Automatic detection and control method and system for false information and comments based on large models and fine-tuning

Abstract

The present invention discloses a method and system for automatic detection and control of false information and comments based on large models and fine-tuning. The method improves the accuracy of false information detection through multimodal feature fusion and low-rank adaptive fine-tuning technology, and effectively identifies complex false information with implicit logical contradictions or factual conflicts by combining text context features, image semantic information and external knowledge bases; it adopts low-rank matrix fine-tuning method to optimize the parameters of large language models, while retaining general capabilities and enhancing the ability to distinguish Chinese false information. The present invention achieves rapid response throughout the entire process through an automated processing mechanism, significantly reducing the cost of manual review, and the modular design supports flexible expansion to different social media platforms. It dynamically labels risk levels, and achieves local shielding and dissemination containment by simulating user behavior for platforms without public interfaces. The overall solution achieves technological breakthroughs in detection accuracy, processing efficiency and system scalability, providing support for large-scale user-generated content management.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and content management technology, and in particular relates to a method and system for automatically detecting and controlling false information and comments based on a large model and fine-tuning. Background Art

[0002] With the surge in user numbers and interactions on social media platforms, these platforms have become crucial sources of information in our daily lives. However, the emergence of false information on these platforms is inevitable. This false information not only severely impacts user experience but can also negatively impact the platform's reputation and even lead to legal risks. Therefore, effectively identifying and blocking this false information is a major challenge facing this method as it helps users access authentic content on social platforms.

[0003] Traditional methods for detecting disinformation rely primarily on manually defined rules or pattern matching, manual review, and machine learning models. However, each of these methods has limitations. While manually defined rules or pattern matching methods can detect some disinformation, they are unable to effectively identify subtle forms of disinformation and newer variants. While manual review can provide a certain degree of accuracy, the surge in content on platforms has led to increasing efficiency and cost issues. Furthermore, existing machine learning models have low accuracy when detecting disinformation with complex sentiment or targeted at specific scenarios, and are prone to false positives or omissions.

[0004] In recent years, with the rapid development of natural language processing (NLP) and deep learning technologies, the application of large-scale pre-trained language models (such as LLaMA) in sentiment analysis and text classification has greatly improved the accuracy of malicious content detection. These technologies provide a deeper understanding of the context and sentiment of text, enabling more accurate identification of false or partially false information. However, how to effectively apply these technologies to content filtering on social media platforms remains an urgent challenge. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes an automatic detection and control method and system for false information and comments based on a large model and fine-tuning to solve the problems existing in the above-mentioned prior art.

[0006] In a first aspect, to achieve the above objectives, the present invention provides a method for automatically detecting and controlling false information and comments based on a large model and fine-tuning, comprising the following steps:

[0007] Acquiring information data from a network platform, wherein the information data includes text, images, and metadata;

[0008] Preprocessing the information data includes text cleaning, extracting text content and image embedded text, and converting the data into structured data and storing it in a database;

[0009] The preprocessed data is input into a fine-tuned large language model, and the false probability of the information data is calculated by fusing text features with image features;

[0010] Perform automated processing on information determined to be false based on the probability of falsehood, including blocking, reporting, or labeling;

[0011] The fine-tuning process of the large language model includes: inserting a low-rank matrix into the query matrix and value matrix of the large language model, and performing supervised training in combination with the labeled dataset.

[0012] Optionally, in the process of acquiring information data from the network platform, a web crawler is used to acquire data, the process including: writing parsing code according to the page structure of the target platform, and regularly executing crawling tasks to acquire posts, comments, and associated metadata;

[0013] The crawled data is classified by platform and stored in JSONL format. The metadata includes the author, release time and source platform identifier.

[0014] Optionally, the preprocessing process includes:

[0015] Use regular expressions to clean illegal characters and HTML tags in text;

[0016] Parse the data to extract text content, author, publication date, and image information;

[0017] Convert the parsed text, image metadata, and cleaned content into structured data and store it in a relational database.

[0018] Optionally, the fine-tuning process of the large language model includes:

[0019] Insert the low-rank matrix into the query matrix and value matrix of the model respectively and ,in , A, B, C, and D are all low-rank adaptation matrices inserted by LoRA, which are used to optimize the original query matrix and numerical matrix in an incremental form;

[0020] The cross-entropy loss function is used for supervised training on a labeled dataset that is manually annotated to distinguish false information from true information.

[0021] Optionally, the process of fusing the text features with the image features includes:

[0022] Extracting text context features and correlating them with entity descriptions retrieved from Internet search engines;

[0023] Generate image description titles and extract embedded text in images;

[0024] Text features, image description titles and embedded text are concatenated into input text, and preset judgment prompt words are added.

[0025] Optionally, the process of performing an automated processing operation on the information determined to be false according to the false probability includes:

[0026] For platforms that support public APIs, submit API requests to report or block the platform;

[0027] For platforms that do not support APIs, mark or hide operations are performed by simulating user behavior, including disguising request headers and switching proxy IPs.

[0028] In a second aspect, the present invention further provides an automatic detection and control system for false information and comments based on a large model and fine-tuning, for implementing an automatic detection and control method for false information and comments based on a large model and fine-tuning, the system comprising:

[0029] A data collection module, which is used to periodically capture user-generated content data from multiple platforms through a web crawler. The data includes text, images, and metadata.

[0030] A preprocessing module is used to clean the data, remove illegal characters and HTML tags, extract text content and image embedded text, and convert the data into a structured format and store it in a database;

[0031] a model processing module, comprising a fine-tuned large language model for fusing text features with image features and calculating the false probability of the information, wherein a low-rank adaptation matrix is ​​inserted into the query matrix and the value matrix of the model;

[0032] A false information determination module, used to generate a determination result based on the false probability;

[0033] An automated control module is used to block, report, or mark information that is determined to be false.

[0034] Optionally, the data acquisition module includes:

[0035] The crawler configuration unit is used to write parsing code based on the page structure of the target platform and define crawling parameters;

[0036] The scheduled execution unit is used to periodically trigger crawler tasks, obtain posts, comments, and related metadata, and store them in JSONL format according to platform classification.

[0037] In a third aspect, the present invention further provides a computer terminal device, comprising:

[0038] one or more processors;

[0039] a memory, coupled to the processor, for storing one or more programs;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement an automatic detection and control method for false information and comments based on large models and fine-tuning.

[0041] In a fourth aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for automatically detecting and controlling false information and comments based on a large model and fine-tuning.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] The present invention provides a method and system for automatically detecting and controlling false information and comments based on a large model and fine-tuning. The present invention improves the accuracy of false information detection through multimodal feature fusion and low-rank adaptive fine-tuning technology. It combines text context features, image semantic information, and external knowledge bases to effectively identify complex false information that implies logical contradictions or factual conflicts. It also uses a low-rank matrix fine-tuning method to optimize the parameters of the large language model, enhancing the ability to discriminate against Chinese false information while retaining general capabilities. The automated processing mechanism enables rapid response throughout the entire process, significantly reducing manual review costs, and the modular design supports flexible expansion to different social media platforms. The front-end interface dynamically labels risk levels and provides batch operation capabilities. For platforms without public interfaces, local blocking and dissemination containment are achieved by simulating user behavior. The overall solution achieves technological breakthroughs in detection accuracy, processing efficiency, and system scalability, providing reliable support for large-scale user-generated content management. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings, which constitute part of the present invention, are provided to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are provided to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0045] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0046] Figure 2 A method flow chart of an embodiment of the present invention;

[0047] Figure 3This is an architecture diagram of an automatic detection and control system for false media information based on front-end and back-end technologies according to an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of the LoRA fine-tuning process based on the LLaMA large model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides an automatic detection and control method for false information and comments based on a large model and fine-tuning, including:

[0053] Acquiring information data from a network platform, wherein the information data includes text, images, and metadata;

[0054] Preprocessing the information data includes text cleaning, extracting text content and image embedded text, and converting the data into structured data and storing it in a database;

[0055] The preprocessed data is input into a fine-tuned large language model, and the false probability of the information data is calculated by fusing text features with image features;

[0056] Perform automated processing on information determined to be false based on the probability of falsehood, including blocking, reporting, or labeling;

[0057] The fine-tuning process of the large language model includes: inserting a low-rank matrix into the query matrix and value matrix of the large language model, and performing supervised training in combination with the labeled dataset.

[0058] like Figure 2 As shown, it specifically includes: S1: using crawler technology to automatically crawl historical information and comment data of multiple platforms in batches at regular intervals;

[0059] S2: Preprocess the crawled information data and fine-tune the large language model;

[0060] S3: Process the information and analyze it using a fine-tuned large language model to determine its authenticity and return the probability that the information is false.

[0061] S4: Automated processing of information determined to be false, including blocking, reporting, or marking.

[0062] Regarding the above S1 to S4, further comprising:

[0063] S1: Through the front-end and back-end Vue+Springboot+MySQL technology, use crawler technology to automatically crawl information and comments on various online social platforms. After configuring the corresponding parameters, crawl information such as: post text content, post image content, post time range, and post publisher;

[0064] S2: After obtaining the relevant information data, the data is organized into a unified JSONL format, cleaned, and illegal characters and empty characters are removed before the data is stored in a relational database.

[0065] S3: Connects to the database, retrieves stored user data, and takes data such as article ID, author, content, comment ID, author, and content as input. It uses a pre-trained large language model to determine the healthiness of the user and returns a health score. The health score is then updated to the database using the comment ID.

[0066] S4: Through front-end and back-end technologies, unhealthy comments with health scores below the specified standard are displayed. Through web crawlers, comments are located according to article and comment IDs and deleted or hidden. Users can select malicious comments in batches for control.

[0067] As an implementation method of this embodiment, in the process of obtaining information data from the network platform, a web crawler is used to obtain data. The process includes: writing parsing code according to the page structure of the target platform, and regularly executing crawling tasks to obtain posts, comments and related metadata;

[0068] The crawled data is classified by platform and stored in JSONL format. The metadata includes the author, release time and source platform identifier.

[0069] Specifically, the above process corresponds to step S1, which includes:

[0070] S1.1: Write code to crawl multiple mainstream media pages using web crawlers to crawl data from each platform, including the source of the information and the page documents of the information;

[0071] S1.2: Use scheduled tasks to execute crawler code regularly and automatically obtain data.

[0072] Furthermore, we first used crawler technology to write code for pages on multiple platforms, encapsulating different platforms into different modules to ensure complete reliability and scalability. We then used the front-end and back-end technologies of Vue+SpringBoot+MySQL to build a complete detection platform. The interaction process with users included but was not limited to the user manually selecting the platform for which information needs to be detected on the front-end page and then entering the corresponding URL. It also supports adding configuration files containing multiple URL formats and detection times as input for batch timed detection.

[0073] Use the crawler module to crawl post data of specific URLs on the social networking platform, such as post text, post author, post picture, post publishing time and post comment data, and save the data in jsonl format.

[0074] As an implementation method in this embodiment, the preprocessing process includes:

[0075] Use regular expressions to clean illegal characters and HTML tags in text;

[0076] Parse the data to extract text content, author, publication date, and image information;

[0077] Convert the parsed text, image metadata, and cleaned content into structured data and store it in a relational database.

[0078] Specifically, the above process corresponds to step S2, which includes:

[0079] S2.1: Use regular expressions and other tools to clean the data and remove irrelevant or interfering content;

[0080] S2.2: Use automated programs to parse the acquired data and identify the authors, text content, and publication dates of articles, images, and information categories;

[0081] S2.3: Create a structured data model and convert the parsed data into a unified format and store it in a relational database.

[0082] Furthermore, Python code is used to process the collected data, remove illegal and unconventional characters such as emoticons, and remove duplicate and missing data. For posts that do not contain pictures or text, the corresponding fields are set to None. Finally, the cleaned and uniformly formatted JSON data is saved in the MySQL database.

[0083] As an implementation method in this embodiment, the fine-tuning process of the large language model includes:

[0084] Insert the low-rank matrix into the query matrix and value matrix of the model respectively and ,in , A, B, C, and D are all low-rank adaptation matrices inserted by LoRA, which are used to optimize the original query matrix and numerical matrix in an incremental form;

[0085] The cross-entropy loss function is used for supervised training on a labeled dataset that is manually annotated to distinguish false information from true information.

[0086] Specifically, the above process corresponds to the process in step S2, and the step S2 further includes:

[0087] S2.4: The fine-tuning process of a large language model includes:

[0088] Data collection and preprocessing: Collect different types of information data from multiple media platforms. This information data is preprocessed through the following steps to obtain the input matrix for the fine-tuning process:

[0089] Text cleaning: remove HTML tags, special characters, irrelevant spaces, etc. from comments.

[0090] Labeling: Manually or semi-automatically label information to distinguish false from true information, providing training data for subsequent classification tasks and feedback tuning processes.

[0091] Word segmentation: Use word segmentation tools to segment information text and match the token format that the LLaMA model can process.

[0092] Extract data from the file and input it into the LLaMA model for embedding to obtain the user vector.

[0093] Furthermore, the training and fine-tuning process of the model is as follows Figure 4 As shown, the text description is as follows:

[0094] Data Collection and Preprocessing: We collected over 14,000 messages and comments from multiple social media platforms, including 8,200 false messages and 5,500 true messages. This data was preprocessed through the following steps to generate the input preprocessed files for the Lora fine-tuning process.

[0095] Text Cleaning: During data preprocessing, to improve the generalization of LLaMA-3 8B in the Chinese malicious comment detection task, this method uses regular expression matching to remove special characters and irrelevant whitespace characters from the raw data to ensure data validity. Furthermore, this method adds a stop word filtering module to the code to remove words that do not contribute significantly or may cause authentication bias, thereby reducing noise.

[0096] Annotation: To ensure efficient annotation, this method uses both automated and manual review. During the automated process, the method performs an initial assessment by calculating the cosine similarity between unannotated text and high-frequency words found in false information. This result provides a preliminary list of false information candidates, which are then further annotated manually to determine if they are genuine or false information.

[0097] Word segmentation: During the word segmentation process, Hugging Face Transformers is used to convert the text into a token format that the model can recognize.

[0098] During the data loading process, Hagging Face datasets are used to load data. Compared with the traditional JSON parsing method, this method does not require manual parsing of JSON files. It loads the data as a training set and supports simple data splitting into appropriate ratios of training and test sets.

[0099] Some of the core formulas included in LLaMA are as follows:

[0100] Self-Attention Mechanism:

[0101] Query, Key, and Value calculations:

[0102] The model embeds the input through a linear transformation matrix , ,and Perform linear transformation to obtain the calculation of Query, Key and Value:

[0103]

[0104] in, is the query matrix, is the bond matrix, is the value matrix, is the embedding matrix of the input, which represents the representation vector of the input data (feature vector of text or image). 、 、 These are parameter matrices learned during the training process, and are used to generate query, key, and value vectors, respectively.

[0105] The model calculates the similarity between the query and the key to get the attention weight. In order to avoid the value being too large, a scaled dot product is used:

[0106]

[0107] Positional embedding:

[0108]

[0109]

[0110] Where i is the position of the word in the sequence, j is the index of the embedding dimension, and d is the embedding dimension. represents the attention function, represents the position embedding function.

[0111] Output and loss function:

[0112] For classification tasks, the output of the model is used as input and the cross entropy loss module is introduced. The calculation process can be expressed as:

[0113]

[0114] in Represents the prediction result of the model, Represents the true label of the sample, N represents the number of samples, and L represents the loss value.

[0115] Lora fine-tuning model: During fine-tuning, the Q (query matrix) and V (value matrix) modules of the LLaMA model are fine-tuned using the Lora method. Lora fine-tuning optimizes the parameters of these modules, making them more suitable for the task of identifying information and comments in Chinese-language scenarios, and improving the accuracy of false information detection. After multiple tests, the optimal parameter settings were selected, and the ranks of the fine-tuning low-rank matrices A, B and C, D were set to 16 to ensure high learning ability while reducing computational complexity. The core process of fine-tuning is as follows:

[0116] By inserting the low-rank matrix into the Query and Value modules of the Transformer model:

[0117]

[0118]

[0119] Among them, Q and V are the frozen matrices in the original model, that is, the parameters of these matrices will not be updated, and they maintain the weights of the original model. , 、 and 、 is the new LoRA-inserted low-rank adaptation matrix used to incrementally optimize the original query matrix and numerical matrix.

[0120] Validation and tuning: Use the labeled test set and generated labeled data to validate the model and evaluate the performance of the model.

[0121] Model deployment and update: Deploy the model trained and fine-tuned through supervised learning into the actual system, and continuously optimize and adjust the model using user feedback.

[0122] As an implementation method of this embodiment, the process of fusing text features with image features includes:

[0123] Extracting text context features and correlating them with entity descriptions retrieved from Internet search engines;

[0124] Generate image description titles and extract embedded text in images;

[0125] Text features, image description titles and embedded text are concatenated into input text, and preset judgment prompt words are added.

[0126] Specifically, the above process corresponds to step S3, which includes:

[0127] S3.1: Use the existing model’s text encoder to extract the text features that need to be discriminated, and obtain the text context features of the information;

[0128] S3.2: To maximize the extraction of semantic information from images, this method performs two transformations on information image features: first, using existing models to generate titles for information images to obtain global semantic information; second, using tools to detect embedded text within information images. Finally, the features are stitched together to obtain image context features.

[0129] S3.3: Use existing methods to extract entities from image context features and text context features to obtain corresponding entity sets, search for corresponding entity descriptions in Internet search engines, and finally splice the entities and description texts into a complete sentence to obtain the input text of the large language model.

[0130] Furthermore, the BERT model's text encoder is used to extract the text features that need to be discriminated, obtaining the text context features of the information. The TagMe tool is used to extract the corresponding entity descriptions, and the splicing calculation process can be described as follows:

[0131]

[0132]

[0133] in, Represents the entity set in the input information text, Represents the text context features of news, Represents the set of textual background knowledge related to information text entities. Regarding the image features of information, in order to maximize the extraction of semantic information in the image, the following two transformations are performed on the image features: the BLIP model is used to generate the title of the information image to obtain global semantic information, and the PaddleOCR model is used to detect the embedded text in the information image. Finally, the TagMe tool is used to extract the corresponding entity descriptions and splice them to obtain the contextual features of the information image. The calculation results are:

[0134]

[0135] in To extract the description title of information image using BLIP, To detect embedded text in information images using the PaddleOCR model, is the textual expression of information images, Represents the set of textual background knowledge related to information image entities;

[0136] As an additional implementation manner in this embodiment, the process of calculating the false probability of the information data includes:

[0137] S3.4: Input the obtained text and the specified prompt into the fine-tuned large language model for judgment, and output the probability value of true or false information.

[0138] S3.5: Use the fine-tuned large language model to analyze the data in the database and obtain a health score for each review;

[0139] S3.6: Synchronously update the obtained true and false scores to the corresponding information data in the database.

[0140] S3.7: Return the results to the user in real time. For information that takes a long time to process, use an asynchronous processing mechanism to immediately return the processing status or preliminary results, and notify the user after the processing is completed.

[0141] Furthermore, the text description obtained in S3.3 and S3.4 and the specified prompt are input into the fine-tuned large language model for judgment, and the output information is the probability value of true or false;

[0142] The result of the authenticity determination is returned to the front-end user through SpringMVC. For information that requires longer processing time, an asynchronous processing mechanism is used to immediately return the processing status or preliminary results, and then notify the user after the processing is completed. At the same time, the authenticity probability field of the corresponding information data is stored in the MySQL database.

[0143] As an implementation method of this embodiment, the process of performing an automated processing operation on information determined to be false according to the false probability includes:

[0144] For platforms that support public APIs, submit API requests to report or block the platform;

[0145] For platforms that do not support APIs, mark or hide operations are performed by simulating user behavior, including disguising request headers and switching proxy IPs.

[0146] Specifically, the above process corresponds to step S4, which includes:

[0147] S4.1: Use web crawlers to write automatic reporting and hiding codes for different media;

[0148] S4.2: Use the front-end to display a list of suspected false information;

[0149] S4.3: Users select false or partially false information in batches and perform operations such as blocking, reporting, or marking.

[0150] Further, such as Figure 3 As shown, the architecture diagram of the automatic detection and control system for media false information based on front-end and back-end technologies is disclosed. The shielding, reporting or marking codes used in this method have been integrated into the corresponding platform module of the S1 step. For platforms that do not provide an official reporting API, the system reports by simulating the manual reporting process. However, in the specific implementation process, some platforms have certain robot detection behaviors. In order to cope with common robot detection mechanisms, the system can use User-Agent disguise, proxy IP rotation, verification code recognition and other technologies. For platforms that provide a reporting API, the system submits a corresponding reporting request to complete the report. In addition, if the platform does not provide a public reporting entrance, the system can use the method of hiding or folding relevant information to block content locally to reduce the spread of false information;

[0151] The front-end interface displays a list of false information returned by the backend, dynamically rendering the information's platform, content tags, summary, publication time, and model rating. Depending on the platform's functionality, corresponding blocking, flagging, and reporting buttons are illuminated. If the platform doesn't have such functionality, the information will appear grayed out. To enhance the user experience, the system uses different colors to indicate information of varying risk levels. Users are also provided with filtering, search, and sorting features to help them quickly locate content.

[0152] Based on this, an embodiment of the present invention provides an automatic detection and control method for false information and comments based on a large model and fine-tuning. It improves the accuracy of false information detection through multimodal feature fusion and low-rank adaptive fine-tuning technology, and combines text context features, image semantic information and external knowledge bases to effectively identify complex false information with implicit logical contradictions or factual conflicts. It uses a low-rank matrix fine-tuning method to optimize the parameters of the large language model, while retaining general capabilities and enhancing the ability to distinguish Chinese false information. The present invention achieves rapid response throughout the entire process through an automated processing mechanism, significantly reducing the cost of manual review, and its modular design supports flexible expansion to different social media platforms. It dynamically labels risk levels, and achieves local shielding and dissemination containment by simulating user behavior for platforms without public interfaces. The overall solution achieves technological breakthroughs in detection accuracy, processing efficiency and system scalability, providing support for large-scale user-generated content management.

[0153] This invention is widely applicable to user-generated content (UGC) management in online environments such as social media platforms, online forums, and e-commerce platforms, particularly in monitoring, filtering, and automatically processing false information and comments. Furthermore, it can quickly extract authentic and valid information from the vast amount of content on social platforms, accurately identify comments using a fine-tuned LLM model, and implement automated blocking, reporting, or tagging functions based on this information. This allows users to more easily filter out false information and enhance user experience.

[0154] Example 2

[0155] In this embodiment, a computer terminal device is provided, including:

[0156] one or more processors;

[0157] a memory, coupled to the processor, for storing one or more programs;

[0158] When the one or more programs are executed by the one or more processors, the one or more processors implement the methods in the above embodiments.

[0159] In this embodiment, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the method in the above embodiment is implemented.

[0160] In this embodiment, an electronic device is further provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute the method in the above embodiment.

[0161] The above program can be executed in a processor or stored in a memory (or computer-readable medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and can be implemented using any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0162] These computer programs can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps of the functions specified in one or more blocks can be implemented by different modules corresponding to different steps.

[0163] This embodiment provides such a device or system. The system is called an automatic detection and control system for false information and comments based on a large model and fine-tuning, and includes:

[0164] A data collection module, which is used to periodically capture user-generated content data from multiple platforms through a web crawler. The data includes text, images, and metadata.

[0165] A preprocessing module is used to clean the data, remove illegal characters and HTML tags, extract text content and image embedded text, and convert the data into a structured format and store it in a database;

[0166] a model processing module, comprising a fine-tuned large language model for fusing text features with image features and calculating the false probability of the information, wherein a low-rank adaptation matrix is ​​inserted into the query matrix and the value matrix of the model;

[0167] A false information determination module, used to generate a determination result based on the false probability;

[0168] An automated control module is used to block, report, or mark information that is determined to be false.

[0169] As an implementation method of this embodiment, the data acquisition module includes:

[0170] The crawler configuration unit is used to write parsing code based on the page structure of the target platform and define crawling parameters;

[0171] The scheduled execution unit is used to periodically trigger crawler tasks, obtain posts, comments, and related metadata, and store them in JSONL format according to platform classification.

[0172] As an implementation method in this embodiment, the preprocessing module includes:

[0173] Regular expression cleaning unit, used to clean illegal characters and HTML tags in text using regular expressions;

[0174] Data parsing unit, used to extract text content, author, publishing time and image information;

[0175] The structured storage unit is used to convert the cleaned data into a structured format and store it in a relational database.

[0176] As an implementation method in this embodiment, the model processing module includes:

[0177] The low-rank adaptation unit is used to insert low-rank matrices ΔQ and ΔV into the query matrix and value matrix of the model, where ΔQ=AB^T, ΔV=CD^T, and A, B, C, and D are all low-rank adaptation matrices inserted by LoRA;

[0178] The supervised training unit is used to train a labeled data set using a cross-entropy loss function, where the data set contains manually labeled false information and real information.

[0179] As an implementation method of this embodiment, the model processing module further includes:

[0180] The multimodal fusion unit is used to extract text context features and associate them with entity descriptions retrieved from Internet search engines, generate image description titles and extract embedded text in images, and then splice the text features with the image features and input them into the large language model;

[0181] The prompt word injection unit is used to add preset judgment prompt words to the input text to guide the model to output false probability.

[0182] As an implementation method in this embodiment, the automation control module includes:

[0183] API call unit, used to submit blocking or reporting requests to platforms that support public APIs;

[0184] The simulation operation unit is used to simulate user behavior on platforms that do not provide APIs and perform marking or hiding operations, including disguising request headers and switching proxy IP addresses.

[0185] The system or device is used to implement the functions of the method in the above-mentioned embodiment. Each module in the system or device corresponds to each step in the method, which has been explained in the method and will not be repeated here.

[0186] Through the above implementation, the problem of automatic detection and control of false information and comments based on large models and fine-tuning in related technologies is solved, thereby ensuring that technical problems are solved.

[0187] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An automatic detection and control method for false information and comments based on large models and fine-tuning, characterized by: The following steps are involved: Acquiring information data from a network platform, wherein the information data includes text, images, and metadata; Preprocessing the information data includes text cleaning, extracting text content and image embedded text, and converting the data into structured data and storing it in a database; The preprocessed data is input into a fine-tuned large language model, and the false probability of the information data is calculated by fusing text features with image features; The fine-tuning process of the large language model includes: Set the rank of the fine-tuned low-rank matrices A, B and C, D to 16 by inserting the low-rank matrices into the Query and Value modules of the Transformer model: Among them, Q and V are the frozen matrices in the original model, that is, the parameters of these matrices will not be updated, and the weights of the original model are maintained. , 、 and 、 is the new LoRA-inserted low-rank adaptation matrix used to incrementally optimize the original query matrix and numerical matrix; Use the labeled test set and the generated labeled data to validate the model and evaluate the performance of the model; Deploy the supervised learning trained and fine-tuned model into the actual system, and continuously optimize and adjust the model using user feedback; The fusion process of the text features and the image features includes: S3.1: Use the existing model’s text encoder to extract the text features that need to be discriminated, and obtain the text context features of the information; S3.2: For the image features of the information, perform the following two transformations: first, use the existing model to generate the title of the information image to obtain global semantic information; second, use tools to detect the embedded text of the information image; finally, splice them together to obtain the image context features; S3.3: Use existing methods to extract entities from image context features and text context features to obtain corresponding entity sets. Search for corresponding entity descriptions in internet search engines. Finally, concatenate the entity and description text into a complete sentence to obtain the input text for the large language model. Furthermore, the BERT model's text encoder is used to extract the text features that need to be discriminated, obtaining the text context features of the information. The TagMe tool is used to extract the corresponding entity descriptions, and the splicing calculation process can be described as follows: in, Represents the entity set in the input information text, Represents the text context features of news, It represents the set of text background knowledge related to information text entities. For the image features of the information, the following two transformations are performed: the BLIP model is used to generate the title of the information image to obtain global semantic information, the PaddleOCR model is used to detect the embedded text of the information image, and finally the TagMe tool is used to extract the corresponding entity description and splice it to obtain the contextual features of the information image. The calculation results are: in To extract the description title of information image using BLIP, To detect embedded text in information images using the PaddleOCR model, is the textual expression of information images, Represents the set of textual background knowledge related to information image entities; The process of calculating the false probability of the information data includes: S3.4: Input the obtained text and the specified prompt into the fine-tuned large language model for judgment, and output the probability value of the information being true or false; S3.5: Use the fine-tuned large language model to analyze the data in the database and obtain a health score for each review; S3.6: Synchronously update the obtained true and false scores to the corresponding information data in the database; S3.7: Return results to the user in real time. For information that requires a long time to process, use an asynchronous processing mechanism to immediately return the processing status or preliminary results and notify the user after processing is complete. Furthermore, the text description obtained in S3.3 and S3.4 and the specified prompt are input into the fine-tuned large language model for judgment, and the output information is the probability value of true or false; The result of the true or false judgment is returned to the front-end user through springMVC, and stored in the true or false probability field of the corresponding information data in the MySQL database; Perform automated processing on information determined to be false based on the probability of falsehood, including: For platforms that support public APIs, submit API requests to report or block the platform; For platforms that do not support APIs, mark or hide operations are performed by simulating user behavior, including disguising request headers and switching proxy IPs.

2. The method according to claim 1, characterized in that In the process of acquiring information data from the network platform, web crawlers are used to obtain data. The process includes: writing parsing code based on the page structure of the target platform, and executing crawling tasks regularly to obtain posts, comments and related metadata; The crawled data is classified by platform and stored in JSONL format. The metadata includes the author, release time and source platform identifier.

3. The method according to claim 1, characterized in that The pre-treatment process includes: Use regular expressions to clean illegal characters and HTML tags in text; Parse the data to extract text content, author, publication date, and image information; Convert the parsed text, image metadata, and cleaned content into structured data and store it in a relational database.

4. An automatic detection and control system for false information and comments based on large models and fine-tuning, characterized by: The system is used to implement the steps of the method according to any one of claims 1 to 3, and the system includes: A data collection module, which is used to periodically capture user-generated content data from multiple platforms through a web crawler. The data includes text, images, and metadata. A preprocessing module is used to clean the data, remove illegal characters and HTML tags, extract text content and image embedded text, and convert the data into a structured format and store it in a database; a model processing module, comprising a fine-tuned large language model for fusing text features with image features and calculating the false probability of the information, wherein a low-rank adaptation matrix is ​​inserted into the query matrix and the value matrix of the model; A false information determination module, used to generate a determination result based on the false probability; An automated control module is used to block, report, or mark information that is determined to be false.

5. The system according to claim 4, characterized in that The data acquisition module includes: The crawler configuration unit is used to write parsing code based on the page structure of the target platform and define crawling parameters; The scheduled execution unit is used to periodically trigger crawler tasks, obtain posts, comments, and related metadata, and store them in JSONL format according to platform classification.

6. A computer terminal device, characterized in that: include: one or more processors; a memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the automatic detection and control method for false information and comments based on large models and fine-tuning as described in any one of claims 1-3.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for automatically detecting and controlling false information and comments based on a large model and fine-tuning as described in any one of claims 1 to 3 is implemented.

Images