Foreign trade enterprise intelligent decision support system based on deep learning

Abstract

The invention discloses a deep learning-based intelligent decision support system for foreign trade enterprises, and particularly relates to the technical field of intelligent decision, which comprises a decision generation module, a decision inspection module, a first decision text acquisition module, a risk inspection module and a decision result acquisition module. According to the method, a decision text library is constructed through a decision inspection module, unstructured texts are converted into numerical vectors, the problems of cross-system data islands and semantic understanding errors are solved, customer decision data sets are integrated through cosine similarity matching, decision deviation caused by decision making through single data is not limited, and the accuracy of decision making is improved. By obtaining data from adjacent target decision vectors and dynamically adjusting decision paths, it is ensured that each decision node can meet preset index requirements, the decision efficiency and accuracy of cross-border e-commerce enterprises are comprehensively improved, the operation cost is effectively reduced, a set of replicable and easy-to-popularize intelligent decision support mode is formed, and industrial upgrading of the foreign trade industry is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent decision-making, more specifically, the present application relates to an intelligent decision-making support system for foreign trade enterprises based on deep learning. BACKGROUND

[0002] At present, the intelligent decision-making support system for foreign trade enterprises is in the key stage of digital transformation. With the intensification of globalization competition and the rise of uncertainty risks, more and more enterprises begin to introduce big data analysis and artificial intelligence technology to cope with market fluctuations, supply chain management and customer demand forecasting challenges. The head enterprises have gradually built data middle platforms, integrated multi-source data such as logistics, customs and customers, and tried to use deep learning models to optimize pricing strategies, risk early warning and other scenarios. However, the overall intelligent level of the industry is uneven, and small and medium-sized enterprises are still dependent on traditional experience-driven decision-making due to insufficient data accumulation and limited technology investment. The system application is mostly limited to basic data visualization, and has not yet formed a closed-loop intelligent decision-making capability.

[0003] The traditional intelligent decision-making system includes a scene creation module, a decision information parameter collection module, a decision information analysis module and a man-machine interaction module. The scene creation module is used for users to create a decision-making scene to provide a suitable decision-making environment for decision-making. The decision information parameter collection module collects decision information parameters by training a large model. The decision information analysis module is used to establish a decision information analysis model to integrate and analyze the collected data. The man-machine interaction module visualizes the decision information through the intelligent decision-making system, and the user provides suggestions for decision feasibility through the decision information.

[0004] However, in actual use, there are still some shortcomings. First, there are errors or biases in the process of data collection and processing, such as incomplete data collection or excessive influence of historical decision data, which affects the accuracy of decision results. Cross-system data islands and semantic understanding errors of unstructured text lead to model prediction bias. In this case, preprocessing technology needs to be introduced to improve data quality. Second, the real-time performance of the existing decision-making is limited, especially in the generalization ability of decision-making for sudden events, which causes great risk to enterprise operation. Based on this situation, risk monitoring and adjustment of the decision analysis model are needed, but the existing decision analysis is mostly limited to historical data training, which leads to a lag in response to sudden events and is difficult to adapt to small sample scenarios in emerging markets. SUMMARY

[0005] Therefore, embodiments of the present application provide a deep learning-based foreign trade enterprise intelligent decision support system, which sets a decision path by acquiring basic information and decision items in the foreign trade enterprise decision-making process, determines a target decision vector, and thereby performs fine and flexible adjustment, maximizes the combination of enterprise decision-making risks and effects, and effectively solves the problems raised in the background technology.

[0006] To achieve the above object, the present application provides the following technical solutions: The decision generation module is configured to acquire the basic information and the decision items in the decision-making process of the foreign trade enterprise and set a decision path corresponding to the decision items; The decision verification module is configured to set a plurality of decision nodes on the decision path, verify the information of each decision node on the decision path, and thereby determine a target decision vector corresponding to the decision node; The first decision text acquisition module is configured to extract decision indicators from the target decision vector, wherein the decision indicators include enterprise internal indicators, enterprise external indicators, and customer indicators, and thereby acquire a first decision text corresponding to the decision indicators; The risk verification module is configured to acquire decision risk features based on the first decision text and determine the decision risk difference degree of the decision risk features in the process corresponding to the decision indicators, and thereby evaluate the corresponding risk level of the first decision text; The decision result acquisition module is configured to combine all risk level results according to the positions of the decision indicators in the decision nodes based on the risk level results, thereby generating a decision evaluation report, and sending the report to a user terminal according to a preset summary mode.

[0007] The technical effects and advantages of the present application are as follows: 1. The decision verification module of the present application constructs a decision text library, converts unstructured text into numerical vectors, solves the problems of cross-system data islands and semantic understanding errors, integrates customer decision data sets through cosine similarity matching, is not limited to decision bias caused by single data, extracts enterprise internal, external, and customer indicators from the target decision vector according to the first decision text acquisition module, covers multi-dimensional information such as financial cost, exchange rate fluctuation, and repeat purchase rate, ensures comprehensive decision basis, and avoids single indicator limitations; 2. The present application can dynamically adjust the decision path by acquiring data from adjacent target decision vectors, ensure that each decision node meets the preset indicator requirements, the data of different decision nodes covers different decision information, integrates the data of multiple decision nodes, can perform more comprehensive analysis, makes the decision more accurate, on the other hand, without manual intervention in data collection and screening, the system can automatically complete indicator filling, improves decision efficiency, at the same time, the deep learning model of the risk verification module analyzes the decision risk features in real time, improves the response speed to unexpected events; 3. This invention analyzes the decision-making process of foreign trade enterprises through market analysis, risk assessment, and decision verification, comprehensively improving the decision-making efficiency and accuracy of cross-border e-commerce enterprises, effectively reducing operating costs, significantly enhancing their competitiveness in the global market, and reducing reliance on large amounts of historical data and professional technologies through automated data integration, path reuse, and visualized decision support, thereby promoting the comprehensive digital transformation of enterprises and forming a replicable and easily promoted intelligent decision support model to facilitate the upgrading of the foreign trade industry. Attached Figure Description

[0008] Fig. 1 This is a schematic diagram of the overall structure of the present invention.

[0009] Fig. 2 This is a flowchart of the first decision text acquisition process of the present invention.

[0010] Fig. 3 This is a schematic diagram of the risk inspection module structure of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0012] As attached Figs. 1-3 The system shown is an intelligent decision support system for foreign trade enterprises based on deep learning. It includes a system operating database, a system central processing unit and a user terminal, as well as a decision generation module, a decision verification module, a first decision text acquisition module, a risk verification module and a decision result acquisition module.

[0013] The specific embodiments of the present invention include the following: The system's operating database includes all data texts from the intelligent decision support system for foreign trade enterprises, and collects information texts output by each module in real time. The system's central processing unit is used to centrally control the information text instructions output throughout the entire control process, and the user information terminal is a device for receiving information outputs from the intelligent decision support system for foreign trade enterprises.

[0014] The decision generation module is used to obtain basic information and decision items for foreign trade enterprises when making decisions, and to set the decision path corresponding to the decision items.

[0015] In this embodiment, it needs to be specifically pointed out that the decision path is a series of steps corresponding to the entire decision-making process around the decision-making project of the foreign trade enterprise, including the decision-making arrangement of each stage, such as data collection, data analysis, decision-making target and decision-making project to be executed. Exemplarily, the market product pricing optimization can be set as the decision path as market demand data collection, cost structure analysis, competitor pricing benchmarking, dynamic pricing model calculation, risk simulation, and final strategy output.

[0016] It needs to be explained that the basic information generally refers to the internal and external environment of the enterprise and the basic information of the customer. The external environment of the enterprise is the data of the global market macro situation and competitive pattern, including but not limited to the GDP growth rate, inflation rate, unemployment rate and exchange rate of the target area. The internal environment of the enterprise is the structured data of the resources, capabilities and processes of the enterprise itself, including but not limited to historical decision-making, profit margin and cash flow status. These contents will represent the specific situation of the foreign trade enterprise. The foreign trade enterprise decision-making is classified according to these situations, and multiple groups of data are obtained.

[0017] The decision-making project refers to the specific decision-making task or problem that needs to be solved through data analysis and model calculation in the operation and management process of the enterprise. Exemplarily, it can be specific to optimize market product pricing, evaluate the risk of cooperation with new suppliers, and formulate quarterly inventory replenishment strategy.

[0018] When making decisions for foreign trade enterprises, setting the decision path corresponding to the decision-making project can convert the fuzzy decision-making target into clear steps and correspond to the system function module one by one, ensuring that the technical tool accurately intervenes in the decision-making link, avoiding the disorder when the decision maker faces a large amount of data, reducing the complexity of decision-making. On the other hand, in the case of successful decision-making of the enterprise, the decision path can be reused for similar decision-making projects, such as inventory strategy in different seasons, which can reduce repetitive work and improve the efficiency of enterprise decision-making.

[0019] The decision verification module is used to set multiple decision nodes on the decision path, and verify the information of each decision node on the decision path, so as to determine the target decision vector corresponding to the decision node.

[0020] In this embodiment, it needs to be specifically pointed out that the decision node is set as follows: A decision text library corresponding to the decision path is constructed. The decision text library includes multiple text vectors corresponding to the current basic information and the decision-making project. Exemplarily, the text vector is represented by any one of the text vector representation forms of one-hot encoding, Word2Vec and global vector. By converting the unstructured text of market research report in the decision text library corresponding to the decision path into a numerical vector, the position dynamics of different words and the relative relationship between texts are determined. The text vector is matched in similarity and the customer decision dataset of the text vector is set according to the value of the similarity matching, wherein the field decision dataset specifically includes extracting the customer historical order frequency and product preference type of the text vector according to the decision item and the basic information corresponding to the text vector, and setting the decision keyword of the text vector according to the customer historical order frequency and the product preference type. The decision keyword is mapped with the text vector, the similarity of the decision keyword is matched, and thus the cosine similarity of the decision keyword is obtained. When the cosine similarity of the decision keyword is greater than the preset similarity threshold, the corresponding text vector is integrated and constitutes the customer decision dataset, wherein the preset similarity threshold is set by the average value of the cosine similarity of the historical data for the current basic information and the decision item. The clustering center of the text vector is extracted from the customer decision dataset, and the clustering center is used as a decision node.

[0021] In the specific implementation of the above scheme, in the intelligent decision process, the clustering center represents the text after the text vector is clustered after the customer decision dataset is constructed. These texts represent the key texts corresponding to the key themes or demand patterns in the decision. For example, in the text vector space, each text sample is represented as a vector, and the clustering center of the cluster is the arithmetic average of all vectors in the cluster.

[0022] It should be explained that the similarity matching specifically includes: comparing the cosine similarity of the nearest set of decision keywords in the distribution position of the text vector in sequence, and setting the direction with the maximum cosine similarity of the decision keyword as the similarity matching direction; calculating the cosine similarity of the decision keyword in the similarity matching direction and outputting it as the value of the similarity matching of the text vector.

[0023] It should be explained that the text vector is represented by one-hot encoding, Word2Vec and global vector. The corresponding text vector can show different needs of the text vector in enterprise decision-making, and according to the different needs, the evaluation of the decision is realized, wherein one-hot encoding maps each word to a unique binary vector, which can intuitively define the decision scenario; Word2Vec generates dense vectors by predicting context words through a shallow neural network, which can capture the similarity of context words, such as cross-border e-commerce and logistics, payment in space; global vector balances the calculation efficiency and semantic accuracy by combining global statistical information and local context, which can more stably represent low-frequency professional words and provide more stable vector conversion for foreign trade terms.

[0024] The strategic decision node relates to the long-term development direction of the enterprise, including the establishment of decision criteria and the matching degree of target market corresponding resources; the tactical decision node is used for making medium and short-term execution strategies, including cost-benefit analysis and risk warning; the decision evaluation node is used for analyzing the capital cost of enterprise value creation evaluation, and the decision support node covers data aggregation, statistical analysis and conclusion at the end of the whole decision cycle.

[0025] When the decision node is set, it is set according to the key nodes on the decision path, so that the decision node can reflect the key moment of data collection, analysis and decision; The decision node includes strategic decision node, tactical decision node, decision evaluation node and decision support node. These nodes will show prominent forms when making decisions for enterprises. Verify these nodes, extract the target decision vector currently needed, and the logical relationship between different decision nodes and the influence degree of decision nodes on the overall decision process.

[0026] It needs to be further explained that the target decision vector is determined as follows: the obtained decision node is verified to determine the type of the current decision node, so that the decision node is any one of the strategic decision node, the tactical decision node, the decision evaluation node and the decision support node; According to the position of the decision node on the decision path, the decision tendency of the decision node is obtained; Compare the decision tendency of the decision node with the decision tendency threshold. When the decision tendency of the decision node is greater than or equal to the decision tendency threshold, it is regarded as a key decision node, and the key decision node is composed of a target decision vector; When the decision tendency of the decision node is less than the decision tendency threshold, it is regarded as an improved decision node, and the improved decision node is calculated by cosine similarity according to the specific content of the decision node and the decision node. The decision node with the maximum cosine similarity is regarded as the target decision vector.

[0027] It needs to be explained that the decision tendency is obtained by extracting the current node position and the total number of nodes, and then compared with the difference between the total number of nodes and the value 1. Because of the position of the decision node in the decision path, the front-end node has a greater impact, and the influence of the nodes at different positions in the decision path on the overall decision is different. By quantifying the position of the node in the path, this difference can be systematically reflected, so that the intelligent decision system can pay more attention to the key link, optimize resource allocation, and ensure the order of decision logic.

[0028] The first decision text acquisition module is used for extracting decision indicators from the target decision vector, wherein the decision indicators include enterprise internal indicators, enterprise external indicators and customer indicators, so as to obtain the first decision text corresponding to the decision indicators.

[0029] In this embodiment, it needs to be specifically pointed out that the enterprise internal indicators are to ensure the feasibility of decision making, such as financial cost and supply chain efficiency; the enterprise external indicators are to avoid environmental risks, such as exchange rate fluctuation and policy risk; and the customer indicators are to ensure demand orientation, such as repeat purchase rate and preference change.

[0030] It needs to be further pointed out that the first decision text is obtained in the following manner: the target decision vector is subjected to index extraction to obtain the enterprise internal indicators, the enterprise external indicators and the customer indicators in the target decision vector; According to the enterprise internal indicators, the enterprise external indicators and the customer indicators, the decision indicators corresponding to the decision nodes are set; If the type and the number of the decision indicators corresponding to the current target decision vector are less than the preset decision number, the adjacent target decision vector of the current target decision vector is obtained, and the data in the adjacent target decision vector except the decision indicators corresponding to the current target decision vector are selected to be added into the decision indicators corresponding to the current target decision vector, so as to determine the number of the decision indicators; The data after the supplement and the processing are output as the first decision text, and the first decision text includes the data stored in the multiple decision nodes, the decision types, the decision indicators and the target relationship.

[0031] It needs to be explained that the decision indicator type includes any one or more of the enterprise internal indicators, the enterprise external indicators and the customer indicators; by obtaining the data from the adjacent target decision vector, the decision path can be dynamically adjusted to ensure that each decision node can meet the preset indicator requirement, and the data of different decision nodes covers different decision information, so that the data of the multiple decision nodes can be integrated for more comprehensive analysis, so that the decision is more accurate, and manual intervention in data collection and screening is not required, the system can automatically complete the indicator filling, the decision efficiency is improved, and it is assumed that an export enterprise is making a decision on external market pricing, the strategic decision node only includes establishment of decision standard and target market corresponding resource matching degree, and the enterprise external indicator of exchange rate situation is lacking. By obtaining the exchange rate situation from the adjacent tactical decision node, the missing data can be completed, so that the decision information description of the enterprise decision is more complete; the data of the decision support node at the end of the entire decision cycle is summarized, but due to the change of the customer, the customer indicator is incomplete, and by obtaining the customer historical order from the adjacent decision evaluation node, the comprehensive evaluation of the customer can be ensured.

[0032] The risk inspection module obtains the decision risk features based on the first decision text, determines the decision risk difference degree of the decision risk features in the process corresponding to the decision indicators, and thus evaluates the corresponding risk level of the first decision text.

[0033] In this embodiment, it needs to be specifically pointed out that the decision risk difference degree is determined as follows: based on the first decision text, the decision risk features and historical decision risk features are extracted, the decision risk features include decision type and decision quantity; The decision type corresponding to the first decision text is compared with the historical decision type, it is judged whether there is the same decision type, if there is, the same decision type is recorded as the common decision type, and the different decision type is recorded as the new decision type; The number of common decision types and new decision types is respectively counted, here the number of each target decision vector corresponding to each common decision type and new decision type is recorded as Xj and Yj, and the number of successful decisions corresponding to each common decision type is extracted and recorded as Vj; According to the number of common decision types, the number of new decision types and the number of successful decisions corresponding to the common decision type, the decision risk model is established by deep learning algorithm, thus the decision risk feature value of the first decision text is calculated, which is specifically expressed as: , Among them, TD represents the decision risk feature value of the first decision text, Xj, Yj and Zj represent the number of common decision types, the number of new decision types and the total number of decisions corresponding to the foreign trade enterprise respectively, Vj i represents the number of successful decisions corresponding to the ith common decision type, n represents the total number of common decision types, wherein the more the number of common decision types is, the more the number of successful decisions corresponding to the common decision type is, and the smaller the decision risk is.

[0034] It needs to be further pointed out that the risk level of the first decision text is specifically evaluated as follows: the decision risk feature value of the first decision text corresponding to each decision index is compared, the maximum decision risk feature value is extracted, and the risk difference degree of the decision index is obtained according to the maximum decision risk feature value, which is specifically expressed as: , Among them, Q represents the risk difference degree of the first decision text, TD max and TD min represent the maximum decision risk feature value and the minimum decision risk feature value respectively; According to the risk difference degree of the first decision text, the risk level is set, specifically, TD min is taken as the first risk lower limit value, is taken as the second risk lower limit value, is taken as the third risk lower limit value, TD max is taken as the third risk upper limit value, and then the decision index is divided into risk levels through the risk level setting.

[0035] The decision result acquisition module combines all risk level results according to the positions of the decision indexes at the decision nodes based on the risk level results, thereby generating a decision evaluation report and sending the report to the user terminal according to a preset summary mode.

[0036] In the embodiment, the decision result acquisition module outputs the risk level in the risk inspection module, if the first decision text does not have risks, the enterprise regards the decision as an executable decision text through the user terminal, if the first decision text has risks, the decision is regarded as an unexecutable text, and the corresponding decision risk characteristic value is converted into a trend chart form to generate a decision evaluation report, the decision indexes are aggregated according to the positions of the decision nodes to obtain a final decision evaluation report which is displayed according to the distribution of the decision indexes and the decision nodes, thereby fully displaying the situations under different decision indexes; the preset summary mode is a combination of report summary, picture summary and chart summary.

[0037] Secondly, only the structures involved in the disclosed embodiment are involved in the drawings of the disclosed embodiment, other structures can be referred to the general design, and the same embodiment and different embodiments of the present application can be combined with each other under the condition of no conflict; Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A deep learning-based intelligent decision support system for foreign trade enterprises, characterized in that, include: The decision generation module is used to obtain basic information and decision items for foreign trade enterprises when making decisions, and to set the decision path corresponding to the decision items; The decision verification module is used to set up multiple decision nodes on the decision path, verify the information of each decision node on the decision path, and thereby determine the target decision vector corresponding to the decision node. The first decision text acquisition module is used to extract decision indicators from the target decision vector, including internal enterprise indicators, external enterprise indicators and customer indicators, thereby obtaining the first decision text corresponding to the decision indicators. The risk verification module obtains decision risk characteristics based on the first decision text and determines the degree of difference in decision risk under the corresponding decision indicators and processes, thereby assessing the risk level corresponding to the first decision text. The decision result acquisition module combines all risk level results according to the position of decision indicators at the decision node based on the risk level results, thereby generating a decision evaluation report, which is then sent to the user terminal according to a preset summary method.

2. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The decision node settings are as follows: Construct a decision text library corresponding to the decision path. The decision text library includes current basic information and multiple text vectors corresponding to the decision items. The text vectors are similarity matched, and a customer decision dataset is set up according to the similarity matching value. The domain decision dataset specifically includes extracting the customer's historical order frequency and product preference type from the text vectors according to the decision items and basic information corresponding to the text vectors. Decision keywords are set for the text vectors according to the customer's historical order frequency and product preference type. The decision keywords are mapped to the text vectors, and similarity matching is performed on the decision keywords to obtain the cosine similarity of the decision keywords. When the cosine similarity of the decision keywords is greater than the preset similarity threshold, the corresponding text vectors are integrated to form the customer decision dataset. Extract cluster centers from text vectors in the customer decision dataset and use these cluster centers as decision nodes.

3. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 2, characterized in that: The similarity matching specifically includes: according to the distribution position of the text vectors, comparing the cosine similarity of the nearest set of decision keywords in turn, and setting the direction in which the cosine similarity of the decision keywords reaches the maximum value as the similarity matching direction; calculating the cosine similarity of the decision keywords in the similarity matching direction and outputting it as the value of the text vector for similarity matching.

4. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The target decision vector is determined as follows: the acquired decision nodes are verified to determine the type of the current decision node, such that the decision node is any one of the strategic decision node, tactical decision node, decision evaluation node, and decision support node; Based on the location of the decision node on the decision path, obtain the decision tendency of the decision node; The decision tendency of a decision node is compared with the decision tendency threshold. If the decision tendency of a decision node is greater than or equal to the decision tendency threshold, it is regarded as a key decision node, and the key decision nodes are combined into a target decision vector. When the decision tendency of a decision node is less than the decision tendency threshold, it is regarded as an improved decision node. The improved decision node is then compared with the decision node by cosine similarity calculation according to the specific content of the decision node, and the decision node with the maximum cosine similarity is taken as the target decision vector.

5. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The first decision text is obtained as follows: the target decision vector is extracted by indicators to obtain the internal enterprise indicators, external enterprise indicators and customer indicators in the target decision vector; Set decision indicators corresponding to decision nodes based on internal enterprise indicators, external enterprise indicators, and customer indicators. If the type and number of decision indicators corresponding to the current target decision vector are both less than the preset number of decisions, then obtain the target decision vectors adjacent to the current target decision vector, and select the data from the adjacent target decision vectors other than the decision indicators corresponding to the current target decision vector to add to the decision indicators corresponding to the current target decision vector, thereby determining the number of decision indicators; The supplemented and processed data is output as the first decision text, which includes multiple decision nodes, decision types, decision indicators, and data stored on the target decision vector.

6. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The degree of difference in decision risk is determined as follows: Based on the first decision text, decision risk features and historical decision risk features are extracted. Decision risk features include decision type and number of decisions. Compare the decision type corresponding to the first decision text with the historical decision types to determine if there are any identical decision types. If so, record the identical decision types as regular decision types and the different decision types as new decision types. The number of common decision types and new decision types are counted separately. Here, the number of common decision types and new decision types corresponding to each objective decision vector are denoted as Xj and Yj, respectively, and the number of successful decisions corresponding to each common decision type is extracted and denoted as Vj. A decision risk model is established using a deep learning algorithm based on the number of common decision types, the number of new decision types, and the number of successful decisions corresponding to common decision types. The decision risk feature value of the first decision text is then calculated, specifically as follows: ， Where TD represents the decision risk characteristic value of the first decision text, Xj, Yj, and Zj represent the number of common decision types, the number of new decision types, and the total number of decisions corresponding to foreign trade enterprises, respectively, and Vj i Let represent the number of successful decisions corresponding to the i-th constant decision type, and n represent the total number of constant decision types.

7. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The first decision text is specifically assessed at the corresponding risk level as follows: The decision risk characteristic values ​​of each decision indicator corresponding to the first decision text are compared, and the extreme value decision risk characteristic value is extracted. Based on the extreme value decision risk characteristic value, the risk difference degree of the decision indicators is obtained, specifically expressed as follows: ， Where Q represents the risk difference degree of the first decision text, and TD max and TD min These represent the maximum and minimum decision risk eigenvalues, respectively. Risk levels are set based on the risk differences in the first decision text, specifically TD min As the first risk lower limit value As the second risk lower limit As the third risk lower limit, TD max As the third risk ceiling, the decision-making indicators are then classified into risk levels through risk level settings.

8. The intelligent decision support system for foreign trade enterprises based on deep learning according to claim 1, characterized in that: The decision result acquisition module outputs the risk level from the risk inspection module and converts it into a trend chart to generate a decision evaluation report. The decision indicators are aggregated according to the position of the decision nodes, and the final decision evaluation report is displayed according to the distribution of decision indicators and decision nodes, thereby fully demonstrating the situation under different decision indicators. The preset summary method includes a combination of report summary, picture summary and chart summary.

Images