Artificial intelligence-based multi-end collaborative scheme preparation and review system and cloud platform

Abstract

The application discloses a scheme compiling and reviewing system and cloud platform based on multi-terminal cooperation of artificial intelligence, and relates to the technical field of artificial intelligence natural language processing and multi-terminal cooperation document compiling.The system adopts a two-level requirement disassembly architecture of preliminary screening and refining, completes requirement through horizontal and vertical two-dimensional multi-round database comparison and calibration, realizes intelligent distribution of multiple compiling terminals by combining block labeling analysis and relying on topology scheduling, and completes the whole process from requirement interpretation to scheme output through intelligent review of semantic matching and deep checking.The application can improve deep requirement recognition coverage, reduce compiling conflict rate, unify review scale, and significantly improve bidding scheme compiling efficiency and response quality.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence natural language processing and multi-terminal collaborative document preparation technology, and in particular to an artificial intelligence-based multi-terminal collaborative solution preparation and review system and cloud platform. Background Technology

[0002] In bidding scenarios, the quality of a bid proposal directly determines the compliance and market competitiveness of the bid response. As the technical complexity and content dimensions of bidding projects continue to increase, bid documents typically require collaboration across multiple terminals by personnel from different roles, including business, technical, and implementation services. Current technologies generally involve three stages in the preparation and review process: the first stage involves manual review of the bidding requirements, breaking down key points, and assigning them to individual preparers based on experience; the second stage involves each preparer independently completing their respective chapters, followed by manual aggregation; and the third stage involves manual review for quality verification. Based on the applicant's research and actual project verification, current technologies exhibit three clear shortcomings: First, the requirement breakdown is too simplistic, leading to a high rate of omission of deep constraints. Existing requirement breakdowns are mostly single-dimensional itemizations, only covering the surface-level requirements explicitly stated in the bidding text. For implicit derivative constraints and hierarchical depth requirements, the coverage rate fluctuates greatly due to the experience of personnel, with the industry average omission rate of deep requirements being approximately 18% to 25%, easily leading to response deviations or even bid rejection. Second, multi-terminal collaborative scheduling relies heavily on human experience, resulting in a high rate of conflict during preparation. Current task allocation and scheduling rely heavily on the subjective judgment of managers, lacking a quantifiable task-person matching model and a dependency-driven automatic scheduling mechanism. Industry statistics show that content and timing conflicts account for approximately 15% of the total work under the manual allocation model, and repeated modifications hinder overall work efficiency. Third, the review process suffers from poor consistency and high labor costs. Current quality reviews heavily depend on the professional capabilities of reviewers, with significant differences in review standards among different personnel. Furthermore, verifying each requirement item by item is labor-intensive and inefficient; the manual review cycle for medium-sized projects typically requires 2-3 working days. To address these issues, NLP-based bidding requirement extraction tools have emerged in the industry, but these tools largely remain at the level of keyword extraction and simple classification, failing to effectively solve the core problems of missing requirements and inefficient collaboration. In addition, existing tender preparation tools only support simple online text sharing among multiple users, lack an end-to-end collaborative task scheduling mechanism for multiple preparation terminals, have not built a multi-agent system collaborative allocation model, and do not have an integrated data interaction architecture for end-to-cloud and cloud-to-cloud collaboration, nor do they have a computing power layering processing logic for cloud-edge collaboration and cloud-edge fusion. The data fragmentation between terminal tasks, cloud databases, and AI computing power further exacerbates the deviation in demand breakdown and the conflict in preparation time. Summary of the Invention

[0003] This invention provides a multi-terminal collaborative solution development and review system and cloud platform based on artificial intelligence to solve the aforementioned technical problems.

[0004] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, comprising: The initial interpretation module is used to perform a preliminary screening of the bidding requirements document through a pre-trained large language model, and output horizontal interpretation information and in-depth requirement mining chain; The first and second comparison modules are used to convert the chain attribute combination of the horizontal demand chain and the horizontal center corresponding to each horizontal interpretation information into a semantic vector and then perform the first similarity matching with the historical demand chain in the existing database to obtain the calibrated first bidding requirements. Simultaneously, all horizontal centers, their corresponding core importance, and the business relationships between horizontal centers are integrated to generate the first center combination. The first center combination is then matched with the set of historical project core requirements in the existing database to obtain the calibrated second bidding requirements. The vertical analysis module is used to perform core node filtering on each vertical interpretation sequence to obtain the vertical center; Perform pairwise alignments on all longitudinal interpretation sequences to obtain longitudinal subsets. At the same time, integrate all longitudinal centers and deep statistical features to generate a second center combination. The third comparison module is used to perform a third matching between the average depth features of each vertical subset and the core node vector group and the second center combination with the historical depth requirement data in the existing database, so as to obtain the third bidding requirements for each vertical subset and the fourth bidding requirements for the second center combination. The roughing and mapping module is used to integrate the first, second, third, and fourth bidding requirements, generate a rough template outline according to the standard structure of the tender document, and combine the capability characteristics of each preparation end to perform block-preparation end mapping matching and preparation sequence with the final preparation sequence number of each block to obtain the collaborative preparation method of each outline area block; The compilation and review module is used to schedule each compilation end to complete the compilation of the corresponding outline area block according to the collaborative compilation method, generate the initial bid document, perform intelligent scoring and review on the initial bid document, and output the review report.

[0005] Preferably, one horizontal interpretation information corresponds to one deep demand mining chain, and one deep demand mining chain corresponds to one vertical interpretation sequence; each vertical interpretation sequence contains at least one vertical center, and each horizontal interpretation information contains at least one horizontal center; the vertical pair contains the vertical depth and vertical center of each of the two vertical interpretation sequences. The horizontal interpretation information set consists of several horizontal interpretation information divided by business domain. The vertical interpretation sequence set contains a vertical interpretation sequence that corresponds one-to-one with each horizontal interpretation information. Each vertical interpretation sequence is obtained by hierarchical structure transformation of the deep requirement mining chain of the corresponding horizontal interpretation information. The deep requirement mining chain is a logically progressive requirement node chain from surface requirements to deep constraints.

[0006] Preferably, the initial interpretation module includes: The macro analysis unit is used to perform macro-theme decomposition on the tender requirements document to obtain several global sub-requirements under each horizontal interpretation information, and then perform fine-grained sentence segmentation and entity extraction on each global sub-requirement to obtain several local detailed requirements points; The horizontal chain building unit is used to generate a horizontal requirement chain based on the business connection sequence, scope of action and enforcement level of each local detailed requirement point under the same global sub-requirement; At the same time, from all the global sub-requirements of the same horizontal interpretation information, the first key sub-requirement and the second key sub-requirement are extracted by comprehensive weight sorting and forced point number statistics, and together they form the horizontal center of the corresponding horizontal interpretation information. The derivative tree construction unit is used to perform semantic clustering on all global sub-requirements under each horizontal interpretation information, using the initial demand mining chain in the initial screening results as the initial skeleton, to obtain several clusters. It matches the first matching sub-requirement and its lower-level derivative sub-requirements corresponding to each cluster from the historical bidding files in the existing database, constructs the derivative subtree of the corresponding cluster according to the hierarchical relationship, and integrates all derivative subtrees to obtain the derivative tree of the corresponding horizontal interpretation information. The top-down and bottom-up analysis and summarization units are used to perform top-down filtering and bottom-up backtracking on the derivative tree to determine the core branch lines from the parent node to the child node. The core branch lines are then combined and summarized to obtain the corresponding horizontal interpretation information depth mining chain.

[0007] Preferably, the vertical analysis unit includes: The analysis sub-unit is used to convert the deep demand mining chain into a vertical interpretation sequence with hierarchical numbering. Then, it calculates the comprehensive core weight of each demand node in the sequence through the node weight calculation formula, and selects nodes with weights greater than or equal to a preset threshold as core nodes. If the weights of all nodes are lower than the preset threshold, then at least one node with the highest weight is selected as the core node to ensure that each sequence has at least one core node. The depth and center sub-units are used to count the number of core nodes at each level in the vertical interpretation sequence. The original vertical depth is calculated by combining the level weights of each level, and then the normalized vertical depth is obtained by normalizing the maximum depth of the total levels. At the same time, the core node at the deepest level in the vertical interpretation sequence is selected as the vertical center.

[0008] Preferably, the up-down analysis and summarization unit includes: The configuration subunit is used to pre-configure the demand intensity weight and logical dependency attributes for each node of the derivative tree, set the demand transmission connection verification conditions for each parent-child node link, and preset the same-level branch degradation and replacement rules and the fallback termination conditions. The fallback termination condition is: the verification is terminated when the valid branch is not selected after falling back to the root node of the derivative tree. The deep demand mining chain is constructed according to the three-level fixed hierarchical structure. The backtracking subunit is used to select the leaf node with the highest demand intensity weight in the corresponding derivative subtree of each cluster as the main anchor point and the second highest leaf node as the secondary anchor point. Starting from the anchor point, the backtracking traversal is performed on the derivative tree from bottom to top, and the traversal path is recorded. Candidate building subunits are used to perform forward and backward association extensions along the logical dependency edges of nodes during traversal. The dependency strength is calculated based on the business dependency annotation rate and the propagation verification pass rate. The dependency strength of nodes with no historical co-occurrence is directly set to 0. Extract nodes that have a strong dependency relationship with the anchor point to form a candidate branch node set; The candidate set adopts the path integrity priority rule, and the main path from the anchor point to the root node and the initial trunk path selected from top to bottom are forcibly retained in their entirety. The size limit is set for associated supplementary nodes. At the same time, implicit logical dependency edges across levels are added, implicit related nodes with excessive level differences and no intermediate connecting nodes are removed, and directed acyclic graph detection is performed simultaneously to eliminate loops and complete the link associations of the candidate branch node set. The verification of the substitute subunit is used to perform a link-by-link validity verification on each candidate branch in the candidate branch node set according to the requirement transmission connection verification conditions. Nodes introduced by implicit dependency edges need to be transmitted and verified separately. If the current candidate branch fails the verification, the substitute branch will be switched according to the same level branch downgrade and replacement rule and the verification will be repeated. If all branches at the same level fail the verification, the process will roll back to the next higher level and reselect the anchor point to perform the verification until a valid branch is found or the rollback termination condition is triggered. The route generation subunit is used to sort and integrate the selected effective branches in a top-down hierarchical order. The main path node is used as the trunk, and the associated supplementary nodes are merged into the corresponding node group according to the hierarchy to obtain the core branch route from the parent node to the child node.

[0009] Preferably, the coarse mapping module includes: The block attribute parsing unit is used to perform attribute parsing on each outline area block in the rough template outline. It generates compilation difficulty tags based on the vertical depth coverage level and the horizontal center association number, dependency attribute tags based on the horizontal requirement chain scope overlap, and requirement mandatory tags based on the proportion of substantive requirements, forming the tag group for the corresponding outline area block. The auxiliary unit is used to determine the compilation thread of each compilation end based on its historical compilation performance data, extract the corresponding expertise, ability level and adaptation time interval of the compilation thread, and construct the capability auxiliary group of the compilation end. The mapping analysis unit is used to encapsulate the content domain attributes, outline position, and compilation task of each outline area block into a block triple, perform adaptation calculation on the block triple and the capability auxiliary group of each compilation end, select the compilation end with the highest adaptation as the mapping execution end of the block, and obtain the initial mapping position of the block in the corresponding compilation end queue. The sorting unit is used to arrange all outline area blocks in the original outline order to generate the first trajectory, and to mark the initial mapping position of each block at the corresponding position of the first trajectory. The topology generation unit is used to construct a dependency directed graph based on the dependency attributes of all outline region blocks, and the topology sorting algorithm is used to adjust the dependency constraints of the first trajectory. For blocks with multiple preceding dependencies, they are placed after the highest sorting position of all preceding blocks to finally generate the compilation trajectory; the sequence number calibration unit is used to calibrate the final position of each outline area block on the compilation trajectory, and assigns a sequence block number to each block in chronological order. The method acquisition unit is used to combine the block number, mapping compilation information and tag group of each outline area block to generate the collaborative compilation method of the corresponding block.

[0010] Preferably, the method acquisition unit includes: The matching subunit is used to obtain the entry with the highest matching degree with the block triplet and the tag group from the pre-constructed double-group-compilation reference table through a hybrid distance matching algorithm, and output the corresponding initial compilation method. The double-group-compilation reference table is constructed based on the compilation records of historical winning projects through clustering and expert verification, and stores the mapping relationship between block feature combinations and corresponding compilation specifications. The additional sub-unit is used to attach the block number and mapping information of the block to the initial compilation method to generate the final collaborative compilation method.

[0011] Preferably, the compilation and review module includes: The compilation unit is used to schedule the corresponding compilation end to complete the compilation of the content of the outline area block in a collaborative compilation manner, and generate the first sub-file of the corresponding block. The layout processing unit is used to assemble all the first sub-files in the outline order and generate the initial tender document after unifying the format specifications; the review unit is used to verify the response coverage of the tender requirements through semantic similarity matching, verify the response depth through title level comparison, verify the format compliance through rule matching, and finally quantify the score according to the scoring rules and generate a review report containing deduction items and optimization suggestions.

[0012] This invention provides a cloud platform for multi-terminal collaborative scheme development and review based on artificial intelligence, used to execute any of the aforementioned multi-terminal collaborative scheme development and review systems based on artificial intelligence.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: 1. Significantly improved coverage of demand breakdown. Through horizontal and vertical dual-dimensional breakdown and derivative tree backtracking mining, the identification coverage of deep hidden demands is improved compared with the existing single-dimensional breakdown, effectively reducing the risk of bid rejection caused by demand omission.

[0014] 2. Significantly improved multi-terminal collaboration efficiency. Through quantitative matching and dependency topology scheduling, the task conflict rate is reduced, the overall development cycle is shortened, and the reliance on the experience of management personnel is decreased.

[0015] 3. Improved consistency and efficiency in review. Intelligent review improves the accuracy of requirement verification, shortens the review cycle, and ensures consistent review standards, avoiding subjective bias from human reviewers.

[0016] 4. This invention establishes a complete end-to-end collaboration, end-to-cloud collaboration, and cloud-to-cloud collaboration framework. It relies on a multi-agent system to complete the intelligent division of labor and scheduling of multiple terminals, and is equipped with a cloud-edge collaboration and cloud-edge fusion hierarchical computing power allocation mode. It opens up the data links of terminal programming, cloud storage, and AI model operation, and eliminates multi-terminal data silos and computing power waste.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a multi-terminal collaborative scheme preparation and review system based on artificial intelligence, as described in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Example 1:

[0021] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, such as... Figure 1 As shown, it includes: The initial interpretation module is used to perform a preliminary screening of the bidding requirements document through a pre-trained large language model, and output horizontal interpretation information and in-depth requirement mining chain; The first and second comparison modules are used to convert the chain attribute combination of the horizontal demand chain and the horizontal center corresponding to each horizontal interpretation information into a semantic vector and then perform the first similarity matching with the historical demand chain in the existing database to obtain the calibrated first bidding requirements. Simultaneously, all horizontal centers, their corresponding core importance, and the business relationships between horizontal centers are integrated to generate the first center combination. The first center combination is then matched with the set of historical project core requirements in the existing database to obtain the calibrated second bidding requirements. The vertical analysis module is used to perform core node filtering on each vertical interpretation sequence to obtain the vertical center; Perform pairwise alignments on all longitudinal interpretation sequences to obtain longitudinal subsets. At the same time, integrate all longitudinal centers and deep statistical features to generate a second center combination. The third comparison module is used to perform a third matching between the average depth features of each vertical subset and the core node vector group and the second center combination with the historical depth requirement data in the existing database, so as to obtain the third bidding requirements for each vertical subset and the fourth bidding requirements for the second center combination. The roughing and mapping module is used to integrate the first, second, third, and fourth bidding requirements, generate a rough template outline according to the standard structure of the tender document, and combine the capability characteristics of each preparation end to perform block-preparation end mapping matching and preparation sequence with the final preparation sequence number of each block to obtain the collaborative preparation method of each outline area block; The compilation and review module is used to schedule each compilation end to complete the compilation of the corresponding outline area block according to the collaborative compilation method, generate the initial bid document, perform intelligent scoring and review on the initial bid document, and output the review report.

[0022] In this embodiment, the bidding requirements document refers to a collection of formal documents issued by the bidding entity that specify all eligibility requirements, technical specifications, business rules, and formatting requirements for the bidding project. It is typically an editable electronic document and includes sub-files such as the bidding announcement, instructions to bidders, technical specifications, scoring rules, and contract templates. The system first extracts the full text, paragraph levels, heading structure, and table content using the pdfplumberPDF parsing library or the python-docxWord parsing library to form structured text data, which serves as input for subsequent processing.

[0023] The unified execution standard for full-text semantic encoding is as follows: the BERT-base-chinese pre-trained model is used to perform semantic encoding on single-sentence text, and the output vector corresponding to the CLStoken is taken as the 768-dimensional semantic vector of the text; for overall texts such as demand chains and demand sets composed of multiple sentences, the arithmetic mean of all single-sentence vectors is taken as the overall semantic vector.

[0024] In this embodiment, the initial interpretation module serves as the system's input, employing a two-stage processing mechanism of initial screening and refinement: the initial screening stage rapidly outputs a rough result (initial screening result) as the initial skeleton, while the refinement stage involves subsequent requirement refinement links to complete fine-grained expansion and verification. The advantage of this design is that while large language models are fast to process, their accuracy is unstable. The subsequent rule-based refinement process can correct initial screening biases and expand deeper requirements, balancing efficiency and accuracy.

[0025] In the initial screening stage, open-source pre-trained large language models (such as DeepSeek-V2 and Qwen-7B-Chat) are used to perform coarse-grained interpretation. The specific operation is as follows: The structured tender requirement document is input into the large language model along with structured prompts. The prompt template is: "Please interpret the following tender requirement document, dividing the information into four business domains: business requirements, technical requirements, service requirements, and format requirements. For each piece of information, starting from the surface requirements, deduce the direct requirements, implicit requirements, and derived constraints layer by layer to generate the corresponding deep requirement mining chain. The output format is JSON, with the structure {"List of Horizontal Interpretation Information":[{"Information Name":"","Deep Requirement Mining Chain":["Node 1","Node 2","Node 3"]}]}."

[0026] The initial screening stage outputs several preliminary horizontal interpretation information and corresponding preliminary in-depth requirement mining chains. For example, after performing an initial screening on the bidding requirements document ZB for a certain government informatization project, three preliminary horizontal interpretation information are obtained: a1 qualification and business requirements, a2 system function requirements, and a3 implementation service requirements. Among them, the preliminary in-depth requirement mining chain corresponding to a1 is [b11 possessing system integration qualifications, b12 providing scanned copies of qualification certificates, and b13 qualification certificates being valid].

[0027] The initial depth requirement mining chain in the preliminary screening results directly serves as the initial hierarchical skeleton of the derivative tree. The nodes obtained from the preliminary screening correspond to the core nodes of levels 1-3 of the derivative tree. Subsequent historical derivatives require expansion and supplementation of deeper-level nodes based on this skeleton. Simultaneously, the preliminary screening results serve as a verification benchmark for the refinement stage. If the node overlap between the final refined output and the preliminary screening results is less than 40%, manual review is triggered to prevent the refinement process from deviating from the core direction of the original tender document. The 40% threshold is determined based on testing with 500 historical winning bid samples, ensuring flexibility in decomposition while controlling deviation risk. Different industries can fine-tune this threshold within ±10% based on their business characteristics.

[0028] The classification correction rules for the initial screening results are as follows: During the refinement stage, the horizontal classification of the initial screening is verified and adjusted. The rule is as follows: calculate the semantic similarity between each preliminary requirement item and the standard corpus of each business domain. If the similarity with the current category is less than 0.75 and the similarity with another category is greater than 0.80, then the item is moved to the corresponding category. After the migration, if a category has fewer than 2 items, they are merged into the adjacent category with the closest semantics. This rule ensures the accuracy of the final horizontal interpretation of information classification and avoids classification bias in the initial screening stage.

[0029] Ultimately, all horizontal interpretation information is aggregated to form a horizontal interpretation information set, and all vertical interpretation sequences are aggregated to form a vertical interpretation sequence set.

[0030] Horizontal interpretation of information involves categorizing bidding requirements by business domain from a broad perspective. Each category corresponds to an independent major direction of bidding requirements, such as qualification and business requirements, system function requirements, etc.

[0031] The deep requirement mining chain is a chain of requirement nodes that are logically progressive and obtained by digging down from the surface requirements layer by layer for a single horizontal interpretation of information, and are connected in a logical manner from general to specific.

[0032] The vertical interpretation sequence is an ordered sequence of the deep demand mining chain after hierarchical structuring. Each node is labeled with its level number and parent node identifier, arranged from highest to lowest level. The transformation rule is: the root node of the deep demand mining chain is level 1, the direct child nodes of the root node are level 2, and so on.

[0033] For example, the deep demand mining chain: data transmission encryption → adoption of national cryptographic algorithms → support for SM2 / SM4 algorithm adaptation → automatic key rotation periodically, can be converted into the following vertical interpretation sequence: Level 1: Data transmission encryption (parent node: none) Level 2: Adoption of national cryptographic algorithms (parent node: data transmission encryption) Level 3: Support for SM2 / SM4 algorithm adaptation (parent node: adoption of national cryptographic algorithms) Level 4: Automatic key rotation periodically (parent node: support for SM2 / SM4 algorithm adaptation).

[0034] The horizontal calibration requirements (first and second tender requirements) form the basis for surface compliance, while the vertical calibration requirements (third and fourth tender requirements) supplement the deep constraints. When there is a conflict between the two, the horizontal surface compliance requirements shall be given priority, and the vertical deep constraints shall be written as optional bonus items in the remarks, and shall not be mandatory to respond to.

[0035] The existing database is a structured database, using MySQL to store metadata (project type, industry, scale, score, etc.) and Milvus vector database to store 768-dimensional semantic vectors (generated by BERT-base-chinese encoding) of the requirement text and solution text, supporting keyword retrieval and cosine similarity matching. The database stores bidding requirement entries, corresponding bid solution fragments, and review scores for over 500 successful bids from the same industry over the past three years. The database construction and annotation specifications are as follows: all requirement entries are decomposed and annotated by at least two professionals with over five years of bidding experience according to a unified standard. Annotation dimensions include requirement level, enforcement level, business domain, and dependencies; annotation results are cross-validated, and only entries with a consistency rate of over 90% are included in the database; the data is subject to quality sampling inspection every quarter, and entries that fail the inspection are re-annotated or removed to ensure data consistency and availability. It should be noted that industry classification filtering is performed before database matching, only retrieving historical requirement data consistent with the industry tags of the current bidding project. When no samples from the same industry are available, the matching weight is automatically reduced and manually reviewed; strong cross-industry matching of derived requirements is prohibited.

[0036] The first comparison uses the chain attribute combination of a single horizontal demand chain as a unit, performing similarity matching with historical demand chains. The chain attribute combination consists of the corresponding horizontal demand chain and the horizontal center. The matching method is as follows: the horizontal demand chain is converted into a 768-dimensional semantic vector after average pooling; cosine similarity is calculated with the vectors of historical demand chains in the database; historical entries with a similarity ≥ 0.80 are matched, and the corresponding standardized descriptions and supplementary refinements of the demand are extracted. The demand points of the original horizontal demand chain are merged with the matched historical supplementary points to remove duplicates, and then logically reorganized to obtain the calibrated first bidding requirements. The purpose of the first comparison is to fill in any missing refinements within a single demand chain.

[0037] The second comparison is performed on the first center combination as a unit, and a global match is performed with the core requirements set of historical projects. The first center combination consists of all horizontal centers, the core importance of each horizontal center, and the business relevance between the horizontal centers.

[0038] In this embodiment, the importance of the horizontal center represents the degree to which a single horizontal center serves as a core node for this type of requirement, and the calculation formula is as follows: ,in, Let be the core importance of the i-th horizontal center, with a value range of [0,1]. The number of times this requirement appears in the tender document; This represents the maximum number of occurrences among all requirement points within the same horizontal interpretation of information. This is used as a core item identifier for scoring. If the requirement appears in the core items of the scoring rules, it takes the value 1; otherwise, it takes the value 0. The requirement is identified as a substantive requirement. If the requirement is marked as a substantive clause, it takes a value of 1; otherwise, it takes a value of 0. All three parameters are normalized to the [0,1] range, ensuring consistent dimensions. The weighting is based on the following: substantive clauses directly determine the risk of bid rejection and have the greatest impact on the bid, therefore they have the highest weight (0.4); the frequency of occurrence and the impact of core scoring items on the degree of importance are roughly equivalent, each accounting for 0.3. The formula has a clear physical meaning: even if a substantive clause appears infrequently, its core importance is higher than that of high-frequency non-substantive clauses, which aligns with the actual logic of bidding processes.

[0039] In this embodiment, the horizontal center business correlation degree characterizes the degree of business coupling between two horizontal centers in the implementation of the solution. It is jointly measured by two independent dimensions: co-occurrence prevalence and dependency strength. The calculation formula is as follows: ,in, The business correlation degree between horizontal center i3 and horizontal center j3, with a value range of [0,1]. This refers to the number of projects in the historical winning bid schemes where the demand points corresponding to horizontal center i3 and horizontal center j3 appear simultaneously. This represents the maximum number of items that appear simultaneously in all horizontal center pairs; In the historical winning bids, the number of projects marked as having a clear business constraint dependency relationship between horizontal center i3 and horizontal center j3 is as follows: (The two requirements did not co-occur in the historical project) or If there are no co-occurring samples in the database, then the business relevance is determined directly. Fractional calculations are not performed to avoid division by zero errors.

[0040] In this embodiment, the core requirement vector group of the first central combination is compared with the core requirement vector group of historical projects in the database to calculate the overall similarity. Historical projects with a similarity ≥ 0.80 are matched, and their core requirement dimensions are extracted as supplements. The original core requirements and historical supplementary requirements are merged and deduplicated to obtain the second bidding requirements. The purpose of the second comparison is to fill in the gaps in the core dimensions at the project level and avoid the absence of an overall requirement framework.

[0041] In this embodiment, similarity analysis is used to group longitudinal interpretation sequences with similar features into the same longitudinal subset. Similarity analysis is performed on longitudinal pairs; each pair consists of two longitudinal interpretation sequences, including their longitudinal depth and longitudinal center. The similarity calculation formula is: ,in, The similarity between sequences i4 and j4 is determined vertically, with values ​​ranging from [0,1]. , Normalized longitudinal depth values ​​for the longitudinal interpretation sequences i4 and j4; The mean cosine similarity between the core node vectors of sequence i4 and j4 is used for longitudinal interpretation. These are the semantic average vectors of all core nodes in the vertical interpretation sequences i4 and j4, respectively. The vertical center matching degree is 1 if the vertical centers of the two sequences belong to the same demand domain (determined by BERT classification), otherwise it is 0.

[0042] Since the depth values ​​have been normalized to [0,1], the depth difference term can be calculated directly using the absolute difference, without further normalization. When similarity... When two vertical interpretation sequences are highly similar, they are determined to be highly similar. The vertical subset is divided using the undirected graph connected component method: each vertical interpretation sequence is regarded as a graph node, and undirected connecting edges are added between sequence pairs with similarity ≥ 0.7. All sequences within the same connected component are assigned to the same vertical subset; if the similarity of a single sequence to all other sequences is less than 0.7, it is treated as a separate vertical subset.

[0043] In this embodiment, all vertical center and depth requirement attributes are integrated to generate a second center combination. The depth requirement attributes are statistical characteristics of all vertical depths, including the project's average normalized vertical depth, maximum normalized vertical depth, and depth distribution percentage. The second center combination represents the core set of requirements for the overall depth dimension of the project.

[0044] In this embodiment, the third comparison module performs two rounds of matching calibration in the vertical dimension: Vertical subset matching: The average normalized vertical depth and core node vector group of each vertical subset are compared with the similar depth requirement subsets in the database. The similarity is calculated with a threshold of 0.80. After matching, historical depth requirement supplementary points are extracted, and after merging and deduplication, the third bidding requirements are obtained.

[0045] Second center combination matching: Perform global matching between the deep core requirement vector group of the second center combination and the deep core requirements of historical projects, with a threshold of 0.80, to supplement the project-level deep core requirements and obtain the fourth bidding requirements.

[0046] Horizontal and vertical decomposition represent two perspectives on the same bidding requirement, not two independent sources of demand, and should not inherently conflict under normal circumstances. Before merging the requirements from four rounds of bidding to generate a rough template outline, the following processing rules are applied: Complete deduplication rule: If the semantic similarity is ≥0.85, retain the one with the most complete expression and the highest enforcement level.

[0047] Partial overlap handling rules: If requirement A is a subordinate specific implementation of requirement B (superior-subordinate relationship), then the superior requirement B is retained, and requirement A is included in the corresponding hierarchical structure as a subordinate child node of B, without duplication or parallel retention.

[0048] If conflicting statements arise from different sources, the original tender document for the corresponding requirement should be traced back first, and the decomposition results should be corrected based on the original tender document. If there is a mismatch in the hierarchy, the hierarchy numbering of the vertical sequence should be adjusted, and conflicting statements should not be retained. If the original document cannot be used for judgment, manual review should be triggered.

[0049] Priority rule: Requirements with the same wording are sorted from highest to lowest level of enforceability: substantive clauses > general clauses > bonus items; at the same level, the original requirements in the tender document have higher priority than historical supplementary requirements. This rule aligns with the essential attributes of both horizontal and vertical dimensions, resolving conflicts by correcting errors in the decomposition process.

[0050] It should be noted that all thresholds matching cosine similarity can be automatically switched based on the project industry and database sample size, and are not fixed thresholds. Example 2:

[0051] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence. The initial interpretation module includes: The macro analysis unit is used to perform macro-theme decomposition on the tender requirements document to obtain several global sub-requirements under each horizontal interpretation information, and then perform fine-grained sentence segmentation and entity extraction on each global sub-requirement to obtain several local detailed requirements points; The horizontal chain building unit is used to generate a horizontal requirement chain based on the business connection sequence, scope of action and enforcement level of each local detailed requirement point under the same global sub-requirement; At the same time, from all the global sub-requirements of the same horizontal interpretation information, the first key sub-requirement and the second key sub-requirement are extracted by comprehensive weight sorting and forced point number statistics, and together they form the horizontal center of the corresponding horizontal interpretation information. The derivative tree construction unit is used to perform semantic clustering on all global sub-requirements under each horizontal interpretation information, using the initial demand mining chain in the initial screening results as the initial skeleton, to obtain several clusters. It matches the first matching sub-requirement and its lower-level derivative sub-requirements corresponding to each cluster from the historical bidding files in the existing database, constructs the derivative subtree of the corresponding cluster according to the hierarchical relationship, and integrates all derivative subtrees to obtain the derivative tree of the corresponding horizontal interpretation information. The top-down and bottom-up analysis and summarization units are used to perform top-down filtering and bottom-up backtracking on the derivative tree to determine the core branch lines from the parent node to the child node. The core branch lines are then combined and summarized to obtain the corresponding horizontal interpretation information depth mining chain.

[0052] The initial interpretation module includes a macro analysis unit, a horizontal chain construction unit, a derivative tree construction unit, and an up-down analysis and summarization unit in sequence according to the data flow. The output of the macro analysis unit is simultaneously sent to the horizontal chain construction unit and the derivative tree construction unit, and the two are processed in parallel: the horizontal chain construction unit outputs the horizontal demand chain and the horizontal center to the first and second comparison units, the derivative tree construction unit outputs the derivative tree to the up-down analysis and summarization unit, and finally the up-down analysis and summarization unit outputs the deep demand mining chain to the vertical analysis unit.

[0053] In this embodiment, the macro-analysis is based on the first-level headings and chapter divisions of the tender text, combined with the semantic topic classification of the BERT-base-chinese pre-trained model, to break down each piece of horizontal interpretation information into several independent first-level sub-topics, i.e., global sub-requirements. For example, the horizontal interpretation information of system functional requirements is macroscopically broken down into four global sub-requirements: system architecture, functional design, performance indicators, and security system.

[0054] Micro-analysis involves segmenting the text corresponding to a single global sub-requirement using the spaCy Chinese sentence segmentation tool. It then leverages BERT-based-Chinese named entity recognition capabilities to extract constraints, parameters, and formatting requirements, breaking them down into independent, locally refined requirements. For example, the global sub-requirement for a security system is micro-decomposed into four locally refined requirements: transmission encryption, storage encryption, authentication, and log auditing. Each locally refined requirement corresponds to a specific constraint in the tender text; for instance, using the SM4 algorithm for data storage encryption is a locally refined requirement.

[0055] A horizontal requirement chain is a chain-like structure of detailed local requirements arranged in the order of business logic. Its construction is based on three dimensions: Connection order: The order is determined according to the sequence of business processes and document chapters corresponding to the requirements, such as identity authentication → permission verification → operation auditing.

[0056] Scope of application: The content coverage area of ​​a single localized refinement requirement in the tender document, matching its chapter position in the standard tender template based on the content attributes of the requirement.

[0057] Depth of Influence: The mandatory level of a single detailed requirement is divided into three levels: Substantive Clauses (level value 1.0, failure to meet will result in rejection), General Clauses (level value 0.5, failure to meet will result in point deduction), and Bonus Items (level value 0.2, meeting these will result in additional points). The level marking is based on the semantic determination of identifying words such as substantive clauses, mandatory, should, and recommended in the tender document.

[0058] The local refinement requirements are sorted according to the connection order, and the scope and depth of each point are marked to generate a horizontal requirement chain.

[0059] The horizontal center is the key requirement point that occupies the core position in a single horizontal interpretation of information and plays a constraining and leading role in other requirement points. It is determined by the first key sub-requirement and the second key sub-requirement, and is selected from two dimensions: overall weight and mandatory constraint, to avoid omission of a single dimension.

[0060] The first key sub-requirement is selected from all global sub-requirements that interpret information in the same horizontal direction, choosing the top 1-2 based on their comprehensive weight. The formula for calculating the comprehensive weight is: ,in, Let be the first key weight of the i-th global sub-requirement, with a value range of [0,1]. The length (number of characters) of the text required by the i-th global sub-requirement; This refers to the total number of characters in all global sub-requirements under the same horizontal interpretation information. Let be the score of the scoring criteria corresponding to the i-th global sub-requirement; The total score is the score of all global sub-requirements corresponding to the scoring rules under the same horizontal interpretation of information; This is used to identify the core objective. If this global sub-requirement is a core objective of the project, it takes the value 1; otherwise, it takes the value 0.

[0061] The second key sub-requirement: Count the number of local detailed requirements with high enforcement level (effect depth ≥ 0.5, i.e., substantive clauses and general clauses) contained under each global sub-requirement, and select the global sub-requirement with the largest number.

[0062] When the first key sub-requirement and the second key sub-requirement are the same global sub-requirement, the horizontal center only contains the core requirement points corresponding to that sub-requirement; when they are different, the horizontal center contains the core requirement points corresponding to both global sub-requirements.

[0063] The derivative tree building unit uses the initial deep demand mining chain obtained from the initial screening as the initial skeleton, and expands to generate candidate derivative demands by matching historical winning bid schemes, providing a data foundation for deep mining.

[0064] Semantic clustering involves performing K-means clustering on all global sub-requirements for each lateral interpretation, resulting in several clusters. The specific implementation standard is as follows: first, L2 normalization is applied to the semantic vectors of all global sub-requirements; then, standard K-means clustering is performed, equivalent to a cosine distance metric. Initial cluster centers are selected using the k-means++ algorithm, with an iterative convergence threshold of cluster center offset less than 1e-4. The number of clusters k is determined using the silhouette coefficient method, which is based on cosine distance calculation: for each sample, the average cosine distance 'a' with other samples in the same cluster and the average cosine distance 'b' with the nearest sample from a different cluster are calculated, along with the silhouette coefficient for each sample. The average of all samples is the overall silhouette coefficient. The iteration range is from k=2 to k=min(10, number of global sub-requirements - 1), selecting the k value with the largest silhouette coefficient. When the number of global sub-requirements is ≤2, each global sub-requirement is treated as a separate cluster. The center vector of each cluster is the arithmetic mean of all global sub-requirement vectors within the cluster.

[0065] In this embodiment, historical matching and derivative tree construction involve performing a two-level matching process from historical winning bid files stored in an existing database. The first matching sub-requirement: Calculate the cosine similarity between the center vector of the current cluster and the global sub-requirement vector in the historical file, and select the historical sub-requirement with the highest similarity; the similarity threshold is set to 0.75. If the highest similarity is lower than 0.75, it is determined that there is no matching item, the cluster is not extended, and a simplified tree structure is constructed.

[0066] Derivative sub-requirements: Extract all subordinate sub-requirements and detailed descriptions under the first matching sub-requirement in the historical files as candidate derived requirements.

[0067] Using the nodes of the initial deep demand mining chain as the upper-level skeleton, and a single first matching sub-requirement as the extended root node of the corresponding branch, with all its derived sub-requirements as child nodes, a derivative subtree is constructed according to the hierarchical relationship; the derivative subtrees of all clusters under the same horizontal interpretation information are integrated to form a complete derivative tree.

[0068] The simplified tree structure definition for no matching items is as follows: the root node is the corresponding horizontal interpretation information, all global sub-requirements within the cluster are the second-level nodes, and the local refinement requirements under each global sub-requirement are the third-level nodes, directly forming a three-level tree structure without additional derivation or extension. This rule covers boundary scenarios and ensures that the process can still be executed normally when there is no historical match.

[0069] It should be noted that the derived requirements are only candidate requirements and will not be directly included in the final requirements list; the subsequent analysis and summarization modules will use a strict verification mechanism to screen valid derived requirements and avoid introducing irrelevant content.

[0070] In this embodiment, the top-down analysis includes two directions: filtering the main branches from top to bottom and backtracking from bottom to top to verify the effectiveness of demand transmission. The two-way cooperation ensures the accuracy of the core branches.

[0071] The top-down selection process starts from the root node of the derived tree and extends layer by layer to form the initial trunk path. The output is used for two purposes: first, to prioritize the inclusion of candidate branch nodes to ensure that the trunk direction is not lost; second, to serve as a verification benchmark for the bottom-up backtracking results. The specific execution steps are as follows: Step 01: Take the root node of the derived tree as the starting point of the path and mark it as the node of the current level; Step 02: Sort all direct child nodes of the current level node in descending order of demand intensity weight; Step 03: Select the child node ranked first and include it in the initial trunk path, marking it as the current node of the next level; Step 04: Repeat steps 02-03 until the leaf node is reached, and output the initial trunk path.

[0072] If the overlap between the candidate branches obtained from the bottom up and the nodes of the initial trunk path is less than 30%, a secondary verification is triggered: the upper limit of the size of the associated nodes in the candidate set is increased by 50%, and candidate construction and screening are re-executed; the maximum number of retries is 2; if the overlap still does not meet the standard after 2 verifications, core branches are forcibly generated based on the initial trunk path, and the cluster is marked as a manually reviewed item. The threshold of 30% and the number of retries are both determined based on testing with 500 historical samples to balance the efficiency of automatic processing and the accuracy of results. Example 3:

[0073] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, wherein the up-down analysis and summarization unit includes: The configuration subunit is used to pre-configure the demand intensity weight and logical dependency attributes for each node of the derivative tree, set the demand transmission connection verification conditions for each parent-child node link, and preset the same-level branch degradation and replacement rules and the fallback termination conditions. The fallback termination condition is: the verification is terminated when the valid branch is not selected after falling back to the root node of the derivative tree. The deep demand mining chain is constructed according to the three-level fixed hierarchical structure. The backtracking subunit is used to select the leaf node with the highest demand intensity weight in the corresponding derivative subtree of each cluster as the main anchor point and the second highest leaf node as the secondary anchor point. Starting from the anchor point, the backtracking traversal is performed on the derivative tree from bottom to top, and the traversal path is recorded. Candidate building subunits are used to perform forward and backward association extensions along the logical dependency edges of nodes during traversal. The dependency strength is calculated based on the business dependency annotation rate and the propagation verification pass rate. The dependency strength of nodes with no historical co-occurrence is directly set to 0. Extract nodes that have a strong dependency relationship with the anchor point to form a candidate branch node set; The candidate set adopts the path integrity priority rule, and the main path from the anchor point to the root node and the initial trunk path selected from top to bottom are forcibly retained in their entirety. The size limit is set for associated supplementary nodes. At the same time, implicit logical dependency edges across levels are added, implicit related nodes with excessive level differences and no intermediate connecting nodes are removed, and directed acyclic graph detection is performed simultaneously to eliminate loops and complete the link associations of the candidate branch node set. The verification of the substitute subunit is used to perform a link-by-link validity verification on each candidate branch in the candidate branch node set according to the requirement transmission connection verification conditions. Nodes introduced by implicit dependency edges need to be transmitted and verified separately. If the current candidate branch fails the verification, the substitute branch will be switched according to the same level branch downgrade and replacement rule and the verification will be repeated. If all branches at the same level fail the verification, the process will roll back to the next higher level and reselect the anchor point to perform the verification until a valid branch is found or the rollback termination condition is triggered. The route generation subunit is used to sort and integrate the selected effective branches in a top-down hierarchical order. The main path node is used as the trunk, and the associated supplementary nodes are merged into the corresponding node group according to the hierarchy to obtain the core branch route from the parent node to the child node.

[0074] In this embodiment, the configuration subunit pre-completes three configurations: Configure demand intensity weights and logical dependency attributes for each node in the derivative tree; Set demand propagation and connection verification conditions for each parent-child node link; Preset the rules for downgrading and replacing branches at the same level, and the conditions for regression and termination.

[0075] Demand intensity weight represents the importance and enforcement level of the demand corresponding to a single node, and the calculation formula is as follows: ,in, , represents the demand intensity weight, with a value range of [0,1]; L is the demand enforcement level value, with substantive requirements taking 1.0, general requirements taking 0.5, and bonus items taking 0.2; This is the score for the requirement corresponding to this node in the scoring criteria. The total score is the detailed scoring criteria for the relevant horizontal interpretation information. This refers to the number of projects in the same industry's historical bidding projects over the past three years where this requirement was explicitly marked as a substantive rejection clause; the statistical scope is consistent with the historical project scope of the existing database, and the marking is based on explicit statements in the bidding text such as substantive requirements and treatment as rejection for failure to meet them. This is the maximum frequency of rejection clauses across all nodes in the same derivative tree. All three parameters in this formula are normalized to the [0,1] interval, ensuring consistent dimensions. A higher rejection frequency indicates a higher compliance risk and stronger actual constraint, thus a higher strength weight. The physical meaning of the parameters perfectly matches the design objectives. Regarding the robustness of maximum value normalization: when the existing database project size reaches over 300 documents, the maximum frequency value of each node tends to stabilize, with an annual fluctuation of less than 5%, meeting the stability requirements of engineering applications; the database is calibrated annually to ensure comparability of the indicators.

[0076] Logical dependency attributes identify the logical constraint relationships between nodes and are divided into three categories: Supporting the parent node: The implementation of child nodes is a necessary condition for the implementation of the parent node; Complementary at the same level: Two nodes at the same level must both support the parent node and this condition must be met simultaneously. Sibling mutual exclusion: Two sibling nodes are alternatives, and only one needs to be chosen.

[0077] The requirement transmission connection verification condition is used to determine the transmission validity of the parent-child node link. It must meet two conditions at the same time: Condition 1: The fine-tuned Chinese text implication recognition model (BERT-base fine-tuned on CNLI dataset) determines that the child node requirement is the lower-level specific implementation of the parent node requirement, with an implication confidence of ≥0.75; Condition 2: The frequency of the parent-child requirement pair as a valid transmission pair in the historical bidding documents is ≥5 times.

[0078] This verification mechanism is based on one-way implication relationship determination, which can accurately distinguish between effective refined propagation and synonymous repetition: synonymous repetition has high bidirectional similarity but the implication relationship is not valid, while cross-domain specific implementations have low literal similarity but the implication relationship is valid. In principle, it solves the logical distortion problem of semantic similarity verification. The determination of the threshold of 0.75 and 5 times is based on the statistical data of 500 historical winning bid documents. The average implication confidence of effective propagation links is 0.86 and the standard deviation is 0.07. Taking the mean minus 1.5 standard deviations is approximately 0.75; the average co-occurrence frequency of effective propagation links is 12 times, and the lower quartile is 5 times. Therefore, setting this threshold can balance recall and precision.

[0079] Same-level branch downgrade and replacement rule: When a candidate branch fails verification, it is sorted by the demand intensity weight of nodes at the same level from high to low, and the branch containing the next node is selected as the replacement. Backtracking termination condition: If all branches at the same level fail verification, backtrack to the next higher level and reselect the anchor point; when backtracking to the root node of the derivative tree still fails to select a valid branch, the backtracking terminates, and a deep demand mining chain is constructed according to a fixed three-level structure: Level 1 is horizontal interpretation information, Level 2 is all global sub-requirements within the cluster, and Level 3 is the corresponding local refinement requirement points. This rule clarifies the hierarchical transformation standard in boundary scenarios, ensuring a unified input format for subsequent modules.

[0080] Backtracking through sub-units to select anchor nodes: For each cluster's derived subtree, the leaf node with the highest demand intensity weight is selected as the primary anchor, and the second highest is selected as the secondary anchor. Leaf nodes refer to nodes with a degree of 0 and no lower-level children. The primary and secondary anchors perform backtracking and verification independently, and finally, the branch with the highest comprehensive score is selected as the core branch. This dual-anchor mechanism reduces the risk of missing path failures when using a single anchor.

[0081] Backtracking starts from the anchor point, iterates through the parent nodes of the current node and visits them until the root node is reached, and records the complete traversal path.

[0082] During the traversal, candidate building units perform forward (child→parent) and backward (parent→child) association extensions along the logical dependency edges of nodes, calculate the dependency strength between nodes, and extract nodes with strong dependencies on anchor points to form a candidate branch node set. Dependency strength is calculated based on two independent dimensions: business dependency annotation rate and propagation validation pass rate, without relying on semantic similarity. The calculation formula is as follows: ,in, Let be the dependency strength between node i and node j, with a value range of [0,1]. The number of projects with two nodes that were manually marked as having strong business constraint dependencies among historical winning bids; For historical winning bids, the number of projects whose demand transmission between two nodes is determined to be valid through implicit model validation; The total number of current projects for the two nodes in their historical winning bids. If there is no co-occurrence in the past, then the dependency strength is determined directly. The two parameters measure the degree of dependency from two independent perspectives: manual business judgment and algorithm verification. They are not inclusive of each other: strong dependency identified manually is a business-level constraint judgment, while effective propagation is a technical-level implication verification. The weighted average of these two parameters measures the true dependency strength between nodes. When the dependency strength... When a strong demand dependency relationship is identified, it is included in the candidate branch node set.

[0083] The candidate set adopts a scale truncation rule that prioritizes path integrity: The first step forces the complete retention of all nodes on the main path from the anchor point to the root node, as well as the initial trunk path nodes selected from top to bottom; the second step supplements the remaining slots with related nodes according to the dependency strength from high to low, with the upper limit of related nodes set at 20, which can be automatically adjusted according to actual needs. For example, when the number of global sub-requirements of a cluster is >50, the upper limit is dynamically scaled to 30; when the number of clusters is <10, the upper limit is scaled to 10; nodes exceeding the upper limit are truncated in descending order of dependency strength.

[0084] The candidate building sub-units are supplemented with implicit logical dependency edges across levels. To control computational complexity, a two-layer screening strategy is adopted: the first step is to screen out node pairs with overlapping business domains through keyword matching and exclude cross-domain node pairs with no overlap; the second step is to perform dependency strength calculation only on node pairs with a level difference of 2 or 3, and nodes with a level difference greater than 3 have no direct implicit dependency by default.

[0085] Hierarchical break handling rules for implicit dependency edges: If the hierarchical difference between an implicitly related node and its corresponding node on the main path is ≥3, and there are no connecting nodes in the intermediate levels, then the implicitly related node is removed; if the hierarchical difference is 2, the node is retained but marked as a cross-level constraint, without disrupting the hierarchical progression of the main path. After adding implicit edges, a directed acyclic graph (DAG) detection is immediately performed: a directed graph is constructed using the Kahn topological sorting algorithm. If a cycle is detected, the implicit edge with the lowest dependency strength in the cycle is removed, and the detection is repeated until no cycle is found. This avoids the logical cycle problem introduced by implicit edges and ensures the feasibility of subsequent traversal.

[0086] In this embodiment, the verification substitute subunit performs a link-by-link validity check on each candidate branch, that is, it sequentially checks whether each parent-child link on the branch meets the requirement transmission and connection verification conditions. Nodes included through implicit dependency edges need to have their transmission validity with adjacent nodes checked individually; if the check fails, the node is removed. If all links meet the requirements, the branch is determined to be valid; if there are links that do not meet the requirements, the substitute branch is switched according to the same-level branch downgrade substitution rule and the verification is repeated; if all branches at the same level fail, the branch is backed up, and the process stops when the back-down termination condition is triggered.

[0087] After each of the main and secondary anchor points produces one valid branch, the final core branch is selected based on the overall score: ,in, is the average implied confidence of all parent-child links on the branch, i.e., the average transmission confidence, with a value range of [0,1]; H is the normalized vertical depth corresponding to the branch. This balances transmission quality and mining depth: a higher average transmission confidence indicates a higher degree of branch effectiveness, and a higher normalized depth indicates deeper mining. Weights are tilted towards transmission quality to avoid over-mining caused by simply pursuing depth. If the combined scores of the main and secondary anchor branches are equal, the branch with the larger normalized vertical depth H is selected first; if the depths are still the same, the branch with the higher average transmission confidence Conf is selected.

[0088] The route generation unit sorts and integrates the selected valid branches in a top-down hierarchical order: the main path node serves as the trunk node of the core branch, and the associated supplementary nodes and valid nodes introduced by implicit edges are merged into the corresponding node group according to their original level; if the semantic similarity of nodes in the same level is ≥0.85, they are merged and deduplicated. The final output is the core branch route labeled with the level and attributes of each node.

[0089] Finally, perform inductive combination on nodes of the same level (with the same level number and the same parent node) in the core branch line, package the core nodes of the same level into node groups, arrange them in hierarchical order, and obtain the corresponding horizontal interpretation information depth demand mining chain. Example 4:

[0090] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, wherein the vertical analysis unit includes: The analysis sub-unit is used to convert the deep demand mining chain into a vertical interpretation sequence with hierarchical numbering. Then, it calculates the comprehensive core weight of each demand node in the sequence through the node weight calculation formula, and selects nodes with weights greater than or equal to a preset threshold as core nodes. If the weights of all nodes are lower than the preset threshold, then at least one node with the highest weight is selected as the core node to ensure that each sequence has at least one core node. The depth and center sub-units are used to count the number of core nodes at each level in the vertical interpretation sequence. The original vertical depth is calculated by combining the level weights of each level, and then the normalized vertical depth is obtained by normalizing the maximum depth of the total levels. At the same time, the core node at the deepest level in the vertical interpretation sequence is selected as the vertical center.

[0091] In this embodiment, after the analysis subunit converts the deep demand mining chain into a vertical interpretation sequence, it calculates the comprehensive core weight of each node and filters out core nodes. The comprehensive core weight characterizes the overall core influence of the demand node on the bidding response, and is composed of three independent dimensions: demand intensity, hierarchical influence range, and path hub degree. The calculation formula is as follows: Non-single-chain: ,in, This represents the core weight of the node, with a value range of [0,1]. This represents the demand intensity weight for that node. Number the level to which the node belongs (level 1 d=1, level 2 d=2, and so on). This is the highest level number in the longitudinal interpretation sequence; The path hub degree is the proportion of the number of core transmission paths passing through this node to the total number of core transmission paths in the sequence, with a value range of [0,1]. The core transmission path refers to the complete path from the root node to each leaf core node.

[0092] If the vertically interpreted sequence is a single-chain structure (each level has 1 node), then the path pivot degree of all nodes is 1, and this dimension loses its distinguishing power. In this case, the relevant formula for a single chain is: .

[0093] Filtering rules: Node weight The nodes are labeled as core nodes. If the weight of all nodes in the sequence is less than 0.6, at least one node with the highest weight is selected as the core node to ensure that each vertical interpretation sequence has at least one core node, consistent with the limitations of the claims. The threshold of 0.6 is determined based on historical labeled data, where the average weight of core requirement nodes is 0.78 and the standard deviation is 0.12. After balancing the screening precision and recall, 0.6 is selected as the threshold, which can cover more than 90% of the real core nodes.

[0094] The depth and center subunits calculate the longitudinal depth and determine the longitudinal center. First, the original longitudinal depth is calculated, and then maximum depth normalization is performed to obtain the normalized longitudinal depth that can be fairly compared across sequences.

[0095] The original vertical depth includes the number of core nodes, reflecting the refinement density within the hierarchy. The calculation formula is as follows: ,in, This is the original longitudinal depth value; The total number of levels in the vertically interpreted sequence; k0 is the level weight of the k0th level, which is equal to the level number, and the deeper the level, the higher the weight; This represents the number of core nodes at level k0.

[0096] The normalized vertical depth eliminates the baseline difference in the total number of levels, and the calculation formula is as follows: ,in, Let H be the total number of nodes at level k0, and let H be the maximum possible original depth of the sequence. After normalization, the value of H ranges from (0,1). The physical meaning of the normalized vertical depth is the combined value of the depth distribution and refinement density of the core requirements. The higher the value, the deeper the core requirements and the higher the degree of refinement. For example, if a sequence has 4 levels, with 1 core node at level 1, 2 core nodes at level 2, 1 core node at level 3, and none at level 4; and the total number of nodes at each level is 1, 3, 3, and 2 respectively, then the original depth is 8, the maximum possible depth is 24, and the normalized depth is 0.33.

[0097] The vertical center is the deepest core node in a single vertical interpretation sequence, representing the finest-grained core constraint in that requirement chain. The node with the highest level number is selected from all core nodes; if there are multiple core nodes at the deepest level, all of them are used as the vertical center. Example 5:

[0098] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, wherein the roughing and mapping module includes: The block attribute parsing unit is used to perform attribute parsing on each outline area block in the rough template outline. It generates compilation difficulty tags based on the vertical depth coverage level and the horizontal center association number, dependency attribute tags based on the horizontal requirement chain scope overlap, and requirement mandatory tags based on the proportion of substantive requirements, forming the tag group for the corresponding outline area block. The auxiliary unit is used to determine the compilation thread of each compilation end based on its historical compilation performance data, extract the corresponding expertise, ability level and adaptation time interval of the compilation thread, and construct the capability auxiliary group of the compilation end. The mapping analysis unit is used to encapsulate the content domain attributes, outline position, and compilation task of each outline area block into a block triple, perform adaptation calculation on the block triple and the capability auxiliary group of each compilation end, select the compilation end with the highest adaptation as the mapping execution end of the block, and obtain the initial mapping position of the block in the corresponding compilation end queue. The sorting unit is used to arrange all outline area blocks in the original outline order to generate the first trajectory, and to mark the initial mapping position of each block at the corresponding position of the first trajectory. The topology generation unit is used to construct a dependency directed graph based on the dependency attributes of all outline region blocks, and the topology sorting algorithm is used to adjust the dependency constraints of the first trajectory. For blocks with multiple preceding dependencies, they are placed after the highest sorting position of all preceding blocks to finally generate the compilation trajectory; the sequence number calibration unit is used to calibrate the final position of each outline area block on the compilation trajectory, and assigns a sequence block number to each block in chronological order. The method acquisition unit is used to combine the block number, mapping compilation information and tag group of each outline area block to generate the collaborative compilation method of the corresponding block.

[0099] In this embodiment, the coarse mapping module is responsible for generating the tender outline, splitting blocks, performing multi-end mapping and timing scheduling, and finally outputting the collaborative compilation method for each block.

[0100] In this embodiment, the first, second, third, and fourth bidding requirements are integrated, and the various bidding requirements are filled into each chapter according to the standard structure of the business section-technical section-service section of the tender documents, forming a rough template outline with hierarchical headings.

[0101] The rough template outline is divided into several independent outline blocks according to content modules and chapter boundaries. Each block corresponds to a complete chapter or sub-chapter and serves as an independent compilation unit.

[0102] The block attribute parsing unit performs multi-dimensional tagging parsing on each outline block, generating a tag group containing tags for compilation difficulty, dependency attributes, and mandatory requirements. The criteria for determining each tag are shown in Table 1 below: Table 1 Judgment Criteria Table

[0103] The overlap of the scope of action is calculated using Jaccard similarity: Let the scope of action of region block i5 and region block j5 be respectively... and Then the degree of overlap The maximum overlap between the current block and all other blocks is used as the criterion for determining dependency tags.

[0104] The auxiliary unit determines the programming threads and constructs capability support groups based on the historical programming performance data of each programming endpoint. (A programming thread refers to an independent task channel that can be handled in parallel by the programming endpoint, corresponding to the job division and task processing sequence of the programming personnel; each programming endpoint can contain one or more programming threads.) Programming performance data is obtained through periodic standard task tests, including areas of expertise, programming quality level (A / B / C, with A being the highest), and suitable time frame. The capability support group includes three components: Staffing attributes: Area of ​​expertise + Staffing quality level; The sorting range: the range of task sequences that this compiler is suitable to handle; Compile a list of available task types to be undertaken.

[0105] The mapping analysis unit encapsulates each outline region block into a block triple: (content domain attribute, outline position number, compilation task content). The block triple is then compared with the capability auxiliary group of each compilation endpoint to calculate the mapping fit. The compilation endpoint with the highest fit is selected as the mapping execution endpoint for that block.

[0106] The formula for calculating the mapping fit is: ,in, The mapping fit is defined, with values ​​ranging from [0,1]. Domain matching score: When the content domain of the block is completely consistent with the domain of the editor, it is judged as a complete match, with a value of 1; when the content domain of the block belongs to a sub-domain of the editor's domain of expertise, or when the semantic overlap reaches more than 80%, it is judged as a partial match, with a value of 0.5; when there is no overlap between the domains, it is judged as a mismatch, with a value of 0. For difficulty adaptation, asymmetric calculation is used: if the compiler's capability level is greater than or equal to the block difficulty level, then 1.0 is used; if the compiler's capability level is less than the block difficulty level, then... ,in, The difficulty level of the block (low=1, medium=2, high=3). The competency levels are defined as follows: C=1, B=2, A=3. The maximum grade difference (value 2); For sequence matching: the sorting range of the compilation end is obtained by mapping its adapted time sequence interval. The start and end nodes of the adapted time sequence interval correspond to the position number interval of the compilation trajectory; if the position number of the block outline falls within this interval, the value is 1, otherwise the value is 0. The asymmetric design of difficulty adaptation ensures that high-capacity compilation ends can undertake low-difficulty tasks, avoiding the unreasonable situation where high-capacity ends get zero points for simple tasks, which is in line with actual business logic.

[0107] After obtaining the mapping compilation end, determine the initial mapping position of the block: if the block outline position is within the sorting range of the compilation end, the initial mapping position is equal to the outline position number; if it is less than the lower limit of the range, take the lower limit of the range; if it is greater than the upper limit of the range, take the upper limit of the range.

[0108] In this embodiment, the sorting unit arranges all outline region blocks in the original outline order to generate a first trajectory, which serves as the initial sorting benchmark.

[0109] The topology generation unit constructs a dependent directed graph based on the dependency attributes of all blocks: if block A depends on block B (A needs to be compiled after B is completed), then a directed edge B→A is added. The Kahn topology sorting algorithm is used to adjust the dependency constraints of the first trajectory to generate the final compiled trajectory. The adjustment algorithm steps are as follows: Perform cycle detection on the dependent directed graph. If a cycle exists, remove the dependency with the lowest dependency strength in the cycle until there are no cycles. Iterate through all blocks. If a block has multiple predecessor dependencies, take the maximum sorting position of all predecessor blocks and place the current block after that position. For position conflicts on the same compilation end, priority is given to the blocks according to their mapping fit from high to low. Blocks with low fit are postponed to the next idle time sequence position in the task queue of the compilation end, and their corresponding positions in the global compilation trajectory are updated synchronously. If the postponement causes the preconditions of subsequent dependent blocks to be unmet, the global positions of the dependent blocks are adjusted synchronously.

[0110] Iterate and adjust until all dependency constraints are satisfied, or until the maximum number of iterations (twice the total number of blocks) is reached. Unsatisfied constraints are marked as manually processed items.

[0111] The algorithm clearly defines the processing rules for multiple pre-dependencies, generates a trajectories based on dependency topology sorting, has clear and reproducible logic, and does not have circular dependencies.

[0112] The serial number calibration unit calibrates the final position of each outline area block on the compilation trajectory, and assigns sequential block numbers such as 1, 2, 3, etc. to each block in order of priority. Example 6:

[0113] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, wherein the scheme acquisition unit includes: The matching subunit is used to obtain the entry with the highest matching degree with the block triplet and the tag group from the pre-constructed double-group-compilation reference table through a hybrid distance matching algorithm, and output the corresponding initial compilation method. The double-group-compilation reference table is constructed based on the compilation records of historical winning projects through clustering and expert verification, and stores the mapping relationship between block feature combinations and corresponding compilation specifications. The additional sub-unit is used to attach the block number and mapping information of the block to the initial compilation method to generate the final collaborative compilation method.

[0114] In this embodiment, the matching subunit matches the initial compilation method from the bi-group compilation comparison table. The bi-group compilation comparison table is constructed based on the compilation records of historical winning bid projects. The construction steps are as follows: Step 11: Extract the triplet features, label groups, and actual compilation methods of all compilation blocks from historical winning bid projects; Step 12: Use the K-means clustering algorithm, with content domain semantic vectors and difficulty level values ​​as features, and Euclidean distance as the metric, to perform clustering according to 8 preset categories, with each category corresponding to a compilation standard level; Step 13: Calculate the review pass rate of the compilation methods corresponding to each feature combination, and select the compilation method with the highest pass rate as the standard mapping item for that category; Step 14: The clustering results are verified by at least 3 senior bid compilation experts with more than 5 years of experience. Deviation items are manually corrected and then entered into the database to form the final comparison table.

[0115] The mapping table stores the relationship between block feature combinations and compilation specifications. An example of the structure is shown in Table 2: Table 2 Structural Examples

[0116] The matching algorithm uses a hybrid distance metric: Hamming distance is used for categorical features (content domain, three types of tags, a total of 4 features), which counts the number of mismatched features and normalizes the result. This is denoted as the categorical distance component; for ordinal features (difficulty level, mapped to 1 / 2 / 3), normalized absolute difference is used. denoted as ordinal distance component Total distance The initial organization method is selected based on the organization method corresponding to the item with the smallest total distance. If the minimum distance is greater than 0.5, the default method is adopted: intermediate position organization + pre-verification + secondary depth.

[0117] The additional sub-unit adds the block number and mapping information to the initial compilation method to generate the final collaborative compilation method, clarifying the execution subject, compilation order, depth requirements and verification rules for each block. Example 7:

[0118] This invention provides a multi-terminal collaborative scheme development and review system based on artificial intelligence, wherein the development and review module includes: The compilation unit is used to schedule the corresponding compilation end to complete the compilation of the content of the outline area block in a collaborative compilation manner, and generate the first sub-file of the corresponding block. A typesetting processing unit, configured to splice all first sub-files in the outline order, unify format specifications, and generate an initial bidding document; a review unit, configured to check the response coverage rate of bidding requirements through semantic similarity matching, check the response depth through title hierarchy comparison, check format compliance through rule matching, and finally perform quantitative scoring according to scoring rules to generate a review report containing deduction items and optimization suggestions.

[0119] The preparation unit dispatches the corresponding preparation terminals to perform content preparation of the outline area blocks according to the collaborative preparation mode. After each block is prepared, an independent first sub-file is output, which includes all content and basic format settings of the corresponding chapter. During the preparation process, the system automatically checks whether the pre-dependent blocks are completed. If not, the preparation permission of the current block is blocked to avoid timing conflicts.

[0120] The typesetting processing unit splices all the first sub-files in the outline order and performs unified format processing: the main text uses Song font with size 12, titles are set in bold at different levels, headers, footers and page numbers are unified, a table of contents is automatically generated, and chart numbering is standardized. The system automatically detects format differences output by multiple terminals, performs format normalization based on the preset standard template, and finally generates a complete initial bidding document.

[0121] The review unit performs fully automatic review from three dimensions to generate a quantitative review report: Requirement response coverage check: Encode each bidding requirement into a semantic vector, calculate the cosine similarity with each paragraph vector of the bidding document, a similarity ≥0.75 is regarded as responded, and a similarity lower than 0.75 is regarded as unresponded, and the response rate is counted.

[0122] Response depth check: For the responded requirements, locate all semantically matched response paragraphs, extract the title hierarchy to which each paragraph belongs, and take the maximum title hierarchy among all matched paragraphs as the response hierarchy depth; compare the response hierarchy depth with the total hierarchy number m6 of the corresponding vertical interpretation sequence of the requirement, and if the response hierarchy depth is less than m6, it is marked as insufficient response depth.

[0123] Format compliance check: Check the format requirements such as page number, binding, volume division, and font of the bidding document through rule matching, and mark non-conforming items. The depth statistical standard for multi-paragraph and cross-chapter responses is clarified to ensure that the review results can be stably reproduced.

[0124] Finally, scoring is performed item by item according to the scoring rules of the bidding document, and the score of each item = the full score of the item × response coefficient (complete in-depth response (response hierarchy ≥ total requirement hierarchy): coefficient 1.0; response hierarchy reaching 70% or more of the total requirement hierarchy: coefficient 0.7; response hierarchy reaching 30% to 70% of the total requirement hierarchy: coefficient 0.4; response hierarchy lower than 30% of the total requirement hierarchy: coefficient 0.1; no matched response content: coefficient 0), and the total score is obtained by summation. The generated review report includes three parts: total score, details of deduction items, and optimization suggestions. Example 8:

[0125] This invention also provides a cloud platform for multi-terminal collaborative solution development and review based on artificial intelligence, deployed on a cloud server, for executing the system described in any of the above embodiments. The cloud platform provides web and mobile access points, supports online collaborative development by multiple locations and teams, and its modular functions can be selected as needed to adapt to bidding projects of different sizes. The platform also supports a manual review interface, allowing for manual verification at key nodes, balancing automation efficiency with business compliance.

[0126] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-terminal collaborative scheme development and review system based on artificial intelligence, characterized in that, include: The initial interpretation module is used to perform a preliminary screening of the bidding requirements document through a pre-trained large language model, and output horizontal interpretation information and in-depth requirement mining chain; The first and second comparison modules are used to convert the chain attribute combination of the horizontal demand chain and the horizontal center corresponding to each horizontal interpretation information into a semantic vector and then perform the first similarity matching with the historical demand chain in the existing database to obtain the calibrated first bidding requirements. Simultaneously, all horizontal centers, their corresponding core importance, and the business relationships between horizontal centers are integrated to generate the first center combination. The first center combination is then matched with the set of historical project core requirements in the existing database to obtain the calibrated second bidding requirements. The vertical analysis module is used to perform core node filtering on each vertical interpretation sequence to obtain the vertical center; Perform pairwise alignments on all longitudinal interpretation sequences to obtain longitudinal subsets. At the same time, integrate all longitudinal centers and deep statistical features to generate a second center combination. The third comparison module is used to perform a third matching between the average depth features of each vertical subset and the core node vector group and the second center combination with the historical depth requirement data in the existing database, so as to obtain the third bidding requirements for each vertical subset and the fourth bidding requirements for the second center combination. The roughing and mapping module is used to integrate the first, second, third, and fourth bidding requirements, generate a rough template outline according to the standard structure of the tender document, and combine the capability characteristics of each preparation end to perform block-preparation end mapping matching and preparation sequence with the final preparation sequence number of each block to obtain the collaborative preparation method of each outline area block; The compilation and review module is used to schedule each compilation end to complete the compilation of the corresponding outline area block according to the collaborative compilation method, generate the initial bid document, perform intelligent scoring and review on the initial bid document, and output the review report.

2. The AI-based multi-terminal collaborative scheme development and review system according to claim 1, characterized in that, One horizontal interpretation information corresponds to one deep demand mining chain, and one deep demand mining chain corresponds to one vertical interpretation sequence; each vertical interpretation sequence contains at least one vertical center, and each horizontal interpretation information contains at least one horizontal center; the vertical pair contains the vertical depth and vertical center of each of the two vertical interpretation sequences. The horizontal interpretation information set consists of several horizontal interpretation information divided by business domain. The vertical interpretation sequence set contains a vertical interpretation sequence that corresponds one-to-one with each horizontal interpretation information. Each vertical interpretation sequence is obtained by hierarchical structure transformation of the deep requirement mining chain of the corresponding horizontal interpretation information. The deep requirement mining chain is a logically progressive requirement node chain from surface requirements to deep constraints.

3. The AI-based multi-terminal collaborative scheme development and review system according to claim 1, characterized in that, The initial interpretation module includes: The macro analysis unit is used to perform macro-theme decomposition on the tender requirements document to obtain several global sub-requirements under each horizontal interpretation information, and then perform fine-grained sentence segmentation and entity extraction on each global sub-requirement to obtain several local detailed requirements points; The horizontal chain building unit is used to generate a horizontal requirement chain based on the business connection sequence, scope of action and enforcement level of each local detailed requirement point under the same global sub-requirement; At the same time, from all the global sub-requirements of the same horizontal interpretation information, the first key sub-requirement and the second key sub-requirement are extracted by comprehensive weight sorting and forced point number statistics, and together they form the horizontal center of the corresponding horizontal interpretation information. The derivative tree construction unit is used to perform semantic clustering on all global sub-requirements under each horizontal interpretation information, using the initial demand mining chain in the initial screening results as the initial skeleton, to obtain several clusters. It matches the first matching sub-requirement and its lower-level derivative sub-requirements corresponding to each cluster from the historical bidding files in the existing database, constructs the derivative subtree of the corresponding cluster according to the hierarchical relationship, and integrates all derivative subtrees to obtain the derivative tree of the corresponding horizontal interpretation information. The top-down and bottom-up analysis and summarization units are used to perform top-down filtering and bottom-up backtracking on the derivative tree to determine the core branch lines from the parent node to the child node. The core branch lines are then combined and summarized to obtain the corresponding horizontal interpretation information depth mining chain.

4. The AI-based multi-terminal collaborative scheme development and review system according to claim 3, characterized in that, The vertical analysis unit includes: The analysis sub-unit is used to convert the deep demand mining chain into a vertical interpretation sequence with hierarchical numbering. Then, it calculates the comprehensive core weight of each demand node in the sequence through the node weight calculation formula, and selects nodes with weights greater than or equal to a preset threshold as core nodes. If the weights of all nodes are lower than the preset threshold, then at least one node with the highest weight is selected as the core node to ensure that each sequence has at least one core node. The depth and center sub-units are used to count the number of core nodes at each level in the vertical interpretation sequence. The original vertical depth is calculated by combining the level weights of each level, and then the normalized vertical depth is obtained by normalizing the maximum depth of the total levels. At the same time, the core node at the deepest level in the vertical interpretation sequence is selected as the vertical center.

5. The AI-based multi-terminal collaborative scheme development and review system according to claim 3, characterized in that, The up-down analysis and summarization unit includes: The configuration subunit is used to pre-configure the demand intensity weight and logical dependency attributes for each node of the derivative tree, set the demand transmission connection verification conditions for each parent-child node link, and preset the same-level branch degradation and replacement rules and the fallback termination conditions. The fallback termination condition is: the verification is terminated when the valid branch is not selected after falling back to the root node of the derivative tree. The deep demand mining chain is constructed according to the three-level fixed hierarchical structure. The backtracking subunit is used to select the leaf node with the highest demand intensity weight in the corresponding derivative subtree of each cluster as the main anchor point and the second highest leaf node as the secondary anchor point. Starting from the anchor point, the backtracking traversal is performed on the derivative tree from bottom to top, and the traversal path is recorded. Candidate building subunits are used to perform forward and backward association extensions along the logical dependency edges of nodes during traversal. The dependency strength is calculated based on the business dependency annotation rate and the propagation verification pass rate. The dependency strength of nodes with no historical co-occurrence is directly set to 0. Extract nodes that have a strong dependency relationship with the anchor point to form a candidate branch node set; The candidate set adopts the path integrity priority rule, and the main path from the anchor point to the root node and the initial trunk path selected from top to bottom are forcibly retained in their entirety. The size limit is set for associated supplementary nodes. At the same time, implicit logical dependency edges across levels are added, implicit related nodes with excessive level differences and no intermediate connecting nodes are removed, and directed acyclic graph detection is performed simultaneously to eliminate loops and complete the link associations of the candidate branch node set. The verification of the substitute subunit is used to perform a link-by-link validity verification on each candidate branch in the candidate branch node set according to the requirement transmission connection verification conditions. Nodes introduced by implicit dependency edges need to be transmitted and verified separately. If the current candidate branch fails the verification, the substitute branch will be switched according to the same level branch downgrade and replacement rule and the verification will be repeated. If all branches at the same level fail the verification, the process will roll back to the next higher level and reselect the anchor point to perform the verification until a valid branch is found or the rollback termination condition is triggered. The route generation subunit is used to sort and integrate the selected effective branches in a top-down hierarchical order. The main path node is used as the trunk, and the associated supplementary nodes are merged into the corresponding node group according to the hierarchy to obtain the core branch route from the parent node to the child node.

6. The multi-terminal collaborative scheme development and review system based on artificial intelligence according to claim 1, characterized in that, The coarse mapping module includes: The block attribute parsing unit is used to perform attribute parsing on each outline area block in the rough template outline. It generates compilation difficulty tags based on the vertical depth coverage level and the horizontal center association number, dependency attribute tags based on the horizontal requirement chain scope overlap, and requirement mandatory tags based on the proportion of substantive requirements, forming the tag group for the corresponding outline area block. The auxiliary unit is used to determine the compilation thread of each compilation end based on its historical compilation performance data, extract the corresponding expertise, ability level and adaptation time interval of the compilation thread, and construct the capability auxiliary group of the compilation end. The mapping analysis unit is used to encapsulate the content domain attributes, outline position, and compilation task of each outline area block into a block triple, perform adaptation calculation on the block triple and the capability auxiliary group of each compilation end, select the compilation end with the highest adaptation as the mapping execution end of the block, and obtain the initial mapping position of the block in the corresponding compilation end queue. The sorting unit is used to arrange all outline area blocks in the original outline order to generate the first trajectory, and to mark the initial mapping position of each block at the corresponding position of the first trajectory. The topology generation unit is used to construct a dependency directed graph based on the dependency attributes of all outline region blocks, and the topology sorting algorithm is used to adjust the dependency constraints of the first trajectory. For blocks with multiple preceding dependencies, they are placed after the highest sorting position of all preceding blocks to finally generate the compilation trajectory; the sequence number calibration unit is used to calibrate the final position of each outline area block on the compilation trajectory, and assigns a sequence block number to each block in chronological order. The method acquisition unit is used to combine the block number, mapping compilation information and tag group of each outline area block to generate the collaborative compilation method of the corresponding block.

7. The AI-based multi-terminal collaborative scheme development and review system according to claim 6, characterized in that, The method of obtaining the unit includes: The matching subunit is used to obtain the entry with the highest matching degree with the block triplet and the tag group from the pre-constructed double-group-compilation reference table through a hybrid distance matching algorithm, and output the corresponding initial compilation method. The double-group-compilation reference table is constructed based on the compilation records of historical winning projects through clustering and expert verification, and stores the mapping relationship between block feature combinations and corresponding compilation specifications. The additional sub-unit is used to attach the block number and mapping information of the block to the initial compilation method to generate the final collaborative compilation method.

8. The multi-terminal collaborative scheme development and review system based on artificial intelligence according to claim 1, characterized in that, The compilation and review module includes: The compilation unit is used to schedule the corresponding compilation end to complete the compilation of the content of the outline area block in a collaborative compilation manner, and generate the first sub-file of the corresponding block. The layout processing unit is used to assemble all the first sub-files in the outline order and generate the initial tender document after unifying the format specifications; the review unit is used to verify the response coverage of the tender requirements through semantic similarity matching, verify the response depth through title level comparison, verify the format compliance through rule matching, and finally quantify the score according to the scoring rules and generate a review report containing deduction items and optimization suggestions.

9. A cloud platform for multi-terminal collaborative scheme development and review based on artificial intelligence, characterized in that, Used to implement the AI-based multi-terminal collaborative scheme preparation and review system as described in any one of claims 1-8.

Images