Supplier relationship type identification method and related device
By combining multimodality and large language models with RPA technology to automatically process supplier relationship identification, the problem of low efficiency of manual analysis in existing technologies is solved, the automatic identification and accuracy improvement of supplier relationship types are achieved, and the scope of bidding collusion is effectively prevented.
Patent Information
- Application Number
- CN202510791995.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-12
Smart Images
Figure CN120634686A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a supplier relationship type identification method and related devices. Background Art
[0002] In public bidding, tendering entities typically require suppliers to be independent and impartial. However, in actual bidding, some suppliers may have relationships, such as equity control or overlapping personnel. This could constitute illegal bid rigging and collusion, seriously disrupting market order and undermining the principle of fair competition.
[0003] In order to ensure the transparency and fairness of bidding activities, the audit department often needs to identify and analyze whether there are any related relationships between bidding companies and give the relationship category when supervising or auditing bidding projects.
[0004] However, current work mainly relies on manual analysis of relationship categories, which leads to low efficiency. Summary of the Invention
[0005] In view of this, the present application provides a supplier relationship type identification method and related devices, which realize the automation of the supplier relationship type identification process and effectively reduce labor costs.
[0006] To solve the above problems, the technical solutions provided by this application are as follows:
[0007] In one aspect, the present application provides a method for identifying a supplier relationship type, the method comprising:
[0008] Obtaining a relationship graph, wherein the relationship graph is used to identify the association relationship between suppliers in the target bidding activity;
[0009] Inputting the relationship image and the first prompt word into a multimodal model to obtain a first-level relationship text, wherein the first prompt word is used to instruct the multimodal model to identify the direct relationships corresponding to the respective suppliers;
[0010] The first-level relationship text and the second prompt word are input into the large language model to obtain the predicted relationship type and corresponding description between the suppliers. The second prompt word is used to instruct the large language model to classify the association relationship between the suppliers based on the relationship type and generate corresponding classification basis.
[0011] In a possible implementation, obtaining the relationship picture includes:
[0012] Use Robotic Process Automation (RPA) technology to traverse all suppliers in the target bidding activity, and input the suppliers into the enterprise relationship query tool to obtain a relationship picture.
[0013] In a possible implementation, before inputting the relationship image and the first prompt word into the multimodal model to obtain the first-level relationship text, the method further includes:
[0014] Acquire training data, where the training data includes multiple relationship graphs and corresponding labels, where the labels are used to identify the direct relationship corresponding to each supplier in each relationship graph;
[0015] Based on the training data and the third prompt word, the initial multimodal model is fine-tuned to obtain a multimodal model. The third prompt word is used to instruct the initial multimodal model to identify the direct relationship corresponding to each supplier in each relationship picture and output the corresponding label.
[0016] In one possible implementation, fine-tuning the initial multimodal model based on the training data and the third prompt word to obtain the multimodal model includes:
[0017] According to the training data and the third prompt word, the parameters of the initial multimodal model are adjusted by the Lora fine-tuning method to obtain the multimodal model.
[0018] In a possible implementation, the relationship type includes:
[0019] Different suppliers with the same person as the person in charge (legal person, investor of a sole proprietorship, partner who executes partnership affairs on behalf of a partnership), different suppliers with a controlling stake (holding more than 50% or relative controlling stake), different suppliers with a shareholding (non-controlling stake), different suppliers with a management relationship (a management-being-managed relationship between other units that do not have a capital contribution or shareholding relationship, such as public institutions), different suppliers with common management personnel (a director, supervisor or senior management of a supplier concurrently serves as a director, supervisor or senior management of another supplier), and different suppliers with no direct weak relationship.
[0020] In one possible implementation, after inputting the first-level relationship text and the second prompt word into a large language model to obtain the predicted relationship types and corresponding descriptions between the suppliers, the method further includes:
[0021] Through the RPA technology, the predicted relationship type and the corresponding description are written into the table record corresponding to the target bidding activity.
[0022] In another aspect, the present application provides a device for identifying a supplier relationship type, the device comprising an acquisition unit and an identification unit:
[0023] The acquisition unit is used to acquire a relationship picture, wherein the relationship picture is used to identify the relationship between the suppliers in the target bidding activity;
[0024] The recognition unit is configured to input the relationship image and the first prompt word into a multimodal model to obtain a first-level relationship text, wherein the first prompt word is used to instruct the multimodal model to recognize the direct relationship corresponding to each of the suppliers;
[0025] The recognition unit is further used to input the first-level relationship text and the second prompt word into the large language model to obtain the predicted relationship type and corresponding description between the suppliers. The second prompt word is used to instruct the large language model to classify the association relationship between the suppliers based on the relationship type and generate corresponding classification basis.
[0026] In a possible implementation, the acquiring unit is specifically configured to:
[0027] Use Robotic Process Automation (RPA) technology to traverse all suppliers in the target bidding activity, and input the suppliers into the enterprise relationship query tool to obtain a relationship picture.
[0028] In a possible implementation, the apparatus further includes a fine-tuning unit, configured to:
[0029] Acquire training data, where the training data includes multiple relationship graphs and corresponding labels, where the labels are used to identify the direct relationship corresponding to each supplier in each relationship graph;
[0030] Based on the training data and the third prompt word, the initial multimodal model is fine-tuned to obtain a multimodal model. The third prompt word is used to instruct the initial multimodal model to identify the direct relationship corresponding to each supplier in each relationship picture and output the corresponding label.
[0031] In a possible implementation, the fine-tuning unit is specifically configured to:
[0032] According to the training data and the third prompt word, the parameters of the initial multimodal model are adjusted by the Lora fine-tuning method to obtain the multimodal model.
[0033] In a possible implementation, the relationship type includes:
[0034] Different suppliers with the same person as the person in charge (legal person, investor of a sole proprietorship, partner who executes partnership affairs on behalf of a partnership), different suppliers with a controlling stake (holding more than 50% or relative controlling stake), different suppliers with a shareholding (non-controlling stake), different suppliers with a management relationship (a management-being-managed relationship between other units that do not have a capital contribution or shareholding relationship, such as public institutions), different suppliers with common management personnel (a director, supervisor or senior management of a supplier concurrently serves as a director, supervisor or senior management of another supplier), and different suppliers with no direct weak relationship.
[0035] In a possible implementation, the apparatus further includes an input unit, which is configured to:
[0036] Through RPA technology, the predicted relationship type and the corresponding description are written into the table record corresponding to the target bidding activity.
[0037] In another aspect, the present application provides a computer device, comprising a processor and a memory:
[0038] The memory is used to store computer programs;
[0039] The processor is configured to execute any one of the above methods according to the computer program.
[0040] On the other hand, the present application provides a computer-readable storage medium, which is used to store a computer program. When the computer program is executed by a computer device, it implements any of the methods described above.
[0041] On the other hand, the present application provides a computer program product comprising a computer program, which, when executed on a computer device, enables the computer device to execute any one of the methods described above.
[0042] As can be seen from the above technical solution, the technical solution first obtains a relationship image used to identify the relationship between each supplier in the target bidding activity, then inputs the relationship image and the first prompt word into the multimodal model, guides the multimodal model through the first prompt word to identify the direct relationship corresponding to each supplier in the relationship image, and obtains the first-level relationship text describing the direct relationship, and finally inputs the first-level relationship text and the second prompt word into the large language model, classifies the relationship between each supplier by the relationship type in the second prompt word, and obtains the predicted relationship type between each supplier and the description of the corresponding classification basis. Among them, the first-level relationship text extracted from the relationship image can provide the large language model with rich information with direct relationships with each supplier, helping it to analyze the key hubs used to connect the relationships between each supplier, so that the second prompt word can guide the large language model to identify the relationship type between suppliers and give the corresponding classification basis, realizing the automatic identification of relationship types, effectively saving labor costs, and improving the accuracy and efficiency of relationship type identification, so as to accurately combat the risks of bid rigging and collusion in the bidding process and accelerate the audit process of the bidding and procurement department. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0044] Figure 1 A flowchart of a method for identifying supplier relationship types provided in an embodiment of the present application;
[0045] Figure 2 This is one of the schematic diagrams of a relationship diagram provided in an embodiment of the present application;
[0046] Figure 3 This is a second schematic diagram of a relationship diagram provided in an embodiment of the present application;
[0047] Figure 4 This is a third schematic diagram of a relationship diagram provided in an embodiment of the present application;
[0048] Figure 5 This is a fourth schematic diagram of a relationship diagram provided in an embodiment of the present application;
[0049] Figure 6 This is a fifth diagram of a relationship diagram provided in an embodiment of the present application;
[0050] Figure 7 This is a sixth schematic diagram of a relationship diagram provided in an embodiment of the present application;
[0051] Figure 8 Schematic diagram seven of a relationship diagram provided in an embodiment of the present application;
[0052] Figure 9 This is a schematic diagram of a relationship diagram provided in an embodiment of the present application;
[0053] Figure 10 A schematic diagram of the overall design of a supplier relationship type identification method provided in an embodiment of the present application;
[0054] Figure 11 A flowchart of a method for identifying an application supplier relationship type provided in an embodiment of the present application;
[0055] Figure 12 A schematic diagram of a supplier relationship type identification device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0056] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.
[0057] As described in the background art, the relationship types between suppliers are currently mainly identified through manual analysis, which results in high labor costs and low efficiency.
[0058] In order to solve the above problems, the present application provides a method for identifying supplier relationship types. First, a relationship image is obtained for identifying the relationship between each supplier in the target bidding activity. Then, the relationship image and the first prompt word are input into a multimodal model. The first prompt word is used to guide the multimodal model to identify the direct relationship corresponding to each supplier in the relationship image, and obtain the first-level relationship text describing the direct relationship. Finally, the first-level relationship text and the second prompt word are input into a large language model. The relationship between each supplier is classified according to the relationship type in the second prompt word, and the predicted relationship type between each supplier and the description of the corresponding classification basis are obtained. Among them, the first-level relationship text extracted from the relationship image can provide the large language model with rich information that has a close relationship with each supplier, helping it to analyze the key hubs used to connect the relationships between each supplier, so that the second prompt word can guide the large language model to identify the relationship type between suppliers and give the corresponding classification basis, realizing the automatic identification of relationship types, effectively saving labor costs, and improving the accuracy and efficiency of relationship type identification, so as to accurately combat the risks of bid rigging and collusion in the bidding process and accelerate the audit process of the bidding and procurement department.
[0059] The solutions provided by the embodiments of the present application relate to the field of computer technology and are specifically described through the following embodiments.
[0060] See also Figure 1 As shown, it is a flow chart of a supplier relationship type identification method provided in an embodiment of the present application. In this embodiment, it can be executed by a computer device as an example.
[0061] S101: Obtain a relationship picture.
[0062] The relationship diagram is used to identify the relationships between suppliers in the target bidding activity. The relationships between suppliers include not only direct relationships, but also other complex relationships formed through connections with other companies or individuals.
[0063] If there is no relationship between the suppliers in a bidding activity, it means that there is no violation in the bidding activity and the suppliers can participate in the bidding normally. If there is a relationship between the suppliers, it means that there is a risk of collusion in the bidding activity. It is necessary to identify the relationship type between the suppliers and conduct an in-depth analysis of the violation type so that appropriate measures can be taken.
[0064] In the embodiment of the present application, there is an association relationship between the suppliers in the target bidding activity. In the prior art, the association relationship between each supplier is obtained by inputting the supplier entities in the target bidding activity into the enterprise relationship query tool at the same time, and the association relationship is clearly displayed in the form of a picture to facilitate auditors to analyze the relationship type.
[0065] refer to Figure 2 As shown, this is one of the schematic diagrams of a relationship diagram provided in an embodiment of the present application. The relationship diagram can be obtained through an enterprise relationship query tool, in which the relationship between supplier A, supplier B and supplier C is displayed through personnel A, B, enterprise a and the corresponding direct relationship (such as personnel A invests 100% in supplier A and is the legal representative and executive director of supplier A).
[0066] In a possible implementation, step S101 includes:
[0067] Use RPA technology to traverse all suppliers in the target bidding activity, input each supplier into the enterprise relationship query tool, and obtain a relationship picture.
[0068] Robotic Process Automation (RPA) is a technology that uses software robots (also known as "digital robots") to automatically perform repetitive, rule-based tasks. The core of RPA technology is to automate business processes by simulating human actions on computers (such as clicking, typing, copying and pasting, and reading data), thereby improving work efficiency, reducing human errors, and lowering labor costs.
[0069] Currently, enterprises generally have a system dedicated to managing bidding and tendering information, which is called a bidding and procurement management system. This type of system usually has the following functions: (1) bidding document management; (2) bidding information management: recording the information of bidding companies, bidding documents, quotations, etc.; (3) bid evaluation process management: recording the evaluation process and results of the bid evaluation committee; (4) contract management: generating a contract after the bid is won and linking it with the bidding information.
[0070] RPA technology can be used to batch export the names of all suppliers in multiple bidding activities from the bidding and procurement management system. Then, RPA can traverse the names of each supplier in the target bidding activity and automatically input multiple supplier names into the enterprise relationship query tool to obtain a relationship picture that describes the relationship between all suppliers in the activity.
[0071] Therefore, using RPA technology instead of manpower to query the relationship picture of suppliers in a bidding activity can avoid omissions, effectively reduce labor costs, and improve the efficiency of querying related relationships.
[0072] S102: Input the relationship image and the first prompt word into the multimodal model to obtain the first-level relationship text.
[0073] The first prompt word is used to indicate the direct relationship between each supplier in the multimodal model recognition relationship graph. The direct relationship between each supplier refers to the person or company that has a direct relationship with the supplier and the corresponding relationship description. For example Figure 2 Among them, the person who has a direct relationship with Supplier A is Person A, who invests 100% in Supplier A and is the legal representative and executive director of Supplier A.
[0074] A multimodal model means that its input can be data of different modalities, that is, relationship pictures and prompt words can both be used as input data and input into the multimodal model to obtain the first-level relationship text used to describe the direct relationship between suppliers.
[0075] As an example, the first prompt is "Please play the role of an enterprise relationship analysis assistant and output the first-level relationships of the entities [Supplier A, Supplier B, Supplier C] respectively." Figure 2 The prompt word is input into the multimodal model, which can output the direct relationship between the three suppliers:
[0076] "Entity 1: Supplier A
[0077] relation:
[0078] Party A->Investment 100%, Legal Representative, Executive Director->Supplier A
[0079] Entity 2: Supplier B
[0080] relation:
[0081] B->Investment 90%, Legal Representative, Executive Director, General Manager->Supplier B
[0082] Entity 3: Supplier C
[0083] relation:
[0084] B->Operator->Supplier C"
[0085] The associations contained in the relationship diagram are relatively complex. The closer the relationship between suppliers, the greater the possibility of bid rigging and collusion. Therefore, the relationship types of close relationships between suppliers are usually identified.
[0086] In the embodiment of the present application, a multimodal model can be used to mine rich information about companies or individuals that have direct relationships with each supplier, and then correspond it with the information of companies or individuals in the relationship descriptions of other suppliers, which can help analyze the potential close associations between suppliers, thereby facilitating the identification of relationship types.
[0087] S103: Input the first-level relationship text and the second prompt word into the large language model to obtain the predicted relationship type and corresponding description between each supplier.
[0088] The second prompt word is used to instruct the large language model to classify the relationships between suppliers based on the relationship type and generate classification basis.
[0089] Relationship types are categorized based on the experience of tender auditors. They can be classified as either direct relationships between suppliers or indirect relationships. Direct relationships can be categorized into equity control, overlapping personnel, and other types based on the specific first-level relationship text descriptions, to facilitate audits by tender auditors.
[0090] In one possible implementation, the relationship types include:
[0091] The person in charge (legal person, investor of a sole proprietorship, partner who represents a partnership in executing partnership affairs) is the same person, there is a controlling relationship (holding more than 50% or relative controlling stake), a shareholding relationship (non-controlling), a management relationship (a management-management relationship between other units that do not have a capital contribution and shareholding relationship, such as a public institution), co-management (a director, supervisor or senior manager of a supplier who also serves as a director, supervisor or senior manager of another supplier), and there is no direct weak relationship.
[0092] refer to Figure 3 The figure shows a second diagram of a relationship diagram provided by an embodiment of the present application. In the figure, the relationship type among Supplier A, Supplier B, and Supplier C is that the person in charge (legal person, investor of a sole proprietorship, partner who represents a partnership in executing partnership affairs) is the same person. Figure 4 The figure shows the third diagram of a relationship diagram provided by the embodiment of the present application. In this diagram, there is a controlling relationship between the suppliers (holding more than 50% or relatively controlling). Figure 5 The figure shows a fourth diagram of a relationship diagram provided in an embodiment of the present application, in which there is a shareholding relationship (non-controlling) between the suppliers. Figure 6 As shown, this is a fifth diagram of a relationship diagram provided by an embodiment of the present application. In this diagram, there is a management relationship between the suppliers (a management and managed relationship between other units that do not have a capital contribution and shareholding relationship, such as a public institution). Figure 7 The figure shows the sixth diagram of a relationship diagram provided by an embodiment of the present application. In the figure, there are common management personnel between the suppliers (the directors, supervisors or senior management personnel of a supplier also serve as directors, supervisors or senior management personnel of another supplier). Figure 8As shown, this is the seventh schematic diagram of a relationship diagram provided in an embodiment of the present application. In this diagram, there are only weak relationships without direct relationships between the suppliers, and the possibility of bid rigging and collusion in this case is relatively small.
[0093] The above six types of relationships are important influencing factors for the risk of bid rigging. When bidding auditors conduct bid rigging risk assessment, corresponding measures can be taken based on the relationship types between suppliers to promote fair competition and optimize the bidding management process.
[0094] In an embodiment of the present application, the first-level relationship text and the second prompt word are input into a large language model. The large language model can predict the relationship type between the suppliers based on the relationship type in the second prompt word and the direct relationship between the suppliers in the first-level relationship text, generate corresponding classification basis, and generate the relationship type and classification basis according to a given required format.
[0095] As an example, the second prompt is "Please judge whether the subject meets one or more of the following six types based on the first-level relationship description of the subject [supplier A, supplier B, supplier C], and give an explanation. The six types are: (1) different units with the same person as the head of the unit (legal person, investor of a sole proprietorship, partner who represents a partnership in executing partnership affairs); (2) different units with a controlling stake (holding more than 50% or relative controlling stake); (3) different units with a shareholding (non-controlling stake); (4) different units with a management relationship (a management and being managed relationship between other units that do not have a capital contribution and shareholding relationship, such as a public institution); (5) a director, supervisor, or senior manager of a supplier who also serves as a director, supervisor, or senior manager of another supplier; (6) other weak relationships without direct relationships. Output json format: {"Type 1":"Subject Relationship Explanation 1", "Type 2":"Subject Relationship Explanation 2"], please do not output other invalid information." The first-level relationship text is:
[0096] 【Entity 1: Supplier A
[0097] relation:
[0098] A-> Operator-> Supplier A
[0099] Entity 2: Supplier B
[0100] relation:
[0101] Party A -> Legal Representative, Executive Director and General Manager -> Supplier B
[0102] Entity 3: Supplier C
[0103] relation:
[0104] A -> Operator -> Supplier C]
[0105] Input the above first-level relationship text and prompt words into the large language model and output:
[0106] Type 1: "Supplier A, Supplier B, and Supplier C are all operated by or legally represented by "A." These are different entities whose head (legal person, investor of a sole proprietorship, or partner representing a partnership in executing partnership affairs) is the same person."
[0107] Therefore, the first-level relationship text extracted from the relationship graph can provide the large language model with rich information that is closely related to each supplier, helping it to analyze the key hubs used to connect the relationships between suppliers. The second prompt word can then guide the large language model to identify the relationship types between suppliers and provide corresponding classification basis, thus realizing the automatic identification of relationship types, effectively saving labor costs, and improving the accuracy and efficiency of relationship type identification, thereby accurately combating the risks of bid rigging and collusion in the bidding process and accelerating the audit process of the bidding and procurement department.
[0108] In a possible implementation, after step S103, the method further includes:
[0109] S104: Using robotic process automation (RPA) technology, the predicted relationship type and the corresponding description are written into the table record corresponding to the target bidding activity.
[0110] The table can include records of multiple bidding activities, with separate fields for the predicted relationship type and corresponding description. RPA can then write these predicted relationship types and descriptions into the corresponding fields in the row containing the target bidding activity. RPA then uses this technology to sequentially write the target bidding activity's purchase number, the predicted relationship type between each supplier, and the corresponding descriptions returned by the supplier relationship identification system into the table. The batch identification results are then emailed to auditors, who then distribute them to procurement personnel for subsequent audits.
[0111] As a result, a fully automated process for relationship type identification is achieved, reducing the time cost of manual analysis, identification, and result entry, and effectively improving the efficiency and accuracy of identification.
[0112] Because the relationships between suppliers in the relationship graph are relatively complex, the multimodal model's own understanding ability cannot well identify the direct relationships corresponding to each supplier, and cannot return them in the given required format. Therefore, the multimodal model needs to be fine-tuned for this specific task.
[0113] In a possible implementation, before step S102, the method includes:
[0114] A1: Obtain training data, which includes multiple relationship images and corresponding labels.
[0115] Among them, the label is used to identify the direct relationship corresponding to each supplier in each relationship graph.
[0116] Through manual labeling, the direct relationships between suppliers in multiple images can be obtained, so that the parameters of the model can be adjusted in a supervised manner through training data, so that the model can generate first-level relationship text to describe the direct relationships between suppliers.
[0117] A2: Fine-tune the initial multimodal model based on the training data and the third prompt word to obtain a multimodal model.
[0118] The third prompt word is used to instruct the initial multimodal model to identify the direct relationship corresponding to each supplier in each relationship image and output the corresponding label.
[0119] Inputting multiple relationship pictures and the third prompt word into the multimodal model can guide the initial multimodal model to identify the direct relationship corresponding to the supplier and generate a preliminary description of the direct relationship. Then, based on the difference between the preliminary description and the label, the parameters of the initial multimodal model are fine-tuned until the initial multimodal model can output the corresponding label, so that the fine-tuned multimodal model can be used to identify the relationship type.
[0120] The initial multimodal model can be the Qwen2.5-VL-32B model. When inputting the third prompt word and training data, the input data needs to be converted to the data format required by the model. The sample code for data conversion is as follows:
[0121] {
[0122] "messages": [
[0123] {
[0124] "content": " Please play the role of an enterprise relationship analysis assistant and output the first-level relationships of the entities [Supplier A, Supplier B, Supplier C] respectively.
[0125] "role": "user"
[0126] },
[0127] {
[0128] "content": "Entity 1: Supplier A
[0129] relation:
[0130] Party A -> 100% investment, legal representative, executive director -> Supplier A
[0131] Entity 2: Supplier B
[0132] relation:
[0133] B -> Investor 90%, Legal Representative, Executive Director, General Manager -> Supplier B
[0134] Entity 3: Supplier C
[0135] relation:
[0136] B->Operator->Supplier C
[0137] "
[0138] "role": "assistant"
[0139] }
[0140] ],
[0141] "images": [
[0142] "mllm_demo_data / XX.png" ]
[0144] },
[0145] Here, messages indicates the text portion of a training data entry, content indicates the specific input or output content, and role indicates the current role. If it is user, it indicates that the current content should be user input, and if it is assistant, it indicates that the current content should be the output of the multimodal model. images indicates that this is the image portion of the training data entry, and mllm_demo_data indicates the path to the image.
[0146] In this application, visual fine-tuning tools, such as the LLaMa-Factory framework, can also be used to fine-tune the multimodal model. Based on this framework, the input data and parameter values of the multimodal model can be manually adjusted.
[0147] In one possible implementation, step A2 includes:
[0148] A21: The training data and the third prompt word are used to adjust the parameters of the initial multimodal model through the Lora fine-tuning method to obtain a multimodal model.
[0149] The Lora fine-tuning method can adjust some parameters. Compared with full parameter adjustment, it has obvious advantages in efficiency, high performance, and can quickly adapt to new tasks and data.
[0150] The embodiment of this application adopts the Lora fine-tuning method to fine-tune the parameters of the multimodal model by setting parameters such as learning rate and number of iterations. Figure 9 , which is the eighth schematic diagram of a relationship picture provided in an embodiment of the present application, inputs the relationship picture and the prompt "Please play the role of an enterprise association relationship analysis assistant, and please output the first-level relationships of the subjects [supplier A, supplier B, supplier C] respectively" into the multimodal model before and after fine-tuning.
[0151] The output content is compared as follows. The output before fine-tuning is:
[0152] According to the enterprise relationship diagram in the picture, the following are the first-level relationships between the three entities:
[0153] Supplier A
[0154] Investors:
[0155] Company A (70% investment)
[0156] A (investment 30%, legal representative, executive director)
[0157] Supplier B
[0158] Investors:
[0159] Company A (70% investment)
[0160] A (investment 30%, legal representative, executive director)
[0161] Supplier C
[0162] Investors:
[0163] Company A (70% investment)
[0164] Supplier C (30% investment, legal representative, executive director)
[0165] Summarize:
[0166] Company A is a joint investor of the three suppliers mentioned above, holding 70% of the shares respectively.
[0167] Person A is a joint investor of the above three suppliers, holding 30% of the shares in each supplier, and serves as the legal representative and executive director of each supplier.
[0168] The output after fine-tuning is:
[0169] Entity 1: Supplier A
[0170] relation:
[0171] Party A -> 30% investment, legal representative, executive director -> Supplier A
[0172] Company A -> Invest 70% -> Supplier A
[0173] Entity 2: Supplier B
[0174] relation:
[0175] Party A -> 30% investment, legal representative, executive director -> Supplier B
[0176] Company A -> Invest 70% -> Supplier B
[0177] Entity 3: Supplier C
[0178] relation:
[0179] Party A -> 30% investment, legal representative, executive director -> Supplier C
[0180] Company A -> Invest 70% -> Supplier C
[0181] The Lora fine-tuning method can save computational costs and significantly improve the efficiency of fine-tuning while maintaining high performance. It is more suitable for scenarios where parameters of large models are adjusted, and can quickly adapt to the task of generating first-level relationship text for suppliers.
[0182] Therefore, by training the multimodal model in a supervised manner, the accuracy of identifying relationship types can be significantly improved, thereby enhancing audit accuracy, ensuring the fairness and transparency of the bidding process, and effectively preventing scope bidding and collusion in bidding.
[0183] In order to more clearly describe the supplier relationship type identification method, the following is an explanation based on a specific implementation scenario. Figure 10 As shown, it is a schematic diagram of the overall design of a supplier relationship type identification method provided in an embodiment of the present application.
[0184] The bidding and procurement management system stores information on each supplier in different bidding activities. Through RPA technology, all supplier names in multiple bidding activities can be traversed from the bidding and procurement management system, and then entered into the enterprise relationship query tool to obtain relationship images corresponding to different bidding activities. Each relationship image is used to describe the association relationship between all suppliers of the corresponding bidding activity. The relationship image is then entered into the supplier relationship identification system. The supplier relationship identification system includes a multimodal model for identifying the direct relationship corresponding to each supplier, and a large model for identifying the relationship type. This system can obtain the relationship type and corresponding description of each supplier in different bidding activities, and finally send it to auditors for audit to combat the risks of bid rigging and collusion.
[0185] refer to Figure 11 FIG. 1 is a flowchart of a method for identifying an application supplier relationship type provided in an embodiment of the present application. The specific steps are as follows:
[0186] 1. The purchasing staff enters the purchasing information into the bidding and purchasing management system. This purchasing information may include the purchase number, purchase name, purchase department, purchase time, purchase price, bidding suppliers, and winning suppliers corresponding to the bidding activity.
[0187] 2. PRA obtains procurement information in batches from the bidding and procurement management system. When an audit is required, RPA uses an interface to obtain procurement information in batches from the bidding and procurement management system and saves it as an execl file F.
[0188] 3. Traverse the procurement information and obtain the names of participating suppliers. Use RPA technology to traverse the procurement information file F and obtain the names of participating suppliers for one bidding activity C.
[0189] 4. PRA automatically logs in to the enterprise relationship query tool and enters multiple supplier names.
[0190] 5. Determine whether a relationship image is found. If no relationship image is found, it means that the suppliers participating in the bid are independent of each other and have no association relationship, which complies with the procurement rules and is ignored. If a relationship image is found, the procurement number, image, and supplier name are entered into the supplier relationship identification system. The supplier relationship identification system involves two model services: a multimodal model and a large language model. The multimodal large model is mainly used to identify the direct relationship corresponding to the supplier entity from the image, and the large language model is mainly used to determine the relationship type and provide an explanation based on the direct relationship identified by the business rules and the multimodal large model.
[0191] 6. Perform model training and deployment before using the multimodal model to adapt the multimodal model to the task described in step 5. The fine-tuning process includes the following steps:
[0192] 6.1. Save supplier names and relationship pictures.
[0193] 6.2. Construct a training dataset. The training dataset includes multiple relationship images and manually labeled relationship types between suppliers.
[0194] 6.3. Training the multimodal model. Input the training dataset and the prompt words that indicate the direct relationship corresponding to each supplier in the relationship image for the multimodal model to identify, and fine-tune its parameters so that the fine-tuned multimodal model can output the corresponding relationship type.
[0195] 6.4. Start the multimodal model and the large language model. Select the trained multimodal model file path and start the model using the VLLM framework. This provides a functional service for identifying the first-level relationships between subject companies based on relationship images. Start the large language model using the VLLM framework and provide a functional interface for classifying and interpreting subject relationships.
[0196] 7. Input the relationship images and supplier names into the multimodal model to generate a fixed-format subject relationship description. The fine-tuned multimodal model has good performance and can accurately output the subject relationship description, which is the first-level relationship text that describes the direct relationship between each supplier.
[0197] 8. Input the subject relationship description and supplier name into the large language model, allowing the large language model to determine the relationship type and generate a relationship description. The large language model determines the relationship type between each supplier based on the description of the direct relationship corresponding to each supplier in the subject relationship and generates a relationship description based on the classification criteria.
[0198] 9. Write the purchase number, relationship type, and relationship description into the table. Use RPA technology to write the purchase number, relationship type, and relationship description returned by the supplier relationship identification system into the table.
[0199] 10. Feedback the results to the auditors. The batch recognition results are emailed to the auditors using RPA technology.
[0200] 11. Auditors distribute the procurement staff to carry out the subsequent audit process.
[0201] Therefore, this application has the following advantages:
[0202] 1. A supplier relationship type identification method based on large language model and multimodal model is proposed;
[0203] 2. Combining RPA technology with generative big models enables automated identification of supplier relationship types;
[0204] 3. In the process of supplier relationship type identification, a two-step approach is proposed: using a multimodal model to extract the direct relationship corresponding to the supplier, and using a large language model for classification, judgment, and generation of explanations;
[0205] 4. A method for fine-tuning the general multimodal model is proposed to improve the accuracy of extracting relational feature information in relational images.
[0206] Based on the above embodiment, the present embodiment provides a supplier relationship type identification device, referring to Figure 12 FIG. 1 is a schematic diagram of a device for identifying supplier relationship types provided in an embodiment of the present application. The device 1300 includes an acquisition unit 1301 and an identification unit 1302.
[0207] The acquisition unit 1301 is used to acquire a relationship graph, where the relationship graph is used to identify the relationship between suppliers in the target bidding activity;
[0208] The recognition unit 1302 is configured to input the relationship image and the first prompt word into a multimodal model to obtain a first-level relationship text, wherein the first prompt word is used to instruct the multimodal model to recognize the direct relationship corresponding to each of the suppliers;
[0209] The recognition unit 1302 is also used to input the first-level relationship text and the second prompt word into the large language model to obtain the predicted relationship type and corresponding description between the suppliers. The second prompt word is used to instruct the large language model to classify the association relationship between the suppliers based on the relationship type and generate corresponding classification basis.
[0210] In a possible implementation, the acquiring unit is specifically configured to:
[0211] Use Robotic Process Automation (RPA) technology to traverse all suppliers in the target bidding activity, and input the suppliers into the enterprise relationship query tool to obtain a relationship picture.
[0212] In a possible implementation, the apparatus further includes a fine-tuning unit, configured to:
[0213] Acquire training data, where the training data includes multiple relationship graphs and corresponding labels, where the labels are used to identify the direct relationship corresponding to each supplier in each relationship graph;
[0214] Based on the training data and the third prompt word, the initial multimodal model is fine-tuned to obtain a multimodal model. The third prompt word is used to instruct the initial multimodal model to identify the direct relationship corresponding to each supplier in each relationship picture and output the corresponding label.
[0215] In a possible implementation, the fine-tuning unit is specifically configured to:
[0216] According to the training data and the third prompt word, the parameters of the initial multimodal model are adjusted by the Lora fine-tuning method to obtain the multimodal model.
[0217] In a possible implementation, the relationship type includes:
[0218] Different suppliers with the same person as the person in charge (legal person, investor of a sole proprietorship, partner who executes partnership affairs on behalf of a partnership), different suppliers with a controlling stake (holding more than 50% or relative controlling stake), different suppliers with a shareholding (non-controlling stake), different suppliers with a management relationship (a management-being-managed relationship between other units that do not have a capital contribution or shareholding relationship, such as public institutions), different suppliers with common management personnel (a director, supervisor or senior management of a supplier concurrently serves as a director, supervisor or senior management of another supplier), and different suppliers with no direct weak relationship.
[0219] In a possible implementation, the apparatus further includes an input unit, which is configured to:
[0220] Through the RPA technology, the predicted relationship type and the corresponding description are written into the table record corresponding to the target bidding activity.
[0221] Based on the above embodiments, an embodiment of the present application provides a computer device, which includes a processor and a memory:
[0222] The memory is used to store computer programs;
[0223] The processor is configured to execute the supplier relationship type identification method according to the computer program.
[0224] Based on the above embodiments, an embodiment of the present application provides a computer-readable storage medium, which is used to store a computer program. When the computer program is executed by a computer device, it implements the above-mentioned supplier relationship type identification method.
[0225] Based on the above embodiments, an embodiment of the present application provides a computer program product including a computer program, which, when executed on a computer device, enables the computer device to execute the above supplier relationship type identification method.
[0226] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.
[0227] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for identifying supplier relationship types, characterized in that: The method comprises: Obtaining a relationship graph, wherein the relationship graph is used to identify the association relationship between suppliers in the target bidding activity; Inputting the relationship image and the first prompt word into a multimodal model to obtain a first-level relationship text, wherein the first prompt word is used to instruct the multimodal model to identify the direct relationships corresponding to the respective suppliers; The first-level relationship text and the second prompt word are input into the large language model to obtain the predicted relationship type and corresponding description between the suppliers. The second prompt word is used to instruct the large language model to classify the association relationship between the suppliers based on the relationship type and generate corresponding classification basis.
2. The method according to claim 1, characterized in that The obtaining of the relationship picture includes: Use Robotic Process Automation (RPA) technology to traverse all suppliers in the target bidding activity, and input the suppliers into the enterprise relationship query tool to obtain a relationship picture.
3. The method according to claim 1, characterized in that Before inputting the relationship picture and the first prompt word into the multimodal model to obtain the first-level relationship text, the method further includes: Acquire training data, where the training data includes multiple relationship graphs and corresponding labels, where the labels are used to identify the direct relationship corresponding to each supplier in each relationship graph; Based on the training data and the third prompt word, the initial multimodal model is fine-tuned to obtain a multimodal model. The third prompt word is used to instruct the initial multimodal model to identify the direct relationship corresponding to each supplier in each relationship picture and output the corresponding label.
4. The method according to claim 3, characterized in that Fine-tuning the initial multimodal model based on the training data and the third prompt word to obtain the multimodal model includes: According to the training data and the third prompt word, the parameters of the initial multimodal model are adjusted by the Lora fine-tuning method to obtain the multimodal model.
5. The method according to claim 1, characterized in that The relationship types include: Different suppliers with the same person as the person in charge (legal person, investor of a sole proprietorship, partner who executes partnership affairs on behalf of a partnership), different suppliers with a controlling stake (holding more than 50% or relative controlling stake), different suppliers with a shareholding (non-controlling stake), different suppliers with a management relationship (a management-being-managed relationship between other units that do not have a capital contribution or shareholding relationship, such as public institutions), different suppliers with common management personnel (a director, supervisor or senior management of a supplier concurrently serves as a director, supervisor or senior management of another supplier), and different suppliers with no direct weak relationship.
6. The method according to claim 2, characterized in that After inputting the first-level relationship text and the second prompt word into the large language model to obtain the predicted relationship types and corresponding descriptions between the suppliers, the method further includes: Through the RPA technology, the predicted relationship type and the corresponding description are written into the table record corresponding to the target bidding activity.
7. A device for identifying supplier relationship types, characterized in that: The device includes an acquisition unit and an identification unit: The acquisition unit is used to acquire a relationship picture, wherein the relationship picture is used to identify the relationship between the suppliers in the target bidding activity; The recognition unit is configured to input the relationship image and the first prompt word into a multimodal model to obtain a first-level relationship text, wherein the first prompt word is used to instruct the multimodal model to recognize the direct relationship corresponding to each of the suppliers; The recognition unit is further used to input the first-level relationship text and the second prompt word into the large language model to obtain the predicted relationship type and corresponding description between the suppliers. The second prompt word is used to instruct the large language model to classify the association relationship between the suppliers based on the relationship type and generate corresponding classification basis.
8. A computer device, characterized in that: The computer device includes a processor and a memory: The memory is used to store computer programs; The processor is configured to execute the method according to any one of claims 1 to 6 according to the computer program.
9. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store a computer program, and when the computer program is executed by a computer device, the computer program implements the method according to any one of claims 1 to 6.
10. A computer program product comprising a computer program, characterized in that When the method is executed on a computer device, the computer device is enabled to execute the method according to any one of claims 1 to 6.