Supplier risk assessment method based on large language model

Abstract

The invention discloses a supplier risk assessment method based on a large language model. The supplier risk assessment method comprises the steps of establishing a database, preprocessing global data, preprocessing local data, constructing a risk assessment model, outputting a decision and a report and correcting the database. Real-time quantitative evaluation and dynamic early warning of supplier risks are achieved, risk grade division and purchase suggestions are automatically generated, a traditional manual regular analysis mode is replaced, the evaluation efficiency is remarkably improved, a global database and a local database are physically separated, sensitive financial information participates in model calculation only in a desensitized score form, and the evaluation efficiency is improved. The demand of manually looking up original sensitive files is avoided, and the protection of commercial secrets is enhanced.

Description

Technical Field

[0001] This invention relates to the technical field of supplier risk assessment, and in particular to a supplier risk assessment method based on a large language model. Background Technology

[0002] In supply chain management, supplier risk assessment is a crucial step in ensuring supply chain stability and reducing procurement risks. Traditional supplier risk assessment methods primarily rely on manual review. However, manual assessment mainly depends on historical data and lacks real-time updates. Supplier financial data, order records, and public opinion information are often scattered across independent systems such as ERP, CRM, and news platforms, resulting in data fragmentation. When a supplier experiences a sudden financial crisis, compliance penalties, or public opinion events, it is difficult to respond promptly. Therefore, traditional manual review methods involve massive amounts of data, insufficient data integration, and are prone to assessment blind spots and subjective biases. This leads to supplier risk assessments deviating from the true situation, and companies may be unable to avoid high-risk suppliers in time, resulting in production disruptions, product delivery delays, and significant economic losses. Summary of the Invention

[0003] To overcome the shortcomings of the prior art, the technical solution adopted in this invention is: a supplier risk assessment method based on a large language model, comprising:

[0004] S1. Establish a database: This includes a global database and a local database. The global database contains basic information about suppliers, financial data, and historical order data. The local database contains supplier-related risks, public opinion news, and corporate credit.

[0005] S2.1 Global Data Preprocessing: Based on basic information and financial data, the supplier level is divided using the One-hot encoding method. Based on historical order data, the number of days of delayed orders is binned into risk levels, and then standardized into a global static score.

[0006] S2.2 Local data preprocessing: Extract key events of the corresponding suppliers based on the local database, associate key events with risk types, determine the text risk tendency of key events, and then standardize them into local dynamic risk values;

[0007] S3. Architecture Risk Assessment Model: Construct a multi-dimensional risk assessment model that integrates global static scoring and local dynamic risk values, iteratively optimize the model, and realize the quantification and dynamic early warning of supplier risks.

[0008] S4. Output Decisions and Reports: Classify supplier risk levels, automatically generate actionable procurement recommendations, and provide explanatory analysis reports;

[0009] S5. Correct Database: Sales staff promptly repair or adjust parameters based on the content of the explanatory analysis report. After the administrator reviews and approves, the system jumps to S1 and regenerates the decision report. The corrected records will be input into the database as new samples, establishing a closed-loop feedback system.

[0010] The present invention is further configured such that the basic information includes supplier qualifications, production capacity, geographical location and years of cooperation; the financial data includes the previous year's output value, asset-liability ratio and credit rating; and the historical order data includes on-time delivery rate, product quality pass rate and order fulfillment rate.

[0011] The present invention is further configured such that the risk types include operational risk, financial risk, compliance risk, public opinion risk, and environmental risk. Operational risk is recorded and quantified based on the supplier's capacity utilization rate and batch product qualification rate. Financial risk is recorded and quantified based on the supplier's credit report, annual report data, and cash flow data. Compliance risk is recorded and quantified based on the supplier's labor dispute cases and safety and environmental penalties. Public opinion risk is recorded and quantified based on sensitive items such as media reports, changes in senior management equity, and changes in legal representatives of the supplier. Environmental risk is recorded and quantified based on force majeure factors such as natural disasters and abnormal weather in the supplier's location.

[0012] The present invention is further configured such that if there are missing values ​​in the financial risk, the item is recorded according to the corresponding highest risk value and marked as having a downgraded confidence level.

[0013] The present invention is further configured such that the public opinion risk is used to perform entity identification on public opinion news and corporate credit text, extract supplier names, event types and scope of impact, output negative / neutral / positive labels, and assign risk values ​​according to label types.

[0014] The present invention is further configured such that the environmental risks are classified into meteorological disasters, geological disasters and epidemic disasters according to the officially announced types of natural disasters, and outputs the corresponding risk values.

[0015] The present invention is further configured such that the explanatory analysis report is associated with risk events through a knowledge graph and generates a language description that includes a combination of risk root causes, similar historical cases, and recommended measures.

[0016] The present invention is further configured such that, in step S5, when the administrator reviews the data, the basis for the correction must be marked, and a correction history record containing the salesperson's and administrator's IDs must be generated.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1. By integrating global static scoring with local dynamic risk values, it enables real-time quantitative assessment and dynamic early warning of supplier risks, automatically generates risk level classifications and procurement recommendations, replacing the traditional manual periodic analysis mode and significantly improving assessment efficiency.

[0019] 2. The global database and local databases are physically separated, and sensitive financial information is only used in the model calculation in the form of anonymized scores, avoiding the need for manual review of original sensitive documents and strengthening the protection of trade secrets. Attached Figure Description

[0020] Figure 1 This is a flowchart of the supplier risk assessment method based on a large language model according to the present invention; Detailed Implementation

[0021] Embodiments of the present invention will now be described with reference to the accompanying drawings. In this process, to ensure clarity and convenience, we may exaggerate the width of lines or the size of constituent elements in the drawings.

[0022] Furthermore, the terms used below are defined based on the functions of this invention and may vary depending on the intentions or conventions of the user or operator. Therefore, these terms are defined based on the entire contents of this specification.

[0023] like Figure 1 As shown, a supplier risk assessment method based on a large language model includes:

[0024] S1. Establish a database: This includes a global database and a local database. The global database contains basic information about suppliers, financial data, and historical order data. The local database contains supplier-related risks, public opinion news, and corporate credit.

[0025] S2.1 Global Data Preprocessing: Based on basic information and financial data, the supplier level is divided using the One-hot encoding method. Based on historical order data, the number of days of delayed orders is binned into risk levels, and then standardized into a global static score.

[0026] S2.2 Local data preprocessing: Extract key events of the corresponding suppliers based on the local database, associate key events with risk types, determine the text risk tendency of key events, and then standardize them into local dynamic risk values;

[0027] S3, Architecture Risk Assessment Model: Construct a multi-dimensional risk assessment model that integrates global static scoring and local dynamic risk values, iteratively optimize the model, realize the quantification and dynamic early warning of supplier risks, and use the model for text risk tendency analysis and natural language generation of explanatory reports in local data preprocessing.

[0028] S4. Output Decisions and Reports: Classify supplier risk levels, automatically generate actionable procurement recommendations, and provide explanatory analysis reports;

[0029] S5. Correct Database: Sales staff promptly repair or adjust parameters based on the content of the explanatory analysis report. After the administrator reviews and approves, the system jumps to S1 and regenerates the decision report. The corrected records will be input into the database as new samples, establishing a closed-loop feedback system.

[0030] In this embodiment, basic information includes supplier qualifications, production capacity, geographical location, and years of cooperation; financial data includes the previous year's output value, asset-liability ratio, and credit rating; historical order data includes on-time delivery rate, product quality pass rate, and order fulfillment rate. Global data is obtained from enterprise forms or queried from internal ERP systems, financial systems, and enterprise information query systems. Suppliers are assigned corresponding numbers based on global data and categorized into Tier 1, Tier 2, and Tier 3 suppliers, with risk coefficients increasing progressively. For example, the risk coefficient for Tier 1 suppliers is 0.8, for Tier 2 suppliers it is 1, and for Tier 3 suppliers it is 1.2. Suppliers are sorted by historical delivery delay days and categorized according to contract terms as compliant with contract terms, slightly delayed, and severely delayed, with the risk value increasing progressively with the number of delay days. For example, the risk value for compliant delivery is 0, the risk threshold for slightly delayed delivery is 1-3, and the risk threshold for severely delayed delivery is ≥4.

[0031] In this embodiment, the risk types include operational risk, financial risk, compliance risk, public opinion risk, and environmental risk, among which:

[0032] Operational risks are recorded and quantified based on the supplier's capacity utilization rate and batch product qualification rate. The main data for operational risks come from production reports, quality inspection reports and order data. The comprehensive evaluation of the supplier's equipment failure frequency, capacity utilization rate, personnel turnover rate and product qualification rate avoids the risk of supply chain disruption caused by the supplier's internal production, delivery or quality management failures.

[0033] Financial risks are recorded and quantified based on the supplier's credit report, annual report data and cash flow data. The main data for financial risks come from financial statements, credit reports, annual reports, tax data and bank statements. If there are missing values ​​in the financial risk data, the item is recorded according to the corresponding highest risk value and marked with a credit rating downgrade, thereby comprehensively evaluating the supplier's financial situation and avoiding the risk of failure to perform due to deteriorating financial situation.

[0034] Compliance risks are recorded and quantified based on labor disputes and safety and environmental penalties imposed on suppliers. The main data on compliance risks come from China Judgments Online, Credit China, and the Ministry of Ecology and Environment's public information platform. Tools such as NLP legal document analysis or keyword crawling are used to comprehensively evaluate the legality of suppliers and avoid operational legality risks caused by violations of laws, regulations, or industry standards.

[0035] The main data for public opinion risk comes from mainstream financial media, relevant industry forums, and enterprise information query systems. Public opinion risk is used to identify entities in public opinion news and enterprise credit texts, extract supplier names, event types, and scope of impact. Public opinion risk mainly targets supplier fire incidents, strike incidents, product counterfeiting, industry scandals, etc., and outputs negative / neutral / positive labels. The risk value is assigned according to the label type. Negative and neutral labels increase the corresponding risk value, while positive labels decrease the corresponding risk value, so as to avoid the risk of damage to the client's reputation due to the fermentation of negative public opinion about suppliers.

[0036] In this embodiment, the risk types include operational risk, financial risk, compliance risk, public opinion risk, and environmental risk, among which: Operational risks are recorded and quantified based on the supplier's capacity utilization rate and batch product qualification rate. The main data for operational risks come from production reports, quality inspection reports and order data. The comprehensive evaluation of the supplier's equipment failure frequency, capacity utilization rate, personnel turnover rate and product qualification rate avoids the risk of supply chain disruption caused by the supplier's internal production, delivery or quality management failures. Financial risks are recorded and quantified based on the supplier's credit report, annual report data and cash flow data. The main data for financial risks come from financial statements, credit reports, annual reports, tax data and bank statements. If there are missing values ​​in the financial risk data, the item is recorded according to the corresponding highest risk value and marked with a credit rating downgrade, thereby comprehensively evaluating the supplier's financial situation and avoiding the risk of failure to perform due to deteriorating financial situation. Compliance risks are recorded and quantified based on labor disputes and safety and environmental penalties imposed on suppliers. The main data on compliance risks come from China Judgments Online, Credit China, and the Ministry of Ecology and Environment's public information platform. Tools such as NLP legal document analysis or keyword crawling are used to comprehensively evaluate the legality of suppliers and avoid operational legality risks caused by violations of laws, regulations, or industry standards. The main data for public opinion risk comes from mainstream financial media, relevant industry forums, and enterprise information query systems. Public opinion risk is used to identify entities in public opinion news and enterprise credit texts, extract supplier names, event types, and scope of impact. Public opinion risk mainly targets supplier fire incidents, strike incidents, product counterfeiting, industry scandals, etc., and outputs negative / neutral / positive labels. The risk value is assigned according to the label type. Negative and neutral labels increase the corresponding risk value, while positive labels decrease the corresponding risk value, so as to avoid the risk of damage to the client's reputation due to the fermentation of negative public opinion about suppliers. Environmental risks are recorded and quantified based on force majeure factors such as natural disasters and abnormal weather in the supplier's location. Environmental risks are categorized into meteorological disasters, geological disasters, and epidemic disasters according to officially published natural disaster types, and corresponding risk values ​​are output. Environmental risk data mainly comes from meteorological bureaus, earthquake data, typhoon paths, etc., to avoid the risk of untimely supply due to force majeure factors. Furthermore, delayed delivery caused by force majeure factors by suppliers is not included in the global database, and the risk value is reset to zero from the date the disaster is eliminated.

[0037] In this embodiment, the risk score is the sum of the global static score and the local dynamic risk value. If the risk score is less than 10, the supplier is judged as low-risk and the normal procurement process continues. If the risk score is between 10 and 30, the supplier is judged as medium-risk and the procurement process needs to be confirmed by the business manager. If the risk score is greater than 30, the supplier is judged as high-risk and the procurement process needs to be confirmed by the supply chain director.

[0038] In this embodiment, the explanatory analysis report is linked to risk events through a knowledge graph and generates a language description that includes the root cause of the risk, similar historical cases, and a combination of recommended measures. When the administrator reviews the report in S5, the basis for correction must be marked and a correction history record containing the salesperson's and administrator's IDs is generated. An open manual review mechanism is provided to provide a feedback process for issues such as supplier complaints, outdated real-time text data, and incorrect model parameters. This provides a reference data basis for training the risk assessment model and effectively reduces the reliance on manual analysis.

[0039] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A supplier risk assessment method based on a large language model, characterized in that, include: S1. Establish a database: This includes a global database and a local database. The global database contains basic information about suppliers, financial data, and historical order data. The local database contains supplier-related risks, public opinion news, and corporate credit. S2.1 Global Data Preprocessing: Based on basic information and financial data, the supplier level is divided using the One-hot encoding method. Based on historical order data, the number of days of delayed orders is binned into risk levels, and then standardized into a global static score. S2.2 Local data preprocessing: Extract key events of the corresponding suppliers based on the local database, associate key events with risk types, determine the text risk tendency of key events, and then standardize them into local dynamic risk values; S3. Architecture Risk Assessment Model: Construct a multi-dimensional risk assessment model that integrates global static scoring and local dynamic risk values, iteratively optimize the model, and realize the quantification and dynamic early warning of supplier risks. S4. Output Decisions and Reports: Classify supplier risk levels, automatically generate actionable procurement recommendations, and provide explanatory analysis reports; S5. Correct Database: Sales staff promptly repair or adjust parameters based on the content of the explanatory analysis report. After the administrator reviews and approves, the system jumps to S1 and regenerates the decision report. The corrected records will be input into the database as new samples, establishing a closed-loop feedback system.

2. The supplier risk assessment method based on a large language model according to claim 1, characterized in that: The basic information includes supplier qualifications, production capacity, geographical location, and years of cooperation; the financial data includes the previous year's output value, asset-liability ratio, and credit rating; and the historical order data includes on-time delivery rate, product quality pass rate, and order fulfillment rate.

3. The supplier risk assessment method based on a large language model according to claim 2, characterized in that: The risk types include operational risk, financial risk, compliance risk, public opinion risk, and environmental risk. Operational risk is recorded and quantified based on the supplier's capacity utilization rate and batch product qualification rate. Financial risk is recorded and quantified based on the supplier's credit report, annual report data, and cash flow data. Compliance risk is recorded and quantified based on the supplier's labor dispute cases and safety and environmental penalties. Public opinion risk is recorded and quantified based on sensitive items such as media reports, changes in senior management equity, and changes in legal representatives. Environmental risk is recorded and quantified based on force majeure factors such as natural disasters and abnormal weather in the supplier's location.

4. The supplier risk assessment method based on a large language model according to claim 3, characterized in that: If any of the financial risks are missing values, the item will be recorded according to the highest corresponding risk value and marked as having a downgraded reliability.

5. The supplier risk assessment method based on a large language model according to claim 3, characterized in that: The aforementioned public opinion risk is used for entity identification in public opinion news and corporate credit texts, extracting supplier names, event types and scope of impact, outputting negative / neutral / positive labels, and assigning risk values ​​based on label types.

6. The supplier risk assessment method based on a large language model according to claim 3, characterized in that: The environmental risks are categorized into meteorological disasters, geological disasters, and epidemic disasters based on the officially announced types of natural disasters, and corresponding risk values ​​are output.

7. The supplier risk assessment method based on a large language model according to claim 1, characterized in that: The explanatory analysis report uses a knowledge graph to link risk events and generates a linguistic description that includes the root causes of the risk, similar historical cases, and a combination of recommended measures.

8. The supplier risk assessment method based on a large language model according to claim 1, characterized in that: In S5, when the administrator reviews the data, he / she needs to mark the basis for the correction and generate a correction history that includes the salesperson's and administrator's IDs.

Images