A general industry chain security risk identification method, system, device and storage medium based on a large language model

Abstract

The application discloses a general industry chain security risk identification method, system and device based on a large language model, and a storage medium, and comprises the following steps: inputting initial news text data into an industry chain news classification model, and extracting industry chain related news text; merging the industry chain related news text and extension information to obtain target news text; extracting a theme from the target news text by using a large language model to obtain theme classification information; searching for background reference content matched with the target news text in a preset knowledge base; inputting the target news text, the theme classification information and the background reference content into the large language model, and outputting a qualitative evaluation result of the target news text in a preset evaluation dimension; counting the qualitative evaluation result according to a preset time window, and calculating an industry chain security risk evaluation index; and performing weighted calculation on the industry chain security risk evaluation index to generate an industry chain security risk index.

Description

Technical Field

[0001] This invention belongs to the field of supply chain security risk, and relates to a general supply chain security risk identification method, system, device and storage medium based on a large language model. Background Technology

[0002] Current technologies for identifying and assessing security risks primarily rely on two analytical paradigms: qualitative and quantitative. Qualitative methods, such as expert interviews, case studies, and scenario simulations, heavily depend on the experience and insights of domain experts and demonstrate unique advantages in identifying risks arising from macroeconomic non-economic factors (such as policy shifts and geopolitical games). However, these methods are highly subjective, susceptible to expert cognitive biases, and difficult to standardize, replicate, and validate, leading to limitations such as high uncertainty and delayed conclusions.

[0003] In contrast, quantitative methods—including input-output models, network analysis, and machine learning algorithms—quantify the transmission of risks within the industrial chain through modeling. However, these models face multiple technical bottlenecks: First, they heavily rely on structured data, with data compilation typically incurring a significant time lag of 3 to 5 years, making it difficult to meet the high timeliness requirements of risk assessment during emergencies; second, traditional models struggle to process massive amounts of unstructured text such as news and policy reports, leading to the easy overlooking of weak risk signals hidden within long texts; finally, existing AI-based risk identification research often simplifies risk into a binary classification problem (i.e., risk or no risk), failing to quantify the impact gradient of risk or depict the dynamic evolution of risk over time. Therefore, how to achieve real-time monitoring and multi-dimensional quantitative assessment of industrial chain security risks in complex and uncertain environments has become a critical issue that urgently needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of strong subjectivity, delayed conclusions, difficulty in handling unstructured text, and single risk quantification dimension in the existing industrial chain security risk identification, and to provide a general industrial chain security risk identification method, system, device and storage medium based on a large language model.

[0005] To achieve the above objectives, the present invention employs the following technical solution: A general supply chain security risk identification method based on a large language model includes the following process: Obtain initial news text data, input the initial news text data into the industry chain news classification model, and extract industry chain related news text; Obtain extended information corresponding to news texts related to the industrial chain, and merge the news texts related to the industrial chain and the extended information to obtain the target news text; A large language model is used to extract themes from the target news text, thereby obtaining the theme classification information corresponding to the target news text. Retrieve background reference content that matches the target news text from a pre-defined knowledge base; Input the target news text, topic classification information, and background reference content into the large language model, and output the qualitative evaluation results of the target news text on the preset evaluation dimensions. Based on the qualitative assessment results collected within a preset time window, the industrial chain security risk assessment indicators are calculated. The supply chain security risk assessment indicators are weighted and calculated to generate the supply chain security risk index.

[0006] Optionally, the training steps for the industry chain news classification model include: The sample news text and the target industry chain description information are processed by feature vectorization and similarity is calculated. Texts whose similarity meets the set threshold condition are extracted as the first associated text. The sample news text is input into a lightweight language model for inference and recognition, and the text that is determined to be relevant is extracted as the second related text; Merge the first and second associated texts to obtain a fine-tuned dataset after manual review and labeling. Low-rank adaptive techniques are used to fine-tune the parameters of a pre-trained large language model, and the incremental parameter matrix of the pre-trained large language model is updated to train a news classification model for the industry chain.

[0007] Optionally, the steps of obtaining extended information corresponding to the industry chain-related news text and merging the industry chain-related news text and extended information to obtain the target news text include: Search online for relevant policy documents and online reports related to news articles related to the industrial chain; By combining relevant policy documents, online reports, and news texts related to the industry chain, the target news text is generated.

[0008] Optionally, the steps of using a large language model to extract topics from the target news text and obtain the topic classification information corresponding to the target news text include: Construct a multi-level prompt word architecture based on few-shot learning; Input the multi-level prompt word architecture into the large language model to extract the top-level and second-level topics of the target news text; The text embedding model is used to map the top-level topics and second-level topics into topic vectors; the cosine similarity of each topic vector is calculated and iterative clustering analysis is performed to identify topic clusters whose similarity meets the preset similarity threshold; The standard topic names in the topic clusters are selected for deduplication and merging to generate topic classification information corresponding to the target news text.

[0009] Optionally, the step of retrieving background reference content matching the target news text from a preset knowledge base includes: Utilize a multilingual text embedding model to convert target news text into dense retrieval vectors; In the risk assessment case library, which serves as a pre-defined knowledge base, the vector similarity of dense retrieval vectors is calculated, and typical historical events that meet the vector similarity requirements are extracted as logical reference cases. In the knowledge base of the industrial chain nodes, which serves as a pre-set knowledge base, the vector similarity of dense retrieval vectors is calculated, and the corresponding industrial chain structural features, regional distribution, and subdivision scale data are extracted as static background information of the industrial chain. By integrating logical reference cases with static background information of the industry chain, background reference content is generated.

[0010] Optionally, the preset evaluation dimensions include certainty, influence, and resilience. The certainty dimension includes historical frequency indicators, trigger clarity indicators, and precedent indicators; The influence dimension includes indicators such as the scope of influence, the severity of keywords, the speed of dissemination, and the authority of the information source. The resilience dimension includes reversibility indicators and decision-making window indicators.

[0011] Optionally, the steps of calculating the supply chain security risk assessment indicators by statistically analyzing the qualitative assessment results according to a preset time window, and then weighting these indicators to generate the supply chain security risk index include: By using indicator functions to statistically analyze the frequency of occurrence of qualitative assessment results under preset assessment dimensions within a preset time window, a three-level indicator vector for supply chain security risks is constructed. Obtain the multi-level index weight matrix determined based on the analytic hierarchy process; By using a multi-level indicator weight matrix, the three-level indicator vectors of industrial chain security risk are weighted and aggregated layer by layer to obtain the industrial chain competitiveness risk indicator, industrial chain resilience risk indicator, and industrial chain control risk indicator, which are then established as industrial chain security risk assessment indicators. The supply chain security risk assessment indicators are multiplied by their corresponding global weights to generate the supply chain security risk index.

[0012] A general supply chain security risk identification system based on a large language model includes: The related news text extraction module is used to obtain initial news text data, input the initial news text data into the industry chain news classification model, and extract industry chain related news text. The target news text acquisition module is used to acquire extended information corresponding to industry chain-related news texts and merge industry chain-related news texts and extended information to obtain target news texts. The topic classification information acquisition module is used to extract topics from target news texts using a large language model, and obtain the topic classification information corresponding to the target news texts. The retrieval module is used to retrieve background reference content that matches the target news text from a preset knowledge base; The qualitative evaluation module is used to input the target news text, topic classification information and background reference content into the large language model, and output the qualitative evaluation results of the target news text on the preset evaluation dimensions. The supply chain security risk assessment indicator calculation module is used to calculate the supply chain security risk assessment indicators by statistically analyzing the qualitative assessment results according to a preset time window. The supply chain security risk index generation module is used to perform weighted calculations on supply chain security risk assessment indicators to generate the supply chain security risk index.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the general supply chain security risk identification method based on a large language model.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the general supply chain security risk identification method based on a large language model.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention first employs a supply chain classification model trained using LoRA parameter efficient fine-tuning technology. By constructing a vertical domain feature instruction set, it overcomes the limitations of traditional NLP models in terms of insufficient semantic understanding depth and length truncation for long texts. Combined with a context enhancement mechanism, it achieves accurate capture of weak risk signals and background information completion in massive amounts of news. Secondly, the core component adopts a decoupled architecture of Retrieval Augmentation (RAG), separating the general logical reasoning capabilities of the large language model from the dynamically updated external professional knowledge base (historical risk cases and supply chain maps). This not only avoids the high computational cost and loss of general capabilities caused by model fine-tuning, but also effectively suppresses the illusion of the large language model by utilizing real-time retrieval of similar cases and supply chain structure background, thereby significantly improving the model's generalization ability to identify emerging and cross-scenario risks. Finally, by constructing a general quantitative indicator system covering three dimensions of competitiveness, resilience, and control, and using indicator functions and the analytic hierarchy process (AHP) to achieve scientific calculation of dynamic risk indices from qualitative and discrete industry chain news texts, this approach fundamentally overcomes the high subjectivity of traditional human expert assessments and the limitations of traditional quantitative models in handling news information. Ultimately, it provides an objective, universal, and highly reusable quantitative assessment system for industry chain security risks. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the general supply chain security risk identification method based on a large language model according to the present invention. Figure 2 This is a schematic diagram of the human-computer collaborative annotation method of the present invention, which combines dual-track parallelism (vector similarity + AI annotation) and expert review. Figure 3 This is a diagram illustrating the supply chain information filtering process of the present invention. Figure 4 This is a schematic diagram of the news text filtering and enhancement process based on the fine-tuning model of the present invention; Figure 5 This is a schematic diagram of the topic modeling and merging process based on prompt word engineering according to the present invention; Figure 6 This is a schematic diagram of the quantitative layer principle for supply chain security risk assessment in this invention. Figure 7 This is a trend diagram showing the overall risk changes in the chemical industry chain according to the present invention; Figure 8 is a comparison of the safety analysis results of ablation experiments using different modules of the present invention; Figure 8a This is a comparison chart of the safety risks in the chemical industry chain for the complete LIRAN assessment of the present invention; Figure 8b This is a comparison chart of safety risks in the chemical industry chain after the removal of module three of the present invention; Figure 8cA comparative diagram of safety risks in the chemical industry chain for assessing the removal of the RAG module of this invention; Figure 8d This invention provides a comparative chart of safety risks in the chemical industry chain for the removal of the subject extraction module. Figure 8e This invention provides a comprehensive diagram for assessing the safety risks of the chemical industry chain using ablation experiments. Figure 8f A three-dimensional comparative diagram of the ablation experiment assessment of the safety risks in the chemical industry chain in this invention; Figure 9 This is a comparison chart of the safety risks in the chemical industry chain and the trend of the chemical price index, as presented in this invention. Figure 10 This is a comparison chart of the fitting trends of safety risks and price indices in the chemical industry chain according to the present invention; Figure 11 This is a schematic diagram illustrating the relationship between chemical industry chain risks and industry chain nodes in this invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Existing technical solutions have significant technical bottlenecks in practical applications, limiting the accuracy of identifying security risks in the industrial chain. The methods are highly subjective, susceptible to cognitive biases, and difficult to standardize, replicate, and verify: Experts from different fields often have significant differences in their assessment conclusions on the same supply chain risk scenario, lacking objective and unified measurement standards. This makes it difficult for companies to form a consistent decision-making basis when integrating multi-dimensional risk information. Moreover, such methods cannot be standardized and reused through fixed processes or algorithms. When facing new supply chain risks (such as raw material supply fluctuations caused by extreme weather events), it is necessary to reassemble an expert team to conduct analysis, which is not only time-consuming and labor-intensive, but may also lead to missed opportunities for optimal risk intervention due to response delays, further amplifying the impact of risks on the stability of the supply chain.

[0020] Training data lag: Existing quantitative models rely heavily on historical statistical data for training. The entire process of data collection, cleaning, integration, and storage inevitably involves time delays, making it difficult to reflect real-time dynamic changes in the supply chain. For example, when a core raw material supplier in the supply chain temporarily suspends operations due to a public health emergency or when international trade policies undergo urgent adjustments, the risk assessment results output by the model based on historical data from weeks or even months ago will fail to accurately capture the current risk status, causing companies to miss the golden window for risk warnings. Furthermore, excessively long data update cycles make it difficult for models to cover emerging risk types such as supply chain transparency risks brought about by blockchain technology applications or decision-making bias risks caused by vulnerabilities in artificial intelligence algorithms. This further weakens the timeliness and foresight of risk identification, failing to meet the real-time and dynamic requirements of supply chain security management.

[0021] Insufficient generalization ability: Existing quantitative models are mostly designed for specific industry chain scenarios or risk types, and model parameters and feature engineering are highly dependent on the historical data distribution of specific fields. When the industry chain structure changes (such as the integration of emerging technologies or the restructuring of supply chain nodes) or the risk transmission path changes, the model struggles to quickly adapt to the new input data distribution, leading to a significant drop in the accuracy of risk identification across scenarios and industries. For example, a risk identification model built for the electronics and information industry chain cannot effectively extract the risk characteristics unique to the biopharmaceutical field (such as monopolies in the supply of key raw materials and the risk of clinical trial failure) when applied to the biopharmaceutical industry chain due to significant differences in the core links, risk triggering factors, and transmission mechanisms of the two types of industry chains. This results in incorrect risk assessment results. Furthermore, overfitting further exacerbates the lack of generalization ability: during the training phase, the model overfits to noise and local features in historical data instead of capturing the general rules of risk identification. This results in the inability to make accurate judgments when faced with novel and unseen risk cases, severely restricting the applicability and scalability of the model in complex and ever-changing industry chain environments.

[0022] Inability to effectively integrate unstructured information: Existing traditional quantitative models mainly rely on structured data and cannot effectively integrate potential risk signals from unstructured information such as news, policy reports, and industry information.

[0023] In conclusion, a single paradigm is insufficient to meet the need for comprehensive and objective identification of supply chain security risks in complex environments. Therefore, combining the flexibility and foresight of qualitative analysis with the precision and real-time nature of quantitative analysis has become a key path to improving risk identification capabilities. This integration is not a simple superposition, but rather the construction of an innovative framework that absorbs the profound insights of qualitative methods and incorporates the objectivity of quantitative methods.

[0024] This invention aims to address the problems of strong subjectivity, significant lag, and weak generalization ability in existing supply chain security risk identification by introducing an innovative model method that combines real-world news data with qualitative experience, thereby achieving accurate assessment of supply chain security risks. Its purpose is: Addressing the issue of high subjectivity: By introducing large text models, over-reliance on individual expert experience is reduced, providing more objective and consistent risk assessment criteria and establishing a reusable standardized analysis process. Simultaneously, leveraging the natural language processing capabilities of large text models, risk elements, relationships, and evolutionary trends within unstructured text data are deeply mined, enriching the information dimensions of risk identification and improving the comprehensiveness and accuracy of risk assessment.

[0025] To address the issues of lagging data updates and poor identification results: By introducing real-time or near-real-time news and policy documents with higher update frequencies, and processing them using a large text model, potential risk signals in the dynamic changes of the industrial chain can be captured in a timely manner, resulting in a better identification effect of industrial chain security risks.

[0026] This invention addresses the weakness of generalization capabilities in previous quantitative models: Existing quantitative methods for assessing supply chain security risks largely rely on data from network nodes within the supply chain to identify vulnerable nodes or links, or use historical national sector input-output data to decompose and measure the structural risks of the studied supply chain. These models suffer from weak generalization capabilities due to data lag and rapid changes in supply chain-related data, failing to provide timely and accurate assessments of the security risks of the targeted supply chain. This invention innovatively proposes a method for structured analysis of news and policy data using a large-scale model, while simultaneously constructing a key indicator system to assess supply chain security risks. This enables accurate and timely assessment of the issue, ensuring precise risk values ​​are obtained for risk prevention and strategy formulation within the targeted supply chain. This contribution not only improves the accuracy and timeliness of supply chain security risk identification but also provides reliable reference support for practical strategy development.

[0027] like Figure 1As shown in the figure, this embodiment describes a general supply chain security risk identification method based on a large language model. The entire process starts with news data input, goes through classification and expansion, theme extraction and merging, and risk assessment, and finally obtains a supply chain security risk index calculated with a set time window and outputs corresponding prevention suggestions.

[0028] The method will be explained in detail below: S1, Industry Chain News Text Classification and Expansion.

[0029] The purpose of this step is to accurately classify news articles relevant to the target industry chain from a massive amount of unlabeled long-form news articles, and to expand the context to address the relevance and completeness issues of the input data for the safety risk assessment system. Because truly relevant information about the safety risks of key industry chains (chemicals, semiconductors, etc.) is often very sparse, and given the time-consuming nature of detailed analysis by large models, classifying news articles relevant to the industry chain before subsequent assessment is crucial. Previously, clustering or classifying text data often employed traditional machine learning models such as Support Vector Machines, Naive Bayes, and Random Forests. However, these models have limited semantic understanding capabilities, perform poorly in classifying long texts, and struggle to adapt to dynamic changes in the context of different industry chains. This is especially true for news data, where texts typically contain numerous details and background descriptions, making it easy for traditional methods to lose important information or introduce biases.

[0030] Therefore, to address the aforementioned issues, this invention designs a deep semantic perception and classification method for long texts. After constructing a feature instruction set for the industry chain domain, this method, from the perspective of logical reasoning, internalizes industry chain-related knowledge into a news classification model by retraining some key parameters of an open-source language model, thereby achieving industry chain news classification. Specifically, it includes two core stages: model fine-tuning and model inference. It integrates low-rank adaptive fine-tuning technology and a vectorized embedding model (BGE-M3) to construct a high-quality industry chain news relevance annotation dataset and train it into a large-scale news classification language model. Simultaneously, it automatically retrieves extended information from the internet as supplementary data, providing sufficient and accurate information for subsequent modules. A detailed illustration is shown below. Figure 2 Figure 3 The similarity calculation formula is shown in Equation 1.

[0031] This step categorizes content related to the target industry chain from the initial news data and supplements it with additional information. This process can be summarized as "Construction of Industry Chain News Classification and Labeling Dataset — Efficient LoRA Fine-tuning — Related Information Expansion".

[0032] First, a dataset for classifying and labeling industry chain news was created. A text embedding model (BGE-M3) was used to transform news text and industry chain description text into high-dimensional dense vectors, and cosine similarity was calculated using Equation 1. Simultaneously, a lightweight LLM model (gemini-3.1-flash-lite-preview) was used with a low generation temperature parameter (Temperature=0.1) for binary classification inference (clue words are shown in Table 1), while industry chain news relevance was simultaneously labeled. Data meeting any of the following conditions underwent expert review, resulting in a high-quality, fine-tuned dataset.

[0033] Formula 1 in, Represents news content vectors, Represents the target industry chain vector. Similarity and Threshold A comparison is then made. This involves using a text embedding model (BGE-M3 model) to convert the news content to be processed into a high-dimensional vector representation. Simultaneously, the descriptive text of the target industry chain ("chemical industry chain", "new energy vehicle industry chain") is converted into a target vector. (Note: All comparison texts must be vectorized using the same data type to ensure consistency in the vector space.) Only texts with a similarity greater than the threshold will be considered similar. Only the most detailed news text will be retained for expert review (see...). Figure 2 Finally, experts in the industry chain manually tagged this part of the reviewed news based on actual business logic (a tag of 1 indicates "relevant" and 0 indicates "irrelevant").

[0034] Secondly, LoRA efficient fine-tuning. This invention employs low-rank adaptive techniques for efficient parameter fine-tuning of open-source large language models. The training principle and process are as follows: Figure 3 The core idea of ​​LoRA is to freeze the original weight parameters of a pre-trained large model and inject trainable low-rank decomposition matrices only on the side of specific fully connected layers. By optimizing these very few incremental parameters, the model's ability to perform downstream classification tasks is achieved.

[0035] Finally, relevant information is expanded. For the news categorized by the fine-tuning model, AI web search tools (such as Bocha web search) are used to retrieve relevant official responses, policy documents, in-depth reports, and other extended information from the internet. These retrieved texts are used as supplementary context and merged with the original news to provide a complete information background for subsequent large-scale model analysis. The specific process for news text filtering and enhancement based on the fine-tuning model is as follows: Figure 4 As shown.

[0036] S2, topic extraction and merging.

[0037] After classifying and enhancing industry chain news, the next challenge is aggregating fragmented news. Because raw news texts often exhibit diverse expressions and varying information densities, directly using them for risk assessment can lead to distorted conclusions. Therefore, LIRAN's theme summarization layer aims to extract and merge themes from industry chain news, providing a clear analytical direction for subsequent news analysis and assessment. This facilitates the extraction of industry chain-level security risk indicators from discrete industry chain news sets by theme. Specifically, by mapping long texts to structured combinations of "risk themes + risk tags," the model can extract interfering factors and focus on core events that substantially impact industry chain security.

[0038] The implementation principle is based on Prompt Engineering, which uses pre-designed prompt templates and a small number of examples to guide the model extraction of top-level and sub-topics, such as... Figure 5 Similar topic merging employs a vector clustering method: first, the topics and tags of news texts are vectorized using the BGE-M3 model, and the cosine similarity of each text is calculated (method reference S1) for clustering; subsequently, a large language model is used to uniformly name topics of the same category. This method ensures the dynamism and accuracy of the topic database, supporting multi-dimensional event classification in supply chain security risk assessment.

[0039] S3, Risk Identification and Assessment.

[0040] The risk identification and assessment module addresses the core issues of how to correctly structure and analyze industry chain news and aggregate discrete news assessment results to extract an industry chain security risk index. It consists of two parts: first, structured analysis of each news item; and second, statistical analysis of the analysis results, constructing an indicator system to calculate weights and periodic risk values. This module is the core of the entire system, combining knowledge base construction, prompt word engineering, and multi-criteria decision-making methods to perform structured analysis on enhanced relevant news and calculate statistical indicators, ultimately obtaining a quantitative risk index for the industry chain. The overall principle and process can be found in [link to documentation]. Figure 6 .

[0041] The process in this step involves: performing structured analysis on the news and calculating risk values. First, it receives news text data, extracted topic results, a risk knowledge base, and a supply chain graph knowledge base. When analyzing individual news items, a large language model combined with cue word engineering is used. Simultaneously, historical assessment cases and supplementary supply chain information are retrieved from the knowledge base using retrieval enhancement generation technology and provided to the large language model. Based on this, the large language model evaluates the news from several dimensions: certainty, impact, resilience, and topic classification. The certainty dimension includes historical frequency, clarity of triggering factors, and precedent. The impact dimension includes the scope of influence, severity keywords, speed of dissemination, and the authority of the information source. The resilience dimension includes reversibility and the decision-making window. After completing the analysis of a single news item, statistical data is collected according to a set time window. Then, the proportion of each specific indicator is calculated according to three criteria: supply chain competitiveness risk, supply chain recovery capability risk, and supply chain control risk. Finally, the weighted average of each supply chain security risk indicator is calculated using the analytic hierarchy process to obtain the supply chain security risk index for this time window, ultimately outputting a comprehensive supply chain security risk value and a periodic risk assessment report.

[0042] This step combines the news topics provided by the topic extraction and classification module with the risk knowledge base and industry chain graph knowledge base additionally constructed in this paper. A powerful comprehensive language model (DeepSeek-V3.2, Qwen-Max) is used to evaluate the news texts. For industry chain news, this study selected 3 primary indicators and 9 secondary indicators to obtain the objective indicators of each relevant news item on the security of the studied industry chain, as shown in Table 2. The objective indicators are based on signal transmission theory, the actual capabilities of the large language model, and the connotation of industry chain security, and were summarized through expert consultation using the Delphi method and relevant literature. Each news item is categorized into one or more aspects of geopolitics, market economy, technological innovation, operational logistics, and policy and law. Finally, the qualitative assessment is converted into a quantitative score by calculating weights using the analytic hierarchy process (AHP) to quantify the final industry chain security risk assessment result.

[0043] Table 1. Analytical Dimensions for Each News Text

[0044] A risk identification architecture based on Retrieval Enhanced Generation (RAG): When processing relevant news texts to identify their impact on supply chain security and potential risks, this patent employs a large language model with hundreds of billions of parameters as the core semantic understanding foundation. Specifically, advanced models including, but not limited to, DeepSeek-V3, Qwen-Max, and GLM-4-9B are selected. These models, pre-trained on massive amounts of general-purpose corpora, possess extremely strong natural language processing, logical reasoning, and contextual understanding capabilities. However, to achieve accurate identification of deep-seated risks in specific supply chains (such as chemicals and new energy vehicles), simply relying on the general knowledge embedded in the model's internal parameters is insufficient. Therefore, this invention constructs a professional knowledge base retrieval enhanced generation system composed of a "supply chain node knowledge base" and a "risk assessment case base," serving as the underlying knowledge support for the model in the quantitative decision-making stage. Technically, the system constructs a knowledge base based on the pgvector vector database for two parts of knowledge. The construction method involves using a text embedding model (such as the BGE-M3 model) to convert text content into high-dimensional vectors hierarchically. During the question-answering phase, the system first retrieves content from these two knowledge bases using vector similarity, then converts the relevant content back into text and adds it to the context. Finally, the large model provides a structured analysis of the news based on the news and this relevant context. This system architecture addresses the illusion problem that large language models are prone to when processing knowledge outside the training domain, as well as the technical shortcomings of traditional model fine-tuning methods, such as high computational resource consumption, delayed knowledge updates, and poor versatility. It enhances the model's reasoning ability through decoupling—the knowledge base provides instant retrieval, avoiding the limitations of resource-intensive fine-tuning, while utilizing the model's long context and autonomous judgment capabilities. The external knowledge base introduced in this patent is mainly divided into two parts: a risk assessment case library, providing the large model with pre-existing cases for reference. When the large model processes a new news item, the system retrieves the most similar historical cases as examples in the "Prompt" (Few-shot examples). This not only standardizes the output format of the large model but also allows the model to make analogical inferences based on the logic of historical precedents, thereby significantly improving the consistency and accuracy of the assessment. The second part consists of the industrial chain structure of different industrial chains and related information (such as upstream and downstream supply and demand relationships, lists of key node enterprises, and core technology dependencies), which helps the large model supplement the data information for risk assessment. For example, when news mentions "a chemical raw material plant has stopped production," the knowledge base can supplement the background information that "this raw material is a core component of new energy batteries," helping the large model accurately deduce potential risk factors that "may affect downstream battery production capacity."

[0045] Why choose RAG instead of model fine-tuning? The reason for not using model fine-tuning is that its approach involves retraining the new model with the new knowledge required, but this consumes too much computing and human resources and may even result in worse performance. Fine-tuned models can only be used for the corresponding problems with the best results.

[0046] In summary, the approach is based on three considerations: First, resource consumption and update lag. The industrial chain structure and risk events are highly dynamic (such as sudden sanctions or raw material price fluctuations). Fine-tuning the model not only requires expensive computing resources, but also necessitates retraining with each knowledge update. Second, the risk of worse fine-tuning results. Excessive fine-tuning on specific vertical domain data often leads to the model losing its original general logical reasoning ability. Third, the advantages of a decoupled architecture. By separating computation from knowledge, the large model serves only as an integrator for reasoning, while the external knowledge base acts as dynamically added storage. This architecture leverages the large language model's extremely long context window and powerful autonomous judgment capabilities, allowing for the real-time injection of the latest knowledge during reasoning.

[0047] The construction of an external knowledge base adopts the idea of ​​decoupling. When the large language model needs new knowledge, it automatically retrieves the constructed knowledge base. The limitation of this method is that the large language model needs to have the ability to make autonomous judgments and have a long context. However, the large language models currently in use have a large enough number of parameters and context, and all have the ability to make autonomous judgments and have long context. Therefore, the method of constructing a knowledge base is used to enhance the model's answer.

[0048] Based on this, the 5W2H analysis method (Table 2) from the thinking analysis method is introduced into news analysis to guide the industry chain news analysis model to obtain the indicators in Table 2 and the textual analysis of the news.

[0049] Table 2. Brief Description of the 5W2H Analysis Method

[0050] After obtaining comprehensive analysis results for each news item, this invention addresses the issue of aggregating discrete news analysis results into an industry chain security risk index by focusing on the time attributes of the news (daily / weekly / monthly, etc.). This invention combines fragmented news analysis with the design of 3 primary indicators, 9 secondary indicators, and 15 tertiary indicators to calculate the industry chain security risk system indicators. Furthermore, expert judgment and multi-criteria decision-making are used to calculate the weights of these indicators (indicators and weights are shown in Table 3).

[0051] Table 3. Supply Chain Security Risk Assessment Index System

[0052] This invention proposes a dynamic text classification and expansion mechanism combining general large-scale model fine-tuning and vectorized similarity matching. The method first uses an embedding model (such as BGE-M3) to vectorize massive amounts of news text and target industry chain description text separately. By calculating cosine similarity and setting a threshold β, a small model is used in parallel to query the relevance of news and industry chains. The results are then combined with expert verification to efficiently and accurately construct the model training dataset. Subsequently, a general language model is trained using LoRA model fine-tuning technology to obtain an industry chain news classification model specifically designed to classify highly relevant news from massive amounts of messy news data. Simultaneously, for the classified news, this invention introduces an AI web search tool (such as Bocha Web Search) to automatically retrieve and integrate relevant official responses, policy documents, reports, and other extended information, merging them into a complete context. This combined approach solves the problem of how to quickly and accurately classify relevant news from massive amounts of news, providing high-quality, highly targeted input data for subsequent analysis.

[0053] This invention proposes a method for automatically summarizing and classifying topics in filtered news by utilizing the natural language understanding capabilities of large language models, combined with prompt engineering and vector clustering techniques. Figure 3 This method guides a large model to extract top-level and secondary topic tags from news articles using pre-designed prompt templates and a few examples. It then uses a vectorization model to represent the extracted topics and performs clustering using cosine similarity calculation. Finally, a large language model is used to uniformly name similar topics. This method can dynamically and accurately establish a structured industry chain risk event topic database, achieving multi-dimensional and standardized management of unstructured news information. This lays the foundation for subsequent targeted risk assessments and demonstrates significant innovation.

[0054] This invention proposes an innovative risk identification architecture that decouples the powerful semantic understanding and logical reasoning capabilities of large language models (such as DeepSeek-V3 and Qwen-Max) with external professional knowledge bases (including risk assessment case libraries and industry chain graph knowledge bases). This method employs Retrieval Enhanced Generation (RAG) technology, using vector databases such as pgvector to retrieve the most similar historical risk cases in real time as reasoning examples (Few-shot examples) when the model analyzes a single news item, supplementing it with background knowledge such as industry chain structure and upstream and downstream relationships. This approach effectively overcomes the illusion problem of large models and avoids the shortcomings of traditional fine-tuning methods, such as high computational consumption and lagging knowledge updates. It achieves accurate and structured analysis of news events across multiple predefined dimensions (such as news classification, historical frequency, scope of impact, severity, dissemination speed, information source authority, reversibility, and decision-making window). This fusion architecture is key to this invention's ability to transform qualitative information into quantitative assessment.

[0055] This invention innovatively designs a complete industrial chain security risk assessment index system (Table 2). This method performs structural analysis of each news item's results, statistically analyzing them over a fixed time window (day / week / month), to calculate more than ten specific indicators. Finally, by using AHP (Advanced Personality Hierarchy Process) to calculate weights and synthesize these indicator scores, a quantitative assessment value is calculated that comprehensively reflects the risks faced by the industrial chain in different dimensions such as competitiveness, resilience, and control. This method of systematically transforming the qualitative analysis results of a large model into a quantifiable, comparable, and traceable risk index constitutes the final output of this invention, possessing high practical value and innovation.

[0056] The general supply chain security risk identification method based on a large language model proposed in this patent has demonstrated good results in practical applications. To verify its effectiveness, we compared this patented method with industry-recognized supply chain risk indicators. The results show that this patent has significant advantages in both risk quantification and timeliness.

[0057] Validation of the industry chain news classification model: The training results of the model trained on the industry chain labeled news dataset were compared before and after training (Table 5, Table 6) and the accuracy was compared with the baseline model.

[0058] Table 5. Comparison of five accuracy experiments before and after training.

[0059] The experiment evaluated the speed and alignment performance of the model during the inference phase before and after training (Table 6).

[0060] Table 6 Comparison of classification efficiency indicators before and after training

[0061] Table 7 Experimental results compared with the baseline model

[0062] Validation of the Chemical Industry Chain: This invention selects the chemical industry chain as the research object to verify the reliability of the method. As a fundamental pillar industry of the national economy, the chemical industry is characterized by its long industrial chain, significant cyclical fluctuations, and high degree of industrial linkage. From the perspective of the industrial chain structure, the upstream segment of the chemical industry chain is highly dependent on the extraction and supply of bulk commodities such as crude oil and natural gas; the midstream segment involves relatively complex refining and chemical synthesis processes; and the downstream segment is widely connected to multiple end-consumer application fields such as automobile manufacturing, electronics, and textiles. In recent years, the chemical industry chain has been frequently disrupted by various external shocks due to the interplay of multiple external factors, including frequent geopolitical conflicts, the competition for production policies among major international oil-producing countries, and the continuous escalation of global trade barriers. Therefore, my country's chemical industry chain urgently needs to construct a methodological system capable of dynamically measuring security risks to enhance the overall resilience of the industrial chain.

[0063] This invention uses 50,000 real news articles from 2020 to 2023 in the OpenNewsArchive database of the OpenDataLab platform as the basis for analysis. This dataset has broad coverage and high neutrality. After screening and classification in S1 of this invention, background noise was removed and 1,610 risk texts highly relevant to the chemical industry chain were identified. Through structured and quantitative analysis in S2 and S3, the safety risk situation of the chemical industry chain was obtained, as shown in the figure.

[0064] The results clearly demonstrate that specific macroeconomic events have a significant impact on supply chain security, thus verifying the effectiveness of the framework model proposed in this invention and its sensitivity to risk. Figure 7As can be seen, the solid black line represents the total risk of the chemical industry chain, exhibiting significant multi-peak fluctuations. The score reached a high point at the end of 2020 and the beginning of 2021, fluctuated multiple times in 2021, and then showed an overall upward trend, remaining at a high level until the end of 2023. The light blue dashed line represents the three-period moving average of the chemical industry chain risk. This line mainly smooths out short-term fluctuations and reveals long-term trends. The curve is relatively flat, indicating that the overall risk fluctuated upwards between 2020 and 2023, and although it stabilized after 2023, it remained at a high level. The fluctuations in the industry chain security risk score are closely related to real-world events. The peak at the end of 2020 and the beginning of 2021 was related to the global COVID-19 pandemic, which led to supply chain disruptions, production stoppages, and logistical bottlenecks, exacerbating the vulnerability of the chemical industry chain. The subsequent decline in fluctuations reflects the positive impact of regulatory policies, but the subsequent upward trend indicates that geopolitical conflicts, energy price fluctuations, and tightening environmental policies continue to impact the industry chain.

[0065] Ablation experiments identify the actual contribution of each module ( Figures 8a-8f ): like Figure 8b As shown, the risk assessment module is removed from LIRAN, and the safety risk assessment of the chemical industry chain relies solely on the capabilities of the large language model itself, while keeping the assessment prompts unchanged.

[0066] like Figure 8c As shown, by removing the knowledge base and RAG module from LIRAN, the evaluation of industry chain news is carried out solely based on the constructed news evaluation index system.

[0067] like Figure 8d As shown, by removing the news topic extraction module from LIRAN, only the time dimension is used as the aggregation basis when calculating the supply chain security risk of discrete supply chain news: (1) The news structured analysis and supply chain risk extraction methods included in the risk assessment module play a key role in ensuring the objectivity and stability of supply chain security risk assessment.

[0068] (2) The knowledge base and RAG module play an important role in preventing security risks from being underestimated by false alarms or missed reports.

[0069] (3) The topic extraction module plays a role in enhancing the model’s attention to the identified topics and reducing the disturbance caused by the superposition of similar news data.

[0070] To further verify the effectiveness of the safety risk assessment results for the chemical industry chain and demonstrate the accuracy of the LIRAN model itself, this invention employs a correlation analysis between the concurrent chemical price index and the safety risk score of the chemical industry chain. Figure 9 and Figure 10As shown.

[0071] Finally, knowledge base analysis was used to determine the correlation between risk information and nodes in the industrial chain. Figure 11 The size of the nodes in the diagram maps the frequency and core nature of the risk causative factors, while the density of the connections reflects the coupling strength between different causative elements. Combining the network graph characteristics, the key safety risks in the chemical industry chain exhibit the following patterns: a "dual-core" characteristic dominated by market economy and operational logistics; barriers of social and legal constraints and technological innovation; and highly networked coupling and cross-domain spillover of safety risk elements. The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not omitted in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0072] In another embodiment of the present invention, a general supply chain security risk identification system based on a large language model is provided. This general supply chain security risk identification system based on a large language model can be used to implement the above-mentioned general supply chain security risk identification method based on a large language model. Specifically, the general supply chain security risk identification system based on a large language model includes a related news text extraction module, a target news text acquisition module, a topic classification information acquisition module, a retrieval module, a qualitative assessment module, a supply chain security risk assessment index calculation module, and a supply chain security risk index generation module.

[0073] The associated news text extraction module is used to obtain initial news text data, input the initial news text data into the industry chain news classification model, and extract industry chain associated news text.

[0074] The target news text acquisition module is used to acquire extended information corresponding to industry chain-related news texts, and merge industry chain-related news texts and extended information to obtain target news texts.

[0075] The topic classification information acquisition module is used to extract topics from target news texts using a large language model, and obtain the topic classification information corresponding to the target news texts.

[0076] The retrieval module is used to retrieve background reference content that matches the target news text from a preset knowledge base.

[0077] The qualitative assessment module is used to input the target news text, topic classification information, and background reference content into the large language model, and output the qualitative assessment results of the target news text on the preset assessment dimensions.

[0078] The supply chain security risk assessment indicator calculation module is used to calculate the supply chain security risk assessment indicators by statistically analyzing the qualitative assessment results according to a preset time window.

[0079] The supply chain security risk index generation module is used to perform weighted calculations on supply chain security risk assessment indicators to generate the supply chain security risk index.

[0080] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, the computer program including program instructions, and the processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., are the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to realize the corresponding method flow or corresponding function; the processor described in the embodiments of the present invention can be used for the operation of a general supply chain security risk identification method based on a large language model, including: acquiring initial news text data, inputting the initial news text data into a supply chain news classification model, and extracting supply chain related news text; acquiring extended information corresponding to the supply chain related news text, and merging the supply chain related news text and extended information to obtain target news text; using a large language model to extract the theme of the target news text, and obtaining the theme classification information corresponding to the target news text; retrieving background reference content matching the target news text in a preset knowledge base; inputting the target news text, theme classification information and background reference content into a large language model, and outputting a qualitative evaluation result for the target news text on a preset evaluation dimension; statistically analyzing the qualitative evaluation results according to a preset time window, and calculating the supply chain security risk assessment index; and performing weighted calculation on the supply chain security risk assessment index to generate a supply chain security risk index.

[0081] In another embodiment, the present invention also provides a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here may include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0082] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the general supply chain security risk identification method based on a large language model in the above embodiments. One or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps: obtaining initial news text data, inputting the initial news text data into a supply chain news classification model, and extracting supply chain-related news text; obtaining extended information corresponding to the supply chain-related news text, and merging the supply chain-related news text and extended information to obtain target news text; using a large language model to extract the theme of the target news text, and obtaining the theme classification information corresponding to the target news text; retrieving background reference content matching the target news text in a preset knowledge base; inputting the target news text, theme classification information, and background reference content into the large language model, and outputting a qualitative assessment result for the target news text on a preset assessment dimension; statistically analyzing the qualitative assessment results according to a preset time window, and calculating the supply chain security risk assessment index; and performing a weighted calculation on the supply chain security risk assessment index to generate a supply chain security risk index.

[0083] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0088] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0090] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0091] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

[0092] It should be understood that the above description is for illustrative purposes and not for limitation. Many embodiments and applications beyond the provided examples will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of this patent should not be determined by reference to the above description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed inventive subject matter.

Claims

1. A general supply chain security risk identification method based on a large language model, characterized in that, The process includes the following: Obtain initial news text data, input the initial news text data into the industry chain news classification model, and extract industry chain related news text; Obtain extended information corresponding to news texts related to the industrial chain, and merge the news texts related to the industrial chain and the extended information to obtain the target news text; A large language model is used to extract themes from the target news text, thereby obtaining the theme classification information corresponding to the target news text. Retrieve background reference content that matches the target news text from a pre-defined knowledge base; Input the target news text, topic classification information, and background reference content into the large language model, and output the qualitative evaluation results of the target news text on the preset evaluation dimensions. Based on the qualitative assessment results collected within a preset time window, the industrial chain security risk assessment indicators are calculated. The supply chain security risk assessment indicators are weighted and calculated to generate the supply chain security risk index.

2. The method for identifying general supply chain security risks based on a large language model as described in claim 1, characterized in that, The training steps for the industry chain news classification model include: The sample news text and the target industry chain description information are processed by feature vectorization and similarity is calculated. Texts whose similarity meets the set threshold condition are extracted as the first associated text. The sample news text is input into a lightweight language model for inference and recognition, and the text that is determined to be relevant is extracted as the second related text; Merge the first and second associated texts to obtain a fine-tuned dataset after manual review and labeling. Low-rank adaptive techniques are used to fine-tune the parameters of a pre-trained large language model, and the incremental parameter matrix of the pre-trained large language model is updated to train a news classification model for the industry chain.

3. The general supply chain security risk identification method based on a large language model according to claim 1, characterized in that, The steps for obtaining extended information corresponding to industry chain-related news texts and merging industry chain-related news texts and extended information to obtain target news texts include: Search online for relevant policy documents and online reports related to news articles related to the industrial chain; By combining relevant policy documents, online reports, and news texts related to the industry chain, the target news text is generated.

4. The general supply chain security risk identification method based on a large language model according to claim 1, characterized in that, The steps for extracting topics from target news text using a large language model to obtain topic classification information for the target news text include: Construct a multi-level prompt word architecture based on few-shot learning; Input the multi-level prompt word architecture into the large language model to extract the top-level and second-level topics of the target news text; The text embedding model is used to map the top-level topics and second-level topics into topic vectors; the cosine similarity of each topic vector is calculated and iterative clustering analysis is performed to identify topic clusters whose similarity meets the preset similarity threshold; The standard topic names in the topic clusters are selected for deduplication and merging to generate topic classification information corresponding to the target news text.

5. The general supply chain security risk identification method based on a large language model according to claim 1, characterized in that, The steps for retrieving background reference content that matches the target news text from a pre-defined knowledge base include: Utilize a multilingual text embedding model to convert target news text into dense retrieval vectors; In the risk assessment case library, which serves as a pre-defined knowledge base, the vector similarity of dense retrieval vectors is calculated, and typical historical events that meet the vector similarity requirements are extracted as logical reference cases. In the knowledge base of the industrial chain nodes, which serves as a pre-set knowledge base, the vector similarity of dense retrieval vectors is calculated, and the corresponding industrial chain structural features, regional distribution, and subdivision scale data are extracted as static background information of the industrial chain. By integrating logical reference cases with static background information of the industry chain, background reference content is generated.

6. The general supply chain security risk identification method based on a large language model according to claim 1, characterized in that, The pre-defined evaluation dimensions include certainty, impact, and flexibility. The certainty dimension includes historical frequency indicators, trigger clarity indicators, and precedent indicators; The influence dimension includes indicators such as the scope of influence, the severity of keywords, the speed of dissemination, and the authority of the information source. The resilience dimension includes reversibility indicators and decision-making window indicators.

7. The general supply chain security risk identification method based on a large language model according to claim 1, characterized in that, The steps for calculating the supply chain security risk assessment index by statistically analyzing the qualitative assessment results within a preset time window, and then weighting these indicators to generate the supply chain security risk index include: By using indicator functions to statistically analyze the frequency of occurrence of qualitative assessment results under preset assessment dimensions within a preset time window, a three-level indicator vector for supply chain security risks is constructed. Obtain the multi-level index weight matrix determined based on the analytic hierarchy process; By using a multi-level indicator weight matrix, the three-level indicator vectors of industrial chain security risk are weighted and aggregated layer by layer to obtain the industrial chain competitiveness risk indicator, industrial chain resilience risk indicator, and industrial chain control risk indicator, which are then established as industrial chain security risk assessment indicators. The supply chain security risk assessment indicators are multiplied by their corresponding global weights to generate the supply chain security risk index.

8. A general supply chain security risk identification system based on a large language model, characterized in that, include: The related news text extraction module is used to obtain initial news text data, input the initial news text data into the industry chain news classification model, and extract industry chain related news text. The target news text acquisition module is used to acquire extended information corresponding to industry chain-related news texts and merge industry chain-related news texts and extended information to obtain target news texts. The topic classification information acquisition module is used to extract topics from target news texts using a large language model, and obtain the topic classification information corresponding to the target news texts. The retrieval module is used to retrieve background reference content that matches the target news text from a preset knowledge base; The qualitative evaluation module is used to input the target news text, topic classification information and background reference content into the large language model, and output the qualitative evaluation results of the target news text on the preset evaluation dimensions. The supply chain security risk assessment indicator calculation module is used to calculate the supply chain security risk assessment indicators by statistically analyzing the qualitative assessment results according to a preset time window. The supply chain security risk index generation module is used to perform weighted calculations on supply chain security risk assessment indicators to generate the supply chain security risk index.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the general supply chain security risk identification method based on a large language model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the general supply chain security risk identification method based on a large language model as described in any one of claims 1 to 7.

Images