Composite model for identifying public health false information and construction method

Abstract

The invention discloses a composite model for identifying public health false information and a construction method, and relates to the technical field of information data processing, and the composite model comprises a data acquisition module, an information processing and analysis module, a false information identification module and a rumor reverse module. An advanced natural language processing technology and a large language model are adopted, so that the system can accurately and quickly recognize false information and rumors from a large amount of information, the accuracy and speed of false information recognition can be improved, meanwhile, a continuous learning and optimization mechanism is included, new false information strategies can be continuously adapted, and the recognition efficiency is improved. And the progress of related technologies and the updating of public health knowledge are promoted, so that false information is intervened in time, the market and social order can be maintained, and the economic loss caused by panic is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of information data processing, and in particular to a composite model and a construction method for identifying false public health information. Background Art

[0002] In the globalized information age, especially in the face of public health crises such as infectious disease outbreaks and epidemics, accurately and rapidly identifying false information and effectively managing and refuting it are essential to protecting public safety and health. The spread of false information may not only cause public panic, but also interfere with the implementation of effective disease prevention and control measures, thereby exacerbating the impact of public health crises.

[0003] Identifying and handling false information is a long-standing and complex challenge in the field of public health. With the widespread use of the Internet and social media, information spreads extremely quickly, causing false information to spread rapidly, bringing unprecedented challenges to public health management. Current technologies mainly rely on keyword filtering, user reports, and expert verification to identify false information, but these methods have problems such as large delays, low accuracy, and difficulty in dealing with large-scale information flows.

[0004] In the field of artificial intelligence, although large language models and deep learning algorithms have been applied to text recognition and processing, they still face several technical problems in practical applications:

[0005] 1. Adaptability and flexibility: Existing models often require extensive customization and training to adapt to specific types of false information, which may not be efficient enough in a rapidly changing public health crisis;

[0006] 2. Depth of contextual understanding: False information often involves complex backgrounds and subtle semantic differences. Existing models have difficulty accurately capturing and understanding these details, especially in cross-cultural and cross-linguistic environments.

[0007] 3. Automatic update and learning capabilities: The form and content of false information are constantly evolving, requiring the model to be able to continuously learn and self-update to adapt to new false information strategies;

[0008] The above technical problems show that although there have been significant progress in the identification and processing of false information, existing technologies still cannot fully meet the urgent needs in the public health field for efficient and accurate identification and refutation of false information.

[0009] In view of the above technical defects, a solution is now proposed. Summary of the invention

[0010] The purpose of the present invention is to adopt advanced natural language processing technology and large language models to enable the system to quickly and accurately identify false information and rumors from a large amount of information, thereby improving the accuracy and speed of false information identification. At the same time, it includes a continuous learning and optimization mechanism, which can continuously adapt to new false information strategies, promote the advancement of related technologies and the updating of public health knowledge, and then intervene in false information in a timely manner, which will help maintain market and social order and reduce economic losses caused by panic.

[0011] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a composite model for identifying false public health information and a construction method thereof, wherein the composite model comprises a data acquisition module, an information processing and analysis module, a false information identification module and a rumor refutation module, wherein the data acquisition module is used to collect text data related to public health on an online platform, the information processing and analysis module performs preprocessing and feature extraction on the acquired data, the false information identification module constructs a customized large language model and performs deep learning training, and the rumor refutation module is used to generate accurate and authoritative information to refute the identified false information, and the construction method specifically comprises the following steps:

[0012] Step 1: Obtain the latest announcements and alerts through the official websites and social media of national and local public health departments, obtain a database of public health emergencies, and characterize the database to obtain a data set;

[0013] Step 2: preprocess and extract features of the acquired data;

[0014] Step 3: Build a large language model based on deep learning algorithms and conduct deep learning training;

[0015] Step 4: Build a composite model based on the big voice model to integrate the spread of false information with epidemiological prediction;

[0016] Step 5: Construct an evaluation index system based on the system performance of the large language model;

[0017] The above system performance evaluation index system includes accuracy (Acc), timeliness (Timeliness), and specificity (Specificity), and its evaluation function is expressed as (E sys (F DI , M CP , Θ, Ψ)):

[0018] [E sys ={Acc(F DI ), Timeliness(M CP ), Specificity(F DI )}];

[0019] Step 6: Conduct empirical analysis and refine crisis response strategies based on the large language model;

[0020] Step 7: Establish a real-time monitoring system to monitor public health information. At the same time, use the constructed detection algorithm to detect false information in real-time monitoring. When false information is detected, immediately execute the response mechanism: [Response (S monitor , Detection(D, A))].

[0021] Furthermore, the database includes text data, audio data and video data, that is, [D = {d 1 , d 2 , ..., d n}], where (D) represents the collected data set, (d i ) represents a single data sample.

[0022] For each data sample (d i ), further describing its feature set: [F i ={f i1 , f i2 , ..., f im}], where (F i ) represents the data sample (d i ) feature set, (f ij ) represents the value of a specific feature.

[0023] The selection of data source is represented by a weight function, considering the importance of different data sources, which can be expressed as: [W s ={w 1 , w 2 , ..., w k}], where (W s ) represents the weights of different data sources, (w k ) represents the weight of the (k)th data source;

[0024] The data set obtained is: selected ={di|di∈D ,满足选择条件共同卫生相关的资讯}].

[0025] Furthermore, the specific process of preprocessing and feature extraction of the acquired data is as follows:

[0026] S101. Data preprocessing: clean the data and remove missing values ​​and outliers: [D cleaned ={di|di∈D selected,满足清洗条件}];

[0027] S102, add a label to each data sample: [L = {l 1, l 2 , ..., l n}], where (L) represents the label set, i.e., real news, manually written news, and fake news generated by GPT, (l i ) represents the data sample (d i ) corresponding labels;

[0028] S103, the specific process of data classification is represented by the classification function: [C(d i )={c 1 , c 2 , ..., c k}], where (C(d i )) represents the data sample (d i ) classification results, (c k ) denotes category (k);

[0029] S104, obtain the data set for model training and testing: [D final ={(Fi,li)|di∈D cleaned ,Fi is the sample feature, Li is the corresponding label}].

[0030] Furthermore, the specific process of building a large language model is as follows:

[0031] S201, define a false information identification function (F DI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, (x) is converted into a feature vector (f(x)) by applying a high-level semantic mapping: [f(x) = F DI (x,Θ)], the mapping uses natural language processing techniques, including semantic understanding and context analysis;

[0032] S202, introduce false information loss function (L DI (f(x), y)), where (y) is the label, and the difference between the model prediction (f(x)) and the true label (y) is quantified by the false information loss function, that is, [L DI (f(x),y) = loss_metric(f(x),y)];

[0033] S203, multi-dimensionally setting the optimization goal of the large speech model, wherein the optimization goal includes minimizing the false information loss function and maximize the accuracy of real information

[0034] S204, Introducing sensitivity indicators Used to focus on specific types of false information: Where (T) represents a specific type of false information category;

[0035] S205. Establish a large language model fine-tuning technology in the public health field based on Low-Rank Adaptation. LoRA is a powerful parameter adjustment strategy, especially suitable for large language models such as the GPT series. By introducing low-rank matrices, model parameters can be adjusted more flexibly to adapt to specific tasks. In the false information identification task, the LoRA fine-tuning strategy is defined as: Where (Θ GPT ) represents the parameter set of a large language model (such as GPT), and (Γ) represents the hyperparameters of LoRA fine-tuning;

[0036] The specific form of the LoRA is: in is a low-rank matrix controlled by a hyperparameter (Γ);

[0037] The introduction of this low-rank matrix makes fine-tuning more flexible and can better adapt to the public health false information identification task while retaining the basic knowledge of the large language model. The hyperparameter (Γ) of LoRA can be designed as a matrix whose elements represent the adjustment amplitude of different model parameters. For each weight parameter (θ i ), corresponding to (γ i ) can control the amplitude of its fine-tuning, and can achieve targeted adjustment of the parameters of different parts of the model to better meet the needs of false information identification.

[0038] Furthermore, the specific process of deep learning training for a large language model is as follows:

[0039] S301. Pre-classification of data and reference of models:

[0040] S302: Define dynamic update strategy Where (M) is the existing model, (D new ) is the newly added data, (α) is the update rate parameter;

[0041] S303, obtain an updated model (M′) by fusing the new data set and the existing model parameters:

[0042] S304: Using the explainability scoring function (ε(M)) and the transparency scoring function To measure the model (M′), we get:

[0043] S305. Define the deep feature conversion function of the model Mapping the input text (x) to a high-dimensional feature space, the deep embedding representation of the knowledge graph (G = (V, E)) is By fusion function Combined to get:

[0044] S305. Define interactive modeling function Where (v i , v j ) are any two nodes in the knowledge graph, which are used to improve the semantic understanding and association analysis capabilities of the model in specific fields:

[0045] S305: Introducing semantic relationship modeling functions based on the fusion of knowledge graph and large language model Where (v i , v j ) represents a node in the knowledge graph, which is used to enhance the semantic understanding ability of the model in a specific field:

[0046] Furthermore, the specific process of data pre-classification and model reference is as follows:

[0047] S3011, divide the collected data into training data (D train ) and verification data (D val );

[0048] S3012, introduce training model (M train ) to train the model, expressed as: [M train =Train(D train )];

[0049] S3013, introduce verification model (M val ) Use the validation set to validate the model, expressed as: [Accuracy val =Validate(M train , D val )];

[0050] S3014. Introduce the data diversity index (Δ(D)) and the data representativeness index (Π(D)). The data diversity index (Δ(D)) is used to measure the coverage breadth of the dataset on multi-dimensional attributes, and the data representativeness index (Π(D)) is used to measure the applicability of the dataset in the target false information identification task, thereby obtaining the maximized dataset index: [maxΔ(D), maxΠ(D)].

[0051] Furthermore, the specific process of constructing the composite model is as follows:

[0052] S401, define a false information source identification function (F SI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, which is used to identify the origin of false information (S origin ): [Sorigin =F SI (x, Θ)]

[0053] S402. Define the propagation dynamics prediction model (F PD (x, Φ)), where (Φ) represents the model parameters, which is used to predict the propagation path and scope of false information in social networks, expressed as (P spread ): [P spread =F PD (x, Φ)]

[0054] S403, combined with epidemiological prediction model (F EP ) integrates the above two models to construct a composite model (M CI (F SI , F PD , F EP )) to analyze the comprehensive impact of false information dissemination on public health and the development of the epidemic (I CI ): [I CI =M CI (F SI (x, Θ), F PD (x,Φ), F EP (y,Ψ))], where (y) is epidemiological data.

[0055] Furthermore, the specific process of conducting empirical analysis and refining crisis response strategies based on the large language model is as follows:

[0056] S501. Use the large language model (M) to represent the crisis case as (Emb(C, M)), and construct a composite model (M CP ), expressed as (Emb(C,M CP ));

[0057] S502: Introducing utility function It is used to measure the effectiveness of response strategies, where (S) represents the crisis response strategy, (I PE ) represents the real information of the event, and the utility function can be expressed as: Where (T) represents the number of execution steps of the strategy, (S t ) represents the specific strategy of step (t), (α t ) and (β t ) is the weight coefficient for weighing effect and cost;

[0058] S502. The optimal crisis response strategy is extracted by solving the following optimization problem, which is expressed as:

[0059]

[0060] [subject to: Constraints on S].

[0061] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0062] 1. This composite model and construction method for identifying false public health information uses advanced natural language processing technology and large language models to enable the system to accurately and quickly identify false information and rumors from a large amount of information, which can improve the accuracy and speed of false information identification. It also includes a continuous learning and optimization mechanism, which can continuously adapt to new false information strategies, promote the advancement of related technologies and the updating of public health knowledge, and then intervene in false information in a timely manner, which will help maintain market and social order and reduce economic losses caused by panic.

[0063] 2. The composite model and construction method for identifying false public health information not only provides correct public health information by automatically generating content to refute false information, but also supports this information with scientific and authoritative data, thereby enhancing the persuasiveness of the information and helping to more effectively correct public misunderstandings. At the same time, by accurately and timely providing scientific health information and correcting false information, it helps to guide the public to adopt correct health behaviors and reduce the adverse consequences caused by believing in false health information. It can serve as a powerful tool for governments and health departments to strengthen communication with the public and improve the efficiency of emergency information dissemination, thereby effectively managing and alleviating crises. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The overall framework structure schematic diagram of the present invention is shown;

[0065] Figure 2 A schematic diagram of the process of identifying false information according to the present invention is shown. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Example:

[0068] like Figure 1-2As shown, a composite model for identifying false public health information and a construction method thereof, the composite model includes a data acquisition module, an information processing and analysis module, a false information identification module and a rumor refutation module, the data acquisition module is used to collect text data related to public health on an online platform, the information processing and analysis module preprocesses and extracts features from the acquired data, the false information identification module constructs a customized large language model and performs deep learning training, and the rumor refutation module is used to generate accurate and authoritative information to refute the identified false information, and the construction method specifically includes the following steps:

[0069] Step 1: Obtain the latest announcements and alerts through the official websites and social media of national and local public health departments, obtain a database of public health emergencies, and characterize the database to obtain a data set;

[0070] The database includes text data, audio data and video data, that is, [D = {d 1 , d 2 , ..., d n}], where (D) represents the collected data set, (d i ) represents a single data sample.

[0071] For each data sample (d i ), further describing its feature set: [F i ={f i1 , f i2 , ..., f im}], where (F i ) represents the data sample (d i ) feature set, (f ij ) represents the value of a specific feature.

[0072] The selection of data source is represented by a weight function, considering the importance of different data sources, which can be expressed as: [W s ={w 1 , w 2 , ..., w k}], where (W s ) represents the weights of different data sources, (w k ) represents the weight of the (k)th data source;

[0073] The obtained data set is: [D selected ={di|di∈D ,满足选择条件共同卫生相关的资讯}].

[0074] Step 2: Preprocess and extract features of the acquired data:

[0075] S101. Data preprocessing: clean the data and remove missing values ​​and outliers: [D cleaned ={di|di∈D selected,满足清洗条件}];

[0076] S102, add a label to each data sample: [L = {l 1 , l 2 , ..., l n}], where (L) represents the label set, i.e., real news, manually written news, and fake news generated by GPT, (l i ) represents the data sample (d i ) corresponding labels;

[0077] S103, the specific process of data classification is represented by the classification function: [C(d i )={c 1 , c 2 , ..., c k}], where (C(d i )) represents the data sample (d i ) classification results, (c k ) denotes category (k);

[0078] S104, obtain the data set for model training and testing: [D final ={(Fi,li)|di∈D cleaned ,Fi is the sample feature, Li is the corresponding label}]

[0079] Step 3: Build a large language model based on deep learning algorithms and conduct deep learning training;

[0080] The specific process of building a large language model is as follows:

[0081] S201, define a false information identification function (F DI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, (x) is converted into a feature vector (f(x)) by applying a high-level semantic mapping: [f(x) = F DI (x, Θ)], the mapping uses natural language processing techniques, including semantic understanding and context analysis;

[0082] S202, introduce false information loss function (L DI (f(x), y)), where (y) is the label, and the difference between the model prediction (f(x)) and the true label (y) is quantified by the false information loss function, that is, [L DI (f(x), y) = loss measure (f(x), y)];

[0083] S203, multi-dimensionally set the optimization goal of the large speech model, the optimization goal includes minimizing the false information loss function and maximize the accuracy of real information

[0084] S204, Introducing sensitivity indicators Used to focus on specific types of false information: Where (T) represents a specific type of false information category;

[0085] S205. Establish a large language model fine-tuning technology in the public health field based on Low-Rank Adaptation. LoRA is a powerful parameter adjustment strategy, especially suitable for large language models such as the GPT series. By introducing low-rank matrices, model parameters can be adjusted more flexibly to adapt to specific tasks. In the false information identification task, the LoRA fine-tuning strategy is defined as: Where (Θ GPT ) represents the parameter set of a large language model (such as GPT), and (Γ) represents the hyperparameters of LoRA fine-tuning;

[0086] The specific form of LoRA is: in is a low-rank matrix controlled by a hyperparameter (Γ);

[0087] The introduction of this low-rank matrix makes fine-tuning more flexible and can better adapt to the public health false information identification task while retaining the basic knowledge of the large language model. The hyperparameter (Γ) of LoRA can be designed as a matrix whose elements represent the adjustment amplitude of different model parameters. For each weight parameter (θ i ), corresponding to (γ i ) can control the amplitude of its fine-tuning, and can achieve targeted adjustment of the parameters of different parts of the model to better meet the needs of false information identification.

[0088] The specific process of deep learning training for a large language model is as follows:

[0089] S301. Pre-classification of data and reference of models:

[0090] S3011, divide the collected data into training data (D train ) and verification data (D val );

[0091] S3012, introduce training model (M train ) to train the model, expressed as: [M train =Train(D train )];

[0092] S3013, introduce verification model (M val ) Use the validation set to validate the model, expressed as: [Accuracy val =Validate(M train , D val )];

[0093] S3014. Introduce the data diversity index (Δ(D)) and the data representativeness index (Π(D)). The data diversity index (Δ(D)) is used to measure the coverage breadth of the dataset on multi-dimensional attributes, and the data representativeness index (Π(D)) is used to measure the applicability of the dataset in the target false information identification task, thereby obtaining the maximized dataset index: [maxΔ(D), maxΠ(D)];

[0094] S302: Define dynamic update strategy Where (M) is the existing model, (D new ) is the newly added data, (α) is the update rate parameter;

[0095] S303, obtain an updated model (M′) by fusing the new data set and the existing model parameters:

[0096] S304: Using the explainability scoring function (ε(M)) and the transparency scoring function To measure the model (M′), we get:

[0097] S305. Define the deep feature conversion function of the model Mapping the input text (x) to a high-dimensional feature space, the deep embedding representation of the knowledge graph (G = (V, E)) is By fusion function Combined to get:

[0098] S305. Define interactive modeling function Where (v i ,v j ) are any two nodes in the knowledge graph, which are used to improve the semantic understanding and association analysis capabilities of the model in specific fields:

[0099] S305: Introducing semantic relationship modeling functions based on the fusion of knowledge graph and large language model Where (v i ,v j ) represents a node in the knowledge graph, which is used to enhance the semantic understanding ability of the model in a specific field:

[0100] Step 4: Build a composite model based on the big voice model to integrate the spread of false information with epidemiological prediction;

[0101] The specific process of building a composite model is as follows:

[0102] S401, define a false information source identification function (F SI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, which is used to identify the origin of false information (S origin ): [S origin =F SI (x,Θ)]

[0103] S402. Define the propagation dynamics prediction model (F PD (x,Φ)), where (φ) represents the model parameters, which is used to predict the propagation path and scope of false information in social networks, expressed as (P spread ): [P spread =F PD (x, Φ)]

[0104] S403, combined with epidemiological prediction model (F EP ) integrates the above two models to construct a composite model (M CI (F SI , F PD , F EP )) to analyze the comprehensive impact of false information dissemination on public health and the development of the epidemic (I CI ): [I CI =M CI (F SI (x, Θ), F PD (x,Φ), F EP (y,Ψ))], where (y) is epidemiological data.

[0105] Step 5: Construct an evaluation index system based on the system performance of the large language model;

[0106] The above system performance evaluation index system includes accuracy (Acc), timeliness (Timeliness), and specificity (Specificity), and its evaluation function is expressed as (E sys (F DI , M CP , Θ, Ψ)):

[0107] [E sys ={Acc(F DI ), Timeliness(M CP ), Specificity(F DI )}];

[0108] Step 6: Conduct empirical analysis and refine crisis response strategies based on the large language model;

[0109] The specific process of conducting empirical analysis and refining crisis response strategies based on the large language model is as follows:

[0110] S501. Use the large language model (M) to represent the crisis case as (Emb(C, M)), and construct a composite model (M CP ), expressed as (Emb(C,M CP ));

[0111] S502: Introducing utility function It is used to measure the effectiveness of response strategies, where (S) represents the crisis response strategy, (I PE ) represents the real information of the event, and the utility function can be expressed as: Where (T) represents the number of execution steps of the strategy, (S t ) represents the specific strategy of step (t), (α t ) and (β t ) is the weight coefficient for weighing effect and cost;

[0112] S502. The optimal crisis response strategy is extracted by solving the following optimization problem, which is expressed as:

[0113]

[0114] [subjet to: Constraints on S].

[0115] Step 7: Establish a real-time monitoring system to monitor public health information. At the same time, use the constructed detection algorithm to detect false information in real-time monitoring. When false information is detected, immediately execute the response mechanism: [Response (S monitor , Detection(D, A))].

[0116] The above formulas are all dimensionless and numerical calculations. The formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0117] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A composite model and construction method for identifying false public health information, characterized in that: The composite model includes a data collection module, an information processing and analysis module, a false information identification module and a rumor refutation module. The data collection module is used to collect text data related to public health on an online platform. The information processing and analysis module performs preprocessing and feature extraction on the acquired data. The false information identification module builds a customized large language model and performs deep learning training. The rumor refutation module is used to generate accurate and authoritative information to refute the identified false information. The construction method specifically includes the following steps: Step 1: Obtain the latest announcements and alerts through the official websites and social media of national and local public health departments, obtain a database of public health emergencies, and characterize the database to obtain a data set; Step 2: preprocessing and feature extraction of the acquired data; Step 3: Build a large language model based on deep learning algorithms and conduct deep learning training; Step 4: Build a composite model based on the big voice model to integrate the spread of false information with epidemiological prediction; Step 5: Construct an evaluation index system based on the system performance of the large language model; Step 6: Conduct empirical analysis and refine crisis response strategies based on the large language model; Step 7: Establish a real-time monitoring system to monitor public health information. At the same time, use the constructed detection algorithm to detect false information in real-time monitoring. When false information is detected, immediately execute the response mechanism: [Response (S monitor , Detection(D, A))].

2. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The database includes text data, audio data and video data, that is, [D = {d1, d2, ..., d n }], where (D) represents the collected data set, (d i ) represents a single data sample. For each data sample (d i ), further describing its feature set: [F i ={f i1 , f i2 , ..., f im }], where (F i ) represents the data sample (d i ) feature set, (f ij ) represents the value of a specific feature. The selection of data source is represented by a weight function, considering the importance of different data sources, which can be expressed as: [W s ={w1, w2, ..., w k }], where (W s ) represents the weights of different data sources, (w k ) represents the weight of the (k)th data source; The data set obtained is: selected ={di|di∈D, 满足选择条件共同卫生 相关的资讯 }].

3. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The specific process of preprocessing and feature extraction of the acquired data is as follows: S101. Data preprocessing: clean the data and remove missing values ​​and outliers: [D cleaned ={di|di∈D selected，满足清洗条件 }]; S102, label each data sample: [L = {l1, l2, ..., l n }], where (L) represents the label set, i.e., real news, manually written news, and fake news generated by GPT, (l i ) represents the data sample (d i ) corresponding labels; S103, the specific process of data classification is represented by the classification function: [C(d i )={c1,c2,...,c k }], where (C(d i )) represents the data sample (d i ) classification results, (c k ) denotes category (k); S104, obtain the data set for model training and testing: [D final ={(Fi,li)|di∈D cleaned , Fi is the sample feature, Li is the corresponding label}].

4. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The specific process of building a large language model is as follows: S201, define a false information identification function (F DI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, (x) is converted into a feature vector (f(x)) by applying a high-level semantic mapping: [f(x) = F DI (x, Θ)]; S202, introduce false information loss function (L DI (f(x), y)), where (y) is the label, and the difference between the model prediction (f(x)) and the true label (y) is quantified by the false information loss function, that is, [L DI (f(x), y) = loss measure (f(x), y)]; S203, multi-dimensionally setting the optimization goal of the large speech model, wherein the optimization goal includes minimizing the false information loss function and maximize the accuracy of real information S204, Introducing sensitivity indicators Used to focus on specific types of false information: Where (T) represents a specific type of false information category; S205. Establish a large language model fine-tuning technology in the public health field based on Low-Rank Adaptation. LoRA is a powerful parameter adjustment strategy, especially suitable for large language models such as the GPT series. By introducing low-rank matrices, model parameters can be adjusted more flexibly to adapt to specific tasks. In the false information identification task, the LoRA fine-tuning strategy is defined as: Where (Θ GPT ) represents the parameter set of a large language model (such as GPT), and (Γ) represents the hyperparameters of LoRA fine-tuning; The specific form of the LoRA is: in is a low-rank matrix controlled by a hyperparameter (Γ).

5. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The specific process of deep learning training for a large language model is as follows: S301. Pre-classification of data and reference of models: S302: Define dynamic update strategy Where (M) is the existing model, (D new ) is the newly added data, (α) is the update rate parameter; S303, obtain an updated model (M′) by fusing the new data set and the existing model parameters: S304: Using the explainability scoring function (ε(M)) and the transparency scoring function To measure the model (M′), we get: S305. Define the deep feature conversion function of the model Mapping the input text (x) to a high-dimensional feature space, the deep embedding representation of the knowledge graph (G = (V, E)) is By fusion function Combined to get: S305. Define interactive modeling function Where (v i , v j ) are any two nodes in the knowledge graph, which are used to improve the semantic understanding and association analysis capabilities of the model in specific fields: S305: Introducing semantic relationship modeling functions based on the fusion of knowledge graph and large language model Where (v i , v j ) represents a node in the knowledge graph, which is used to enhance the semantic understanding ability of the model in a specific field:

6. The composite model and construction method for identifying false public health information according to claim 5, characterized in that: The specific process of data pre-classification and model reference is as follows: S3011, divide the collected data into training data (D train ) and verification data (D val ); S3012, introduce training model (M train ) to train the model, expressed as: [M train =Train(D train )]; S3013, introduce verification model (M val ) Use the validation set to validate the model, expressed as: [Accuracy val =Validate(M train , D val )]; S3014. Introduce the data diversity index (Δ(D)) and the data representativeness index (Π(D)). The data diversity index (Δ(D)) is used to measure the coverage breadth of the dataset on multi-dimensional attributes, and the data representativeness index (Π(D)) is used to measure the applicability of the dataset in the target false information identification task, thereby obtaining the maximized dataset index: [maxΔ(D), maxΠ(D)].

7. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The specific process of building a composite model is as follows: S401, define a false information source identification function (F SI (x, Θ)), where (x) is the input data and (Θ) is the model parameter, which is used to identify the origin of false information (S origin ): [S origin =F SI (x, Θ)] S402. Define the propagation dynamics prediction model (F PD (x, Φ)), where (Φ) represents the model parameters, which is used to predict the propagation path and scope of false information in social networks, expressed as (P spread ): [P spread =F PD (x, Φ)] S403, combined with epidemiological prediction model (F EP ) integrates the above two models to construct a composite model (M CI (F SI , F PD , F EP )) to analyze the comprehensive impact of false information dissemination on public health and the development of the epidemic (I CI ): [I CI =M CI (F SI (x, Θ), F PD (x,Φ), F EP (y, ψ))], where (y) is epidemiological data.

8. The composite model and construction method for identifying false public health information according to claim 1, characterized in that: The specific process of conducting empirical analysis and refining crisis response strategies based on the large language model is as follows: S501. Use the large language model (M) to represent the crisis case as (Emb(C,M)), and construct a composite model (M CP ), expressed as (Emb(C,M CP )); S502: Introducing utility function It is used to measure the effectiveness of response strategies, where (S) represents the crisis response strategy, (I PE ) represents the real information of the event, and the utility function can be expressed as: Where (T) represents the number of execution steps of the strategy, (S t ) represents the specific strategy of step (t), (α t ) and (β t ) is the weight coefficient for weighing effect and cost; S502. The optimal crisis response strategy is extracted by solving the following optimization problem, which is expressed as: [subject to: Constraints on S].

Images