Search

CN-121998746-A - Bid-recruitment purchase compliance and risk decision method based on bipolar theory

CN121998746ACN 121998746 ACN121998746 ACN 121998746ACN-121998746-A

Abstract

The invention discloses a bid-recruitment purchase compliance and risk decision method based on a bipolar theory, which comprises the steps of firstly carrying out structuring and atomization treatment on bid documents and laws and regulations through DeepOCR and a large language model, then fusing internal atomic arguments and external evidences to construct a bidirectional argumentation graph containing supporting relations and attack relations, then carrying out iterative computation on the standing strength score of each arguments through an aggregation influence model to realize quantitative risk assessment, finally generating structured audit opinions with traceable evidence chains, and automatically distributing the structured audit opinions to expert review or automatic approval channels according to risk grades. And precipitating expert correction opinions into new knowledge through a man-machine cooperative feedback mechanism, so as to realize continuous self-evolution of the bidding knowledge body network. The method and the system remarkably improve the accuracy, the interpretability and the adaptability of compliance judgment, and are suitable for scenes such as government purchasing, engineering bidding, enterprise internal control and the like.

Inventors

  • ZHANG ZHE

Assignees

  • 杭州科技职业技术学院(杭州开放大学、杭州远程教育中心、杭州社区大学、杭州市民大学、杭州广播电视中等专业学校)

Dates

Publication Date
20260508
Application Date
20260116

Claims (10)

  1. 1. The bid-recruitment purchase compliance and risk decision method based on the bipolar theory is characterized by comprising the following steps of: S1, acquiring legal and regulation treaty related to bidding purchase and electronic documents of historical bidding cases; s2, carrying out structural pretreatment on the electronic document to generate a structured text with clear hierarchy; S3, inputting the structured text into a large language model, splitting the structured text into an atomization logic unit, and converting the atomization logic unit into structured JSON data; S4, constructing a bidding knowledge ontology network in a graph database based on the structured JSON data; s5, receiving an electronic document of the bid-inviting requirement submitted by the purchasing demand party; s6, extracting an atomic argument set from the electronic document of the bid request by using a large language model; s7, analyzing the supporting and attacking relation inside the atomic discussion point set to generate an internal evidence set; s8, based on each atomic argument in the atomic argument set, searching relevant external evidence in the bidding knowledge ontology network to generate an external evidence set; S9, combining the internal evidence set and the external evidence set to construct a bipolar weighted arguments graph containing internal and external correlations; S10, running an aggregation influence model on the bipolar weighted arguments, and calculating the standing strength of each arguments to obtain a final standing strength mapping table; S11, in the bipolar weighted argumentation, combining the final standpoint intensity mapping table, evaluating the risk level according to the standpoint intensity of each arguments, constructing an influence tracing algorithm, and tracing a core causal evidence chain; And S12, generating a structured audit opinion text based on the causal evidence chain.
  2. 2. The method of claim 1, wherein building a bid ontology network in a graph database based on the structured JSON data comprises: S401, regarding each historical bidding case node, using key value pairs in the structured JSON data as attributes of the historical bidding case node, converting the dispute focus text into semantic embedded vectors through a large language model, and storing the semantic embedded vectors as node vector attributes; s402, traversing all historical bidding case nodes, finding legal regulation treaty nodes corresponding to the historical bidding case nodes according to legal association basis attributes of each historical bidding case node, and creating an explanation relationship between the two nodes; S403, traversing risk tag arrays of all historical bidding case nodes, and establishing a risk relation between the historical bidding case nodes and risk tag nodes preset in the map aiming at each tag of the arrays so as to facilitate quick retrieval of all related cases through risk types in the future; S404, circularly executing the processing flow until all the historical bidding case nodes finish map mapping; s405, for the past expert to audit and finally adopt the dispute clauses, creating an expert experience knowledge node, and embedding the text of the expert opinion text and the core opinion text into vectors, and taking the expert ID, the project ID and the timestamp as attributes of the expert experience knowledge node; S406, for each expert experience knowledge node, traversing each legal regulation provision node in the atlas, calculating cosine similarity between a text embedded vector of the expert experience knowledge node and a text vector of the legal regulation provision node, and creating a supporting relationship between the expert experience knowledge node and the legal regulation provision node when the similarity is greater than a preset threshold; S407, traversing each historical bidding case node in the atlas for each expert experience knowledge node, analyzing whether an applicable relation, a reject relation or an irrelevant relation exists between the expert experience knowledge node and the historical bidding case node by using LLM, if the applicable relation exists, creating the applicable relation between the two nodes, and if the applicable relation exists, creating the reject relation between the two nodes; S408, traversing all legal provision nodes in the graph database, and establishing conflict relations between the two legal provision nodes.
  3. 3. The method of claim 2, wherein the bid ontology network constructed in S4 comprises the following node types and weighted relationships between nodes: the node types comprise legal regulation provision nodes, historical bidding case nodes and expert experience knowledge nodes; the weighted relation among the nodes comprises: (1) The interpretation relation points to legal regulation treaty nodes from historical bidding case nodes, the specific interpretation of the case to the legal regulation treaty is represented, and the weight of the interpretation relation is set to be 0.9; (2) Supporting relation, namely pointing from expert experience knowledge nodes to legal regulation provision nodes, representing the applicability of expert experience to support the legal provision, wherein the weight of the supporting relation is 0.7; (3) The conflict relation is established between two legal regulation treaty nodes, which indicates that the two have direct legal conflict, and the weight of the conflict relation is set to be 1.0; (4) The application relation points to a historical bidding case node from an expert experience knowledge node, which indicates that expert experience is applicable to case situation, and weight of the application relation is 0.8; (5) Reject relation, namely, pointing from a historical bidding case node to an expert experience knowledge node, wherein the case rejects expert experience, and the weight of the reject relation is set to 0.85.
  4. 4. The method of claim 2, wherein the specific process of establishing a conflicting relationship between two legal provision nodes comprises the steps of: S4081, traversing all legal regulation nodes in a graph database, generating text vectors based on regularization description corresponding to each legal regulation node, calculating semantic similarity between any two legal regulation nodes, and screening out all node pairs with similar semantics or related subjects as candidate processing objects; S4082, judging whether the respective space jurisdictions of the node pairs are overlapped or not according to the node pairs processed currently, if the two jurisdictions are not overlapped, judging that no conflict exists to skip the node pair, and continuing to process the next node pair, if so, entering S4083; s4083 comparing the efficacy levels of the node pairs to determine whether they are at the same efficacy level: If the efficacy levels are different, judging the higher efficacy level as an upper method and the lower efficacy level as a lower method according to a preset efficacy priority rule, creating a conflict relation pointing from the lower method to the upper method when the upper method and the lower method are semantically similar but have potential conflicts, otherwise creating a conflict relation pointing from the upper method to the lower method; when the efficacy levels are the same, then compare if the date of effectiveness of the node pairs is the same: If the effective dates are different, the legal and regulation regulations with the later effective dates are regarded as new laws, the earlier legal and regulation regulations are regarded as old laws, and a relation edge pointing from the old laws to the new laws is created in the graph database, so that the old laws are replaced by the new laws; If the effective dates are the same, no relation is created; And S4084, after one node pair is circularly processed and the corresponding relation is created, continuing to process the next node pair until all candidate node pairs are processed.
  5. 5. A method according to claim 3, wherein S9 specifically comprises: S901, initializing an empty graph data structure for storing nodes and edges; S902, traversing an atomic argument set, adding each atomic argument as an independent internal argument node into a node set of a graph data structure, and carrying all attribute information of the node set; S903, traversing the external evidence set, adding each external evidence in the external evidence set as an independent external evidence node into the node set of the graph data structure, and carrying all the attributes of the external evidence in the bidding knowledge ontology network; S904, traversing an internal evidence set, creating a directed edge between corresponding internal argument nodes in a graph data structure for each internal relation, wherein the type of the edge is support or attack, and setting a weight value of the edge; S905, traversing the combined internal evidence set and external evidence set, and creating a directed edge between the corresponding internal argument node and external evidence node in the graph data structure for each internal and external cross relation in the combined internal evidence set and external evidence set; S906, traversing all external evidence node pairs in an external evidence set, judging whether a weighted relation exists between the external evidence node pairs by querying a graph database, and if so, adding the relation as an edge into a graph data structure, and keeping the original relation type and weight unchanged; S907. After all nodes and relationships have been successfully added, the complete bipolar weighted arguments are returned.
  6. 6. The method of claim 1, wherein S10 comprises: S1001, inputting a bipolar weighted argumentation chart, a basic strength mapping table, an influence factor alpha, the maximum iteration number and a convergence threshold; s1002, initializing a standing strength mapping table, and setting the initial standing strength of each node as the corresponding basic strength; S1003, traversing all supporters of each node in the bipolar weighted arguments, and calculating the total supporting influence; traversing all attackers, and calculating total adverse influence; S1004, calculating the standing strength of the node in the next round according to the basic strength, the total supporting influence and the total counterintuitive influence; S1005, performing normalization processing on the calculated standing strength by applying a Tanh activation function, and compressing the standing strength into a [ -1, 1] interval; S1006, storing the normalized standing strength into a standing strength mapping table, and taking the standing strength as the standing strength of the node after the updating of the round; S1007, comparing absolute difference values of the standing intensities before and after updating, and if the change of any node is greater than a preset convergence threshold value, setting a convergence mark as false; After one round of iteration is completed, judging whether the convergence mark is still true; if true, the whole bipolar weighted arguments are judged to have reached a stable state, iteration is terminated in advance, and a final standing strength mapping table is output, and if false, S1003 to S1007 are repeatedly executed until convergence conditions are met or the maximum number of iterations is reached, and the final standing strength mapping table is output.
  7. 7. The method of claim 6, wherein the node's position strength at the next round is calculated as: Wherein, the For the next round of position strength of node e, BS (e) is the base strength, alpha is the influencing factor, And The weights of the supporting relationship and the attacking relationship respectively, And The supporters and aggressors are respectively from the standpoint of the current turn.
  8. 8. The method of claim 1, wherein the trace back core causal evidence chain comprises: S1101, inputting an atomic arguments set to be interpreted, a bipolar weighted arguments graph, a final stand strength mapping table and a maximum backtracking depth, taking the atomic arguments as current target risk nodes, initializing an attack chain and a support chain, respectively used for storing key nodes in an attack path and a support path, and respectively executing S1102 and S1104; S1102, screening out relation edges with the type of attack from all input edges of the current target risk nodes, calculating the actual attack influence of the source node of each edge, selecting the node with the largest actual attack influence as the strongest attacker of the layer, and adding the node to an attack chain; S1103, taking the strongest attacker as a new current target risk node, repeatedly executing S1102, continuing to trace the upstream attacker upwards until no effective attacker exists or the maximum tracing depth is reached, and completing the construction of an attack chain; s1104, selecting relation edges with the type being support from all input edges of the current target risk nodes, calculating the actual support influence of source nodes of each edge, selecting the node with the largest actual support influence as the strongest support of the layer, and adding the node to a support chain; s1105, taking the strongest supporters as new current target risk nodes, repeatedly executing S1104, continuing to trace back upstream supporters until no effective supporters exist or the maximum tracing depth is reached, and completing construction of a support chain; And S1106, after the attack chain and the support chain are traced, the target risk node, the final stand strength score, the support chain and the attack chain which need to be interpreted are packaged into a structured argumentation.
  9. 9. The method as recited in claim 1, further comprising: Creating two empty task lists, namely a high-priority task list and a low-priority task list, which are respectively used for storing the examination task needing expert deep intervention and the examination task capable of passing through quickly; Traversing each audit opinion in the structured audit opinion text, judging that the audit opinion has high dispute for the stroke insurance audit opinion and the dispute audit opinion according to the standing intensity score corresponding to the audit opinion, and adding the audit opinion into a high-priority task list; Sending the audit opinion in the high-priority task list to an expert deep examination workbench for manual review and intervention; And directly pushing audit opinions in the low-priority task list to an automatic approval channel to realize quick circulation.
  10. 10. The method as recited in claim 9, further comprising: and storing the adopted expert opinion into a bidding knowledge ontology network in the form of expert experience knowledge nodes to realize the self-evolution updating of the knowledge graph.

Description

Bid-recruitment purchase compliance and risk decision method based on bipolar theory Technical Field The invention relates to the technical field of artificial intelligence, in particular to a bid-recruitment purchase compliance and risk decision method based on a bipolar theory. Background Bidding purchasing is an important way of enterprise resource allocation, and legal program and fair result are required to be ensured. However, in practice, the quality of the bidding documents is uneven, and bidder ring bid strings, complaints question and even legal disputes are often caused by problems of fuzzy clauses, improper qualification condition setting, unreasonable scoring standards and the like. Traditional compliance inspection mainly relies on manual experts to check rules one by one, and has the bottlenecks of low efficiency, strong subjectivity, limited knowledge coverage and the like. In recent years, some organizations have attempted to introduce Natural Language Processing (NLP) and knowledge-graph technology to aid in censoring. For example, french reference inspection is achieved by matching keywords by a rules engine, or high risk expressions are identified using a pre-trained language model. However, these methods have the following limitations: (1) The depth reasoning capability is lacking, namely, only surface semantic matching can be performed, and the supporting/conflict logic relation between the arguments cannot be modeled; (2) Internal and external knowledge splitting, wherein the internal bidding clause and external laws, cases and expert experience do not form a unified reasoning framework; (3) The result can not be explained, the large model outputs 'black box' judgment, and a traceable evidence chain is difficult to provide for an expert to review; (4) The dynamic adaptability is poor, and the knowledge system can not be automatically updated according to the newly issued regulations or typical cases. In addition, the existing systems mostly adopt unidirectional "applicability judgment", neglecting the two-way arguments that are common in bidding scenes-that is, the same clause may be supported or rejected by multiple authorities at the same time. How to quantify this complex impact and generate audit opinions with deterministic assessment becomes a key technical challenge in the current intelligent compliance field. Disclosure of Invention In view of the above, the present invention has been made to provide a bid-to-bid procurement compliance and risk decision method based on bipolar theory that overcomes or at least partially solves the above-described problems, In order to achieve the above purpose, the present invention adopts the following technical scheme: A bid-recruitment purchase compliance and risk decision method based on bipolar theory comprises the following steps: S1, acquiring legal and regulation treaty related to bidding purchase and electronic documents of historical bidding cases; s2, carrying out structural pretreatment on the electronic document to generate a structured text with clear hierarchy; S3, inputting the structured text into a large language model, splitting the structured text into an atomization logic unit, and converting the atomization logic unit into structured JSON data; S4, constructing a bidding knowledge ontology network in a graph database based on the structured JSON data; s5, receiving an electronic document of the bid-inviting requirement submitted by the purchasing demand party; s6, extracting an atomic argument set from the electronic document of the bid request by using a large language model; s7, analyzing the supporting and attacking relation inside the atomic discussion point set to generate an internal evidence set; s8, based on each atomic argument in the atomic argument set, searching relevant external evidence in the bidding knowledge ontology network to generate an external evidence set; S9, combining the internal evidence set and the external evidence set to construct a bipolar weighted arguments graph containing internal and external correlations; S10, running an aggregation influence model on the bipolar weighted arguments, and calculating the standing strength of each arguments to obtain a final standing strength mapping table; S11, in the bipolar weighted argumentation, combining the final standpoint intensity mapping table, evaluating the risk level according to the standpoint intensity of each arguments, constructing an influence tracing algorithm, and tracing a core causal evidence chain; And S12, generating a structured audit opinion text based on the causal evidence chain. Preferably, constructing a bidding knowledge ontology network in a graph database based on the structured JSON data includes: S401, regarding each historical bidding case node, using key value pairs in the structured JSON data as attributes of the historical bidding case node, converting the dispute focus text into semantic embedded vectors through a la