Data dynamic monitoring methods, devices, and equipment based on intelligent agents

Abstract

This application discloses a data dynamic supervision method, apparatus, and equipment based on intelligent agents, relating to the field of data security governance. The method includes: extracting keywords from public health data and multi-source online public opinion data using a pre-trained language model, and calculating the similarity between the extracted keywords to be classified and public opinion keywords; conducting a risk assessment of public health data to obtain an inherent risk value; calculating the product of the source weight of the public opinion data and the keyword similarity score to obtain the source credibility; generating an attenuation factor based on the time interval between the release time and the current time of the public opinion data and an exponential decay function, and calculating the product of the attenuation factor and the source credibility to obtain a public opinion similarity score; fusing the public opinion similarity score with the inherent risk value, and determining the target data level based on the target risk interval where the fused comprehensive risk value falls. This application can improve the timeliness and accuracy of public health data classification.

Description

Technical Field

[0001] This application relates to the field of data security governance, and in particular to methods, devices and equipment for dynamic data supervision based on intelligent agents. Background Technology

[0002] With the continuous development of intelligent management of public health data and information security technology, the classification, grading and security management of public health data, especially infectious disease surveillance data, has become a core component of the public health emergency response system.

[0003] Currently, the classification and grading of infectious disease surveillance data is usually based on pre-set static rules and fixed thresholds. For example, methods such as static classification and grading based on fixed rules, classification models based on traditional machine learning, and sensitive information identification methods combined with NLP (Natural Language Processing) technology, while capable of data classification and grading, have the following limitations: they cannot adapt to the dynamic impact of epidemic development and public opinion fluctuations on data sensitivity; they lack the ability to integrate and analyze multi-source information (such as public opinion trends and public concerns); the grading results are lagging, making it difficult to support real-time or near-real-time adjustments to security strategies; and they are highly dependent on manual intervention and inefficient.

[0004] Furthermore, the development of the epidemic and public concerns are constantly changing, and static systems cannot perceive these changes in the external environment, leading to a disconnect between data classification results and the actual situation. For example, during a stable period of the epidemic, some data may not require a high level of protection; however, once related public opinion intensifies, the sensitivity of the same data increases sharply. At this time, because static systems cannot adjust in real time, they are prone to data leakage due to insufficient protection, or data sharing may be hindered due to excessive protection, thus affecting the balance between data flow and security. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a data dynamic monitoring method, device, and equipment based on intelligent agents, which can dynamically adjust the data classification level according to the evolution of public opinion, and can mine the semantic relationship between public health data and public opinion data, thereby significantly improving the accuracy of risk perception. At the same time, it overcomes the one-sidedness of data classification from a single risk perspective; furthermore, it improves the timeliness of data classification and overcomes the problems of strong reliance on manual methods and low efficiency, thus enabling accurate and timely public health early warning processing. The specific solution is as follows:

[0006] Firstly, this application discloses a data dynamic supervision method based on intelligent agents, including:

[0007] Acquire public health data and multi-source online public opinion data related to the public health data;

[0008] Using a pre-trained language model, keywords are extracted from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords. The similarity between the keywords to be classified and the public opinion keywords is calculated to obtain keyword similarity scores.

[0009] The public health data is risk-assessed based on multiple data grading and evaluation dimensions to obtain an inherent risk value.

[0010] The data sources of the multi-source online public opinion data are determined, and the product of the source weight corresponding to the data source and the keyword similarity score is calculated to obtain the source credibility of the multi-source online public opinion data.

[0011] The time interval between the release time of the multi-source network public opinion data and the current time is determined, and the time interval is input into the exponential decay function to generate the failure decay factor. Then, the product of the failure decay factor and the source credibility is calculated to obtain the public opinion similarity score.

[0012] The public opinion similarity score is fused with the inherent risk value to obtain a comprehensive risk value. Based on the target risk range in which the comprehensive risk value is located, the target data level corresponding to the public health data is determined, and the public health early warning processing is determined based on the target data level.

[0013] Optionally, the step of extracting keywords from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords includes:

[0014] The public health data and the multi-source online public opinion data are preprocessed to obtain multiple words to be classified and segmented and multiple public opinion words; the preprocessing operation includes unified encoding and word segmentation processing.

[0015] Multiple words to be classified and multiple words to be public opinion are input into a preset multi-channel for scoring processing to obtain multiple scores for a single word to be classified and multiple public opinion scores for a single word to be public opinion; the preset multi-channel includes a statistical channel for the base text keyword extraction algorithm, an entity importance channel based on a preset named entity recognition model, and a dictionary matching channel based on a preset dictionary;

[0016] Based on the multiple scores corresponding to each of the words to be classified and the channel weights of each channel in the preset multi-channel, the comprehensive weight of each of the words to be classified is calculated by weighted fusion to obtain the comprehensive weight of the words to be classified.

[0017] Based on the multiple public opinion scores corresponding to each of the aforementioned public opinion segments and the multiple channel weights, the comprehensive weight of each of the aforementioned public opinion segments is calculated by weighted fusion to obtain the comprehensive public opinion weight.

[0018] Based on the comprehensive weight to be classified, keywords from multiple words to be classified are filtered to obtain a set of keywords to be classified that include the comprehensive weight to be classified. Based on the comprehensive weight to be classified, keywords from multiple words to be classified are filtered to obtain a set of keywords to be classified that include the comprehensive weight to be classified.

[0019] Optionally, the step of filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of keywords to be classified containing the comprehensive weight, and filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of public opinion keywords containing the comprehensive weight, includes:

[0020] The multiple weights to be classified and the multiple public opinion weights are sorted in descending order to obtain the sorted weights to be classified and the sorted public opinion weights.

[0021] From the sorted weights to be graded, select the first preset number of comprehensive weights to be graded, and use the words to be graded corresponding to the first preset number of comprehensive weights to be graded as keywords to obtain a set of keywords to be graded that includes the comprehensive weights to be graded.

[0022] From the sorted public opinion weights, a preset number of comprehensive public opinion weights are selected, and the public opinion words corresponding to the preset number of comprehensive public opinion weights are used as keywords to obtain a set of public opinion keywords containing the comprehensive public opinion weights.

[0023] Optionally, the step of filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of keywords to be classified containing the comprehensive weight, and filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of public opinion keywords containing the comprehensive weight, includes:

[0024] The multiple weights to be classified and the multiple public opinion weights are sorted in descending order to obtain the sorted weights to be classified and the sorted public opinion weights.

[0025] The maximum values ​​of the sorted weights to be classified and the sorted public opinion weights are determined respectively, thus obtaining the maximum value of the weights to be classified and the maximum value of the public opinion weights.

[0026] Calculate the product of the maximum value of the weight to be classified and the preset cutoff factor, and the product of the maximum value of the public opinion weight and the preset cutoff factor to obtain the first product and the second product;

[0027] Determine the comprehensive weight to be classified that is greater than or equal to the first product among the weights to be classified after sorting, and take the words to be classified corresponding to the comprehensive weights to be classified that are greater than or equal to the first product as keywords, to obtain a set of keywords to be classified that contain the comprehensive weights to be classified.

[0028] Determine the comprehensive public opinion weights that are greater than or equal to the second product among the sorted public opinion weights, and use the public opinion word segments corresponding to the comprehensive public opinion weights that are greater than or equal to the second product as keywords to obtain a set of public opinion keywords containing the comprehensive public opinion weights.

[0029] Optionally, calculating the similarity between the keyword to be classified and the public opinion keyword to obtain a keyword similarity score includes:

[0030] The keywords in the set of keywords to be classified and the set of public opinion keywords are encoded using an encoding model that includes a word segmenter, resulting in multiple word vectors to be classified and multiple public opinion word vectors.

[0031] Each of the word vectors to be classified and each of the public opinion word vectors are subjected to mean pooling to obtain the pooled word vectors to be classified and the pooled public opinion vectors.

[0032] The cosine similarity algorithm is used to calculate the similarity between the pooled vector to be classified and the pooled sentiment vector to obtain the cosine similarity score.

[0033] The similarity between the set of keywords to be classified and the set of public opinion keywords is calculated using a weighted Jaccard similarity algorithm based on the comprehensive weight of the keywords to be classified and the comprehensive weight of public opinion, and a set similarity score is obtained.

[0034] The topic probability distributions of the keyword set to be classified and the public opinion keyword set are constructed using the BERTopic model, respectively, to obtain the topic probability distribution to be classified and the public opinion probability distribution. The similarity between the topic probability distribution to be classified and the public opinion probability distribution is calculated using JS divergence to obtain the topic distribution similarity score.

[0035] Optionally, the formula for calculating the public opinion similarity score is as follows:

[0036] ;

[0037] in, ;

[0038] In the formula, The public opinion similarity score is given. The time decay coefficient is denoted by t, which represents the time interval. For attenuation adjustment parameters, This indicates the source weight corresponding to data source c. The cosine similarity score is... The similarity score of the set, The similarity score for the topic distribution is... The cosine similarity score corresponds to the similarity coefficient. The similarity coefficient corresponds to the set similarity score. The similarity coefficient corresponds to the similarity score of the topic distribution.

[0039] Optionally, the step of fusing the public opinion similarity score with the inherent risk value to obtain a comprehensive risk value includes:

[0040] The public opinion similarity score and the inherent risk value are normalized respectively to obtain the normalized public opinion similarity value and the normalized risk value;

[0041] The normalized public opinion similarity value and the normalized risk value are weighted and fused to obtain a comprehensive risk value;

[0042] The formula for calculating the comprehensive risk value is as follows:

[0043] ;

[0044] In the formula, R is the comprehensive risk value. Indicates the fusion weight coefficient. This represents the normalized risk value. This represents the normalized public opinion similarity value.

[0045] Secondly, this application discloses a data dynamic monitoring device based on an intelligent agent, comprising:

[0046] The data acquisition module is used to acquire public health data and multi-source online public opinion data related to the public health data;

[0047] The keyword extraction module is used to extract keywords from the public health data and the multi-source online public opinion data using a pre-trained language model, respectively, to obtain keywords to be classified and public opinion keywords.

[0048] The first calculation module is used to calculate the similarity between the keyword to be classified and the public opinion keyword, and obtain the keyword similarity score;

[0049] The risk assessment module is used to assess the risk of the public health data based on multiple data classification and assessment dimensions to obtain the inherent risk value.

[0050] The second calculation module is used to determine the data source of the multi-source online public opinion data, and calculate the product of the source weight corresponding to the data source and the keyword similarity score to obtain the source credibility of the multi-source online public opinion data.

[0051] The time interval determination module is used to determine the time interval between the release time of the multi-source online public opinion data and the current time;

[0052] The third calculation module is used to input the time interval into the exponential decay function to generate a failure decay factor, and then calculate the product of the failure decay factor and the source credibility to obtain the public opinion similarity score.

[0053] The fusion module is used to fuse the public opinion similarity score with the inherent risk value to obtain a comprehensive risk value;

[0054] The data classification module is used to determine the target data level corresponding to the public health data based on the target risk range in which the comprehensive risk value is located, so as to determine whether to carry out public health early warning processing based on the target data level.

[0055] Thirdly, this application discloses an electronic device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the aforementioned agent-based dynamic data monitoring method.

[0056] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned agent-based dynamic data monitoring method.

[0057] As can be seen, this application is applied to an artificial intelligence entity. First, it acquires public health data and related multi-source online public opinion data. Then, it uses a pre-trained language model to extract keywords from both the public health data and the multi-source online public opinion data, obtaining keywords to be classified and public opinion keywords. It then calculates the similarity between the keywords to be classified and the public opinion keywords, obtaining a keyword similarity score. Next, it performs a risk assessment on the public health data based on multiple data classification and evaluation dimensions to obtain an inherent risk value. Finally, it determines the data source of the multi-source online public opinion data and calculates the source weight corresponding to the data source and the keywords. The source credibility of the multi-source online public opinion data is obtained by multiplying the similarity scores. Next, the time interval between the release time of the multi-source online public opinion data and the current time is determined, and the time interval is input into the exponential decay function to generate the failure decay factor. Then, the product of the failure decay factor and the source credibility is calculated to obtain the public opinion similarity score. Finally, the public opinion similarity score is fused with the inherent risk value to obtain the comprehensive risk value. Based on the target risk interval in which the comprehensive risk value is located, the target data level corresponding to the public health data is determined, and based on the target data level, it is determined whether to carry out public health early warning processing.

[0058] This application first uses a pre-trained language model to calculate the similarity between different keywords in public health data and multi-source online public opinion data. Then, it calculates the product of the source weight corresponding to the source of the public opinion data and the keyword similarity to obtain the source credibility of the public opinion data. Next, it generates an invalidation decay factor based on the lag time of the public opinion data (i.e., the time interval between the release time and the current time) and the exponential decay function. It then obtains the public opinion similarity score by calculating the product of the invalidation decay factor and the source credibility. Finally, it merges the public opinion similarity score with the inherent risk value (the value obtained after risk assessment of public health data based on multiple data classification and evaluation dimensions). Based on the risk range of the comprehensive risk value obtained after fusion, it determines the data level of the public health data so as to determine whether to carry out public health early warning processing based on the data level. As can be seen, this application integrates dynamic, multi-source online public opinion data into the foundation of public health data, enabling the data classification results to be combined with the actual situation. It allows for dynamic adjustment of data classification levels based on the evolution of public opinion. Furthermore, by calculating the similarity between the keywords to be classified and the keywords in public opinion, the semantic relationship between public health data and public opinion data can be mined, significantly improving the accuracy of risk perception. Moreover, this application not only quantifies the inherent risks of public health data through multiple data classification assessment dimensions but also quantifies its external risks through public opinion analysis. It also introduces an attenuation factor based on the timeliness and source credibility of public opinion, thus overcoming the one-sidedness of classification from a single risk perspective. In addition, automated data classification through artificial intelligence not only improves the timeliness of data classification but also overcomes the problems of strong reliance on manual methods and low efficiency, thereby enabling accurate and timely public health early warning processing. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application 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 embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0060] Figure 1 This is a flowchart of a data dynamic monitoring method based on intelligent agents disclosed in this application;

[0061] Figure 2 This application discloses a specific data dynamic supervision method based on intelligent agents.

[0062] Figure 3 This is a schematic diagram of a specific agent-based dynamic classification process for public health data disclosed in this application;

[0063] Figure 4This is a schematic diagram of the structure of a data dynamic monitoring device based on an intelligent agent disclosed in this application;

[0064] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0065] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] This application discloses a data dynamic monitoring method based on intelligent agents. See also Figure 1 As shown, the method includes:

[0067] Step S11: Obtain public health data and multi-source online public opinion data related to the public health data.

[0068] It should be noted that the dynamic monitoring scheme for public health data proposed in this application is specifically applied to an artificial intelligence agent (i.e., an AI Agent). When dynamically monitoring public health data, the AI ​​Agent can acquire relevant public health data (such as infectious disease monitoring and early warning data), as well as multi-source online public opinion data related to this public health data (such as daily public opinion data covering news reports, social media trending topics, and public concerns related to the epidemic). The data sources / channels of the multi-source online public opinion data include, but are not limited to, news reports, social media, forum discussions, and official announcements.

[0069] Step S12: Use a pre-trained language model to extract keywords from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords, and calculate the similarity between the keywords to be classified and the public opinion keywords to obtain keyword similarity scores.

[0070] In this embodiment, after obtaining public health data and related multi-source online public opinion data, a pre-trained language model (i.e., LLM, Large Language Model), such as the Qwen3-32B model, is further used to perform semantic parsing and keyword extraction on the public health data and multi-source online public opinion data, respectively, to extract key terms that can reflect the core content and sensitive characteristics of the data, thereby obtaining corresponding keywords to be classified and public opinion keywords. The semantic similarity between the keywords to be classified and the public opinion keywords is calculated to generate corresponding keyword similarity scores, which are used to measure the degree of correlation between the public health data to be classified and the public opinion hotspots.

[0071] Specifically, the step of extracting keywords from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords may include: performing preprocessing operations on the public health data and the multi-source online public opinion data to obtain multiple words to be classified and multiple public opinion words; the preprocessing operations include unified encoding and word segmentation processing; inputting the multiple words to be classified and the multiple public opinion words into a preset multi-channel for scoring processing to obtain multiple scores corresponding to a single word to be classified and multiple public opinion scores corresponding to a single public opinion word; the preset multi-channel includes a statistical channel of the base text keyword extraction algorithm, an entity importance channel based on a preset named entity recognition model, and a dictionary matching channel based on a preset dictionary. Channel allocation; based on the multiple scores corresponding to each of the words to be classified and the channel weights corresponding to each channel in the preset multi-channel, the comprehensive weight of each of the words to be classified is calculated by weighted fusion to obtain the comprehensive weight of the words to be classified; based on the multiple public opinion scores corresponding to each of the public opinion words and the multiple channel weights, the comprehensive weight of each of the public opinion words is calculated by weighted fusion to obtain the comprehensive weight of the public opinion; based on the comprehensive weight of the words to be classified, the keywords in the multiple words to be classified are filtered to obtain a set of keywords to be classified containing the comprehensive weight of the words to be classified, and based on the comprehensive weight of the public opinion, the keywords in the multiple words to be classified are filtered to obtain a set of public opinion keywords containing the comprehensive weight of the public opinion.In this embodiment, public health data and related multi-source online public opinion data can be preprocessed separately. For example, a medical-specific word segmentation dictionary can be used to uniformly encode and segment infectious disease monitoring and early warning data. Then, noise such as stop words, numbers, and emoticons in the segmented data can be removed, and time data and place name data can be regularized. At the same time, fields involving identity identifiers can be desensitized. After the above preprocessing operations, the corresponding words to be classified and multiple public opinion words can be obtained. Then, the words to be classified and multiple public opinion words are input into a preset multi-channel for scoring processing. The preset multi-channel includes three different channels: a statistical channel based on text keyword extraction algorithms (such as TF-IDF (term frequency–inverse document frequency, a commonly used weighting technique for information retrieval and data mining), TextRank (a graph-based ranking algorithm for text), etc.), and a channel based on a preset named entity recognition model (i.e., Named The system includes an entity importance channel for Entity Recognition (NER) and a dictionary matching channel based on a pre-defined dictionary. The statistical channel is used to generate high-weight candidate words or phrases, the entity importance channel is used to extract key entities based on NER in the medical field (such as "disease / symptom / department / place name / institution / population"), and the dictionary matching channel is used to match and label with "sensitive words / high-risk topic thesaurus" (such as case trajectory, cluster events, specific vulnerable groups, key places, etc.). Specifically, the score after the statistical channel can be represented as S_stat(t)∈[0,1] (after Min-Max normalization, where t represents the token), the score after the entity importance channel can be represented as I_ent(t)∈{0,1,2,3,4} (non-entity=0, ordinary entity=1, crowd / place entity=2, identity + location combination=3, high-sensitivity triplet (simultaneous appearance of the three elements of "identity identifier + time + precise location")=4), and the score after the dictionary matching channel can be represented as M_dict(t)∈{0,1,2} (miss=0, ordinary sensitivity=1, high risk=2).

[0072] Next, the scores obtained in each channel are normalized, that is, the scores of the three channels are compressed to the same dimension, such as [0,1]. The normalized scores of the three channels are expressed as: N_stat(t) = S_stat(t) (already located between 0 and 1), N_ent(t) = I_ent(t) / 4 (i.e., divided by the maximum entity importance level 4), N_dict(t) = M_dict(t) / 2 (i.e., divided by the maximum matching strength level 4).

[0073] Furthermore, based on the scores and channel weights output from the three channels, the comprehensive weights of each segmented word to be graded and the sentiment segmented word are calculated separately using a weighted fusion method. This yields the corresponding comprehensive weights for the segmented word and the sentiment segmentation. The formula for calculating the comprehensive weights is as follows:

[0074] ;

[0075] in, , and These are the channel weights for the three channels, and they satisfy... , The initial value can be set to w = [0.5, 0.3, 0.2]. It should be noted that the weights of each channel are learnable weights and can be optimized and adjusted according to actual application needs. For example, training can be performed using a small number of manually labeled samples (during training, loss functions such as cross-entropy, mean squared error, and gradient descent can be used for parameter optimization). After obtaining the above comprehensive weights, Wk(t) can be used to filter keywords from multiple word segments, thus obtaining a keyword set containing the corresponding comprehensive weights = {(t, Wk(t)) | t is the filtered word segment}.

[0076] In one specific implementation, the step of filtering keywords from multiple segmented words to be classified based on the comprehensive weight to be classified, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified, and filtering keywords from multiple segmented words based on the comprehensive weight to be classified, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified, specifically may include: sorting the multiple comprehensive weights to be classified and the multiple comprehensive weights to be classified in descending order, obtaining sorted weights to be classified and sorted public opinion weights; filtering out a first preset number of comprehensive weights to be classified from the sorted weights to be classified, and using the segmented words to be classified corresponding to the first preset number of comprehensive weights to be classified as keywords, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified; filtering out a first preset number of comprehensive weights to be classified from the sorted weights to be classified, and using the segmented words to be classified corresponding to the first preset number of comprehensive weights to be classified as keywords, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified; and filtering out a first preset number of comprehensive weights to be classified to be classified, and using the segmented words to be classified corresponding to the first preset number of comprehensive weights to be classified as keywords, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified. In this embodiment, the multiple comprehensive weights to be graded and the multiple comprehensive weights of public opinion are first sorted in descending order, that is, the comprehensive weights Wk(t) are sorted in descending order. Then, the words corresponding to the top preset number (e.g. Top-K, K=20~50) comprehensive weights in the sorted weights are used as keywords, thereby obtaining a set of keywords containing the corresponding comprehensive weights Wk(t).

[0077] For example, when public health data is medical report document data and multi-source online public opinion data is social media discussion data of the day, the medical report document data (such as "A hospital reported a cluster of influenza outbreaks involving 15 elderly patients, whose main symptoms were fever and cough") is first preprocessed, and candidate words are extracted from the preprocessed data through three channels. For example, the candidate words and corresponding scores obtained after extraction through the statistical channel are influenza (0.8), cluster (0.6), outbreak (0.7), elderly (0.5), and patient (0.4); the candidate words and corresponding levels obtained after extraction through the entity importance channel are influenza (disease, level 1), elderly patients (population, level 2), and hospital (location, level 2); the candidate words and corresponding levels obtained after extraction through the dictionary matching channel are cluster outbreak (high risk, level 2) and influenza (common sensitivity, level 1). Similarly, the social media discussion data of the day (such as "a case trajectory has appeared in a certain community, involving multiple close contacts, and a high-risk area has been identified") is preprocessed, and candidate words are extracted from the preprocessed data through three channels. For example, the candidate words and corresponding scores obtained after extraction through the statistical channel are: case trajectory (0.9), close contacts (0.8), high risk (0.7), and community (0.5); the candidate words and corresponding levels obtained after extraction through the entity importance channel are: community (location, level 2) and close contacts (population, level 2); the candidate words and corresponding levels obtained after extraction through the dictionary matching channel are: case trajectory (high risk, level 2), close contacts (high risk, level 2), and high-risk area (high risk, level 2). Next, normalization is performed within each channel. The normalization formulas are: N_stat(t) = S_stat(t) ∈ [0,1]; N_ent(t) = I_ent(t) / 4 ∈ [0,1]; N_dict(t) = M_dict(t) / 2 ∈ [0,1]. The normalization results for the medical report document data are shown in Table 1.

[0078] Table 1. Normalization results of medical report document data

[0079] Candidate words N_stat N_ent N_dict Case trajectory 0.9 0 1.0 (2 / 2) Close contacts 0.8 0.5 (2 / 4) 1.0 (2 / 2) High-risk areas 0.7 0 1.0 (2 / 2) residential area 0.5 0.5 (2 / 4) 0

[0080] The normalized results of the social media discussion data for that day are shown in Table 2:

[0081] Table 2. Normalization results of social media discussion data for the day.

[0082] Candidate words N_stat N_ent N_dict Case trajectory 0.9 0 1.0 (2 / 2) Close contacts 0.8 0.5 (2 / 4) 1.0 (2 / 2) High-risk areas 0.7 0 1.0 (2 / 2) residential area 0.5 0.5 (2 / 4) 0

[0083] Furthermore, a weight fusion calculation is performed based on the three channel weights w' = [0.4, 0.4, 0.2] corresponding to the medical report document data (corresponding to the statistical channel, dictionary matching channel, and entity importance channel, respectively), and the three channel weights w'' = [0.3, 0.3, 0.4] corresponding to the social media discussion data of the day. The calculation formula is as follows: The weighted fusion results of the medical report document data are as follows: Influenza: Wk = 0.4×0.8 + 0.4×0.25 + 0.2×0.5 = 0.32 + 0.10 + 0.10 = 0.52; Cluster outbreaks: Wk = 0.4×0.65 + 0.4×0 + 0.2×1.0 = 0.26 + 0 + 0.20 = 0.46; Elderly patients: Wk = 0.4×0.45 + 0.4×0.5 + 0.2×0 = 0.18 + 0.20 + 0 = 0.38; Hospitals: Wk = 0.4×0.3 + 0.4×0.5 + 0.2×0 = 0.12 + 0.20 + 0 =0.32; Next, all weight fusion results are sorted to obtain the sorted results: influenza (0.52), cluster outbreak (0.46), elderly patients (0.38), and hospitals (0.32). Similarly, the weighted fusion result of the social media discussion data for that day is calculated as follows: Case trajectory: Wk = 0.3×0.9 + 0.3×0 + 0.4×1.0 = 0.27 + 0 + 0.40 = 0.67; Close contacts: Wk = 0.3×0.8 + 0.3×0.5 + 0.4×1.0 = 0.24 + 0.15 + 0.40 = 0.79; High-risk areas: Wk = 0.3×0.7 + 0.3×0 + 0.4×1.0 = 0.21 + 0 + 0.40 = 0.61; Community: Wk = 0.3×0.5 + 0.3×0.5 + 0.4×0 = 0.15 + 0.15 + 0 =0.30; Next, all weight fusion results are sorted to obtain the sorted results: close contacts (0.79), case trajectory (0.67), high-risk area (0.61), community (0.30). Finally, the top-3 keywords are obtained to obtain the keyword set: {(close contacts, 0.79), (case trajectory, 0.67), (high-risk area, 0.61)}.

[0084] In another specific implementation, the step of filtering keywords from multiple segmented words to be classified based on the comprehensive weight to be classified, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified, and filtering keywords from multiple segmented words based on the comprehensive weight to be classified, obtaining a set of keywords to be classified that includes the comprehensive weight to be classified, specifically may include: sorting the multiple comprehensive weights to be classified and the multiple comprehensive weights to be classified in descending order, obtaining sorted weights to be classified and sorted weights to be classified; determining the maximum value among the sorted weights to be classified and the sorted weights to be classified, obtaining the maximum value of the weights to be classified and the maximum value of the weights to be classified; calculating the... The first product and the second product are obtained by multiplying the maximum value of the weight to be classified by the preset cutoff factor and the maximum value of the public opinion weight by the preset cutoff factor. The comprehensive weight to be classified that is greater than or equal to the first product among the sorted weights to be classified are determined, and the words to be classified corresponding to the comprehensive weights to be classified that are greater than or equal to the first product are used as keywords, thus obtaining a set of keywords to be classified that include the comprehensive weights to be classified. The comprehensive public opinion weight that is greater than or equal to the second product among the sorted public opinion weights is determined, and the public opinion words corresponding to the comprehensive public opinion weights that are greater than or equal to the second product are used as keywords, thus obtaining a set of public opinion keywords that include the comprehensive public opinion weights. In this embodiment, the comprehensive weights Wk(t) (such as the comprehensive weights to be graded or the comprehensive weights of public opinion) can be sorted in descending order first. Then, the maximum value among the sorted weights, i.e., max(Wk), is determined, and the product of the maximum value and the preset truncation factor (θ) is calculated, i.e., θ×max(Wk), such as 0.3×max(Wk). Next, the sorted weights are dynamically truncated, retaining the word segments with Wk(t)≥θ, and the word segments with Wk(t)≥θ are used as keywords, thereby obtaining the filtered keyword set.

[0085] In this embodiment, calculating the similarity between the keywords to be classified and the public opinion keywords to obtain a keyword similarity score may specifically include: encoding each keyword in the set of keywords to be classified and the set of public opinion keywords using an encoding model containing a word segmenter, to obtain multiple word vectors to be classified and multiple public opinion word vectors; performing mean pooling on each of the word vectors to be classified and each of the public opinion word vectors, to obtain pooled word vectors to be classified and pooled public opinion vectors; calculating the similarity between the pooled word vectors to be classified and the pooled public opinion vectors using a cosine similarity algorithm, to obtain a cosine similarity score; calculating the similarity between the set of keywords to be classified and the set of public opinion keywords using a weighted Jaccard similarity algorithm based on the comprehensive weights of the keywords to be classified and the comprehensive weights of the public opinion keywords, to obtain a set similarity score; constructing the topic probability distributions of the set of keywords to be classified and the set of public opinion keywords using a BERTopic model, to obtain the topic probability distributions to be classified and the public opinion probability distributions, and calculating the similarity between the topic probability distributions to be classified and the public opinion probability distributions using JS divergence, to obtain a topic distribution similarity score. In this embodiment, an encoding model containing a tokenizer, such as the RoBERTa model (Robustly optimized BERT approach, a topic model based on Transformer architecture and BERT pre-trained model), is first used to classify the keyword sets to be graded. }) and public opinion keyword set ( The keywords in the text are encoded to obtain multiple word vectors to be classified and multiple public opinion word vectors. , This represents the (i)th keyword. Keywords The weight of keywords in the set to be graded. Keywords The weights in the set of public opinion keywords are determined; then, mean pooling is performed on the word vectors to be graded and the public opinion word vectors to represent them semantically as a whole, which is used for subsequent similarity calculation and risk assessment, resulting in the pooled word vectors to be graded and the pooled public opinion vectors. Then, the similarity between the two vectors (i.e., the pooled word vectors to be graded and the pooled public opinion vectors) is calculated using three different similarity calculation methods.

[0086] In one specific implementation, the cosine similarity algorithm can be used to quantify the angle between the semantic vectors (such as 768-dimensional semantic vectors) of two texts in a high-dimensional space, thereby measuring the "directional consistency" of their overall semantics. The smaller the angle, the more similar their semantics. The specific calculation formula is as follows:

[0087] ;

[0088] In the formula, A and B are two n-dimensional vectors (corresponding to the pooled vector to be graded and the pooled public opinion vector, respectively), the numerator is the dot product of the two vectors, and the denominator is the product of the magnitudes of the two vectors. The range is [-1, 1]. The closer to 1, the more similar the semantics; the closer to -1, the more opposite the semantics; and 0, the irrelevant.

[0089] It should be noted that, compared to Euclidean distance, cosine similarity focuses more on the "direction" of vectors rather than their "length," making it suitable for semantic matching of text (for example, "environmental pollution" and "environmental pollution" have different lengths but the same direction, resulting in higher cosine similarity). The cosine similarity algorithm can effectively capture deep semantic relationships in text, providing important quantitative evidence.

[0090] For example, when the vector to be graded after pooling is The pooled sentiment vector is First, calculate the dot product of the two vectors. = (0.8×0.6) + (0.2×0.3) + (0.5×0.7) + (0.3×0.4) = 0.48 + 0.06 + 0.35 + 0.12 = 1.01; Next, calculate the vector magnitude:

[0091] ;

[0092] .

[0093] Furthermore, the cosine similarity is calculated: cosine_sim = 1.01 / (0.959 × 1.049) = 1.01 / 1.006 ≈ 1.004. It should be noted that due to the precision limitations of floating-point calculations, the result may slightly exceed 1, and truncation can be performed in practical applications. The values ​​and meanings of different cosine similarity are as follows: 1.0: Completely identical direction (semantically identical); 0.8-0.9: Highly similar (strong semantic association); 0.5-0.7: Moderately similar (some semantic association); 0.0-0.3: Basically dissimilar (large semantic differences). It is evident that high cosine similarity indicates a high degree of deep semantic correlation between the public health data to be classified and public opinion hotspots, posing a risk of being amplified by public opinion; low cosine similarity indicates a weaker semantic association, with limited amplification effect on public opinion.

[0094] In another specific implementation, the weight mapping function is first defined:

[0095] ;

[0096] ;

[0097] Next, the similarity between the set of keywords to be classified and the set of public opinion keywords is calculated using the weighted Jaccard similarity algorithm (a text similarity measurement method), and the set similarity score is obtained. The specific calculation formula is as follows:

[0098] ;

[0099] In the formula, the numerator (intersection part) is: Where (C) is the union of all keywords: (C = t | t ∈ A or t ∈ B); the denominator (the union part) is: .

[0100] For example, when the set of keywords to be classified is A={(epidemic, 0.8),(prevention and control, 0.6)}, and the set of public opinion keywords is B={(epidemic, 0.7),(cases, 0.5)}, first construct the complete set of terms C={epidemic,prevention and control,cases}, and then calculate the numerator (intersection part): And the denominator (union part): Thus, the final weighted Jaccard similarity is obtained: .

[0101] Compared to the traditional Jaccard similarity algorithm (which ignores weights): This solution takes into account the differences in the importance of keywords, so that the high weight contribution of "epidemic" is correctly reflected, and the results are more in line with the intuition of semantic importance.

[0102] In another specific implementation, the topic probability distributions of the keyword set A to be classified and the public opinion keyword set B can be constructed separately using the BERTopic model. For example, the keyword set A to be classified (such as the explanatory text and field descriptions of "SARS close contact trajectory table") can be clustered using the BERTopic model to obtain T topics (such as "close contact management", "patient privacy", "material allocation"), and their topic probability distributions are as follows: ,satisfy ( Let A represent the probability that A belongs to the i-th topic. Similarly, by clustering the set of public opinion keywords B (such as Weibo posts and news comments related to "SARS close contact trajectory leak") using the BERTopic model (using the same topic dictionary as A to ensure consistency in topic dimensions), the probability distribution of the T topics obtained is as follows: ,satisfy ( (This represents the probability that B belongs to the i-th topic).

[0103] In a specific disease control scenario, the BERTopic model is used to cluster the keyword set A corresponding to disease control texts into three core topics: Topic 1 ( ): Close contact management (including keywords such as "close contact trajectory, epidemiological investigation, isolation"), Topic 2 ( Patient privacy (including keywords such as "patient ID, medical records, information leakage"), Topic 3 ( ): If the allocation of resources (including keywords such as "vaccines, masks, and inventory") is a given topic, then the topic distribution of A (close contact trajectory data) is P = [0.85, 0.10, 0.05] (85% probability belongs to "close contact management", 10% to "patient privacy", and 5% to "resource allocation"); similarly, the topic distribution of the public opinion keyword set B (leakage of SARS close contact trajectory) on the same day is Q = [0.70, 0.25, 0.05] (70% to "close contact management", 25% to "patient privacy", and 5% to "resource allocation").

[0104] Next, the similarity between the probability distribution P of the topic to be classified and the probability distribution Q of public opinion is calculated using Jensen-Shannon divergence (JS divergence), resulting in a topic distribution similarity score. It should be noted that JS divergence is a symmetric indicator that measures the "distance" between two probability distributions. It is based on KL divergence but solves the problem of KL divergence's asymmetry and potential infinity. KL divergence (Kullback-Leibler divergence) measures the difference between a distribution P and another distribution Q, and its specific calculation formula is as follows:

[0105] ;

[0106] The KL divergence is asymmetric (KL(P||Q)≠KL(Q||P)), and when and When KL divergence is infinite, it is unsuitable for directly measuring the distance between two distributions. JS divergence addresses the asymmetry and infinity issues of KL divergence. JS divergence symmetrizes KL divergence through a "mixture distribution M," with the following steps: Step 1: Define the mixture distribution M as the average of P and Q. The specific calculation formula is: M = (P + Q) / 2, meaning that for each topic i, the probability of its mixture distribution is... And satisfy (because ,and Step 2: Calculate the JS divergence. The JS divergence is the average of the KL divergence from P to M and the KL divergence from Q to M. The specific calculation formula is: JS(P || Q) = [KL(P || M) + KL(Q || M)] / 2; Substituting this into the KL divergence formula, the expanded result is:

[0107] ;

[0108] In the formula, Let B be the probability of the set of public opinion keywords B in topic i; Let B be the probability of the set of public opinion keywords B in topic i; The probability of a mixed distribution in topic i ( ); It is a base-2 logarithm (ensuring that the JS divergence value ranges from 0 to 2). (between). When P=Q (the two distributions are completely identical), JS(P||Q)=0 (minimum distance); when P and Q are completely unrelated, JS(P||Q)=1 (maximum distance); the range of values ​​is: JS(P||Q)∈[0, 1] (the greater the correlation, the smaller the distance value).

[0109] It should be noted that JS divergence is a "distance" (the smaller the value, the closer the distribution). In order to perform multi-indicator fusion with cosine similarity score and set similarity score, JS divergence needs to be converted to "similarity" (the larger the value, the closer the distribution). The specific conversion formula is: Sim_topic = 1 - JS(P || Q); the range after conversion is: Sim_topic∈[0, 1], that is, consistent with the value range of cosine similarity score and set similarity score, which facilitates subsequent multi-indicator fusion.

[0110] In a specific disease control scenario, when P=[0.85, 0.10, 0.05] and Q=[0.70, 0.25, 0.05], the specific calculation process for the topic distribution similarity score is as follows: Step 1: Calculate the mixture distribution M. , , Therefore, M = [0.775, 0.175, 0.05]; Step 2: Calculate KL(P||M), the specific calculation formula is:

[0111] ;

[0112] Substituting P and M, we get: Topic 1: ≈ ≈0.85×0.137≈0.116; Topic 2: ≈ ≈0.10×(-0.811)≈-0.081; Topic 3: = = 0.05×0 = 0, therefore KL (P||M)≈0.116 -0.081 + 0 = 0.035; Step 3: Calculate KL (Q||M), the specific calculation formula is:

[0113] ;

[0114] Substituting Q and M, we get: Topic 1: ≈ ≈0.70×(-0.152)≈-0.106; Topic 2: ≈ ≈0.25×0.515≈0.129; Topic 3: = 0, therefore KL(Q||M)≈-0.106 + 0.129 + 0 = 0.023. Step 4: Calculate JS(P||Q), JS(P||Q) = [KL(P||M) + KL(Q||M)] / 2≈(0.035 + 0.023) / 2 =0.029; Step 5: Calculate the topic distribution similarity Sim_topic, Sim_topic = 1 - 0.029 = 0.971.

[0115] Understandably, Sim_topic≈0.971 (close to 1) indicates that the thematic distribution of "close contact trajectory data" and "public opinion on the leakage of close contact trajectory" is highly consistent, further verifying the strong correlation between the two on the themes of "close contact management" and "patient privacy". Since JS(P||Q)=JS(Q||P), the JS divergence is symmetric, thus avoiding the asymmetry problem of KL divergence. Therefore, it can more objectively measure the bidirectional topic association between "data and public opinion". Furthermore, the JS divergence range is fixed in [0,1], with stable values, and will not become infinitely large due to the probability of a certain topic being 0, which is suitable for disease control scenarios (such as the probability of the topic "material allocation" being 0 in close contact data). In addition, compared with the weighted Jaccard similarity algorithm based on "term matching", topic distribution pays more attention to "macro semantic direction", which can avoid misjudgment caused by individual keyword mismatch (such as "close contact trajectory" in the data and "close contact route" in public opinion, the terms are different but the topics are the same, so the Sim_topic value is still high). Therefore, the topic distribution similarity has higher anti-interference ability.

[0116] Step S13: Conduct a risk assessment on the public health data based on multiple data classification and evaluation dimensions to obtain the inherent risk value.

[0117] In this embodiment, public health data can be risk-assessed based on multiple data grading and assessment dimensions to obtain an inherent risk value D. The data grading and assessment dimensions are the grading and identification impact dimensions determined based on the inherent risks present in the public health data.

[0118] Furthermore, different weights can be assigned to different data classification and assessment dimensions, and the inherent risk value can be calculated through a weighted average. For example, according to the requirements of the "Implementation Rules / Guidelines for Classification and Grading of Disease Prevention and Control Data," the inherent risk of disease prevention and control data is quantified, resulting in five data classification and assessment dimensions, including sensitivity level benchmark f1, identifiability / reducibility f2, scope of impact f3, potential harm level f4, and compliance trigger factor f5. The definitions and scoring criteria for each dimension (scores have been normalized to the [0,1] interval, supporting linear interpolation or step functions) are shown in Table 3.

[0119] Table 3 Definitions and Scoring Criteria for Graded Assessment Dimensions

[0120] Dimension definition Scoring Criteria (Example) f1. Sensitivity Level Benchmark The legal sensitivity of a data type reflects its inherent risk. - Sensitive personal health information (such as case details, genetic data) → 1.0 - Group statistics but including individual characteristics → 0.7 - Completely anonymized regional statistics → 0.3 - Public epidemic reports → 0.0 f2. Identifiability / Reversibility Can data identify a specific individual, directly or indirectly? - Includes name, ID number, and mobile phone number → 1.0 - Includes quasi-identifiers (age + gender + region) with granularity < township → 0.8 - Only contains fuzzy spatiotemporal labels (e.g., city level + month) → 0.5 - K-anonymized (k≥5) or differential privacy processed → 0.2 - Completely anonymized and irreversible → 0.0 f3. ImpactScope The data involves a large population, covers a wide geographical area, and has a high degree of cross-institutional sharing. - Coverage of the entire national population → 1.0 - Coverage of the entire province → 0.8 - Coverage of prefecture-level cities → 0.6 - Coverage of districts and counties → 0.4 - Internal use by the institution → 0.2 - Individual records only → 0.1 f4. Harm Potential The impact of disclosure or misuse on individual rights and social order - Causes discrimination / personal harm / social panic → 1.0 - Causes economic loss or privacy violation → 0.7 - Affects public trust in public health decision-making → 0.5 - Minor impact → 0.3 - Almost no impact → 0.0 f5. Compliance Trigger Does it trigger specific clauses or sensitive period control strategies? - Sensitive personal information clauses triggering personal information protection → 1.0 - Data export security management regulations triggering → 0.9 - Response period for major public health events triggering → 0.8 - Industry confidentiality regulations triggering → 0.7 - No special triggers → 0.0

[0121] Next, the disease prevention and control data are assessed for risk using the five data grading and evaluation dimensions in Table 3 to obtain the inherent risk value D. The specific calculation formula is as follows:

[0122] ;

[0123] In the formula, The weights for each dimension can be initially determined using the Delphi (expert consultation) method and can be dynamically optimized based on the actual application scenario. Risk scores are assigned to the data across each dimension.

[0124] In one specific implementation, the following steps can be used to... Optimize and adjust. Step 1: Determine the initial weights using the Delphi (expert consultation) method: = [0.30, 0.25, 0.15, 0.20, 0.10]; Step 2: Adjust the initial weights based on historical event data fitting. Perform correction: ,in These are historical event correction values, among which The process is guided by expert experience to prevent subjective bias. Step 3: The corrected weights obtained in Step 2 are automatically fine-tuned based on online feedback (such as alarms and audit feedback) to achieve dynamic optimization. For example, based on information such as the frequency of data access alarms, the number of anomalies found in audits, and the results of user reports / compliance reviews recorded during system operation, the weights of high-scoring but low-alarm data are slightly reduced, while the weights of relevant dimensions are increased for low-scoring but high-frequency alarm data. This ensures that the weights are both consistent with domain experience and can be dynamically optimized with the environment. For example, the Exponential Weighted Moving Average (EMA) method can be used for weight fine-tuning.

[0125] Step S14: Determine the data source of the multi-source online public opinion data, and calculate the product of the source weight corresponding to the data source and the keyword similarity score to obtain the source credibility of the multi-source online public opinion data.

[0126] In this embodiment, the data source c of the multi-source online public opinion data is first determined (i.e., the channel / source of public opinion dissemination, such as "official CDC", "CCTV News", "Weibo personal account", etc.), and then the source weight corresponding to data source c is calculated. (That is, the credibility weight of channel C, where the source weight of official channels is greater than that of unofficial channels, which needs to be set in advance according to business rules, such as official disease control centers.) =1.0, Weibo personal =0.3) is multiplied by the keyword similarity score above to obtain the source credibility of multi-source online public opinion data.

[0127] Step S15: Determine the time interval between the release time of the multi-source network public opinion data and the current time, and input the time interval into the exponential decay function to generate the failure decay factor. Then calculate the product of the failure decay factor and the source credibility to obtain the public opinion similarity score.

[0128] In this embodiment, after obtaining the source credibility of multi-source network public opinion data, the time interval t' between the public opinion release time and the current time is determined, that is, the lag days, in days; then, the time interval t' is input into the exponential decay function to generate the failure decay factor, and then the product of the failure decay factor and the source credibility is calculated to obtain the public opinion similarity score.

[0129] Specifically, the formula for calculating the public opinion similarity score is as follows:

[0130] ;

[0131] in, ;

[0132] In the formula, The public opinion similarity score is given. The time decay coefficient (controls the decay rate) The higher the value, the faster the public opinion heat dissipates, such as =0.2, which aligns with the characteristics of rapid escalation and dissipation of public opinion regarding the epidemic; t' represents the aforementioned time interval. For attenuation adjustment parameters (used to further adjust the "steepness" of attenuation, when...) =1, the exponential decay function simplifies to , This indicates the source weight corresponding to data source c (e.g., official > mainstream media > social media platforms). The cosine similarity score (representing the overall semantic association) is given. The similarity score for the set (representing keyword matching and importance) is given. The similarity score for the topic distribution (representing macro-level topic association) is given. The cosine similarity score corresponds to the similarity coefficient. The similarity coefficient corresponds to the set similarity score. The similarity coefficient corresponding to the topic distribution similarity score satisfies the following conditions: It can be adjusted according to the scenario, such as This highlights the core role of semantic similarity.

[0133] in, This represents the basic comprehensive similarity (without considering timeliness or source) between a piece of public opinion data from channel c and the public health data to be classified. It is achieved by weighting and fusing the similarity across three dimensions: semantics, terms, and topics, into a single basic similarity score. This score provides the "original strength" for subsequent timeliness decay correction. For example, when a public opinion item "close contact trajectory leakage" and "close contact trajectory data" are similar... =0.8、 =0.6、 =0.9, and Then: =0.4×0.8+0.3×0.6+0.3×0.9=0.32+0.18+0.27=0.77.

[0134] Specifically, This indicates that different versions of the same public opinion are disseminated through different channels, weighted and merged according to the credibility of the channel / source, thereby avoiding interference from low-credibility sources (such as rumors on social media platforms) in risk assessment. For example, when "close contact trajectory leakage" is simultaneously reported on the official website of the CDC (… ) and Weibo personal accounts ( When it spreads, then: =1.0×0.77+0.3×0.65=0.77+0.195=0.965.

[0135] It should be noted that this application uses an exponential decay function to simulate the timeliness of public opinion. The longer the public opinion has been released, the smaller its impact on the current risk, and the decay rate increases from... and control. This represents the time-decrease factor (i.e., the exponential decay function, which is 1 when t'=0 and rapidly approaches 0 as t' increases, reflecting the characteristic that the impact of public opinion on the current day is the greatest, while the impact of old public opinion is negligible), and It can simulate timeliness, better aligning with the "rapid decay followed by slow decay" pattern of public opinion dissemination (e.g., a sharp drop in public opinion intensity on the day of release, followed by a slow decay after 3 days), and can more precisely demonstrate the dynamic impact of timeliness on risk. For example, if the public opinion data is released on the same day (i.e., t'=0), then the timeliness decay factor... If the public opinion data was released 3 days ago (i.e.) =0, =0.2, =1), then the time decay factor .

[0136] In addition, the public opinion similarity score This indicates the similarity of public opinion risks after adjustments for time and credibility. In a disease control scenario, if the keyword set B containing "close contact trajectories leaked" was published 3 days ago, and simultaneously on the official disease control center (… ) and Weibo personal ( The propagation occurs, and t'=3. =0.2, =1, then: = ×(1.0×0.77+0.3×0.65)≈0.549×0.965≈0.530. By considering the release time of public opinion (time decay) + including non-authoritative sources (credibility weighting), the risk similarity of the data can be reduced from the basic similarity of 0.77 (i.e., It dropped to 0.53 (i.e.) This approach better aligns with the actual risk impact of "old public opinion + some non-authoritative dissemination." Furthermore, the dynamic correction method described above overcomes the inherent shortcomings of traditional static methods, such as delayed response and a disconnect between protection efforts and actual risks.

[0137] Step S16: The public opinion similarity score is fused with the inherent risk value to obtain a comprehensive risk value, and the target data level corresponding to the public health data is determined based on the target risk range in which the comprehensive risk value is located, so as to determine whether to carry out public health early warning processing based on the target data level.

[0138] In this embodiment, after obtaining the public opinion similarity score and the inherent risk value, the two values ​​are fused to obtain a comprehensive risk value. Then, the target risk range in which the comprehensive risk value is located is determined, and the target data level corresponding to the public health data is determined based on the target risk range. Then, it is determined whether the target data level exceeds the preset warning level. If the target data level exceeds the preset warning level, the corresponding public health warning processing is carried out.

[0139] Specifically, see Table 3, where the data level is determined based on the pre-set interval.

[0140] Table 4 Data Level Determination Rules

[0141] Data types Risk range Data Level Core data >0.8 L4 Particularly important data [0.5,0.8] L3 Important data [0.35,0.5) L2 General Data <0.35 L1

[0142] For example, if the target risk range where the comprehensive risk value R is located is [0.5, 0.8], then the target data level corresponding to the current public health data is determined to be L3. For risk ranges exceeding the L2 data level (i.e., the risk ranges corresponding to L3 and L4), corresponding alarm information can be generated and sent to the corresponding terminal.

[0143] Furthermore, to improve the speed of emergency response to different types of abnormal data, different data security policies (such as access control, data masking rules, and sharing scope) can be pre-set for different data levels. For example, the strictest access control, full data masking, and encryption can be implemented for Level 4; partial data masking, limited sharing, and enhanced auditing can be implemented for Level 3; sharing can be opened under compliance and basic protection can be retained for Level 2; and visitor information can be recorded under compliance. Through the above methods, the linkage and dynamic adjustment of the hierarchical results and differentiated protection policies can be achieved. After determining the corresponding data security policies, information such as the highest weight dimension, keywords / topics, comprehensive risk value, target risk range, target data level, and policy version number of the target data security policy can be recorded, thus forming an interpretable and auditable closed-loop mechanism. It also supports online optimization and adaptive adjustment of weights and thresholds based on audit feedback.

[0144] As can be seen, the embodiments of this application first use a pre-trained language model to calculate the similarity between different keywords in public health data and multi-source network public opinion data. Then, the source weight corresponding to the source of public opinion data is calculated as a product of the keyword similarity to obtain the source credibility of the public opinion data. Next, an invalidation decay factor is generated based on the lag time of the public opinion data (i.e., the time interval between the release time and the current time) and the exponential decay function. The public opinion similarity score is obtained by calculating the product of the invalidation decay factor and the source credibility. Finally, the public opinion similarity score is fused with the inherent risk value (the value obtained after risk assessment of public health data based on multiple data classification and evaluation dimensions). The data level of public health data is determined based on the risk range of the comprehensive risk value obtained after fusion, so as to determine whether to carry out public health early warning processing based on the data level. As can be seen, this application embodiment, based on public health data, also integrates dynamically changing multi-source online public opinion data, enabling the data classification results to be combined with the actual situation. It can dynamically adjust the data classification level according to the evolution of public opinion. Furthermore, by calculating the similarity between the keywords to be classified and the keywords in public opinion, the semantic relationship between public health data and public opinion data can be mined, thereby significantly improving the accuracy of risk perception. Moreover, this application embodiment not only quantifies the inherent risks of public health data through multiple data classification assessment dimensions, but also quantifies its external risks through public opinion analysis, and introduces an attenuation factor based on the timeliness and source credibility of public opinion, thereby overcoming the one-sidedness of classification from a single risk perspective. In addition, automated data classification through artificial intelligence not only improves the timeliness of data classification, but also overcomes the problems of strong reliance on manual methods and low efficiency, thus enabling accurate and timely public health early warning processing.

[0145] For details, see Figure 2 As shown, the AI ​​(Artificial Intelligence) big data model is used to extract keywords and calculate similarity between infectious disease monitoring and early warning data and multi-source online public opinion data on the same day. The similarity value is then fused with the inherent risk value obtained after risk assessment of the monitoring and early warning data according to preset guidelines to obtain a comprehensive risk value. Based on the target risk range where the comprehensive risk value is located, the target data level corresponding to the infectious disease monitoring and early warning data is determined, so as to determine whether to carry out public health early warning processing based on the target data level.

[0146] This application discloses a specific data dynamic supervision method based on intelligent agents. See [link to relevant documentation] Figure 3 As shown, the method includes:

[0147] Step S21: Obtain public health data and multi-source online public opinion data related to the public health data.

[0148] Step S22: Use a pre-trained language model to extract keywords from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords, and calculate the similarity between the keywords to be classified and the public opinion keywords to obtain keyword similarity scores.

[0149] Step S23: Conduct a risk assessment on the public health data based on multiple data classification and evaluation dimensions to obtain the inherent risk value.

[0150] Step S24: Determine the data source of the multi-source online public opinion data, and calculate the product of the source weight corresponding to the data source and the keyword similarity score to obtain the source credibility of the multi-source online public opinion data.

[0151] Step S25: Determine the time interval between the release time of the multi-source network public opinion data and the current time, and input the time interval into the exponential decay function to generate the failure decay factor. Then calculate the product of the failure decay factor and the source credibility to obtain the public opinion similarity score.

[0152] Step S26: Normalize the public opinion similarity score and the inherent risk value respectively to obtain the normalized public opinion similarity value and the normalized risk value.

[0153] In this embodiment, after obtaining the public opinion similarity score... After determining the inherent risk value D, considering the potential shifts in distribution across different datasets, the two scores can be normalized, for example, by using the Min-Max standard normalization method. By linearly mapping D to the [0,1] interval, we obtain the normalized public opinion similarity value. and normalized risk value .

[0154] Among them, the normalized risk value The specific calculation formula is as follows:

[0155] ;

[0156] In the formula, , These represent the minimum and maximum risk values ​​in the current entire database, respectively. They can be recalculated monthly or quarterly to maintain comparability; if Then the normalized risk value Set it to 0.5.

[0157] For example, as shown in Table 4, after conducting a risk assessment on data objects in the field of disease prevention and control through five data classification assessment dimensions (sensitivity f1, identifiability f2, scope of impact f3, degree of harm f4, and compliance trigger f5), five risk scores were obtained, which were 1.0, 1.0, 0.6, 1.0, and 1.0, respectively.

[0158] Table 5 Data Level Determination Rules

[0159] Data objects List of HIV-positive cases in a certain city (including name, ID number, and address) f1 (sensitivity) 1.0 (Statutory Sensitive Personal Information) f2 (identifiability) 1.0 (Direct Recognition) f3 (Scope of Impact) 0.6 (covering the entire city, approximately 5 million people) f4 (Level of hazard) 1.0 (The leak led to severe discrimination and social panic) f5 (Compliance Trigger) 1.0 (Triggering Sensitive Personal Information Clause)

[0160] Furthermore, the weights for each dimension are w = [0.30, 0.25, 0.15, 0.20, 0.10]; therefore, the inherent risk value D = 0.3×1 + 0.25×1 + 0.15×0.6 + 0.2×1 + 0.1×1 = 0.3 + 0.25 + 0.09 + 0.2 + 0.1 = 0.94. Additionally, in the current database... =0.98, =0.20, then the normalized risk value for: This is considered high-risk.

[0161] Step S27: The normalized public opinion similarity value and the normalized risk value are weighted and fused to obtain a comprehensive risk value. The target data level corresponding to the public health data is determined based on the target risk range in which the comprehensive risk value is located, so as to determine whether to carry out public health early warning processing based on the target data level.

[0162] In this embodiment, the normalized public opinion similarity value is... and normalized risk value A weighted fusion is performed to obtain a comprehensive risk value; wherein, the formula for calculating the comprehensive risk value is:

[0163] ;

[0164] In the formula, R is the comprehensive risk value. Indicates the fusion weight coefficient. This represents the normalized risk value. This represents the normalized public opinion similarity value.

[0165] For more detailed processing procedures of steps S21 to S25, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0166] As can be seen, the embodiments of this application use intelligent agents (such as AI agents) to dynamically monitor public health data. Based on public health data, multi-source online public opinion data is combined to perceive changes in external public opinion in real time and dynamically classify public health data according to changes in public opinion. In addition, this mechanism ensures that during periods of public opinion sensitivity, high-risk data can automatically trigger stricter protection strategies to effectively prevent information leakage; while during stable periods, restrictions are appropriately relaxed, thereby significantly improving data sharing efficiency and utilization flexibility. At the same time, the entire process is automated through intelligent agents, which greatly shortens decision-making time, enhances the real-time performance, accuracy and compliance of public health data management, and achieves a dynamic balance between security and effectiveness.

[0167] Accordingly, this application also discloses a data dynamic monitoring device based on intelligent agents, see [link to relevant documentation]. Figure 4 As shown, the device includes:

[0168] Data acquisition module 11 is used to acquire public health data and multi-source online public opinion data related to the public health data;

[0169] Keyword extraction module 12 is used to extract keywords from the public health data and the multi-source online public opinion data using a pre-trained language model, to obtain keywords to be classified and public opinion keywords;

[0170] The first calculation module 13 is used to calculate the similarity between the keyword to be classified and the public opinion keyword, and obtain the keyword similarity score;

[0171] Risk assessment module 14 is used to conduct risk assessment on the public health data based on multiple data classification and assessment dimensions to obtain an inherent risk value;

[0172] The second calculation module 15 is used to determine the data source of the multi-source online public opinion data, and calculate the product of the source weight corresponding to the data source and the keyword similarity score to obtain the source credibility of the multi-source online public opinion data.

[0173] The time interval determination module 16 is used to determine the time interval between the release time of the multi-source network public opinion data and the current time;

[0174] The third calculation module 17 is used to input the time interval into the exponential decay function to generate a failure decay factor, and then calculate the product of the failure decay factor and the source credibility to obtain the public opinion similarity score.

[0175] The fusion module 18 is used to fuse the public opinion similarity score with the inherent risk value to obtain a comprehensive risk value;

[0176] The data classification module 19 is used to determine the target data level corresponding to the public health data based on the target risk range in which the comprehensive risk value is located, so as to determine whether to carry out public health early warning processing based on the target data level.

[0177] The specific workflow of each of the above modules can be found in the relevant content disclosed in the foregoing embodiments, and will not be repeated here.

[0178] As can be seen, this application embodiment, based on public health data, also integrates dynamically changing multi-source online public opinion data, enabling the data classification results to be combined with the actual situation. It can dynamically adjust the data classification level according to the evolution of public opinion. Furthermore, by calculating the similarity between the keywords to be classified and the keywords in public opinion, the semantic relationship between public health data and public opinion data can be mined, thereby significantly improving the accuracy of risk perception. Moreover, this application embodiment not only quantifies the inherent risks of public health data through multiple data classification assessment dimensions, but also quantifies its external risks through public opinion analysis, and introduces an attenuation factor based on the timeliness and source credibility of public opinion, thereby overcoming the one-sidedness of classification from a single risk perspective. In addition, automated data classification through artificial intelligence not only improves the timeliness of data classification, but also overcomes the problems of strong reliance on manual methods and low efficiency, thus enabling accurate and timely public health early warning processing.

[0179] In some specific embodiments, the keyword extraction module 12 may further include:

[0180] The preprocessing unit is used to perform preprocessing operations on the public health data and the multi-source online public opinion data respectively, to obtain multiple words to be classified and segmented and multiple public opinion words; the preprocessing operations include unified encoding and word segmentation processing;

[0181] The scoring processing unit is used to input multiple words to be classified and multiple public opinion words into a preset multi-channel for scoring processing, so as to obtain multiple scores corresponding to a single word to be classified and multiple public opinion scores corresponding to a single public opinion word; the preset multi-channel includes a statistical channel of the base text keyword extraction algorithm, an entity importance channel based on a preset named entity recognition model, and a dictionary matching channel based on a preset dictionary;

[0182] The first calculation unit is used to calculate the comprehensive weight of each of the words to be classified based on the multiple scores corresponding to each word to be classified and the channel weights corresponding to each channel in the preset multi-channel, and to obtain the comprehensive weight of the words to be classified by weighted fusion.

[0183] The second calculation unit is used to calculate the comprehensive weight of each public opinion word based on the multiple public opinion scores corresponding to each of the aforementioned public opinion words and the multiple channel weights, and to obtain the comprehensive public opinion weight by weighted fusion.

[0184] The first filtering unit is used to filter keywords in multiple word segments to be classified based on the comprehensive weight to be classified, to obtain a set of keywords to be classified that includes the comprehensive weight to be classified, and to filter keywords in multiple word segments to be classified based on the comprehensive weight to be classified, to obtain a set of keywords to be classified based on the comprehensive weight to be classified, to obtain a set of keywords to be classified based on the comprehensive weight to be classified.

[0185] In some specific embodiments, the first screening unit may specifically include:

[0186] The first sorting unit is used to sort the multiple weights to be graded and the multiple public opinion weights in descending order to obtain the sorted weights to be graded and the sorted public opinion weights.

[0187] The second filtering unit is used to filter out the first preset number of comprehensive weights to be classified from the sorted weights to be classified, and to use the words to be classified corresponding to the first preset number of comprehensive weights to be classified as keywords, so as to obtain a set of keywords to be classified that includes the comprehensive weights to be classified.

[0188] The third filtering unit is used to filter out a preset number of comprehensive public opinion weights from the sorted public opinion weights, and use the public opinion word segments corresponding to the preset number of comprehensive public opinion weights as keywords to obtain a set of public opinion keywords containing the comprehensive public opinion weights.

[0189] In some specific embodiments, the first screening unit may specifically include:

[0190] The second sorting unit is used to sort the multiple weights to be graded and the multiple public opinion weights in descending order to obtain the sorted weights to be graded and the sorted public opinion weights.

[0191] The first determining unit is used to determine the maximum value of the sorted weight to be graded and the sorted public opinion weight respectively, so as to obtain the maximum value of the weight to be graded and the maximum value of the public opinion weight.

[0192] The third calculation unit is used to calculate the product of the maximum value of the weight to be classified and the preset cutoff factor, and the product of the maximum value of the public opinion weight and the preset cutoff factor, to obtain the first product and the second product.

[0193] The second determining unit is used to determine the comprehensive weight to be classified that is greater than or equal to the first product among the sorted weights to be classified, and to take the words to be classified corresponding to the comprehensive weights to be classified that are greater than or equal to the first product as keywords, so as to obtain a set of keywords to be classified that includes the comprehensive weights to be classified.

[0194] The third determining unit is used to determine the comprehensive public opinion weight that is greater than or equal to the second product among the sorted public opinion weights, and to take the public opinion word segment corresponding to the comprehensive public opinion weight that is greater than or equal to the second product as keywords, so as to obtain a set of public opinion keywords containing the comprehensive public opinion weight.

[0195] In some specific embodiments, the first computing module 13 may specifically include:

[0196] The encoding unit is used to encode each keyword in the set of keywords to be classified and the set of public opinion keywords using an encoding model that includes a word segmenter, so as to obtain multiple word vectors to be classified and multiple public opinion word vectors.

[0197] The mean pooling processing unit is used to perform mean pooling processing on each of the word vectors to be classified and each of the public opinion word vectors respectively, to obtain the pooled word vectors to be classified and the pooled public opinion vectors.

[0198] The first similarity calculation unit is used to calculate the similarity between the pooled vector to be classified and the pooled public opinion vector using the cosine similarity algorithm, and obtain the cosine similarity score.

[0199] The second similarity calculation unit is used to calculate the similarity between the set of keywords to be classified and the set of public opinion keywords using a weighted Jaccard similarity algorithm based on the comprehensive weight of the keywords to be classified and the comprehensive weight of public opinion, and to obtain a set similarity score.

[0200] The construction unit is used to construct the topic probability distribution of the set of keywords to be classified and the set of public opinion keywords respectively through the BERTopic model, so as to obtain the topic probability distribution and the public opinion probability distribution to be classified.

[0201] The third similarity calculation unit is used to calculate the similarity between the probability distribution of the topic to be classified and the probability distribution of public opinion through JS divergence, and obtain the topic distribution similarity score.

[0202] In some specific embodiments, the formula for calculating the public opinion similarity score is as follows:

[0203] ;

[0204] in, ;

[0205] In the formula, The public opinion similarity score is given. Here, t' represents the time decay coefficient. For attenuation adjustment parameters, This indicates the source weight corresponding to data source c. The cosine similarity score is... The similarity score of the set, The similarity score for the topic distribution is... The cosine similarity score corresponds to the similarity coefficient. The similarity coefficient corresponds to the set similarity score. The similarity coefficient corresponds to the similarity score of the topic distribution.

[0206] In some specific embodiments, the fusion module 18 may specifically include:

[0207] The normalization unit is used to normalize the public opinion similarity score and the inherent risk value respectively, so as to obtain the normalized public opinion similarity value and the normalized risk value.

[0208] The weighted fusion unit is used to perform weighted fusion of the normalized public opinion similarity value and the normalized risk value to obtain a comprehensive risk value;

[0209] The formula for calculating the comprehensive risk value is as follows:

[0210] ;

[0211] In the formula, R is the comprehensive risk value. Indicates the fusion weight coefficient. This represents the normalized risk value. This represents the normalized public opinion similarity value.

[0212] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0213] Figure 5This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the agent-based data dynamic monitoring method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0214] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0215] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0216] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the agent-based data dynamic monitoring method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0217] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned agent-based data dynamic supervision method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0218] Furthermore, embodiments of this application also disclose a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the aforementioned agent-based data dynamic monitoring method.

[0219] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0220] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0221] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0222] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only 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.

[0223] The above provides a detailed description of the data dynamic monitoring method, apparatus, and device based on intelligent agents provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data dynamic monitoring method based on intelligent agents, characterized in that, include: Acquire public health data and multi-source online public opinion data related to the public health data; Using a pre-trained language model, keywords are extracted from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords. The similarity between the keywords to be classified and the public opinion keywords is calculated to obtain keyword similarity scores. The public health data is risk-assessed based on multiple data grading and evaluation dimensions to obtain an inherent risk value. The data sources of the multi-source online public opinion data are determined, and the product of the source weight corresponding to the data source and the keyword similarity score is calculated to obtain the source credibility of the multi-source online public opinion data. The time interval between the release time of the multi-source network public opinion data and the current time is determined, and the time interval is input into the exponential decay function to generate the failure decay factor. Then, the product of the failure decay factor and the source credibility is calculated to obtain the public opinion similarity score. The public opinion similarity score is fused with the inherent risk value to obtain a comprehensive risk value. Based on the target risk range in which the comprehensive risk value is located, the target data level corresponding to the public health data is determined, and the public health early warning processing is determined based on the target data level.

2. The data dynamic monitoring method based on intelligent agents according to claim 1, characterized in that, The process involves extracting keywords from the public health data and the multi-source online public opinion data to obtain keywords to be classified and public opinion keywords, including: The public health data and the multi-source online public opinion data are preprocessed to obtain multiple words to be classified and segmented and multiple public opinion words; the preprocessing operation includes unified encoding and word segmentation processing. Multiple words to be classified and multiple words to be public opinion are input into a preset multi-channel for scoring processing to obtain multiple scores for a single word to be classified and multiple public opinion scores for a single word to be public opinion; the preset multi-channel includes a statistical channel for the base text keyword extraction algorithm, an entity importance channel based on a preset named entity recognition model, and a dictionary matching channel based on a preset dictionary; Based on the multiple scores corresponding to each of the words to be classified and the channel weights of each channel in the preset multi-channel, the comprehensive weight of each of the words to be classified is calculated by weighted fusion to obtain the comprehensive weight of the words to be classified. Based on the multiple public opinion scores corresponding to each of the aforementioned public opinion segments and the multiple channel weights, the comprehensive weight of each of the aforementioned public opinion segments is calculated by weighted fusion to obtain the comprehensive public opinion weight. Based on the comprehensive weight to be classified, keywords from multiple words to be classified are filtered to obtain a set of keywords to be classified that include the comprehensive weight to be classified. Based on the comprehensive weight to be classified, keywords from multiple words to be classified are filtered to obtain a set of keywords to be classified that include the comprehensive weight to be classified.

3. The data dynamic monitoring method based on intelligent agents according to claim 2, characterized in that, The process involves filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of keywords to be classified that includes the comprehensive weight, and then filtering keywords from multiple segmented words based on the comprehensive weight of public opinion to obtain a set of public opinion keywords that includes the comprehensive weight of public opinion, including: The multiple weights to be classified and the multiple public opinion weights are sorted in descending order to obtain the sorted weights to be classified and the sorted public opinion weights. From the sorted weights to be graded, select the first preset number of comprehensive weights to be graded, and use the words to be graded corresponding to the first preset number of comprehensive weights to be graded as keywords to obtain a set of keywords to be graded that includes the comprehensive weights to be graded. From the sorted public opinion weights, a preset number of comprehensive public opinion weights are selected, and the public opinion words corresponding to the preset number of comprehensive public opinion weights are used as keywords to obtain a set of public opinion keywords containing the comprehensive public opinion weights.

4. The data dynamic monitoring method based on intelligent agents according to claim 2, characterized in that, The process involves filtering keywords from multiple segmented words to be classified based on the comprehensive weight to obtain a set of keywords to be classified that includes the comprehensive weight, and then filtering keywords from multiple segmented words based on the comprehensive weight of public opinion to obtain a set of public opinion keywords that includes the comprehensive weight of public opinion, including: The multiple weights to be classified and the multiple public opinion weights are sorted in descending order to obtain the sorted weights to be classified and the sorted public opinion weights. The maximum values ​​of the sorted weights to be classified and the sorted public opinion weights are determined respectively, thus obtaining the maximum value of the weights to be classified and the maximum value of the public opinion weights. Calculate the product of the maximum value of the weight to be classified and the preset cutoff factor, and the product of the maximum value of the public opinion weight and the preset cutoff factor to obtain the first product and the second product; Determine the comprehensive weight to be classified that is greater than or equal to the first product among the weights to be classified after sorting, and take the words to be classified corresponding to the comprehensive weights to be classified that are greater than or equal to the first product as keywords, to obtain a set of keywords to be classified that contain the comprehensive weights to be classified. Determine the comprehensive public opinion weights that are greater than or equal to the second product among the sorted public opinion weights, and use the public opinion word segments corresponding to the comprehensive public opinion weights that are greater than or equal to the second product as keywords to obtain a set of public opinion keywords containing the comprehensive public opinion weights.

5. The data dynamic monitoring method based on intelligent agents according to claim 2, characterized in that, The calculation of the similarity between the keyword to be classified and the public opinion keyword to obtain a keyword similarity score includes: The keywords in the set of keywords to be classified and the set of public opinion keywords are encoded using an encoding model that includes a word segmenter, resulting in multiple word vectors to be classified and multiple public opinion word vectors. Each of the word vectors to be classified and each of the public opinion word vectors are subjected to mean pooling to obtain the pooled word vectors to be classified and the pooled public opinion vectors. The cosine similarity algorithm is used to calculate the similarity between the pooled vector to be classified and the pooled sentiment vector to obtain the cosine similarity score. The similarity between the set of keywords to be classified and the set of public opinion keywords is calculated using a weighted Jaccard similarity algorithm based on the comprehensive weight of the keywords to be classified and the comprehensive weight of public opinion, and a set similarity score is obtained. The topic probability distributions of the keyword set to be classified and the public opinion keyword set are constructed using the BERTopic model, respectively, to obtain the topic probability distribution to be classified and the public opinion probability distribution. The similarity between the topic probability distribution to be classified and the public opinion probability distribution is calculated using JS divergence to obtain the topic distribution similarity score.

6. The data dynamic monitoring method based on intelligent agents according to claim 5, characterized in that, The formula for calculating the public opinion similarity score is as follows: ； in, ; In the formula, The public opinion similarity score is given. Here, t' represents the time decay coefficient. For attenuation adjustment parameters, This indicates the source weight corresponding to data source c. The cosine similarity score is... The similarity score of the set, The similarity score for the topic distribution is... The cosine similarity score corresponds to the similarity coefficient. The similarity coefficient corresponds to the set similarity score. The similarity coefficient corresponds to the similarity score of the topic distribution; This indicates the basic overall similarity between a certain piece of public opinion data in channel c and the public health data to be classified.

7. The data dynamic monitoring method based on intelligent agents according to any one of claims 1 to 6, characterized in that, The process of fusing the public opinion similarity score with the inherent risk value to obtain a comprehensive risk value includes: The public opinion similarity score and the inherent risk value are normalized respectively to obtain the normalized public opinion similarity value and the normalized risk value; The normalized public opinion similarity value and the normalized risk value are weighted and fused to obtain a comprehensive risk value; The formula for calculating the comprehensive risk value is as follows: ； In the formula, R is the comprehensive risk value. Indicates the fusion weight coefficient. This represents the normalized risk value. This represents the normalized public opinion similarity value.

8. A data dynamic monitoring device based on intelligent agents, characterized in that, include: The data acquisition module is used to acquire public health data and multi-source online public opinion data related to the public health data; The keyword extraction module is used to extract keywords from the public health data and the multi-source online public opinion data using a pre-trained language model, respectively, to obtain keywords to be classified and public opinion keywords. The first calculation module is used to calculate the similarity between the keyword to be classified and the public opinion keyword, and obtain the keyword similarity score; The risk assessment module is used to assess the risk of the public health data based on multiple data classification and assessment dimensions to obtain the inherent risk value. The second calculation module is used to determine the data source of the multi-source online public opinion data, and calculate the product of the source weight corresponding to the data source and the keyword similarity score to obtain the source credibility of the multi-source online public opinion data. The time interval determination module is used to determine the time interval between the release time of the multi-source online public opinion data and the current time; The third calculation module is used to input the time interval into the exponential decay function to generate a failure decay factor, and then calculate the product of the failure decay factor and the source credibility to obtain the public opinion similarity score. The fusion module is used to fuse the public opinion similarity score with the inherent risk value to obtain a comprehensive risk value; The data classification module is used to determine the target data level corresponding to the public health data based on the target risk range in which the comprehensive risk value is located, so as to determine whether to carry out public health early warning processing based on the target data level.

9. An electronic device, characterized in that, It includes a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the agent-based data dynamic supervision method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the agent-based data dynamic supervision method as described in any one of claims 1 to 7.

Images