A hotspot event early warning method and device

Abstract

The application discloses a hotspot event early warning method and device, belonging to the field of natural language processing, including the following steps: based on domain knowledge, constructing a correlation index network, combining background modeling information, and assigning weights to the indexes according to importance; constructing a basic intelligent algorithm small model, solidifying the network index system into a business data small model to form a small model set; taking the large language model and the small model set as the basis, taking the large language model as the scheduling model, scheduling the basic intelligent algorithm small model and the business data small model, and discovering the hotspot event indicators; performing indicator weighting calculation based on the analysis formula to obtain the hotspot event early warning probability, and giving early warning information when the early warning threshold is exceeded; and comparing with the real situation, based on the difference of the comparison, and through the adaptive feedback of the large model, the index network is optimized. The application realizes the hotspot event early warning driven by the index network and the large model, and can effectively improve the timeliness and interpretability of the hotspot event early warning.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and more specifically, to a method and apparatus for early warning of hot events. Background Technology

[0002] With the rapid development of information technology, social media and digital platforms are generating an unprecedented deluge of cross-domain, heterogeneous data. In this context, hotspot event early warning has evolved from a theoretical concept into a key technological paradigm integrating information science and social computing. Its core lies in constructing a closed-loop system of dynamic perception, risk assessment, and decision support: by real-time analysis of diverse information sources such as social network semantic streams and news corpora, it accurately captures abnormal patterns and generates risk transmission path maps and tiered early warning labels, thus providing a quantifiable risk prediction window for emergency management and strategic decision-making. Hotspot event early warning can play a significant role in multiple fields. In public safety, it can achieve second-level triggering of post-earthquake emergency response based on social media text mining; in the financial sector, it can significantly improve the accuracy of black swan event prediction by constructing investor sentiment-market volatility correlation models; in social management, it can preemptively block the fermentation chain of public opinion based on group interaction behavior clustering algorithms; in business competition, it can help companies seize market opportunities through dynamic tracking of brand voice evolution. Its essence has transcended its tool-like nature, evolving into the intelligent nerve center of social governance.

[0003] With the development of artificial intelligence technology, early warning of hot events has transformed from a passive response mechanism to a proactive prevention mechanism, undergoing a three-stage leap: from static analysis dominated by symbolic logic, to dynamic knowledge distillation through data-model co-evolution, and finally to autonomous decision-making based on generative reasoning. Specifically, early systems relied on a "rule-matching" closed loop formed by manually created semantic rule bases, which was deeply mired in semantic gaps, leading to a vicious cycle of contextual misjudgment and false alarms-ignores; as the marginal cost of computing resources approached zero, machine learning architectures propelled early warning logic from expert-driven to data-expert dual-driven, significantly improving portability and probabilistic quantification accuracy; the topological innovation of deep neural networks ushered in a new era of evolutionary logic mining, with graph neural networks visualizing risk transmission paths, event graphs deconstructing the internal logic of events, and human-machine collaboration mechanisms upgrading early warning to decision empowerment; currently, the cognitive revolution driven by large language models achieves adaptive question-and-answer transformation for event prediction through prompting engineering, integrates causal reasoning to construct dynamic knowledge graphs, and completes multi-granular joint modeling of event sequences, marking the early warning system's shift from event recognition to social evolution deduction, becoming a "cognitive accelerator" of digital civilization.

[0004] In general, while early warning systems for hot events have achieved some success, current early warning technologies suffer from a fundamental bottleneck due to a mismatch in cognitive frameworks. The outbreak of a hot event is essentially a dynamic game system involving multiple agents, encompassing social psychology, technological dissemination, and environmental variables. Its evolutionary trajectory exhibits strong nonlinearity and path dependence. Traditional reductionist modeling based on indicator systems is too idealistic and cannot simulate the cascading propagation mechanism of risks in complex networks. This leads to a continuous decline in the effectiveness of early warning models in identifying systemic risks. Therefore, there is still a long way to go before practical implementation and engineering applications, which indirectly affects the advancement and progress of early warning work for hot events. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for early warning of hot events. Targeting the complex adaptive characteristics of hot events, it conducts modeling analysis from the perspective of multiple groups, characterizes the logical connections in event early warning, constructs an indicator network to describe the behavioral interactions between groups, and realizes a hot event early warning method driven by both the indicator network and the large model through intelligent collaborative reasoning of large and small models. This can effectively improve the timeliness and interpretability of hot event early warning.

[0006] The objective of this invention is achieved through the following solution: A method for early warning of hot events includes the following steps: Based on domain knowledge, a network of related indicators is constructed. Combined with background modeling information, the indicators are hierarchically and weighted according to their importance to form a network indicator system with coupled influence. Starting from the need for early warning of hot events, we construct basic intelligent algorithm mini-models, solidify the network indicator system into business data mini-models, and form a set of mini-models. Based on the large language model and the set of small models, the large language model is used as the scheduling model to schedule the basic intelligent algorithm small models and business data small models to discover hot event symptoms and form a hot event symptom set. Based on the aforementioned hotspot event symptom set, a weighted calculation of the symptom is performed using an analytical formula to obtain the early warning probability of the hotspot event. If the probability exceeds the early warning threshold, an early warning message is given and compared with the actual situation. Based on the differences in the comparison, the indicator network is optimized through adaptive feedback of a large model.

[0007] Furthermore, the construction of the associated indicator network, combined with background modeling information, involves hierarchically and hierarchically assigning weights to the indicators according to their importance, forming a network indicator system with coupled influences. This process specifically includes the following sub-steps: A dynamic indicator network is constructed, which includes multiple dynamic indicator subnetworks. Each subnetwork contains multiple types of indicators, which are divided into two layers according to their importance. Some types of indicators are located at the bottom layer, and other types of indicators are located at the top layer. The bottom layer is more important than the top layer. The indicators in each layer influence each other, i.e., they are mutually coupled. The indicators in the same subnetwork influence each other, i.e., they are self-coupled. The relationships within and between sub-networks are reflected through the superposition of indicators. At the same time, the superposition is constrained by background information. The same background information has a positive correlation, while different background information has a negative correlation. The superposition effect is divided into two categories: strengthening and weakening.

[0008] Furthermore, the basic intelligent algorithm mini-model specifically includes a spatiotemporal analysis model, an event detection model, and an event analysis model.

[0009] Furthermore, the business data mini-model specifically includes hierarchical indicators, correlation indicators, and group weight mini-models.

[0010] Furthermore, the process of using a large language model as a scheduling model to schedule basic intelligent algorithm mini-models and business data mini-models to discover hotspot event symptoms, and based on the hotspot event symptom set, performing symptom weighted calculations based on analytical formulas to obtain the hotspot event warning probability, issuing warning information if it exceeds the warning threshold, and comparing it with the actual situation, and based on the differences in the comparison, optimizing the indicator network through adaptive feedback of the large model, specifically includes the following sub-steps: Input layer: Extracts the data needed by the model from open source sources; Small model understanding layer: The text-based small model is based on the Roberta pre-trained language model to perform event detection and event classification; the spatiotemporal small model is based on the ray method, which determines the location of the target activity, i.e. whether it is active in the region of interest, by calculating the number of intersections between points and rays. Large Model Inference Layer: A general scheduling model is set up, which schedules the text-based small models fixed in the small model understanding layer. These models perform event detection and classification on all text data acquired from the input layer to obtain event sentences. The spatiotemporal small models fixed in the small model understanding layer determine the current target activity location and transcribe it into text information. Then, the text is vectorized using a vector embedding model to obtain text embedding vectors. Finally, the scheduling model schedules the business data small models, performing vector calculations on the text embedding vectors and the indicator embedding vectors in the business data small models to analyze semantic similarity. If the semantic similarity exceeds a preset threshold, it is determined to be a hotspot event symptom. Hotspot Event Early Warning Layer: This layer performs weighted calculations on the symptoms to obtain the probability of a hotspot event warning and feeds back the warning results. Specifically, it uses a large-scale model prompt word correction method to transform the warning results into a large-scale model prompt word optimization strategy. This includes optimization strategies for indicators and indicator weights in the indicator network. For example, if a hotspot event is warned of but does not actually occur, the large-scale model prompt adds descriptions to reduce indicator weights and the number of indicators involved in the calculation. If a hotspot event is not warned of but actually occurs, the large-scale model prompt adds descriptions to increase indicator weights, generate similar new indicators based on the indicator network and the current event, and increase the number of indicators involved in the calculation. Through the automatic prompt word optimization strategy, the hotspot event early warning evaluation model is dynamically constructed, continuously iteratively optimizing the hotspot event early warning effect. The generated new indicators are solidified into the underlying indicator network, transforming the static indicator network manually constructed based on domain expert experience into a continuously optimized dynamic indicator network.

[0011] Furthermore, the data obtained at the input layer specifically includes two categories: text data and spatiotemporal data.

[0012] Furthermore, the event detection results are divided into two categories: events and non-events. The event classification results are further divided into six categories: security, politics, economy, diplomacy, military, and technology.

[0013] Furthermore, the overall scheduling model specifically includes the large language model Qwen3-30B.

[0014] Furthermore, the step of performing grouped and weighted calculations on the symptoms to obtain the early warning probability of hotspot events specifically includes the following sub-steps: First, calculate the independent early warning probability P1 of the dynamic index subnetwork: (1); In the formula, This represents the probability of issuing early warnings for hot events based on independent predictions of pre-selected indicators from a dynamic indicator subnetwork, i.e., the probability of independent early warnings from a dynamic indicator subnetwork. G represents the number of dynamic indicator subnetworks. This represents the predictive impact factor of hotspot events based on the s-th subnetwork. The specific calculation formula is as follows (2): (2); In the formula, This represents the weight of the m-th primary indicator, where M represents the total number of primary indicators. Primary indicators are removed as needed, and c is a minimum value representing a low-probability event. and , Denotes the influence factor of the m-th primary indicator and The specific calculation formula is as follows (3): (3); In the formula, This represents the weight of the e-th tertiary indicator for the m-th primary indicator. The positive or negative weight represents the positive or negative effect of the indicator on the conclusion, n represents the total number of tertiary indicators of the m-th primary indicator, and z represents the total number of tertiary indicators with positive weights of the m-th primary indicator. It is a constant that adjusts on demand, representing the maximum number of events related to the third-level indicators of the s-th dynamic indicator subnetwork. This represents the number of events related to the e-th tertiary indicator of the m-th primary indicator. The value is greater than Then let ; Then, calculate the joint early warning probability P2 of the dynamic index subnetwork: (4); In the formula, This represents the probability of a hotspot event warning based on correlation indicators within and between dynamic indicator subnetworks, i.e., the joint warning probability of dynamic indicator subnetworks. The specific calculation formula is as follows (5): (5); In the formula, This represents the predictive impact factor of hotspot events based on the correlation index between the k-th subnetwork and the h-th subnetwork. Time indicates the dynamic indicator within the subnetwork. Time indicates dynamic indicators between subnetworks; Let represent the association weight between the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, and satisfy ... F represents the total number of third-level correlation indicators in the k-th sub-network, D represents the total number of third-level correlation indicators in the h-th sub-network, and N represents the maximum number of events among the related indicators. The specific calculation formula is as follows (6): (6); In the formula, and These represent the maximum number of events associated with the third-level indicators of the k-th and h-th sub-networks, respectively, and can be adjusted as needed; in equation (5) and These represent the quantified event values ​​associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The value is greater than or The value is greater than Then let or , and The calculation formulas are as follows: (7) - (8) (7); (8); In the formula, and Let represent the number of events associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The polarity factor of the index is represented by the following formula (9): (9); Finally, the probability of a hotspot event warning, P3, is calculated using a linear weighting method, as shown in equation (10): (10); In the formula, P3 represents the probability of a hotspot event warning. , It is a scaling factor, and ; Set early warning thresholds as needed. ,when In such cases, early warnings are issued for trending events.

[0015] A hotspot event early warning device includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method described in any of the preceding methods.

[0016] The beneficial effects of this invention include: (1) Based on the pain points of the lack of correlation between indicators and the inflexible reasoning and analysis process in the existing hot event early warning, this invention proposes a complete hot event early warning framework that goes from indicator network construction to collaborative reasoning between large and small models, and then to dual-drive by indicator network and large model. Compared with the existing technology, it can improve the completeness of symptom detection and the timeliness of hot event early warning.

[0017] (2) This invention proposes an indicator network construction method and uses the indicator network as the basis for early warning of hot events. Compared with the prior art, the indicator system model of this invention is closer to the real situation and usage environment, the model has a high degree of fidelity, the early warning process is more in line with reality, and the early warning results are more effective, which can effectively improve the accuracy of early warning of hot events.

[0018] (3) This invention proposes a new paradigm for early warning of hot events based on the collaboration of large and small models. It can dynamically optimize the indicator network in real time based on the early warning results during the early warning process, making the early warning process more autonomous. Compared with the existing technology, it has higher interpretability and continuous iteration. The positive closed-loop early warning process is also more convenient for engineers to use in practice.

[0019] This invention is applicable to early warning of hot-spot events. The core of this invention is to construct a complete hot-spot event early warning framework driven by both an indicator network and a large-scale model, thereby achieving early warning of hot-spot events and improving the completeness of symptom detection and the timeliness of early warning. This invention is effective for any matter related to early warning of hot-spot events. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a general framework diagram of a hotspot event early warning method and device according to an embodiment of the present invention; Figure 2 This is a diagram of the index network structure in an embodiment of the present invention; Figure 3 This is a diagram of the collaborative early warning framework for large and small models in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the principle of calculating the probability of early warning for hot events in an embodiment of the present invention. Detailed Implementation

[0022] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0023] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 limitations, 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.

[0024] See Figure 1The overall framework of this invention's hotspot event early warning method and device consists of four parts: background modeling, dynamic indicator network construction, collaborative scheduling of large and small models, and hotspot event early warning. First, a background model is established for different analysis objects, mainly constructed from three aspects: social system, economic output, and military capability. The social system mainly includes socialism and capitalism; economic output is mainly reflected in large and small; and military capability is mainly reflected in strong and weak, supporting the correlation analysis between indicator networks. Then, a static indicator network is constructed based on domain expert experience to support hotspot event early warning analysis. During the hotspot event early warning analysis process, new indicators are automatically and continuously generated and solidified into the indicator network, forming a dynamic indicator network. The dynamic indicator network consists of multiple dynamic indicator sub-networks, each containing six categories of indicators: military, political, diplomatic, security, economic, and technological. These are divided into two layers according to their importance: security, politics, and economy are at the bottom layer, while diplomacy, military, and technology are at the top layer. The underlying layers are more important than the upper layers, and the indicators at each layer influence each other (i.e., mutual coupling). Indicators within the same sub-network also influence each other (i.e., self-coupling). Secondly, a classic intelligent algorithm mini-model for hot event early warning is constructed, solidifying the dynamic indicator network into a business data mini-model. This classic intelligent algorithm mini-model includes a spatiotemporal analysis model, an event detection model, and an event analysis model. The business data mini-model includes hierarchical indicators, correlated indicators, and grouping weights. Based on an open-source large language model, hot event symptom mining is performed through collaborative scheduling of large and small models. Finally, the symptoms are weighted to obtain the hot event early warning probability. Based on the difference between the hot event early warning result and the actual occurrence, the indicators and indicator weights in the indicator network are adaptively fed back to complete the hot event early warning and feedback.

[0025] See Figure 2The indicator network structure of this invention is mainly composed of multiple dynamic indicator sub-networks. Each sub-network is divided into three levels of indicators. The sub-networks themselves and the sub-networks are interconnected through the mutual influence of the indicators, forming a unified dynamic indicator network. The first-level indicators include six aspects: security, politics, economy, military, diplomacy, and science and technology. The first-level indicators are further decomposed into second-level indicators: security includes energy security, territorial security, and social security; politics includes statements, activities, and resumes of political figures; economy includes economic indices and living standards indices; military includes military deployments and military operations; diplomacy includes diplomatic activities and diplomatic statements; and science and technology includes high-tech output and high-tech exposure. The second-level indicators are further decomposed into tertiary indicators: energy security includes energy reserves, energy price fluctuations, and energy diversity; territorial security includes resource development area, land encroachment, and illegal immigration; and social security includes terrorist activities and demonstrations. The factors influencing political activity and public opinion include: political figures' statements (including important pronouncements, expressions of strong opinions, and important meetings); political figures' resumes (including past and current positions, and promotion paths); economic indices (including GDP, per capita disposable income, and income ratio); living standards indices (including unemployment rate and price index); military deployments (including mobilization orders and logistical support plans); military operations (including troop deployments and exercises); diplomatic activities (including diplomatic visits and participation in international conferences); diplomatic statements (including foreign policy statements and declarations on international events); high-tech output (including annual output quantity, annual output growth rate, and the scope of high-tech coverage); and high-tech exposure (including the quantity and growth rate of high-tech exposure). The relationships within and between sub-networks are reflected through the cumulative effects of these indicators. These cumulative effects are constrained by background information such as social systems, economic scale, and military capabilities. Similar background information primarily exhibits a positive correlation, while different background information primarily exhibits a negative correlation. The cumulative effects are mainly categorized into two types: strengthening and weakening. Specifically, if the price index and unemployment rate have a cumulative effect, and different sub-networks have the same social system, similar economic size, or military capabilities, a higher price index and a higher unemployment rate have a strengthening relationship, while a lower price index and a lower unemployment rate have a weakening relationship; a higher price index and a lower unemployment rate also have a weakening relationship; and a lower price index and a higher unemployment rate also have a weakening relationship. Conversely, if different sub-networks have different social systems, significantly different economic sizes, or military capabilities, a higher price index and a higher unemployment rate have a weakening relationship, while a lower price index and a lower unemployment rate also have a weakening relationship; a higher price index and a lower unemployment rate also have a weakening relationship. Within the same sub-network, that is, those with the same social system, economic size, and military capabilities, the relationship is unaffected by background information.

[0026] See Figure 3This invention's large-scale model collaborative early warning framework is mainly divided into an input layer, a small model understanding layer, a large model inference layer, and a hot event early warning layer. The input layer includes data extracted from open-source channels such as news reports, social media, and various networks, including text data and spatiotemporal data. The small model understanding layer includes text-based small models and spatiotemporal small models. The text-based small models are based on the Roberta pre-trained language model for event detection and classification. Event detection results are categorized as events or non-events, and event classification results are categorized into six types: security, politics, economy, diplomacy, military, and technology. The spatiotemporal small models are based on the ray casting method, determining the target's location (i.e., whether it is active within the area of ​​interest) by calculating the number of intersections between a point and a ray. The large model inference layer first uses an open-source large language model, such as Qwen3-30B, as the overall scheduling model, scheduling the small model understanding layer's fixed... The text-based mini-model performs event detection and classification on all text data acquired from the input layer, resulting in six categories of event sentences: security, politics, economy, diplomacy, military, and technology. The scheduling mini-model, based on the spatiotemporal mini-model fixed in the understanding layer, determines the current target activity location and transcribes it into text information. This text is then vectorized using a vector embedding model, such as BGE, to obtain text embedding vectors. The scheduling model automatically schedules the business data mini-model to perform vector calculations on the obtained text embedding vectors and the indicator embedding vectors in the business data mini-model. It analyzes the semantic similarity between the six categories of event sentences (security, politics, economy, diplomacy, military, and technology) and the target activity information and indicators. If the semantic similarity exceeds a preset threshold, such as 80%, it is determined to be a hotspot event symptom. The hotspot event warning layer then performs weighted grouping calculations on the symptom symptom to obtain the hotspot event warning probability. See details in [link to relevant documentation]. Figure 4 The system then feeds back the warning results, specifically through a large-scale model prompt word correction approach. This involves optimizing the prompt words in the large-scale model, primarily by optimizing the indicators and their weights within the indicator network. For example, if a warning of a hot event occurs but it doesn't, the large-scale model prompt includes descriptions of reducing indicator weights and decreasing the number of indicators involved in the calculation. Conversely, if a warning of a hot event occurs but it does, the large-scale model prompt includes descriptions of increasing indicator weights, generating new indicators similar to those in the current event based on the indicator network, and increasing the number of indicators involved in the calculation. Through this automatic prompt word optimization strategy, the system dynamically constructs a hot event warning assessment model, continuously iterates to optimize the warning effect, and embeds the generated new indicators into the underlying indicator network. This transforms the static indicator network, manually constructed based on domain expert experience, into a continuously optimized dynamic indicator network.

[0027] See Figure 4The hotspot event early warning system is primarily based on dynamic indicator sub-networks. It calculates the probability of a hotspot event being warned and determines whether this probability exceeds a warning threshold. If it does, a warning is issued. The specific calculation process consists of three steps: calculating the independent warning probability P1 of the dynamic indicator sub-networks, calculating the joint warning probability P2 of the dynamic indicator sub-networks, and calculating the hotspot event warning probability P3. Specifically, the independent warning probability calculation involves calculating the warning probability for each of the N different dynamic indicator sub-networks and taking the average as the independent warning probability; the joint warning probability calculation involves performing a one-to-one association calculation on each of the N different dynamic indicator sub-networks (including the sub-networks themselves) without repetition and taking the average as the joint warning probability; and the hotspot event warning probability calculation is a weighted average of the independent warning probabilities and the joint warning probabilities of the dynamic indicator sub-networks. Specifically: First, the formula for calculating the independent early warning probability of the dynamic indicator sub-network is as follows: (1); This represents the probability of issuing early warnings for hot events based on independent predictions of pre-selected indicators from a dynamic indicator subnetwork, i.e., the probability of independent early warnings from a dynamic indicator subnetwork. G represents the number of dynamic indicator subnetworks. This represents the predictive impact factor of hotspot events based on the s-th subnetwork. The specific calculation formula is shown in equation (2) below: (2); This represents the weight of the m-th primary indicator, where M represents the total number of primary indicators (which can be reduced as needed), and c is a minimum value representing a low-probability event. and , Denotes the influence factor of the m-th primary indicator and The specific calculation formula is shown in equation (3) below: (3); This represents the weight of the e-th tertiary indicator for the m-th primary indicator. The positive or negative weight represents the positive or negative effect of the indicator on the conclusion, n represents the total number of tertiary indicators of the m-th primary indicator, and z represents the total number of tertiary indicators with positive weights of the m-th primary indicator. It is a constant that can be adjusted as needed, representing the maximum number of events related to the third-level indicators of the s-th dynamic indicator subnetwork. This represents the number of events related to the e-th tertiary indicator of the m-th primary indicator. The value is greater than Then let .

[0028] Then, the formula for calculating the joint early warning probability of the dynamic indicator sub-network is as follows: (4); This represents the probability of a hotspot event warning based on correlation indicators within and between dynamic indicator subnetworks, i.e., the joint warning probability of dynamic indicator subnetworks. The specific calculation formula is shown in equation (5) below: (5); in, Represents the relationship based on the k-th subnetwork and the h-th subnetwork ( Time indicates the dynamic indicator within the subnetwork. The time represents the influence factor of hotspot event prediction on the correlation indicators between dynamic indicator sub-networks. Let represent the association weight between the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, and satisfy ... F represents the total number of third-level correlation indicators in the k-th sub-network, D represents the total number of third-level correlation indicators in the h-th sub-network, and N represents the maximum number of events among the related indicators. The specific calculation formula is shown in the following formula (6): (6); and represents the maximum number of events associated with the third-level indicators of the k-th and h-th sub-networks, respectively, and can be adjusted as needed. In equation (5) and These represent the quantified event values ​​associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The value is greater than or The value is greater than Then let or , and The calculation formulas are shown in equations (7) to (8) below: (7); (8); and Let represent the number of events associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The polarity factor of the index is represented by the following formula (9): (9); Finally, the probability P3 of the hotspot event warning is calculated using a linear weighting method, as shown in the following formula (10): (10); The probability of early warning for hot events. , It is a scaling factor, and .

[0029] Set early warning thresholds as needed. Generally, the value is taken as 0.75. In such cases, early warnings are issued for trending events.

[0030] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0031] Example 1 A method for early warning of hot events includes the following steps: Based on domain knowledge, a network of related indicators is constructed. Combined with background modeling information, the indicators are hierarchically and weighted according to their importance to form a network indicator system with coupled influence. Starting from the need for early warning of hot events, we construct basic intelligent algorithm mini-models, solidify the network indicator system into business data mini-models, and form a set of mini-models. Based on the large language model and the set of small models, the large language model is used as the scheduling model to schedule the basic intelligent algorithm small models and business data small models to discover hot event symptoms and form a hot event symptom set. Based on the aforementioned hotspot event symptom set, a weighted calculation of the symptom is performed using an analytical formula to obtain the early warning probability of the hotspot event. If the probability exceeds the early warning threshold, an early warning message is given and compared with the actual situation. Based on the differences in the comparison, the indicator network is optimized through adaptive feedback of a large model.

[0032] Example 2 Based on Example 1, the construction of the associated indicator network, combined with background modeling information, involves hierarchically and hierarchically assigning weights to the indicators according to their importance, forming a network indicator system with coupled influences. This specifically includes the following sub-steps: A dynamic indicator network is constructed, which includes multiple dynamic indicator subnetworks. Each subnetwork contains multiple types of indicators, which are divided into two layers according to their importance. Some types of indicators are located at the bottom layer, and other types of indicators are located at the top layer. The bottom layer is more important than the top layer. The indicators in each layer influence each other, i.e., they are mutually coupled. The indicators in the same subnetwork influence each other, i.e., they are self-coupled. The relationships within and between sub-networks are reflected through the superposition of indicators. At the same time, the superposition is constrained by background information. The same background information has a positive correlation, while different background information has a negative correlation. The superposition effect is divided into two categories: strengthening and weakening.

[0033] Example 3 Based on Example 1, the basic intelligent algorithm mini-model specifically includes a spatiotemporal analysis model, an event detection model, and an event analysis model.

[0034] Example 4 Based on Example 1, the business data mini-model specifically includes hierarchical indicators, correlation indicators, and group weight mini-models.

[0035] Example 5 Based on Example 1, the large language model is used as the scheduling model to schedule the basic intelligent algorithm small model and the business data small model to discover hot event symptoms. Based on the hot event symptom set, a weighted calculation of the symptoms is performed based on the analytical formula to obtain the hot event warning probability. If the warning threshold is exceeded, a warning message is given and compared with the actual situation. Based on the differences in the comparison, the indicator network is optimized through adaptive feedback of the large model. The specific steps include: Input layer: Extracts the data needed by the model from open source sources; Small model understanding layer: The text-based small model is based on the Roberta pre-trained language model to perform event detection and event classification; the spatiotemporal small model is based on the ray method, which determines the location of the target activity, i.e. whether it is active in the region of interest, by calculating the number of intersections between points and rays. Large Model Inference Layer: A general scheduling model is set up, which schedules the text-based small models fixed in the small model understanding layer. These models perform event detection and classification on all text data acquired from the input layer to obtain event sentences. The spatiotemporal small models fixed in the small model understanding layer determine the current target activity location and transcribe it into text information. Then, the text is vectorized using a vector embedding model to obtain text embedding vectors. Finally, the scheduling model schedules the business data small models, performing vector calculations on the text embedding vectors and the indicator embedding vectors in the business data small models to analyze semantic similarity. If the semantic similarity exceeds a preset threshold, it is determined to be a hotspot event symptom. Hotspot Event Early Warning Layer: This layer performs weighted calculations on the symptoms to obtain the probability of a hotspot event warning and feeds back the warning results. Specifically, it uses a large-scale model prompt word correction method to transform the warning results into a large-scale model prompt word optimization strategy. This includes optimization strategies for indicators and indicator weights in the indicator network. For example, if a hotspot event is warned of but does not actually occur, the large-scale model prompt adds descriptions to reduce indicator weights and the number of indicators involved in the calculation. If a hotspot event is not warned of but actually occurs, the large-scale model prompt adds descriptions to increase indicator weights, generate similar new indicators based on the indicator network and the current event, and increase the number of indicators involved in the calculation. Through the automatic prompt word optimization strategy, the hotspot event early warning evaluation model is dynamically constructed, continuously iteratively optimizing the hotspot event early warning effect. The generated new indicators are solidified into the underlying indicator network, transforming the static indicator network manually constructed based on domain expert experience into a continuously optimized dynamic indicator network.

[0036] Example 6 Based on Example 1, the data obtained at the input layer specifically includes two categories: text data and spatiotemporal data.

[0037] Example 7 Based on Example 1, the event detection results are divided into two categories: events and non-events. The event classification results are divided into six categories: security, politics, economy, diplomacy, military, and technology.

[0038] Example 8 Based on Example 1, the overall scheduling model specifically includes the large language model Qwen3-30B.

[0039] Example 9 Based on Example 5, the step of performing grouped weighted calculations on symptoms to obtain the early warning probability of hotspot events specifically includes the following sub-steps: First, calculate the independent early warning probability P1 of the dynamic index subnetwork: (1); In the formula, This represents the probability of issuing early warnings for hot events based on independent predictions of pre-selected indicators from a dynamic indicator subnetwork, i.e., the probability of independent early warnings from a dynamic indicator subnetwork. G represents the number of dynamic indicator subnetworks. This represents the predictive impact factor of hotspot events based on the s-th subnetwork. The specific calculation formula is as follows (2): (2); In the formula, This represents the weight of the m-th primary indicator, where M represents the total number of primary indicators. Primary indicators are removed as needed, and c is a minimum value representing a low-probability event. and , Denotes the influence factor of the m-th primary indicator and The specific calculation formula is as follows (3): (3); In the formula, This represents the weight of the e-th tertiary indicator for the m-th primary indicator. The positive or negative weight represents the positive or negative effect of the indicator on the conclusion, n represents the total number of tertiary indicators of the m-th primary indicator, and z represents the total number of tertiary indicators with positive weights of the m-th primary indicator. It is a constant that adjusts on demand, representing the maximum number of events related to the third-level indicators of the s-th dynamic indicator subnetwork. This represents the number of events related to the e-th tertiary indicator of the m-th primary indicator. The value is greater than Then let ; Then, calculate the joint early warning probability P2 of the dynamic index subnetwork: (4); In the formula, This represents the probability of a hotspot event warning based on correlation indicators within and between dynamic indicator subnetworks, i.e., the joint warning probability of dynamic indicator subnetworks. The specific calculation formula is as follows (5): (5); In the formula, This represents the predictive impact factor of hotspot events based on the correlation index between the k-th subnetwork and the h-th subnetwork. Time indicates the dynamic indicator within the subnetwork. Time indicates dynamic indicators between subnetworks; Let represent the association weight between the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, and satisfy ... F represents the total number of third-level correlation indicators in the k-th sub-network, D represents the total number of third-level correlation indicators in the h-th sub-network, and N represents the maximum number of events among the related indicators. The specific calculation formula is as follows (6): (6); In the formula, and These represent the maximum number of events associated with the third-level indicators of the k-th and h-th sub-networks, respectively, and can be adjusted as needed; in equation (5) and These represent the quantified event values ​​associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The value is greater than or The value is greater than Then let or , and The calculation formulas are as follows: (7) - (8) (7); (8); In the formula, and Let represent the number of events associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The polarity factor of the index is represented by the following formula (9): (9); Finally, the probability of a hotspot event warning, P3, is calculated using a linear weighting method, as shown in equation (10): (10); In the formula, P3 represents the probability of a hotspot event warning. , It is a scaling factor, and ; Set early warning thresholds as needed. ,when In such cases, early warnings are issued for trending events.

[0040] Example 10 A hotspot event early warning device includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded by the processor, it executes the method described in any one of Embodiments 1 to 9.

[0041] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0042] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0043] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A method for early warning of hot events, characterized in that, Includes the following steps: Based on domain knowledge, a network of related indicators is constructed. Combined with background modeling information, the indicators are hierarchically and weighted according to their importance to form a network indicator system with coupled influence. Starting from the need for early warning of hot events, we construct basic intelligent algorithm mini-models, solidify the network indicator system into business data mini-models, and form a set of mini-models. Based on the large language model and the set of small models, the large language model is used as the scheduling model to schedule the basic intelligent algorithm small models and business data small models to discover hot event symptoms and form a hot event symptom set. Based on the aforementioned hotspot event symptom set, a weighted calculation of the symptom is performed using an analytical formula to obtain the early warning probability of the hotspot event. If the probability exceeds the early warning threshold, an early warning message is given and compared with the actual situation. Based on the differences in the comparison, the indicator network is optimized through adaptive feedback of a large model.

2. The hotspot event early warning method according to claim 1, characterized in that, The construction of the associated indicator network, combined with background modeling information, involves hierarchically and hierarchically assigning weights to the indicators according to their importance, forming a network indicator system with coupled influences. This process specifically includes the following sub-steps: A dynamic indicator network is constructed, which includes multiple dynamic indicator subnetworks. Each subnetwork contains multiple types of indicators, which are divided into two layers according to their importance. Some types of indicators are located at the bottom layer, and other types of indicators are located at the top layer. The bottom layer is more important than the top layer. The indicators in each layer influence each other, i.e., they are mutually coupled. The indicators in the same subnetwork influence each other, i.e., they are self-coupled. The relationships within and between sub-networks are reflected through the superposition of indicators. At the same time, the superposition is constrained by background information. The same background information has a positive correlation, while different background information has a negative correlation. The superposition effect is divided into two categories: strengthening and weakening.

3. The hotspot event early warning method according to claim 1, characterized in that, The basic intelligent algorithm mini-models specifically include a spatiotemporal analysis model, an event detection model, and an event analysis model.

4. The hotspot event early warning method according to claim 1, characterized in that, The business data mini-model specifically includes hierarchical indicators, correlation indicators, and group weight mini-models.

5. The hotspot event early warning method according to claim 1, characterized in that, The process involves using a large language model as a scheduling model to schedule basic intelligent algorithm small models and business data small models to discover hotspot event symptoms. Based on the hotspot event symptom set, a weighted calculation of the symptoms is performed using analytical formulas to obtain the hotspot event warning probability. If the probability exceeds the warning threshold, a warning message is issued and compared with the actual situation. Based on the differences in the comparison, the indicator network is optimized through adaptive feedback of the large model. This process specifically includes the following sub-steps: Input layer: Extracts the data needed by the model from open source sources; Small model understanding layer: The text-based small model is based on the Roberta pre-trained language model to perform event detection and event classification; the spatiotemporal small model is based on the ray method, which determines the location of the target activity, i.e. whether it is active in the region of interest, by calculating the number of intersections between points and rays. Large Model Inference Layer: A general scheduling model is set up, which schedules the text-based small models fixed in the small model understanding layer. These models perform event detection and classification on all text data acquired from the input layer to obtain event sentences. The spatiotemporal small models fixed in the small model understanding layer determine the current target activity location and transcribe it into text information. Then, the text is vectorized using a vector embedding model to obtain text embedding vectors. Finally, the scheduling model schedules the business data small models, performing vector calculations on the text embedding vectors and the indicator embedding vectors in the business data small models to analyze semantic similarity. If the semantic similarity exceeds a preset threshold, it is determined to be a hotspot event symptom. Hotspot Event Early Warning Layer: This layer performs weighted calculations on the symptoms to obtain the probability of a hotspot event warning and feeds back the warning results. Specifically, it uses a large-scale model prompt word correction method to transform the warning results into a large-scale model prompt word optimization strategy. This includes optimization strategies for indicators and indicator weights in the indicator network. For example, if a hotspot event is warned of but does not actually occur, the large-scale model prompt adds descriptions to reduce indicator weights and the number of indicators involved in the calculation. If a hotspot event is not warned of but actually occurs, the large-scale model prompt adds descriptions to increase indicator weights, generate similar new indicators based on the indicator network and the current event, and increase the number of indicators involved in the calculation. Through the automatic prompt word optimization strategy, the hotspot event early warning evaluation model is dynamically constructed, continuously iteratively optimizing the hotspot event early warning effect. The generated new indicators are solidified into the underlying indicator network, transforming the static indicator network manually constructed based on domain expert experience into a continuously optimized dynamic indicator network.

6. The hotspot event early warning method according to claim 1, characterized in that, The data obtained at the input layer specifically includes two categories: text data and spatiotemporal data.

7. The hotspot event early warning method according to claim 1, characterized in that, Event detection results are categorized into two types: events and non-events. Event classification results are further divided into six categories: security, politics, economy, diplomacy, military, and technology.

8. The hotspot event early warning method according to claim 1, characterized in that, The overall scheduling model specifically includes the large language model Qwen3-30B.

9. The hotspot event early warning method according to claim 5, characterized in that, The step of performing grouped and weighted calculations on the symptoms to obtain the early warning probability of hotspot events specifically includes the following sub-steps: First, calculate the independent early warning probability P1 of the dynamic index subnetwork: （1）； In the formula, This represents the probability of issuing early warnings for hot events based on independent predictions of pre-selected indicators from a dynamic indicator subnetwork, i.e., the probability of independent early warnings from a dynamic indicator subnetwork. G represents the number of dynamic indicator subnetworks. This represents the predictive impact factor of hotspot events based on the s-th subnetwork. The specific calculation formula is as follows (2): （2）； In the formula, This represents the weight of the m-th primary indicator, where M represents the total number of primary indicators. Primary indicators are removed as needed, and c is a minimum value representing a low-probability event. and , Denotes the influence factor of the m-th primary indicator and The specific calculation formula is as follows (3): （3）； In the formula, This represents the weight of the e-th tertiary indicator for the m-th primary indicator. The positive or negative weight represents the positive or negative effect of the indicator on the conclusion, n represents the total number of tertiary indicators of the m-th primary indicator, and z represents the total number of tertiary indicators with positive weights of the m-th primary indicator. It is a constant that adjusts on demand, representing the maximum number of events related to the third-level indicators of the s-th dynamic indicator subnetwork. This represents the number of events related to the e-th tertiary indicator of the m-th primary indicator. The value is greater than Then let ; Then, calculate the joint early warning probability P2 of the dynamic index subnetwork: （4）； In the formula, This represents the probability of a hotspot event warning based on correlation indicators within and between dynamic indicator subnetworks, i.e., the joint warning probability of dynamic indicator subnetworks. The specific calculation formula is as follows (5): （5）； In the formula, This represents the predictive impact factor of hotspot events based on the correlation index between the k-th subnetwork and the h-th subnetwork. Time indicates the dynamic indicator within the subnetwork. Time indicates dynamic indicators between subnetworks; Let represent the association weight between the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, and satisfy ... F represents the total number of third-level correlation indicators in the k-th sub-network, D represents the total number of third-level correlation indicators in the h-th sub-network, and N represents the maximum number of events among the related indicators. The specific calculation formula is as follows (6): （6）； In the formula, and These represent the maximum number of events associated with the third-level indicators of the k-th and h-th sub-networks, respectively, and can be adjusted as needed; in equation (5) and These represent the quantified event values ​​associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The value is greater than or The value is greater than Then let or , and The calculation formulas are as follows: (7) - (8) （7）； （8）； In the formula, and Let represent the number of events associated with the i-th tertiary indicator of the k-th subnetwork and the j-th tertiary indicator of the h-th subnetwork, respectively. The polarity factor of the index is represented by the following formula (9): （9）； Finally, the probability of a hotspot event warning, P3, is calculated using a linear weighting method, as shown in equation (10): （10）； In the formula, P3 represents the probability of a hotspot event warning. , It is a scaling factor, and ; Set early warning thresholds as needed. ,when In such cases, early warnings are issued for trending events.

10. A hotspot event early warning device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method as described in any one of claims 1 to 9.

Images