Risk transmission evaluation method and system based on knowledge graph and semantic reasoning

Abstract

The method and system for risk transmission evaluation based on knowledge graph and semantic reasoning acquire an international event data set to be evaluated; the event text description of each international event data is matched with a preset risk type set, the semantic correlation weight between each international event data and each preset risk type is calculated, and an event-type weight matrix is generated based on all the international event data; a three-level knowledge graph is constructed; the time decay weight of each event node is calculated according to the time difference between the time stamp corresponding to each event node and the evaluation reference time; the node score of each risk type node is calculated; the node scores of the risk type nodes are multiplied by the preset edge weights of the respective target risk result nodes and then summed to obtain a comprehensive risk score; based on the comprehensive risk score and the weights of the nodes and directed edges in the three-level knowledge graph, a risk transmission evaluation report including a risk transmission path and a contribution degree ranking is generated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a risk transmission assessment method and system based on knowledge graphs and semantic reasoning. Background Technology

[0002] In terms of international conflict and security risk assessment, existing risk transmission assessment methods mainly include the following three categories: First, rule-based reasoning systems based on expert experience, which identify risk events by matching predefined rules; second, trend prediction methods based on statistical models, which use time series analysis or regression models to model historical data; and third, relational query methods based on knowledge graphs, which store entity relationships through graph structures and perform simple traversal analysis.

[0003] However, the above technologies still have significant limitations in practical applications: Rule-based matching methods have limited processing capabilities for unstructured international event texts, making it difficult to automatically extract semantic associations of risk transmission from large-scale open-source intelligence, resulting in a low degree of automation in risk analysis. Statistical model-based methods typically rely on structured input data, cannot directly process event text descriptions, and lack sufficient depth of understanding of text semantics, making it difficult to accurately capture implicit relationships between events. Existing knowledge graph methods mostly adopt a flat graph structure, directly linking events with risk outcomes. They lack hierarchical modeling of the intermediate processes of risk transmission, resulting in unclear transmission chains and difficult-to-interpret prediction results.

[0004] In addition, most existing methods do not fully consider the impact of the time attribute of events on risk transmission, and often treat all events as equally important, resulting in insufficient timeliness of risk scoring.

[0005] In summary, there is still considerable room for improvement in the intelligence and accuracy of existing technologies in assessing the transmission of international conflict risks. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a risk transmission assessment method based on knowledge graphs and semantic reasoning, comprising the following steps: Obtain the dataset of international events to be evaluated, where each international event data entry includes at least an event ID, an event text description, and its corresponding timestamp; The event text description of each international event data is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data and each preset risk type, and an event-type weight matrix is ​​generated based on all international event data. A three-level knowledge graph is constructed based on an event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk outcome nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk outcome nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and iteratively adjusted during the graph propagation calculation process. Calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time; For each risk type node, calculate the product of the initial activation value, time decay weight, and semantic relevance weight of each event node associated with it, and sum the products corresponding to the same risk type node to obtain the node score of that risk type node. The node scores of each risk type node are multiplied by the preset edge weights pointing to the target risk result nodes, and then summed to obtain the comprehensive risk score. Based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, which includes the risk transmission path and contribution ranking.

[0007] Optional, preset risk types include military buildup, diplomatic crisis, economic sanctions, cyberattacks, territorial changes, political instability, economic pressure, and foreign interference.

[0008] Optionally, obtain the international event dataset to be evaluated, including: Obtain raw international event data, and perform format unification and noise reduction processing on the event descriptions in the raw international event data to obtain basic event text descriptions; The time information in the raw international event data is standardized to obtain a timestamp in a unified format. Based on the original international event data, corresponding event IDs are generated, and the basic event text descriptions and their corresponding timestamps are associated and combined with the corresponding event IDs to obtain the basic international event dataset; Data augmentation is performed on the basic international event dataset to obtain the international event dataset to be evaluated.

[0009] Optionally, data augmentation processing can be performed on the basic international event dataset to obtain the international event dataset to be evaluated, including: Semantic parsing is performed on the textual descriptions of basic events in the basic international event dataset to identify numerical parameters and the degree of correlation. Based on a preset severity grading system, numerical parameters are replaced with values ​​corresponding to higher or lower levels, and severity descriptions are replaced with expressions of stronger or weaker intensity to generate multiple intermediate texts. The large language model is invoked to process the data of each intermediate text and generate enhanced event samples; the entity information in the enhanced event samples is consistent with the entity information in the basic event text description. Calculate the semantic similarity between each enhanced event sample and the corresponding basic event text description, and verify the validity of the enhanced event samples based on a preset similarity threshold; For each enhanced event sample that passes the validity check, a new event ID is generated. The enhanced event sample that passes the validity check, the new event ID, and its corresponding original timestamp are then associated and combined and added to the basic international event dataset to generate the international event dataset to be evaluated.

[0010] Optionally, the event text description of each international event data point is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data point and each preset risk type, including: The event text description of each international event data and the description text of each preset risk type are encoded into semantic vectors by a pre-trained model to generate a similarity matrix. Based on the timestamp corresponding to each international event data, the weights of the similarity matrix are optimized to obtain the semantic correlation weights between each international event data and each preset risk type. Alternatively, a structured prompt template can be constructed, which includes role definition, task description, event text description, preset risk type set and output format constraints. Based on the structured prompt template, a large language model is called to perform semantic reasoning on the event text description of each international event data. The semantic correlation weights of each preset risk type are parsed from the returned semantic reasoning results, and the parsing results are numerically verified and corrected to obtain the semantic correlation weights between each international event data and each preset risk type.

[0011] Optionally, the initial activation value for each event node is the same preset constant.

[0012] Optionally, based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time, the time decay weight of each event node is calculated, including: Obtain the time difference between the timestamp of each event node and the evaluation baseline time, as well as the configurable decay rate parameter; The time decay weight, which decreases exponentially with the increase of time difference, is calculated based on the time difference and decay rate parameters using the exponential decay function, and is then stored in the corresponding event node.

[0013] Optionally, based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, including the risk transmission path and contribution ranking, comprising: Starting from the target risk outcome node, backtrack along the directed edges of the three-level knowledge graph to extract all complete transmission paths from the event node through the risk type node to the target risk outcome node; For each complete transmission path, the contribution score of the path is obtained by multiplying the time decay weight of the event node, the semantic correlation weight between the event node and the risk type node, and the preset edge weight between the risk type node and the target risk result node. All complete transmission paths are sorted in descending order based on their contribution scores, and the set of paths with contribution scores greater than a preset contribution score threshold or the set of paths with a cumulative contribution score percentage greater than a preset percentage threshold are selected as the main transmission paths. For each event node, the sum of the products of all semantically related weights associated with it and the preset edge weights of the corresponding risk type nodes is multiplied by the time decay weight of the event node to obtain the overall contribution of the event node. Key event nodes are identified based on their overall contribution and the degree of convergence of each event node in the transmission path. Generate a risk transmission assessment report that includes a comprehensive risk score, a ranking of the contribution of nodes for each risk type, a list of key event nodes, and a sequence of nodes along the main transmission path.

[0014] Optionally, a three-level knowledge graph is constructed based on the event-type weight matrix, including at least event nodes, risk type nodes, and target risk outcome nodes, including: Create a corresponding event node for each piece of international event data, create a corresponding risk type node for each preset risk type, and create a unique target risk outcome node, with the specific target risk outcome being the outbreak of war. For a corresponding event node whose semantic relevance weight is greater than a preset weight threshold, a directed edge is created between the event node and the corresponding risk type node, and the semantic relevance weight is used as the weight of the directed edge. For each risk type node, create a directed edge pointing to a unique target risk outcome node.

[0015] Corresponding to the aforementioned risk transmission assessment method based on knowledge graphs and semantic reasoning, this invention provides a risk transmission assessment system based on knowledge graphs and semantic reasoning, comprising: The dataset acquisition module acquires the international event dataset to be evaluated, where each international event data item includes at least an event ID, an event text description, and its corresponding timestamp; The semantic matching module is used to semantically match the event text description of each international event data with a preset risk type set, and calculate the semantic correlation weight between each international event data and each preset risk type. The weight matrix generation module is used to generate an event-type weight matrix based on all international event data; The knowledge graph construction module is used to construct a three-level knowledge graph based on the event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk result nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk result nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and are iteratively adjusted during the graph propagation calculation process. The first calculation module is used to calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation benchmark time. The second calculation module is used to calculate the product of the initial activation value, time decay weight and semantic relevance weight of each event node associated with each risk type node, and to sum the products corresponding to the same risk type node to obtain the node score of that risk type node. The third calculation module is used to multiply the node scores of each risk type node by the preset edge weights of their respective nodes pointing to the target risk result nodes and then sum them to obtain a comprehensive risk score. The risk transmission assessment report generation module is used to generate a risk transmission assessment report, including risk transmission path and contribution ranking, based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By semantically matching international event text descriptions with preset risk types and directly calculating semantically relevant weights, the automatic conversion from unstructured text to structured risk semantics is achieved, improving the automation level and semantic understanding depth of the assessment. By constructing a three-level knowledge graph of event nodes, risk type nodes, and target risk outcome nodes, the risk transmission path is explicitly decomposed into a hierarchical structure, solving the problem of unclear transmission chains and poor interpretability caused by flat graph structures. The introduction of time decay weights in node score calculation makes recent events and historical events present differentiated contributions in risk scoring, enhancing timeliness. Through the semantically relevant weight edges between event nodes and risk type nodes, the preset edge weights between risk type nodes and target risk outcome nodes, and the two-stage weighted aggregation scoring mechanism, semantic weights, time weights, and graph propagation weights work synergistically, realizing the layer-by-layer transmission calculation from event semantics to risk type and then to comprehensive risk. This invention significantly improves the automation level, accuracy, interpretability, and timeliness of international conflict risk transmission assessment.

[0017] (2) The preset risk types are military mobilization, diplomatic crisis, economic sanctions, cyber attack, territorial change, political turmoil, economic pressure and foreign intervention. This provides a standardized and domain-adaptive risk classification system for international conflict risk assessment, which is convenient to call directly when performing subsequent semantic matching and map construction, thus improving the pertinence and consistency of the assessment.

[0018] (3) By standardizing the format and denoising the event descriptions in the original international event data, the interference of inconsistent formats and noise information in the original international event data on the accuracy of subsequent semantic matching was eliminated; the time information was standardized to provide a standardized data foundation for the unified calculation of the time decay weight; the basic international event dataset was augmented to expand the event sample size while maintaining the original data correlation, thereby improving the data foundation quality and assessment stability of the subsequent risk transmission assessment.

[0019] (4) By replacing numerical parameters with corresponding numerical values ​​of higher or lower levels and replacing degree descriptions with expressions of stronger or weaker intensity, intermediate texts covering different degrees of severity were generated while preserving the original semantic framework. The large language model was called to generate enhanced event samples and constrain the consistency of entity information, so that the enhanced samples and the original events maintained the core semantic consistency. Qualified samples were selected through semantic similarity verification to supplement the dataset, which improved the sample diversity and semantic coverage of the dataset without excessively increasing the cost of manual annotation, and provided more sufficient data support for subsequent map construction.

[0020] (5) By encoding event text and risk type descriptions into semantic vectors and calculating similarity through a pre-trained model, deep semantic association calculation is achieved by utilizing the semantic representation capability of the pre-trained model; the weights of the similarity matrix are optimized based on timestamps, so that the semantic matching results can be dynamically adjusted in combination with time factors; or by calling a large language model through structured prompts to perform semantic reasoning and parse the weights, the reasoning capability of the large language model is used to directly parse the association strength between events and risk types from the text level. Both of the above methods realize direct semantic processing of unstructured event text descriptions without relying on manual rules or structured feature engineering, thereby improving the automation level of semantic related weight calculation and the depth of semantic understanding.

[0021] (6) The initial activation value of each event node is set to the same preset constant, so that all event nodes have a unified activation benchmark in the initial state of graph propagation calculation. This avoids unfair weighting between event nodes due to differences in initial activation values, and ensures that the difference in subsequent node scores comes only from the differentiated contributions of semantic related weights and time decay weights. This ensures the consistency and comparability of the calculation of event node contribution in risk transmission assessment.

[0022] (7) The time decay weight, which decreases exponentially with the increase of time difference, is calculated by the exponential decay function based on the time difference and the configurable decay rate parameter. This makes the time difference between the event occurrence time and the assessment benchmark time quantified into a continuously decreasing weight value. The weight is stored on the corresponding event node, so that the time attribute is embedded in the graph node in the form of a weight. In the subsequent graph propagation, it can directly participate in the weighting calculation, realize the differentiated impact of the event time attribute on the risk score, and improve the timeliness of the assessment.

[0023] (8) Extract all complete transmission paths by backtracking from the target risk outcome node, and visualize the abstract risk transmission process as a traceable node sequence; by multiplying the time decay weight, semantic relevance weight and risk type node of the event node on the path with the preset edge weight of the target risk outcome node to obtain the path contribution score, the multi-weight collaborative quantification of the risk contribution of a single transmission path is realized; by screening the main transmission paths through the contribution score and determining the key event nodes according to the comprehensive contribution and node convergence, the assessment report not only includes the comprehensive risk score, but also provides the ability to visualize and track the transmission path and locate key nodes, which significantly improves the interpretability and decision support value.

[0024] (9) An event node, a risk type node, and a unique target risk outcome (war outbreak) node were created for each international event data point, establishing a three-layer node architecture of event-risk type-outcome. For corresponding node pairs with semantic relevance weights greater than a preset threshold, directed edges were created with the semantic relevance weight as the edge weight, avoiding interference from noisy connections with low semantic relevance to graph propagation. Directed edges pointing to the target risk outcome node were created for each risk type node, establishing a standardized transmission channel from risk type to assessment result. The constructed three-layer knowledge graph maintains the integrity of the transmission chain while having a clear hierarchical structure and controllable connection density, providing a structured graph data foundation for graph propagation computation and path backtracking. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a simplified flowchart of an embodiment of the risk transmission assessment method based on knowledge graphs and semantic reasoning of the present invention; Figure 2 This is a framework diagram of an embodiment of the risk transmission assessment system based on knowledge graphs and semantic reasoning of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] like Figure 1 As shown, the present invention provides a risk transmission assessment method based on knowledge graphs and semantic reasoning, which includes the following steps: Obtain the dataset of international events to be evaluated, where each international event data entry includes at least an event ID, an event text description, and its corresponding timestamp; The event text description of each international event data is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data and each preset risk type, and an event-type weight matrix is ​​generated based on all international event data. A three-level knowledge graph is constructed based on an event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk outcome nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk outcome nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and iteratively adjusted during the graph propagation calculation process. Calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time; Phase 1 Propagation: For each risk type node, calculate the product of the initial activation value, time decay weight, and semantic relevance weight of each associated event node, and sum the products corresponding to the same risk type node to obtain the node score for that risk type node. The specific calculation formula is as follows: ;in, For the i-th international event data Time decay weight, For the i-th international event data With the j-th preset risk type Semantic relevance weights between them Represents the data of the i-th international event. The initial activation value of the corresponding event node; The second stage of propagation involves multiplying the node scores of each risk type node by the preset edge weights pointing to the target risk outcome node, and then summing the results to obtain the comprehensive risk score. The specific calculation formula is as follows: ;in, For the j-th preset risk type The preset edge weights corresponding to the risk type nodes pointing to the target risk result nodes; furthermore, it is also possible to... Perform normalization processing to map it to the interval [0, 1]; Based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, which includes the risk transmission path and contribution ranking.

[0028] This invention achieves automated conversion from unstructured text to structured risk semantics by semantically matching international event text descriptions with preset risk types and directly calculating semantically relevant weights, thus improving the automation level and semantic understanding depth of the assessment. By constructing a three-level knowledge graph of event nodes, risk type nodes, and target risk outcome nodes, the risk transmission path is explicitly decomposed into a hierarchical structure, solving the problems of unclear transmission chains and poor interpretability caused by flat graph structures. Introducing time decay weights into node score calculations allows recent and historical events to contribute differently to risk scoring, enhancing timeliness. Through semantically relevant weight edges between event nodes and risk type nodes, preset edge weights between risk type nodes and target risk outcome nodes, and a two-stage weighted aggregation scoring mechanism, semantic weights, time weights, and graph propagation weights work synergistically, achieving layer-by-layer transmission calculation from event semantics to risk type and then to comprehensive risk. This invention significantly improves the automation, accuracy, interpretability, and timeliness of international conflict risk transmission assessment.

[0029] In this embodiment, the preset risk types include military mobilization, diplomatic crisis, economic sanctions, cyberattacks, territorial changes, political turmoil, economic pressure, and foreign interference.

[0030] The preset risk types of this invention are military mobilization, diplomatic crisis, economic sanctions, cyberattack, territorial changes, political turmoil, economic pressure and foreign intervention. It provides a standardized and domain-adaptive risk classification system for international conflict risk assessment, which can be directly called upon in subsequent semantic matching and map construction, thereby improving the pertinence and consistency of the assessment.

[0031] In this embodiment, obtaining the international event dataset to be evaluated includes: The process involves acquiring raw international event data and batch processing (32-64 entries per batch) to standardize the format of event descriptions (parsing unstructured or semi-structured data from different data sources into plain text) and remove noise (removing original HTML tags, special characters, garbled text, etc.) to obtain basic event text descriptions. Preferably, the raw international event data is sourced from highly reliable data sources, including international news media texts, publicly available intelligence from official media, satellite imagery, and authoritative international data platforms (such as GDELT: Global Database of Events, Language, and Tone). Specifically, large-scale models can be used for event data extraction, and JSONL format can be used for streaming storage. Furthermore, the event descriptions in the raw international event data can be deduplicated using algorithms such as Locality Sensitive Hashing (LSH) or SimHash to calculate the similarity of the basic event text descriptions, filtering out highly repetitive events and retaining only one valid record. The time information in the raw international event data is standardized to obtain a unified format timestamp. Specifically, various time formats from different sources are parsed and converted into the ISO 8601 format of the Coordinated Universal Time standard. For events with missing timestamps but known source time, the collection time can be added as a marker to complete the timestamp. Based on the original international event data, corresponding event IDs are generated, and the basic event text descriptions and their corresponding timestamps are associated and combined with the corresponding event IDs to obtain the basic international event dataset; Data augmentation is performed on the basic international event dataset to obtain the international event dataset to be evaluated.

[0032] This invention eliminates the interference of inconsistent formats and noise in the original international event data on the accuracy of subsequent semantic matching by standardizing the format of event descriptions and denoising the data. It also standardizes the time information to provide a standardized data foundation for the unified calculation of time decay weights. Furthermore, it performs data augmentation on the basic international event dataset to expand the event sample size while maintaining the original data correlation, thereby improving the data foundation quality and assessment stability for subsequent risk transmission assessment.

[0033] In this embodiment, data augmentation processing is performed on the basic international event dataset to obtain the international event dataset to be evaluated, including: Semantic parsing is performed on the text descriptions of basic events in the basic international event dataset to identify numerical parameters and associated degree descriptions. For example, dependency parsing and named entity recognition are performed on each basic event text description to locate numerical parameters that represent quantifiable degrees such as scale, quantity, and level, including: military size (e.g., "100,000"), number of weapons, economic amount, sanctions magnitude, time window length, and troop movement frequency. At the same time, the units associated with the numerical values ​​(e.g., "ten thousand", "army", "hundred million US dollars", "day") and comparative descriptions (e.g., "large-scale", "partial", "comprehensive") are extracted to construct modifiable parameter slots. Based on a preset severity grading system, the values ​​are adjusted by multiples of the grade or by a fixed step size. Numerical parameters are replaced with values ​​corresponding to higher or lower grades, and severity descriptions are replaced with expressions of stronger or weaker intensity to generate multiple intermediate texts. The preset severity grading system sets severity levels for each numerical parameter based on domain knowledge. For example, for military size: small (+20%), medium (+50%), large (+100%), and extreme (+200%). Severity levels can be pre-configured and can be flexibly adjusted for different application scenarios. The large language model is invoked to process the data of each intermediate text to generate enhanced event samples. The entity information in the enhanced event samples is consistent with the entity information in the basic event text description. Specifically, the entire event sentence is rewritten using a predefined template or the large language model to ensure that the modified text is natural, fluent and semantically consistent, and that the entity information (such as subject and location) remains unchanged. Calculate the semantic similarity between each enhanced event sample and the corresponding basic event text description, and verify the validity of the enhanced event samples based on a preset similarity threshold (preferably 0.85); For each enhanced event sample that passes the validity verification (preferably an enhanced event sample with a similarity greater than a preset threshold), a new event ID is generated. The enhanced event sample that passes the validity verification, the new event ID, and its corresponding original timestamp are then associated and combined to supplement the basic international event dataset, generating the international event dataset to be evaluated. Furthermore, for each enhanced event sample that passes the validity verification, an `augment_info` field can be added, annotating the original event ID, enhancement level (e.g., "severity - moderate"), modified parameters, and change ratio of the enhanced event sample. This field is then associated and combined with the enhanced event sample that passes the validity verification, the new event ID, and its corresponding original timestamp to supplement the basic international event dataset, generating the international event dataset to be evaluated, facilitating traceability and analysis.

[0034] This invention generates intermediate text covering different severity levels by replacing numerical parameters with corresponding values ​​at higher or lower levels and replacing degree descriptions with expressions of stronger or weaker intensity, while preserving the original semantic framework. It calls a large language model to generate enhanced event samples and constrains the consistency of entity information, ensuring that the enhanced samples maintain core semantic consistency with the original events. By filtering qualified samples through semantic similarity verification to supplement the dataset, this invention improves the sample diversity and semantic coverage of the dataset without excessively increasing the cost of manual annotation, providing more sufficient data support for subsequent graph construction.

[0035] In this embodiment, the event text description of each international event data is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data and each preset risk type, including: The event text description of each international event data point and the description text of each preset risk type are encoded into semantic vectors using a pre-trained model (obtaining the 768-dimensional vector representation corresponding to the [CLS] tag), generating an (N×M)-dimensional similarity matrix (N is the number of international event data points, and M is the number of preset risk types). Based on the timestamp corresponding to each international event data point, the similarity matrix is ​​weighted to obtain the semantic relevance weights between each international event data point and each preset risk type. The specific calculation formula for weight optimization is as follows: ;in, Represents semantic relevance weights. Represents the similarity matrix. This indicates the configurable attenuation rate parameter. This represents the time difference between the timestamp corresponding to each international event data point (i.e., the timestamp of the event node corresponding to each international event data point) and the evaluation baseline time, converted into a value in days.

[0036] Furthermore, a similarity threshold (e.g., 0.15) can be set to filter out weak associations below the threshold and reduce noise interference; and softmax normalization can be performed on the corresponding similarities of all preset risk types in the event text description of each international event data to ensure that the weight sum is 1 and output standardized weights in the range of [0, 1]; through an expert sampling verification mechanism, abnormally high / low weights (e.g., confidence < 0.6) results are manually reviewed to establish a feedback closed-loop optimization model.

[0037] Alternatively, semantic matching can be performed between the event text description of each international event data point and a preset risk type set to calculate the semantic correlation weight between each international event data point and each preset risk type, including: A structured prompt template is constructed, which includes role definition (e.g., you are a risk analysis expert), task description (e.g., please evaluate the semantic relevance between the event text description of the following international event data and the preset risk types), event text description, preset risk type set, and output format constraints. Based on the structured prompt template, a large language model is invoked to perform semantic reasoning on the event text description of each international event data. The semantic relevance weights of each preset risk type are parsed from the returned semantic reasoning results, and the parsing results are numerically verified and corrected to obtain the semantic relevance weights between each international event data and each preset risk type.

[0038] In this embodiment, the preset risk type set includes preset risk types and standard description text corresponding to each preset risk type (which can be predefined by domain experts based on historical cases and professional knowledge).

[0039] This invention encodes event text and risk type descriptions into semantic vectors using a pre-trained model and calculates their similarity, leveraging the semantic representation capabilities of the pre-trained model to achieve deep semantic association calculation. It optimizes the weights of the similarity matrix based on timestamps, allowing the semantic matching results to be dynamically adjusted in conjunction with time factors. Alternatively, it invokes a large language model through structured prompts to perform semantic reasoning and parse the weights, utilizing the reasoning capabilities of the large language model to directly analyze the association strength between events and risk types at the text level. Both of these methods achieve direct semantic processing of unstructured event text descriptions without relying on manual rules or structured feature engineering, improving the automation level of semantic weight calculation and the depth of semantic understanding.

[0040] In this embodiment, the initial activation value of each event node is the same preset constant.

[0041] This invention sets the initial activation value of each event node to the same preset constant, so that all event nodes have a unified activation benchmark in the initial state of graph propagation calculation. This avoids unfair weighting between event nodes due to differences in initial activation values, and ensures that the difference in subsequent node scores comes only from the differentiated contributions of semantic relevance weight and time decay weight. This guarantees the consistency and comparability of the calculation of event node contribution in risk transmission assessment.

[0042] In this embodiment, the time decay weight of each event node is calculated based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time, including: Obtain the time difference between the timestamp of each event node and the evaluation baseline time, as well as the configurable decay rate parameter; The time decay weight, which decreases exponentially with increasing time difference, is calculated using the exponential decay function based on the time difference and decay rate parameters. The time decay weight is then stored in the corresponding event node; time decay weight The specific calculation formula is as follows: ;in, This indicates the configurable attenuation rate parameter. This represents the time difference between the timestamp of each event node and the evaluation baseline time, converted into a value in days.

[0043] This invention calculates a time decay weight that decreases exponentially with increasing time difference using an exponential decay function based on the time difference and a configurable decay rate parameter. This quantifies the time difference between the event occurrence time and the assessment benchmark time as a continuously decreasing weight value. This weight is stored on the corresponding event node, allowing the time attribute to be embedded into the graph node in the form of a weight. In subsequent graph propagation, it can directly participate in the weighted calculation, realizing the differentiated impact of the event time attribute on the risk score and improving the timeliness of the assessment.

[0044] In this embodiment, based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, including risk transmission paths and contribution rankings, comprising: Starting from the target risk outcome node, backtrack along the directed edges of the three-level knowledge graph to extract all complete transmission paths from the event node through the risk type node to the target risk outcome node; For each complete transmission path, the contribution score of the path is obtained by multiplying the time decay weight of the event node, the semantic correlation weight between the event node and the risk type node, and the preset edge weight between the risk type node and the target risk result node. All complete transmission paths are sorted in descending order based on their contribution scores. The set of paths with contribution scores greater than a preset contribution score (preferably the top 20%) or the set of paths with a cumulative contribution score percentage greater than a preset percentage threshold (preferably 80%) are selected as the main transmission paths. For each event node, the sum of the products of all semantically related weights associated with it and the preset edge weights of the corresponding risk type nodes is multiplied by the time decay weight of the event node to obtain the overall contribution of the event node. Key event nodes are determined based on their overall contribution and the degree of convergence of each event node in the transmission path (a "bridging" event that serves as a convergence point for multiple paths is considered a key event if removing the event node would significantly decrease the final risk score). Specifically, key event nodes can be determined by quantifying the degree of convergence of each event node in the transmission path by the number of times each event node appears as an intermediate node in all transmission paths, or by calculating the statistical distribution of the contribution of each event and identifying the event nodes corresponding to outlier events with an overall contribution significantly higher than the average level as key event nodes. Generate a risk transmission assessment report that includes a comprehensive risk score, a ranking of the contribution of nodes for each risk type, a list of key event nodes, and a sequence of nodes along the main transmission path.

[0045] Furthermore, the risk transmission assessment report also includes the risk level (determined based on the comprehensive risk score), as follows: Low risk: The comprehensive risk score ranges from [0, 0.3]. The impact of the event transmission chain is weak, the probability of the result is low, and routine monitoring is maintained.

[0046] Moderate risk: The comprehensive risk score ranges from [0.3 to 0.6]: There are obvious transmission events and paths, and it is necessary to continuously monitor and analyze key events and types, and prepare corresponding contingency plans.

[0047] High risk: The comprehensive risk score ranges from [0.6, 0.8]: Multiple strongly correlated transmission paths are activated, the probability of the result is high, and an early warning should be issued immediately and action should be taken.

[0048] High risk: The comprehensive risk score ranges from extremely high risk [0.8, 1.0]: multiple paths are highly aggregated, a landmark event occurs, the outcome is highly likely, and an emergency response state is entered.

[0049] This invention extracts all complete transmission paths by tracing back from the target risk outcome node, concretizing the abstract risk transmission process into a traceable sequence of nodes. By multiplying the time decay weight, semantic relevance weight, and risk type node of the event node on the path with the preset edge weight of the target risk outcome node, a path contribution score is obtained, realizing multi-weight collaborative quantification of the risk contribution of a single transmission path. The main transmission paths are selected through contribution scores, and key event nodes are determined based on comprehensive contribution and node convergence. This makes the assessment report not only include a comprehensive risk score, but also provides the ability to visualize and track the transmission path and locate key nodes, significantly improving interpretability and decision support value.

[0050] In this embodiment, a three-level knowledge graph is constructed based on an event-type weight matrix, comprising at least event nodes, risk type nodes, and target risk outcome nodes, including: Create a corresponding event node for each piece of international event data (the node attributes include event_id and timestamp), create a corresponding risk type node for each preset risk type, and create a unique target risk result node (representing the target risk to be assessed in the end), the target risk result being the outbreak of war; For a corresponding event node whose semantic relevance weight is greater than a preset weight threshold (preferably 0.1), a directed edge is created between the event node and the corresponding risk type node, and the semantic relevance weight is used as the weight of the directed edge. For each risk type node, create a directed edge pointing to a unique target risk outcome node.

[0051] This invention creates an event node for each international event data point, a risk type node for each preset risk type, and a unique target risk outcome node (war outbreak), establishing a three-layer node architecture of event-risk type-outcome. For node pairs with semantic relevance weights greater than a preset threshold, directed edges are created using this semantic relevance weight as the edge weight, avoiding interference from noisy connections with low semantic relevance on graph propagation. A directed edge pointing to the target risk outcome node is created for each risk type node, establishing a standardized transmission channel from risk type to assessment result. The constructed three-layer knowledge graph maintains the integrity of the transmission chain while possessing a clear hierarchical structure and controllable connection density, providing a structured graph data foundation for graph propagation computation and path backtracking.

[0052] In this embodiment, the initial activation value of each event node is the same preset constant (preferably 1).

[0053] This invention sets the initial activation value of each event node to the same preset constant, so that all event nodes have a unified activation benchmark in the initial state of graph propagation calculation. This avoids unfair weighting between event nodes due to differences in initial activation values, and ensures that the difference in subsequent node scores comes only from the differentiated contributions of semantic relevance weight and time decay weight. This guarantees the consistency and comparability of the calculation of event node contribution in risk transmission assessment.

[0054] like Figure 2 As shown, the present invention also provides a risk transmission assessment system based on knowledge graphs and semantic reasoning, which includes: The dataset acquisition module 10 acquires the international event dataset to be evaluated, wherein each international event data includes at least an event ID, an event text description and its corresponding timestamp; The semantic matching module 20 is used to semantically match the event text description of each international event data with a preset risk type set, and calculate the semantic correlation weight between each international event data and each preset risk type. The weight matrix generation module 30 is used to generate an event-type weight matrix based on all international event data; The knowledge graph construction module 40 is used to construct a three-level knowledge graph based on the event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk result nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk result nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and iteratively adjusted during the graph propagation calculation process. The first calculation module 50 is used to calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation benchmark time. The second calculation module 60 is used to calculate the product of the initial activation value, time decay weight and semantic relevance weight of each event node associated with each risk type node, and to sum the products corresponding to the same risk type node to obtain the node score of that risk type node. The third calculation module 70 is used to multiply the node scores of each risk type node by the preset edge weights of their respective points to the target risk result nodes and then sum them to obtain a comprehensive risk score. The risk transmission assessment report generation module 80 is used to generate a risk transmission assessment report, including risk transmission path and contribution ranking, based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph.

[0055] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0056] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0057] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A risk transmission assessment method based on knowledge graphs and semantic reasoning, characterized in that, Includes the following steps: Obtain the dataset of international events to be evaluated, where each international event data entry includes at least an event ID, an event text description, and its corresponding timestamp; The event text description of each international event data is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data and each preset risk type, and an event-type weight matrix is ​​generated based on all international event data. A three-level knowledge graph is constructed based on an event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk outcome nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk outcome nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and iteratively adjusted during the graph propagation calculation process. Calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time; For each risk type node, calculate the product of the initial activation value, time decay weight, and semantic relevance weight of each event node associated with it, and sum the products corresponding to the same risk type node to obtain the node score of that risk type node. The node scores of each risk type node are multiplied by the preset edge weights pointing to the target risk result nodes, and then summed to obtain the comprehensive risk score. Based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, which includes the risk transmission path and contribution ranking.

2. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, The preset risk types include military buildup, diplomatic crisis, economic sanctions, cyberattacks, territorial changes, political instability, economic pressure, and foreign interference.

3. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, Obtain the dataset of international events to be evaluated, including: Obtain raw international event data, and perform format unification and noise reduction processing on the event descriptions in the raw international event data to obtain basic event text descriptions; The time information in the raw international event data is standardized to obtain a timestamp in a unified format. Based on the original international event data, corresponding event IDs are generated, and the basic event text descriptions and their corresponding timestamps are associated and combined with the corresponding event IDs to obtain the basic international event dataset; Data augmentation is performed on the basic international event dataset to obtain the international event dataset to be evaluated.

4. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 3, characterized in that, Data augmentation was performed on the basic international event dataset to obtain the international event dataset to be evaluated, including: Semantic parsing is performed on the textual descriptions of basic events in the basic international event dataset to identify numerical parameters and the degree of correlation. Based on a preset severity grading system, numerical parameters are replaced with values ​​corresponding to higher or lower levels, and severity descriptions are replaced with expressions of stronger or weaker intensity to generate multiple intermediate texts. The large language model is invoked to process the data of each intermediate text and generate enhanced event samples; the entity information in the enhanced event samples is consistent with the entity information in the basic event text description. Calculate the semantic similarity between each enhanced event sample and the corresponding basic event text description, and verify the validity of the enhanced event samples based on a preset similarity threshold; For each enhanced event sample that passes the validity check, a new event ID is generated. The enhanced event sample that passes the validity check, the new event ID, and its corresponding original timestamp are then associated and combined and added to the basic international event dataset to generate the international event dataset to be evaluated.

5. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, The event text description of each international event data point is semantically matched with a preset risk type set to calculate the semantic correlation weight between each international event data point and each preset risk type, including: The event text description of each international event data and the description text of each preset risk type are encoded into semantic vectors by a pre-trained model to generate a similarity matrix. Based on the timestamp corresponding to each international event data, the weights of the similarity matrix are optimized to obtain the semantic correlation weights between each international event data and each preset risk type. Alternatively, a structured prompt template can be constructed, which includes role definition, task description, event text description, preset risk type set and output format constraints. Based on the structured prompt template, a large language model is called to perform semantic reasoning on the event text description of each international event data. The semantic correlation weights of each preset risk type are parsed from the returned semantic reasoning results, and the parsing results are numerically verified and corrected to obtain the semantic correlation weights between each international event data and each preset risk type.

6. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, The initial activation value of each event node is the same preset constant.

7. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, Based on the time difference between the timestamp corresponding to each event node and the evaluation baseline time, calculate the time decay weight of each event node, including: Obtain the time difference between the timestamp of each event node and the evaluation baseline time, as well as the configurable decay rate parameter; The time decay weight, which decreases exponentially with the increase of time difference, is calculated based on the time difference and decay rate parameters using the exponential decay function, and is then stored in the corresponding event node.

8. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, Based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph, a risk transmission assessment report is generated, including risk transmission paths and contribution rankings, comprising: Starting from the target risk outcome node, backtrack along the directed edges of the three-level knowledge graph to extract all complete transmission paths from the event node through the risk type node to the target risk outcome node; For each complete transmission path, the contribution score of the path is obtained by multiplying the time decay weight of the event node, the semantic correlation weight between the event node and the risk type node, and the preset edge weight between the risk type node and the target risk result node. All complete transmission paths are sorted in descending order based on their contribution scores, and the set of paths with contribution scores greater than a preset contribution score threshold or the set of paths with a cumulative contribution score percentage greater than a preset percentage threshold are selected as the main transmission paths. For each event node, the sum of the products of all semantically related weights associated with it and the preset edge weights of the corresponding risk type nodes is multiplied by the time decay weight of the event node to obtain the overall contribution of the event node. Key event nodes are identified based on their overall contribution and the degree of convergence of each event node in the transmission path. Generate a risk transmission assessment report that includes a comprehensive risk score, a ranking of the contribution of nodes for each risk type, a list of key event nodes, and a sequence of nodes along the main transmission path.

9. The risk transmission assessment method based on knowledge graph and semantic reasoning according to claim 1, characterized in that, Construct a three-level knowledge graph based on an event-type weight matrix, comprising at least event nodes, risk type nodes, and target risk outcome nodes, including: Create a corresponding event node for each piece of international event data, create a corresponding risk type node for each preset risk type, and create a unique target risk outcome node, with the specific target risk outcome being the outbreak of war. For a corresponding event node whose semantic relevance weight is greater than a preset weight threshold, a directed edge is created between the event node and the corresponding risk type node, and the semantic relevance weight is used as the weight of the directed edge. For each risk type node, create a directed edge pointing to a unique target risk outcome node.

10. A risk transmission assessment system based on knowledge graphs and semantic reasoning, characterized in that, include: The dataset acquisition module acquires the international event dataset to be evaluated, where each international event data item includes at least an event ID, an event text description, and its corresponding timestamp; The semantic matching module is used to semantically match the event text description of each international event data with a preset risk type set, and calculate the semantic correlation weight between each international event data and each preset risk type. The weight matrix generation module is used to generate an event-type weight matrix based on all international event data; The knowledge graph construction module is used to construct a three-level knowledge graph based on the event-type weight matrix, which includes at least event nodes, risk type nodes, and target risk result nodes. The directed edges between event nodes and risk type nodes are assigned semantically related weights, and the directed edges between risk type nodes and target risk result nodes are configured with preset edge weights. These preset edge weights are initialized based on historical statistical data and are iteratively adjusted during the graph propagation calculation process. The first calculation module is used to calculate the time decay weight of each event node based on the time difference between the timestamp corresponding to each event node and the evaluation benchmark time. The second calculation module is used to calculate the product of the initial activation value, time decay weight and semantic relevance weight of each event node associated with each risk type node, and to sum the products corresponding to the same risk type node to obtain the node score of that risk type node. The third calculation module is used to multiply the node scores of each risk type node by the preset edge weights of their respective nodes pointing to the target risk result nodes and then sum them to obtain a comprehensive risk score. The risk transmission assessment report generation module is used to generate a risk transmission assessment report, including risk transmission path and contribution ranking, based on the comprehensive risk score and the weights of each node and directed edge in the three-level knowledge graph.

Images