An intelligent auxiliary bid evaluation method and system based on AI strategy reasoning

Abstract

The application provides an intelligent auxiliary bid evaluation method and system based on AI strategy reasoning, relates to the technical field of intelligent bid evaluation and bidding information processing, and the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in that: the method is characterized in

Description

Technical Field

[0001] This invention relates to the field of intelligent bid evaluation and bidding information processing technology, and in particular to an intelligent assisted bid evaluation method and system based on AI strategy reasoning. Background Technology

[0002] In engineering construction, government procurement, and large-scale equipment bidding, the bid evaluation process typically involves a large number of complex and heterogeneous bid documents. These documents include not only textual descriptions but also price lists, qualification certificates, performance materials, and unstructured technical documents such as system architecture diagrams and flowcharts. With the continuous increase in the size and complexity of bid documents, existing bid evaluation systems exhibit the following shortcomings in actual operation: Most existing bid evaluation systems focus on comparing surface fields or matching text in bid documents, making it difficult to uniformly model the multimodal evidence relationships within the bid documents. This results in a lack of stable and locatable mapping between the scoring results and specific evidence, making it difficult to achieve automatic review and consistency verification at the system level in the event of a dispute.

[0003] Secondly, in complex scenarios involving multiple bidders, there may be implicit homology relationships between bid documents formed through content rewriting, structural adjustment, or template reuse. Such relationships cannot be identified simply by text similarity, and existing bid evaluation technologies cannot reliably detect them without relying on human experience.

[0004] Therefore, we propose an intelligent assisted bidding evaluation method and system based on AI strategy reasoning; the information disclosed in the background section is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent assisted evaluation method and system based on AI strategy reasoning, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An intelligent assisted bidding evaluation method based on AI strategy reasoning includes the following steps: S1. Obtain multiple bid data corresponding to the same bid evaluation project, perform format parsing and field alignment on the bid data, generate bid feature vectors for each bidder, and segment the locatable and referable text paragraphs, table cells and attachment items into a set of evidence fragments; S2. Extract multimodal semantic fingerprints from the evidence fragment set, and simultaneously extract document DNA features representing the editing source to form a homology detection fingerprint set of the bidders; S3. Based on the Siamese neural network, perform distance measurement on the fingerprint sets of the same source detection of different bidders and calculate the same source probability; when the same source probability meets the preset threshold condition, output the suspected same source risk and its corresponding evidence fragment identifier. S4. Obtain the evaluation rule text and extract the scoring items, rejection items and mutual exclusion conditions. Compile the evaluation rules into a computable intermediate rule representation and generate a constraint set. The constraint set is used to verify the consistency of the scoring results. S5. Based on the bid feature vector, evidence fragment set, and bid rigging risk warning information, execute AI strategy reasoning to generate a scoring candidate set; optimize the scoring candidate set to optimize the preset objective function under the premise of satisfying the constraint set, and output the optimal scoring scheme and the evidence fragment references corresponding to each scoring item. S6. Perform rule consistency verification on the optimal scoring scheme. If there is a conflict, output the conflict location information and trigger candidate backtracking or local re-inference; output the final auxiliary evaluation result.

[0007] S1 specifically involves: acquiring multiple sets of bidding data corresponding to the same bidding project. The bidding data includes at least the text of the bidding documents, the price list data, the qualification and performance proof data, and the bidding submission metadata, and establishing a bidding data set according to the bidding party; performing format parsing, field standardization, and missing data completion on the bidding data set to obtain structured bidding data, and forming a unified field view that can be used for subsequent feature extraction; segmenting the text of the bidding documents, the price list data, and the qualification and performance proof data into evidence fragment sets according to paragraphs, clauses, table units, and attachment items, and generating a unique evidence fragment identifier for each evidence fragment to support subsequent location and citation; extracting bidding feature vectors based on the structured bidding data, and associating and storing the bidding feature vectors and evidence fragment sets according to the bidding party to obtain a consistent input structure of bidding party, bidding feature vectors, and evidence fragment sets.

[0008] S2 specifically involves: reading the set of evidence fragments, determining the processing mode for each evidence fragment according to its type, and performing necessary normalization preprocessing to ensure comparability of subsequent features; extracting textual semantic features from textual evidence fragments to generate textual semantic fingerprint components, and binding the textual semantic fingerprint components with the corresponding evidence fragment identifiers; extracting topological structure features from unstructured graphic evidence fragments to generate topological structure fingerprint components, and binding the topological structure fingerprint components with the corresponding evidence fragment identifiers; extracting document DNA features from the evidence fragments and their carrying files, where document DNA features include at least layout rendering features and residual features of the generation environment; fusing the textual semantic fingerprint components, topological structure fingerprint components, and document DNA features to form a homologous detection fingerprint set, and storing it according to the bidding party.

[0009] S3 specifically involves: constructing pairwise similarity detection pairs for each bidder based on the similarity detection fingerprint set, and establishing a fingerprint pair to be calculated as input for each similarity detection pair; inputting each similarity detection pair into a Siamese neural network to obtain a fingerprint distance measurement result, which is used to characterize the similarity between the two bidders on the similarity detection fingerprint set; calculating the similarity probability based on the fingerprint distance measurement result, and comparing the similarity probability with a preset threshold condition to obtain a similarity determination result; when the similarity determination result meets the preset threshold condition, generating a bid-rigging risk warning message, and outputting a set of evidence fragment identifiers corresponding to the bid-rigging risk warning message to form a risk evidence chain that can be verified and located.

[0010] S4 specifically involves: obtaining the evaluation rule text corresponding to the evaluation project, and segmenting the evaluation rule text at the clause level to form a rule clause set; extracting scoring items, veto items, and mutually exclusive conditions from the rule clause set, and mapping them to computable scoring item fields and decision fields to form a rule element structure; compiling and generating an intermediate rule representation based on the rule element structure, so that the intermediate rule representation can be directly called by the subsequent consistency verification process; generating a constraint set from the intermediate rule representation, and performing an integrity verification on the constraint set to ensure that each scoring item and each veto item has an executable decision constraint in the constraint set.

[0011] S5 specifically involves: inputting the bid feature vector, evidence fragment set, bid-rigging risk warning information, and constraint set into the AI ​​strategy reasoning process to generate a scoring candidate set, which contains multiple candidate scoring schemes; for each candidate scoring scheme in the scoring candidate set, generating an evidence fragment identifier reference set corresponding to each scoring item, so that each scoring item has locatable evidence support; under the premise of satisfying the constraint set, performing strategy optimization on the scoring candidate set to optimize the preset objective function, and in the strategy optimization process, incorporating bid-rigging risk warning information as a risk constraint or penalty term to participate in the solution, so as to suppress high-risk candidate scoring schemes; determining the optimal scoring scheme from the scoring candidate set based on the strategy optimization results, and outputting the optimal scoring scheme and its corresponding evidence fragment identifier reference set as input for subsequent consistency verification.

[0012] S6 specifically involves: performing rule consistency verification on the optimal scoring scheme based on the constraint set to obtain the consistency verification result; when the consistency verification result indicates a conflict, generating conflict location information, which includes at least the rule clause identifier that triggered the conflict, the corresponding scoring item field, and the associated evidence fragment identifier; triggering candidate rollback or partial re-reasoning based on the conflict location information, re-executing AI strategy reasoning and strategy optimization on the scoring items affected by the conflict to generate a new optimal scoring scheme that satisfies the constraint set; and outputting the final auxiliary evaluation result, which includes at least the scoring details, bid-rigging risk warning information, evidence fragment identifier reference set, and reasoning process summary to support review and result traceability.

[0013] An intelligent assisted bidding evaluation system based on AI strategy reasoning includes: The module includes: bidding data processing module, evidence fragment generation and management module, bidding feature construction module, multimodal homology detection fingerprint construction module, homology detection and bid rigging risk analysis module, evaluation rule parsing and constraint generation module, AI strategy reasoning and scoring candidate generation module, strategy optimization and optimal scoring scheme determination module, consistency verification and conflict handling module, and result output and traceability management module.

[0014] The beneficial effects of this invention are as follows: This invention segments bidding data into uniquely identified evidence fragments and uses these fragments for identification-level referencing during scoring, risk analysis, and consistency verification. This allows each scoring item to be precisely traced to its corresponding document location and evidence source, avoiding the problem of discrepancies between scoring criteria and original materials in existing bidding systems. This improves the stability and verifiability of scoring results at the system level. Furthermore, this invention integrates textual semantic features, unstructured graphical topological features, and document DNA features to construct a homology detection fingerprint set. Based on Siamese neural networks, it performs distance measurement and probability determination, enabling the identification of potential homology relationships even after bidding content has been rewritten, rearranged, or template adjusted. This enhances the system-level ability to identify hidden bid-rigging risks and reduces reliance on human experience.

[0015] This invention generates intermediate rule representations and constraint sets by performing clause-level parsing and formal modeling of the evaluation rules. This enables the scoring process to perform rule verification item by item in the computer system and automatically locate the source of the problem when rule conflicts or missing evidence occur, thereby avoiding the scoring inconsistencies caused by misunderstandings of rules or configuration errors in existing technologies. An AI strategy reasoning mechanism is introduced into the generation and selection of scoring candidates. Under the premise of satisfying rule constraints, the scoring results, risk factors, and evidence completeness are comprehensively optimized. This ensures that the scoring scheme no longer relies solely on a single score result, but achieves synergistic optimization of compliance, low risk, and evidence sufficiency at the system level, improving the overall decision-making quality in complex bidding scenarios.

[0016] This invention introduces a rule consistency verification and conflict location mechanism in the scoring output stage. When a scoring result is detected to violate constraints or have evidence defects, it can trigger candidate rollback or local re-reasoning based on the scope of the conflict, and automatically mark manual review when necessary. This avoids the system from falling into an inexplicable or unreproducible abnormal state, and improves the reliability and sustainable operation capability of the intelligent bidding system in real engineering environments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an intelligent assisted evaluation method based on AI strategy reasoning according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: As Figure 1 As shown, this embodiment provides an intelligent assisted evaluation method based on AI strategy reasoning, including the following steps: S1. Obtain multiple sets of bidding data corresponding to the same bidding project. The bidding data shall include at least the text of the bidding documents, the price list data, the qualification and performance certificate data, and the bidding submission metadata. Perform format parsing and field alignment on the bidding data, generate the bidding feature vector of each bidder, and divide the locatable and referable text paragraphs, table cells and attachment items into evidence fragment sets. S2. Extract multimodal semantic fingerprints from the evidence fragment set. The multimodal semantic fingerprints include at least textual semantic features and topological structural features of unstructured graphics. At the same time, extract document DNA features that represent the editing source. The document DNA features include at least layout rendering features and residual features of the generation environment, forming a set of homologous detection fingerprints for the bidders. S3. Based on the Siamese neural network, perform distance measurement on the fingerprint sets of the same source detection of different bidders and calculate the probability of same source; when the probability of same source meets the preset threshold condition, output the suspected risk of same source and its corresponding evidence fragment identifier to form a bid-rigging risk warning information; S4. Obtain the evaluation rule text and extract the scoring items, rejection items and mutual exclusion conditions. Compile the evaluation rules into a computable intermediate rule representation and generate a constraint set. The constraint set is used to verify the consistency of the scoring results. S5. Based on the bid feature vector, evidence fragment set, and bid rigging risk warning information, execute AI strategy reasoning to generate a scoring candidate set; optimize the scoring candidate set to optimize the preset objective function under the premise of satisfying the constraint set, and output the optimal scoring scheme and the evidence fragment references corresponding to each scoring item. S6. Perform rule consistency verification on the optimal scoring scheme. If there is a conflict, output the conflict location information and trigger candidate rollback or local re-reasoning. Output the final auxiliary evaluation result. The final auxiliary evaluation result includes at least the scoring details, bid-rigging risk warning, evidence fragment references and reasoning process summary to support review.

[0020] S1 specifically includes the following sub-steps: S110. Bidding data acquisition: The computer equipment acquires multiple bidding data for the same bidding project from the bidding and procurement platform interface, file upload directory or bidding data receiving end corresponding to the bidding project, and establishes a bidding data set with "bidding project ID + bidder ID" as the archiving key. The bidding data includes at least the following: the text of the bidding documents, the price list data, the qualification and performance certification data, and the bidding submission metadata. The bidding submission metadata includes at least the submission time, the submission account, the client type, the submission IP address, the file generation time, and the file format type. When there are missing files, duplicate submissions, failed attachment decompression, or unsupported formats, the corresponding entries will be recorded in the data collection anomaly list, and the original files and the reasons for the anomalies will be retained to support subsequent review.

[0021] S120, Format Parsing and Field Alignment: Perform format parsing and field standardization on the bidding data set to generate structured bidding data; wherein, format parsing includes at least: extracting text paragraphs and chapter levels from PDF / Word, extracting table row and column structure and cell content from Excel, extracting readable text and page sequence information from scanned images, and parsing attachment item tree from compressed file attachments; When generating structured bidding data, the bidding fields are aligned according to a preset field mapping table, and synonymous fields are uniformly mapped to standard fields, including but not limited to mapping "quotation / total bid price / total price" to "total bid price", and mapping "project manager / project leader" to "key personnel-project leader"; the field values ​​are normalized in terms of scope and unit, including at least unifying the currency and decimal places of the amount, the date format, and the unit of work period to days; a missing field completion strategy is implemented: fields that can be derived from other fields are completed by calculation, and fields that cannot be derived are marked as missing and written with a missing flag instead of being filled by speculation, in order to avoid generating inaccurate data.

[0022] S130. Evidence Fragment Segmentation and Identification Generation: The verifiable content in the structured bidding data is segmented into a set of evidence fragments, and an evidence fragment identifier is generated for each fragment to support location citation; wherein the granularity of evidence fragment segmentation meets the requirement of "directly supporting the review of scoring items or veto items," and includes at least: The tender document text is divided into textual evidence segments according to "chapter / section / clause / paragraph"; The quotation list is segmented into tabular evidence fragments by “table name + row number + column number (or cell coordinates)”; The qualification and performance certificates are divided into "certificate items / performance items / attachment items" to form attachment evidence fragments; For each piece of evidence, locatable citation information is generated. This locatable citation information includes at least the filename, page number, paragraph number or cell coordinates, and attachment path. When generating evidence fragment identifiers, a stable combination of content and locatable information is used to ensure that evidence fragment identifiers within the same evaluation project are unique and reproducible. Specifically: in, Indicates any fragment of evidence, For identifying evidence fragments, For hash functions, This is a summary of the evidence fragment. This refers to the locatable citation information for evidence fragments. For the project ID, These are used as splicing symbols to ensure that the identifiers of evidence fragments can be stably referenced and verified in subsequent reasoning processes.

[0023] S140. Bidding Feature Vector Generation and Evidence Association: Based on structured bidding data, generate bidding feature vectors corresponding to each bidder, and associate and store the bidding feature vectors with a set of evidence fragments to form a unified input structure. The bid feature vector should include at least the following feature families: price features (including total bid price, itemized unit price statistics and abnormally low price ratio), qualification features (including qualification level, certificate validity period and scope of application), performance features (including performance quantity, performance amount range and project type label), and contract performance features (including construction period, key personnel configuration completeness and equipment list completeness), and optionally include bid submission metadata features (including submission time distribution and document generation environment label). Numerical normalization is performed on numerical features, encoding is performed on enumerated features, and vectorization is performed on text fields to form comparable semantic feature components. When storing related data, the bidder ID is used as the primary key to bind "bid feature vector - evidence fragment set (including evidence fragment identifier)" and an inverted index is built to support evidence retrieval for subsequent reasoning. The inverted index includes at least the mapping relationship of "standard field or scoring item field → evidence fragment identifier set", so that subsequent scoring reasoning can directly retrieve the evidence fragment corresponding to the target field.

[0024] S2 specifically includes the following sub-steps: S210, Evidence Fragment Loading and Modality Normalization: Read the set of evidence fragments generated in step S130, and determine its processing modality based on the type field and the format of the carrying file for each evidence fragment. The processing modality includes at least text modality, table modality, unstructured graphic modality and attachment metadata modality. Among them, the text modality corresponds at least to the text evidence fragments of the tender document and the text content transcribed from the table unit; the table modality corresponds at least to the table evidence fragments of the quotation list; the unstructured graphic modality corresponds at least to the graphic evidence fragments of CAD drawings, system architecture diagrams, and flowcharts; and the attachment meta-information modality corresponds at least to the attribute information of the attachment files. After completing the modality determination, normalization preprocessing is performed on different processing modalities: character encoding unification, full-width and half-width character standardization, and whitespace character cleanup are performed on the text modality. For tabular modalities, numerical values ​​are removed to the thousands place, unit sign normalization is performed, empty cells are explicitly marked, and row and column indexes are consistent. For unstructured graphical modalities, resolution or vector scale unification, noise suppression, and line width or primitive simplification are performed. When any evidence fragment fails in normalization preprocessing or subsequent feature extraction, the evidence fragment identifier of that evidence fragment is retained. With locatable reference information The missing feature markers and failure reasons are recorded so that the missing components can be handled consistently in subsequent fusion processes.

[0025] S220. Text Semantic Feature Extraction: For evidence fragments belonging to the text modality in the evidence fragment set, extract text semantic features after completing normalization preprocessing to generate text semantic features. ; Among them, text semantic features It is obtained by vectorizing the content of the evidence fragments. The vectorization method is an implementable text vector generation method, including at least one of the vectorization methods based on term statistics or based on a pre-trained text encoder. The obtained Normalization is performed to ensure comparability between different pieces of evidence, and each piece of evidence is identified as a separate piece of evidence. For key Bind and store with corresponding evidence fragments to form a system that contains at least [number of fragments]. The text feature records are used for subsequent homologous detection fingerprint generation.

[0026] S230. Unstructured Graph Topology Feature Extraction: For evidence fragments belonging to the unstructured graph modality in the evidence fragment set, after normalization preprocessing, the graph evidence fragments are parsed into topology graphs. , where the set of nodes Used to represent primitives, symbols, block frames, connection points, or arrow endpoints in a graphic; edge set. Used to represent the connection relationships of lines, arrow directions, or the connection relationships between graphic elements in a graphic; when the graphic evidence fragment is a vector graphic, it is generated based on the endpoint adjacency and connection relationships of the vector objects. and When the graphic evidence fragment is a raster graphic, the endpoints and intersections are extracted through binarization and skeletonization. And generate based on pixel connectivity ; Obtaining the topology map Afterwards, from Calculate topological features and generate topological fingerprint components. The topological features include at least one of the following: number of nodes, number of edges, degree distribution statistics, number of connected components, number of cycles, and length of the longest path; for the obtained Normalization was performed to ensure comparability, and the evidence was identified as fragments. The key will be bound and stored with the corresponding evidence fragment, forming a record containing at least... The topological feature records are used for subsequent homologous detection fingerprint generation.

[0027] S240. Document DNA Feature Extraction and Homology Detection Fingerprint Set Generation: Extract and characterize document DNA features representing the editing source of evidence fragments and their accompanying documents, and generate document DNA features. ; The document DNA features include at least: layout rendering features and generation environment residual features, wherein the layout rendering features include at least one of the following: line spacing and character spacing statistics, paragraph indentation statistics, font subset fingerprint and page object density distribution; and the generation environment residual features include at least one of the following: Producer (production tool) or Creator (creation tool) identifier, software version number, PDF object structure mode, embedded font and image compression parameters. In generation Then, marked with evidence fragments. The key will be bound to the corresponding evidence fragment and stored; subsequently, the text semantic features will be... Topological fingerprint components With document DNA characteristics By fusing the samples, we obtain the homology detection fingerprint vector of evidence fragment e. The fusion method is to normalize after splicing, specifically:

[0028] in, For same-origin fingerprint detection, For vector concatenation, For normalization operator; when , or When missing components exist, the missing components are set to zero vectors or skipped before normalization is performed based on the feature missing markers recorded in step S210, to ensure that... It can be generated stably; Finally, for the set of evidence fragments from the same bidder, the corresponding [evidence fragment name] is calculated for each evidence fragment. The fingerprints are then aggregated to form a set of homology detection fingerprints for the bidder, and the set of homology detection fingerprints is associated with the bidder's ID and stored as input for subsequent homology detection pair construction and homology probability calculation.

[0029] Document DNA feature extraction, specifically implemented using the following underlying parsing techniques: Extraction of residual features from the generated environment: Instead of relying on the rendering results of the PDF reader, this method directly parses the cross-reference table and trailer of the PDF file. It obtains the production tool identifier by reading the / Producer and / Creator fields from the document information dictionary; and retrieves the original creation time and modification history after obfuscation or erasure operations by parsing the / XMP metadata stream.

[0030] Layout rendering feature extraction: Traverse the PDF content stream, extract operands between text display operators, and calculate the small jitter variance of average character spacing and line spacing; simultaneously extract font descriptors from page objects, and statistically analyze the glyph width table of embedded font subsets as font rendering fingerprints. These underlying features retain specific quantitative statistical patterns even after the document is "printed as PDF".

[0031] S3 specifically includes the following sub-steps: S310. Construction of Homologous Detection Pairs: Based on the homologous detection fingerprint set generated in step S240, construct homologous detection pairs for each pair of bidders; wherein, for any bidder b, its homologous detection fingerprint set is denoted as... , The homogeneous detection fingerprint vectors corresponding to each evidence fragment in the bidder's evidence fragment set. Composition; for the bidding party The comparison object is determined as and The set of pairs is compared; to reduce computational overhead and ensure deployability, candidate bidder pairs can be optionally screened based on bid submission metadata features or simplified fingerprints to generate a set of bidder pairs to be detected. For each bidder pair to be detected, a computational input structure is established, which at least includes the bidder pair identifier and the set of identical detection fingerprints on both sides. , .

[0032] S320, Twin Neural Network Distance Metric: For each bidder to be detected... Fingerprint vectors are detected using evidence fragment-level homology. For input, and The fingerprint vectors are input into a parameter-shared Siamese neural network to obtain the corresponding embedding vector z(e); the fingerprint distance metric for evidence fragment pairs is calculated based on the embedding vectors, and the fingerprint distance metric is used to characterize the similarity between the two evidence fragments in the same-origin fingerprint detection; the fingerprint distance metric is implemented using at least one of Euclidean distance or cosine distance; in the scenario of set-to-set comparison, the fingerprint distance is calculated using the fingerprint distance metric for evidence fragment pairs. and After calculating the distance set of the evidence fragments, the Top-K evidence fragment pairs with the smallest distance are selected as strong evidence pairs, and the distance value corresponding to each strong evidence pair and the identifiers of the evidence fragments on both sides are recorded. .

[0033] Specifically, the parameter-sharing Siamese neural network comprises two structurally identical sub-networks, each a multilayer perceptron (MLP) structure. This MLP structure contains at least three fully connected layers, connected by ReLU activation functions, and a dropout layer is introduced before the output layer to prevent overfitting. The dimension of the input layer is the same as the homogeneous detection fingerprint vector. The dimensions are consistent, and the output layer dimension is... (e.g., 64 or 128) are used to map high-dimensional fingerprints to low-dimensional Euclidean space.

[0034] S330. Homology Probability Calculation and Threshold Determination: Based on the distance values ​​of the Top-K strong evidence pairs obtained in step S320. Generate bidders The aggregation distance D, the aggregation distance D is the pair The aggregation results, including at least one of taking the mean or the minimum value; after obtaining the aggregation distance D, the homology probability P is calculated, and the correspondence between the homology probability P and the aggregation distance D is defined as follows: Where P is the homology probability and D is the aggregation distance. It is an exponential function; subsequently, the homology probability P is compared with a preset dual threshold condition, which includes a high threshold. With low threshold ,and ;when When it is determined to be of high homology risk, When it is determined to be a suspected homology risk, when The evidence is classified as having low homology risk; K is the preset number of strong evidence pairs to be selected.

[0035] When the number of valid homologous fingerprint vectors is insufficient to form a preset number of Top-K strong evidence pairs, the bidder pair is marked as insufficient evidence and a manual review prompt is triggered or the risk level is reduced to suppress false alarms.

[0036] S340, Generation and Evidence Binding Output of Bid-rigging Risk Warning Information: When the threshold determination result of step S330 is high homology risk or suspected homology risk, generate bid-rigging risk warning information and bind and output the bid-rigging risk warning information with the verifiable and locatable risk evidence chain; The risk evidence chain consists of a set of evidence fragment identifiers corresponding to the Top-K strong evidence pairs, and includes at least the evidence fragments corresponding to the bidders on both sides. It further outputs locatable citation information corresponding to the evidence fragment identifier. This allows reviewers to locate files by filename, page number, paragraph number, or cell coordinates; where subscripts are used. The bids correspond to the two bidders in the homologous detection pair. A sorted index for strong evidence pairs; The risk warning information for bid rigging should include at least the bidder's identification, the probability of same source P, the risk level, the trigger threshold range, the Top-K evidence fragment identification set and its locatable reference information, and write the above fields into the risk record table to form a traceable record of same source detection results, which will be used for risk warning and review support in the subsequent bid evaluation process.

[0037] S4 specifically includes the following sub-steps: S410. Acquisition of Bidding Rule Text: The computer equipment acquires the bidding rule text corresponding to the bidding project and performs clause-level segmentation of the bidding rule text to form a set of rule clauses. The clause-level segmentation is performed according to the chapter titles, clause numbering styles, and paragraph boundaries in the bidding rule text, including at least segmentation by the level of "chapter / section / clause / item", or, if the number is missing, segmentation by paragraph boundaries containing keywords such as "scoring criteria, scoring items, veto clauses, bonus items, deduction items, response requirements, mutual exclusion, and not simultaneous". A rule clause identifier is generated for each rule clause and stored in association with the bidding project ID and rule document version number for subsequent constraint generation, conflict location, and result traceability.

[0038] S420. Rule Element Extraction and Scoring Item Mapping: Perform rule element extraction on the set of rule clauses to obtain the rule element structure; wherein, the rule element extraction adopts an implementable extraction method, including at least the extraction method based on keywords and regular expression templates and the extraction method based on structured table reading, and optionally uses a semantic parsing model to perform consistency verification on the extraction results; the rule element structure includes at least scoring item elements, veto item elements, and mutually exclusive condition elements: The scoring item elements should at least include a scoring item identifier, a maximum score, a scoring method, a scope of application, a required type of evidence, and a scoring item field; among which, the scoring method should at least include one of the following: adding points, deducting points, interval scoring, and tiered scoring. The elements of a veto should at least include the expression of the veto condition, the trigger field, the evidence requirements, the scope of application, and the veto field; the elements of a mutually exclusive condition should at least include the mutually exclusive group identifier, the list of clauses within the group, the trigger priority, and the exclusivity strategy. Based on the field mapping table, the scoring item fields and veto item fields are mapped to the standard fields in the structured bidding data generated in step S120, so that the field references in the rule element structure are consistent with the standard fields in the structured bidding data, in order to avoid the inability to locate fields and cause subsequent constraints to be unexecutable.

[0039] S430, Rule Intermediate Representation Generation: Based on the rule element structure, compile and generate rule intermediate representations so that the rule intermediate representations can be directly executed by computer devices for judgment and scoring; wherein, the rule intermediate representations adopt an executable data structure representation, and each rule intermediate representation record includes at least: rule clause identifier, rule type, field reference, operator, threshold or range, score or weight, evidence requirement, scope of application and priority. The rule types include at least scoring item rules, veto item rules, and mutually exclusive condition rules; the operators include at least comparison operators, existence operators, range operators, inclusion operators, and logical mutually exclusive operators, and include at least one or more of the following: ≥, ≤, =, in_range (range inclusion), contains (set inclusion), exists (existence determination), AND (logical AND), OR (logical OR), and XOR (logical XOR); when the rule clause contains tiered scoring or segmented scoring, the middle of the rule further includes the scoring segment definition and corresponding score mapping to ensure that the scoring rule can be directly executed.

[0040] In this embodiment, the intermediate representation (IR) of the rule is stored in the form of a JSON object or an abstract syntax tree (AST) node. For example, the IR structure of a rule that "the total bid price must be less than the maximum price limit" is as follows: {"rule_id":"R001","type":"threshold_constraint","field_ref":"total_bid_price","operator":"LT","value_source":"max_limit_variable","action_on_fail":"reject"}. This structured representation allows subsequent validation engines to directly read field references and operators without re-parsening the text.

[0041] S440. Constraint Set Generation and Integrity Verification: A constraint set is generated from the intermediate representation of the rules. The constraint set is used to perform consistency verification on subsequent scoring results. The constraint set includes at least: threshold constraints, mutual exclusion constraints, and integrity constraints. Threshold constraints are used to limit the standard fields or scoring item fields of the structured bidding data to meet the threshold or interval conditions. Mutual exclusion constraints are used to limit the rules within the same mutual exclusion group from taking effect at the same time or to select the effective rules according to the exclusive strategy. Integrity constraints are used to limit the necessary fields and necessary evidence to exist and be referenced by evidence fragment identifiers. After generating the constraint set, integrity and executability checks are performed on the constraint set: for each scoring item field and veto item field, it is checked whether there is a corresponding constraint, and for each constraint, it is checked whether its field references, operators, and thresholds or ranges are complete; further, consistency checks are performed to handle rule conflicts and priorities: when there are multiple rule constraints for the same field reference that are contradictory, the effective constraint is determined according to the priority in the intermediate representation of the rule and a conflict record is generated. The conflict record includes at least the conflicting rule clause identifier, conflicting field reference, and conflict reason; When there are fields that cannot be mapped, thresholds that are missing, or mutual exclusion groups that are incomplete, the corresponding clauses will be marked as rule defects and manual supplementation or rollback to the default template constraints will be triggered to ensure that the constraint set can be used continuously for subsequent scoring candidate generation and consistency verification processes.

[0042] S5 specifically includes the following sub-steps: S510, Generation of scoring candidate set (AI strategy reasoning): Input the bidding feature vector generated in step S140, the evidence fragment set generated in step S130, the bid rigging risk warning information generated in step S340, and the constraint set generated in step S440 into the AI ​​strategy reasoning process to generate a scoring candidate set. The AI ​​strategy reasoning process takes the bid evaluation status as input and outputs a sequence of strategy actions. The bid evaluation status includes at least: the bid feature vector, the available field reference information of the constraint set, the risk level or homology probability of the bid rigging risk warning information, and the evidence availability marker of the evidence fragment set. The strategy actions include at least: retrieving a set of candidate evidence fragment identifiers for a certain scoring item field, calculating the candidate score for that scoring item field based on the rule intermediate representation, and generating missing or review-required markers for scoring item fields with insufficient evidence or missing fields; Based on different strategy action sequences, a preset number of candidate scoring schemes are generated to form a scoring candidate set. The candidate scoring schemes are then subjected to initial screening. The initial screening rules include at least removing candidate scoring schemes that trigger veto constraints and removing candidate scoring schemes whose evidence missing ratio exceeds a preset threshold, so as to ensure that the candidate scoring schemes entering subsequent optimization have an executable and verifiable basis.

[0043] Specifically, the AI ​​policy reasoning process is implemented using the Monte Carlo Tree Search (MCTS) algorithm or the Beam Search algorithm.

[0044] State definition: The initial state is an empty scoring scheme, and the intermediate state is a scheme in which some scoring items have been assigned values.

[0045] Action definition: The action space is defined for the currently unrated rating item field. Select a piece of evidence from the set of candidate evidence. And assign scores.

[0046] Reasoning process: Starting from the root node, the system expands based on a preset heuristic strategy (such as prioritizing high-scoring items) to simulate and generate multiple complete scoring paths (i.e., strategy action sequences). Each complete path corresponds to a candidate scoring scheme.

[0047] S520, Generation of evidence citations for candidate scoring schemes: For each candidate scoring scheme in the candidate scoring scheme set, an evidence fragment identifier citation set is generated for each scoring item field; wherein, the inverted index established in step S140 is used as the evidence retrieval entry point, and the candidate evidence set matching the target scoring item field is retrieved from the evidence fragment set according to the mapping relationship of "scoring item field → evidence fragment identifier set"; For each scoring item field, a preset number of Top-K evidence fragment identifiers are selected from the candidate evidence set based on field matching degree, evidence type matching, or text similarity as references, and the corresponding locatable reference information is bound simultaneously. To support the verification and positioning; When the candidate evidence retrieved for a certain scoring item field is insufficient to constitute a Top-K citation, a missing marker is generated for that scoring item field, and the scoring item field is added to the review list. This ensures that each candidate scoring scheme meets the coverage requirement that "each scoring item field has evidence fragment identification citation or missing markers," thereby avoiding the suspension of scoring item evidence.

[0048] S530, Strategy Optimization Solution: Under the premise of satisfying the constraint set, perform strategy optimization on the scoring candidate set to output the optimal scoring scheme; firstly, based on the constraint set, perform feasibility filtering on the scoring candidate set, and determine and eliminate candidate scoring schemes that trigger veto constraints, violate mutual exclusion constraints, or violate necessary evidence integrity constraints as infeasible, thus obtaining a feasible candidate set; then, calculate the objective function value for each candidate scoring scheme in the feasible candidate set, and select the candidate scoring scheme with the optimal objective function value as the optimal scoring scheme; the objective function is used to weigh the total score, bid-rigging risk, and lack of evidence, and is defined as: in, Let the objective function value of the candidate scoring scheme be... As candidate scoring schemes, For the feasible candidate set, The total score for the candidate scoring schemes. The risk cost of bid rigging for candidate scoring schemes is derived from the risk level or homologous probability mapping of bid rigging risk warning information. The cost of missing evidence for candidate scoring schemes is determined by the number of missing markers or the proportion of missing markers. and As a tradeoff coefficient and as a system parameter configuration; Optionally, iterative improvements are performed on a preset number of candidate scoring schemes with the highest objective function values: the evidence fragment identifier reference set is reselected or the candidate score is recalculated for the scoring item fields in the review list, and the constraint set verification and objective function calculation are re-executed after each improvement to improve the compliance, low risk and completeness of evidence of the optimal scoring scheme.

[0049] When solving for the optimal scoring scheme, if the candidate scoring set is large, a genetic algorithm is used for iterative optimization. The specific steps include: encoding each candidate scoring scheme as a chromosome, where each gene location corresponds to an evidence citation index for a scoring item; and using the objective function... As a fitness function, the individual with the highest fitness is output as the optimal scoring scheme after a preset number of iterations through selection, crossover (exchanging subsets of scoring items of different schemes) and mutation (randomly changing the evidence reference of a scoring item).

[0050] S540. Output of Optimal Scoring Scheme: Based on the optimal scoring scheme determined in step S530, output the core scoring content of the auxiliary evaluation results; wherein, the optimal scoring scheme shall at least output: scoring details (scoring item fields and corresponding scores), evidence fragment identification reference set corresponding to each scoring item field and its locatable reference information. Information on bid-rigging risks (including risk level or probability of similarity and its Top-K risk evidence chain), evidence missing markers and a list of documents requiring review, and constraint fulfillment markers (including pass or fail and the reasons). The aforementioned fields of the optimal scoring scheme are then written into the scoring result record table, enabling subsequent consistency verification and review processes to trace and locate the scoring source, evidence chain, and risk warnings based on the scoring result record table.

[0051] S6 specifically includes the following sub-steps: S610, Rule Consistency Verification Execution: Perform rule consistency verification on the optimal scoring scheme output in step S540 based on the constraint set generated in step S440 to obtain the consistency verification result; wherein, the input of the consistency verification includes at least: the scoring details of the optimal scoring scheme (scoring item field → score), the evidence fragment identification reference set and its locatable reference information loc(e), missing markers and review list, bid rigging risk warning information, and structured bidding data; The consistency check is performed by checking each constraint set one by one, reading the field references, operators, thresholds or ranges, evidence requirements and applicable scope of each constraint, and obtaining the corresponding field values ​​and evidence existence from the structured bidding data and evidence fragment set to determine whether the constraint is satisfied. Constraints include at least threshold constraints, mutual exclusion constraints, and integrity constraints. Threshold constraints are used to verify whether field values ​​meet threshold or range conditions. Mutual exclusion constraints are used to verify whether rules within a mutual exclusion group are effective according to the exclusive strategy. Integrity constraints are used to verify whether necessary fields and necessary evidence exist and can be referenced by evidence fragment identifiers. The consistency verification output includes at least the consistency verification result (pass or fail) and a list of violated constraints. The list of violated constraints is used to list the violated constraint entries and serve as input for subsequent conflict localization.

[0052] S620. Conflict Location Information Generation: When the consistency check result is unsuccessful, conflict location information is generated based on the list of violated constraints, and the conflict location information is written to the conflict record table to support traceability. The conflict location information for each violated constraint item includes at least: the corresponding rule clause identifier or constraint identifier, the field reference that triggered the conflict (consistent with the standard fields or scoring item fields of the structured bidding data), the conflict type, the cause of the conflict, and a set of associated evidence fragment identifiers and their locatable reference information. ; The conflict types include at least threshold conflict, mutual exclusion conflict and integrity conflict. Threshold conflict is used to characterize that the field value does not meet the threshold or interval condition. Mutual exclusion conflict is used to characterize that rules in a mutual exclusion group are in effect at the same time or that the effective rule is not selected according to the exclusive strategy. Integrity conflict is used to characterize that necessary fields are missing, necessary evidence is missing or evidence cannot be located. The conflict cause is used to explain the specific reasons that trigger the conflict, including missing fields, missing evidence, field values ​​exceeding thresholds, or conflicts with mutually exclusive rules, so that subsequent rollbacks or partial re-inferences can be made in a targeted manner based on the scope of the conflict.

[0053] S630, Candidate Backoff or Local Re-inference: Based on the conflict location information, candidate backoff or local re-inference is triggered to generate an updated optimal scoring scheme that satisfies the constraint set; wherein, the backoff source of the candidate backoff is clearly defined as the scoring candidate set generated in step S510 or the Top-N feasible candidate scoring schemes retained during the optimization process in step S530. The computer device selects candidate scoring schemes that have not triggered the same conflict type or have the second-best objective function value from the backoff source as backoff candidates; N is the preset number of backoff candidates to be retained; Local re-inference uses conflict location information to determine the set of affected fields. The set of affected fields includes at least the scoring field and the rejection field involved in the conflict location information. The computer device only re-executes evidence retrieval and evidence fragment identification reference generation for the set of affected fields, and recalculates the candidate scores of the affected fields according to the intermediate representation of the rules. Then, it re-executes strategy optimization on the updated candidate scoring scheme to generate an updated optimal scoring scheme. After candidate rollback or partial re-reasoning, the consistency check in step S610 is re-executed to verify whether the updated optimal scoring scheme satisfies the constraint set. To avoid infinite loops, a maximum number of re-reasoning rounds is set. When conflicts still exist after reaching the preset round limit, a mark requiring manual review is generated, and the current optimal feasible candidate scoring scheme and its conflict location information are output for manual processing. When the conflict type is rule defect, field cannot be mapped, or necessary evidence is missing, manual supplementation or a mark requiring review is triggered first rather than continuing re-reasoning to ensure that the process can be implemented.

[0054] Specifically, the system maintains a field dependency graph, where nodes represent scoring item fields, and directed edges represent computational dependencies or mutual exclusion constraints between fields. Once conflicting fields are identified, the system performs a breadth-first search (BFS) based on this dependency graph, adding all downstream dependent nodes to the "affected field set," thereby ensuring that the re-inference process covers all scoring items that may experience cascading effects due to changes.

[0055] S640. Final Auxiliary Bid Evaluation Result Output: Output the final auxiliary bid evaluation result and generate a traceable record; wherein, the final auxiliary bid evaluation result includes at least: Final scoring details (scoring item fields → score and total score), evidence fragment identification and citation set and its locatable citation information The following information is required: risk warning information on bid rigging (risk level or probability of homology and its Top-K risk evidence chain), consistency verification results (pass or fail), conflict handling records (including the reasons for triggering rollback or partial re-inference, the rollback candidate identifier or re-inference round used), missing markers and a list of items to be reviewed, and markers that need to be manually reviewed (if triggered). The final auxiliary evaluation results will be written into the scoring result record table, and linked to the risk record table and conflict record table, so that reviewers can locate, verify and trace the scoring basis, risk warning and conflict handling process based on evidence fragment identification and locatable reference information.

[0056] Example 2: This example provides an intelligent assisted evaluation system based on AI strategy reasoning, including: The bidding data processing module is used to acquire multiple bidding data corresponding to the same bidding project and establish a bidding data set according to the bidding project ID and the bidder ID; wherein, the bidding data includes at least the bidding document text, quotation list data, qualification and performance certificate data, and bidding submission metadata; the bidding data processing module is also used to perform format parsing, field standardization, field alignment and missing tag processing on the bidding data to generate structured bidding data.

[0057] The evidence fragment generation and management module is used to divide the structured bidding data into a set of evidence fragments that can be located and referenced. The evidence fragments include at least text evidence fragments, table evidence fragments, and attachment evidence fragments. The evidence fragment generation and management module is also used to generate a unique evidence fragment identifier for each evidence fragment and associate the evidence fragment identifier with locatable reference information to support subsequent evidence retrieval and verification.

[0058] The bidding feature construction module is used to generate a bidding feature vector based on the structured bidding data. The bidding feature vector includes at least price features, qualification features, performance features, and contract performance features. The bidding feature construction module is also used to associate and store the bidding feature vector with the evidence fragment set according to the bidding party, and to establish a mapping relationship between the scoring item field and the evidence fragment identifier set.

[0059] The multimodal homology detection fingerprint construction module is used to perform modality determination and normalization preprocessing on the evidence fragment set, and extract text semantic features, unstructured graph topological features and document DNA features respectively; the multimodal homology detection fingerprint construction module is also used to fuse the text semantic features, the topological features and the document DNA features to generate evidence fragment-level homology detection fingerprint vectors, and to summarize them according to the bidders to form a homology detection fingerprint set.

[0060] The homology detection and bid-rigging risk analysis module is used to construct homology detection pairs between bidders based on the homology detection fingerprint set, and input the homology detection pairs into a Siamese neural network to perform distance measurement; the homology detection and bid-rigging risk analysis module is also used to calculate the homology probability of bidder pairs based on the evidence fragment level distance results, and generate bid-rigging risk warning information and corresponding risk evidence chains according to preset threshold conditions.

[0061] The bid evaluation rule parsing and constraint generation module is used to obtain the bid evaluation rule text and perform clause-level segmentation to generate a set of rule clauses. The bid evaluation rule parsing and constraint generation module is also used to extract scoring items, veto items and mutually exclusive conditions from the set of rule clauses, generate intermediate rule representations, and generate a constraint set from the intermediate rule representations. The constraint set is used to limit the range of values ​​of scoring items, mutual exclusion relationships and the completeness of necessary evidence.

[0062] The AI ​​strategy reasoning and scoring candidate generation module is used to perform AI strategy reasoning based on the bidding feature vector, the evidence fragment set, the bid rigging risk warning information, and the constraint set to generate a scoring candidate set; wherein, each scoring candidate scheme includes at least a scoring item field, a corresponding candidate score, an evidence fragment identifier reference set, and a missing mark or a mark that needs to be reviewed.

[0063] The strategy optimization and optimal scoring scheme determination module is used to perform strategy optimization on the scoring candidate set under the premise of satisfying the constraint set, and determine the optimal scoring scheme by comprehensively considering the scoring results, bid-rigging risk and evidence deficiency; the strategy optimization and optimal scoring scheme determination module is also used to perform local re-inference or candidate backtracking on scoring items with conflicts or insufficient evidence.

[0064] The consistency verification and conflict handling module is used to perform rule consistency verification on the optimal scoring scheme based on the constraint set, and generate conflict location information when the verification fails; wherein, the conflict location information includes at least the identifier of the violated rule clause, the field reference that triggered the conflict, the conflict type, and the identifier and locatable reference information of the associated evidence fragment; the consistency verification and conflict handling module is also used to trigger candidate rollback, partial re-reasoning, or manual review marking under preset conditions.

[0065] The result output and traceability management module is used to output the final auxiliary evaluation result, which includes at least the scoring details, evidence fragment identification reference set, bid rigging risk warning information, consistency verification result, and conflict handling record. The result output and traceability management module is also used to write the final auxiliary evaluation result into the scoring result record table and associate it with the risk record table and conflict record table to support review and traceability.

[0066] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0067] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0068] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent assisted evaluation method based on AI strategy reasoning, characterized in that, Includes the following steps: S1. Obtain multiple bid data corresponding to the same bid evaluation project, perform format parsing and field alignment on the bid data, generate bid feature vectors for each bidder, and segment the locatable and referable text paragraphs, table cells and attachment items into a set of evidence fragments; S2. Extract multimodal semantic fingerprints from the evidence fragment set, and simultaneously extract document DNA features representing the editing source to form a homology detection fingerprint set for the bidders, specifically: The system reads a set of evidence fragments, determines the processing mode for each fragment according to its type, and performs necessary normalization preprocessing to ensure comparability of subsequent features. For text-based evidence fragments, it extracts textual semantic features, generates textual semantic fingerprint components, and binds these components to the corresponding evidence fragment identifiers. For unstructured graphical evidence fragments, it extracts topological features, generates topological fingerprint components, and binds these components to the corresponding evidence fragment identifiers. Finally, it extracts document DNA features from the evidence fragments and their supporting files. Document DNA features include at least layout rendering features and residual features from the generation environment. The semantic fingerprint components of the text, the topological fingerprint components, and the document DNA features are fused to form a homology detection fingerprint set, which is then aggregated and stored according to the bidder. S3. Based on the Siamese neural network, perform distance measurement on the fingerprint sets of the same source detection of different bidders and calculate the probability of same source. When the homology probability meets the preset threshold condition, output the suspected homology risk and its corresponding evidence fragment identifier; S4. Obtain the evaluation rule text and extract the scoring items, rejection items, and mutual exclusion conditions. Compile the evaluation rules into a computable intermediate rule representation and generate a constraint set. The constraint set is used to verify the consistency of the scoring results, specifically: Obtain the evaluation rule text corresponding to the evaluation project, and perform clause-level segmentation of the evaluation rule text to form a rule clause set; extract scoring items, veto items, and mutually exclusive conditions from the rule clause set, and map them into computable scoring item fields and decision fields to form a rule element structure; compile and generate an intermediate rule representation based on the rule element structure, so that the intermediate rule representation can be directly called by the subsequent consistency verification process; generate a constraint set from the intermediate rule representation, and perform integrity verification on the constraint set to ensure that each scoring item and each veto item has an executable decision constraint in the constraint set; S5. Based on the bid feature vector, evidence fragment set, and bid rigging risk warning information, execute AI strategy reasoning to generate a scoring candidate set; optimize the scoring candidate set to optimize the preset objective function while satisfying the constraint set, and output the optimal scoring scheme and the evidence fragment references corresponding to each scoring item, specifically: The bidding feature vector, evidence fragment set, bid-rigging risk warning information, and constraint set are input into the AI ​​strategy reasoning process to generate a scoring candidate set, which contains multiple candidate scoring schemes. For each candidate scoring scheme in the scoring candidate set, an evidence fragment identifier reference set corresponding to each scoring item is generated, so that each scoring item has locatable evidence support. Under the premise of satisfying the constraint set, the scoring candidate set is optimized to optimize the preset objective function. In the process of strategy optimization, bid-rigging risk warning information is used as a risk constraint or penalty term to participate in the solution, so as to suppress high-risk candidate scoring schemes.

2. The intelligent assisted evaluation method based on AI strategy reasoning according to claim 1, characterized in that, It also includes S6, which performs rule consistency verification on the optimal scoring scheme. If there is a conflict, it outputs the conflict location information and triggers candidate backtracking or local re-inference; and outputs the final auxiliary evaluation result.

3. The intelligent assisted evaluation method based on AI strategy reasoning according to claim 1, characterized in that, S1 specifically refers to: Obtain multiple sets of bidding data corresponding to the same bidding project. The bidding data shall include at least the text of the bidding documents, the price list data, the qualification and performance certificate data, and the bidding submission metadata, and establish a bidding data set according to the bidder. The bidding data set is parsed, fields are standardized and missing data is filled to obtain structured bidding data and form a unified field view that can be used for subsequent feature extraction. The tender document text, quotation list data, qualification and performance certificate data are divided into evidence fragment sets according to paragraphs, clauses, table units, and attachment items, and a unique evidence fragment identifier is generated for each evidence fragment to support subsequent location and reference. Based on the structured bidding data, the bidding feature vector is extracted, and the bidding feature vector and the evidence fragment set are associated and stored according to the bidding party, so as to obtain a consistent input structure of bidding party, bidding feature vector, and evidence fragment set.

4. The intelligent assisted evaluation method based on AI strategy reasoning according to claim 1, characterized in that, S3 specifically refers to: Based on the set of homology detection fingerprints, construct homology detection pairs for each pair of bidders, and establish a fingerprint pair to be calculated for each homology detection pair. Each pair of homologous detections is input into the Siamese neural network to obtain the fingerprint distance metric, which is used to characterize the similarity between the two bidders on the set of homologous detection fingerprints. The homology probability is calculated based on the fingerprint distance measurement result, and the homology probability is compared with a preset threshold condition to obtain the homology determination result. When the homology determination result meets the preset threshold condition, a bid-rigging risk warning message is generated, and a set of evidence fragment identifiers corresponding to the bid-rigging risk warning message is output to form a risk evidence chain that can be verified and located.

5. The intelligent assisted evaluation method based on AI strategy reasoning according to claim 1, characterized in that, S5 also includes: Based on the strategy optimization results, the optimal scoring scheme is determined from the scoring candidate set, and the optimal scoring scheme and its corresponding evidence fragment identifier reference set are output as input for subsequent consistency verification.

6. The intelligent assisted evaluation method based on AI strategy reasoning according to claim 2, characterized in that, S6 specifically refers to: The optimal scoring scheme is subjected to rule consistency verification based on the constraint set, and the consistency verification result is obtained. When the consistency verification result indicates a conflict, conflict location information is generated. The conflict location information includes at least the rule clause identifier that triggered the conflict, the corresponding scoring item field, and the associated evidence fragment identifier. Based on conflict location information, candidate backoff or local re-inference is triggered, and AI policy reasoning and policy optimization are re-executed for the scoring items affected by the conflict in order to generate a new optimal scoring scheme that satisfies the constraint set. Output the final auxiliary evaluation results, which should include at least a detailed score breakdown, information on bid rigging risks, a set of evidence fragments and references, and a summary of the reasoning process, to support review, verification, and result traceability.

7. An intelligent assisted bidding evaluation system based on AI strategy reasoning, employing an intelligent assisted bidding evaluation method based on AI strategy reasoning according to any one of claims 1-6, characterized in that, include: The module includes: bidding data processing module, evidence fragment generation and management module, bidding feature construction module, multimodal homology detection fingerprint construction module, homology detection and bid rigging risk analysis module, evaluation rule parsing and constraint generation module, AI strategy reasoning and scoring candidate generation module, strategy optimization and optimal scoring scheme determination module, consistency verification and conflict handling module, and result output and traceability management module.

Images