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

Abstract

The invention discloses an automatic detection and control method and system for false information and comments based on a large model and fine tuning, and the method comprises the steps: improving the detection accuracy of the false information through multi-modal feature fusion and a low-rank adaptive fine tuning technology, and combining text context features, image semantic information and an external knowledge base; complex false information implying logic conflicts or factual conflicts can be effectively identified; a low-rank matrix fine tuning method is adopted to optimize large language model parameters, and the ability to discriminate Chinese false information is enhanced while the general ability is reserved. According to the method, full-process quick response is realized through an automatic processing mechanism, the manual auditing cost is remarkably reduced, and the modular design supports flexible expansion to different social media platforms. The risk level is dynamically labeled, and local shielding and propagation restraint are realized by simulating user behaviors for a platform without a public interface. The whole scheme realizes technical breakthrough in detection precision, processing efficiency and system expansibility, and provides support for large-scale user generation content management.

Description

Technical Field

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

[0002] With the explosion in the number of users and interactions on social media platforms, information platforms such as social media have become an important source of information in people's daily lives. However, false information content inevitably appears on the platforms. Such false information not only seriously affects the user experience but may also have a negative impact on the reputation of the platform and even lead to legal risks. Therefore, how to effectively identify and block such false information is a major challenge faced by this method when users obtain real content on social platforms.

[0003] Traditional false information detection methods mainly rely on manually defined rules or pattern matching, manual review, and machine learning models. However, each of these methods has certain limitations. Manually defined rules or pattern matching methods can detect some false information, but they cannot effectively identify some implicit false information and newer false information variants. Although manual review can provide a certain degree of accuracy, with the explosion of platform content, the efficiency and cost issues of manual review have become increasingly serious. In addition, existing machine learning models have low accuracy in identifying false information with complex emotions or for specific scenarios, and are prone to misjudgment or missed judgment.

[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 information identification. These technologies can more deeply understand the context and sentiment information of the text, thereby more accurately identifying false or partially false information. However, how to efficiently apply these technologies to content filtering on social media platforms remains an urgent problem to be solved. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a method and system for automatically detecting and controlling false information and comments based on large models and fine-tuning to solve the problems existing in the above prior art.

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

[0007] Obtain information data in a network platform, where the information data includes text, images, and metadata;

[0008] Preprocess the information data, where the preprocessing includes text cleaning, extracting text content and image-embedded text, and converting it into structured data for storage in a database;

[0009] Input the preprocessed data into a fine-tuned large language model, and calculate the false probability of the information data by fusing text features and image features;

[0010] Perform automated processing operations on the information determined to be false according to the false probability, including blocking, reporting, or marking;

[0011] Among them, the fine-tuning process of the large language model includes: inserting low-rank matrices 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 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 the crawling task to obtain posts, comments, and associated metadata;

[0013] Classify the crawled data by platform and store it in JSONL format. The metadata includes the author, publication time, and source platform identifier.

[0014] Optionally, the preprocessing process includes:

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

[0016] Parse the data to extract text content, author, publication time, 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 low-rank matrices into the query matrix and value matrix of the model and , where , A, B, C, and D are all low-rank adaptation matrices inserted by LoRA, which are used to optimize the original query matrix and value matrix in an incremental form;

[0020] Use the cross-entropy loss function to perform supervised training on the labeled dataset, and the labeled dataset is manually labeled to distinguish false information from true information.

[0021] Optionally, the fusion process of the text features and image features includes:

[0022] Extract the context features of the text and associate with the entity descriptions retrieved from the Internet search engine;

[0023] Generate an image description title and extract the embedded text in the image;

[0024] Concatenate the text features, the image description title, and the embedded text into the input text, and add a preset judgment prompt word.

[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 perform reporting or blocking operations;

[0027] For platforms that do not support APIs, perform marking or hiding operations by simulating user behavior, including disguising the request header and switching the proxy IP.

[0028] In a second aspect, the present invention also 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 includes:

[0029] A data collection module for periodically scraping user-generated content data of multiple platforms through a web crawler. The data includes text, images, and metadata;

[0030] A preprocessing module for cleaning the data, removing illegal characters and HTML tags, extracting the text content and the image embedded text, and converting the data into a structured format for storage in a database;

[0031] A model processing module, including a fine-tuned large language model, for fusing text features and image features, 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 for generating a determination result according to the false probability;

[0033] An automated control module for performing blocking, reporting, or marking operations on the information determined to be false.

[0034] Optionally, the data collection module includes:

[0035] A crawler configuration unit for writing parsing code according to the page structure of the target platform and defining crawling parameters;

[0036] A timed execution unit for periodically triggering the crawler task, obtaining posts, comments, and associated metadata, and storing them in JSONL format classified by platform.

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

[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 a large model and fine-tuning.

[0041] In a fourth aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements an automatic detection and control method for 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 an automatic detection and control method and system for false information and comments based on a large model and fine-tuning. The present invention improves the accuracy of false information detection through multi-modal feature fusion and low-rank adaptation fine-tuning technology, combines text context features, image semantic information and external knowledge bases, and effectively identifies complex false information with implicit logical contradictions or factual conflicts; adopts a low-rank matrix fine-tuning method to optimize the parameters of the large language model, and enhances the discriminative ability for Chinese false information while retaining general capabilities. The automated processing mechanism enables a full-process rapid response, significantly reducing the manual review cost. The modular design supports flexible expansion to different social media platforms. The front-end interface dynamically marks the risk level and provides a batch operation function. For platforms without public interfaces, local shielding and dissemination containment are achieved by simulating user behaviors. 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 that form a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0045] Figure 1 is a schematic flowchart of the method according to an embodiment of the present invention;

[0046] Figure 2 is a framework diagram of the method according to an embodiment of the present invention;

[0047] Figure 3Architecture diagram of the automatic detection and control system for media false 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 according to an embodiment of the present invention. Specific implementation manners

[0049] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

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

[0051] Embodiment 1

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

[0053] Obtain information data in the network platform, where the information data includes text, images, and metadata;

[0054] Preprocess the information data, where the preprocessing includes text cleaning, extracting text content and text embedded in images, and converting it into structured data for storage in a database;

[0055] Input the preprocessed data into the fine-tuned large language model, and calculate the false probability of the information data by fusing text features and image features;

[0056] Perform automated processing operations on the information determined to be false according to the false probability, including blocking, reporting, or marking;

[0057] Among them, 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 an annotated data set.

[0058] As Figure 2 shown, specifically including: S1: Regularly and batch-automatically crawl the historical information and comment data of multiple platforms using web crawler technology;

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

[0060] S3: Process the information and use the fine-tuned large language model to analyze the information, judge the authenticity of the information, and return the probability that the information is false;

[0061] S4: Automatically process the information determined to be false, including operations such as blocking, reporting, or marking.

[0062] For the above S1 to S4, it further includes:

[0063] S1: Through the front-end and back-end vue + springboot + mysql technology, use web crawler technology to automatically crawl information and comments on various online social platforms. After configuring the corresponding parameters, crawl, for example: the text content of the post, the picture content of the post, the time range of the post, the publisher of the post;

[0064] S2: After obtaining the relevant information data, organize the data into a unified JSONL format and clean it. After removing illegal characters and null characters, store the data in a relational database;

[0065] S3: Connect to the database, retrieve the saved user data, use data such as article id, article author, article content, comment id, comment author, and comment content as input, judge its health level through a pre-trained large language model and return a health score, and update the obtained health score to the database through the comment id;

[0066] S4: Through the front-end and back-end technology, display the unhealthy comments with a health score lower than the specified standard. Use web crawlers to locate comments according to article and comment ids and perform deletion and hiding operations. Users can batch-select malicious comments for control.

[0067] As an implementation method in this embodiment, in the process of obtaining information data from the network platform, web crawlers are used to obtain data. The process includes: writing parsing code according to the page structure of the target platform, and periodically executing the crawling task to obtain posts, comments, and associated metadata;

[0068] Classify the crawled data by platform and store it in JSONL format. The metadata includes the author, publication time, and source platform identifier.

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

[0070] S1.1: Write code using web crawlers for the pages of multiple mainstream media, and crawl data on each platform. The data includes the source of the information and the page document of the information;

[0071] S1.2: Use a scheduled task to periodically execute the crawler code to automatically obtain data.

[0072] Furthermore, first, code is written for the pages of multiple platforms using web crawler technology, and different platforms are encapsulated into different modules to ensure complete reliability and extensibility. Then, a complete detection platform is built using front-end and back-end technologies vue + springboot + mysql. The interaction process with users includes, but is not limited to, the user being able to manually check the platforms for which information needs to be detected on the front-end page, and then inputting the corresponding url. It also supports adding a configuration file containing multiple url formats and detection times as input for batch scheduled detection.

[0073] Use the crawler module to crawl the post data of specific urls on the online social platform, such as the text of the post, the author of the post, the pictures of the post, the release time of the post, and the comment data of the post, 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 the text;

[0076] Parse the data to extract text content, author, release time, 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, and S2 includes:

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

[0080] S2.2: Use an automated program to parse the obtained data, identify the author, text content, release time, pictures, and classification of the information of the article;

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

[0082] Furthermore, use python code to process the collected data, remove illegal non-conventional characters such as emojis, and at the same time remove duplicate and missing data. For posts without pictures or without text, set the corresponding fields to None. Finally, save the cleaned and uniformly formatted jsonl data to the mysql database.

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

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

[0085] Use the cross-entropy loss function to perform supervised training on the labeled dataset, and the labeled dataset distinguishes false information from true information through manual labeling.

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

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

[0088] Data collection and preprocessing: Collect information data of different categories from multiple media platforms. These information data are 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 the comments.

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

[0091] Word segmentation processing: Use a word segmentation tool to segment the information text to match the tokens 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 Figure 4 shown, and the text description is as follows:

[0094] Data collection and preprocessing: A total of more than 14,000 pieces of information and comment data were collected from multiple online social platforms, including 8,200 pieces of false information and 5,500 pieces of true information. These information data are preprocessed through the following steps to obtain the input preprocessing file for Lora fine-tuning.

[0095] Text cleaning: In the data preprocessing stage, to improve the generalization ability of LLaMA-3 8B in the Chinese malicious comment recognition task. For the original data, this method uses regular expression matching to remove some special characters and irrelevant empty characters in the text to ensure the validity of the data. In addition, this method adds a stop word filtering module in the code to remove some words with no significant contribution or words that may cause certification bias in the text, reducing noise.

[0096] Annotation: To ensure the annotation efficiency, this method uses an automated and manual review method for annotation. In the automated process, this method makes an initial judgment by calculating the cosine similarity between the unannotated text and the high-frequency words that appear in the false information. The result will provide a preliminary list of false information candidates. After that, manual review is carried out on these false information candidate lists to determine whether they are true false information.

[0097] Word segmentation processing: In the word segmentation process, Hugging Face Transformers is used to convert the text into a token format recognizable by the model.

[0098] In the process of data loading, Hagging Face datasets is used to load the data. Compared with the traditional json parsing method, this method does not require manual parsing of jsonl files and directly loads them as the training set. At the same time, it supports the simple splitting of data into appropriate training set and test set ratios.

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

[0100] Self-attention mechanism:

[0101] Calculation of Query, Key, and Value:

[0102] The model will linearly transform the input embedding through the linear transformation matrix , , and to perform the linear transformation and obtain the calculation of Query, Key, and Value:

[0103]

[0104] Among them, is the query matrix, is the key matrix, is the value matrix, is the input embedding matrix, representing the representation vector of the input data (the feature vector of text or image). , , They are all parameter matrices learned during the training process and are used to generate vectors for Query, Key, and Value respectively.

[0105] The model calculates the similarity between the Query and the Key to obtain the attention weights. To avoid overly large values, scaled dot product is used:

[0106]

[0107] Position 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 the classification task, the output result of the model is used as the input, and the cross-entropy loss module is introduced. The calculation process can be expressed as:

[0113]

[0114] where 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 the fine-tuning process, the Lora method is used to fine-tune the Q (query matrix) and V (value matrix) modules of the LLaMA model. Lora fine-tuning optimizes the parameters of these modules to make them more suitable for information recognition tasks and comment recognition tasks in the Chinese scenario, and improves the judgment accuracy of false information. After multiple tests, better parameter settings are selected. The ranks of the fine-tuning low-rank matrices A, B and C, D are set to 16 to ensure high learning ability under the premise of relatively low computational cost. The specific core fine-tuning process is as follows:

[0116] By inserting low-rank matrices 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 、 are the low-rank adaptation matrices inserted by the new LoRA, which are used to optimize the original query matrix and value matrix in an incremental form.

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

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

[0122] As an implementation method in this embodiment, the fusion process of the text feature and the image feature includes:

[0123] Extract text context features and associate entity descriptions retrieved from the Internet search engine;

[0124] Generate an image description title and extract the embedded text in the image;

[0125] Concatenate the text feature, the image description title and the embedded text into the input text, and add a preset judgment prompt word.

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

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

[0128] For the image features of the information, in order to extract the semantic information in the image to the greatest extent, this method performs the following two conversions on the image features of the information: one is to use the existing model to generate the title of the information image to obtain the global semantic information; the other is to use tools to detect the embedded text in the information image. Finally, perform concatenation to obtain the image context features;

[0129] S3.3: Use the existing method to extract entities from the image context features and the text context features to obtain the corresponding entity set, search for the corresponding entity descriptions in the Internet search engine respectively, and finally concatenate the entity and the description text to form a complete sentence to obtain the input text of the large language model.

[0130] Further, a text encoder using the BERT model is employed to extract the text features of the information to be discriminated, obtaining the text context features of the information. The process of using the TagMe tool to extract the corresponding entity descriptions and perform splicing calculations can be described as follows:

[0131]

[0132]

[0133] Among them, represents the set of entities in the input information text, represents the text context features of the news, represents the set of text background knowledge related to the entities in the information text. For the image features of the information, in order to extract the semantic information in the image to the greatest extent, the following two transformations are performed on the image features of the information: Use the BLIP model to generate the title of the information image to obtain global semantic information, and use the PaddleOCR model to detect the embedded text in the information image. Finally, the TagMe tool is also used to extract the corresponding entity descriptions and perform splicing to obtain the information image context features. The calculation result is:

[0134]

[0135] Among them is the descriptive title of the information image extracted by using BLIP, is the embedded text in the information image detected by using the PaddleOCR model, is the text expression of the information image, represents the set of text background knowledge related to the entities in the information image;

[0136] As an additional implementation method 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 the information being true or false.

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

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

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

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

[0142] Return the result of true or false judgment to the front-end user in the way of springMVC. For information that takes a long time to process, adopt an asynchronous processing mechanism, immediately return the processing status or preliminary result, and notify the user after the processing is completed. At the same time, store it in the true / false probability field of the corresponding information data in the MySQL database.

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

[0144] For platforms that support public APIs, submit API requests to perform reporting or blocking operations;

[0145] For platforms that do not support APIs, perform marking or hiding operations by simulating user behavior, including disguising the request header and switching the proxy IP.

[0146] Specifically, the above process corresponds to step S4, and S4 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 batch-select false or partially false information and perform operations such as blocking, reporting, or marking.

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

[0151] The front - end interface is used to display a list of false information returned by the back - end, and dynamically render information including the affiliated platform, content tags, content abstract, release time, model score. According to the functions provided by the affiliated platform, corresponding blocking, marking, and reporting buttons are lit. If the function does not exist on the platform, it will be shown in gray. To improve the user experience, the system marks information with different risk levels in different colors and provides functions such as filtering, searching, and sorting to help users quickly locate content.

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

[0153] The present invention is widely applicable to the management of user - generated content (UGC) in network environments such as social media platforms, online forums, and e - commerce platforms, especially in scenarios of monitoring, filtering, and automatic processing of false information and comments. And the present invention can quickly extract true and effective information from the massive content of social platforms, accurately identify comments through a fine - tuned LLM model, and on this basis, realize automated functions such as blocking, reporting, or marking. Users can more easily filter out false information and ensure the user experience.

[0154] Embodiment Two

[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 also provided, on which a computer program is stored. When the computer program is executed by a processor, the methods in the above embodiments are implemented.

[0160] In this embodiment, an electronic device is also provided, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the methods in the above embodiments.

[0161] The above program can run in a processor, or can also be stored in a memory (or referred to as a computer-readable medium). The computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The 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 cassette tapes, magnetic 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 operation steps are executed on the computer or other programmable device to generate computer-implemented processing. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or boxes Figure 1 Steps corresponding to different steps can be implemented by different modules to implement the functions specified in one box or multiple boxes.

[0163] In this embodiment, such a device or system is provided. The system is called an automatic detection and control system for false information and comments based on large models and fine-tuning, including:

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

[0165] A preprocessing module, configured 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] The model processing module includes a fine-tuned large language model for fusing text features and image features and calculating the false probability of the information, where a low-rank adaptation matrix is inserted into the query matrix and value matrix of the model;

[0167] The false information determination module is used to generate a determination result according to the false probability;

[0168] The automatic control module is used to perform operations such as shielding, reporting, or marking on the information determined to be false.

[0169] As an implementation manner in this embodiment, the data acquisition module includes:

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

[0171] The timed execution unit is used to periodically trigger the crawler task, obtain posts, comments, and associated metadata, and store them in JSONL format classified by platform.

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

[0173] The regular cleaning unit is used to clean illegal characters and HTML tags in the text through regular expressions;

[0174] The data parsing unit is used to extract text content, author, publication 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 manner 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 the labeled data set through the cross-entropy loss function, and the data set contains artificially labeled false information and true information.

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

[0180] The multi-modal fusion unit is used to extract text context features, associate entity descriptions retrieved from the Internet search engine, generate image description titles, extract image embedded text, splice text features and image features, and input them into the large language model;

[0181] A prompt injection unit for adding a preset judgment prompt to the input text to guide the model to output a false probability.

[0182] As an implementation manner in this embodiment, the automatic control module includes:

[0183] An API call unit for submitting a shielding or reporting request to a platform that supports public APIs;

[0184] A simulated operation unit for simulating user behavior to perform marking or hiding operations on platforms that do not provide APIs, including disguising request headers and switching proxy IP addresses.

[0185] This system or device is used to implement the functions of the method in the above embodiment. Each module in this system or device corresponds to each step in the method. Those that have been described in the method will not be elaborated here.

[0186] Through the above implementation manner, the problems of automatic detection and control of false information and comments based on large models and fine-tuning in the related art are solved, so that the technical problems can be guaranteed to be solved.

[0187] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An automatic detection and control method for false information and comments based on large models and fine-tuning, characterized in that, Including the following steps: Obtain information data in the network platform, where the information data includes text, images, and metadata; Preprocess the information data, where the preprocessing includes text cleaning, extracting text content and image-embedded text, and converting it into structured data for storage in a database; Input the preprocessed data into a fine-tuned large language model, and calculate the false probability of the information data by fusing text features and image features; Perform automated processing operations on the information determined to be false according to the false probability, including blocking, reporting, or marking; Among them, 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 data set.

2. The method according to claim 1, wherein In the process of obtaining information data in the network platform, a web crawler is used to obtain data, and the process includes: writing parsing code according to the page structure of the target platform, and periodically executing a crawling task to obtain posts, comments, and associated metadata; Classify the crawled data by platform and store it in JSONL format, where the metadata includes the author, publication time, and source platform identifier.

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

4. The method according to claim 1, wherein The fine-tuning process of the large language model includes: Insert low-rank matrices into the query matrix and value matrix of the model and , where , A, B, C, and D are all low-rank adaptation matrices inserted by LoRA, which are used to optimize the original query matrix and value matrix in an incremental form; Use the cross-entropy loss function to perform supervised training on the labeled data set, and the labeled data set is manually labeled to distinguish false information from true information.

5. The method according to claim 1, wherein The fusion process of the text features and image features includes: Extract text context features and associate entity descriptions retrieved from the Internet search engine; Generate an image description title and extract the embedded text in the image; Concatenate the text features, image description title, and embedded text into an input text, and add a preset determination prompt word.

6. The method according to claim 1, characterized in that The process of performing automated processing operations on the information determined to be false according to the false probability includes: For platforms that support public APIs, submit API requests to perform reporting or blocking operations; For platforms that do not support APIs, perform marking or hiding operations by simulating user behavior, including disguising the request header and switching the proxy IP.

7. An automatic detection and control system for false information and comments based on large models and fine-tuning, characterized in that, The system includes: A data collection module for periodically scraping user-generated content data from multiple platforms through a web crawler, where the data includes text, images, and metadata; A preprocessing module for cleaning the data, removing illegal characters and HTML tags, extracting text content and image-embedded text, and converting the data into a structured format for storage in a database; A model processing module, including a fine-tuned large language model, for fusing text features and image features and calculating the false probability of the information, where a low-rank adaptation matrix is inserted into the query matrix and value matrix of the model; A false information determination module for generating a determination result according to the false probability; An automated control module for performing blocking, reporting, or marking operations on the information determined to be false.

8. The system according to claim 7, wherein The data collection module includes: A crawler configuration unit for writing parsing code according to the page structure of the target platform and defining crawling parameters; A scheduled execution unit for periodically triggering crawler tasks, obtaining posts, comments, and associated metadata, and storing them in JSONL format classified by platform.

9. A computer terminal device, characterized in that, Comprising: 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 method for automatically detecting and controlling false information and comments based on large models and fine-tuning as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the method for automatically detecting and controlling false information and comments based on large models and fine-tuning as described in any one of claims 1-6 is implemented.

Images