False information recognition language model based on large language model

Abstract

The invention discloses a false information recognition language model based on a large language model, and relates to the technical field of information recognition, and the false information recognition language model comprises a data acquisition unit, a knowledge graph construction unit, an interactive modeling unit and a model optimization unit. In order to enhance the understanding and judgment ability of the model to complex false information scenes, and in order to improve the professional adaptability of the model in the public health field, the knowledge graph construction unit is used to obtain professional knowledge of each field to construct a field specific corpus, and a knowledge graph is integrated to provide richer and more accurate public health related information. Through the measures, the model can capture complex narration related to public health and subtle changes in the context, so that false information is identified more accurately, and powerful support is provided for public health crisis emergency and risk communication.

Description

Technical Field

[0001] The present invention relates to the technical field of information recognition, and in particular to a false information recognition language model based on a large language model. Background Art

[0002] With the popularization of the Internet and communication technology, the network information environment has become complex, causing a large amount of information to flood the network. Therefore, the identification of false information is particularly important. In the invention patent with application number 202410575897.4, "a training method, device, equipment and storage medium for a false information identification model are disclosed; the method includes: using a false information identification model to identify false information on short-term noise samples, and obtaining corresponding identification results; based on the sample label and the identification result, determining the loss value of the first loss function; using the false information identification model to identify false information on long-term noise samples, and obtaining corresponding identification results; based on the sample label and the identification result, determining the first loss value of the second loss function; based on the loss value of the first loss function and the first loss value of the second loss function, updating the model parameters of the false information identification model. Through this application, the accuracy and recall rate of the false information identification model can be improved.";

[0003] The above-mentioned existing technologies solve the problem that limited false data can easily cause model overfitting and other problems. However, the model cannot effectively integrate the expertise in the field of public health into the large language model during operation, which reduces the accuracy and reliability of false information identification and intervention. Especially when it comes to multicultural and multilingual environments, the model may not be able to accurately identify and interpret public health professional terms or regional language features. At the same time, rule-based systems and traditional machine learning models often rely on fixed data sets and predefined rules. This method is difficult to adapt to the rapid changes in false information content and strategies. Summary of the invention

[0004] The purpose of the present invention is to provide a false information identification language model based on a large language model to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a false information identification language model based on a large language model;

[0006] A data collection unit, wherein the data collection unit stores the collected data samples into a data set, sets a corresponding weight for each data sample, cleans the data samples in the data set, and stores them into a test set and a training set respectively;

[0007] A knowledge graph construction unit, wherein the knowledge graph construction unit determines a plurality of attributes corresponding to each knowledge data, analyzes the data diversity index and the data representativeness index according to the data and the corresponding attributes, stores the data in the corresponding domain-specific corpus according to the involved domains, and uses the TransE model to construct a knowledge graph according to the data in the domain-specific corpus;

[0008] An interactive modeling unit, wherein the interactive modeling unit sequentially transmits the data in the training set to the fake information identification model, combines the received input text and the deep embedded representation of the knowledge graph using a fusion function, generates a corresponding interactive modeling function according to the nodes in the knowledge graph, and transmits the interactive modeling function to the model;

[0009] The model optimization unit converts the input text data into a feature vector, analyzes it using a false information recognition model, adjusts each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy through a parameter fine-tuning algorithm, completes model training, and analyzes the model prediction results and actual results according to the evaluation function to determine the performance of the model.

[0010] Preferably, the data collection unit includes a sample extraction module and a weight setting module. The sample extraction module uses a crawler tool to extract public health related data from various information platforms, and stores it as multiple data samples in a data set. The weight setting module extracts features of all data samples, stores them in a feature set, and sets corresponding weights for each data sample according to the feature set.

[0011] Preferably, the data acquisition unit also includes a data partitioning module, which cleans the data samples in the data set, removes missing values ​​and outliers, generates corresponding labels according to the source and characteristics of each data sample, stores 20% of the data samples in the data set in the test set, and stores the remaining data samples in the training set.

[0012] Preferably, the knowledge graph construction unit includes an indicator analysis module, which obtains knowledge data from the professional knowledge database according to the set collection time, determines multiple attributes corresponding to each knowledge data, and analyzes the data diversity index and the data representativeness index according to the data and the corresponding attributes.

[0013] Preferably, the knowledge graph construction unit also includes a data storage module and an entity recognition module. If the two indicators of the data storage module reach the set thresholds, the acquired knowledge data will be stored in the corresponding domain-specific corpus according to the involved fields. Otherwise, the knowledge data will be re-acquired. The entity recognition module uses TexSmart software to identify entities and the relationships between entities in the domain-specific corpus, and uses the TransE model to construct a knowledge graph based on the entities and relationships.

[0014] Preferably, the interactive modeling unit includes a model building module and a function definition module. The model building module uses deep learning technology to build a false information recognition model, and transmits the data in the training set to the false information recognition model in sequence. The function definition module defines the deep feature conversion function in the false information recognition model. After receiving the input text, the input text is mapped to the corresponding high-dimensional feature space using the deep feature conversion function, and the deep embedding representation of the knowledge graph is determined. The two are combined using a fusion function, and the fusion function is specifically:

[0015] F comlined (x,G)=C(J LLM (x; Θ LLM ),ε kg (G; θ KG ))

[0016] Among them, F comlined (x,G) represents the fusion result of input text and knowledge graph, J LLM (x; Θ LLM ) represents the deep feature conversion function corresponding to the input text x, ε kg (G; θ KG ) represents the deep embedding representation corresponding to the knowledge graph G, and C(.,.) represents the fusion function.

[0017] Preferably, the interactive modeling unit also includes a node selection module, which selects any two nodes in the knowledge graph, defines corresponding interactive modeling functions according to the node data, repeats the operation until each node has an interactive modeling function with other nodes, and transmits the interactive modeling function to the model.

[0018] Preferably, the model optimization unit includes a sample analysis module, a parameter update module and a result evaluation module. The sample analysis module converts the input text data into a feature vector, and then analyzes it using a false information recognition model to determine the optimization target and sensitivity index, wherein the optimization target includes minimizing the false information loss function and maximizing the accuracy of true information. The parameter update module uses a parameter fine-tuning algorithm to adjust each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy, and then updates these parameters in the model. The result evaluation module uses the false information recognition model to analyze all data samples in the training set to determine the model parameters. After completing the model training, the data samples in the test data set are input into the model, and the model prediction results and actual results are analyzed according to the evaluation function to determine the accuracy, timeliness and specificity corresponding to the model. The parameter fine-tuning algorithm is specifically as follows:

[0019]

[0020] in, represents the adjusted model parameter set, Θ GPT represents the existing model parameter set, Γ represents the hyperparameters of LoRA fine-tuning, and R(Γ) represents the low-rank matrix.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The present invention uses an interactive modeling unit to deeply process and analyze text, images and social media data to enhance the model's understanding and judgment capabilities of complex false information scenarios. In order to improve the professional adaptability of the model in the field of public health, a knowledge graph construction unit is used to obtain professional knowledge in various fields to construct a domain-specific corpus, and a knowledge graph is integrated to provide richer and more accurate public health-related information. Through these measures, the model can capture subtle changes in complex narratives and contexts related to public health, thereby more accurately identifying false information and providing strong support for public health crisis emergency response and risk communication. In the model optimization stage, the model optimization unit is used to learn from new data and update the model's parameters to gradually improve the performance and accuracy of the model, making it more adaptable to the needs of the public health field. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Provides a schematic diagram of the overall system flow for an embodiment of the present invention;

[0024] Figure 2 An internal module block diagram of a data acquisition unit provided by an embodiment of the present invention;

[0025] Figure 3An internal module block diagram of a knowledge graph construction unit provided in an embodiment of the present invention;

[0026] Figure 4 An internal module block diagram of an interactive modeling unit provided in an embodiment of the present invention;

[0027] Figure 5 This is a block diagram of the internal modules of the model optimization unit provided in an embodiment of the present invention.

[0028] In the figure: 1. Data acquisition unit; 101. Sample extraction module; 102. Weight setting module; 103. Data partitioning module; 2. Knowledge graph construction unit; 201. Index analysis module; 202. Data storage module; 203. Entity recognition module; 3. Interaction modeling unit; 301. Model construction module; 302. Function definition module; 303. Node selection module; 4. Model optimization unit; 401. Sample analysis module; 402. Parameter update module; 403. Result evaluation module. DETAILED DESCRIPTION

[0029] 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.

[0030] See also Figure 1-5 ,The present invention provides a technical solution: a false information identification language model based on a large language model;

[0031] Data collection unit 1, data collection unit 1 stores the collected data samples into a data set, sets a corresponding weight for each data sample, cleans the data samples in the data set and stores them into a test set and a training set respectively;

[0032] The knowledge graph construction unit 2 determines multiple attributes corresponding to each knowledge data, analyzes the data diversity index and the data representativeness index according to the data and the corresponding attributes, stores the data in the corresponding domain-specific corpus according to the involved fields, and uses the TransE model to construct a knowledge graph according to the data in the domain-specific corpus;

[0033] Interaction modeling unit 3, which sequentially transmits the data in the training set to the fake information identification model, combines the received input text and the deep embedding representation of the knowledge graph using a fusion function, generates a corresponding interaction modeling function according to the nodes in the knowledge graph, and transmits the interaction modeling function to the model;

[0034] Model optimization unit 4, after converting the input text data into a feature vector, uses the false information recognition model to analyze it, adjusts each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy through the parameter fine-tuning algorithm, completes the model training, analyzes the model prediction results and actual results according to the evaluation function, and determines the performance of the model.

[0035] The data collection unit 1 includes a sample extraction module 101 and a weight setting module 102. The sample extraction module 101 uses a crawler tool to extract public health related data from various information platforms and stores them as multiple data samples in a data set. The weight setting module 102 extracts features of all data samples, stores them in a feature set, and sets a corresponding weight for each data sample according to the feature set.

[0036] The data collection unit 1 also includes a data partitioning module 103, which cleans the data samples in the data set, removes missing values ​​and outliers, generates corresponding labels according to the source and characteristics of each data sample, stores 20% of the data samples in the data set into the test set, and stores the remaining data samples into the training set;

[0037] The knowledge graph construction unit 2 includes an indicator analysis module 201. After acquiring knowledge data from the professional knowledge database according to the set collection time, the indicator analysis module 201 determines multiple attributes corresponding to each knowledge data, and analyzes the data diversity index and the data representativeness index according to the data and the corresponding attributes;

[0038] The knowledge graph construction unit 2 also includes a data storage module 202 and an entity recognition module 203. If the two indicators of the data storage module 202 reach the set threshold, the acquired knowledge data is stored in the corresponding domain-specific corpus according to the involved domains. Otherwise, the knowledge data is acquired again. The entity recognition module 203 uses TexSmart software to identify entities and relationships between entities in the domain-specific corpus, and uses the TransE model to construct a knowledge graph based on the entities and relationships.

[0039] The interactive modeling unit 3 includes a model building module 301 and a function definition module 302. The model building module 301 uses deep learning technology to build a false information recognition model, and transmits the data in the training set to the false information recognition model in sequence. The function definition module 302 defines the deep feature conversion function in the false information recognition model. After receiving the input text, the deep feature conversion function is used to map the input text to the corresponding high-dimensional feature space, and the deep embedding representation of the knowledge graph is determined. The fusion function is used to combine the two. The fusion function is specifically:

[0040] F comlined (x,G)=C(J LLM (x; Θ LLM ),ε kg (G; θ KG ))

[0041] Among them, F comlined (x, G) represents the fusion result of input text and knowledge graph, J LLM (x; Θ LLM ) represents the deep feature conversion function corresponding to the input text x, ε kg (G; θ KG ) represents the deep embedding representation corresponding to the knowledge graph G, and C(.,.) represents the fusion function;

[0042] The interactive modeling unit 3 also includes a node selection module 303. The node selection module 303 selects two nodes in the knowledge graph at random, defines the corresponding interactive modeling function according to the node data, repeats the operation until each node has an interactive modeling function with other nodes, and transmits the interactive modeling function to the model. The interactive modeling function is specifically:

[0043]

[0044] Among them, v i and v j represents any two nodes in the knowledge graph, E represents the node set in the knowledge graph, Θ Sem Represents the relationship between nodes, R semantic and J S em(vi,vj;Θ S em) all represent interaction modeling functions;

[0045] The model optimization unit 4 includes a sample analysis module 401, a parameter updating module 402 and a result evaluation module 403. The sample analysis module 401 converts the input text data into a feature vector, and then analyzes it using a false information recognition model to determine the optimization target and sensitivity index, wherein the optimization target includes minimizing the false information loss function and maximizing the accuracy of true information. The parameter updating module 402 uses a parameter fine-tuning algorithm to adjust each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy, and then updates these parameters in the model. The result evaluation module 403 uses the false information recognition model to analyze all data samples in the training set to determine the model parameters. After completing the model training, the data samples in the test data set are input into the model, and the model prediction results and actual results are analyzed according to the evaluation function to determine the accuracy, timeliness and specificity corresponding to the model. The parameter fine-tuning algorithm is specifically as follows:

[0046]

[0047] in, represents the adjusted model parameter set, Θ GPT represents the existing model parameter set, Γ represents the hyperparameters of LoRA fine-tuning, and R(Γ) represents the low-rank matrix.

[0048] Working principle: The present invention obtains data samples through the sample extraction module 101 in the data acquisition unit 1, the weight setting module 102 sets the corresponding weight for each sample, and the data division module 103 divides all data samples and stores them in the training set and the test set respectively. The indicator analysis module 201 in the knowledge graph construction unit 2 extracts knowledge data and determines the attributes of each knowledge data. The data storage module 202 stores the acquired knowledge data in the corresponding field-specific corpus according to the involved fields. The entity recognition module 203 constructs a knowledge graph based on these data. The model construction module 301 in the interactive modeling unit 3 uses the false information recognition model to analyze the data in the training set. The function definition module 302 combines the input text with the knowledge graph. The node selection module 303 defines the corresponding interactive modeling function for each node in the knowledge graph. The sample analysis module 401 in the model optimization unit 4 determines the optimization target and sensitivity index corresponding to the model. The parameter update module 402 adjusts each parameter in the model. The result evaluation module 403 analyzes the performance of the current model.

[0049] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0050] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A false information identification language model based on a large language model, characterized by: A data collection unit (1), wherein the data collection unit (1) stores the collected data samples in a data set, sets a corresponding weight for each data sample, cleans the data samples in the data set, and stores them in a test set and a training set respectively; A knowledge graph construction unit (2), wherein the knowledge graph construction unit (2) determines a plurality of attributes corresponding to each knowledge data, analyzes a data diversity index and a data representativeness index according to the data and the corresponding attributes, stores the data in a corresponding domain-specific corpus according to the domain involved, and constructs a knowledge graph based on the data in the domain-specific corpus using a TransE model; An interactive modeling unit (3), wherein the interactive modeling unit (3) sequentially transmits the data in the training set to the false information identification model, combines the received input text and the deep embedding representation of the knowledge graph using a fusion function, generates a corresponding interactive modeling function according to the nodes in the knowledge graph, and transmits the interactive modeling function to the model; The model optimization unit (4) converts the input text data into a feature vector, analyzes the feature vector using a false information recognition model, adjusts each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy through a parameter fine-tuning algorithm, completes model training, and analyzes the model prediction results and actual results according to an evaluation function to determine the performance of the model.

2. The false information identification language model based on a large language model according to claim 1, characterized in that: The data collection unit (1) comprises a sample extraction module (101) and a weight setting module (102). The sample extraction module (101) uses a crawler tool to extract public health related data from various information platforms and stores the data as multiple data samples in a data set. The weight setting module (102) extracts features of all data samples, stores them in a feature set, and sets a corresponding weight for each data sample according to the feature set.

3. The false information identification language model based on a large language model according to claim 2, characterized in that: The data collection unit (1) also includes a data partitioning module (103), which cleans the data samples in the data set, removes missing values ​​and outliers, generates corresponding labels according to the source and characteristics of each data sample, stores 20% of the data samples in the data set in the test set, and stores the remaining data samples in the training set.

4. The false information identification language model based on a large language model according to claim 1, characterized in that: The knowledge graph construction unit (2) comprises an indicator analysis module (201), which acquires knowledge data from a professional knowledge database according to a set collection time, determines a plurality of attributes corresponding to each knowledge data, and analyzes data diversity indicators and data representativeness indicators according to the data and the corresponding attributes.

5. The false information identification language model based on a large language model according to claim 4, characterized in that: The knowledge graph construction unit (2) also includes a data storage module (202) and an entity recognition module (203). If the two indicators of the data storage module (202) reach the set thresholds, the acquired knowledge data is stored in the corresponding domain-specific corpus according to the involved domains. Otherwise, the knowledge data is re-acquired. The entity recognition module (203) uses TexSmart software to identify entities and relationships between entities in the domain-specific corpus, and uses the TransE model to construct a knowledge graph based on the entities and relationships.

6. The false information identification language model based on a large language model according to claim 1, characterized in that: The interactive modeling unit (3) comprises a model construction module (301) and a function definition module (302). The model construction module (301) uses deep learning technology to construct a false information recognition model, and transmits the data in the training set to the false information recognition model in sequence. The function definition module (302) defines a deep feature conversion function in the false information recognition model, and after receiving the input text, uses the deep feature conversion function to map the input text to the corresponding high-dimensional feature space, and determines the deep embedding representation of the knowledge graph, and combines the two using a fusion function.

7. The false information identification language model based on a large language model according to claim 6, characterized in that: The interactive modeling unit (3) also includes a node selection module (303), which selects two nodes in the knowledge graph at random, defines corresponding interactive modeling functions according to the node data, repeats the operation until each node has an interactive modeling function with other nodes, and transmits the interactive modeling function to the model.

8. The false information identification language model based on a large language model according to claim 1, characterized in that: The model optimization unit (4) includes a sample analysis module (401), a parameter updating module (402) and a result evaluation module (403). The sample analysis module (401) converts the input text data into a feature vector, and then analyzes the feature vector using a false information recognition model to determine an optimization target and a sensitivity index, wherein the optimization target includes minimizing the false information loss function and maximizing the accuracy of true information. The parameter updating module (402) uses a parameter fine-tuning algorithm to adjust each parameter in the false information recognition model according to the original parameter set of the model and the hyperparameters of the LoRA fine-tuning strategy, and then updates these parameters in the model. The result evaluation module (403) uses the false information recognition model to analyze all data samples in the training set to determine the model parameters. After completing the model training, the data samples in the test data set are input into the model, and the model prediction results and actual results are analyzed according to the evaluation function to determine the accuracy, timeliness and specificity of the model.

Images