A surrounding string risk identification method based on micro-service cooperation and natural language processing

Abstract

The present application relates to the technical field of bidding risk monitoring, and particularly relates to a method for identifying surrounding bidding risks based on micro-service cooperation and natural language processing, which comprises a method executed by a micro-service cluster, including locating response commitment statements in different bidding texts through a natural language processing service, implementing reversible counterfactual transformation on conditional scopes or exceptional scopes, and composing a counterfactual constraint test set, recording semantic repair operations and cross-section propagation sequences, and associating a repair path homology evidence set; sending to a cooperative review service, disabling semantic repair micro-services one by one, and calling the remaining semantic repair micro-services to regenerate the repair path homology evidence set; and outputting surrounding bidding risk identification results. The present application can reduce misjudgments caused by industry general expressions, mandatory expressions in bidding documents and common template responses, and improve the pertinence, explainability and evidence positioning accuracy of surrounding bidding risk identification.

Description

Technical Field

[0001] This invention relates to the field of bidding risk monitoring technology, and in particular to a method for identifying bid-rigging risks based on microservice collaboration and natural language processing. Background Technology

[0002] With the widespread adoption of electronic bidding, the number and complexity of bid documents are constantly increasing. Utilizing information technology to identify bid-rigging and collusion risks has become a crucial requirement for bidding supervision. Existing technologies typically determine whether there are abnormal connections between bid documents based on features such as keyword overlap, sentence similarity, consistent document attributes, or price correlation. However, bidding documents themselves have standardized response requirements, and it is reasonable for different bidders to use the same technical terminology, format templates, or commitment statements. Simply relying on static text similarity can easily lead to misjudgments.

[0003] Meanwhile, jointly prepared tender documents can reduce surface similarity through methods such as synonym substitution, word order adjustment, paragraph shifting, condition splitting, and partial rewriting, making it difficult for traditional detection methods to identify their deep logical connections. Although some natural language processing methods can analyze the semantics of tender texts, their judgment results are easily affected by the expression preferences and decoding strategies of a single language model, and they usually only output risk scores, making it difficult to explain how abnormal relationships are formed or to trace back to specific locations in the original tender text.

[0004] Furthermore, existing technologies have limited research on the consequential modifications to tender documents after they have been subjected to the same semantic intervention in order to restore the compatibility of tender constraints. They also cannot effectively utilize cross-segment repair order, operation triggering dependence, and heterogeneous model ablation results to distinguish between true homogeneous features and model random biases. Summary of the Invention

[0005] This invention provides a method for identifying bid-rigging risks based on microservice collaboration and natural language processing. It can proactively reveal the internal constraints of the bid text and perform cross-model verification and source tracing of the identified evidence.

[0006] A method for identifying tag-rigging risks based on microservice collaboration and natural language processing, executed by a microservice cluster, the method comprising:

[0007] S1. Locate the commitment statements in different bid texts that respond to the same bidding constraint through natural language processing services, perform the same reversible counterfactual transformation on the conditional scope or exception scope of the commitment statements, and form a counterfactual constraint test set by the transformed commitment statements, transformation types and associated reference positions;

[0008] S2. Send the counterfactual constraint test set to multiple independent semantic repair microservices. Each semantic repair microservice restores the transformed commitment statement to a state compatible with the original bidding constraints. Record the semantic repair operations and cross-segment propagation order of non-mandatory response content during the restoration process. Associate the abnormally consistent semantic repair operations and cross-segment propagation order between different bidding texts as a set of evidence of homology of repair paths.

[0009] S3. The repair path homology evidence set is sent to the collaborative review service. The collaborative review service disables the semantic repair microservices that participated in generating the repair path homology evidence set one by one, and calls the remaining semantic repair microservices to regenerate the repair path homology evidence set. When the regenerated repair path homology evidence set continuously points to the same bidding entity and can be traced back to the associated reference position, the collusion risk identification result is output.

[0010] Optionally, the natural language processing service performs clause parsing on the tender documents, extracts tender constraint clauses with unique identifiers, and generates semantic matching templates based on the core predicates and constraint parameters in the clauses. For each semantic matching template, sentence vector similarity calculation and keyword co-occurrence retrieval are used in each tender text to lock down candidate commitment statements that respond to the same tender constraint, and negation or non-committee statements are eliminated through dependency parsing to obtain accurate commitment statements.

[0011] Optionally, S1 further includes:

[0012] Semantic role labeling and dependency analysis are performed on commitment statements to identify clauses or phrases introduced by conditional conjunctions as conditional scopes and modifiers introduced by exception exclusion words as exception scopes.

[0013] For the conditional scopes responding to the same tender constraint in different tender texts, the same reversible counterfactual transformation is applied. The transformation includes replacing the antecedents that are true with the antecedents that are false, or removing the condition-guided structure, generating a transformed commitment statement that violates the original constraint, and recording the original conditional word sequence before the transformation to maintain the reversibility of the transformation.

[0014] For exception scopes responding to the same tender constraint in different tender documents, apply the same reversible counterfactual transformation, which includes replacing the exception scope with a general scope statement or deleting the entire exception modifier, generating a transformed commitment statement, and recording the original exception word sequence before the transformation;

[0015] Record the transformation type for each transformation;

[0016] Record the associated reference position of the transformed commitment statement in the original bid text. The associated reference position includes at least the bid text identifier, the chapter number, and the start and end character offsets of the statement.

[0017] Optionally, the transformation type is at least selected from condition inversion, condition removal, exception generalization, and exception removal. The transformed commitment statements of all bid texts under the same bidding constraint, their respective transformation types, and associated reference positions are encapsulated into a test sample, and all test samples are collected to form the counterfactual constraint test set.

[0018] Optionally, the multiple independent semantic repair microservices are deployed as independent service instances based on different pre-trained language models and employing differentiated decoding strategies, and all test samples in the counterfactual constraint test set are fully distributed to each semantic repair microservice via message middleware.

[0019] After receiving the test samples, each semantic repair microservice takes the original text of the bidding constraint clause to which the transformed commitment statement belongs as the repair constraint condition, and generates the repaired commitment statement through constraint decoding or mask language modeling. This makes the repair result logically compatible with the bidding constraint clause, while maintaining compatibility with other non-binding content of the original bid text.

[0020] Optionally, during the generation of the post-repair commitment statement, every associated modification triggered by this repair and applied to the non-mandatory response content is monitored and recorded. The non-mandatory response content includes statements, data items, or related commitments in the original tender text that do not directly violate the tender constraints but are indirectly adjusted to maintain the internal consistency of the text or the coherence of cross-paragraph references.

[0021] Optionally, for each associated modification, the operation type of the semantic repair operation, the position of the operation object in the text, and the timestamp of the operation trigger are extracted, and a cross-segment propagation order is constructed according to the triggering dependency relationship between operations. The cross-segment propagation order is represented by a directed acyclic graph of the operation occurrence, where the node is the semantic repair operation and the edge represents the dependency relationship between the preceding operation and the subsequent operation.

[0022] For different bid texts responding to the same bidding constraint, the semantic repair microservice compares the semantic repair operation sequence and cross-segment propagation order corresponding to each of the different bid texts. When the standardized editing similarity and graph structure isomorphism of the sequence both exceed the preset threshold, it is determined that there is an abnormally consistent repair propagation mode between the two.

[0023] The identifiers of any two or more bid texts that meet the above-mentioned anomaly consistency conditions, the identifiers of the bid constraints to which they belong, the instance identifier of this semantic repair microservice, the common semantic repair operation mode, and the cross-segment propagation subgraph are associated as a repair path homology evidence; the repair path homology evidence output by all semantic repair microservices is collected to form the repair path homology evidence set.

[0024] Optionally, the set of evidence for the homology of repair paths is sent to the collaborative review service via a message middleware. The collaborative review service extracts instance identifiers of the semantic repair microservices that participated in generating the homology evidence for each repair path from the evidence set and constructs a microservice-evidence mapping table. For each semantic repair microservice instance involved in the microservice-evidence mapping table, the collaborative review service sends a deactivation command to the service registry center to temporarily unregister the service of the corresponding instance, so that the corresponding instance no longer receives subsequent test set distribution requests. After each semantic repair microservice instance is deactivated, the collaborative review service redistributes the counterfactual constraint test set to the remaining semantic repair microservice instances that have not been deactivated, receives the regenerated homology evidence for the repair paths, and performs consistency analysis on the newly generated evidence.

[0025] Optionally, in the regenerated evidence of homology of the repair path, if at least one piece of evidence points to a set of bidding entities whose intersection with the set of bidding entities pointed to by the original evidence before the suspension is not empty, and the cross-segment propagation order that is determined to be abnormally consistent in the evidence can be traced back to the associated reference position recorded in step S1, then it is determined that the suspension operation has not eliminated the homology characteristics of the bidding entities.

[0026] Optionally, after completing the deactivation and regeneration test of all semantic repair microservice instances, the collaborative review service will count the cumulative number of times each bidding entity is determined to still retain homology characteristics throughout the entire ablation review process;

[0027] When the cumulative number of bids by any bidder reaches a preset threshold and can be traced back to the same set of associated reference locations, a bid-rigging risk identification result is generated. The result includes at least the bidder's identifier, the identifiers of other bidders that are related to it, the corresponding list of associated reference locations, and the risk confidence level.

[0028] The beneficial effects of this invention are:

[0029] This invention constructs a counterfactual constraint test set that violates the original constraints by applying the same reversible counterfactual transformation to the conditional or exception scope of different bid texts for the same bidding constraint clause. It records the transformation type, original word sequence, and associated reference positions, allowing different bid texts to be tested under a unified and traceable semantic intervention condition. Compared to directly comparing the word, sentence structure, or text similarity of bid documents, this invention can proactively trigger potential conditional dependencies, exception restrictions, numerical harmonies, and cross-paragraph references within the bid texts, transforming common compilation features masked by synonym rewriting, word order adjustment, paragraph migration, or partial rewriting into observable repair behavior. Because the counterfactual transformation only applies to the commitment statements corresponding to the same bidding constraint and can restore the original text position, it can reduce misjudgments caused by industry-standard expressions, mandatory wording in bidding documents, and conventional template responses, improving the targeting, interpretability, and evidence location accuracy of bid-rigging risk identification.

[0030] This invention distributes a complete counterfactual constraint test set to multiple independent semantic repair microservices employing different pre-trained language models and differentiated decoding strategies. It not only analyzes the post-repair commitment statements but also continuously records non-mandatory associated modifications triggered by the repair. Furthermore, it utilizes the semantic repair operation sequence and operation trigger dependencies to form a cross-segment propagation order, thereby elevating traditional result text similarity judgment to repair process homology judgment. Further, the collaborative review service deactivates the semantic repair microservice instances one by one, regenerates homology evidence of the repair path in the remaining instances, and determines the risk outcome based on the intersection of bidding entities, the backtracking results of the cross-segment propagation order, and the number of times evidence is retained at the same associated citation position. This controlled ablation review mechanism effectively distinguishes between anomalous consistency caused by a single language model's expression preference or decoding strategy and stable common compilation features that still exist across heterogeneous models, reducing the risk of model bias directly leading to collusion conclusions. Simultaneously, it forms a complete traceability chain from the risk subject, homologous related subjects, ablation rounds, repair propagation process, to the original bidding text position, improving the stability of the identification results and the efficiency of regulatory review. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0033] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0034] This invention proactively injects reversible counterfactual transformations to observe the repair propagation path generated when disturbances are eliminated in each bid text. Even if the text has undergone synonym rewriting, word order adjustment, or term replacement, as long as multiple bid documents originate from the same draft, their cross-paragraph citations, exception adjustments, and related commitments often produce an unusually consistent repair order. By ablating and verifying microservices one by one, we can avoid the direct formation of risk conclusions from a single model bias.

[0035] like Figure 1 As shown, a method for identifying tag-rigging risks based on microservice collaboration and natural language processing is executed by a microservice cluster. The method includes:

[0036] S1. Locate the commitment statements in different bid texts that respond to the same bidding constraint through natural language processing services, apply the same reversible counterfactual transformation to the conditional scope or exception scope of the commitment statements, and form a counterfactual constraint test set by combining the transformed commitment statements, transformation types and associated reference positions.

[0037] S11. Generate tender constraint clauses and semantic matching templates with unique identifiers.

[0038] Tender documents read by natural language processing service and multiple bid documents corresponding to the tender documents. ,in, The tender document indicating the current procurement task; Indicates the first One bid document, ; This indicates the total number of bid documents corresponding to the current procurement task.

[0039] First, the layout hierarchy of the tender documents is restored. For body paragraphs, clause boundaries are determined based on heading numbers, indentation levels, paragraph ends, and punctuation marks. For table content, multi-level table headers, row headings, and corresponding cell contents are sequentially connected to form clause text that can independently express the meaning of constraints. For cross-paragraph references, the referencing relationships are established with the superior clauses they refer to, in accordance with the aforementioned requirements and the provisions of this section.

[0040] When a paragraph contains two or more constraints with independent subject-verb relationships, it is split into multiple tender constraint clauses using a semicolon, ordinal number, or parallel core predicates as the dividing point. After clause parsing, the following is obtained:

[0041] ;

[0042] in, This represents the set of tender constraints. Indicates the first One bidding constraint clause, ; This represents the total number of tender constraint clauses parsed from the tender documents.

[0043] This embodiment employs a multi-task natural language understanding model to process the tender constraint clauses. The multi-task natural language understanding model includes a Chinese word embedding layer, a shared Transformer encoder, a semantic element annotation layer, a semantic role annotation layer, a dependency parsing layer, and a commitment type discrimination layer.

[0044] The shared Transformer encoder is set to 12 layers, with 12 self-attention heads per layer. The hidden layer dimension is set to 768, the feedforward network intermediate dimension is set to 3072, and the dropout ratio is set to 0.1. The maximum length of a single input is set to 512 words. For terms exceeding 512 words, sliding encoding is performed with a window length of 384 words and an adjacent window overlap length of 128 words. Weighted merging is then performed based on the distance of each word in the overlapping region from the center of the window.

[0045] The semantic element annotation layer employs a linear mapping layer and a conditional random field layer to identify core predicates, constraint objects, numerical values, units, ranges, time limits, conditional components, and exception components according to the BIO annotation method. The semantic role annotation layer employs a linear mapping layer and a conditional random field layer to identify agent roles, patient roles, behavioral roles, conditional roles, and exception roles.

[0046] The dependency parsing layer employs a dual affine dependency parsing structure, calculating the representations of each word as dependency child nodes and dependency parent nodes, and outputting the dependency parent node and dependency relation type. The commitment type discrimination layer consists of a fully connected layer, a non-linear activation layer, and a Softmax output layer, used to distinguish between mandatory constraints, general statements, background descriptions, and non-constraint proposals.

[0047] The training corpus was constructed from anonymized tender documents and corresponding bid documents. The training data included at least 30,000 tender clauses and 120,000 pairs of tender clause-bid response statements. Two annotators labeled the core predicates, constraint parameters, semantic roles, dependency relations, conditional scope, exception scope, and commitment type, respectively. When the annotation results of the two annotators were inconsistent, a third annotator reviewed and determined the final annotation result.

[0048] The training sample record table can be set up as shown in the example below:

[0049] Table 1 shows examples of model training sample records.

[0050] Clause text The project lead has at least five years of experience in similar projects Core predicate has Constraint object Project lead Constraint parameter At least five years, experience in similar projects Condition component None Exception component None Commitment type Mandatory constraint Bid response statement The project lead we plan to assign has eight years of experience in similar projects Response relationship Positive response

[0051] The training data was divided into training, validation, and test sets in an 8:1:1 ratio. The model was trained using the AdamW optimizer with an initial learning rate of 2×10⁻⁻⁻⁶. 5 The batch size is set to 32, and the maximum training time is 10 epochs. Training stops when the combined loss on the validation set fails to decrease for two consecutive epochs, and the model parameters with the lowest combined loss on the validation set are saved.

[0052] The overall loss of multi-task training , is represented as:

[0053] ;

[0054] in, The task number represents the task sequence number. The four tasks are semantic element annotation, semantic role annotation, dependency parsing, and commitment type determination, respectively. Indicates the first Training loss for each task; Indicates the first The trainable task weight parameters corresponding to each task are initially set to 0 and trained together with the model parameters. The contribution used to automatically adjust the corresponding task loss based on the degree of uncertainty during the training process of each task.

[0055] By using the aforementioned trainable task weight parameters, there is no need to manually fix the loss weights for different tasks, which can avoid the situation where the number of semantic annotation samples and the number of dependency syntax samples are inconsistent, causing one task to overly dominate the model training.

[0056] After the model training is complete, the shared Transformer encoder and the output layers of each task are registered as the same model version. The Natural Language Processing service loads the same model version when executing S11 to S13 to prevent version deviations between the core predicate extraction results, dependency parsing results, and semantic role annotation results.

[0057] For each tender constraint clause, a unique identifier is generated based on the tender document identifier, clause section number, and standardized clause text. , is represented as:

[0058] ;

[0059] in, This represents the SHA-256 digest operation; Indicates the identifier for the tender document; Indicates the first Chapter number of each tender constraint clause; Represents a text normalization function; This indicates a string concatenation operation.

[0060] When the text normalization function is executed, it sequentially converts full-width characters to half-width characters, merges consecutive spaces, eliminates line breaks, standardizes number formats, and standardizes punctuation, but does not delete negative words, conditional conjunctions, exception words, or units of measurement. This ensures that the same clause can still be identified stably when the page layout changes, while avoiding confusion caused by deleting words with substantial semantic meaning.

[0061] The natural language processing service identifies core predicates and constraint parameters from the tender constraints and generates semantic matching templates in the following format. , is represented as:

[0062] ;

[0063] in, This indicates the core predicates identified from the tender constraints. Indicates the constraint object that the core predicate points to; Represents the set of constraint parameters associated with the core predicate; This indicates the semantic roles and dependencies in the tender constraints.

[0064] Constraint parameters are further recorded according to parameter type as numerical parameters, range parameters, time parameters, qualification parameters, object parameters, or behavioral status parameters. For example, when a tender constraint clause states that the project manager should have more than five years of experience in similar projects, the core predicate is "have," the constraint object is the project manager, and the constraint parameters are "more than five years" and "similar project experience."

[0065] S12. Lock in the precise commitment statement that responds to the same tender constraint.

[0066] For each semantic matching template, sentence vector similarity calculation and keyword co-occurrence retrieval are performed in each bidding text to obtain candidate commitment statements that respond to the corresponding bidding constraints.

[0067] S121. Sentence segmentation and sentence vector model of tender text.

[0068] The tender text is segmented into sentences according to chapter number, paragraph mark, period, semicolon, and table cell boundaries to obtain the searchable statements. .

[0069] in, Indicates the first In the tender document, regarding the first The first semantic matching template retrieved the first One query to be searched; This indicates the sequence number of the statement to be retrieved in the corresponding bid text.

[0070] The sentence vector model employs a dual-tower semantic encoding structure. The tender constraints corresponding to the semantic matching template and the sentence to be retrieved are respectively input into a Transformer encoder with shared parameters; the encoding results are processed through an attention pooling layer and a 256-dimensional linear projection layer before being normalized to obtain the corresponding sentence vector.

[0071] The sentence vector model uses labeled tender clauses and bid response statements for comparative learning. Statement pairs belonging to the same constraint-response relationship are designated as positive samples, while statement pairs not belonging to the same constraint-response relationship but with similar core predicates are designated as difficult negative samples. Contrastive learning loss... , is represented as:

[0072] ;

[0073] in, This indicates the number of positive sample pairs in the training batch, set to 64. Indicates the sequence number of the sample pair in the training batch; Indicates the first The sentence vector of each tender constraint clause; Indicates the relationship with the first The correct vector of bid response statements corresponding to each bidding constraint clause; Indicates the first in the training batch A vector of bid response statements; Represents the cosine similarity function; This represents the temperature coefficient.

[0074] The temperature coefficient was compared on the validation set with values ​​of 0.03, 0.05, 0.07, and 0.10, and the value with the highest accuracy when the validation set recall was not lower than 95% was used as the final value. In this embodiment, the validation results showed the highest accuracy when the value was 0.05, so the temperature coefficient was set to 0.05.

[0075] After the model is trained, the sentence vectors of the tender constraint clauses corresponding to each semantic matching template are pre-calculated and associated with unique identifiers for storage. When performing recognition on the tender text, only the sentence vectors of the statements to be retrieved need to be calculated online, reducing repetitive encoding operations.

[0076] S122. Sentence vector similarity calculation.

[0077] Sentence vector similarity between semantic matching template and the query statement , is represented as:

[0078] ;

[0079] in, Indicates the first Each semantic matching template corresponds to a sentence vector of the tender constraint clause; The sentence vector representing the statement to be retrieved; This represents the L2 norm.

[0080] In practice, this similarity function first reads the sentence vector corresponding to the semantic matching template from the template storage area, and then performs a dot product operation with the sentence vector of the current statement to be retrieved. Once the sentence vector has been normalized, the dot product result is the cosine similarity.

[0081] S123. Keyword co-occurrence search.

[0082] A set of template keywords is generated from the core predicate, constraint objects, and constraint parameters, and the keyword matching degree is calculated using a weighted co-occurrence method.

[0083] Keyword weight is: ;

[0084] in, Keywords Distinguishing weights; This indicates the total number of bid response statements used to generate keyword statistics. Indicates the presence of keywords The number of bid response statements.

[0085] Keyword co-occurrence score , is represented as:

[0086] ;

[0087] in, This represents a set of template keywords consisting of the core predicate, constraint objects, and constraint parameters of a semantic matching template. This represents the set of keywords for the retrieved statement after word segmentation and synonym normalization.

[0088] The normalized thesaurus is pre-built from synonym pairs found in historical bidding terms and bid response statements. Each synonym pair is confirmed by two business annotation personnel, who record the standard word, synonym, applicable part of speech, and applicable business category.

[0089] For example, "provide" and "submit" are only synonymous under the "materials delivery" category, and "possess" and "have" are synonymous under the "qualification conditions" category, avoiding unconditional substitution across business categories.

[0090] S124. Fusion retrieval and threshold determination.

[0091] The fusion matching score is:

[0092] ;

[0093] in, This represents the fusion matching score between the semantic matching template and the query statement; This represents the fusion weight corresponding to the sentence vector similarity.

[0094] The fusion weights are used to search the validation set in increments of 0.1 to 0.9. For each candidate value, the precision, recall, and F1 score of the constraint response matching task are calculated, and the value with the highest F1 score that results in a recall of at least 95% is selected. In this embodiment, the F1 score is highest when the fusion weight is 0.7, so it is set to 0.7.

[0095] When the fusion matching score reaches the matching threshold, the corresponding query statement is identified as a candidate commitment statement. The matching threshold is validated in increments of 0.01 between 0.50 and 0.90, with a recall rate of no less than 95% as a constraint, and the value with the highest precision is selected. In this embodiment, the matching threshold obtained is 0.72.

[0096] When the number of statements reaching the matching threshold exceeds 5, the top 5 statements are retained according to their fusion matching scores from highest to lowest. The number of statements to retain is determined by the validation set: the coverage of correct response statements is tested when retaining the top 1 to top 10 statements respectively. When the number of statements to retain is increased from 5 to 6, the coverage improvement is less than 0.5 percentage points, so the number of statements to retain is set to 5.

[0097] S125. Removal of negative and non-committal statements.

[0098] The candidate commitment statements are input into a multi-task natural language understanding model, and the probability of the candidate commitments is obtained through a commitment type discrimination layer. , is represented as:

[0099] ;

[0100] in, This represents the commitment category score output by the commitment type discrimination layer; This represents the non-commitment category score output by the commitment type discrimination layer.

[0101] The commitment threshold is validated in increments of 0.01, ranging from 0.50 to 0.90, and the value with the highest recall rate that achieves a precision of at least 92% is selected. In this embodiment, the commitment threshold is set to 0.65.

[0102] When the probability of a candidate commitment is below 0.65, the corresponding candidate commitment statement is removed. For statements with a candidate commitment probability of 0.65 or higher, further checks are performed based on dependency parsing results to determine whether the negation word directly modifies the core predicate or the modal component corresponding to the core predicate. When negation structures such as "no commitment," "cannot provide," or "no response" dominate the core predicate, the corresponding statement is removed.

[0103] In cases where the bidding constraints themselves are prohibitive, such as prohibiting subcontracting, they will only be retained if the negative structure in the bid response statement is consistent with the semantic direction of the prohibitive constraint, thereby avoiding the misinterpretation of our commitment not to subcontract as a negative response.

[0104] After completing the above processing, in the first... In the bid document, the response to the first The precise commitment statements of each tender constraint clause are denoted as When there are multiple valid response statements in the same bid document, the statement with the highest matching score and passed commitment determination will be identified as the main commitment statement, and the remaining statements will be saved together with the main commitment statement as related commitment statements.

[0105] Compared to retrieval methods that only use sentence vector similarity, this step introduces keyword co-occurrence verification of core predicates and constraint parameters, and further utilizes dependency syntax relations to eliminate negative responses. This approach can prevent statements with similar content themes but opposite commitment directions from entering the subsequent counterfactual transformation process.

[0106] S13. Identify the conditional and exception scopes in precise commitment statements.

[0107] Semantic role labeling and dependency analysis are performed on precise commitment statements to identify conditional and exception scopes.

[0108] S131. Scope triggers the pre-construction of the thesaurus.

[0109] Conditional conjunctions, conditional phrases, and exception exclusion words are pre-collected from manually annotated tender terms and bid response statements. Two annotators each determine whether candidate trigger words have stable conditional guidance or exception exclusion functions, and a third annotator handles inconsistent results.

[0110] The scope trigger dictionary should at least record the trigger word, trigger type, common syntactic positions, dependency relations, and scope location rules. An example is shown below:

[0111] Table 2 Examples of Scope Triggering Dictionaries

[0112] If, in case, such as Conditional conjunction At the start of the subordinate clause Obtain the dependency subtree of the predicate connected by the trigger word When, at the time of... Temporal condition Before the main clause or in the middle of the main clause Obtain the dependency subtree corresponding to the temporal adverbial Under the condition of... State condition Before the core predicate Obtain the prepositional object and its modifiers Except for... Exception exclusion Before the main clause or in the middle of the main clause Obtain the object corresponding to "except" and its modifiers Unless Exception condition At the start of the subordinate clause Obtain the dependency subtree of the predicate connected by "unless" But... is not subject to this restriction Exception exclusion After the main clause Obtain the transition component and the excluded object Special circumstances excepted Exception exclusion At the end of the sentence Obtain the nominal component governed by "excepted"

[0113] S132. Conditional scope identification.

[0114] When a conditional conjunction appears in a precise commitment statement, the conditional role is first determined using semantic role annotation, and then the conditional predicate connected by the conditional conjunction is determined using dependency parsing. Using this conditional predicate as the root node, all its dependent descendant nodes are obtained, excluding nodes that are parallel to the main commitment predicate and not governed by the conditional conjunction.

[0115] Conditional scope , is represented as:

[0116] ;

[0117] in, This represents the set of character positions corresponding to conditional predicates, conditional conjunctions, and their dependent descendant nodes. Indicates the starting character position of the conditional scope; Indicates the position of the end character of the conditional scope; This indicates that text should be extracted according to a given range of character positions.

[0118] The statement "If the acceptance result is not satisfactory, we will complete the rectification within two working days" uses "if" as the conditional conjunction and "not satisfactory" as the conditional predicate, with the corresponding scope being "if the acceptance result is not satisfactory".

[0119] The statement "Under the condition that the tenderer requires early delivery, we will adjust the production plan" will be identified as a condition-guided structure under the condition that the tenderer requires early delivery. The scope of the condition will be determined as "Under the condition that the tenderer requires early delivery".

[0120] S133. Exception scope identification.

[0121] When an exception exclusion term appears in a precise commitment statement, first determine the excluded object governed by the exception exclusion term, then obtain the modifiers, quantitative modifiers, and parallel components of the excluded object. Exception Scope , is represented as:

[0122] ;

[0123] in, This represents the set of character positions corresponding to the exception excluder, the excluded object, and their dependent modifier nodes; Indicates the starting character position of the exception scope; Indicates the position of the end character of the exception scope.

[0124] The exception to the statement "Except for force majeure, we will complete the delivery on schedule" applies only to situations other than force majeure. The exception to the statement "We will bear all costs, but this does not apply to services temporarily added by the tendering party" applies only to services temporarily added by the tendering party.

[0125] When the same precise commitment statement exists in both conditional and exception scopes, both scopes are recorded separately. When a counterfactual transformation is subsequently performed on the conditional scope, the exception scope remains unchanged; conversely, when a counterfactual transformation is performed on the exception scope, the conditional scope remains unchanged, thus avoiding a single transformation altering two semantic factors simultaneously.

[0126] When scopes are nested, the outer and inner scopes are recorded separately according to their dependency hierarchy, with counterfactual transformations performed first on the innermost scope. This approach reduces the collateral impact of outer scope transformations on the semantic boundaries of inner scopes.

[0127] S14. Perform a reversible counterfactual transformation and construct a counterfactual constraint test set.

[0128] For precise commitment statements responding to the same tender constraints, perform a reversible counterfactual transformation according to the same transformation type to generate the transformed commitment statement, and record the transformation type, original word sequence, and associated reference position.

[0129] S141. Counterfactual transformation rule base.

[0130] A counterfactual transformation rule base is pre-established. This base consists of a conditional transformation rule table and an exception transformation rule table. Each rule records the triggering structure, transformation type, original expression pattern, alternative expression pattern, and reversibility check rules.

[0131] The conditional transformation rule is represented as follows:

[0132] Table 3 Examples of Condition Transformation Rules

[0133] If condition A is satisfied If condition A is not satisfied Condition inversion When state A is established When state A is not established Condition inversion Under the condition that matter A has been completed Under the condition that matter A has not been completed Condition inversion If condition A is met, perform action B Perform action B Condition removal Perform action B under condition A Perform action B Condition removal

[0134] The exception transformation rule is represented as follows:

[0135] Table 4 Examples of Exception Transformation Rules

[0136] Except for object A For all objects Exception generalization The case of object A is excepted No distinction between specific cases Exception generalization But object A is not subject to this restriction Under all applicable circumstances Exception generalization Except for object A, perform action B Perform action B Exception elimination Perform action B, except for object A Perform action B Exception elimination

[0137] Counterfactual transformation rules are derived from conditional and exception sentences in manually annotated corpora. Each rule undergoes both forward transformation and reverse recovery in the validation corpus. Only rules that can accurately recover the original word sequence without altering text outside their scope are included in the counterfactual transformation rule library.

[0138] S142. Reversible counterfactual transformation of conditional scope.

[0139] For each precise commitment statement under the same tender constraint clause, if the corresponding condition scope can find the corresponding condition flip rule in the counterfactual transformation rule base, the condition flip is performed on all condition scopes.

[0140] Condition flipping is performed by swapping true and false states, including pairwise substitutions such as satisfied-unsatisfied, existent-nonexistent, completed-incomplete, reached-not reached, and higher-than-not higher.

[0141] When any precise commitment statement cannot safely execute condition flipping, condition removal is uniformly executed on all precise commitment statements responding to the same tender constraint clause, avoiding different transformation types used in different tender texts.

[0142] The changed commitment statement is:

[0143] ;

[0144] in, This represents the transformed commitment statement; This represents a counterfactual transformation function that performs text replacement based on a uniform transformation type. This indicates the transformation type selected uniformly for the same bidding constraint clause; when performing condition transformation, the uniform transformation type is either condition reversal or condition removal.

[0145] When performing condition flipping, the counterfactual transformation function first locates the segment to be replaced based on the start and end character positions of the condition scope. Then, it selects a paired expression from the rule base that matches the core predicate and polarity state of the condition, replacing the state expression in the condition scope. When performing condition removal, it deletes conditional conjunctions, conditional clauses, and the corresponding punctuation marks, and corrects the initial letter position and punctuation connections of the remaining main clause.

[0146] S143. Reversible counterfactual transformations of exception scope.

[0147] When each precise commitment statement responding to the same tender constraint has a replaceable general scope expression, exception generalization is uniformly applied, replacing the specific exception object with a general scope expression such as all objects, all applicable situations, or no distinction between specific situations.

[0148] When any exception scope cannot form a grammatically complete general scope expression, exception removal is uniformly performed on all precise commitment statements responding to the same tender constraint clause, deleting exception exclusion words, exception objects and corresponding modifiers.

[0149] Exception transformation is also performed using the aforementioned counterfactual transformation function, with the uniform transformation type being either exception generalization or exception removal. After transformation, dependency parsing is used to re-check whether the transformed commitment statement still contains the core predicate and its constraint objects. If the core predicate or constraint objects are missing, the current rule is abandoned, and the exception removal rule is used for re-transformation.

[0150] S144. Reversibility recording and recovery verification.

[0151] A transformation record is generated each time a counterfactual transformation is performed. , is represented as:

[0152] ;

[0153] in, Indicates the starting character offset of the replaced scope in the original bid text; This indicates the offset of the replaced scope from the end character in the original bid text; This represents the original sequence of condition words or the original sequence of exception words saved before the transformation; This represents the sequence of replacement words written into the transformed commitment statement; Represents a summary of text content outside its scope.

[0154] Execute the recovery function on the transformed commitment statement:

[0155] ;

[0156] in, This represents the commitment statement recovered from the transformed record; This represents a reversible recovery function.

[0157] The reversible recovery function first checks whether the content summary of the text outside the scope is consistent with the transformation record. If the check passes, it deletes the replacement word sequence and writes back the original condition word sequence or the original exception word sequence at the corresponding character position. When the recovered commitment statement is completely consistent with the exact commitment statement, the current counterfactual transformation is determined to have passed the reversibility check; otherwise, the current counterfactual transformation is revoked, and the corresponding result is not written into the counterfactual constraint test set.

[0158] S145. Associated reference location record.

[0159] Each transformed commitment statement is associated with a reference point in the original tender text:

[0160] ;

[0161] in, Indicates the associated reference position of the transformed commitment statement; Indicates the first The text identifier of each bid document; The chapter number indicating the precise commitment statement.

[0162] Text identifiers are generated based on the procurement task number, the anonymous bidder's ID, and a summary of the bid document content. Chapter numbers are recorded according to the actual heading level of the bid text; for table content, the chapter number further records the table number, row number, and column number. Character offsets are calculated based on the original bid text before normalization, allowing for direct backtracking to the original document later.

[0163] S146. Counterfactual constraint test suite encapsulation.

[0164] The following test samples will be encapsulated: All modified commitment statements, modification types, and associated reference locations in response to the same tender constraints.

[0165] ;

[0166] in, Indicates that for the first Test samples generated from a set of tender constraints; Indicates that for the first The number of tender documents containing validly transformed commitment statements for each tender constraint clause.

[0167] All test samples are compiled as follows:

[0168] ;

[0169] in, Represents the counterfactual constraint test set; This means that the corresponding test sample will only be written into the counterfactual constraint test set if at least two bid documents respond to the same tender constraint clause and pass the reversibility check.

[0170] The minimum number of valid bid texts is set to 2 because with only one bid text, it is impossible to compare the semantic repair operations and cross-segment propagation order of different bid texts in subsequent steps. This test sample does not have the cross-text comparison conditions required for identifying bid rigging risks.

[0171] This invention does not directly modify the original bid text, nor does it randomly replace words. Instead, it uses the same bidding constraint clauses as the control benchmark and applies the same, reversible, and localizable counterfactual transformations to conditional or exceptional scopes with the same semantic function in different bid texts. The resulting counterfactual constraint test set can control for differences in transformation types, making the repair operations and cross-segment propagation order generated by subsequent semantic repair microservices comparable. At the same time, by saving the original word sequence, verifying the content summary, and recording the character position, it ensures that each transformation can restore and backtrack to the original bid text.

[0172] S2. Send the counterfactual constraint test set to multiple independent semantic repair microservices. Each semantic repair microservice restores the transformed commitment statement to a state compatible with the original bidding constraints. Record the semantic repair operations and cross-segment propagation order of non-mandatory response content during the restoration process. Associate the unusually consistent semantic repair operations and cross-segment propagation order among different bidding texts as a set of evidence of homology of repair paths.

[0173] S21. Construct independent semantic repair microservices and fully distribute the counterfactual constraint test suite.

[0174] Multiple independent semantic repair microservices are configured. Each semantic repair microservice adopts a different pre-trained language model infrastructure, independent training parameters, independent model version, and differentiated decoding strategy to avoid correlation deviations caused by multiple services sharing the same model weights or the same decoding path.

[0175] This embodiment sets up at least three semantic repair microservices, and their configuration examples are shown in Table 5.

[0176] Table 5. Example configuration of semantic repair microservice

[0177] First semantic repair microservice Encoder-decoder Transformer 12 encoder layers, 12 decoder layers, hidden layer dimension 768, 12 self-attention heads Input the bidding constraint clause, the transformed commitment statement and the associated context, and directly generate the repaired commitment statement Constrained beam search, with beam width set to 5 Second semantic repair microservice Autoregressive Transformer 24 decoding layers, hidden layer dimension 1024, 16 self-attention heads Serialize the repair task into an instruction prefix, and generate the repaired commitment statement word by word Nucleus sampling, cumulative probability set to 0.90, temperature set to 0.70 Third semantic repair microservice Bidirectional masked Transformer 12 encoding layers, hidden layer dimension 768, 12 self-attention heads Mask the tokens that violate the bidding constraint and their dependent associated tokens, and perform iterative filling Mask confidence ranking, with a maximum of 8 rounds of filling

[0178] The bundle width of the first semantic repair microservice was validated at 3, 5, and 8. When the bundle width was increased from 5 to 8, the improvement in bidding constraint compatibility was less than 0.4 percentage points, while the average inference time increased by more than 30%. Therefore, the bundle width was set to 5.

[0179] The cumulative probabilities of the second semantic repair microservice were validated at 0.85, 0.90, and 0.95, respectively. A value of 0.90 was chosen to minimize the rewriting rate of irrelevant content while maintaining candidate diversity. Temperature was validated at 0.50, 0.70, and 0.90, respectively. A value of 0.70 yielded the best balance between constraint compatibility and repair expression diversity, and was therefore chosen.

[0180] The maximum number of padding rounds for the third semantic repair microservice was tested with 4, 6, 8, and 10 rounds. When the number of padding rounds increased from 8 to 10, the proportion of unresolved conflict statements decreased by less than 0.2 percentage points, while the probability of error propagation increased. Therefore, the maximum number of padding rounds was set to 8.

[0181] S211. Pre-construction of the semantic repair training sample library.

[0182] The semantic repair training sample library is pre-built from historical tender documents and corresponding compliant bid response texts. First, a correspondence is established between the tender constraint clauses and the commitment statements that correctly respond to the tender constraint clauses. Second, according to the condition flipping, condition removal, exception generalization, and exception removal methods in S14, the correct commitment statements are reversibly transformed into counterfactual statements to form transformed commitment statements. Third, while retaining the original bid text writing habits, the annotation personnel restore the transformed commitment statements to repaired commitment statements that are compatible with the tender constraint clauses, and record other content that needs to be modified synchronously to maintain cross-section consistency.

[0183] The semantic repair training sample library contains at least 120,000 direct repair samples and 30,000 propagation repair samples with associated modifications. Data from different procurement projects are isolated according to project identifiers and then divided into training, validation, and test sets in an 8:1:1 ratio to prevent similar texts from the same procurement project from being included in both the training and test sets simultaneously.

[0184] The semantic repair training sample record is shown in Table 6.

[0185] Table 6 shows examples of semantic repair training sample records.

[0186] Bidding constraint clause Except for force majeure circumstances, the bidder shall complete the delivery within five working days Transformation type Exception elimination Transformed commitment statement We will complete the delivery within five working days Associated context Subsequent chapters specify that the delivery period will be extended when force majeure occurs Target repaired commitment statement Except for force majeure circumstances, we will complete the delivery within five working days Associated modification object Delay description in the delivery schedule Associated modification type Associated commitment coordination Trigger relationship Exception scope restoration triggers delay description restoration Source location Bid text identifier, chapter number and character offset

[0187] The three semantic repair microservices are trained independently using a semantic repair training sample library. The encoder-decoder model takes the tender constraints, the transformed commitment statement, and the associated context as encoder input, and the repaired commitment statement and the associated modification results as decoding targets. The autoregressive model concatenates the above content into a repair instruction sequence, and uses the target repair text as the subsequent prediction sequence. The bidirectional masking model sets the condition words, exception words, numerical words, and dependency words that need to be restored as mask positions, and uses the correct lexical units as training targets.

[0188] The training loss for each semantic repair microservice is:

[0189] ;

[0190] In the formula, Indicates the instance number of the semantic repair microservice; Indicates the first The training loss of the semantic repair microservice; This indicates the text generation loss after repair; This indicates the loss of semantic implication between the repaired text and the tender constraints. This indicates the text retention loss in non-modified regions; This represents the semantic implication loss coefficient; This represents the text preservation loss coefficient.

[0191] The semantic implication loss coefficient was validated with a step size of 0.05 between 0.10 and 0.50. A value of 0.35 resulted in the highest tender constraint compatibility without causing significant over-rewriting. Therefore, [the coefficient is used]. The value was set to 0.35. The text preservation loss coefficient was validated in steps of 0.05, ranging from 0.05 to 0.40. A value of 0.20 resulted in the lowest rate of incorrect rewriting of non-constrained content; therefore, it was set to... Set it to 0.20.

[0192] The model was trained using the AdamW optimizer, with the initial learning rate set to 2×10⁻ for both the encoder-decoder and bidirectional mask models. 5 The initial learning rate for the autoregressive model was set to 1×10⁻ 5 The batch size is set to 32, and the maximum training time is 8 epochs. Training stops when the validation set loss does not decrease for two consecutive epochs, and the model version with the highest constraint compatibility and the lowest non-constraint content error rewriting rate is saved.

[0193] After training, each semantic repair microservice registers a basic model summary, a fine-tuning parameter summary, a decoding strategy identifier, and a service instance identifier. Different semantic repair microservices do not share inference caches, candidate sequences, or intermediate attention results.

[0194] S212. Complete distribution of the counterfactual constraint test set.

[0195] Each semantic repair microservice is assigned an independent persistent consumption queue via a message middleware. Distribution messages are generated for each test sample in the counterfactual constraint test set.

[0196] ;

[0197] in, Indicates sending to the A test sample message for a semantic repair microservice; This indicates the distribution number of the corresponding test sample in the counterfactual constraint test set; This represents the total number of test samples in the counterfactual constraint test set; This indicates a complete summary generated based on the content of the test sample.

[0198] The message middleware uses a publish-subscribe approach, writing the same test sample to the persistent consumption queue of each semantic repair microservice. After receiving a message, the semantic repair microservice recalculates the integrity digest; it only returns an acknowledgment message if the recalculated integrity digest matches the integrity digest in the original message.

[0199] The confirmation waiting time is determined by multiplying the 99.9th percentile processing time obtained from message middleware stress testing by 1.5. The 99.9th percentile processing time obtained from the test is 18.6 seconds, which multiplied by 1.5 is 27.9 seconds. After rounding up, the confirmation waiting time is set to 30 seconds. If no confirmation message is received after 30 seconds, it will be resent, with a maximum of 3 resentments. The number of resentments was determined through fault injection testing. After three consecutive resentments, the recovery rate of instantaneous transmission failures exceeded 99.99%, and further increasing the number of resentments improved the recovery rate by less than 0.01 percentage points.

[0200] When a semantic repair microservice receives a distribution sequence number that covers 1 to... Only when all integrity summaries pass verification is the semantic repair microservice allowed to enter the repair processing phase. This ensures that each independent service instance receives the same and complete counterfactual constraint test set.

[0201] S22. Generate a post-repair commitment statement under the constraints of the tender.

[0202] After receiving the test sample, each semantic repair microservice reads the original text of the bidding constraint clause, the transformation type, the associated reference position, and the transformation record for each transformed commitment statement in the test sample.

[0203] Based on the location of related citations, the repair context is composed of the paragraph containing the transformed commitment statement, the two adjacent paragraphs before and after it, the table header and peer data of the table containing the statement, and other chapter content explicitly referenced by the statement. The scope of the two adjacent paragraphs before and after it is determined through a validation set: using one adjacent paragraph results in insufficient coverage of cross-paragraph related information; using three adjacent paragraphs significantly increases the proportion of irrelevant content; and using two adjacent paragraphs yields the highest combined result of cross-paragraph related recall and irrelevant content proportion.

[0204] S221. Constraint Repair Candidate Generation.

[0205] The first semantic repair microservice converts the core predicates, constraint objects, constraint parameters, conditional scope, and exception scope in the bidding constraints into decoded constraints. During the bundle search process, if a candidate sequence lacks necessary constraint parameters, changes the constraint direction, or generates a negation structure that conflicts with the bidding constraints, the expansion of the corresponding candidate sequence is stopped.

[0206] The second semantic repair microservice serializes the bidding constraints, transformed commitment statements, and repair context into instruction prefixes. It generates multiple candidate repair results through kernel sampling and performs constraint consistency checks upon completion of each semantic segment. Candidate results that fail the constraint consistency check do not proceed to the next decoding stage.

[0207] The third semantic repair microservice locates the positions of flipped, removed, generalized, or eliminated lexical units based on the transformation records, sets the corresponding positions and their dependent related words as masks, and fills them in descending order of mask prediction confidence. After completing one round of filling, dependency parsing is re-executed; if there are still missing conditional scopes, missing exception scopes, or conflicting constraint parameters, the next round of filling is initiated.

[0208] Each semantic repair microservice calculates the comprehensive evaluation value of the candidate repair results according to the following formula. , is represented as:

[0209] ;

[0210] in, This represents the conditional generation probability of the candidate repair results output by the corresponding semantic repair microservice; This indicates the repair context read from the original bid text; A semantic implication function representing the relationship between candidate repair results and tender constraints; A function representing the compatibility between candidate repair results and the repair context; A standardized edit distance function representing the candidate repair result and the transformed commitment statement; Indicates the generation probability weights; Indicates semantic implication weight; Indicates context-compatible weights; This indicates that the distance penalty weight has been rewritten.

[0211] The four weights were obtained through a validation set coordinate search, with an adjustment step of 0.05. A generation probability weight of 0.45 maintains text fluency; therefore, [the weight is set to -the value is missing here]. The semantic implication weight is set to 0.45; the constraint conflict false negative rate is lowest when the semantic implication weight is 0.30, therefore... Set to 0.30; the cross-segment conflict rate is lowest when the context compatibility weight is 0.20, therefore... Setting it to 0.20; a rewrite distance penalty weight of 0.05 can suppress irrelevant extensions without hindering necessary repairs, therefore... Set to 0.05.

[0212] The semantic implication function is executed by a natural language inference model. This model employs a 12-layer bidirectional Transformer encoder, which concatenates and encodes the bidding constraints with the candidate repair results. The output consists of three classes of probabilities: implication, neutrality, and conflict, after passing through a fully connected layer. The probability of the implication class is used as the output of the semantic implication function.

[0213] The compatibility function pairs the candidate repair results with each statement in the repair context, determines whether there are numerical, temporal, subject, conditional, or reference contradictions, and then outputs the proportion of context statements without contradictions.

[0214] The standardized edit distance function corresponds to the standardized edit similarity calculation in S24. First, the semantic repair operation sequences corresponding to the two bid texts to be compared are read, and standardized processing is performed according to unified rules on operation type, operation object category, numerical change direction, relative propagation position, and subject role. Second, a dynamic programming matrix is ​​established to transform one of the standardized semantic repair operation sequences into the other, and the cumulative cost corresponding to the insertion, deletion, and replacement operations is calculated sequentially. The cost of identical operations is zero, while different operations are assigned values ​​according to a preset edit cost table. The transformation path with the minimum cumulative cost is selected at each matrix position until the two standardized semantic repair operation sequences are fully aligned. Finally, the minimum cumulative edit cost is divided by the larger of the two sequences to obtain the standardized edit distance. The smaller this distance, the closer the two semantic repair operation sequences are.

[0215] S222. Acceptance of repair results.

[0216] Candidate repair results must simultaneously meet the following conditions:

[0217] ;

[0218] in, Indicates the semantic implication threshold; This indicates a context-compatible threshold.

[0219] The semantic entailment threshold was validated in steps of 0.01 between 0.60 and 0.95. The value with the highest recall was selected while maintaining a conflict resolution accuracy of at least 95%. The context compatibility threshold was validated in steps of 0.01 between 0.60 and 0.95. Under the condition that the false release rate due to cross-segment conflicts is no higher than 5%, the value with the highest recall rate was selected. .

[0220] The candidate repair result that simultaneously meets both thresholds and has the highest comprehensive evaluation value is determined as the post-repair commitment statement, denoted as... If none of the candidate results reach the two thresholds, the corresponding semantic repair microservice will mark the test sample as a repair failure and will not generate repair path homology evidence based on the test sample.

[0221] This step not only requires the repaired commitment statements to be compatible with the tender constraints, but also ensures that the repaired result is compatible with the numerical values, time limits, subjects, and cross-section references in the original tender text by repairing the context constraint generation process, thereby reducing the possibility of creating new contradictions in other chapters by only correcting a single sentence.

[0222] S23. Monitor related modifications and form a semantic repair operation sequence.

[0223] Each semantic remediation microservice is configured with a remediation operation monitor and a text patch executor. Semantic remediation microservices must not directly overwrite the original tender text; instead, they must first submit the proposed modifications to the text patch executor. The text patch executor generates an operation record before submitting the modifications, and only after the operation record passes verification is it applied to the corresponding text copy.

[0224] Direct repairs to the transformed commitment statement itself are not considered incidental modifications. For content outside the scope of the transformed commitment statement and not modified due to a direct violation of bidding constraints, adjustments made to maintain internal textual consistency, numerical consistency, cross-paragraph reference coherence, or consistency of related commitments are considered incidental modifications.

[0225] The objects that are modified include the main text statements, table data items, time nodes, quantity parameters, reference numbers, pronoun references, and associated commitments.

[0226] S231. Semantic Repair Operation Type Library.

[0227] The semantic repair operation type library is pre-built based on manual repair records in the semantic repair training sample library. First, the actual editing operations are collected; second, they are categorized according to the operation purpose, target, and triggering reason; third, two annotators confirm the operation type, and a third annotator handles inconsistent results.

[0228] Examples of semantic repair operation type libraries are shown in Table 7.

[0229] Table 7 Examples of Semantic Repair Operation Types

[0230] OP01 Polarity repair Affirmative word, negative word or modal word Inconsistent polarity of associated statements after conditional reversal and restoration OP02 Conditional recovery conditional conjunctions and conditional clauses Main commitment condition scope restoration OP03 Exception recovery Exception exclusion terms and exception objects Main commitment exception scope recovery OP04 Numerical Coordination Numerical value, unit or range Changes in the main commitment parameters caused inconsistencies in the table values. OP05 Time coordination Date, deadline, or time sequence The delivery deadline after the fix conflicted with the schedule. OP06 Reference target repair Chapter number, clause number, or form number The original reference becomes invalid after the referenced content is restored. OP07 Refers to repair Pronouns or omitted subjects The subject's reference became ambiguous after the repair. OP08 Related commitment coordination Other commitment statements Logical conflicts arise in related commitments after the primary commitment is restored. OP09 Data item synchronization Table cell or list field After the main text is repaired, the corresponding data items need to be synchronized.

[0231] For each associated modification, a semantic repair operation record is generated:

[0232] ;

[0233] in, Indicates the first The semantic repair microservice records the first semantic repair statement. One semantic repair operation; Indicates the sequence number of the semantic repair operation; Indicates the operation type code; Indicates the position of the object being operated on within the original tender text; Indicates the differences in text before and after the modification; Represents the logical timestamp of the operation; This represents the set of identifiers of the preceding operations that triggered the current operation; Indicates the content category of the object being operated on.

[0234] The location of the operation object is recorded according to the tender text identifier, chapter number, paragraph number, table row and column position, and start and end character offset. Text differences should at least record deleted content, inserted content, and replaced content. The content category of the operation object should at least be set as body text statements, data items, time items, reference items, and related commitments.

[0235] S232. Logical timestamp determination.

[0236] To avoid the impact of clock discrepancies between different servers on the cross-segment propagation order, this embodiment does not directly use physical clocks to determine the order of operations, but instead uses logical timestamps based on trigger relationships:

[0237] ;

[0238] in, Indicates the sequence number of any preceding operation that triggered the current operation; Represents the logical timestamp of the corresponding preceding operation; This indicates that there is no preceding triggering operation.

[0239] When a semantic repair operation is triggered by multiple preceding operations, its logical timestamp is the maximum value of the logical timestamps of all preceding operations plus 1. For operations with the same logical timestamp that are independent of each other, they are arranged from front to back according to their original text positions; if the original text positions are also the same, they are arranged according to the lexicographical order of the operation identifiers, thus obtaining a deterministic sequence of semantic repair operations.

[0240] S24. Construct a cross-segment propagation order and identify abnormally consistent repair propagation patterns.

[0241] S241. Construct the cross-segment propagation order.

[0242] For each corrected commitment statement, the corresponding semantic correction operation is treated as a node, and the triggering dependencies between operations are treated as directed edges, constructing a directed acyclic graph:

[0243] ;

[0244] in, Indicates the first A semantic repair microservice constructs a cross-segment propagation order for the corresponding bid text; This represents the set of nodes for semantic repair operations. This represents the set of triggering dependency edges between semantic repair operations.

[0245] When the set of identifiers of the preceding operations of a certain operation records the identifier of another operation, and the logical timestamp of the preceding operation is less than the logical timestamp of the following operation, a directed edge is established between the two operations. Since each directed edge points from the smaller logical timestamp to the larger logical timestamp, a closed loop will not be formed.

[0246] After the cross-segment propagation order is established, nodes are converted into a semantic repair operation sequence according to logical timestamps and deterministic sorting rules. Sequence standardization processing includes: converting synonymous modification actions into unified operation codes; converting specific values ​​into numerical types and directions of change; converting absolute chapter numbers into positional relationships relative to the main commitment statement; replacing specific subject names with subject roles; and replacing specific times with time types and sequence relationships.

[0247] Even when different tender documents differ in specific words, chapter numbers, and numerical values ​​after standardization, it is still possible to compare their repair operation objectives and dissemination structures.

[0248] S242. Standardized editing similarity calculation.

[0249] For two different bid texts responding to the same tender constraints, their standardized semantic repair operation sequences are compared using weighted edit distance:

[0250] ;

[0251] in, This indicates that two different bid texts are in the first... Standardized edit similarity under a semantic repair microservice; and Indicates an index for two different bid documents; and These represent the standardized semantic repair operation sequences corresponding to the two bid texts; This represents the weighted edit distance function; and These represent the lengths of the two standardized semantic repair operation sequences, respectively.

[0252] The weighted edit distance function is implemented using dynamic programming. When transforming a normalized sequence of semantic repair operations into another sequence, the costs corresponding to insertion, deletion, and replacement operations are accumulated, and the transformation path with the minimum total cost is selected.

[0253] Examples of editing costs are shown in Table 8.

[0254] Table 8. Examples of editing costs for standardized semantic repair operations

[0255] The operation type, object category, and propagation location are all the same. 0 Same operation type but different object category 0.35 The operation types belong to the same coordination category but are different in specific terms. 0.60 The operation types are completely different 1.00 Insert an operation 0.80 Delete an operation 0.80

[0256] The replacement costs for operations of the same type but different object categories were validated at 0.20, 0.35, and 0.50, respectively. A value of 0.35 was chosen to distinguish between homologous repairs and accidental similar repairs. The replacement costs for different operation types within the same coordination category were validated at 0.40, 0.60, and 0.80, respectively. A value of 0.60 resulted in the highest F1 score, so this value was chosen. The replacement costs for completely different operation types were set to 1.00 as a normalization baseline. Insertion and deletion costs were validated between 0.60 and 1.00 with a step size of 0.10. A value of 0.80 provided appropriate tolerance for missing individual related modifications and did not amplify short sequence similarity; therefore, both were set to 0.80.

[0257] S243. Calculation of graph structural isomorphism.

[0258] The VF2 matching algorithm with node label constraints is used to search for the maximum common isomorphic subgraph in two cross-segment propagation orders. A node correspondence is only allowed when two nodes have the same operation type, object category, and positional relationship relative to the main commitment statement.

[0259] The graph structural isomorphism is:

[0260] ;

[0261] in, Indicates the graph structural isomorphism between two cross-segment propagation orders; The set of nodes representing the maximum common isomorphic subgraph; Let f(x) represent the set of directed edges of the greatest common isomorphic subgraph. Indicates the number of elements in the set; The weights represent the proportion of nodes that are isomorphic.

[0262] The node isomorphic weights were validated with a step size of 0.1, ranging from 0.2 to 0.8. A weight of 0.4 preserves the contribution of operational node consistency while also giving higher weights to propagation edges with causal significance; therefore, [the weight is determined]. When set to 0.4, the weight of the corresponding edge isomorphism ratio is 0.6.

[0263] S244. Judgment of abnormal consistency conditions.

[0264] Two bid documents are deemed to have an abnormally consistent repair propagation pattern when the following conditions are met simultaneously:

[0265] ;

[0266] in, This represents the standardized edit similarity threshold; The threshold for graph structure isomorphism.

[0267] The standardized editing similarity threshold was validated with a step size of 0.01 between 0.50 and 0.95. The value with the highest recall rate was selected while maintaining a false positive rate of no more than 5%. The graph structure isomorphism threshold was independently validated within the same search interval. The value with the highest recall was selected, provided the false positive rate was no higher than 5%. .

[0268] To avoid accidental high similarity when only one or two related modifications are involved, both cross-segment propagation sequences participating in the comparison must have at least three semantic repair operation nodes, and the nodes must involve at least two content categories. The minimum number of nodes was tested with 2, 3, and 4 respectively. With 2 nodes, the accidental consistency rate was relatively high; with 4 nodes, the false negative rate increased significantly; and with 3 nodes, the overall result was the best. Therefore, the minimum number of nodes was set to 3.

[0269] This step does not use the literal similarity of the repaired text as the basis for determining common origin, but rather compares the semantic repair operation sequence and the triggering dependency structure between operations simultaneously. Even if different bidding texts have undergone synonym rewriting, paragraph adjustment, or numerical replacement, as long as their internal commitment relationships originate from the same draft, they may still produce similar chain repair sequences and cross-segment propagation structures after undergoing the same counterfactual transformation.

[0270] S25. Generate and compile evidence of homology of the repair path.

[0271] For different bid texts responding to the same tender constraints, pairwise comparisons are performed within the same semantic repair microservice. Anomaly consistency relationships are established between bid texts that meet the anomaly consistency criteria.

[0272] When only two bid texts meet the abnormal consistency condition, they are directly identified as a common candidate group; when more than two bid texts meet the abnormal consistency condition, the maximum clique search method is used to determine the bid text group, requiring that any two bid texts in the group simultaneously reach the standardized editing similarity threshold and the graph structure isomorphism threshold, so as to avoid incorrectly classifying two dissimilar bid texts into the same group through a single intermediate bid text.

[0273] For each homology candidate group, generate one piece of homology evidence for the repair path:

[0274] ;

[0275] in, Indicates the first Each semantic repair microservice generates evidence of the homology of repair paths for homology candidate groups. This represents the set of bid text indexes within a shared candidate group. This represents the instance identifier of the corresponding semantic repair microservice; This indicates a standardized semantic repair operation pattern shared by homologous candidate groups; This represents a cross-segment propagation subgraph shared by homologous candidate groups; This represents the similarity records among the bid texts within the same candidate group; It indicates the source location, operation timestamp, model version, and transformation record related to the evidence.

[0276] A common standardized semantic repair operation pattern is obtained by performing multi-sequence alignment on each operation sequence within the group, preserving the operation types and relative order that appear in all sequences. A common cross-segment propagation subgraph is obtained by iteratively performing maximum common isomorphic subgraph matching on the cross-segment propagation order within the group, preserving the operation nodes and triggering dependency edges common to all bidding texts.

[0277] Each repair path homology evidence is associated with the test sample identifier, the unique identifier of the bidding constraint clause, the bid text identifier, the semantic repair microservice instance identifier, the basic model summary, the decoding strategy identifier, the post-repair commitment statement, the semantic repair operation record, the common cross-segment propagation subgraph, and the original text location.

[0278] After all semantic repair microservices have completed their processing, evidence of homology of repair paths is aggregated according to the following formula:

[0279] ;

[0280] in, This represents a set of evidence indicating the homology of repair paths; Indicates the first A set of evidence of homology of repair paths output by each semantic repair microservice; This indicates the total number of semantic repair microservices participating in this repair process.

[0281] During the aggregation process, evidence generated by different semantic repair microservices for the same tender text group is not merged. Instead, the service instance identifier, model version, and decoding strategy corresponding to each piece of evidence are retained. This enables subsequent collaborative review services to deactivate semantic repair microservices one by one, regenerate a set of evidence for the commonality of repair paths, and determine whether the risk results depend on a specific model.

[0282] This step transforms the traditional judgment of text similarity in the result into judgment of homology in the repair process by independently repairing heterogeneous models, monitoring related modifications, sorting logical time, comparing sequence and causal structure, and solidifying evidence sources. It can identify common compilation relationships masked by synonym rewriting, word order adjustment, and local rewriting, while reducing the possibility of bias in a single language model directly leading to the risk of collusion.

[0283] S3. Send the repair path homology evidence set to the collaborative review service. The collaborative review service will disable the semantic repair microservices that participated in generating the repair path homology evidence set one by one, and call the remaining semantic repair microservices to regenerate the repair path homology evidence set. When the regenerated repair path homology evidence set continuously points to the same bidding entity and can be traced back to the associated reference position, output the collusion risk identification result.

[0284] S31. Construct a microservice-evidence mapping table.

[0285] The collaborative review service subscribes to the topic of repair path homology evidence in the message middleware and receives the repair path homology evidence set generated by S2. The message middleware adopts a publish-subscribe model, sending the repair path homology evidence set as an indivisible batch message to the collaborative review service, and writing the evidence set batch identifier, counterfactual constraint test set version, generation time, semantic repair microservice configuration version, and message verification value into the message header.

[0286] After receiving the batch message, the collaborative review service first verifies the integrity of the evidence set for the same origin of the repair path based on the message verification value, and then performs an idempotency check based on the evidence set batch identifier. If the same evidence set batch identifier already exists in the local review task library, the review task will not be created again; if it does not exist, the corresponding collaborative review task will be created, and the version of the counterfactual constraint test set, the configuration version of the semantic repair microservice, and the anomaly consistency judgment threshold in S2 used in this collaborative review will be frozen to prevent the introduction of additional variables due to changes in test samples or judgment conditions during the process of disabling one by one.

[0287] For each piece of evidence related to the same source of repair paths in the evidence set, the collaborative review service reads the semantic repair microservice instance identifier, bidding constraint identifier, set of bidding entities, common semantic repair operation mode, cross-segment propagation subgraph, and associated reference position associated with the evidence, and establishes an inverted index mapping according to the semantic repair microservice instance identifier to obtain the microservice-evidence mapping relationship:

[0288] ;

[0289] in, This represents a semantic repair microservice instance. The corresponding set of evidence for the homology of the repair path; This represents a piece of evidence of the homology of repair paths within a set of evidence of the homology of repair paths; This function represents the instance identifier extraction function for reading the semantic repair microservice instance identifier from the homology evidence of the repair path.

[0290] The instance identifier extraction function sequentially reads the service registration identifier, model version identifier, and decoding strategy identifier from the evidence of origin of the repair path. It then concatenates these three identifiers and verifies their consistency with the instance record in the service registry. Only when all three match is the evidence assigned to the corresponding semantic repair microservice instance. If the service registration identifier matches but the model version identifier or decoding strategy identifier differs, they are considered different semantic repair microservice instances to ensure that subsequent deactivation operations can accurately isolate the specific repair mechanism that generated the evidence.

[0291] The collaborative review service establishes a microservice-evidence mapping table based on the semantic repair microservice instance identifier and evidence identifier. This table is automatically generated using the homology evidence of the repair path output by S2, without relying on manual presets.

[0292] Table 9. Examples of Microservice - Evidence Mapping

[0293] SRV-A-01 Encoder-decoder model A Constraint search EV-001 TC-018 <![CDATA[{B1,B3}]]> SG-001 LG-006 SRV-B-02 Masked Language Model B Iterative mask filling EV-002 TC-018 <![CDATA[{B1,B3}]]> SG-004 LG-006 SRV-C-01 Generative Language Model C Limited sampling EV-003 TC-027 <![CDATA[{B2,B4,B5}]]> SG-009 LG-011

[0294] The basic model summary is derived from the model structure and version registered by the semantic repair microservice in the service registry. The associated reference location group identifier is obtained by hashing the bid text identifier, chapter number, and start and end character offsets recorded in S1 in a fixed order. This hash value is only used for fast indexing; the unhashped associated reference location is still used during actual backtracking.

[0295] The collaborative review service further extracts all semantically repaired microservice instances involved in the microservice-evidence mapping table, forming a sequence of instances to be ablated. The sequence of instances to be ablated is arranged in lexicographical order of instance identifiers to eliminate the randomness of execution order; at the same time, it saves the registration information, running status, model version, decoding strategy, and message consumption position of each instance before it was deactivated, for subsequent use in registration restoration and audit reproduction.

[0296] S32. Deactivate semantic repair microservice instances one by one and regenerate repair path homology evidence.

[0297] The collaborative review service employs a single-instance, round-by-round deactivation approach for ablation review. Only one semantic repair microservice instance is deactivated in each round. After evidence regeneration and result solidification are completed in that round, the instance is restored before proceeding to the next deactivation round. This ensures that only one controlled variable exists across different rounds: the currently deactivated semantic repair microservice instance.

[0298] At the start of the r-th deactivation round, the collaborative review service sends a deactivation command to the service registry, carrying the instance identifier, review task identifier, and registration version number. The service registry changes the service status of the corresponding instance from distributable to review deactivated, cancels its message consumption route, and returns a deactivation confirmation message. The deactivation only removes the instance's service registration and message consumption permissions; it does not delete model files, runtime logs, or intermediate evidence, so that the original running state can be restored after the review is completed.

[0299] To prevent test samples that had already entered the message queue before the deactivation command took effect from being processed by the deactivated instance, the collaborative review service executes a message emptying barrier after receiving the deactivation confirmation message. The service registry reads the last confirmed consumption position of the deactivated instance, and the message middleware reads the number of pending messages for the corresponding consumer of that instance; deactivation is considered complete only when the number of pending messages is zero and the last confirmed consumption position of the instance no longer changes.

[0300] Message queuing wait time is set as follows:

[0301] ;

[0302] in, Indicates the message queue waiting time; This represents the 99th percentile of the processing time of the most recent 1,000 test samples in the semantic repair microservice before decommissioning. This indicates the processing time for a single test sample from being received by the semantic repair microservice to generating the repair result; Indicates the minimum waiting time; This indicates a fixed redundancy time for service registration status propagation and message consumption point synchronization.

[0303] The minimum waiting time is obtained by measuring the service registration propagation time in the test environment. Specifically, one hundred instance deregistration operations are performed consecutively, and the time from the issuance of the deactivation command to all message distribution nodes ceasing to route messages to that instance is recorded. When the measured maximum propagation time does not exceed twenty seconds, a ten-second redundancy is added to this maximum value, and the minimum waiting time is set to thirty seconds. If the measured maximum propagation time in the actual deployment environment exceeds twenty seconds, the minimum waiting time is adjusted to this maximum propagation time plus ten seconds.

[0304] Example After the deactivation is completed, the set of instances eligible for semantic repair in this round is:

[0305] ;

[0306] In the formula, Indicates that the instance is disabled. The set of semantic repair microservice instances that will continue to participate in this round of review; This represents the set of instances to be ablated, extracted from the microservice-evidence mapping table.

[0307] The collaborative review service republishes the counterfactual constraint test suite generated by S1 and whose version has been frozen to the message middleware, and uses a broadcast distribution method to ensure that every other instance receives all test samples. To ensure that the regenerated results can be directly compared with the results before decommissioning, each test sample continues to use the test sample identifier, tender constraint identifier, transformed commitment statement, transformation type, and associated reference position generated by S1, without re-performing the counterfactual transformation.

[0308] The expected number of test sample confirmations in this round , is represented as:

[0309] ;

[0310] in, This represents the counterfactual constraint test set generated by S1; This indicates the number of elements in the set.

[0311] Only after the actual number of confirmed responses received equals the expected number, and all non-discontinued instances return repair results for all test samples, does the collaborative review service initiate the regeneration of homology evidence for this round of repair paths. The regeneration process continues to use the standardized edit similarity threshold and graph structure isomorphism threshold determined by S2, without readjusting the thresholds due to the discontinuation of a particular instance, thus avoiding simultaneous changes in ablation variables and judgment criteria.

[0312] If the actual number of confirmed responses received within the specified processing time limit is less than the expected number, this round will be marked as incomplete distribution, will not be counted in the cumulative number of bids for any bidder, and this round of testing will be re-executed. The specified processing time limit is set according to the highest 99th percentile of the historical processing time of each non-disabled instance.

[0313] ;

[0314] in, This indicates the specified processing time limit for a single deactivation round; Representation of instances Batch processing time required to complete the processing of the entire counterfactual constraint test set; This indicates the fixed redundancy time for evidence gathering, message confirmation, and database writing.

[0315] The 60-second redundancy time was obtained by measuring the evidence gathering time, message confirmation time, and database write time separately. In one hundred consecutive complete test batches, the maximum total time for the three operations was 47 seconds. After rounding up to the nearest ten seconds and adding a 10-second safety margin, it was set to 60 seconds.

[0316] The same anomaly consistency analysis process as S2 regenerates the current cycle's repair path homology evidence set. , is represented as:

[0317] ;

[0318] in, This indicates the function for regenerating evidence.

[0319] The evidence regeneration function distributes the counterfactual constraint test set in sequence and receives the semantic repair operation sequence and cross-segment propagation order re-output by each instance. It compares the operation standardization rules, weighted edit distance function and VF2 matching algorithm with node label constraints that are the same as those in S2. Then, the bidding texts that simultaneously meet the standardized edit similarity threshold and graph structure isomorphism threshold are re-associated as evidence of homology of repair paths.

[0320] Table 10 shows examples of deactivation review records.

[0321] AR-001 SRV-A-01 10:20:31 2 240 240 efficient EVS-A01 AR-002 SRV-B-02 10:43:18 2 240 240 efficient EVS-B02 AR-003 SRV-C-01 11:06:42 2 240 238 Incomplete distribution — AR-003-R1 SRV-C-01 11:13:27 2 240 240 efficient EVS-C01

[0322] After each round of results is finalized, the collaborative review service sends a recovery command to the service registry, using the registration information saved before the start of this round to restore the service registration, model version, decoding strategy, and message consumption configuration of the disabled instance. Only after the service registry confirms that the instance has been restored to a distributable state will the collaborative review service begin the next round of deactivation operations.

[0323] S33, Consistency analysis of execution evidence and backtracking of related citation positions.

[0324] For each original repair path homology evidence generated by the instance before it was deactivated, the collaborative review service reads its bidding constraint identifier, bidding entity set, cross-segment propagation order of anomalies, and associated reference position from the microservice-evidence mapping relationship; then it searches for new repair path homology evidence with the same bidding constraint identifier in the regenerated repair path homology evidence set.

[0325] The collaborative review service does not directly use the literal similarity of the reconstructed commitment statements as the basis for ablation review. Instead, it simultaneously performs the judgment of the intersection of bidding entities and the backtracking of cross-segment propagation order. For any original reconstruction path homology evidence and new reconstruction path homology evidence, the intersection of bidding entities... , is represented as:

[0326] ;

[0327] in, This indicates evidence of homology of the regenerated repair path; This represents the function for extracting the set of bidding entities.

[0328] The specific scheme for the function to extract the set of bidders is as follows: First, it receives a piece of evidence of origin for the repair path and reads the data structure, evidence identifier, and candidate group fields of the evidence. Based on the set of bid text indexes recorded in the evidence of origin for the repair path, it queries the correspondence between each bid text index and the bidder identifier. It then performs format standardization and validity checks on the obtained bidder identifiers, removing duplicate identifiers, empty identifiers, and abnormal identifiers that cannot correspond to the original bid text. Finally, it sorts the bidder identifiers according to a fixed order and outputs the set of bidders pointed to by the evidence of origin for the repair path. This set is used to calculate the intersection of bidders between the evidence before deactivation and the regenerated evidence.

[0329] When the intersection of the bidding entities is empty, the two do not constitute evidence of consistency; when it is not empty, the collaborative review service further performs reverse backtracking on the cross-segment propagation order in the evidence of homology of the new repair path.

[0330] The reverse backtracking starts from the terminal semantic repair operation node in the cross-segment propagation subgraph, and visits the preceding nodes layer by layer in the reverse direction of the triggering dependency edges until the root operation node with no preceding triggering node is reached. For each node visited, the position of the operation object, the modified statement identifier, the upstream reference statement identifier, and the original text position are read. If the operation object belongs to intermediate text generated during the repair process, backtracking continues according to the source field of that operation; if the operation object directly corresponds to the original bidding text, a backtracking position triplet is generated:

[0331] ;

[0332] in, Represents a backtracking position triple; Indicates the tender text identifier; Indicates the chapter number; and These represent the starting character offset and ending character offset of the statement in the original bid text, respectively.

[0333] The collaborative review service sorts all the position triples obtained from the backtracking process according to the bid text identifier, chapter number, and starting character offset to obtain the backtracking position set of the new evidence. Since S1 has already recorded the associated reference positions in the original bid text, and this step uses the same version of the frozen bid text, position matching adopts exact triple matching without setting a fuzzy position threshold. Only when the bid text identifier, chapter number, starting character offset, and ending character offset are all the same, is it determined that the abnormally consistent cross-segment propagation order in the new evidence can be traced back to the associated reference position recorded by S1.

[0334] Regarding the bidding entities and the rounds of discontinuation, the status of retention of homogeneity characteristics is as follows: , is represented as:

[0335] ;

[0336] in, This indicates the function that backtracks to the associated reference position based on the triggering dependency edge; The set of S1-related reference locations representing the evidence of homology in the original repair path; Represents the empty set. This describes the function for extracting tender constraint identifiers. The specific approach is as follows: It receives the evidence of homology of the repair path to be analyzed and reads the test sample identifier associated with that evidence. Based on the test sample identifier, it queries the counterfactual constraint test set generated by S1 to obtain the unique identifier of the tender constraint clause corresponding to the test sample. Next, it verifies the consistency between the unique identifier of the queried tender constraint clause and the fixed tender constraint identifier in the evidence of homology of the repair path. Finally, when the two identifiers match, it outputs the tender constraint identifier. When the two identifiers are inconsistent, missing, or cannot correspond to a tender constraint clause, the corresponding evidence is marked as having an abnormal constraint source and is not included in subsequent evidence consistency analysis.

[0337] The related reference position backtracking function runs in the following order: reading the end node—reverse traversal triggering dependent edges—parsing the node source field—locating the original bid text—outputting the position triplet—performing precise position verification. During the traversal, a set of visited nodes is set; a node is added to this set upon its first visit, and subsequent visits to the same node are not repeated to prevent abnormal data from causing loop processing. If the integrity check detects a dependency cycle, the corresponding evidence is marked as structural anomaly and is not included in the cumulative count.

[0338] when When the shutdown operation fails to eliminate the homology characteristics of the corresponding bidding entities, it is determined that the other bidding entities in this round, the bidding constraint identifier, the associated reference location group, the original evidence identifier, the new evidence identifier, and the semantic repair microservice instance that was not shut down are recorded. This result indicates that the abnormal consistency is not solely caused by a single pre-trained language model or a single decoding strategy that was shut down.

[0339] This step employs a dual constraint: the intersection of bidding entities is non-empty, and the cross-segment propagation order can be traced back to the original associated reference position. Existing text consistency reviews typically only compare risk scores before and after discontinuation or the similarity of repaired texts, easily mistaking shared language preferences of the model for signs of collusion. This step requires new evidence to return to the same original text position in the S1 record along the triggering dependency of semantic repair operations, transforming the review object from the surface similarity of generated texts to the same response structure in the original bidding text, thereby eliminating accidental consistency caused solely by language model expression habits.

[0340] S34. Count the cumulative number of homology features and generate the risk identification results of bid rigging.

[0341] After completing the deactivation and evidence regeneration of all semantic repair microservice instances, the collaborative review service aggregates the source identity retention status of each valid round according to the bidding entity and associated reference location group. For bidding entities in the deactivation round... The set of related reference locations that can be traced back is first sorted according to the tender text identifier, chapter number, start character offset, and end character offset, and then the sorting result is used as the related reference location group. This ensures that records with different positional orders but identical actual positions can be grouped into the same group.

[0342] Cumulative number of bids by the same related reference location group , is represented as:

[0343] ;

[0344] in, Evidence indicating the similarity of the original repair path involves the bidding entity. All valid deactivation rounds; This indicates an indicator function; it takes the value of one if the condition within the parentheses is true, and zero otherwise. Indicates the bidding entity During the decommissioning round The associated reference position group obtained by backtracking.

[0345] Number of valid shutdown rounds for bidding entities , is represented as:

[0346] ;

[0347] The threshold for the cumulative number of homology features is set as follows:

[0348] ;

[0349] in, Indicates the bidding entity The corresponding threshold for the cumulative number of homology features; Indicates the ablation retention ratio threshold; This represents the function for rounding up.

[0350] The specific scheme for the ablation retention ratio threshold is as follows: Historical projects whose conclusions have been confirmed by bidding supervision personnel are selected as the validation set. Positive samples are combinations of bidding entities that share a common source of data or have engaged in bid rigging, while negative samples are combinations of bidding entities that have been independently prepared and verified. Data is divided into projects to prevent the same project from appearing simultaneously in both the threshold selection data and the validation data. Candidate ratios are validated one by one in increments of 0.05 between 0.50 and 0.90. The candidate value with the highest recall rate is selected from those with a false positive rate not exceeding 5%. During the validation process, the false positive rate corresponding to 0.60 was 4.3%, and the recall rate was higher than other candidate values ​​that met the false positive rate constraint. Therefore, the ablation retention ratio threshold was set to 0.60. When historical confirmation data is lacking, 0.60 is used as the initial value, and the value is recalibrated using the same method after accumulating no fewer than fifty manually reviewed projects.

[0351] The minimum threshold of 2 is used to prevent a risk conclusion from being drawn simply because only one non-deactivated instance accidentally reproduces the original evidence. The semantic repair microservice system deploys at least three independent instances using different pre-trained language models or different decoding strategies. Therefore, any bidding entity can only enter the risk identification result generation stage if it retains homology characteristics in at least two valid deactivation rounds.

[0352] Bidding entities The main associated reference location group , is represented as:

[0353] ;

[0354] When multiple related reference location groups have the same maximum cumulative count, the location group with more bidding constraint clauses is selected; if they are still the same, the location group with the smaller sum of starting character offsets is selected. This fixed selection rule is only used to ensure the uniqueness of the output result and does not change the cumulative count.

[0355] The maximum cumulative number of times a bidder can trace back to the same associated reference location group. , is represented as:

[0356] ;

[0357] The collaborative review service generates the risk identification results for bid rigging among the corresponding bidding entities when the following conditions are met:

[0358] ;

[0359] The set of other bidding entities that share a common origin with the bidding entity is as follows:

[0360] ;

[0361] in, Indicates the relationship with the bidding entity The set of other bidding entities that share a common origin; Indicates the bidding entity The homology feature retention status is 1 and the associated reference position group is equal to the set of deactivated rounds of the main associated reference position group; and These represent the rounds of discontinuation. Evidence of homology between the original repair path and the new repair path.

[0362] To calculate the risk confidence level, the homology characteristic ablation retention rate is first calculated. , is represented as:

[0363] ;

[0364] Next, calculate the concentration of related reference locations:

[0365] ;

[0366] in, Indicates the bidding entity The concentration of related reference locations; when the denominator is zero, it is set to zero.

[0367] For each other bidding entity in the homology-related entity set, count the number of valid deactivation rounds in which it co-occurs with the current bidding entity under the main association reference position group, and take the minimum proportion as the entity association stability:

[0368] ;

[0369] in, Indicates the bidding entity Stability of the entity's relationship with other bidding entities; This refers to other bidding entities within the set of entities with similar origins.

[0370] The final risk confidence level was calculated using a conservative aggregation method:

[0371] ;

[0372] in, Indicates the bidding entity The risk confidence level ranges from zero to one.

[0373] Using a minimum function allows the risk confidence level to be simultaneously constrained by the ablation retention rate, the concentration of related citation locations, and the stability of subject associations, without the need for manually setting weighting coefficients. Only when all three indicators are high will the risk confidence level be high, thus avoiding the problem of an abnormally high one indicator masking deficiencies in other dimensions of evidence.

[0374] The risk identification results of bid rigging generated by the collaborative review service shall at least record the bidder identifier, the identifiers of other bidders with the same source, the bidding constraint identifier, the list of related reference locations and the risk confidence level; it may also further record the main related reference location group, the number of effective deactivation rounds, the maximum cumulative number of times in the same location, the cumulative number threshold, the identifier of the semantic repair microservice instance participating in the review, the same source evidence identifier of the original repair path and the same source evidence identifier of the new repair path.

[0375] Table 11 Examples of Results for Identifying Risks of Bid Collusion

[0376] RR-001 <![CDATA[B1]]> <![CDATA[B3]]> TC-018 LG-006 3 3 2 0.93 RR-002 <![CDATA[B3]]> <![CDATA[B1]]> TC-018 LG-006 3 3 2 0.93 RR-003 <![CDATA[B2]]> <![CDATA[B4、B5]]> TC-027 LG-011 4 3 3 0.75

[0377] The core of this step lies in using semantic repair microservice instances to create a controlled ablation environment by disabling them one by one, and using the number of times evidence at the same associated reference position is continuously retained as the basis for risk conclusions. This approach can distinguish between anomalous consistency caused by a certain language model or decoding strategy and the homology characteristics of bidding texts that remain stable across different models and decoding strategies. Simultaneously, the output retains a complete traceability chain from risk conclusions, ablation rounds, homology evidence in the repair path, cross-segment propagation order, to the S1 associated reference position, thereby improving the stability, interpretability, and operability of the collusion risk identification results.

[0378] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0379] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing, characterized in that, Executed by a microservice cluster, the method includes: S1. Locate the commitment statements in different bid texts that respond to the same bidding constraint through natural language processing services, perform the same reversible counterfactual transformation on the conditional scope or exception scope of the commitment statements, and form a counterfactual constraint test set by the transformed commitment statements, transformation types and associated reference positions; S2. Send the counterfactual constraint test set to multiple independent semantic repair microservices. Each semantic repair microservice restores the transformed commitment statement to a state compatible with the original bidding constraints. Record the semantic repair operations and cross-segment propagation order of non-mandatory response content during the restoration process. Associate the abnormally consistent semantic repair operations and cross-segment propagation order between different bidding texts as a set of evidence of homology of repair paths. S3. The repair path homology evidence set is sent to the collaborative review service. The collaborative review service disables the semantic repair microservices that participated in generating the repair path homology evidence set one by one, and calls the remaining semantic repair microservices to regenerate the repair path homology evidence set. When the regenerated repair path homology evidence set continuously points to the same bidding entity and can be traced back to the associated reference position, the collusion risk identification result is output.

2. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 1, characterized in that, The natural language processing service parses the tender documents, extracts the tender constraint clauses with unique identifiers, and generates semantic matching templates based on the core predicates and constraint parameters in the clauses. For each semantic matching template, sentence vector similarity calculation and keyword co-occurrence retrieval are used in each bidding text to identify candidate commitment statements that respond to the same bidding constraints. Negative or non-committal statements are then eliminated through dependency parsing to obtain accurate commitment statements.

3. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 2, characterized in that, S1 further includes: Semantic role labeling and dependency analysis are performed on commitment statements to identify clauses or phrases introduced by conditional conjunctions as conditional scopes and modifiers introduced by exception exclusion words as exception scopes. For the conditional scopes responding to the same tender constraint in different tender texts, the same reversible counterfactual transformation is applied. The transformation includes replacing the antecedents that are true with the antecedents that are false, or removing the condition-guided structure, generating a transformed commitment statement that violates the original constraint, and recording the original conditional word sequence before the transformation to maintain the reversibility of the transformation. For exception scopes responding to the same tender constraint in different tender documents, apply the same reversible counterfactual transformation, which includes replacing the exception scope with a general scope statement or deleting the entire exception modifier, generating a transformed commitment statement, and recording the original exception word sequence before the transformation; Record the transformation type for each transformation; Record the associated reference position of the transformed commitment statement in the original bid text. The associated reference position includes at least the bid text identifier, the chapter number, and the start and end character offsets of the statement.

4. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 3, characterized in that, The transformation type is selected from at least condition inversion, condition removal, exception generalization, and exception removal. The transformed commitment statements of all bid texts under the same bidding constraint, their respective transformation types, and associated reference positions are encapsulated into a test sample. All test samples are collected to form the counterfactual constraint test set.

5. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 1, characterized in that, The multiple independent semantic repair microservices are deployed as independent service instances based on different pre-trained language models and employing differentiated decoding strategies. All test samples in the counterfactual constraint test set are distributed to each semantic repair microservice via message middleware. After receiving the test samples, each semantic repair microservice takes the original text of the bidding constraint clause to which the transformed commitment statement belongs as the repair constraint condition, and generates the repaired commitment statement through constraint decoding or mask language modeling. This makes the repair result logically compatible with the bidding constraint clause, while maintaining compatibility with other non-binding content of the original bid text.

6. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 5, characterized in that, During the generation of the post-repair commitment statement, every associated modification triggered by this repair and affecting the non-mandatory response content is monitored and recorded. The non-mandatory response content includes statements, data items, or related commitments in the original tender text that do not directly violate the tender constraints but are indirectly adjusted to maintain the consistency within the text or the coherence of cross-paragraph references.

7. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 6, characterized in that, For each associated modification, the operation type, the position of the operation object in the text, and the timestamp of the operation trigger are extracted for the semantic repair operation. A cross-segment propagation order is constructed based on the triggering dependency relationship between operations. The cross-segment propagation order is represented by a directed acyclic graph of the operation occurrence, where nodes are semantic repair operations and edges represent the dependency relationship between the preceding operation and the subsequent operation. For different bid texts responding to the same bidding constraint, the semantic repair microservice compares the semantic repair operation sequence and cross-segment propagation order corresponding to each of the different bid texts. When the standardized editing similarity and graph structure isomorphism of the sequence both exceed the preset threshold, it is determined that there is an abnormally consistent repair propagation mode between the two. The identifiers of any two or more bid texts that meet the above-mentioned anomaly consistency conditions, the identifiers of the bid constraints to which they belong, the instance identifier of this semantic repair microservice, the common semantic repair operation mode, and the cross-segment propagation subgraph are associated as a repair path homology evidence; the repair path homology evidence output by all semantic repair microservices is collected to form the repair path homology evidence set.

8. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 1, characterized in that, The evidence set of homology of the repair path is sent to the collaborative review service via message middleware. The collaborative review service extracts the instance identifiers of the semantic repair microservices that participated in generating the homology evidence of each repair path from the evidence set and constructs a microservice-evidence mapping table. For each semantic repair microservice instance involved in the microservice-evidence mapping table, the collaborative review service sends a deactivation command to the service registry to temporarily unregister the corresponding instance, preventing it from receiving subsequent test set distribution requests. After each semantic repair microservice instance is deactivated, the collaborative review service redistributes the counterfactual constraint test set to the remaining semantic repair microservice instances that have not been deactivated, receives the regenerated repair path homology evidence, and performs consistency analysis on the newly generated evidence.

9. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 8, characterized in that, In the regenerated evidence of homology of the repair path, if at least one piece of evidence points to a set of bidding entities whose intersection with the set of bidding entities pointed to by the original evidence before the suspension is not empty, and the cross-segment propagation order that is determined to be abnormally consistent in the evidence can be traced back to the associated reference position recorded in step S1, then it is determined that the suspension operation has not eliminated the homology characteristics of the bidding entities.

10. The method for identifying collusion and bid-rigging risks based on microservice collaboration and natural language processing according to claim 9, characterized in that, After completing the deactivation and regeneration test of all semantic repair microservice instances, the collaborative review service will count the cumulative number of times each bidding entity is determined to still retain homology characteristics throughout the entire ablation review process. When the cumulative number of bids by any bidder reaches a preset threshold and can be traced back to the same set of associated reference locations, a bid-rigging risk identification result is generated. The result includes at least the bidder's identifier, the identifiers of other bidders that are related to it, the corresponding list of associated reference locations, and the risk confidence level.

Images