Auxiliary decision-making system for agricultural economic cross-border cooperation English negotiation

Abstract

The invention relates to an agricultural economy cross-border cooperation English negotiation auxiliary decision-making system, in particular to the field of data processing and analysis, and the system can capture multi-source policy data in real time, construct a policy chain reaction network through dynamic network modeling, effectively evaluate the trade dependency relationship between participating countries, and improve the accuracy of the trade dependency relationship between the participating countries. The system generates a multi-dimensional policy risk score vector based on historical policy data through a risk prediction engine, constructs a multi-objective optimization decision model in combination with a preset negotiation constraint condition, and can output accurate English negotiation policy scripts including clause revision suggestions and interaction sequential logic through a policy optimization generation module. Therefore, accurate decision support is provided for cross-border cooperation negotiation, the cross-border agricultural economic cooperation negotiation method has remarkable advantages in the aspects of improving the decision efficiency of cross-border agricultural economic cooperation, reducing policy risks and optimizing negotiation strategies, the negotiation effect and the agreement achievement rate can be remarkably improved, and smooth proceeding of international trade is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing analysis, more particularly, the present application relates to an agricultural economic cross-border cooperation English negotiation auxiliary decision-making system. BACKGROUND

[0002] The negotiation process of global agricultural product cross-border trade is significantly affected by external policy factors, and both parties need to quickly adjust their strategies in a rapidly changing policy environment. As trade policies continue to change, suppliers and buyers must respond in real time by adjusting prices, quantities, and delivery terms based on the latest policy changes. However, due to the dynamic nature of policies and the complexity of the global economic environment, relevant policy information often updates slowly, causing the system to fail to reflect changing policies in a timely manner during negotiations. This delay not only increases the difficulty of negotiations, but also may lead to unfavorable trade conditions and the failure of agreements, even causing supply chain disruptions, economic losses, and other serious consequences.

[0003] Existing technical solutions cannot effectively address the time difference between external policy changes and internal data, severely affecting the efficiency of cross-border trade negotiations. Most current decision-making systems rely on historical data and static models for prediction, but these models lack real-time dynamic prediction capabilities and cannot quickly integrate multiple policy information, especially during political events or economic policy changes. The system fails to respond in a timely manner, resulting in a significant gap between the recommended strategies and the actual situation. In particular, in cases where tariff, quarantine, and other policy factors change rapidly, the system's strategies cannot adapt in time, thus failing to provide effective support for negotiations. More seriously, existing systems rely on manual intervention for data processing, with scattered data acquisition methods and low update frequency, causing the decision-making process to lag, affecting the efficiency and accuracy of negotiations. Therefore, there is an urgent need for a system that integrates real-time data capture, automated data integration, and intelligent prediction capabilities to respond to changing external policy environments and ensure timely and accurate strategy recommendations during cross-border trade negotiations. SUMMARY

[0004] The present application provides an agricultural economic cross-border cooperation English negotiation auxiliary decision-making system to solve the problems in the background art by setting up a real-time data capture module, a dynamic network modeling module, a risk prediction engine module, and a strategy optimization generation module.

[0005] The technical solution of the present application to solve the above technical problems is as follows: specifically includes: a real-time data capture module, a dynamic network modeling module, a risk prediction engine module, and a strategy optimization generation module.

[0006] Real-time data capture module: in response to a policy update trigger event or a user negotiation request, capture multi-source policy data streams and trade-related data, perform cross-language semantic alignment operations, and generate a structured policy event stream containing standardized triples;

[0007] Dynamic network modeling module: based on the structured policy event stream, construct a policy chain reaction network, which takes participating countries as nodes and trade dependence as dynamic edge weights, calculate node centrality indicators, and output network topology relationship matrices;

[0008] Risk prediction engine module: receive network topology relationship matrices and historical policy time series data, call network-corrected time series prediction models, perform joint training operations, and generate multi-dimensional policy risk score vectors containing time-sensitive risk variables;

[0009] Strategy optimization generation module: based on multi-dimensional policy risk score vectors and pre-set negotiation constraint conditions, construct a multi-objective optimization decision model, perform constraint satisfaction solving operations, and output English negotiation strategy scripts containing clause revision suggestions and interaction timing logic;

[0010] In a preferred embodiment, the capture step of multi-source policy data streams in the real-time data capture module is:

[0011] First, capture external policy update events or user's active request signals, and use them as trigger conditions to build a dynamic data source crawling queue. Each data source calculates its priority based on the source's authority and historical response delay, as well as the degree of association with specific keywords, and initiates requests through parallel crawling engines according to priority, outputting raw policy data streams;

[0012] The multi-source policy data stream in the real-time data capture module includes government API announcement data, agricultural economic news streams, and international organization policy library data.

[0013] In a preferred embodiment, the specific operation of performing the cross-language semantic alignment operation is:

[0014] For captured raw data, use heterogeneous parsers for classification processing. Specifically: for structured data, extract key-value pairs; for unstructured text, use "agricultural ontology injection BERT" for semantic role labeling, use an ontology knowledge base specific to the agricultural domain for cross-language semantic analysis, and generate triples for events;

[0015] Each event is converted to an atomic event after semantic role labeling, and its impact value is quantified. At the same time, a timestamp and a spatial stamp are added to all generated events to form a structured event set;

[0016] The compression and verification operation of generating the event stream is specifically:

[0017] The similarity between each event is calculated by the redundancy eliminator, and if the event similarity exceeds a set threshold, the events are merged. The similarity calculation integrates the similarity of event semantics and the proximity in time;

[0018] Secondly, check if there are mutually contradictory events, and determine which event is more credible through the credibility arbitration rule, and finally select the most credible event as the final output.

[0019] In a preferred embodiment, in the dynamic network modeling module, the specific operation of constructing a policy chain reaction network is:

[0020] A1, map the policy subject to a specific sovereign state, model the association between each event and country, generate a subject-event association matrix, in addition, generate a baseline adjacency matrix according to historical trade data to represent the trade relationship between countries, and consider the influence of time change on edge weight through a time decay factor;

[0021] A2, for each event, generate an impact propagation matrix based on its impact value and historical trade data. After the event occurs, the edge weight will be updated according to the propagation matrix, while considering the decay and time factors of policy effect;

[0022] A3, construct a policy sensitivity matrix to measure the sensitivity of each country to policy changes, then calculate the heterocentricity of each country by weighting the active influence and passive sensitivity of the node;

[0023] A4, generate a core subnetwork from the dynamic adjacency matrix, node centrality and sensitivity matrix obtained by A1, A2 and A3, filter out unimportant edges and nodes, select the core subnetwork according to the set threshold, and construct a topology relationship matrix, which contains the structure information of the dynamic network.

[0024] In a preferred embodiment, in the risk prediction engine module, the specific operation of receiving the network topology relationship matrix and historical policy time series data is:

[0025] Extract relevant features from the input topology structure, and combine them with historical event data to form a spatio-temporal joint input tensor, which specifically includes extracting the upper triangular elements from the core adjacency matrix, and combining them with historical event features, similarity of disturbance history and spatial network features to construct a new spatio-temporal input tensor.

[0026] In a preferred embodiment, the specific operation of calling the time series prediction model of network correction is:

[0027] By processing the spatio-temporal joint input tensor through a double-flow gating mechanism, on the one hand, the time flow uses an LSTM unit to process the historical time sequence, and on the other hand, the space flow uses a graph convolution network to process the static topological features, and through the gating mechanism, the fusion mode at each time step is determined, and the sensitivity of the model to the spatio-temporal features is adjusted.

[0028] In a preferred embodiment, the specific steps of performing the joint training operation are:

[0029] The fused state is mapped to different risk dimensions by a decoder,

[0030] and is jointly optimized through multi-task learning, and the loss function combines the mean square error and the KL divergence to ensure that the difference between the model output and the target value is minimized, and the gradient consistency of the risk distribution is constrained through the KL divergence.

[0031] In a preferred embodiment, the multi-dimensional policy risk score vector specifically includes:

[0032] Based on the fused risk vector, a time-sensitive risk vector is generated, which combines the current risk vector and the information of historical disturbance events, processes the disturbance history through convolution operation, and adjusts the weight according to the disturbance intensity.

[0033] In a preferred embodiment, the specific operation of constructing a multi-objective optimization decision model is:

[0034] Based on the multi-dimensional risk score data of the time-sensitive risk vector and the preset negotiation constraint condition, a dynamic feasible region containing rigid constraint boundary and risk-driven elastic constraint is constructed; wherein the boundary radius of the elastic constraint is adaptively adjusted according to the gradient change of the risk vector, and the adjustment amplitude is in a negative correlation relationship with the risk mutation intensity, forming a dynamically scalable decision space.

[0035] In a preferred embodiment, the specific operation of generating an English negotiation strategy script is:

[0036] By constructing an interactive event state diagram through the Pareto optimization solution set and historical negotiation data, an interactive path is generated using the weighted calculation of state transition probability and solution set approximation degree; and inputting the path and real-time risk change into the semantic fusion generator, combining with the cultural conflict detection mechanism to dynamically modify the basic template, finally outputting the English negotiation strategy script that conforms to the cross-border legal norms and cultural adaptation.

[0037] The agricultural economic cross-border cooperation English negotiation auxiliary decision system has the advantages that the system can capture multi-source policy data in real time, construct a policy chain reaction network through dynamic network modeling, effectively evaluate trade dependence between participating countries, generate a multi-dimensional policy risk score vector based on historical policy data through a risk prediction engine, construct a multi-objective optimization decision model combined with preset negotiation constraints, and output accurate English negotiation strategy scripts, including clause revision suggestions and interactive timing logic, through a strategy optimization generation module, so that the system can provide accurate decision support for cross-border cooperation negotiations, thereby significantly improving the negotiation efficiency, reducing policy risks and optimizing negotiation strategies of cross-border agricultural economic cooperation, significantly improving the negotiation effect and agreement achievement rate, and promoting the smooth progress of international trade. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The method flowchart of the present application is shown in the figure.

[0039] Figure 2 The system structure block diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0042] In the description of the present application, the term "for example" is used to mean "serving as an example, instance, or illustration." Any embodiment described as "for example" in this application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present application. It will be apparent, however, to one skilled in the art that the present application can be practiced without using these specific details. In other instances, well-known structures and processes are not elaborated upon in order to avoid unnecessary detail, which can obscure the description of the present application. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0043] Embodiment 1

[0044] The embodiment provides an agricultural economic cross-border cooperation English negotiation auxiliary decision-making system as shown in the following, and specifically comprises a real-time data capturing module, a dynamic network modeling module, a risk prediction engine module, and a strategy optimization generation module. Figures 1-2

[0045] The real-time data capturing module: in response to a policy update trigger event or a user negotiation request, the module captures multi-source policy data streams and trade association data, performs cross-language semantic alignment operations, and generates a structured policy event stream containing standardized triples (policy subject, regulatory behavior, and impact value). In cross-border agricultural negotiations, due to the dynamic, heterogeneous, and language diversity of multi-source policy data, traditional data collection methods often face problems of time delay and semantic deviation. Therefore, the module converts raw data into structured event streams through a hierarchical collaborative processing mechanism, providing accurate input data to support subsequent policy network modeling.

[0046] The dynamic network modeling module: based on the structured policy event stream, the module constructs a policy chain reaction network with participating countries as nodes and trade dependence as dynamic edge weights, calculates node centrality indicators, and outputs network topology relationship matrices. Since policy events produce chain reactions in cross-border agricultural trade, traditional static network models cannot capture dynamic dependence relationships. The module takes the structured event stream (P-Stream) output by the real-time data capturing module as input, constructs a dynamic policy propagation network, and introduces a dependence decay factor and an event impact propagation operator to quantify the dynamic evolution of trade relationships.

[0047] ​Risk prediction engine module: receive network topology relationship matrix and historical policy time series data, call network corrected time series prediction model, perform joint training operation, generate multi-dimensional policy risk score vector, multi-dimensional policy risk score vector contains time sensitive risk variable, due to policy risk prediction needs to capture time dimension evolution law and space dimension network effect at the same time, traditional time series model ignores the topology dependence of policy transmission, this module takes the topology relationship matrix (including core network, centrality vector, disturbance intensity) output by dynamic network modeling module and historical policy event stream of real-time data capture module as input, constructs network corrected joint prediction framework; The breakthrough is the design of topology-time coupling, which realizes cross-dimensional information fusion.

[0048] Strategy optimization generation module: according to multi-dimensional policy risk score vector and preset negotiation constraint condition, construct multi-objective optimization decision model, execute constraint satisfaction solving operation, output English negotiation strategy script, script contains clause revision suggestion and interactive time sequence logic, this module solves the problem that traditional optimization model ignores the dynamic nature of risk and language interaction causality, realizes risk-constraint-language co-evolution type decision.

[0049] In this embodiment, it is specifically necessary to explain the real-time data capture module, the capture steps of multi-source policy data stream are:

[0050] First, capture external policy update events (such as government API state code change) or user's active request signal REQ user = 1, and take it as the trigger condition to build a dynamic data source crawling queue Q = [S1, S2,..., S n ](S i is data source object), each data source calculates its priority according to the authority degree and historical response delay of the source, and the association degree with the specific keyword, its calculation formula is:

[0051]

[0052] Where, π(S i ) represents the data source S i dispatch priority, the larger the value, the higher the crawling priority, for example: 1. (prioritize crawling of authoritative data sources), 2. (prioritize crawling of efficient data sources), 3. (burst hot data source immediately occupies the queue); C i represents the source authority degree (preset constant, government API is 0.9, news media is 0.7), represents the historical average response delay, unit is second, and (N times of request average response time), D(t) represents the time window [t k-Δt,t k ] keyword density (e.g. "tariff" "quarantine") to monitor the suddenness of policy hotspots, represents the instantaneous change rate of keyword density at t k , which is used to identify the suddenness of policy events (e.g. sudden release of new tariff policy), t k represents the timestamp of the current priority calculation, which is used to identify the trigger time of dynamic scheduling, Δt represents the keyword statistical time window length (e.g. 5 minutes), and the Δt setting rule is Δt = 300s (5 minutes) in news stream and Δt = 60s (1 minute) in API, which is used to balance noise and signal capture, α represents the weight coefficient of authority and time delay, α = 0.6, which is used to control the data source authority C i and response efficiency on priority, β represents the weight coefficient of keyword mutation rate, β = 0.4, which is used to control the sensitivity of keyword density change rate on priority, and α + β = 1 (to ensure priority calculation normalization), and the specific value of α / β is adjusted according to business scenarios: in policy negotiation scenarios, β ≥ 0.5 (emphasis on suddenness), in data archiving scenarios, α ≥ 0.8 (emphasis on authority); in addition, when a data source suddenly generates high-frequency keywords (e.g. sudden news: "new tariffs on soybeans"), item significantly increases π(S i ), realizing second-level hotspot preemption; government API has C i = 0.9 and low latency, always maintaining high basic priority, ensuring core policy is not missed, and initiating requests by parallel crawling engine according to priority, outputting raw policy data stream Raw = {Text lang , Time, SourceID}, where Raw represents the raw multi-modal data stream, Text lang represents the raw text content (including language labels), e.g. {"en":"New tariffs on soybeans"}, Time represents the event timestamp in ISO8601 format: 2025-04-15T08:30:00Z, SourceID represents the unique identifier of the data source, e.g. USDA-API. This design combines the authority of the data source and the mutation rate of the content for hybrid scheduling, which can significantly improve the timeliness of capture, about 37% higher than traditional polling mechanism;

[0053] The multi-source policy data stream in the real-time data capture module includes government API announcement data, agricultural economic news stream, and international organization policy library data;

[0054] The specific operation of performing cross-language semantic alignment operation is:

[0055] For the captured raw data Raw, a heterogeneous parser is used for classification processing, specifically: for structured data (such as policy information provided by API), key-value pairs are extracted, and the data structure is represented as: <PolicyEntity, EffectiveDate, KV>, where PolicyEntity represents the policy subject entity, such as US_Customs (U.S. Customs), the generation rule is to extract from the issuer field of API and map to a unified identifier through the ontology library, EffectiveDate represents the policy effective date, such as 2025-06-30T00:00:00Z, and KV represents the key-value pair set, such as {"tariff_rate":0.15,"quarantine_rule":"A1"}; and for unstructured text (such as news reports), the "agricultural ontology-injected BERT" (Agro-BERT) will be used for semantic role labeling Embed=Agro-BERT(Text lang ;θ ontology ), where Embed represents the semantic embedding vector, i.e. the high-dimensional mathematical representation of the input text after processing by the Agro-BERT model, which is used to capture the deep semantics of agricultural policy text, with the following characteristics: dimension: 768 dimensions (baseline output dimension of BERT), data structure: floating-point vector v∈R 768 ;Text lang represents the input multilingual text data, which serves as the unstructured text to be parsed and contains multilingual agricultural policy content (such as English tariff announcements and Chinese quarantine regulations), with the following characteristics: cross-language support: compatible with ISO639-1 language codes (such as en, zh, es), text format: original string (contains agricultural professional terms such as "tariff" and "quarantine"); θ ontology represents the injected agricultural trade ontology parameter, which is used to guide the model to identify professional entities in the agricultural field (such as "soybean"→"soybean"), regulatory behaviors (such as "import ban"→"import ban"), and cross-language mapping relationships, and where W term represents the agricultural term vector table (covering 5000+ professional terms), M lang represents the cross-language alignment matrix (realizing Chinese and English term vector space projection); in the formula, ; represents the conditional parameter separator, which is used to declare θ ontology is the conditional input parameter of the model (i.e. controls the text parsing rules), and is not the direct processing object, which uses the ontology knowledge base specific to the agricultural field for cross-language semantic analysis, generating event triples;

[0056] Each event is converted into an atomic event (e.g., a country announces to increase tariffs) after semantic role labeling, E k = <Subject, Action k , ImpactVal k >, where E k represents the kth atomic event, which is the smallest logical unit of a policy event, and k ∈ [1, n] (n is the total number of events), <> represents a triple encapsulator, i.e., an event consists of three structured data of fixed dimensions, Subject represents a policy subject (PolicySubject), which refers to the entity that issues or is affected by the policy: a country / region (e.g., "USA"); an international organization (e.g., "WTO"); a government agency (e.g., "USDA"), Action k represents a regulatory action, which refers to the type of operation of the policy on the agricultural economy: tariff adjustment; quarantine change; import / export quota setting, ImpactVal k represents the impact value, which is used to quantify the strength of the economic impact caused by the policy, and for numerical explicit policies (such as tariff adjustment):

[0057]

[0058] where NewValue represents the new numerical value extracted from the policy text (e.g., 15 in "tariff increased to 15%"), BaseValue represents the benchmark value extracted from the historical database (e.g., the current tariff of 10%), ± represents the sign determined by the semantic of the regulatory action ("tax increase" takes +, "tax reduction" takes -), η represents the policy type weight (preset constant: tariff class = 0.9, quarantine class = 0.7, subsidy class = 0.6), Conf sem represents the semantic confidence (range [0, 1], output by Agro-BERT classifier);

[0059] For non-numerical policies (such as import / export ban):

[0060]

[0061] where ω j represents the impact dimension weight (economic = 0.4, supply chain = 0.3, legal = 0.2, environmental = 0.08, social = 0.02), I j represents the dimension impact value (output by the pre-trained model, algorithm as follows), GeoScale represents the geographical impact range coefficient (multinational trade agreement = 1.0, bilateral = 0.6, unilateral = 0.3);

[0062] The calculation formula of the dimension impact value I j is as follows:

[0063]

[0064] where, denotes the attention mechanism-based impact value classifier (Θ is the model parameter), EntityDensity denotes the matching density of the agricultural entity word frequency (such as soybean, quarantine) in the policy text and the domain dictionary D j , μ denotes a nonlinear amplification factor, and μ = 0.5;

[0065] All ImpactVal k A time decay factor needs to be superimposed (to avoid the interference of expired events on network modeling):

[0066]

[0067] where, t k denotes the event occurrence timestamp (taken from P-Stream), τ denotes the half-life period (τ = 30 days for the tariff class and τ = 7 days for the embargo class), and quantifies the impact value (for example, an increase of 10% in the tariff has an impact value of 0.1), while a timestamp and a spatial stamp are added to all generated events, thereby forming a structured event set Events = {E1, E2,..., En}, each event containing a unique time-space stamp Time m , compared with traditional keyword extraction, through dynamic numerical conversion rules (tariff class) and semantic intensity classifiers (non-numerical class), the computable expression of policy impact is realized, providing atomic-level input for subsequent risk modeling; k

[0068] Taking Events as input, to realize timely and accurate information support for decision-making in cross-border agricultural negotiations, the compression and verification operation of the event stream is specifically:

[0069] The similarity between each event is calculated by the redundancy eliminator, and the calculation formula is:

[0070]

[0071] where, Sim ij denotes the comprehensive similarity of event i and event j (used to determine whether to merge the events), its value range is [0, 1] (the larger the value, the more similar), i, j denote the unique identifiers of the two events to be compared, and i, j ∈ {1, 2,..., m} (m is the total number of events), λ1 denotes the weight of semantic similarity, which is a fixed value of 0.7, Embed i denotes the semantic embedding vector of event i, which is generated by the ontology-injected BERT (dimension 768), Embed j denotes the semantic embedding vector of event j, which is generated by the ontology-injected BERT (dimension 768), and cos(·,·) denotes the cosine similarity calculation,​ λ2 represents the weight of time proximity, which is a fixed value of 0.3, Time i denotes the occurrence timestamp of event i, Time j denotes the occurrence timestamp of event j, |·| denotes the absolute value of time difference, τ denotes the maximum valid time window, min(·,τ) denotes the time difference truncation processing, if the time difference exceeds τ, then take τ, if the event similarity exceeds the set threshold (Sim ij > 0.85), then these events are merged wherein denotes the event merging operator, that is, taking the maximum ImpactVal in the two events as the attribute of the new event, ImpactVal represents the event impact value, which is the quantified policy impact scale ([0,1]) in the execution of cross-language semantic alignment operation, to ensure the conciseness and accuracy of information, the calculation of similarity integrates the similarity of event semantics and the proximity in time;

[0072] Secondly, check whether there are mutually contradictory events (for example, two different message sources declare opposite policies in the same country), and determine which event is more credible through the credibility arbitration rule, and finally select the most credible event as the final output, the expression of the credibility arbitration rule is:

[0073]

[0074] wherein, E k denotes a single event in the conflict event set, whose data source is the event set to be verified generated in the previous step (Events), ConflictSet denotes the conflict event set of the same policy subject (for example: A country publishes tax increase / decrease policy events in the same period), whose data source is the conflict event group aggregated by the system in real time, C source denotes the data source authority (the higher the numerical value, the more reliable the source), whose value range is a floating point number [0,1], v denotes the timeliness decay factor, which is a fixed constant v=0.05, Recency denotes the event timeliness index (the larger the numerical value, the more recent the event), whose value range is a floating point number ≥0, the event generation timestamp calculation formula is: Recency=t current -t event , e -v·Recency denotes a time decay function (the older the event, the smaller the value), whose value range is a floating point number (0,1], denotes the normalized timeliness score (compressing timeliness to the interval (0,1)), E final denotes the final event retained after arbitration;

[0075] The output form of the structured policy event stream is:

[0076] P-Stream = {E k ′};

[0077] The conflict determination condition is:

[0078] Two events E i and E j are included in ConflictSet only when they satisfy the following conditions:

[0079] 1) Same subject (e.g. both "CountryA");

[0080] 2) Mutually exclusive actions (e.g. "raise tariffs" vs "reduce tariffs");

[0081] 3) Overlapping time windows (||t i -t j ||≤300s);

[0082] Arbitration priority ranking:

[0083] Final event selection is based on a composite credibility score = C source × timeliness score, with the higher score winning;

[0084] Credibility dominance: An authoritative media event (C source = 0.9) scores higher than a social media new event (C source = 0.3) even if delayed by 10 minutes;

[0085] Timeliness sensitivity: v = 0.05 means timeliness score decays by 50% every 200 seconds (≈3 minutes).

[0086] In this embodiment, it is specifically necessary to explain the dynamic network modeling module, and the specific operation of constructing the policy chain reaction network is:

[0087] A1, P-Stream = {E k}(policy subject k , regulatory behavior k , impact value k ) as an input object, and then map the policy subject to a specific sovereign state c i ∈C = {c1,..., c p}, by modeling the association between each event and country, generate the subject-event association matrix Γ ∈ {0, 1} p×m :

[0088]

[0089] Where γ ij represents the association between country c i and event Ej the association strength of c i , g(c j ) represents the implicit influence probability of agricultural trade dependence, with a value range of [0, 1], which identifies whether each policy subject is involved in an event or whether the event implicitly affects a certain country, in addition, a baseline adjacency matrix A (0) ∈R p×p :

[0090]

[0091] wherein, represents the elements in the baseline adjacency matrix A (0) , i.e., the initial trade dependence of c i to c j , i, j represent the country index subscript, T ij represents the annual import volume of agricultural products from the source country c i to the target country c j , max(T i* ) represents the row maximum value of T ij , which refers to the maximum value of the import volume of c i to all c j when i is fixed (non-global maximum), ζ represents the time decay factor, which is 0.05, used to control the decay strength of historical data on the current weight, t curr represents the current timestamp, t last represents the trade data update time corresponding to T ij , i.e., the latest statistical time point of T ij , ||·|| represents the absolute value operator, i.e., the absolute value of the time difference, e (·) represents the natural exponential function, which is used to realize the time decay effect (the larger t curr -t last , the smaller the weight), represents scaling the import volume by the maximum trade volume of the source country to avoid large country data from drowning small country relationships (e.g., China imports 10 9 million US dollars of American soybeans vs. 10 7 million US dollars of Cambodian rice), using the relative proportion within the row (non-global maximum) to retain the sensitivity of small-scale trade relationships; e -ζ·| t curr -t last | represents the worse time effectiveness of trade data (t last earlier), the initial edge weight exponentially decays, and its effect is to ensure that the trade data one year ago only retains e -0.05×365 ≈2×10 -8It emphasizes the dominance of recent data, represents trade relations between countries, and considers the impact of time changes on border rights through a time decay factor, thereby outputting the country set C and the baseline network A. (0) Event correlation matrix Γ;

[0092] A2, A (0) P-Stream and Γ are used as input objects for each event E. k Based on its impact value and historical trade data, an impact propagation matrix Π is generated. (k) ∈R p×p :

[0093]

[0094] in, This represents the event impact propagation value, used to quantify event E. k For node c i to c j Trade dependence shock intensity, γ ik This represents the policy subject-event correlation factor, used to identify country c. i With Event E k The direct correlation (1: direct correlation; 0: unrelated), Indicates the normalized impact value of the event. That is, event E k The relative strength of the original impact values ​​(extracted from the real-time data capture module), where ι represents the direct dependency weight, which is a fixed value of 0.7, used to control the contribution ratio of direct and indirect dependencies in the model. Representing the edge weight (historical state) in event E k Before it happened, the country c i For c j The trade dependence is defined by the value in the range [0, +∞), where h represents the intermediate node identifier, i.e., traversing all possible inter-country trade paths. This represents the indirect dependency transmission term, used to calculate the dependency through a third country c. h The potential policy impact on the intensity of transmission, ||a i* ||2 represents node c i The outward strength, i.e., for country c i Calculate the L2 norm of all outgoing edge weights, ||a *j ||2 represents node c j The strength of the incoming edge, i.e., the strength of the edge for country c j Calculate the L2 norm of all incoming edge weights. denotes the standardized neighbor covariance, which is used to eliminate the network structure bias and prevent high connectivity nodes from dominating the results, and its value range is [-1, 1], the matrix reflects the propagation effect of the influence between countries, and especially, the direct influence (through the relevance of the event) and the indirect influence (through the propagation of the historical trade network) are considered, after the event occurs, the edge weight will be updated according to the propagation matrix: the incremental network ΔA (k) :

[0095] wherein, denotes the edge weight increment of node i to node j after the kth event, which is used to quantify the dynamic change of the trade dependence between the two countries due to the occurrence of the policy event, a positive value indicates that the dependence is enhanced, and a negative value indicates that the dependence is weakened, denotes the event E k , and the impact strength of node (i, j) (from the impact propagation matrix Π (k) ) is used to measure the direct influence strength of a single event on the trade relationship of a specific country, which integrates the direct correlation and indirect conduction effect (matrix product term), κ represents a policy effect decay index, which is 1.2, which is used to control the speed of the decay of the policy influence over time: κ>1 indicates that the decay is faster than the exponential distribution, which reflects the timeliness of the policy, Δt k denotes the time difference of the event occurrence, Δt k = t curr -t k , that is, the difference between the current time t curr and the time t k when the event E k occurs, τ represents a half-life parameter, which is a time reference required for the policy effect to be halved: τ=30 days indicates that if there is no new event, the impact strength decays by 50% every 30 days, and both the decay of the policy effect and the time factor are considered;

[0096] A3, A (k) = A (k-1) + ΔA (k) (iterative update) is taken as the input object, and a policy sensitivity matrix ∑∈R p ×p :

[0097]

[0098] wherein, σ ij denotes the policy sensitivity of country c i to c j , which is used to quantify the conduction strength of the policy change between the two countries (value range [0, 1]), denotes the edge weight of the dynamic network at time stamp k, that is, the current trade dependence strength (for example: the dependence of the United States on Brazil soybean tariff), and The initial edge weights represent the historical baseline network, used as a reference point before policy shocks. This represents the element-wise summation operation of a row vector, used to calculate the number of countries, c. i The total change in dependence on all trading partners, p represents the total number of sovereign states used to determine the matrix dimension (e.g., p = 15 when negotiations involve 15 countries), and ε represents the smoothing coefficient to prevent division by zero, with a value of ε = 10. -5 This involves measuring the sensitivity of each country to policy changes, and then calculating the heterogeneity centrality Ω∈R for each country by weighting the active influence and passive sensitivity of the nodes. p :

[0099]

[0100] Where, ω i This represents the bimodal centrality value of node i, used to quantify the overall influence of the node (country). The larger the value, the more crucial it is in the policy transmission network. This represents an element of the dynamic adjacency matrix, with values ​​ranging from [0,1], used to describe the trade dependence strength of node i on node j after the k-th event iteration (from the real-time updated A). (k) ), where φ represents the power decay exponent, which is a fixed hyperparameter of 0.9, used to suppress the effects of weak connections: making The weak edge weights decay faster, highlighting core dependencies. λ represents the sensitivity amplification factor, a fixed hyperparameter of 1.5, used to enhance the weights of passive sensitivity. This addresses the problem of traditional centrality models neglecting the "degree of influence." σ ji The elements of the policy sensitivity matrix are represented by [0,1], which are used to describe the sensitivity of node i to policy changes of node j (the element in the j-th row and i-th column of ∑). p represents the total number of sovereign state nodes, i.e. the network size parameter, which is determined by the set of states C in the input event stream. Here, active influence refers to the direct influence of a state on other states, while passive sensitivity refers to the state's ability to respond to external policy shocks.

[0101] A4. The dynamic adjacency matrix A of the current iteration step. (k) Using the heterogeneity centrality Ω and policy sensitivity matrix ∑ as inputs, and the dynamic adjacency matrix, node centrality, and sensitivity matrix obtained from A1, A2, and A3, a core subnet is generated. Unimportant edges and nodes are filtered out, and the core subnet θ = median(A) is selected based on a set threshold. (k) ):

[0102]

[0103] in, A represents (k) The element, namely the country c in the k-th step.i to c j The edge weights, ω, directly participate in threshold judgment and assignment logic. i The element representing Ω, i.e., country c i The centrality strength is used to participate in the second branch condition judgment. Furthermore, Ω represents the node centrality vector, used to determine whether a node is "qualified to indirectly influence". σ represents the 75th percentile of Ω, used to determine whether a node belongs to a high-influence group. ij The element of ∑ is c. i For c j The policy sensitivity, with a value range of [0,1], functions in the second branch in relation to ω. i The product, serving as a new weight for low-strength connections, represents the policy sensitivity matrix, used to quantify the impact of policy changes on the relationships between nodes. θ represents the edge weight filtering threshold, calculated as θ = median(A (k) This function distinguishes between strong and weak connections. Its purpose is to dynamically set the threshold value for distinguishing "strong connections" (taking the median of all edge weights to avoid interference from outliers).

[0104] Based on the above formula, the core subnet generation rule is: in the first branch (strong connection preservation), when At that time, retain the original edge weights. The purpose is to maintain the authenticity of high dependencies. In the second branch (high-influence node correction), if the edge weight is insufficient to reach the threshold... But the starting node c i Centrality Supergroup Then use σ ij ·ω i The purpose of replacing the original edge weights is to weight low-intensity relationships that may trigger cascading effects (node ​​influence × relationship sensitivity). In the third branch (weak link removal), values ​​are set to zero when the above conditions are not met. This is to compress noisy data, highlight key topological structures, and construct a topological relationship matrix T∈R. (p+2)×(p+2) :

[0105]

[0106] in, The matrix represents the core adjacency matrix, i.e., the dynamic trade dependencies between countries. It is used to preserve the connections between important nodes and reduce noise interference. Ω represents the centrality vector, used to quantify the active influence of each country in policy dissemination. Its role is to identify key policy nodes (such as major trading nations) and provide a priority basis for strategy optimization. 0 represents the zero vector, i.e., a placeholder, to maintain matrix dimension alignment. Its role is to maintain the consistency of the matrix structure. ψdenotes the transpose of the centrality vector, which serves to symmetrize the node influence indicator, and denotes the network variability strength, and (k) || F , which measures the degree of global network disturbance caused by policy events, and ψ denotes the transpose of the zero vector, which is used to maintain the symmetry of the matrix structure; the matrix contains the structural information of the dynamic network, reflecting the policy influence and interdependence between countries.

[0107] In this embodiment, the risk prediction engine module specifically needs to be described, and the specific operation of receiving the network topology relationship matrix and the historical policy time series data is as follows:

[0108] The (exported from the dynamic network modeling module), X hist = [x t-1 , x t-2 ,..., x t-τ ](historical event feature vector, x t = <influence, policy type>) is taken as an input object, and denotes the time window length: control model backtracking period (example: τ = 30 days), influence time series feature dimension, relevant features are extracted from the input topology structure, and are combined with historical event data to form a spatio-temporal joint input tensor, which specifically includes extracting the upper triangular elements u ∈ R p(p-1)2 (from the core adjacency matrix , removing redundant symmetry), and combining with the historical event features, disturbance history similarity, and spatial network features to construct a new spatio-temporal input tensor Z ∈ R τ×d :

[0109] Z[:,1:g] = X hist

[0110]

[0111] Wherein, g represents the historical feature dimension, and the calculation method is dim(x t ) = 2 (influence value + policy type), Z[:,1:g] represents a τ × g matrix, which is used to inject original historical event data, Z[:,g+1] represents a τ × 1 vector, i.e. disturbance-history similarity: Ω hist denotes the cosine similarity weighted disturbance strength ρ of the centrality vector of the last time; Z[:,g+2:d] represents a matrix, i.e. network topology broadcast: extending the static network features to the time series dimension; 1 τ denotes a vector of all 1s, and d represents the total feature dimension, which is calculated as denotes vector broadcast concatenation for making static network data available for the time series model, and <.,.> denotes vector inner product for quantifying the time correlation of centrality vectors;

[0112] The specific operation of calling the time series prediction model of network correction is as follows:

[0113] The spatio-temporal input tensor is taken as an input, and the spatio-temporal joint input tensor is processed through a double-flow gating mechanism. On the one hand, the time flow uses an LSTM unit to process the historical time series Z[:,1:g], and on the other hand, the space flow uses a graph convolution network (GCN) to process the static topological features Z[:,g+2:d]. Through the gating mechanism, the fusion mode of each time step t is determined, and the sensitivity of the model to spatio-temporal features is adjusted, especially the model adaptation capability when processing disturbances. The expression is as follows:

[0114]

[0115] wherein, denotes the time flow input slice, i.e., the pure historical event features processed by the LSTM, and the value range is Z[t,1:g]∈R g , c t-1 denotes the historical state memory, denotes the time flow output state, i.e., the time series pattern features extracted by the LSTM, denotes the space flow input slice, i.e., the network topological features processed by the GCN, and the value range is Z[t,g+2:d]∈R d-g-1 , ρ denotes the network disturbance intensity for quantifying the global influence degree of policy events, and the value range is [0, +∞), W s denotes the spatial transformation weight for projecting the topological features to the state space, denotes the space flow output state, i.e., the network structure features extracted by the GCN, W g denotes the learnable gating weight for calculating the disturbance gating coefficient, and σ denotes the sigmoid function for generating the space flow dominant weight [0, 1], and ⊙ denotes Hadamard product for realizing element-wise weighted fusion of state vectors, c t denotes the fusion state output, i.e., the dynamic combination result of spatio-temporal features;

[0116] In addition, the dynamic regulation mechanism is as follows:

[0117] The ρ-driven gate: σ(W g ρ) converts the disturbance intensity into a fusion weight;

[0118] This design solves the pain points of traditional models: the static fusion ratio cannot respond to sudden policy disturbances.

[0119] The specific steps of performing the joint training operation are:

[0120] The fused state c t As input, the fused state is mapped to different risk dimensions, such as customs, quarantine, and logistics, by a decoder, and its expression is:

[0121]

[0122] where c t represents a spatio-temporal fusion state vector, i.e., encoding the spatio-temporal features of the policy risk at the current time, v cent represents the centrality column vector of the topology matrix T, i.e., injecting the prior knowledge of key node influence, || represents vector concatenation, i.e., fusing dynamic state and static topology features, W d represents a learnable decoding weight, i.e., mapping the fused features to the risk dimension, Softmax(·) represents probability normalization, i.e., generating a probability distribution of three risk categories, with a value range of [0, 1] 3 , Φ represents a risk dimension scaling matrix, and φ i represents a learnable scaling factor, used to control the sensitivity of customs / quarantine / logistics risk, and φ i > 0 (optimized during training), r t (·) represents the customs / quarantine / logistics risk value, used to quantify the intensity of policy risk in negotiations in each dimension;

[0123] and are jointly optimized through multi-task learning, and the loss function combines mean square error (MSE) and KL divergence to minimize the gap between the model output and the target value, and the KL divergence is used to constrain the gradient consistency of the risk distribution, and the expression of the loss function is:

[0124]

[0125] where L represents the joint loss function, used to guide the optimization of model parameters, r t represents the model predicted risk value, i.e., the model output to be optimized, represents the true risk label, used as a benchmark for supervised learning, MSE(·) represents the mean square error, used to constrain the absolute error between the predicted value and the label, α, β represent the loss term weight coefficients, used to balance the prediction accuracy and distribution consistency, and α = 0.8, β = 0.2 (adjustable), KL(·||·) represents the KL divergence, used to measure the difference between the risk distribution and the perturbation gradient, represents the total risk score represents the gradient of risk to perturbation intensity, i.e., characterizes the rate of change of global risk with respect to network perturbation ρ, in addition, the KL divergence term forces that high-risk categories (such as tariffs) will significantly dominate the total risk change (i.e. is positively correlated with ), to address the lagging problem of traditional models in responding to sudden policy changes; in addition, by injecting the centrality vector as prior knowledge into the decoder, the weight of the risk of key nodes is enhanced;

[0126] The multi-dimensional policy risk score vector specifically includes:

[0127] Based on the fused risk vector Generate a time-sensitive risk vector R t :

[0128] R t = r t + λ·Δ causal ;

[0129] wherein R t represents the final risk vector: a real-time risk score of three dimensions of tariffs, quarantine, and logistics, which serves to output the core indicators for the strategy optimization module, r t represents the basic risk vector: which serves to represent the original risk prediction without event correction at the current time, λ represents the perturbation adaptive weight, and λ = 1-e -ρ , wherein ρ represents the current perturbation intensity, which measures the overall change of the network caused by policy events, e -ρ represents the negative exponential function, which ensures that λ ∈ [0, 1): 1、 2、 λ serves to dynamically adjust the correction weight according to the perturbation intensity ρ Δ causal represents the historical perturbation event convolution, and wherein Conv1D(·) represents a one-dimensional causal convolution, which serves to extract the time-dependent pattern in historical data, r t-i represents the historical risk vector: which serves to provide risk data at the past k time points, I(·) represents the event screening condition: which serves to select only significant perturbation events for calculation, represents the historical perturbation gradient, and The role is to quantify the disturbance intensity change rate at t-i time, k represents the length of the sliding window, which plays the role of controlling the historical backtracking range (default k = 10, about 30 minutes), this process combines the current risk vector and the information of historical disturbance events, processes the disturbance history through convolution operation, and adjusts the weight according to the disturbance intensity, this process ensures the strict consistency between time steps, so as to improve the accuracy and stability of the prediction.

[0130] In this embodiment, the specific need to be explained is the strategy optimization generation module, and the specific operation of constructing the multi-objective optimization decision model is:

[0131] Based on the multi-dimensional risk score data of the time-sensitive risk vector and the preset negotiation constraint condition, a dynamic feasible region F(R t ) containing rigid constraint boundary and risk-driven elastic constraint is constructed:

[0132]

[0133] Wherein, F(R t ) represents the dynamic feasible region, which is used to define the current acceptable negotiation term range, x represents the decision variable: negotiation term vector (such as price, delivery period, payment method, etc.), R n represents n-dimensional real number space, A represents constraint coefficient matrix (m x n), b com represents rigid constraint boundary (m x 1), represents vector inequality, that is, each element of Ax is ≤b com corresponding element, x0 represents initial negotiation term (n x 1), for example, enterprise preset benchmark term (historical agreement or cost optimal solution), ||·||1 represents L1 norm, δ(R t ) represents risk gradient adaptive relaxation factor, and κ represents risk sensitivity coefficient, which is 0.8, δ0 represents initial constraint radius, and δ0>0, e · represents exponential function, and ▽R t represents risk vector gradient: ||·||2 represents L2 norm, and ||R t ||2 represents risk mutation intensity → the greater the risk, the greater the allowed negotiation term deviation; wherein the boundary radius of the elastic constraint is adaptively adjusted according to the gradient change of the risk vector, and the adjustment amplitude is negatively related to the risk mutation intensity, forming a dynamic scalable decision space;

[0134] The specific operation of generating English negotiation strategy script is:

[0135] An interactive event state diagram is constructed by the Pareto optimization solution set and historical negotiation data, and the expression of the Pareto optimization solution set is:

[0136]

[0137] where P * denotes the Pareto optimal solution set, x denotes the negotiation clause vector, F denotes the dynamic feasible region, μ denotes the economic objective weight, and where ζ denotes the Sigmoid function steepness coefficient, ρ denotes the network disturbance intensity, e -· denotes the nonlinear transformation, f econ denotes the economic objective function, and f safe denotes the safety objective function, and where x vol denotes the agricultural product transaction volume, Δt deliv denotes the logistics time deviation, max(f*) denotes the historical maximum value of the objective function, and

[0138]

[0139] where S k+1 denotes the k+1 step interaction state (node) ∈ {offer, clause negotiation, concession, signing}, s j denotes the next interaction state to be selected, denotes the weight coefficient, and (default 0.6), P(s j |s k ) denotes the historical transition probability of s k → s j , which is obtained by training historical negotiation data, denotes the negotiation strategy vector ∈ R n corresponding to the candidate state s j , such as price, delivery time, and other parameters, denotes the Pareto optimal solution set; and the path and the real-time risk change amount are input into the semantic fusion generator, and the expression is:

[0140]

[0141] wherein Script denotes the final generated English negotiation strategy script, denotes a temporal concatenation operation: merging T interaction steps templates, and denotes a semantic fusion generator: performing natural language synthesis and logical verification, denotes an interaction state s k corresponding base clause template library (e.g. denotes a tensor concatenation operation: horizontally connecting templates with risk modification information, and denotes a Sigmoid activation function, W r denotes a risk projection matrix, and t,k denotes a real-time risk change amount, and wherein Δt k denotes a time window length of the kth interaction, dynamically modifies the base templates in combination with the cultural conflict detection mechanism, and finally outputs an English negotiation strategy script that conforms to cross-border legal norms and cultural adaptation.

[0142] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0143] Those skilled in the art should understand that embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0144] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0145] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0146] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0147] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such additional variations and modifications as fall within the scope of the application.

[0148] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. An agricultural economic cross-border cooperation English negotiation auxiliary decision-making system, characterized in that, Specifically comprising: The real-time data capture module, the dynamic network modeling module, the risk prediction engine module and the strategy optimization generation module; The real-time data capture module: in response to a policy update trigger event or a user negotiation request, capturing multi-source policy data streams and trade-related data, performing cross-language semantic alignment operations, and generating a structured policy event stream containing standardized triples; The dynamic network modeling module: based on the structured policy event stream, a policy chain reaction network is constructed, which takes participating countries as nodes and trade dependence as dynamic edge weights, calculates node centrality indicators, and outputs a network topology relationship matrix; The risk prediction engine module: receiving the network topology relationship matrix and historical policy time series data, calling a network-corrected time series prediction model, performing joint training operations, and generating a multi-dimensional policy risk score vector containing time-sensitive risk variables; The strategy optimization generation module: based on the multi-dimensional policy risk score vector and the preset negotiation constraint conditions, a multi-objective optimization decision model is constructed, a constraint satisfaction solving operation is performed, and an English negotiation strategy script is output, which contains clause revision suggestions and interactive timing logic.

2. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 1, characterized in that: The capture step of the multi-source policy data stream in the real-time data capture module is: First, capture external policy update events or user's active request signals, and use them as trigger conditions to build a dynamic data source crawling queue. Each data source calculates its priority based on the source's authority and historical response delay, as well as the degree of association with specific keywords, and initiates requests through a parallel crawling engine according to priority to output raw policy data streams; The multi-source policy data stream in the real-time data capture module includes government API announcement data, agricultural economic news streams, and international organization policy library data.

3. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 2, characterized in that: The specific operation of performing the cross-language semantic alignment operation is: For the captured raw data, use heterogeneous parsers for classification processing. Specifically: for structured data, extract key-value pairs; for unstructured text, use "agricultural ontology injection BERT" for semantic role labeling, use an ontology knowledge base specific to the agricultural domain for cross-language semantic analysis, and generate triples for events; Each event is converted into an atomic event after semantic role labeling, and its impact value is quantified. At the same time, a timestamp and a spatial stamp are added to all generated events to form a structured event set, The compression and verification operation of generating event streams is specifically: Calculate the similarity between each event through a redundancy eliminator. If the event similarity exceeds a certain threshold, these events will be merged. The similarity calculation combines the similarity of event semantics and the proximity in time; Secondly, check if there are mutually contradictory events, and determine which event is more credible through credibility arbitration rules, and finally select the most credible event as the final output.

4. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 3, characterized in that: In the dynamic network modeling module, the specific operation of constructing the policy chain reaction network is: A1, mapping policy subjects to specific sovereign states, generating a subject-event association matrix by modeling the association between each event and state, in addition, generating a baseline adjacency matrix based on historical trade data to represent trade relationships between countries, and considering the impact of time changes on edge weights through a time decay factor; A2, for each event, generate an impact propagation matrix based on its impact value and historical trade data, after the event occurs, the edge weight will be updated according to the propagation matrix, while considering the attenuation of policy effect and time factor; A3, construct a policy sensitivity matrix to measure the sensitivity of each country to policy changes, then calculate the heterocentricity of each country by weighting the active influence and passive sensitivity of the node; A4, generate a core subnet by using the dynamic adjacency matrix, node centrality and sensitivity matrix obtained by A1, A2 and A3, filter out unimportant edges and nodes, select the core subnet according to the set threshold, and construct a topology relationship matrix; The matrix contains the structure information of the dynamic network.

5. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 4, characterized in that: In the risk prediction engine module, the specific operation of receiving the network topology relationship matrix and the historical policy time series data is: Extract relevant features from the input topology structure, and combine them with historical event data to form a spatio-temporal joint input tensor, which specifically includes extracting the upper triangular elements from the core adjacency matrix, and combining them with historical event features, similarity of disturbance history and spatial network features to construct a new spatio-temporal input tensor.

6. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 5, characterized in that: The specific operation of calling the network corrected time series prediction model is: Process the spatio-temporal joint input tensor through a double-flow gating mechanism, on the one hand, the time flow uses LSTM unit to process the historical time series, on the other hand, the space flow uses graph convolution network to process the static topology features, through the gating mechanism, the fusion mode of each time step is determined, and then the sensitivity of the model to spatio-temporal features is adjusted.

7. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 6, characterized in that: The specific steps of performing joint training operation are: The fused state is mapped to different risk dimensions through the decoder, and is optimized through multi-task learning, the loss function combines mean square error and KL divergence to ensure that the difference between the model output and the target value is minimized, and the gradient consistency of the risk distribution is constrained through KL divergence.

8. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 7, characterized in that: The multi-dimensional policy risk score vector specifically includes: Based on the fused risk vector, a time-sensitive risk vector is generated, which combines the current risk vector and the information of historical disturbance events, processes the disturbance history through convolution operation, and adjusts the weight according to the disturbance intensity.

9. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 8, characterized in that: The specific operation of constructing a multi-objective optimization decision model is: Based on the multi-dimensional risk score data of the time-sensitive risk vector and the preset negotiation constraint condition, a dynamic feasible region containing rigid constraint boundary and risk-driven elastic constraint is constructed; wherein the boundary radius of the elastic constraint is adaptively adjusted according to the gradient change of the risk vector, and the adjustment amplitude is in a negative correlation relationship with the risk mutation intensity, forming a dynamically scalable decision space.

10. The agro-economic cross-border cooperation English negotiation aided decision system according to claim 9, characterized in that: The specific operation of generating an English negotiation strategy script is: Construct an interactive event state diagram through the Pareto optimization solution set and historical negotiation data, and generate an interactive path by using the weighted calculation of state transition probability and solution set approximation degree; The path is input into a semantic fusion generator combined with a real-time risk change amount, and a culture conflict detection mechanism is used to dynamically modify the basic template, and finally an English negotiation strategy script that conforms to cross-border legal norms and cultural adaptation is output.

Images