Multi-agent analysis and decision generation method and system for government-enterprise think tank service

Abstract

This invention discloses a multi-agent analysis and decision generation method and system for government and enterprise think tank services, relating to the field of multi-agent analysis technology. The method includes: generating an anchor point list based on original evidence; determining the role of the original evidence based on the anchor point list to generate evidence role cards; selecting high-buoyancy evidence role cards based on the evidence role cards; obtaining a set of missing anchor point types based on the high-buoyancy evidence role cards and a preset anchor point type set; configuring limited-use tags based on the set of missing anchor point types and generating a ballast list corresponding to the high-buoyancy evidence; constructing multiple candidate seat mapping tables based on the evidence usage stage, evidence role cards, and ballast lists in the multi-agent link; and constructing an initial particle swarm through the routing order corresponding to each piece of original evidence in the candidate seat mapping tables and the reachable seat set corresponding to each piece of original evidence. This invention can improve the traceability of evidence during the analysis process.

Description

Technical Field

[0001] This invention relates to the field of multi-agent analysis technology, and in particular to a multi-agent analysis and decision generation method and system for government and enterprise think tank services. Background Technology

[0002] Government and enterprise think tank services typically require the comprehensive analysis of multi-source heterogeneous data, including policy texts, public announcements, statistical data, industry research reports, publicly available corporate information, news, public opinion, expert minutes, project materials, and business records, within a short period of time, to generate analytical conclusions, proposed solutions, and decision-making outcomes. With the development of large-scale modeling and multi-agent technologies, more and more think tank analysis tasks are being completed using multi-agent collaborative processing. Different analysis nodes undertake different responsibilities, such as background gathering, problem guidance, rationale analysis, solution generation, and decision generation. Original evidence is transmitted, cited, summarized, and aggregated among multiple analysis nodes according to a predetermined link, ultimately forming actionable decision recommendations. In the above process, there is both factual evidence containing information on the subject, behavior, and source, and abstract evidence such as descriptions of industry trends, development directions, macro-level assessments, and empirical summaries. Different types of evidence are repeatedly cited and participate in the generation of candidate conclusions in multiple rounds of analysis.

[0003] In the process of multi-agent analysis and decision generation, existing technologies typically input different types of evidence into various analysis nodes for processing. However, they lack structured control mechanisms for the completeness of evidence boundaries and the scope of evidence's use. This leads to some abstract evidence that is only applicable to background explanation or problem guidance gradually entering the reason generation stage during multiple rounds of induction, summarization, rewriting, and citation. It then further evolves into the reasoning basis for candidate conclusions, weakening the correspondence between the reasoning content in candidate conclusions and factual evidence. This can easily result in the reasoning area being occupied by abstract evidence, thereby reducing the traceability, interpretability, and execution reliability of the decision results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that lack structured control mechanisms for the completeness of evidence boundaries and the scope of evidence use, and to propose a multi-agent analysis and decision generation method and system for government and enterprise think tank services.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: Multi-agent analysis and decision generation methods for government and enterprise think tank services include: S1. Generate an anchor point list based on the original evidence, and determine the role of the original evidence according to the anchor point list to generate evidence role cards. S2. Filter out high-buoyancy evidence role cards based on evidence role cards, and obtain a set of missing anchor point types based on high-buoyancy evidence role cards and preset anchor point type sets; configure limited-use tags according to the set of missing anchor point types, and generate ballast manifests corresponding to high-buoyancy evidence; S3. Based on the evidence usage stage, evidence role card and ballast list in the multi-agent link, construct multiple candidate seat mapping tables. Construct the initial particle swarm through the routing order corresponding to each original piece of evidence and the set of reachable seats corresponding to each original piece of evidence in the candidate seat mapping table. S4. Based on the fitness values ​​corresponding to the candidate seat mapping table, the initial particle swarm is iteratively updated to determine the target candidate seat mapping table. S5. Based on the routing order in the target candidate seat mapping table, perform multi-agent analysis on the original evidence to obtain candidate conclusions; perform seat occupancy audit on the candidate conclusions to obtain a set of executable decision candidates; S6. Based on the set of executable decision candidates, implement the decisions for the government and enterprise think tank service recipients to obtain the decision results.

[0006] Preferably, the evidence role card is generated, including: Multiple pieces of original evidence were obtained; among them, the original evidence came from government and enterprise think tank service data sources; Anchor points are extracted from each piece of original evidence to obtain an anchor point list; the anchor point list includes subject anchor points, behavior anchor points, and source anchor points. Determine the evidentiary role of the original evidence based on the list of anchor points; Based on the list of anchor points, the original evidence is assigned a role to determine the evidence role; among which, the evidence role includes occupiable evidence and high buoyancy evidence. Generate evidence role cards based on evidence roles.

[0007] Preferably, the set of missing anchor point types is obtained, including: From all the evidence role cards, select the evidence role cards that are high buoyancy evidence to obtain a set of high buoyancy evidence role cards; Obtain the High Buoyancy Evidence Character Card from the High Buoyancy Evidence Character Card Collection; Retrieve a set of preset anchor point types; the set of preset anchor point types includes main anchor point types, behavior anchor point types, and source anchor point types; The list of anchor points in the high buoyancy evidence character card is compared item by item with the preset anchor point type set to obtain the set of missing anchor point types.

[0008] Preferably, generating the ballast manifest corresponding to the high buoyancy evidence includes: Based on the set of missing anchor point types, identify the anchor point gap items corresponding to high buoyancy evidence; among them, anchor point gap items include subject gap items, behavior gap items, and source gap items; Based on the anchor point gap item corresponding to the high buoyancy evidence, configure a limited-use label for the high buoyancy evidence; the limited-use label includes a background seat label and an introductory seat label; Based on the evidence identifier, anchor point gap item, and limited purpose label of the high buoyancy evidence, a ballast manifest corresponding to the high buoyancy evidence is generated.

[0009] Preferably, constructing the initial particle swarm includes: Based on the evidence usage stage in the multi-agent link, a set of seats is constructed; wherein, the set of seats includes background seats, introductory seats, reason seats, alternative seats, and decision seats; Based on evidence role cards, ballast lists, and seat sets, multiple candidate seat mapping tables are constructed; each candidate seat mapping table includes the reachable seat set and routing order corresponding to each piece of original evidence. Define the candidate seat mapping table as a particle in the particle swarm algorithm; The initial velocity of the particle is determined by the routing order corresponding to each piece of original evidence in the candidate seat mapping table; The initial position of the particle is determined by the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table; Construct an initial particle swarm based on the initial velocity and initial position of the particles.

[0010] Preferably, determining the target candidate seat mapping table includes: Based on the evidence role cards and ballast slips, the data on seat violations was determined; Based on the seat violation data, calculate the fitness value corresponding to the candidate seat mapping table; Based on the fitness value, the optimal position of the individual particle and the global optimal position are determined; The initial particle swarm is iteratively updated based on the optimal position of each individual particle and the global optimal position, resulting in the particle swarm after the iteration. Based on the particle swarm after the iteration, a target candidate seat mapping table is determined.

[0011] Preferably, candidate conclusions are obtained, including: According to the routing order in the target candidate seat mapping table, each piece of original evidence is distributed to the corresponding multi-agent analysis node; wherein, the multi-agent analysis node includes background analysis node, introduction analysis node, reason analysis node, scheme analysis node and decision analysis node; Based on the set of reachable seats in the target candidate seat mapping table, determine the available seats for the original evidence in the multi-agent analysis node; Based on available seats, multi-agent analysis is performed on the original evidence to obtain candidate conclusions; each candidate conclusion includes a reasoning area and a background area.

[0012] Preferably, the set of executable decision candidates is obtained, including: Based on the evidence role card, determine the evidence role corresponding to the reasoning section; Based on the evidence role corresponding to the reasoning area, the candidate conclusions are audited for their place in the decision-making process to obtain executable decision candidates. By summing up all executable decision candidates, we obtain the executable decision candidate set.

[0013] Preferably, the decision result includes: Based on the set of executable decision candidates, determine the target executable decision candidates; Generate a decision action field based on the candidate decision actions from the target executable decision candidates; Based on the reasoning section of the actionable decision candidates, determine the list of justification evidence; Based on the ballast manifest, generate a high buoyancy evidence limitation column; Generate a decision instruction package based on the decision action field, the list of reasons and evidence, and the high buoyancy evidence limit field. The decision instruction package is used to implement decisions for government and enterprise think tank service recipients and obtain decision results.

[0014] To address the aforementioned problems, this invention also provides a multi-agent analysis and decision generation system for government and enterprise think tank services, the system comprising: The anchor point creation module generates an anchor point list based on the original evidence, and determines the role of the original evidence according to the anchor point list to generate evidence role cards. The gap ballast module filters out high-buoyancy evidence role cards based on evidence role cards, and obtains a set of missing anchor point types based on high-buoyancy evidence role cards and a preset set of anchor point types; it configures limited-use tags according to the set of missing anchor point types and generates the ballast list corresponding to the high-buoyancy evidence; The seat swarm building module constructs multiple candidate seat mapping tables based on the evidence usage stage, evidence role cards, and ballast list in the multi-agent link. It then constructs an initial particle swarm by using the routing order corresponding to each piece of original evidence and the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table. The particle swarm optimization module iteratively updates the initial particle swarm based on the fitness values ​​corresponding to the candidate seat mapping table to determine the target candidate seat mapping table. The seat audit module performs multi-agent analysis on the original evidence based on the routing order in the target candidate seat mapping table to obtain candidate conclusions; it then performs seat occupancy audit on the candidate conclusions to obtain a set of executable decision candidates. The decision output module executes decisions for government and enterprise think tank service recipients based on the set of executable decision candidates, and obtains the decision results.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By constructing an anchor list containing subject anchors, behavior anchors, and source anchors, and determining the role of original evidence based on anchor coverage, evidence role cards are generated, enabling automatic identification and classification management of the completeness of boundaries for different types of evidence. By distinguishing between evidence with clear boundaries and evidence with missing boundaries, evidence quality stratification can be completed before multi-agent analysis begins, providing a unified basis for subsequent evidence flow control. This reduces the probability of abstract content directly entering the decision-making basis chain, and improves the traceability of evidence and the credibility of conclusions during the analysis process.

[0016] 2. By identifying the types of missing anchor points for high-buoyancy evidence and generating ballast lists containing anchor point gaps and limited-use labels, a usage constraint mechanism for high-buoyancy evidence is established, ensuring that evidence with missing boundaries can only enter the background explanation or question guidance stage. By pre-limiting the scope of evidence use, abstract descriptions, trend judgments, and empirical content are prevented from gradually evolving into decision-making reasons during multiple rounds of induction and citation, effectively suppressing the problem of non-factual content occupying the reasoning area and improving the correspondence and explanatory power between the reasoning content and the original evidence.

[0017] 3. By constructing a candidate seat mapping table and combining it with particle swarm optimization to determine the target evidence routing scheme, and performing seat occupancy audit after the candidate conclusion is generated, the evidence cited in the reasoning area is checked item by item, and only the decision candidate results supported by occupancy evidence are retained. This achieves closed-loop control of the entire process from evidence classification and distribution to conclusion review, and can promptly detect and eliminate citations that do not comply with the seat rules, ensuring that the final decision result is based on evidence with clear sources and boundaries, thereby improving the reliability and effectiveness of the decision result. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a multi-agent analysis and decision generation method for government and enterprise think tank services provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] This embodiment provides a multi-agent analysis and decision generation method for government and enterprise think tank services. (See also...) Figure 1 Specifically, including: S1. Generate an anchor point list based on the original evidence, and determine the role of the original evidence according to the anchor point list to generate evidence role cards. In an embodiment of the present invention, generating an evidence role card includes: Multiple pieces of original evidence were obtained; among them, the original evidence came from government and enterprise think tank service data sources; Original evidence refers to data units from government and enterprise think tank service data sources that are used in subsequent analysis and processing; government and enterprise think tank service data sources refer to data sets that can provide information sources for government and enterprise think tank analysis. Specifically, the system receives data collection tasks from government and enterprise think tank service data sources, determines the data categories to be collected based on the data collection tasks, and uses policy texts, announcement texts, statistical data texts, publicly disclosed corporate texts, news texts, public opinion texts, and expert minutes texts as candidate sources. It then performs format conversion, character cleaning, paragraph segmentation, and source marking on the text, table, and record content from the candidate sources, deletes blank and duplicate data that cannot be identified, retains data fragments that can express complete semantics, treats each data fragment as a piece of original evidence, and generates fields for each piece of original evidence, including evidence identifier, source name, source type, collection address, collection batch, and original text content. Multiple pieces of original evidence form a set of original evidence.

[0021] Anchor points are extracted from each piece of original evidence to obtain an anchor point list; the anchor point list includes subject anchor points, behavior anchor points, and source anchor points. Anchor point extraction refers to the process of identifying and extracting information elements from original evidence to define the boundaries of the evidence's content; an anchor point list refers to a data list recording the anchor points extracted from the original evidence; a subject anchor point refers to information elements in the original evidence used to indicate the object, institution, enterprise, project, or region; a behavior anchor point refers to information elements in the original evidence used to indicate actions, business actions, policy actions, transaction actions, or management actions; and a source anchor point refers to information elements in the original evidence used to indicate the source name, document name, announcement number, issuing agency, data link, or collection number. Specifically, semantic component identification is performed on each piece of original evidence in the original evidence set. Words representing objects, institutions, enterprises, projects, and regions are extracted and written into subject anchors. Action words representing declaration, construction, release, procurement, punishment, investment, cooperation, governance, support, evaluation, disclosure, and transaction, along with their objects, are extracted and written into behavior anchors. Content representing document names, announcement numbers, issuing institutions, data links, collection numbers, and source names is extracted and written into source anchors. Subject anchors, behavior anchors, and source anchors are associated and stored according to the evidence identifiers of the corresponding original evidence to obtain the anchor list of the original evidence. When there are multiple subject anchors, multiple behavior anchors, or multiple source anchors in the same piece of original evidence, anchors belonging to the same semantic sentence or the same data record are grouped into the same anchor group, and all anchor groups are jointly written into the anchor list of the original evidence.

[0022] Determine the evidentiary role of the original evidence based on the list of anchor points; Specifically, after obtaining the list of anchor points corresponding to the original evidence, the relationship analysis is performed on the subject anchor points, behavior anchor points, and source anchor points in the list. The subject anchor points are used as the objects of behavior execution, the behavior anchor points are used as the action descriptions corresponding to the subject anchor points, and the source anchor points are used as the information sources corresponding to the subject anchor points and behavior anchor points. Anchor point association groups are established based on the subject anchor points, behavior anchor points, and source anchor points appearing in the same semantic fragment, the same data record, the same announcement item, or the same table row. For each anchor point association group, the number of subject anchor points, behavior anchor points, and source anchor points contained therein are counted to determine the anchor point coverage corresponding to the anchor point association group. The anchor point coverage includes subject behavior coverage, subject source coverage, behavior source coverage, subject behavior source coverage, and single anchor point coverage.

[0023] Based on the list of anchor points, the original evidence is assigned a role to determine the evidence role; among which, the evidence role includes occupiable evidence and high buoyancy evidence. Specifically, based on the anchor point coverage corresponding to the anchor point association group, a boundary integrity analysis is performed on the original evidence. If the anchor point association group simultaneously contains subject anchor points, behavior anchor points, and source anchor points, the original evidence is determined to have complete evidence boundaries. If the anchor point association group contains subject anchor points and behavior anchor points but lacks source anchor points, the original evidence is determined to have object behavior boundaries. If the anchor point association group contains subject anchor points and source anchor points but lacks behavior anchor points, the original evidence is determined to have object source boundaries. If the anchor point association group contains behavior anchor points and source anchor points but lacks subject anchor points, the original evidence is determined to have matter source boundaries. If the anchor point association group only contains subject anchor points, only contains behavior anchor points, only contains source anchor points, or does not contain an anchor point association group, the original evidence is determined to have a boundary missing state. The boundary integrity analysis results are then associated and stored with the evidence identifier corresponding to the original evidence.

[0024] Generate evidence role cards based on evidence roles.

[0025] Evidence role refers to the result of classifying original evidence according to the anchor list; placeable evidence refers to original evidence that meets the requirements for citing conclusions and reasons in the anchor list and can be included in the reasoning area; high buoyancy evidence refers to original evidence that does not meet the requirements for citing conclusions and reasons in the anchor list and is only used for background explanation or question guidance; evidence role card refers to the data carrier that records the correspondence between original evidence, anchor list and evidence role.

[0026] Specifically, the role of evidence is determined based on the boundary integrity analysis results. If the boundary integrity analysis result of the original evidence is a complete evidence boundary, object behavior boundary, object source boundary, or matter source boundary, the evidence role of the original evidence is determined to be placeable evidence. If the boundary integrity analysis result of the original evidence is a boundary missing state, the evidence role of the original evidence is determined to be high buoyancy evidence. The evidence identifier, anchor list, anchor association group, boundary integrity analysis result, and evidence role of the original evidence are combined to generate evidence role cards for the original evidence, and all evidence role cards are written into the evidence role card set.

[0027] For example, the original content of original evidence E001 is a notice issued by the Development and Reform Commission of a district in Beijing in March 2026 regarding the application for an artificial intelligence industry project. It specifies that Technology Co., Ltd. A is applying for a smart manufacturing algorithm platform construction project. After anchor point extraction, the subject anchor point is Technology Co., Ltd. A, the behavior anchor point is the application for the smart manufacturing algorithm platform construction project, and the source anchor point is the notice of the district's artificial intelligence industry project application. All three anchor points are located in the same announcement entry. Therefore, anchor point association group G001 is established, and the original evidence is determined to have complete evidence boundaries, with the evidence role being placeable evidence. The original content of original evidence E002 is "Accelerating the empowerment of the real economy by artificial intelligence and promoting high-quality industrial development." After anchor point extraction, no clear subject anchor point is obtained, and the behavior anchor point... The original evidence, E003, is for promoting high-quality industrial development. Its source anchor is a policy interpretation article. Anchor group G002 only contains behavioral and source anchors but lacks a subject anchor. Therefore, this original evidence is determined to have a source boundary, and its role is placeholder evidence. The original content of E003 states that the artificial intelligence industry has broad development prospects and related opportunities should be continuously monitored. After anchor extraction, no subject anchor or clear behavioral anchor was obtained. Its source anchor is page 5 of the expert minutes. Anchor group G003 only contains the source anchor. Therefore, this original evidence is determined to have a missing boundary, and its role is high-buoyancy evidence. An evidence role card is generated by combining the evidence identifier of E003, the anchor list, the anchor group, the boundary completeness analysis results, and the evidence role.

[0028] S2. Filter out high-buoyancy evidence role cards based on evidence role cards, and obtain a set of missing anchor point types based on high-buoyancy evidence role cards and preset anchor point type sets; configure limited-use tags according to the set of missing anchor point types, and generate ballast manifests corresponding to high-buoyancy evidence; In embodiments of the present invention, a set of missing anchor point types is obtained, including: From all the evidence role cards, select the evidence role cards that are high buoyancy evidence to obtain a set of high buoyancy evidence role cards; The content of the evidence role field is matched. When the content of the evidence role field is high buoyancy evidence, the evidence identifier, anchor list, anchor association group, boundary integrity analysis result, and evidence role of the evidence role card are written into the high buoyancy evidence role card cache area. When the content of the evidence role field is placeable evidence, the evidence role card is not written into the high buoyancy evidence role card cache area. After completing the above matching process for all evidence role cards in the evidence role card set, all evidence role cards in the high buoyancy evidence role card cache area are deduplicated according to the evidence identifier and arranged according to the generation order of the evidence identifier to obtain the high buoyancy evidence role card set.

[0029] Obtain the High Buoyancy Evidence Character Card from the High Buoyancy Evidence Character Card Collection; The high buoyancy evidence role card set refers to the data set formed by the evidence role cards selected from the evidence role card set that are high buoyancy evidence. Each data unit in the high buoyancy evidence role card set corresponds to an original piece of evidence with missing boundaries, and saves the evidence identifier, anchor list, anchor association group, boundary integrity analysis results and evidence role corresponding to the original evidence, so as to serve as the data basis for subsequent identification of missing anchors. Locate a high-buoyancy evidence role card and use it as the current processing object. Extract the evidence identifier, anchor list, anchor association group, boundary integrity analysis results, and evidence role from the current processing object. Write the extracted data into the current processing record. The current processing record is used to carry out subsequent missing anchor type comparison, anchor gap item generation, limited purpose label configuration, and ballast list generation. After the current processing object completes its subsequent processing, locate the next high-buoyancy evidence role card according to the order of evidence identifiers, until each high-buoyancy evidence role card in the high-buoyancy evidence role card set has a corresponding current processing record.

[0030] Retrieve a set of preset anchor point types; the set of preset anchor point types includes main anchor point types, behavior anchor point types, and source anchor point types; The pre-defined anchor point type set refers to the set of anchor point types used to describe the requirements for forming a complete evidence boundary. It includes subject anchor point types, behavior anchor point types, and source anchor point types. The subject anchor point type is used to describe the data type of the initiator, executor, associate, or object of the behavior in the original evidence. The behavior anchor point type is used to describe the data type of the event action, business action, policy action, management action, or transaction action in the original evidence. The source anchor point type is used to describe the data type of the source location, source subject, source document, or source record of the original evidence. The subject anchor point type, behavior anchor point type, and source anchor point type together constitute the type set required for a complete evidence boundary. Extract the anchor point list from the current processing record and establish a preset anchor point type set. The preset anchor point type set consists of subject anchor point type, behavior anchor point type, and source anchor point type. The subject anchor point type is used for the subject anchor point in the corresponding anchor point list, the behavior anchor point type is used for the behavior anchor point in the corresponding anchor point list, and the source anchor point type is used for the source anchor point in the corresponding anchor point list. Write the preset anchor point type set into the current processing record so that the current processing record simultaneously contains evidence identifier, anchor point list, anchor point association group, boundary integrity analysis result, evidence role, and preset anchor point type set.

[0031] The list of anchor points in the high buoyancy evidence character card is compared item by item with the preset anchor point type set to obtain the set of missing anchor point types.

[0032] The missing anchor type set refers to the data set formed by anchor types that do not appear in the anchor list, determined based on the item-by-item comparison results. When there is no main anchor in the anchor list, the main anchor type is written into the missing anchor type set. When there is no behavior anchor in the anchor list, the behavior anchor type is written into the missing anchor type set. When there is no source anchor in the anchor list, the source anchor type is written into the missing anchor type set. The missing anchor type set is used to characterize the boundary components that are missing in the current high buoyancy evidence and serves as the data basis for generating anchor gap items and configuring limited-use labels.

[0033] Using the list of anchor points in the current processing record as the comparison object and the preset set of anchor point types in the current processing record as the comparison benchmark, the main anchor point type, behavior anchor point type, and source anchor point type are compared item by item. When there is no main anchor point in the anchor point list, the main anchor point type is written into the missing anchor point type set. When there is no behavior anchor point in the anchor point list, the behavior anchor point type is written into the missing anchor point type set. When there is no source anchor point in the anchor point list, the source anchor point type is written into the missing anchor point type set. The missing anchor point type set is associated with the evidence identifier in the current processing record and stored to obtain the missing anchor point type set corresponding to the high buoyancy evidence role card.

[0034] For example, the original evidence corresponding to the high buoyancy evidence role card C002 is that low-altitude logistics scenarios have significant application potential, and future attention can be paid to drone delivery, warehouse inspection, and park transportation. The anchor point list of this original evidence has extracted the source anchor point as page 18 of the industry observation report R2026, but no subject anchor points corresponding to specific company names, project names, or regional names have been extracted, nor have any behavioral anchor points corresponding to specific matters such as construction, procurement, delivery, operation, or contract signing been extracted. After establishing a preset anchor point type set, we obtain the subject anchor point type, behavioral anchor point type, and source anchor point type. Comparing this anchor point list with the preset anchor point type set item by item, we find that the subject anchor point type has no corresponding data in the anchor point list, the behavioral anchor point type has no corresponding data in the anchor point list, and the source anchor point type has corresponding data in the anchor point list. Therefore, the missing anchor point type set is the subject anchor point type and the behavioral anchor point type.

[0035] In an embodiment of the present invention, generating the ballast manifest corresponding to the high buoyancy evidence includes: Based on the set of missing anchor point types, identify the anchor point gap items corresponding to high buoyancy evidence; among them, anchor point gap items include subject gap items, behavior gap items, and source gap items; Anchor point gaps refer to boundary missing information records determined by the set of missing anchor point types. They are used to characterize the components missing from the boundary of high buoyancy evidence relative to complete evidence. Anchor point gaps are associated with the evidence identifier of the corresponding high buoyancy evidence and serve as the data basis for subsequent use, configuration restrictions, and seat allocation control. Subject gaps refer to missing records formed when high buoyancy evidence lacks subject anchors. Subject gaps indicate that the current high buoyancy evidence cannot determine the initiator, executor, participant, or associated object of the action, and therefore cannot establish a clear object boundary. Action gaps refer to missing records formed when high buoyancy evidence lacks action anchors. Action gaps indicate that the current high buoyancy evidence cannot determine the specific matters, business content, execution content, or processing content implemented by the corresponding object, and therefore cannot establish a clear action boundary. Source gaps refer to missing records formed when high buoyancy evidence lacks source anchors. Source gaps indicate that the current high buoyancy evidence cannot determine the information source location, source carrier, source record, or source subject, and therefore cannot establish a clear source boundary. Based on the set of missing anchor point types corresponding to high buoyancy evidence, name conversion and record generation are performed on each missing anchor point type in the set. If the set of missing anchor point types includes a subject anchor point type, a subject gap item is generated and written into the anchor gap item field of the high buoyancy evidence. If the set of missing anchor point types includes a behavior anchor point type, a behavior gap item is generated and written into the anchor gap item field of the high buoyancy evidence. If the set of missing anchor point types includes a source anchor point type, a source gap item is generated and written into the anchor gap item field of the high buoyancy evidence. The generated subject gap item, behavior gap item, and source gap item are associated with the evidence identifier of the high buoyancy evidence to obtain the anchor gap item record corresponding to the high buoyancy evidence.

[0036] Based on the anchor point gap item corresponding to the high buoyancy evidence, configure a limited-use label for the high buoyancy evidence; the limited-use label includes a background seat label and an introductory seat label; The restricted use label refers to the use restriction record generated based on the anchor gap item. It is used to characterize the analysis positions in which high buoyancy evidence is allowed to participate and the analysis positions in which it is prohibited from participating, and serves as a constraint basis in the subsequent seat mapping construction process. The background seat label refers to the use identifier that allows high buoyancy evidence to enter the background description area. When high buoyancy evidence carries a background seat label, the corresponding content can be used to describe industry phenomena, development directions, environmental information, or reference information, but is not used to form a decision-making rationale. The introductory seat label refers to the use identifier that allows high buoyancy evidence to enter the issue guidance area. When high buoyancy evidence carries an introductory seat label, the corresponding content can be used to generate matters to be supplemented, matters to be verified, or matters to be analyzed, but is not used to form a decision-making basis. Based on the anchor gap items corresponding to the high buoyancy evidence, a limited-use label is configured for the high buoyancy evidence. When there are subject gap items, behavior gap items, or source gap items in the anchor gap items, the background position label is written into the limited-use label field of the high buoyancy evidence, and the introductory position label is also written into the limited-use label field of the high buoyancy evidence. The background position label is used to indicate that the high buoyancy evidence can only enter the background description position, and the introductory position label is used to indicate that the high buoyancy evidence can only enter the question guidance position. The limited-use label field is associated with the evidence identifier and anchor gap item record of the high buoyancy evidence to obtain the usage limitation record corresponding to the high buoyancy evidence.

[0037] Based on the evidence identifier, anchor point gap item, and limited purpose label of the high buoyancy evidence, a ballast manifest corresponding to the high buoyancy evidence is generated.

[0038] Ballast slip refers to the constraint record generated for a single piece of high buoyancy evidence. It includes the evidence identifier of the corresponding high buoyancy evidence, anchor point gap item, limited purpose label, missing boundary type and usage restriction information. Ballast slip is used to record the missing boundary components of high buoyancy evidence and to limit the range of accessible positions of high buoyancy evidence in the subsequent multi-agent analysis link, thereby ensuring that high buoyancy evidence only participates in analysis and processing at the position corresponding to the limited purpose label.

[0039] Based on the evidence identifier, anchor gap item, and limited purpose label of the high buoyancy evidence, a ballast manifest corresponding to the high buoyancy evidence is generated. The evidence identifier is written into the evidence identifier field of the ballast manifest. The generated gap items from the subject gap item, behavior gap item, and source gap item are written into the gap field of the ballast manifest. The background seat label and the introductory seat label are written into the purpose field of the ballast manifest. The anchor list in the evidence role card is written into the original boundary field of the ballast manifest. The boundary integrity analysis results in the evidence role card are written into the boundary status field of the ballast manifest. The ballast manifest is associated with the corresponding high buoyancy evidence role card and stored. The ballast manifests corresponding to each piece of high buoyancy evidence are summarized to obtain a ballast manifest set.

[0040] For example, the evidence identifier for high buoyancy evidence is E008, corresponding to the original content that the park's warehousing and distribution scenario has significant potential for automation transformation, focusing on unmanned vehicle transfer, intelligent sorting, and inventory counting. The missing anchor point type set for this high buoyancy evidence includes subject anchor point type and behavior anchor point type. Based on this missing anchor point type set, subject gap items and behavior gap items are generated, but no source gap item is generated. The source anchor point for this high buoyancy evidence is page 6 of industry research record R18. Therefore, the background seat label and the introductory seat label are written into the limited purpose label field to generate ballast sheet B008. Ballast sheet B008 includes evidence identifier E008, anchor point gap items are subject gap items and behavior gap items, limited purpose labels are background seat labels and introductory seat labels, the original boundary field is the source anchor point page 6 of industry research record R18, and the boundary status field is the boundary missing status.

[0041] S3. Based on the evidence usage stage, evidence role card and ballast list in the multi-agent link, construct multiple candidate seat mapping tables. Construct the initial particle swarm through the routing order corresponding to each original piece of evidence and the set of reachable seats corresponding to each original piece of evidence in the candidate seat mapping table. In an embodiment of the present invention, constructing an initial particle swarm includes: Based on the evidence usage stage in the multi-agent link, a set of seats is constructed; wherein, the set of seats includes background seats, introductory seats, reason seats, alternative seats, and decision seats; A multi-agent link refers to a processing path formed by multiple analysis nodes according to evidence transmission relationships. Each analysis node is used to perform one of the following processes: background preparation, problem guidance, reason generation, solution formation, or decision generation. The evidence usage stage refers to the processing stage that original evidence can enter in the multi-agent link, including the background explanation stage, problem guidance stage, reason generation stage, solution formation stage, and decision generation stage. The seat set refers to the set of evidence entry positions obtained according to the evidence usage stage. The background seat refers to the position used to receive evidence in the background explanation stage; the introductory seat refers to the position used to receive evidence in the problem guidance stage; the reason seat refers to the position used to receive evidence that supports candidate conclusions; the solution seat refers to the position used to receive evidence participating in the solution formation process; and the decision seat refers to the position used to receive evidence participating in the decision generation process. Based on the usage stages of evidence processing in the multi-agent link—background preparation, problem guidance, reason generation, solution formation, and decision generation—seat names are established corresponding to each usage stage. The evidence entry position corresponding to background preparation is determined as the background seat, the evidence entry position corresponding to problem guidance is determined as the introductory seat, the evidence entry position corresponding to reason generation is determined as the reason seat, the evidence entry position corresponding to solution formation is determined as the solution seat, and the evidence entry position corresponding to decision generation is determined as the decision seat. The background seat, introductory seat, reason seat, solution seat, and decision seat are written into the seat set, and each seat in the seat set is written with a seat name, corresponding usage stage, acceptable evidence role, and output content type, so that the seat set can serve as the seat source for the subsequent candidate seat mapping table.

[0042] Based on evidence role cards, ballast lists, and seat sets, multiple candidate seat mapping tables are constructed; each candidate seat mapping table includes the reachable seat set and routing order corresponding to each piece of original evidence. The candidate seat mapping table is a data table used to record the accessible relationship between original evidence and seat set; the reachable seat set is the set of allowed seats for a single original evidence record in the candidate seat mapping table; the routing order refers to the flow order of a single original evidence in the reachable seat set; Based on the evidence identifier, evidence role, and anchor point list in the evidence role card, and combined with the anchor point gap item and limited purpose label in the ballast sheet, a set of reachable seats and a routing order are generated for each piece of original evidence. When the evidence role is occupiable evidence, the reason seat, solution seat, and decision seat are written into the set of reachable seats corresponding to the original evidence, and the routing order corresponding to the original evidence is generated in the order of the reason seat, solution seat, and decision seat. When the evidence role is high buoyancy evidence, the background seat and introduction seat are written into the set of reachable seats corresponding to the original evidence according to the limited purpose label in the corresponding ballast sheet, and the routing order corresponding to the original evidence is generated in the order of the background seat and introduction seat. The evidence identifier, set of reachable seats, and routing order corresponding to each piece of original evidence are written into the candidate seat mapping table. By changing the routing order arrangement and reachable seat combination method corresponding to different pieces of original evidence, multiple candidate seat mapping tables are formed.

[0043] Define the candidate seat mapping table as a particle in the particle swarm algorithm; Each candidate seat mapping table is treated as a particle in the particle swarm optimization algorithm. A particle data record is established for each particle, which includes a particle identifier, a candidate seat mapping table identifier, a list of original evidence identifiers, a set of reachable seats corresponding to each piece of original evidence, a routing order corresponding to each piece of original evidence, and a ballast reference relationship corresponding to the candidate seat mapping table. The particle position is jointly represented by the set of reachable seats and the routing order of all original evidence in the candidate seat mapping table, so that the same particle can completely represent the arrangement result of a piece of evidence entering different seats and flowing between different seats.

[0044] The initial velocity of the particle is determined by the routing order corresponding to each piece of original evidence in the candidate seat mapping table; For each particle's candidate seat mapping table, the routing order corresponding to each piece of original evidence is extracted. The sequential movement relationship between two adjacent seats in the routing order is taken as the routing movement item of the original evidence. The routing movement items of all original evidence in the candidate seat mapping table are summarized into the initial velocity data of the particle. The reachable seat set and routing order corresponding to each piece of original evidence are written into the initial position data of the particle. The particle identifier, initial velocity data and initial position data are stored together to form a particle record for participating in the particle swarm algorithm iteration.

[0045] The initial position of the particle is determined by the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table; For each particle's candidate seat mapping table, the reachable seat set corresponding to each original piece of evidence in the candidate seat mapping table is extracted. The evidence identifier of each original piece of evidence is used as the position index, and the reachable seat set and routing order corresponding to the original piece of evidence are used as the position content. All position indices and their corresponding position content are combined into the particle's initial position data. In the same particle, the seat name, number of seats, seat arrangement relationship and routing order of the reachable seat set corresponding to each original piece of evidence are retained, so that the particle's initial position can represent the complete seat allocation state corresponding to the candidate seat mapping table.

[0046] Construct an initial particle swarm based on the initial velocity and initial position of the particles.

[0047] In particle swarm optimization (PSO), a particle refers to the solution representation of a candidate seat mapping table in PSO; the initial position refers to the set of reachable seats and the combination of routing order corresponding to the particle before iterative update; the initial particle swarm refers to the set of data to be updated consisting of multiple particles with initial velocities and initial positions.

[0048] Write the particle identifier, candidate seat mapping table identifier, initial velocity data, and initial position data of each particle into a particle record. Aggregate multiple particle records into a particle record set according to the generation order of the candidate seat mapping table. Set an evaluation state for each particle record in the particle record set and associate the evaluation state with the candidate seat mapping table identifier of the particle to obtain the initial particle swarm.

[0049] For example, the evidence usage stage includes background seats, introductory seats, reasoning seats, alternative seats, and decision seats. Evidence role card A001 corresponds to original evidence E001, and its evidence role is occupiable evidence. The reachable seat set includes reasoning seats, alternative seats, and decision seats, with the routing order being reasoning seats to alternative seats to decision seats. Evidence role card A002 corresponds to original evidence E002, and its evidence role is high buoyancy evidence. The limited purpose label in ballast manifest B002 is a background seat label and an introductory seat label. The reachable seat set includes background seats and introductory seats, with the routing order being background seats to introductory seats. Based on the above data, a candidate seat mapping table M00 is constructed. 1. M001 records the reachable seat set and routing order corresponding to E001 and E002. M001 is defined as particle P001. The routing order of E001 and E002 are used together as the initial velocity data of P001. The reachable seat set of E001, the routing order of E001, the reachable seat set of E002, and the routing order of E002 are used together as the initial position data of P001. Particles P002 and P003 are generated in the same way for the candidate seat mapping tables M002 and M003. P001, P002, and P003 are summarized to obtain the initial particle swarm.

[0050] S4. Based on the fitness values ​​corresponding to the candidate seat mapping table, the initial particle swarm is iteratively updated to determine the target candidate seat mapping table. In an embodiment of the present invention, determining the target candidate seat mapping table includes: Based on the evidence role cards and ballast slips, the data on seat violations was determined; For each candidate seat mapping table, based on the evidence role in the evidence role card and the limited purpose label in the ballast list, the reachable seat set corresponding to each piece of original evidence in the candidate seat mapping table is checked. When the evidence role is high buoyancy evidence and the reachable seat set includes a reason seat, a violation record of high buoyancy evidence entering a reason seat is generated. When the evidence role is high buoyancy evidence and the reachable seat set includes seats other than background seats and introductory seats, a violation record of high buoyancy evidence exceeding the limited purpose label is generated. When the evidence role is placeable evidence and the reachable seat set does not include a reason seat, a violation record of placeable evidence not entering a reason seat is generated. The candidate seat mapping table identifier, original evidence identifier, evidence role, reachable seat set, limited purpose label, and violation type are written into the seat violation data.

[0051] Based on the seat violation data, calculate the fitness value corresponding to the candidate seat mapping table; Seat violation data refers to abnormal records obtained by checking the candidate seat mapping table based on the evidence role cards and ballast lists. It includes data on high buoyancy evidence entering the reason seat, data on high buoyancy evidence entering the seat corresponding to the non-restricted purpose label, and data on occupiable evidence not entering the reason seat. The fitness value refers to the evaluation value calculated based on the seat violation data. It is used to indicate the degree to which the candidate seat mapping table meets the seat allocation requirements. The fewer the seat violation data, the better the corresponding fitness value. Count the violation records of seats with high buoyancy evidence, high buoyancy evidence usage, and missing seats with occupiable evidence in the seat violation data to obtain the total number of violation records. Use the sum of the original total number of evidence in the candidate seat mapping table and 1 as the denominator and 1 as the numerator to obtain the base value. Multiply the base value by the total number of violation records to obtain the violation deduction value. Calculate the fitness value: fitness value = 1 - violation deduction value. It should be noted that the calculation process of fitness value conforms to the relationship between stable state and perturbation state in physics. When a physical object is in a stable state, its internal components are distributed in an orderly manner according to predetermined constraints, and the corresponding state has high stability. When external perturbations or internal anomalies increase, the components deviate from the original constraint relationship, and the overall state gradually moves away from the stable state. The degree of stability decreases as the number of perturbations increases. In this embodiment, the original evidence in the candidate seat mapping table is equivalent to the components in the stable state, and the seat violation data is equivalent to the perturbation factors that disrupt the stable state. Situations such as high buoyancy evidence entering a non-allowed seat or missing reasons for occupyable evidence in a seat will lead to a change in the constraint relationship between evidence and seat. Since the system is disrupted, the total number of violation records is used as the measure of perturbation intensity. The perturbation intensity is normalized using the total number of original evidence records, so that candidate seat mapping tables of different sizes can be compared under the same evaluation scale. When the total number of violation records increases, the perturbation intensity increases, the degree to which the candidate seat mapping table deviates from the constraint state increases, and the corresponding fitness value decreases. When the total number of violation records decreases, the perturbation intensity decreases, the degree to which the candidate seat mapping table approaches the constraint state increases, and the corresponding fitness value increases. Therefore, the result calculated by combining the total number of violation records and the total number of original evidence records can reflect the degree to which the candidate seat mapping table satisfies the evidence seat constraint relationship, and thus serve as the fitness value corresponding to the candidate seat mapping table.

[0052] Based on the fitness value, the optimal position of the individual particle and the global optimal position are determined; The optimal position of an individual particle refers to the position in the candidate seat mapping table corresponding to the optimal fitness value of a single particle in each iteration; the optimal position of the global particle swarm refers to the position in the candidate seat mapping table corresponding to the optimal fitness value of all particles in each iteration. Based on the fitness value of each particle, the historical position records of each particle are compared. The particle position corresponding to the highest fitness value obtained by the same particle before the current iteration is determined as the individual optimal position of that particle. The individual optimal position of the particle is associated and stored with the corresponding particle identifier, candidate seat mapping table identifier, fitness value and iteration number. The individual optimal positions of each particle in the initial particle swarm are compared, and the position with the highest fitness value among all the individual optimal positions of particles is determined as the global optimal position.

[0053] The initial particle swarm is iteratively updated based on the optimal position of each individual particle and the global optimal position, resulting in the particle swarm after the iteration. Iterative update refers to the process of updating the velocity and position data of a particle based on its current position, current velocity, individual optimal position, and global optimal position; the particle swarm after iteration refers to the set of particles obtained after the iterative update is completed. The process involves comparing the reachable seat set and routing order at the particle's current position with the reachable seat set and routing order at the particle's optimal position to obtain an individual correction term. It also involves comparing the reachable seat set and routing order at the particle's current position with the reachable seat set and routing order at the globally optimal position to obtain a global correction term. The particle's velocity data is updated based on the current velocity, the individual correction term, and the global correction term. The reachable seat set and routing order at the particle's current position are then adjusted based on the updated velocity data to obtain the updated particle position. The seat violation data and fitness value are recalculated for the updated particle position. The updated particle position, updated velocity data, recalculated fitness value, and corresponding iteration number are written into the particle record. This process is repeated until the particle position no longer generates a new reachable seat set or a new routing order, resulting in the particle swarm after the iteration ends.

[0054] Based on the particle swarm after the iteration, a target candidate seat mapping table is determined.

[0055] The target candidate seat mapping table refers to the candidate seat mapping table determined by comparing fitness values ​​among multiple candidate seat mapping tables. It is used as the data basis for subsequent evidence routing and seat control. The particle with the highest final fitness value is identified as the target particle. The candidate seat mapping table identifier corresponding to the target particle is extracted. The candidate seat mapping table corresponding to the target particle is located according to the candidate seat mapping table identifier. This candidate seat mapping table is identified as the target candidate seat mapping table. The reachable seat set and routing order corresponding to each piece of original evidence in the target candidate seat mapping table are used as the evidence distribution basis for subsequent multi-agent analysis.

[0056] S5. Based on the routing order in the target candidate seat mapping table, perform multi-agent analysis on the original evidence to obtain candidate conclusions; perform seat occupancy audit on the candidate conclusions to obtain a set of executable decision candidates; In embodiments of the present invention, candidate conclusions are obtained, including: According to the routing order in the target candidate seat mapping table, each piece of original evidence is distributed to the corresponding multi-agent analysis node; wherein, the multi-agent analysis node includes background analysis node, introduction analysis node, reason analysis node, scheme analysis node and decision analysis node; Multi-agent analysis nodes refer to analysis and processing units formed according to different analysis responsibilities, which are used to receive raw evidence and output corresponding analysis results; background analysis nodes refer to analysis and processing units used to organize background information, environmental information, reference information, and supplementary information; introductory analysis nodes refer to analysis and processing units used to generate content guiding the analysis of matters to be analyzed, matters to be concerned about, and questions; justification analysis nodes refer to analysis and processing units used to generate justification content that can support candidate conclusions; solution analysis nodes refer to analysis and processing units used to generate processing paths, execution paths, or implementation paths; decision analysis nodes refer to analysis and processing units used to generate candidate decision actions based on justification content and solution content. Based on the routing order in the target candidate seat mapping table, extract the evidence identifier, reachable seat set, and routing order corresponding to each piece of original evidence. Distribute the background seats to the background analysis node, the introductory seats to the introductory analysis node, the reason seats to the reason analysis node, the solution seats to the solution analysis node, and the decision seats to the decision analysis node. Send each piece of original evidence to the corresponding multi-agent analysis node in sequence according to the routing order, and carry the evidence role card, anchor list, ballast list, and target candidate seat mapping table record corresponding to the original evidence when sending.

[0057] Based on the set of reachable seats in the target candidate seat mapping table, determine the available seats for the original evidence in the multi-agent analysis node; Available seats refer to the range of seats that the original evidence is allowed to enter according to the target candidate seat mapping table; multi-agent analysis refers to the process in which the original evidence flows sequentially among multiple analysis nodes and completes analysis and processing according to the routing order corresponding to the target candidate seat mapping table. The set of reachable seats corresponding to the original evidence is matched with the seats corresponding to the multi-agent analysis node. When the set of reachable seats of the original evidence contains the seat corresponding to the multi-agent analysis node, the seat is determined as the available seat of the original evidence in the multi-agent analysis node. When the set of reachable seats of the original evidence does not contain the seat corresponding to the multi-agent analysis node, the input of the original evidence into the multi-agent analysis node is stopped, and the evidence identifier, target candidate seat mapping table identifier and seat name not entered are recorded.

[0058] Based on available seats, multi-agent analysis is performed on the original evidence to obtain candidate conclusions; each candidate conclusion includes a reasoning area and a background area.

[0059] Candidate conclusions refer to the intermediate analysis results generated after the multi-agent analysis is completed; the reasoning area refers to the content area in the candidate conclusions used to store the basis for the formation of the conclusions, and the content in this area comes from the original evidence that enters the reasoning position; the background area refers to the content area in the candidate conclusions used to store the background description, environmental description and guiding description, and the content in this area comes from the original evidence that enters the background position or the introductory position.

[0060] In the background analysis node, background content is generated based on the original evidence entering the background position. In the introduction analysis node, the content of the matter to be analyzed is generated based on the original evidence entering the introduction position. In the reason analysis node, the reason content is generated based on the original evidence entering the reason position. In the solution analysis node, candidate solutions are generated based on the original evidence entering the solution position. In the decision analysis node, candidate conclusions are generated based on the reason content and candidate solutions. The reason content output by the reason analysis node is written into the reason area of ​​the candidate conclusions. The background content output by the background analysis node and the content of the matter to be analyzed output by the introduction analysis node are written into the background area of ​​the candidate conclusions, thus obtaining candidate conclusions including the reason area and the background area.

[0061] In embodiments of the present invention, an executable decision candidate set is obtained, including: Based on the evidence role card, determine the evidence role corresponding to the reasoning section; The evidence role corresponding to the reasoning section refers to the evidence role classification result in the evidence role card corresponding to the original evidence that forms the content of the reasoning section; the evidence roles corresponding to the reasoning section include placeable evidence and high buoyancy evidence; placement refers to the state in which the original evidence enters the corresponding position according to the target candidate position mapping table and participates in the generation of candidate conclusions; audit result refers to the inspection result obtained after the placement audit is completed, which includes pass result and fail result; pass result means that all the content in the reasoning section is the inspection result provided by placeable evidence; fail result means that the content in the reasoning section contains inspection results from high buoyancy evidence; Based on the reasoning section in the candidate conclusions, a reasoning source table is established, dividing the reasoning section into several reasoning content units. Each reasoning content unit retains the corresponding text content, reasoning number, citation position, and original evidence identifier. For each original evidence identifier, it is matched in the evidence role card set to obtain the evidence role card that is the same as the original evidence identifier. From the evidence role card, the evidence role, anchor list, boundary integrity analysis result, and evidence identifier are extracted. The extracted evidence role is written into the evidence role field of the corresponding reasoning content unit in the reasoning source table, the anchor list is written into the anchor field of the corresponding reasoning content unit, and the boundary integrity analysis result is written into the boundary field of the corresponding reasoning content unit. The reasoning number, reasoning content unit, original evidence identifier, evidence role, anchor list, and boundary integrity analysis result are combined to form the reasoning area evidence role record.

[0062] Based on the evidence role corresponding to the reasoning area, the candidate conclusions are audited for their place in the decision-making process to obtain executable decision candidates. Seat-occupancy audit refers to the process of examining candidate conclusions based on evidence roles and seat usage rules. It is used to determine whether the reasons for forming a candidate conclusion are provided by the original evidence that is allowed to enter the reason seat. Executable decision candidate refers to a candidate conclusion that has passed the seat-occupancy audit, where all content in its reasoning area comes from occupiable evidence and there is no situation where high-buoyancy evidence occupies the reason seat. Executable decision candidate set refers to a data set formed by summarizing multiple executable decision candidates. Each evidence role record in the reasoning area of ​​the candidate conclusion is examined. When the evidence role in a reasoning area evidence role record is placeable evidence, the record is written to the pass record table. When the evidence role in a reasoning area evidence role record is high buoyancy evidence, the record is written to the placeholder anomaly record table. Based on the original evidence identifier in the reasoning area evidence role record, the corresponding ballast document is located in the ballast document set. The anchor gap item and restricted use label are extracted from the ballast document. The original evidence identifier, reason number, anchor gap item, restricted use label, and anomaly seat name of the high buoyancy evidence are written to the placeholder anomaly record table. When the placeholder anomaly record table corresponding to the candidate conclusion is empty, the candidate conclusion is marked as an executable decision candidate. When the placeholder anomaly record table corresponding to the candidate conclusion is not empty, the candidate conclusion is marked as a non-executable candidate. The placeholder anomaly record table is written to the audit description field of the candidate conclusion.

[0063] By summing up all executable decision candidates, we obtain the executable decision candidate set.

[0064] The process of generating executable decision candidates refers to the process of retaining candidate conclusions with audit results of "pass" as executable decision candidates, and removing candidate conclusions with audit results of "fail" from the reasoning area reference link.

[0065] Extract the candidate conclusions marked as executable decision candidates, generate a candidate number for each executable decision candidate, and write the candidate number, candidate decision action, reasoning area, background area, reasoning source table, pass record table, original evidence identifier list, target candidate seat mapping table identifier, and analysis node identifier that generated the candidate conclusion into the executable decision candidate record. Write all executable decision candidate records into the executable decision candidate set in the order in which the candidate conclusions are generated, while retaining the audit description field of non-executable candidates as non-executable records, so that the executable decision candidate set only contains candidate conclusions whose reasoning area is supported by placeholder evidence.

[0066] S6. Based on the set of executable decision candidates, implement the decisions for the government and enterprise think tank service recipients to obtain the decision results.

[0067] In embodiments of the present invention, obtaining the decision result includes: Based on the set of executable decision candidates, determine the target executable decision candidates; Traverse the executable decision candidate records in the executable decision candidate set, extract the candidate number, candidate decision action, reason area, background area, reason source table, pass record table, original evidence identifier list, and target candidate seat mapping table identifier from each executable decision candidate record, and perform an association check between the candidate decision action and the corresponding reason area. If a complete reason source table exists in the reason area and the pass record table is not empty, retain the executable decision candidate record as a candidate. Arrange the candidates in order of candidate number, and determine the candidate at the top of the list as the target executable decision candidate.

[0068] Generate a decision action field based on the candidate decision actions from the target executable decision candidates; The target executable decision candidate refers to the candidate record selected from the set of executable decision candidates to generate the decision instruction package. It includes the candidate decision action, the reason area, the background area, the reason source table, and the audit data. The candidate decision action refers to the executable processing content recorded in the target executable decision candidate, which is used to indicate the specific processing method to be taken for the government and enterprise think tank service objects. The decision action field refers to the data field formed after the candidate decision action is written into the decision instruction package, which is used to record the final processing content. Extract candidate decision actions from the target candidate records, write the processing object in the candidate decision action into the object subfield of the decision action field, write the processing type in the candidate decision action into the type subfield of the decision action field, write the execution content in the candidate decision action into the content subfield of the decision action field, write the candidate number corresponding to the target executable decision candidate into the source subfield of the decision action field, and combine the object subfield, type subfield, content subfield and source subfield to form the decision action field.

[0069] Based on the reasoning section of the actionable decision candidates, determine the list of justification evidence; Extract the reason area from the target executable decision candidate, and split the reason area according to reason number, reason content, and original evidence identifier to obtain multiple reason content units. For each reason content unit, locate the corresponding evidence role card in the evidence role card set according to the original evidence identifier, and extract the anchor point list, boundary integrity analysis results, and evidence role from the evidence role card. Write the reason number, reason content, original evidence identifier, anchor point list, boundary integrity analysis results, and evidence role into the reason evidence record. Summarize all reason evidence records in order of reason number to obtain the reason evidence list.

[0070] Based on the ballast manifest, generate a high buoyancy evidence limitation column; Iterate through the ballast manifests in the set, extracting the high buoyancy evidence identifier, anchor point gap item, restricted use label, original boundary field, and boundary status field from each ballast manifest. Write the high buoyancy evidence identifier into the evidence identifier field of the high buoyancy evidence restriction column, write the anchor point gap item into the gap field of the high buoyancy evidence restriction column, write the restricted use label into the use field of the high buoyancy evidence restriction column, write the original boundary field into the boundary field of the high buoyancy evidence restriction column, and write the boundary status field into the status field of the high buoyancy evidence restriction column. Then, summarize the data records corresponding to each ballast manifest in the order of the high buoyancy evidence identifier to generate the high buoyancy evidence restriction column.

[0071] Generate a decision instruction package based on the decision action field, the list of reasons and evidence, and the high buoyancy evidence limit field. The justification and evidence list refers to the evidence list generated based on the justification area of ​​the target executable decision candidate. It includes the original evidence identifier, justification content, anchor list, and evidence role supporting the candidate decision action. The high buoyancy evidence limitation column refers to the restrictive description column generated based on the ballast list. It is used to record the evidence identifier, anchor gap item, and limited use label of high buoyancy evidence, and indicates that high buoyancy evidence only enters the background description or question guidance position. The decision instruction package refers to the data object formed by the decision action field, the justification and evidence list, and the high buoyancy evidence limitation column. It is used to carry the data required for decision execution for government and enterprise think tank service targets. The decision action field, the list of reasons and evidence, and the high buoyancy evidence limit column are written into the same data object to generate a decision instruction package. In the decision instruction package, the decision action field is used as the execution content, the list of reasons and evidence is used as the execution basis, and the high buoyancy evidence limit column is used as the boundary description of the non-reason area. The decision instruction package is associated with the candidate number of the target executable decision candidate, the target candidate seat mapping table identifier, and the government-enterprise think tank service object identifier to obtain a decision instruction package for executing decisions on government-enterprise think tank service objects.

[0072] The decision instruction package is used to implement decisions for government and enterprise think tank service recipients and obtain decision results.

[0073] The service targets of government-enterprise think tanks refer to the matters that need to be analyzed and processed for decision-making, including project selection, resource allocation, industry analysis, risk assessment, or solution selection. Decision execution refers to the process of forming execution results for government-enterprise think tank service targets based on the decision action fields and the list of reasons and evidence in the decision instruction package. Decision results refer to the output results obtained after the decision execution is completed, which are used to represent the adoption results, ranking results, configuration results, or disposal results corresponding to the government-enterprise think tank service targets.

[0074] The decision instruction package is parsed to extract the government-enterprise think tank service object identifier, decision action field, reason evidence list, and high buoyancy evidence limitation column. Based on the government-enterprise think tank service object identifier, the government-enterprise think tank service object to be processed is located. The processing object, processing type, and execution content in the decision action field are written into the decision execution record corresponding to the government-enterprise think tank service object. The original evidence identifier, reason content, anchor point list, and evidence role in the reason evidence list are written into the basis field of the decision execution record. The high buoyancy evidence identifier, anchor point gap item, and limited purpose label in the high buoyancy evidence limitation column are written into the boundary description field of the decision execution record. The corresponding processing is performed on the government-enterprise think tank service object according to the execution content in the decision action field. After processing, a decision result containing the government-enterprise think tank service object identifier, execution content, basis field, boundary description field, and execution status is generated.

[0075] For example, the set of executable decision candidates contains three executable decision candidates, numbered C001, C002, and C003. The decision action corresponding to candidate C001 is to use supply chain optimization solution A as the execution solution. The justification evidence list includes original evidence E015, E021, and E034. The high buoyancy evidence constraint column includes the subject gap item and background position label corresponding to high buoyancy evidence E089. After traversing the set of executable decision candidates, C001 is determined as the target executable decision candidate. Solution A is written into the decision action field, and the justification content corresponding to E015, E021, and E034 is written into the justification evidence list. The subject gap item and background position label corresponding to E089 are written into the high buoyancy evidence. The system defines a limit field, generates a decision instruction package D001, parses the decision instruction package D001, locates the service object P001, writes the execution content corresponding to scheme A into the decision execution record of service object P001, writes the reason content corresponding to E015, E021 and E034 into the basis field, and writes the main gap item and background seat label corresponding to E089 into the boundary description field. After completing the execution processing of scheme A, the system generates a decision result R001. R001 records the service object identifier P001, the execution content as scheme A, the basis field containing E015, E021 and E034, the boundary description field containing E089, and the execution status as completed. R001 is then associated with the decision instruction package D001 and stored together.

[0076] It should be noted that in the multi-agent analysis and decision generation process for government and enterprise think tank services, some abstract evidence has the characteristics of wide interpretability, broad applicability, and easy repeated citation. During multiple rounds of analysis and transmission, it is easy for background information to gradually transform into justification, resulting in the phenomenon of abstract evidence occupying space. This causes the justification area in the candidate conclusion to be occupied by abstract evidence, thereby weakening the supporting role of factual evidence for the candidate conclusion. Therefore, after completing the seat occupancy audit, this solution only retains executable decision candidates whose justification area is entirely supported by occupancy evidence. Based on the set of executable decision candidates, the government and enterprise think tank service objects are executed. This ensures that the decision basis entering the decision execution stage comes from original evidence with clear boundaries and complete evidentiary support relationships, avoiding the impact of the phenomenon of abstract evidence occupying space on the decision-making process. This ensures that the decision results have a clear evidence source chain, traceable justification, and stable execution boundaries, ultimately resulting in interpretable decision results.

[0077] It should be noted that the multi-agent analysis in this invention does not involve multiple agents processing data simultaneously. Rather, it involves the phased flow and collaborative processing of original evidence through background analysis nodes, introduction analysis nodes, reason analysis nodes, solution analysis nodes, and decision analysis nodes, according to the routing order determined by the target candidate seat mapping table. Each analysis node undertakes different analytical responsibilities: the background analysis node is responsible for generating background content, the introduction analysis node is responsible for generating the matters to be analyzed, the reason analysis node is responsible for generating reason content, the solution analysis node is responsible for generating candidate processing solutions, and the decision analysis node is responsible for generating candidate decision actions. Each analysis node uses the analysis results output by the preceding node to continue to complete subsequent reasoning and jointly form candidate conclusions. Therefore, the core of this invention is not that a single analysis node completes all analysis tasks, but rather that through evidence transfer, division of responsibilities, result inheritance, and chain collaboration among multiple analysis nodes, it achieves hierarchical processing and joint decision-making of complex analysis tasks, thereby constituting a complete multi-agent analysis process.

[0078] One embodiment of the present invention also provides a multi-agent analysis and decision generation system for government and enterprise think tank services; in this embodiment, the functions of each module / unit are as follows: The anchor point creation module generates an anchor point list based on the original evidence, and determines the role of the original evidence according to the anchor point list to generate evidence role cards. The gap ballast module filters out high-buoyancy evidence role cards based on evidence role cards, and obtains a set of missing anchor point types based on high-buoyancy evidence role cards and a preset set of anchor point types; it configures limited-use tags according to the set of missing anchor point types and generates the ballast list corresponding to the high-buoyancy evidence; The seat swarm building module constructs multiple candidate seat mapping tables based on the evidence usage stage, evidence role cards, and ballast list in the multi-agent link. It then constructs an initial particle swarm by using the routing order corresponding to each piece of original evidence and the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table. The particle swarm optimization module iteratively updates the initial particle swarm based on the fitness values ​​corresponding to the candidate seat mapping table to determine the target candidate seat mapping table. The seat audit module performs multi-agent analysis on the original evidence based on the routing order in the target candidate seat mapping table to obtain candidate conclusions; it then performs seat occupancy audit on the candidate conclusions to obtain a set of executable decision candidates. The decision output module executes decisions for government and enterprise think tank service recipients based on the set of executable decision candidates, and obtains the decision results.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multi-agent analysis and decision generation for government and enterprise think tank service, characterized in that, The steps include: S1. Generate an anchor point list based on the original evidence, and determine the role of the original evidence according to the anchor point list to generate evidence role cards. S2. Filter out high-buoyancy evidence role cards based on evidence role cards, and obtain a set of missing anchor point types based on high-buoyancy evidence role cards and preset anchor point type sets; configure limited-use tags according to the set of missing anchor point types, and generate ballast manifests corresponding to high-buoyancy evidence; S3. Based on the evidence usage stage, evidence role card and ballast list in the multi-agent link, construct multiple candidate seat mapping tables. Construct the initial particle swarm through the routing order corresponding to each original piece of evidence and the set of reachable seats corresponding to each original piece of evidence in the candidate seat mapping table. S4. Based on the fitness values ​​corresponding to the candidate seat mapping table, the initial particle swarm is iteratively updated to determine the target candidate seat mapping table. S5. Based on the routing order in the target candidate seat mapping table, perform multi-agent analysis on the original evidence to obtain candidate conclusions; Perform a seat-occupancy audit on the candidate conclusions to obtain a set of executable decision candidates; S6. Based on the set of executable decision candidates, implement the decisions for the government and enterprise think tank service recipients to obtain the decision results.

2. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 1, characterized in that, Generate evidence character cards, including: Multiple pieces of original evidence were obtained; among them, the original evidence came from government and enterprise think tank service data sources; Anchor points are extracted from each piece of original evidence to obtain an anchor point list; the anchor point list includes subject anchor points, behavior anchor points, and source anchor points. Determine the evidentiary role of the original evidence based on the list of anchor points; Based on the list of anchor points, the original evidence is assigned a role to determine the evidence role; among which, the evidence role includes occupiable evidence and high buoyancy evidence. Generate evidence role cards based on evidence roles.

3. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 2, characterized in that, The set of missing anchor point types is obtained, including: From all the evidence role cards, select the evidence role cards that are high buoyancy evidence to obtain a set of high buoyancy evidence role cards; Obtain the High Buoyancy Evidence Character Card from the High Buoyancy Evidence Character Card Collection; Retrieve a set of preset anchor point types; the set of preset anchor point types includes main anchor point types, behavior anchor point types, and source anchor point types; The list of anchor points in the high buoyancy evidence character card is compared item by item with the preset anchor point type set to obtain the set of missing anchor point types.

4. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 3, characterized in that, Generate the ballast manifest corresponding to the high buoyancy evidence, including: Based on the set of missing anchor point types, identify the anchor point gap items corresponding to high buoyancy evidence; among them, anchor point gap items include subject gap items, behavior gap items, and source gap items; Based on the anchor point gap item corresponding to the high buoyancy evidence, configure a limited-use label for the high buoyancy evidence; the limited-use label includes a background seat label and an introductory seat label; Based on the evidence identifier, anchor point gap item, and limited purpose label of the high buoyancy evidence, a ballast manifest corresponding to the high buoyancy evidence is generated.

5. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 1, characterized in that, Constructing the initial particle swarm includes: Based on the evidence usage stage in the multi-agent link, a set of seats is constructed; wherein, the set of seats includes background seats, introductory seats, reason seats, alternative seats, and decision seats; Based on evidence role cards, ballast lists, and seat sets, multiple candidate seat mapping tables are constructed; each candidate seat mapping table includes the reachable seat set and routing order corresponding to each piece of original evidence. Define the candidate seat mapping table as a particle in the particle swarm algorithm; The initial velocity of the particle is determined by the routing order corresponding to each piece of original evidence in the candidate seat mapping table; The initial position of the particle is determined by the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table; Construct an initial particle swarm based on the initial velocity and initial position of the particles.

6. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 1, characterized in that, Determine the target candidate seat mapping table, including: Based on the evidence role cards and ballast slips, the data on seat violations was determined; Based on the seat violation data, calculate the fitness value corresponding to the candidate seat mapping table; Based on the fitness value, the optimal position of the individual particle and the global optimal position are determined; The initial particle swarm is iteratively updated based on the optimal position of each individual particle and the global optimal position, resulting in the particle swarm after the iteration. Based on the particle swarm after the iteration, a target candidate seat mapping table is determined.

7. The multi-agent analysis and decision generation method for government and enterprise wisdom library service according to claim 1, characterized in that, Candidate conclusions were obtained, including: According to the routing order in the target candidate seat mapping table, each piece of original evidence is distributed to the corresponding multi-agent analysis node; wherein, the multi-agent analysis node includes background analysis node, introduction analysis node, reason analysis node, scheme analysis node and decision analysis node; Based on the set of reachable seats in the target candidate seat mapping table, determine the available seats for the original evidence in the multi-agent analysis node; Based on available seats, multi-agent analysis is performed on the original evidence to obtain candidate conclusions; each candidate conclusion includes a reasoning area and a background area.

8. The multi-agent analysis and decision generation method for government and enterprise think tank service according to claim 7, characterized in that, The set of executable decision candidates is obtained, including: Based on the evidence role card, determine the evidence role corresponding to the reasoning section; Based on the evidence role corresponding to the reasoning area, the candidate conclusions are audited for their place in the decision-making process to obtain executable decision candidates. By summing up all executable decision candidates, we obtain the executable decision candidate set.

9. The multi-agent analysis and decision generation method for government and enterprise wisdom library service according to claim 1, characterized in that, The decision results include: Based on the set of executable decision candidates, determine the target executable decision candidates; Generate a decision action field based on the candidate decision actions from the target executable decision candidates; Based on the reasoning section of the actionable decision candidates, determine the list of justification evidence; Based on the ballast manifest, generate a high buoyancy evidence limitation column; Generate a decision instruction package based on the decision action field, the list of reasons and evidence, and the high buoyancy evidence limit field. The decision instruction package is used to implement decisions for government and enterprise think tank service recipients and obtain decision results.

10. A system for applying the multi-agent analysis and decision generation method for government and enterprise think tank services as described in any one of claims 1-9, characterized in that, The system includes: The anchor point creation module generates an anchor point list based on the original evidence, and determines the role of the original evidence according to the anchor point list to generate evidence role cards. The gap ballast module filters out high-buoyancy evidence role cards based on evidence role cards, and obtains a set of missing anchor point types based on high-buoyancy evidence role cards and a preset set of anchor point types; it configures limited-use tags according to the set of missing anchor point types and generates the ballast list corresponding to the high-buoyancy evidence; The seat swarm building module constructs multiple candidate seat mapping tables based on the evidence usage stage, evidence role cards, and ballast list in the multi-agent link. It then constructs an initial particle swarm by using the routing order corresponding to each piece of original evidence and the set of reachable seats corresponding to each piece of original evidence in the candidate seat mapping table. The particle swarm optimization module iteratively updates the initial particle swarm based on the fitness values ​​corresponding to the candidate seat mapping table to determine the target candidate seat mapping table. The seat audit module performs multi-agent analysis on the original evidence based on the routing order in the target candidate seat mapping table to obtain candidate conclusions; it then performs seat occupancy audit on the candidate conclusions to obtain a set of executable decision candidates. The decision output module executes decisions for government and enterprise think tank service recipients based on the set of executable decision candidates, and obtains the decision results.

Images