An information propagation risk assessment method and device based on agricultural network information

Abstract

This disclosure discloses a method and apparatus for assessing information dissemination risks based on agricultural network information, belonging to the field of information processing. In this embodiment, spatiotemporal phenological feature information is extracted from the agricultural network information to be analyzed, and a semantic feature tensor is constructed based on this information. The phenological deviation of the semantic feature tensor is calculated based on a pre-constructed agricultural phenological knowledge base. The semantic feature tensor is input into a pre-constructed classification model to determine a semantic compliance score and obtain the corresponding classified event entity. The classified event entity is input into a pre-determined potential energy propagation model to obtain risk potential energy information of the classified event entity on the causal graph of the agricultural industry chain. Based on this risk potential energy information, the systemic risk of the classified event entity across the entire agricultural industry chain is determined. The phenological deviation, semantic compliance score, and systemic risk across the entire industry chain are weighted to obtain a comprehensive risk score for the agricultural network information.

Description

Technical Field

[0001] This article relates to information processing technology, and in particular to a method and device for assessing the risk of information dissemination based on agricultural network information. Background Technology

[0002] The network information intelligent analysis technology in related technologies mainly focuses on: 1) general classification and feature fusion, that is, using deep learning models (such as BERT, LLM) to perform semantic encoding and classification of text, and processing feature extraction, classification, sentiment analysis and other tasks in parallel to improve the overall classification accuracy; 2) knowledge assistance and logical constraints, that is, introducing knowledge graph (KG) tools, using mathematical tools such as statistical models to calculate, comparing the deviation between the model prediction results and the ideal results of the knowledge graph, and using the deviation as a constraint term to correct the output, so as to ensure the logical rationality of the output results of the text analysis model; 3) spatiotemporal features and dynamic classification mechanisms, using geographic coordinates and time windows to match the spatiotemporal attributes and dynamic classification system contained in the text information, and realizing some more complex spatiotemporal indexes.

[0003] The above methods have the following limitations when applied to the agricultural field: 1) They cannot address the spatiotemporal sensitivity of agricultural online information: Agricultural online information is highly sensitive in time and space. The same information may have vastly different effects at different times and in different places. For example, agricultural technical guidance information on the internet may be "high-value information" during the busy farming season, but "junk information" or even "misleading information" during the fallow season. Related technologies lack a mechanism to combine the semantics of agricultural online information with agricultural seasons, phenology, and geographical location. 2) They cannot adapt to the complex scenarios of agricultural business risk transmission: Due to the numerous links in the agricultural industry chain and the complexity of the involved links, the risks reflected in agricultural online information also have significant chain reactions, which may trigger incidental risks in other areas, thus causing a series of public opinion events. The related technologies, which only score and assess the risk of a single text, cannot quantify its potential risk ripples on the entire industry chain, and therefore have significant limitations.

[0004] In summary, how to effectively process agricultural network information and conduct industry risk analysis remains an unresolved issue. Summary of the Invention

[0005] To address the limitations of existing network information analysis methods in analyzing the spatiotemporal sensitivity of agricultural network information, their inability to adapt to complex scenarios of agricultural business risk transmission, and the significant limitations of single-text scoring and risk assessment methods in quantifying the potential risk ripples on the entire industry chain, thus failing to achieve effective industry risk analysis of agricultural network information dissemination, this application provides a method and apparatus for risk assessment of information dissemination based on agricultural network information.

[0006] To achieve the above objectives, embodiments of this application provide a method for assessing the risk of information dissemination based on agricultural network information, including: Spatiotemporal phenological features are extracted from the agricultural network information to be analyzed, and a semantic feature tensor is constructed based on the extracted spatiotemporal phenological features. The spatiotemporal phenological features include: semantic vectors, and the time and space entities involved in the semantic vectors. The phenological deviation of the semantic feature tensor is calculated based on a pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. The constructed semantic feature tensor is input into a pre-constructed classification model to determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor. The semantic compliance score is the sum of the probabilities that the semantic feature tensor falls into the risk category. Input the classified event entity corresponding to the semantic feature tensor into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entity on the pre-constructed causal graph of the agricultural industry chain. Based on the risk potential energy information, determine the systemic risk of the classified event entity in the entire agricultural industry chain. By weighting phenological deviation, semantic compliance score, and systemic risk across the entire industry chain, a comprehensive risk score for agricultural network information is obtained.

[0007] On the other hand, embodiments of this application also provide a computer storage medium storing a computer program, which, when executed by a processor, implements the above-described information dissemination risk assessment method based on agricultural network information.

[0008] Furthermore, embodiments of this application also provide a risk assessment device, including: an extraction and processing unit, a deviation calculation unit, a classification analysis unit, a potential risk analysis unit, and a comprehensive scoring unit; wherein, The extraction and processing unit is set to extract the spatiotemporal phenological feature information from the agricultural network information to be analyzed, and construct a semantic feature tensor based on the extracted spatiotemporal phenological feature information. The spatiotemporal phenological feature information includes: semantic vectors, and the time entities and spatial entities involved in the semantic vectors. The deviation calculation unit is set to calculate the phenological deviation of the semantic feature tensor based on the pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. The classification analysis unit is set up as follows: input the constructed semantic feature tensor into the pre-constructed classification model, determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor, wherein the semantic compliance score is the sum of the probabilities of the semantic feature tensor falling into the risk category; The potential energy risk analysis unit is set up as follows: inputting the classified event entity corresponding to the semantic feature tensor into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entity on the pre-constructed causal graph of the agricultural industry chain, and determining the systemic risk of the classified event entity in the entire agricultural industry chain based on the risk potential energy information. The comprehensive scoring unit is set up to calculate the comprehensive risk score information of agricultural network information by weighting the phenological deviation, semantic compliance score and systemic risk of the entire industry chain.

[0009] This disclosure considers the different effects of the same agricultural information at different times and spaces, and combines the semantics of network information with temporal and spatial entities to construct a semantic feature tensor of agricultural information with spatiotemporal constraints. Based on phenological deviation, the probability of agricultural information being false information is calculated for the agricultural information corresponding to the semantic feature tensor. A semantic compliance score is determined through a classification model, realizing the probability prediction of agricultural risks. The classified event entities corresponding to the semantic feature tensor are processed through a potential energy propagation model to quantify the potential risk ripples of classified event entities on the entire agricultural industry chain. By combining phenological deviation, semantic compliance score, and systemic risk of the entire industry chain, a comprehensive risk assessment of agricultural network information applicable to complex scenarios of agricultural business risk transmission is realized.

[0010] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. Other advantages of this application can be realized and obtained by means of the solutions described in the description and the accompanying drawings. Attached Figure Description

[0011] The accompanying drawings are used to provide an understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0012] Figure 1 This is a flowchart of an embodiment of the information dissemination risk assessment method based on agricultural network information in this disclosure; Figure 2 This is a structural block diagram of an information dissemination risk assessment device based on agricultural network information, as described in this embodiment of the disclosure. Figure 3 This is a schematic diagram illustrating the process of risk assessment based on agricultural network information in an application example. Detailed Implementation

[0013] This application describes several embodiments, but these descriptions are exemplary and not limiting, and it will be apparent to those skilled in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with or in lieu of any other feature or element in any other embodiment.

[0014] This application includes and contemplates combinations of features and elements known to those skilled in the art. The embodiments, features, and elements disclosed in this application can also be combined with any conventional features or elements to form unique inventive solutions. Any feature or element of any embodiment can also be combined with features or elements from other inventive solutions to form another unique inventive solution. Therefore, it should be understood that any feature shown and / or discussed in this application can be implemented individually or in any suitable combination. Therefore, the embodiments are not limited except by the limitations imposed by the appended claims and their equivalents. Furthermore, various modifications and changes can be made within the scope of the appended claims.

[0015] Furthermore, in describing representative embodiments, the specification may have presented methods and / or processes as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that it does not depend on such a specific order. As will be understood by those skilled in the art, other sequences of steps are also possible. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Moreover, the claims concerning the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of this application.

[0016] Figure 1 This is a flowchart of an embodiment of the information dissemination risk assessment method based on agricultural network information, as shown below. Figure 1 As shown, it includes: Step 101: Extract the spatiotemporal phenological features from the agricultural network information to be analyzed, and construct a semantic feature tensor based on the extracted spatiotemporal phenological features. The spatiotemporal phenological features include: semantic vectors, and the time and space entities involved in the semantic vectors. Step 102: Calculate the phenological deviation of the semantic feature tensor based on the pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. Step 103: Input the constructed semantic feature tensor into the pre-constructed classification model, determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor, wherein the semantic compliance score is the sum of the probabilities of the semantic feature tensor falling into the risk category; Step 104: Input the classification event entity corresponding to the semantic feature tensor into the pre-determined potential energy propagation model to obtain the risk potential energy information of the classification event entity on the pre-constructed causal graph of the agricultural industry chain, and determine the systemic risk of the classification event entity in the entire agricultural industry chain based on the risk potential energy information. Step 105: Perform weighted calculations on phenological deviation, semantic compliance score, and systemic risk of the entire industry chain to obtain comprehensive risk score information for agricultural network information.

[0017] This disclosure considers the different effects of the same agricultural information at different times and spaces, and combines the semantics of network information with temporal and spatial entities to construct a semantic feature tensor of agricultural information with spatiotemporal constraints. Based on phenological deviation, the probability of agricultural information being false information is calculated for the agricultural information corresponding to the semantic feature tensor. A semantic compliance score is determined through a classification model, realizing the probability prediction of agricultural risks. The classified event entities corresponding to the semantic feature tensor are processed through a potential energy propagation model to quantify the potential risk ripples of classified event entities on the entire agricultural industry chain. By combining phenological deviation, semantic compliance score, and systemic risk of the entire industry chain, a comprehensive risk assessment of agricultural network information applicable to complex scenarios of agricultural business risk transmission is realized.

[0018] In one exemplary embodiment, the agricultural phenology knowledge base can be constructed using online news about agriculture as source data, but is not limited to any agricultural phenology knowledge base constructed based on agricultural information.

[0019] In this exemplary embodiment, the data structure of the agricultural phenology knowledge base can be: , C represents the set of agricultural production entities, R represents the set of geographical regions, T represents the time window, S represents the growth stage (such as the jointing stage and the grain-filling stage), and E represents the suitable environmental feature vector of the growth stage. The data structure in this exemplary instance may be, but is not limited to, the above composition, and may be obtained by technicians through processing based on the compositional characteristics of agricultural network information.

[0020] In this exemplary embodiment, an agricultural phenology knowledge mapping is defined by constructing an agricultural phenology knowledge base.

[0021] In one exemplary embodiment, spatiotemporal phenological feature information is extracted from the agricultural network information to be analyzed, and a semantic feature tensor is constructed based on the extracted spatiotemporal phenological feature information, including: The BERT model is used to extract the basic semantic vectors and the temporal and spatial entities involved in the semantic vectors from the agricultural network information to be analyzed. The temporal and spatial entities involved in the extracted semantic vectors are weighted using a pre-defined weight matrix. The extracted semantic vectors and the temporal and spatial entities involved in the weighted semantic vectors are multimodal spliced ​​and fused through a layer normalization model to form a semantic feature tensor for constructing agricultural network information with spatiotemporal constraints.

[0022] The feature suppression caused by the difference in dimensions between temporal and spatial entities relative to semantic vectors (for example, temporal entities are generated using trigonometric functions, and their numerical range is relatively small; if directly fused with semantics, they will be "submerged" by the large values ​​in the semantic vector, reducing the weight of the temporal vector in numerical analysis) is addressed by using a layer normalization model (LayerNorm). This model solves the problem of temporal and spatial features being submerged by large semantic values ​​due to their small dimensions, ensuring that the three types of features have the same weight during training. This ensures that phenological constraint features and geospatial features receive equal attention as textual features during training, effectively solving the problem of inconsistent dimensions when fusing multi-source heterogeneous agricultural features, and realizing the engineered processing of weak signals of phenological constraint features in agricultural texts. The above-described method of constructing semantic feature tensors is only an exemplary embodiment of this application. The embodiments disclosed in this application include, but are not limited to, constructing semantic feature tensors in the above manner. The composition of the semantic feature tensor can also be changed and adjusted by those skilled in the art according to the content of the agricultural network information to be analyzed.

[0023] This embodiment of the disclosure constructs a semantic feature tensor based on agricultural network information, consisting of semantic vectors and the temporal and spatial entities involved. It puts semantics, time, and space into the same learnable representation, providing a foundation for subsequent integrated inferences such as phenological deviation, semantic compliance scoring, and systemic risks across the entire industry chain.

[0024] In one exemplary embodiment, the semantic feature tensor The expression can be, but is not limited to: ; In the formula, This indicates a splicing operation. It represents the Hadamah accumulation. , This represents the learnable weight matrix. For semantic feature tensors, As the initial semantic vector, It is a time entity (a time vector with phenological constraints). It is a spatial entity (a geographic vector with spatial characteristics).

[0025] Before performing multimodal concatenation and fusion of the extracted semantic vectors and the temporal and spatial entities involved in the semantic vectors using a layer normalization model (LayerNorm), this exemplary embodiment method further includes: Using sinusoidal positional coding, time entity codes are mapped to periodic vectors; In this context, the position of the sinusoidal position code is the index value after the time entity is mapped to the agricultural cycle, the dimension of the sinusoidal position code is the vector index value of the time entity to be encoded and mapped, and the total dimension of the sinusoidal position code vector is the number of solar terms included in the predetermined agricultural cycle.

[0026] The above encoding mapping is only an exemplary embodiment of this application. This exemplary embodiment can encode and map time entities in other ways, as long as normalization processing can be achieved. In related technologies, time entities are generally timestamp encoded. This exemplary embodiment uses sinusoidal positional encoding to first map the time entity to an "agricultural cycle / solar term index," and then uses sinusoidal positional encoding to form a periodic vector. When the total dimension equals the number of solar terms in the agricultural cycle (e.g., 24), the seasonal patterns of agriculture can be hard-coded into the representation space, obtaining specific constraints in the agricultural field corresponding to agricultural network information, distinct from general time series. In this exemplary embodiment, when the agricultural cycle includes 24 solar terms, the total dimension of the sinusoidal positional encoding vector is 24, and t is the index value of the time entity after mapping to the agricultural cycle. For example, if the date value of the time entity is February 4th, and the agricultural cycle is set to 24 solar terms, then the time entity can be mapped to the "Beginning of Spring" solar term in the agricultural cycle. Since "Beginning of Spring" is the first of the 24 solar terms, the value of t is 1 at this time.

[0027] In one exemplary embodiment, before multimodal splicing and fusion of the extracted semantic vectors and the temporal and spatial entities involved in the semantic vectors using a layer normalization model (LayerNorm), the method in this exemplary embodiment further includes: Map spatial entities to spatial vectors that include latitude, longitude, and climate zone attributes.

[0028] This exemplary embodiment, through the encoding and mapping of time and space information, can not only normalize spatiotemporal data, but also ensure that the encoding of spatiotemporal data remains smooth and continuous in the agricultural cycle, eliminating boundary abrupt changes and making it more in line with the natural law of agricultural cycles, thus ensuring the continuity and unity of the extracted time and space entities.

[0029] In one exemplary embodiment, for each semantic feature tensor, its phenological deviation is obtained through the following processing: Identify the entities that the semantic feature tensor matches in the agricultural phenology knowledge base; for each matched entity, calculate the deviation between the semantic feature tensor and the entity; compare the calculated deviation values ​​and determine the maximum deviation value as the phenological deviation.

[0030] In this exemplary embodiment, after an entity is found in the agricultural phenology knowledge base, the deviation between the semantic feature tensor and the found entity is calculated, and the maximum deviation is taken as the phenological deviation of the semantic feature tensor. Based on the agricultural phenology knowledge base, the phenological deviation of the constructed semantic feature tensor is calculated to obtain the first layer of risk features of agricultural network information.

[0031] In one exemplary embodiment, a classification model based on a continuous neural tree can be constructed, but is not limited to, in the following ways: An initial continuous neural tree classifier is constructed based on an existing set of agricultural ontology (which may be referenced, but is not limited to, books such as the "Thesaurus of Chinese Agricultural Science and Technology"), with a depth of L. The initial set of all leaf nodes of the neural tree classifier ( Based on semantics, risk types are classified into security categories. Subsets and risk types are risk categories ( (A subset of the data is used to obtain a classification model.)

[0032] This exemplary embodiment constructs a continuous neural tree classifier using an agricultural ontology set in the agricultural field, and then divides it into a subset of safety categories and a subset of risk types, forming a tree classification label system specific to the agricultural field.

[0033] In one exemplary embodiment, the classification event entity corresponding to the semantic feature tensor can be obtained, but is not limited to, through the following processing: The semantic feature tensor is input into the neural tree classifier of the classification model to determine the routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier. Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the safe category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the safe category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. For the set of leaf nodes of the neural tree classifier, calculate the uncertainty value of the semantic feature tensor in the neural tree classifier based on the probability that the semantic feature tensor enters the subset of the safe category; When the text set in the semantic feature tensor When the first preset condition is met, the classification label value corresponding to the leaf node with the smallest uncertainty value in the leaf node set of the semantic feature tensor is taken as the classification event entity corresponding to the semantic feature tensor. The first preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window.

[0034] Compared to related technologies such as hard decision trees or tree index-based post-selection models, this exemplary embodiment is based on a continuous neural tree classifier. It calculates the routing probability for each leaf node, and obtains the leaf node probability by accumulating the routing probabilities along the path, thus enabling the probability calculation of subsequent risk and safe subsets. The uncertainty is calculated based on the probability distribution of entering the safe subset, and then combined with a splitting threshold within a time window to form a first preset condition, thereby determining the classification event entity corresponding to the semantic feature tensor.

[0035] In this exemplary embodiment, the first preset condition may be, but is limited to, ,in, This represents the uncertainty value of the semantic feature tensor in the neural tree classifier. To determine the splitting threshold within a preset time window, a semantic vector of the semantic feature tensor is constructed, consisting of k texts, where k is the number of texts contained in the semantic vector. In this exemplary embodiment, Can be determined by a person skilled in the art based on the uncertainty value set up.

[0036] In this exemplary embodiment, the routing probability of the semantic feature tensor reaching the leaf node of the neural tree classifier is determined by a pre-defined activation function and a pre-defined set of decision parameters from the semantic feature tensor to the node. For example, the routing probability is obtained by multiplying the weighting coefficient in the decision parameters of the semantic feature tensor with the activation function after adding the correction coefficient in the decision parameters.

[0037] In one exemplary embodiment, the semantic compliance score is calculated through the following process: Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the risk category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the risk category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. The probabilities of all semantic feature tensors falling into a subset of risk categories are summed, and the sum is used as the semantic compliance score.

[0038] Compared with the scoring of relevant technical rules and the calculation of probability for a single category, this exemplary embodiment adds up the probabilities of all semantic feature tensors entering a subset of the risk category, and determines the semantic compliance score based on the sum obtained by the addition.

[0039] In one embodiment, semantic compliance scoring may be performed using an aggregation calculation method, but is not limited to.

[0040] In one exemplary embodiment, when the semantic feature vector comes from a pre-trained model such as BERT, the above classification model can be constructed by adding classification types; if the semantic feature tensor has a graph structure (such as a knowledge graph or dependency syntax tree), the above classification model can also include, but is not limited to, a model built based on a graph neural network (GNN); the above classification models can all be based on existing agricultural ontology sets, and can be classified by those skilled in the art into risk types and safety categories according to semantics ( Subsets and risk types are risk categories ( The requirements of the subset are used for construction; based on the different classification models constructed, those skilled in the art can use corresponding methods to calculate the classification event entities and semantic compliance scores according to the functional characteristics of different types of classification models constructed.

[0041] In one exemplary embodiment, step 103 further includes: When semantic feature tensor Text collection in When the second preset condition is met, it is determined that the semantic feature tensor falls into the gap between nodes or forms a high-entropy cluster. The second preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window. When determining whether a semantic feature tensor falls into the gap between nodes or forms a high-entropy cluster, the probability of the semantic feature tensor falling into all leaf nodes of the neural tree classifier is determined, and a new subgroup leaf node is generated below the leaf node with the highest probability. The weights of the newly generated leaf nodes are initialized using methods including, but not limited to, clustering algorithms (such as K-Means). Then, the classification label value corresponding to the newly generated leaf node is used as the semantic feature tensor. The corresponding category of event entity.

[0042] In this exemplary embodiment, the second preset condition may be, but is not limited to, the following: When the semantic feature tensor Text collection in satisfy When a semantic feature tensor falls into a node gap or forms a high-entropy cluster, the existing neural tree classifier cannot accurately describe the classification content. The probability of the semantic feature tensor falling into all leaf nodes of the neural tree classifier is determined, including the probability of the semantic feature tensor entering each leaf node in the subset of the risk category and the probability of the semantic feature tensor entering each leaf node in the subset of the safe category. This exemplary embodiment calculates the uncertainty based on the probability distribution of entering the safe subset, and then combines it with a splitting threshold within a time window to form a second preset condition, thus realizing the determination that the semantic feature tensor falls into a node gap or forms a high-entropy cluster. When the semantic feature tensor falls into a node gap / high-entropy cluster, the leaf node with the highest probability is found first, and then a new leaf is generated under it. The weight of the new leaf is initialized with clustering (such as K-Means), and the new word discovery mechanism is started through the above processing classification model, automatically splitting to generate new temporary subgroup leaf nodes, realizing the evolution of "unknown category" into a specific "new classification concept label".

[0043] In one exemplary embodiment, an agricultural industry chain causal graph is constructed based on the following processing: Constructing a directed weighted graph ,in, Indicates the links in the agricultural industry chain, edge This represents a transmission relationship, with each transmission edge having a weight. Represents the conduction strength (i.e., the probability of node i influencing node j); in this exemplary embodiment, during the construction process... The National Bureau of Statistics' "Classification of National Economic Industries" and the Ministry of Agriculture and Rural Affairs' "List of Key Monitored Agricultural Products" can be used as references to determine this. It can be constructed based on agricultural economics knowledge and through expert rule extraction; edge weights Agricultural economics experts can determine this based on historical data statistical correction and calibration; mapping the agricultural industry chain links (seeds / planting / processing / logistics / sales / price / public opinion / regulation, etc.) using the "National Economic Industry Classification" and the "List of Key Monitored Agricultural Products", and using the following rules for hyperedge mapping: 1) If node i (agricultural industry chain link i) is an upstream input or constraint of node j (agricultural industry chain link j), then a candidate edge (i→j) is constructed; the upstream relationship can be determined according to the industry chain flowchart / input-output logic; 2) If the risk indicator of node i has a significant Granger causal / lag correlation with the risk indicator of node j in the historical sequence, and the amount of significant Granger causal / lag correlation exceeds the threshold θ, then a candidate edge (i→j) is constructed; 3) If a person skilled in the art scores the candidate edges obtained by the above two methods based on expert rules as greater than or equal to the pre-set τ value, then the candidate edge is included in the transmission relationship E.

[0044] In this exemplary embodiment, the nodes of the agricultural industry chain causal graph directly correspond to the links in the agricultural industry chain (such as seeds, planting, processing, logistics, and sales), and the weight of edge E is... It provides transmission strength (or probability of impact) and can provide reliable industry classifications / directories and expert-calibrated paths, thus realizing the construction of a map for risk analysis that conforms to agricultural network information.

[0045] In one exemplary embodiment, the classification event entity corresponding to the semantic feature tensor is input into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classification event entity on a pre-constructed causal graph of the agricultural industry chain. Based on the risk potential energy information, the systemic risk of the classification event entity in the entire agricultural industry chain is determined, including: Based on the potential energy propagation model, according to the pre-set ripple decay factor and the set of all leaf nodes connected to leaf node i in the agricultural industry chain causal graph, the risk potential energy of leaf node i at a predetermined time is calculated, where i is the classification leaf node corresponding to the classification event entity of the semantic feature tensor in the agricultural industry chain causal graph. The calculated risk potential energy is multiplied by the pre-defined importance score of leaf node i in the agricultural industry chain, and then iterated K times to obtain the systemic risk value of the entire industry chain.

[0046] This exemplary embodiment uses the obtained classified event entities as anchor points in the causal graph of the agricultural industry chain (i.e., classified event entity → leaf node i of the agricultural industry chain graph), instead of manually specifying risk sources. This automatic alignment mechanism—textual semantics → classified event entity → leaf node of the agricultural industry chain graph—reduces manual settings, improves scalability and consistency, and achieves coupled processing of event entity classification and risk inference using semantic feature tensors. The aforementioned inference algorithm is not PageRank or shortest path, but a potential energy propagation model. It uses a "ripple decay factor" to propagate risk potential energy along adjacent nodes and performs K iterations to obtain the systematic risk, i.e., the systematic risk value of the entire industry chain. Compared to the inference of related technology probability products / graph ranking, this exemplary embodiment realizes the expression of physical potential energy—ripple decay—iterative convergence, achieving the analysis of the ripple effect of event entity risk on the industry chain graph.

[0047] In addition to the potential energy propagation model described above, this exemplary example can also employ other models capable of analyzing the ripple effect of event entity risk on the supply chain map. For example, the Joint Energy-based Model (JEM, a variant of EBM) reinterprets the classifier as an energy function to analyze the ripple effect of event entity risk on the supply chain map. Other models that can achieve this analysis include MeanField variational inference models or PageRank models, as long as they can analyze the ripple effect of event entity risk on the supply chain map.

[0048] In this exemplary embodiment, the potential energy propagation model is used to combine the initial risk of the semantic feature tensor under its corresponding classified event entity. (This value can be set to the above-mentioned phenological deviation value) The risk potential of leaf node i at time t+1 can be calculated using the following formula, but is not limited to, the application of causal graphs of the agricultural industry chain. The expression is: ; In the formula, The initial risk value for leaf node i can be the phenological deviation value of the semantic feature tensor; Let be the risk potential energy of leaf node j at time t, where leaf node j is the leaf node connected to leaf node i in the causal graph of the agricultural industry chain; It is the set of all leaf nodes connected to leaf node i in the causal graph of the agricultural industry chain; The ripple attenuation factor is a pre-set value (to prevent infinite propagation, with a value of 0-1). and The degree of the leaf node (the number of edges connecting the leaf node) is a pre-defined value used for normalization. This refers to the weight of the edge connecting leaf nodes i and j. In this exemplary embodiment, the initial risk potential of each leaf node can be set to the phenological deviation of the semantic tensor feature corresponding to that leaf node. Then, for each leaf node, its risk potential at time t can be calculated. For example, in the causal graph of the agricultural industry chain, the initial risk value of leaf node A is the phenological deviation of the semantic tensor feature corresponding to leaf node A, and the initial risk value of leaf node B is the phenological deviation of the semantic tensor feature corresponding to leaf node B. Assuming that leaf node A is only connected to leaf node B, then at t=1, the initial risk values ​​of leaf node A and leaf node B can be substituted into the above expression to calculate the risk potential of leaf node A at time 1. In the case of t=2, the initial risk value of leaf node A and the risk potential energy of leaf node B at time 1 can be used. The risk potential energy of leaf node A at time 2 is calculated using the above expression. ;in, The risk value of leaf node B can be calculated using the above expression based on the initial risk value of leaf node B and the initial risk values ​​of other leaf nodes connected to leaf node B. Similarly, the risk potential energy of each leaf node at subsequent time points can be obtained; the interval between adjacent time points can be set as needed. The risk potential energy is iterated K times (K can be greater than or equal to 3; in practical applications, it can be manually adjusted by technicians based on the calculation results; if the result is not ideal, the number of iterations can be increased), and the systemic risk value of the entire industry chain is calculated using the following formula. : ; Where i is the classification leaf node corresponding to the initial semantic feature tensor in the causal graph of the agricultural industry chain, and V is the set of leaf nodes in the causal graph of the agricultural industry chain. Let i be the risk potential energy of leaf node i at time k (the Kth time). The importance score of leaf node i in the agricultural industry chain is determined by expert scoring when constructing the causal graph of the agricultural industry chain, and each leaf node has an importance weight.

[0049] This exemplary embodiment explicitly incorporates node importance scoring into the calculation of the systemic risk value of the entire industry chain. Risk potential energy is multiplied by node importance and accumulated iteratively. This approach, which previously only considered path probability or structural center analysis, amplifies key links, demonstrating the amplification effect of key links. Based on the classified event entities corresponding to the semantic feature tensors generated by the above steps, a potential energy propagation model from physics is introduced to calculate the ripple effect of event entity risk on the industry chain graph. This allows agricultural website information risk assessment to directly quantify its systemic impact on upstream and downstream industries, rather than being limited to the agricultural website information itself, thus enhancing decision-making value.

[0050] In one exemplary embodiment, a weighted calculation is performed on phenological deviation, semantic compliance score, and systemic risk across the entire industry chain to obtain comprehensive risk score information for agricultural network information. This includes calculations based on the following formula: ; in , , For the pre-set weighting coefficients, Indicates the degree of phenological deviation. Indicates semantic compliance score, This indicates systemic risk across the entire industry chain.

[0051] This exemplary embodiment integrates risk scoring information by using phenological deviation (for falsification / counterintuitive reasoning) and semantic compliance scoring (for risk category probability), i.e., the systemic risk of the entire industry chain affected by ripples. The three main factors are agriculturally specific and have a clear division of labor, achieving a closed loop of the same agricultural network information across three dimensions: semantics, phenological patterns, and industry chain spillover effects. The integrated risk scoring information can naturally be broken down into the contributions (Δscore) of each factor and can be traced back to "hit phenological entities, the probability distribution of risk leaf nodes, and the set / intensity of nodes in the industry chain affected by ripples," making it more interpretable and scientifically designed than other technologies.

[0052] This exemplary embodiment is when Greater than the preset scoring threshold If the message is flagged, it will be submitted to a human reviewer for a decision on its fate (posting or deletion, etc.); otherwise, if the message is not flagged, it will be automatically deleted. Less than the scoring threshold At that time, it will directly enter the corresponding section of the website's publishing database (by category entity). (Confirmed) and ready for release.

[0053] In one exemplary embodiment, the comprehensive risk scoring information further includes weighting terms, either one or any combination of the following: The factors include the authority of the information source, the difficulty coefficient of risk prevention and control, the regional vulnerability factor, and the policy adaptability coefficient.

[0054] When this exemplary embodiment integrates risk scoring information including a weighted factor of source authority, China plus , Indicates the authority of the source. The weighting coefficients are pre-set; when the comprehensive risk scoring information includes a weighted term for the risk prevention and control difficulty coefficient factor, China plus , This represents the risk control difficulty factor. These are pre-set weighting coefficients; when the comprehensive risk scoring information includes a weighted term for the regional vulnerability factor, China plus , Indicating regional vulnerability factors, The weighting coefficients are pre-set; when the comprehensive risk scoring information includes a weighted term for the policy fit coefficient, China plus , Indicates the policy fit coefficient. These are pre-set weighting coefficients.

[0055] In this exemplary embodiment, the source authority factor can be a correction coefficient determined based on the identity attributes of the publishing entity (official, research, or non-governmental) and the accuracy rate of its historical posts; the risk prevention and control difficulty factor can be based on the difficulty of agricultural technology or resource scarcity, measuring the ease with which a specific agricultural event can be intervened and resolved; the regional vulnerability factor can be combined with the infrastructure and economic level of the information's origin to assess the region's ability to absorb and withstand specific risks; and the policy adaptation coefficient can be used to assess the degree of fit between the information content and current agricultural laws, regulations, and industry policies.

[0056] This exemplary embodiment also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the above-described information dissemination risk assessment method based on agricultural network information.

[0057] This exemplary embodiment also provides a terminal, including: a memory and a processor, wherein the memory stores a computer program; wherein, The processor is configured to execute computer programs in memory; When a computer program is executed by a processor, it implements the information dissemination risk assessment method based on agricultural network information as described above.

[0058] Figure 2 This is a structural block diagram of an information dissemination risk assessment device based on agricultural network information, as described in an embodiment of this disclosure. Figure 2 As shown, it includes: an extraction and processing unit, a deviation calculation unit, a classification analysis unit, a potential risk analysis unit, and a comprehensive scoring unit; among which, The extraction and processing unit is set to extract the spatiotemporal phenological feature information from the agricultural network information to be analyzed, and construct a semantic feature tensor based on the extracted spatiotemporal phenological feature information. The spatiotemporal phenological feature information includes: semantic vectors, and the time entities and spatial entities involved in the semantic vectors. The deviation calculation unit is set to calculate the phenological deviation of the semantic feature tensor based on the pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. The classification analysis unit is set up as follows: input the constructed semantic feature tensor into the pre-constructed classification model, determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor, wherein the semantic compliance score is the sum of the probabilities of the semantic feature tensor falling into the risk category; The potential energy risk analysis unit is set up as follows: inputting the classified event entity corresponding to the semantic feature tensor into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entity on the pre-constructed causal graph of the agricultural industry chain, and determining the systemic risk of the classified event entity in the entire agricultural industry chain based on the risk potential energy information. The comprehensive scoring unit is set up to calculate the comprehensive risk score information of agricultural network information by weighting the phenological deviation, semantic compliance score and systemic risk of the entire industry chain.

[0059] In one exemplary embodiment, the extraction processing unit is configured as follows: The BERT model is used to extract the basic semantic vectors and the temporal and spatial entities involved in the semantic vectors from the agricultural network information to be analyzed. The temporal and spatial entities involved in the extracted semantic vectors are weighted using a pre-defined weight matrix. The extracted semantic vectors and the temporal and spatial entities involved in the weighted semantic vectors are multimodal spliced ​​and fused using a layer normalization model to form a semantic feature tensor.

[0060] In one exemplary embodiment, for each semantic feature tensor, its phenological deviation is obtained through the following processing: Identify the entities that the semantic feature tensor matches in the agricultural phenology knowledge base; for each matched entity, calculate the deviation between the semantic feature tensor and the entity; compare the calculated deviation values ​​and determine the maximum deviation value as the phenological deviation.

[0061] In one exemplary embodiment, the classification model includes a security category and a risk category, and the classification analysis unit is configured to obtain the classified event entity corresponding to the semantic feature tensor through the following processing: The semantic feature tensor is input into the neural tree classifier of the classification model to determine the routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier. Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the safe category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the safe category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. For the set of leaf nodes of the neural tree classifier, calculate the uncertainty value of the semantic feature tensor in the neural tree classifier based on the probability that the semantic feature tensor enters the subset of the safe category; When the text set in the semantic feature tensor When the first preset condition is met, the classification label value corresponding to the leaf node with the smallest uncertainty value in the leaf node set of the semantic feature tensor is taken as the classification event entity corresponding to the semantic feature tensor. The first preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window.

[0062] In one exemplary embodiment, the classification analysis unit is configured to obtain a semantic compliance score through the following processing: Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the risk category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the risk category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. The probabilities of all semantic feature tensors falling into a subset of the risk category are summed, and the sum is used as the semantic compliance score. In an exemplary embodiment, the classification analysis unit is further configured as follows: When semantic feature tensor Text collection in When the second preset condition is met, it is determined that the semantic feature tensor falls into the node gap or forms a high-entropy cluster. The second preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window. When determining whether a semantic feature tensor falls into the gap between nodes or forms a high-entropy cluster, the probability of the semantic feature tensor falling into all leaf nodes of the neural tree classifier is determined, and a new subgroup of leaf nodes is generated below the leaf node with the highest probability. Below the leaf node that has the largest product of all route probabilities on the path from the semantic feature tensor to the leaf node, generate a new subgroup of leaf nodes. After initializing the weights of the newly generated leaf nodes, the classification label value corresponding to the newly generated leaf node is used as the classification event entity corresponding to the semantic feature tensor.

[0063] In one exemplary embodiment, the potential energy risk analysis unit is configured as follows: Based on the potential energy propagation model, according to the pre-set ripple decay factor and the set of all leaf nodes connected to leaf node i in the causal graph of the agricultural industry chain, the risk potential energy of leaf node i at a predetermined time is calculated, where i is the classification leaf node corresponding to the semantic feature tensor in the causal graph of the agricultural industry chain. The calculated risk potential energy is multiplied by the pre-defined importance score of leaf node i in the agricultural industry chain, and then iterated K times to obtain the systemic risk value of the entire industry chain.

[0064] In one exemplary embodiment, the comprehensive risk scoring information further includes weighting terms, either one or any combination of the following: The factors include the authority of the information source, the difficulty coefficient of risk prevention and control, the regional vulnerability factor, and the policy adaptability coefficient.

[0065] The following application examples briefly illustrate the embodiments of this disclosure. These application examples are only used to illustrate the embodiments of this disclosure and are not intended to limit the scope of protection of the embodiments of this disclosure.

[0066] Application Examples Figure 3 This is a schematic diagram illustrating the process of risk assessment based on agricultural network information in an application example, such as... Figure 3 As shown, it includes the following four parts: constructing semantic feature tensors and determining phenological deviation, event entity classification and semantic compliance analysis based on continuous neural trees, risk ripple inference based on the causal graph of the agricultural industry chain, and comprehensive risk scoring of agricultural network information; among them, Constructing semantic feature tensors and determining phenological deviation includes: Obtain agricultural network information to be analyzed; The acquired agricultural network information is encoded to obtain time vectors, semantic vectors, and spatial vectors. This exemplary embodiment of encoding the acquired agricultural network information includes: extracting basic semantic vectors and the time and spatial entities involved in the semantic vectors from the agricultural network information to be analyzed; encoding the extracted time and spatial entities involved in the semantic vectors to obtain time and spatial vectors; including using sinusoidal position encoding to map time entities into periodic vectors and mapping spatial entities into spatial vectors containing latitude, longitude, and climate zone attributes to the agricultural network information. Construct a semantic feature tensor based on the obtained time vector, semantic vector, and spatial vector; The phenological deviation is calculated based on the phenological knowledge base of agricultural seasons, using the semantic feature tensor.

[0067] After constructing the semantic feature tensor in this exemplary embodiment, in addition to inputting it into the agricultural phenology knowledge base for processing, it is also input into the classification model for event entity classification and semantic compliance analysis.

[0068] Event entity classification and semantic compliance analysis based on continuous neural trees, including: Input the semantic feature tensor into the classification model based on continuous neural trees; The semantic compliance score is determined based on the probability that the semantic feature tensor falls into a subset of the risk category.

[0069] For the set of leaf nodes of the neural tree classifier, calculate the uncertainty value of the semantic feature tensor in the neural tree classifier based on the probability that the semantic feature tensor enters the subset of the safe category; Based on the uncertainty value, the classification event entity corresponding to the semantic feature tensor is determined.

[0070] After determining the categorized event entities, this exemplary embodiment inputs the determined categorized event entities into the agricultural industry chain causal graph for subsequent risk ripple analysis.

[0071] Risk ripple projection based on the causal graph of the agricultural industry chain includes: Input the classification event entity corresponding to the semantic feature tensor into the agricultural industry chain causal graph to determine the set of all leaf nodes connected to leaf node i in the agricultural industry chain causal graph, where i is the classification leaf node corresponding to the semantic feature tensor in the agricultural industry chain causal graph. Based on the pre-set ripple decay factor and the set of all leaf nodes connected to leaf node i in the causal graph of the agricultural industry chain, calculate the risk potential energy of leaf node i at a predetermined time. Based on the calculated risk potential and the pre-defined importance score of leaf node i in the agricultural industry chain, the systemic risk of the classified event entity in the entire agricultural industry chain is determined.

[0072] This exemplary embodiment obtains phenological deviation, semantic compliance score, and systemic risk of the entire industry chain based on the above processing. It then performs weighted calculations using pre-set weighting coefficients to obtain comprehensive risk score information.

[0073] The comprehensive risk assessment of agricultural network information includes: By weighting phenological deviation, semantic compliance score, and systemic risk across the entire industry chain, a comprehensive risk score for agricultural network information is obtained.

[0074] This exemplary embodiment further includes processing the obtained comprehensive risk score information, including publishing the comprehensive risk score information or reviewing the comprehensive risk score information. This exemplary embodiment can directly integrate the comprehensive risk score information into business orchestration; for example, if it exceeds a threshold, it is marked and enters manual review (publishing / deleting); if it is below the threshold, it is entered into the publishing database by category. Through a human-machine collaborative processing link, the risk of accidental deletion / omission is reduced, and interpretability and auditability are improved.

[0075] This exemplary embodiment retrieves a news article containing agricultural network information from an agricultural self-media platform. The article is titled "Major Good News! A New Corn Variety Can Be Planted in Late Autumn in a Certain Location, Increasing Yield by 30%!", published on October 25th, with the location specified as a certain location. The risk assessment method of this exemplary embodiment extracts and fuses multiple features from the retrieved information: 1) Extracting various features from the text, namely semantic features (text) containing key information such as "high yield," "late autumn planting," and "new corn variety"; time feature (t): October 25th; spatial feature: a certain location; 2) Performing tensor fusion on the various features, that is, encoding the semantic features using BERT to form a semantic vector. Transforming temporal features into temporal entities through sinusoidal position encoding. (Also known as the agricultural phenological vector, representing the Cold Dew solar term in late autumn), it transforms spatial features into spatial entities based on latitude and longitude. (Spatial vector), by fusing the above three types of vectors, a semantic feature tensor of spatiotemporal semantic fusion is formed. 3) Based on the "Encyclopedia of Chinese Agriculture" (ISBN: 9787109003491), a knowledge base of agricultural seasons and phenology, Kpheno, was constructed, and semantic feature tensors were used to construct the knowledge base. The phenological deviation is calculated by entering the data into this agricultural phenological knowledge base. =0.92, close to the highest risk value, indicating that the semantics of this information conflict significantly with spatiotemporal knowledge. This result is corroborated by the actual situation (late autumn in October, corn in a certain region is in the harvest and straw disposal stage, making corn planting impossible). At this point, the system has determined that this information carries a significant scientific risk. Based on the "FAO AGROVOC Thesaurus," a continuous neural tree (CNT) of depth 15 is constructed. Based on manual selection, the risk types of the leaf nodes of the CNT are divided into two main subtrees: safe categories (positive events) and risk categories (negative events). 1) Calculation Routing in CNT, i.e. In the classification path of a neural tree, this semantic feature tensor has high... If the routing is positive, the routing tends to favor the risk category subtree branch; 2) Based on the final routing result, the semantic feature tensor is finally assigned to the "fake agricultural technology promotion" leaf node of the risk category subtree. Based on the classification confidence of the leaf node, the semantic compliance score of the input semantic feature tensor is calculated. =0.85, 3) Based on the classification results of the semantic feature tensor, the system classifies the event entity of the semantic feature tensor as "false propaganda about late corn planting".

[0076] Based on the National Bureau of Statistics' "Classification of National Economic Industries" and the Ministry of Agriculture and Rural Affairs' "List of Key Monitored Agricultural Products," and combined with expert scoring and evaluation, a causal graph G for the agricultural industry chain was constructed. 1) The categorized event entities of the obtained semantic feature tensors were injected into the causal graph G of the agricultural industry chain. The initial potential energy propagation starting nodes were "planting technology" and "agricultural input sales." Risk potential energy was iterated based on these two initial nodes. 2) The iteration results showed that, starting from "planting technology," the risk potential energy was transmitted to the "seed / pesticide sales" leaf node with a high transmission weight. However, starting from "agricultural input sales," the risk potential energy was transmitted to the "market price" leaf node, but with a low transmission weight. This indicates that the information may affect the reputation of the relevant companies involved, but has little impact on actual production output and market prices. 3) Based on the calculation model, after system iteration calculation, the final systemic risk of the entire industry chain was determined. =0.4.

[0077] According to the three risk source weights pre-set by those skilled in the art, they are respectively , , The comprehensive risk score information for this piece of information can be calculated. If the score is converted to a percentage and is 78.8, which is higher than the information risk threshold of 70 on the aforementioned portal website, the system will mark the agricultural network information as a red risk label, push it to the "Planting Technology" section of the website group, and suggest manual review of the information.

[0078] This exemplary embodiment primarily achieves the following effects: Improved accuracy: By introducing agricultural timing and geographical constraints, it effectively solves the problems of misjudging the regional applicability of agricultural technology information, such as "outdated information" and "oranges grown south of the Huai River turning into trifoliate oranges"; Forward-looking early warning: Through risk ripple analysis, it can not only identify current public opinion but also predict the potential impact of such public opinion on the downstream of the industry chain; Adaptability: Utilizing a continuous neural tree mechanism, emerging agricultural concepts (such as new varieties and new diseases) can be discovered without frequent manual retraining; Using the above solution, in terms of portal website content operation and maintenance alone, the overall workload of content processing and review has been reduced by about 40%, and the number of public opinion complaints caused by website content publication has decreased by 30% compared to before, demonstrating good application results.

[0079] In another exemplary embodiment of this application, for an agricultural self-media platform, the title of an article published on March 15th is: "A company releases a new rice variety, claiming that it can achieve high yields in saline-alkali land, with a yield of 800 kg per mu!" The abstract of the text is: "This rice variety has super salt and alkali resistance and can be planted in heavily saline-alkali land with a soil pH value of 8.5 or higher. It does not require complex soil improvement and reduces planting costs by 40%. It has already been demonstrated in 1,000 mu in this region." The risk assessment method of this exemplary embodiment performs multi-source feature extraction and fusion on the above-mentioned crawled information: 1) Extract various features from the text, namely semantic features (text) including "saline-alkali land", "new rice variety", "yield of 800 kg per mu", "no soil improvement required", "cost reduction of 40%", time features (t): March 15th, corresponding to the second month of the lunar calendar, spring; spatial features: saline-alkali land, corresponding to the northern region, for example, latitude and longitude: (38.0, 115.5); 2) Perform tensor fusion on the various features, that is, fuse the above three types of vectors to form a spatiotemporal semantic feature tensor. ;3) Transform the semantic feature tensor The phenological deviation was calculated by inputting the information into the agricultural phenological knowledge base. The calculated deviation was close to the highest risk value, indicating a significant semantic conflict between the information and spatiotemporal knowledge. This was corroborated by the actual situation (the statement "In mid-March, the soil temperature in Hebei's saline-alkali land is below 15℃, unsuitable for rice sowing, no soil improvement needed" is seriously inconsistent with common agronomical knowledge; the yield of 800 kg per mu in saline-alkali land far exceeds the conventional level, and the pH of 8.5 exceeds the limit for salt-tolerant rice). Therefore, the system determined that this information posed a significant scientific risk. Based on the "FAO AGROVOC Thesaurus," a continuous neural tree (CNT) was constructed. Based on manual selection, the risk types of the leaf nodes of the CNT were divided into two main subtrees: safe categories (positive events) and risk categories (negative events). Routing in CNT, i.e. In the classification path of a neural tree, this semantic feature tensor has high... If the routing path is determined, the semantic feature tensor is ultimately assigned to the leaf node of the risk category subtree, which is "false advertising of saline-alkali land adaptability". Based on the classification confidence of this leaf node, the semantic compliance score of the input semantic feature tensor is calculated. Based on the classification result of the semantic feature tensor, the system assigns the classification event entity of the semantic feature tensor to "false advertising of saline-alkali land adaptability". Referring to the above example, an agricultural industry chain causal graph is constructed. Based on the classification event entity of the obtained semantic feature tensor, it is injected into the agricultural industry chain causal graph, anchored to the initial potential energy propagation starting node. Risk potential energy is iterated based on the anchored initial node to finally determine the systemic risk of the entire industry chain. Based on the pre-set weights of the three risk sources, the comprehensive risk score of this information can be calculated. If it exceeds the risk threshold of the portal website information, the system marks this agricultural network information with a red risk label and suggests manual review of the information. This exemplary embodiment effectively solves the problem of misjudging the applicability of seasons and regions by introducing agricultural timing and geographical constraints; through risk ripple analysis, it can not only identify current public opinion, but also predict the potential impact of public opinion on the downstream of the industrial chain; by using the continuous neural tree mechanism, emerging agricultural concepts can be discovered without frequent manual retraining; using the above solution, in terms of portal website content operation and maintenance alone, the overall workload of content processing and review has been reduced, and the number of public opinion complaints caused by website content publication has decreased compared with before, demonstrating good application results.

[0080] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term "computer storage medium" includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

Claims

1. A method for assessing the risk of information dissemination based on agricultural network information, characterized in that, include: Spatiotemporal phenological features are extracted from the agricultural network information to be analyzed, and a semantic feature tensor is constructed based on the extracted spatiotemporal phenological features. The spatiotemporal phenological features include: semantic vectors, and the time and space entities involved in the semantic vectors. The phenological deviation of the semantic feature tensor is calculated based on a pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. The constructed semantic feature tensor is input into a pre-constructed classification model to determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor. The semantic compliance score is the sum of the probabilities that the semantic feature tensor falls into the risk category. Input the classified event entities corresponding to the semantic feature tensor into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entities on the pre-constructed causal graph of the agricultural industry chain. Based on the risk potential energy information, determine the systemic risk of the classified event entities in the entire agricultural industry chain. By weighting phenological deviation, semantic compliance score, and systemic risk across the entire industry chain, a comprehensive risk score for agricultural network information is obtained.

2. The information dissemination risk assessment method according to claim 1, characterized in that, The process of extracting spatiotemporal phenological features from the agricultural network information to be analyzed, and constructing a semantic feature tensor based on the extracted spatiotemporal phenological features, includes: The BERT model is used to extract the basic semantic vectors and the temporal and spatial entities involved in the semantic vectors from the agricultural network information to be analyzed. The temporal and spatial entities involved in the extracted semantic vectors are weighted using a pre-defined weight matrix. The extracted semantic vectors and the temporal and spatial entities involved in the weighted semantic vectors are multimodal spliced ​​and fused using a layer normalization model to form the semantic feature tensor.

3. The information dissemination risk assessment method according to claim 2, characterized in that, Before performing multimodal splicing and fusion of the extracted semantic vectors and the temporal and spatial entities involved in the weighted semantic vectors through a layer normalization model, the information dissemination risk assessment method further includes: The temporal entity code is mapped into a periodic vector using sinusoidal positional coding; Wherein, the position of the sinusoidal position code is the index value after the time entity is mapped to the agricultural cycle, the dimension of the sinusoidal position code is the vector index value of the time entity to be encoded and mapped, and the total dimension of the sinusoidal position code vector is the number of solar terms included in the predetermined agricultural cycle.

4. The information dissemination risk assessment method according to claim 1, characterized in that, For each of the semantic feature tensors, its phenological deviation is obtained through the following processing: Identify the entities that the semantic feature tensor matches in the agricultural phenology knowledge base; for each matched entity, calculate the deviation between the semantic feature tensor and the entity. By comparing the calculated deviation values, the maximum value of the deviation is determined as the phenological deviation.

5. The information dissemination risk assessment method according to claim 1, characterized in that, The classification model includes a security category and a risk category, and the classification event entities corresponding to the semantic feature tensor are obtained through the following processing: The semantic feature tensor is input into the neural tree classifier of the classification model to determine the routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier; Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the safe category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the safe category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. For the set of leaf nodes of the neural tree classifier, calculate the uncertainty value of the semantic feature tensor in the neural tree classifier based on the probability that the semantic feature tensor enters a subset of the safe category; When the semantic feature tensor contains the text set When the first preset condition is met, the classification label value corresponding to the leaf node with the smallest uncertainty value in the leaf node set of the semantic feature tensor is taken as the classification event entity corresponding to the semantic feature tensor. The first preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window.

6. The information dissemination risk assessment method according to claim 5, characterized in that, The semantic compliance score is calculated through the following process: Based on the determined routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, the probability of the semantic feature tensor entering a subset of the risk category is calculated respectively. For any semantic feature tensor, the probability of it entering any leaf node in the subset of the risk category is the product of all routing probabilities on the path from the semantic feature tensor to that leaf node. The probabilities of all the semantic feature tensors falling into a subset of the risk category are summed, and the sum is used as the semantic compliance score.

7. The information dissemination risk assessment method according to claim 5, characterized in that, The risk assessment method also includes: When the semantic feature tensor contains the text set When the second preset condition is met, it is determined that the semantic feature tensor falls into the node gap or forms a high-entropy cluster. The second preset condition is determined based on the uncertainty value of the semantic feature tensor in the neural tree classifier and the splitting threshold within a preset time window. When the semantic feature tensor falls into the gap between nodes or forms a high-entropy cluster, the probability of the semantic feature tensor falling into each leaf node of the neural tree classifier is determined according to the routing probability of the semantic feature tensor reaching each leaf node of the neural tree classifier, and a new subgroup of leaf nodes is generated below the leaf node with the highest probability. After initializing the weights of the newly generated leaf nodes, the classification label value corresponding to the newly generated leaf node is used as the classification event entity corresponding to the semantic feature tensor.

8. The information dissemination risk assessment method according to claim 6, characterized in that, The step of inputting the classified event entities corresponding to the semantic feature tensors into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entities on a pre-constructed causal graph of the agricultural industry chain, and determining the systemic risk of the classified event entities across the entire agricultural industry chain based on the risk potential energy information, includes: Based on the aforementioned potential energy propagation model, according to the pre-set ripple decay factor and the set of all leaf nodes connected to leaf node i in the causal graph of the agricultural industry chain, the risk potential energy of leaf node i at a predetermined time is calculated, where leaf node i is the leaf node corresponding to the semantic feature tensor in the causal graph of the agricultural industry chain; the calculated risk potential energy is multiplied by the pre-set importance score of leaf node i in the agricultural industry chain and then iterated K times to obtain the systematic risk value of the entire industry chain.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the information dissemination risk assessment method based on agricultural network information as described in any one of claims 1 to 8.

10. A device for assessing the risk of information dissemination based on agricultural network information, characterized in that, include: The system includes an extraction and processing unit, a deviation calculation unit, a classification analysis unit, a potential risk analysis unit, and a comprehensive scoring unit; among which, The extraction and processing unit is set to extract the spatiotemporal phenological feature information from the agricultural network information to be analyzed, and construct a semantic feature tensor based on the extracted spatiotemporal phenological feature information. The spatiotemporal phenological feature information includes: semantic vectors, and the time entities and spatial entities involved in the semantic vectors. The deviation calculation unit is set to calculate the phenological deviation of the semantic feature tensor based on the pre-built agricultural phenological knowledge base. The higher the phenological deviation, the greater the probability that the agricultural information corresponding to the semantic feature tensor is false information. The classification analysis unit is set up as follows: input the constructed semantic feature tensor into the pre-constructed classification model, determine the semantic compliance score and obtain the classification event entity corresponding to the semantic feature tensor, wherein the semantic compliance score is the sum of the probabilities of the semantic feature tensor falling into the risk category; The potential energy risk analysis unit is set up as follows: inputting the classified event entity corresponding to the semantic feature tensor into a pre-determined potential energy propagation model to obtain the risk potential energy information of the classified event entity on the pre-constructed causal graph of the agricultural industry chain, and determining the systemic risk of the classified event entity in the entire agricultural industry chain based on the risk potential energy information. The comprehensive scoring unit is set up to calculate the comprehensive risk score information of agricultural network information by weighting the phenological deviation, semantic compliance score and systemic risk of the entire industry chain.

Images