update the corresponding clause fragment identifiers and risk feature association data within the risk triggering path range, resulting in an updated risk feature association set. This updated risk feature association set is input into the improved BOCPD model. Using the clause fragment identifiers corresponding to the updated or new clause data as the starting point for incremental recalculation, the model regenerates the cross-clause risk observation sequence, semantic vortex unit set, risk rule deviation change point set, and risk evidence chain within the corresponding risk triggering path range. The regenerated risk assessment results replace the risk assessment results for the corresponding clause fragment identifiers in the risk assessment result set, generating an updated risk assessment result set.

Images