Rumor detection data enhancement method and device based on multi-dimensional focus question generation

Abstract

The application discloses a rumor detection data enhancement method and device based on a multi-dimensional focus question generation, and the method comprises the following steps: acquiring rumor data; constructing a rumor detection data set; extracting a multi-dimensional semantic focus; generating a focus question corresponding to the multi-dimensional semantic focus based on the multi-dimensional semantic focus; integrating the rumor detection data and the focus question to obtain rumor detection enhancement input; predicting the authenticity of the rumor; a model predicts a rumor authenticity label; confidence weight optimization of the label is performed through ensemble learning to obtain a final label. The device comprises an acquisition rumor data module, a rumor detection data set construction module, a multi-dimensional semantic focus extraction module and the like. The application introduces a multi-dimensional semantic focus question, performs data enhancement on original rumor data, and prompts through a semantic question, so that a prediction model can comprehensively understand rumor information, and thus can better predict authenticity labels, and provides a more comprehensive, efficient and superior solution for network public opinion monitoring and information guarantee.

Description

Technical Field

[0001] This invention belongs to the field of online public opinion analysis technology, specifically a method and device for enhancing rumor detection data based on multi-dimensional focus issues. Background Technology

[0002] With the widespread use of social media and the rapid development of artificial intelligence, rumors spread much faster, placing higher demands on the credibility of information. False information can trigger unnecessary panic, social discontent, or malicious behavior, especially during critical moments such as public health crises or emergencies, where accurate information is crucial for public safety. Rumor detection technology helps identify and prevent the spread of false information. By identifying and correcting rumors, society can better trust and rely on information disseminated through various channels, enhancing the overall credibility of online information.

[0003] Current research on rumor detection methods often faces challenges such as adversarial attacks and contextual complexity, causing traditional supervised learning methods to perform poorly in certain situations. Some of the main problems with existing rumor detection methods include:

[0004] 1. Supervised learning methods are limited in their learning approach: Traditional supervised learning methods for rumor detection rely solely on single labels from the supervised corpus as the supervisory signal. This makes the rumor detection performance susceptible to noise in the corpus, resulting in poor robustness in rumor identification and unstable rumor detection performance.

[0005] 2. Traditional methods struggle to handle the complex semantics of rumors: News rumors are semantically complex, and multi-faceted analysis at different semantic levels is crucial for improving the comprehensiveness and reliability of rumor detection. However, traditional methods often neglect the necessity of comprehensive analysis from multiple semantic perspectives, leading to inaccurate rumor detection results and lowering the credibility of rumor detection methods.

[0006] 3. Data Augmentation Paradigms for Rumor Detection Need Further Exploration: Data augmentation is an effective means of improving rumor detection performance, especially by enhancing rumor detection data. By introducing different semantic perspectives into the dataset, rumors can be examined more comprehensively, thereby improving the level of identification of false information. However, current mainstream rumor detection data augmentation methods focus on the semantic representation of rumors, neglecting to use guiding questions to plan auxiliary methods for determining the authenticity of rumors.

[0007] In conclusion, while multi-semantic analysis and data augmentation have significant advantages in rumor detection, the challenges and difficulties they face require the use of advanced technologies and comprehensive methods to overcome. Summary of the Invention

[0008] The purpose of this invention is to provide a method and apparatus for enhancing rumor detection data based on multi-dimensional focus questions, targeting the field of online public opinion analysis technology. This method introduces multi-dimensional semantic focus questions to enhance the original rumor data. Through semantic question prompts, the prediction model can more comprehensively understand the rumor information, thereby better predicting the authenticity of the claims and providing a more comprehensive, efficient, and superior solution for online public opinion monitoring and information security.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] Technical Solution: The rumor detection data enhancement method and device based on multi-dimensional semantic focus question generation described in this invention are aimed at the field of online public opinion analysis technology, including a rumor detection data enhancement method based on multi-dimensional focus question generation, and a corresponding rumor detection data enhancement device based on multi-dimensional focus question generation.

[0011] The main steps of the rumor detection data augmentation method based on multi-dimensional focus questions are as follows:

[0012] (1) Obtaining Rumor Data: A large amount of rumor text data to be detected is legally and compliantly crawled from various public information platforms on the Internet, including news media, official media, self-media, e-commerce websites, etc. In the process of obtaining rumor data to be detected, an adaptive dynamic crawling strategy is adopted, which includes the following features: adaptive crawling frequency adjustment, real-time monitoring and hot spot tracking, content priority setting, and automated anomaly detection and response;

[0013] (2) Constructing a rumor detection dataset: Based on the above rumor text data, a subset of rumors is selected evenly, and the authenticity label of each rumor text to be tested in the rumor subset is manually labeled. Multiple semantic focuses and a series of focus issues are manually summarized from different semantic levels of the rumor text to obtain a rumor detection dataset for the learning and optimization of the subsequent rumor detection model.

[0014] (3) Extracting multidimensional semantic focus: Based on the rumor detection dataset constructed above, given a rumor text data T to be tested for authenticity, a series of multidimensional semantic focus Fs of the rumor text from different angles and semantic levels are extracted. In the process of extracting multidimensional semantic focus of the rumor text, two semantic focus exploration approaches are considered: continuous semantic span and metadata indication;

[0015] (4) Generating Multidimensional Focus Questions: Based on the above-mentioned multidimensional semantic focus, a series of focus questions corresponding to the multidimensional semantic focus are generated for each rumor text data. Specifically, a self-generated autoregressive encoder and decoder are used to generate the multidimensional focus questions.

[0016] (5) Input enhancement for question integration: Construct an integrated prompt template, and based on the prompt template, integrate the original rumor text input to be tested with the above-mentioned focus questions, and perform semantic enhancement on the rumor text input to be tested to obtain the enhanced rumor text input;

[0017] (6) Rumor authenticity prediction: Construct a rumor prediction model based on Transformer, predict the rumor detection label set of the above multiple inputs through the model, and learn and optimize based on the above rumor detection dataset;

[0018] (7) Rumor Authenticity Integration: Based on the aforementioned set of multiple rumor detection labels, the confidence weights of the label results are iteratively updated through ensemble learning to obtain the final label for the rumor detection task. Specifically, an ensemble learning-based rumor authenticity integration strategy is designed. This strategy aims to update the confidence weights of the labels in real time during each iteration, select one or more sets of best-performing prediction results, and update the confidence weights of the label predictions based on their error rates. Finally, the label for the rumor detection task is determined by weighting the confidence weights.

[0019] Preferably, step (1) of the method execution process, the rumor data acquisition process, adopts an adaptive dynamic crawling strategy, which includes the following characteristics and elements: adaptive crawling frequency adjustment, real-time monitoring and hotspot tracking, content priority setting, and automated anomaly detection and response, as follows:

[0020] Adaptive crawling frequency adjustment: The crawler's access frequency is automatically adjusted based on the frequency of website content updates. For news websites and trending social media, where content updates are frequent, the crawler can be set to access them more frequently; while for websites that are updated less frequently, the crawling frequency is reduced to decrease the pressure on the server.

[0021] Real-time monitoring and trend tracking: By monitoring trending topics and events on the Internet in real time, we can quickly identify areas or themes that may generate rumors and increase crawling efforts for these areas or themes.

[0022] Content Priority Settings: Different priorities are set for different types of news content. Content with frequent user interaction and rapid dissemination is given a higher crawling priority; content with less interaction is given a lower priority.

[0023] Automated anomaly detection and response: Identify website anti-crawler mechanisms and other anomalies, including access timeouts and server refusal to respond, and automatically adjust crawler strategies to ensure the continuity and integrity of data crawling.

[0024] Feedback-driven optimization: Combining artificial intelligence and machine learning technologies, the web crawler learns and optimizes itself based on historical crawling results. By analyzing the quality and relevance of crawled content and various metrics during the crawling process, it continuously adjusts its crawling algorithms and strategies to improve the efficiency and accuracy of future crawling.

[0025] Preferably, step (2) of the method execution process, namely the construction of the rumor detection dataset, introduces a collaborative annotation mechanism, allowing different annotators to communicate and discuss during the annotation process. This collective wisdom improves the accuracy and consistency of annotation, especially in cases where it is difficult to determine authenticity; the collaborative annotation mode ensures less annotation bias noise. Furthermore, a feedback loop and expert review mechanism are introduced. Annotators can review the annotated data and provide feedback on others' annotation results, forming a positive feedback loop. After annotation is completed, linguistics experts are promptly invited for final review, and the annotation results are randomly checked and verified to ensure the high quality of the rumor detection dataset.

[0026] Preferably, in step (3) of the method execution process, during the extraction of multidimensional semantic focus, two major semantic focus exploration approaches are considered: continuous semantic span and metadata indication, as follows:

[0027] First, regarding the continuous semantic span, a syntax parsing tool was used to parse the rumor text into a parse tree, which was then used to segment the rumor text into M distinct semantic subunits. Set as

[0028] Secondly, regarding metadata indicators, N pieces of metadata about the rumor text data are obtained, including the propagator, publication date, and publication location of the rumor text data. These are then integrated with the original rumor text and the aforementioned metadata using a template to obtain N integrated rumor texts. Set as

[0029] Finally, the above semantic subunits and the integrated rumor text Combined into F =

[0030] [F l ;F r ], as the multidimensional semantic focus set of the rumor text Each focus is represented as f. k L = M + N, used for generating subsequent semantic focus issues.

[0031] Preferably, step (4) in the method execution flow: generating a multidimensional focus problem, the process of which is as follows:

[0032] Given a piece of rumor text data T to be detected and its semantic focus set. For each semantic focus f k Combine the rumor text T to generate a sequence with each f k The corresponding focus issue q k This leads to the set of focus issues corresponding to the multidimensional semantic focus set F. Specifically, the text T to be detected and each semantic focus f in the multidimensional semantic focus set F mentioned above will be used. k By concatenating the text, we obtain the concatenated text sequence. The input is then fed into a focus question generation model, which employs a self-generated autoregressive encoder and decoder, including BART and T5 language models. This model calculates the conditional probability distribution of the language model corresponding to the input. This focus-point generation model will be based on the conditional probability distribution of the language model. Generate each semantic focus f k The corresponding focus issue q k Given the rumor text T to be detected and the semantic focus f k This focus question generation model generates focus question q word by word. k The formula is as follows:

[0033]

[0034] Wherein, the left side of the equation represents the text to be detected T and the semantic focus f. k Generate focus q k Language model conditional probability; the right side of the equation represents the focus problem q. k The product of the conditional probabilities generated for each character in the text; For semantic focus f k The sequence of the text T to be detected is concatenated with the input of the focus problem generation model.

[0035] Preferably, step (5) in the method execution flow: input enhancement for question integration, which aims to semantically enhance the rumor text input to be tested, and its specific steps are as follows:

[0036] First, an integrated prompt template is constructed. Then, based on this prompt template, the original rumor text input to be tested is integrated with the aforementioned key issues to obtain the semantically enhanced rumor text to be tested, as follows:

[0037] First and foremost, the prompt template used to integrate the rumor text to be tested with the key issue must meet the following two conditions:

[0038] 1. A slot containing the rumor text data to be tested [SLOT of Rumor], a slot containing the focus question [SLOT of FocalQ], and a label slot [SLOT of Label] are used to output candidate answers for the labels to be predicted;

[0039] 2. This template must specify the following requirements: Based on the focus issue, the reasonableness and authenticity of the rumor text to be tested must be evaluated, that is, by predicting candidate answers for the label slot, and then obtaining the final rumor detection true or false label.

[0040] In the implementation phase, various prompts can be used as integrated prompt templates. Here is a typical example: "From the semantic perspective of the question: [SLOT of FocalQ], can the truthfulness of the question [SLOT of Rumor] be guaranteed? Yes or no? Please answer: [SLOT of Label]".

[0041] Based on the above steps, the integration of the rumor text to be tested and the focus issue is completed, resulting in the integrated input representation Prompt([T;q)). k This is used as input for subsequent rumor prediction model learning.

[0042] Preferably, step (6) in the execution flow of the method: prediction of the veracity of rumors, the process of which is as follows:

[0043] First, a Transformer-based rumor prediction model is constructed for rumor detection and label prediction. This model employs a prompting learning framework based on the T5 generative model, and its parameters are optimized through backpropagation on a dataset.

[0044] This prediction model is used to predict the above integrated input representation Prompt([T:q)). k The input consists of multiple sets of rumor detection labels. The label prediction involves two stages: language model probability prediction and discrete answer-label alignment. The specific process is as follows:

[0045] 1. Language Model Probability Prediction: The generation probability of each character in the target sequence is obtained through this T5 generative model. The specific process is as follows:

[0046] First, through the T5 encoding embedding layer Layer E The input Prompt([T; q) is a sequence of characters. k Embedded into a high-dimensional vector The formula is as follows:

[0047]

[0048] Subsequently, through multiple layers of consecutive Transformer blocks (TransBlock) k Encode the input to obtain a multi-layered deep representation of the initial input. Where N is the number of Transformer layers in the T5 language model, which is 12, as shown in the formula below:

[0049]

[0050] Finally, based on this depth representation The vocabulary probability distribution p(w) of the j-th character in the target sequence is obtained through a fully connected layer containing weights W and biases b and a normalized exponential function SoftMax. j The formula is as follows:

[0051]

[0052] 2. Discrete Answer and Label Alignment: Given a candidate answer and label mapping table Labels of {"Real, Valid, Trustworthy, Accurate, Yes": real; "False, Untrustworthy, False, No": fake}, based on the character probability distribution of the target sequence, calculate and integrate the probability distribution P of the characters corresponding to the label slots [SLOT of Label]. LM (w [SLOT of Label] |[T;q k ], θ), calculate the average cross-entropy with candidate answers. This involves predicting the fit between the prediction and candidate answers to determine the candidate answers. Then, based on the candidate answers, a candidate answer label mapping table is searched and matched to derive the final rumor detection label L mapped to the intermediate answer. gold The formula for the above process is as follows:

[0053]

[0054] Where θ is the set of trainable parameters for the T5 language model, and L gold The final label is obtained through the Labels mapping table, L gold ∈{real,fake}, note L gold It is aimed at a specific focus problem q k The authenticity.

[0055] Preferably, step (7) in the method execution flow is: rumor authenticity integration, wherein a rumor authenticity integration strategy based on ensemble learning is designed, and the specific process of the strategy is as follows:

[0056] Initially, the label prediction result L corresponds to each focus question t. t The confidence weight w of the result tThey are initially set to be equal, and a weak rumor veracity predictor M is set for each focus issue t. t ;

[0057] In each round of iterative optimization, the error rate is calculated for each specific semantic focus problem t. And based on error rate Update its corresponding weak rumor veracity predictor M t

[0058] Simultaneously update the confidence weight w corresponding to each focus question t. t The details are as follows:

[0059] First, based on the error rate of this focal problem t Calculate the weight balance gradient α t The formula is as follows:

[0060]

[0061] Furthermore, based on the weight balancing gradient α t For the correctly predicted label result L t The confidence weight w corresponding to the focal issue t t Update the confidence weight w t The formula is as follows:

[0062] Conversely, based on the weighted balancing gradient α t For the incorrectly predicted label result L t The confidence weight w corresponding to the focal issue t t Update the confidence weight w t The formula is as follows:

[0063]

[0064] Once the confidence weights of all semantic questions have stabilized, a weak rumor veracity predictor M is set up for each of the aforementioned focus questions t. t The confidence weights of all weak rumor predictors are summed by ∑w. t *M t To build the ultimate powerful rumor veracity predictor Right now

[0065] Finally, through Weak rumor veracity predictor M for each focal issue t t The confidence weights are weighted and summed, and the true / false labels with the larger weighting are regarded as the final labels for this rumor detection task.

[0066] Preferably, the rumor detection data enhancement device based on multi-dimensional focus issues is a specific implementation of the rumor detection data enhancement method, and the device includes the following modules:

[0067] The rumor data acquisition module is used to legally and compliantly crawl a large amount of rumor text data to be detected from major public opinion platforms on the Internet through web crawling technology.

[0068] A rumor detection dataset module is constructed to uniformly select a subset of rumors based on rumor text data, manually label the truth or falsehood of each rumor text to be tested in the rumor subset, and manually summarize multiple semantic focuses and a series of focus issues from different semantic levels of the rumor text to obtain a rumor detection dataset for the learning and optimization of subsequent rumor prediction models.

[0069] The module for extracting multidimensional semantic focus is used to extract a series of multidimensional semantic focus from different angles and semantic levels of a rumor text to be detected, based on the obtained rumor detection dataset. In the process of extracting multidimensional semantic focus of the rumor text, two semantic focus exploration approaches are considered: continuous semantic span and metadata indication.

[0070] The module for generating multidimensional focus questions is used to generate a series of focus questions corresponding to the multidimensional semantic focus for each piece of rumor text data;

[0071] The input enhancement module for question integration is used to construct an integrated prompt template. Based on this prompt template, the original rumor text input to be tested and the focus question are integrated, and the semantics of the rumor text input to be tested are enhanced to obtain the enhanced rumor text input.

[0072] The rumor authenticity prediction module is used to build a Transformer-based rumor prediction model. This model predicts multiple sets of input rumor detection labels and learns and optimizes them based on the rumor detection dataset.

[0073] The rumor authenticity integration module is used to iteratively update the confidence weights of the label results based on multiple sets of rumor detection labels through ensemble learning, so as to obtain the final label for the rumor detection task.

[0074] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the rumor detection data enhancement method based on multi-dimensional focus question generation.

[0075] Specifically, for the computer program written in Python, mainstream deep learning frameworks such as PyTorch or Tensorflow can be selected as the training frameworks for the focus question generation model and the Transformer-based rumor prediction model in this method.

[0076] The present invention has at least the following beneficial technical effects:

[0077] 1. High-quality dataset construction: By manually constructing a rumor detection dataset containing multidimensional focus questions and detection labels, the method avoids the difficulties of relying on large-scale annotation and improves the efficiency of dataset construction. At the same time, the multidimensional semantic focus questions obtained during the annotation process also provide the model with richer learning information.

[0078] 2. Enhanced adaptability to complex contexts: The rumor data acquisition employs a dynamic crawling strategy, adaptively adjusting the crawling frequency, monitoring hot topics in real time, and setting content priorities. This allows the method to better adapt to changes in online public opinion and maintain adaptability to complex contexts;

[0079] 3. Improve the accuracy of rumor detection: By introducing a multi-dimensional semantic focus question generation method, the system can more comprehensively understand rumor information. This helps to accurately identify false information, improve the accuracy of rumor detection, and provide users with a more trustworthy information environment;

[0080] 4. Improved robustness of rumor detection: In the process of semantic focus extraction, multi-dimensional semantic focus extraction based on continuous semantic span and metadata indication is considered, which improves the system's ability to mine the semantics of rumors, and can more flexibly adapt to different forms of rumors, effectively dealing with the challenges of abnormal noise in the data.

[0081] 5. Advanced ensemble learning strategy for integrating rumor and debunking: By iteratively updating the confidence weights of labels through ensemble learning, the best-performing prediction results can be selectively integrated in different iterations, improving the credibility and stability of the final labels.

[0082] In practical applications, this rumor detection data enhancement method and device based on multi-dimensional semantic focus issues is expected to provide more efficient and comprehensive solutions for fields such as online public opinion monitoring and information credibility assurance, maintain the normal order of social information dissemination, and reduce the negative impact of false information on society.

[0083] Of course, implementing this invention in any online public opinion analysis and rumor detection scenario does not necessarily require achieving all the advantages and significance described above at the same time. Attached Figure Description

[0084] Figure 1A flowchart of a method for enhancing rumor detection data based on multi-dimensional semantic focus issues.

[0085] Figure 2 This is a block diagram of a rumor detection data enhancement device generated based on multi-dimensional semantic focus issues. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to a specific embodiment and the corresponding accompanying drawings. It should be noted that the embodiments described below are merely some embodiments of the method of the present invention, and not embodiments compatible with all rumor detection scenarios. Referring to the embodiments demonstrated by the method of the present invention, all other embodiments obtained by those skilled in the art without any inventive step are also within the protection scope of the present invention.

[0087] To address a series of problems existing in the design of data augmentation methods for online rumor detection, this invention discloses a rumor detection data augmentation method and apparatus based on multi-dimensional focus question generation. For details on the rumor detection data augmentation method based on multi-dimensional focus question generation, please refer to the appendix to the specification. Figure 1 As shown, the rumor detection data enhancement device based on multi-dimensional focus question generation is described in the attached instruction manual. Figure 2 As shown.

[0088] It should be noted that the embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed form. Those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features.

[0089] The method described in this invention, in conjunction with its embodiments and appendices, Figure 1 The method flow is described in detail, and the specific methods include:

[0090] First, rumor data was acquired: a large amount of rumor text data to be detected was obtained from major internet opinion platforms, news media, official media, self-media, and e-commerce websites using legal and compliant web scraping technology. During this process, an adaptive dynamic scraping strategy was adopted, including adaptive adjustment of scraping frequency, real-time monitoring and hot topic tracking, content priority settings, and automated anomaly detection and response mechanisms.

[0091] In the process of acquiring rumor data, an adaptive dynamic crawling strategy is adopted. This strategy includes the following characteristics and elements: adaptive adjustment of crawling frequency, real-time monitoring and hotspot tracking, content priority setting, and automated anomaly detection and response, specifically:

[0092] In some embodiments, adaptive crawling frequency adjustment can be employed, intelligently adjusting the crawler's access frequency based on the frequency of website content updates. For platforms with rapidly updated content, such as news websites and trending social media, the crawler can be configured to access more frequently to ensure timely information acquisition; while for websites with less frequent updates, the crawling frequency can be appropriately reduced to decrease the load on the server.

[0093] In some embodiments, real-time monitoring and hotspot tracking can be employed. By monitoring hot topics and events on the Internet in real time, areas or themes that may generate rumors can be quickly identified, and the crawling efforts for these areas or themes can be increased in a targeted manner to ensure that relevant information is captured in a timely manner.

[0094] In some embodiments, content priority settings can be adopted, setting different priorities according to different types of news content. Content with frequent user interaction and fast dissemination speed is given a higher crawling priority; conversely, content with less interaction is given a lower priority to allocate crawling resources more effectively.

[0095] In some embodiments, automated anomaly detection and handling can be employed. The crawler should have the ability to automatically detect website anti-crawler mechanisms and other anomalies, including but not limited to access timeouts and server refusal to respond. Upon detecting anomalies, the crawler should be able to intelligently adjust its strategy to ensure the continuity and integrity of data crawling.

[0096] In some embodiments, feedback-driven optimization can be employed, combining artificial intelligence and machine learning techniques. The crawler should learn and optimize itself based on historical crawling results. By comprehensively analyzing the quality and relevance of crawled content, as well as various indicators during the crawling process, the crawler can continuously adjust its crawling algorithms and strategies, improving the efficiency and accuracy of future continuous crawling.

[0097] Furthermore, a rumor detection dataset is constructed: a subset of rumors is evenly selected from the aforementioned rumor text data and manually labeled to determine the truth or falsehood label of each rumor text to be tested. In addition, multiple semantic focuses and related issues are summarized from different semantic levels of the rumor texts to form a rumor detection dataset for subsequent rumor prediction model learning and optimization.

[0098] In some embodiments, a collaborative annotation mechanism can be selectively introduced for the annotation of rumor detection datasets, allowing different annotators to communicate and discuss during the annotation process. This collective wisdom improves the accuracy and consistency of annotation, especially in cases where it is difficult to determine authenticity; this collaborative annotation mode helps reduce annotation bias and noise. Furthermore, a feedback loop and expert review mechanism are introduced. During the annotation process, annotators can review the annotated data and provide feedback on others' annotations, forming a positive feedback loop. After annotation is completed, experts in the field of linguistics are promptly invited to conduct a final review, and the annotation results are sampled and verified proportionally to ensure the high quality of the rumor detection dataset.

[0099] It is important to note that the two steps described above—acquiring rumor data and constructing a rumor detection dataset—are based on content released by authoritative official media. The data crawling and labeling work strictly adheres to the definition of a rumor used in this invention (i.e., information generated and spread among the population whose truth value cannot be confirmed or is intentionally false. It typically arises in emergency situations, easily causing public panic, disrupting social order, reducing government credibility, and even endangering national security). Data lacking factual basis and whose authenticity cannot be determined has been filtered out.

[0100] Furthermore, multidimensional semantic focus is extracted: using the constructed rumor detection dataset, multidimensional semantic focus is extracted for each rumor text to be detected, taking into account two semantic focus exploration approaches: continuous semantic span and metadata indication.

[0101] In some embodiments, the two exploration approaches, namely continuous semantic span and metadata indication, can normally be selectively used in implementation.

[0102] Regarding the continuous semantic span, a syntax parsing tool was used to parse the rumor text into a parse tree, which was then used to segment the rumor text into M distinct semantic subunits f. i l , set as

[0103] Regarding metadata instructions, obtain N pieces of metadata about the rumor text data, including the propagator, publication date, and publication location of the rumor text data. Then, integrate the original rumor text and the aforementioned metadata using a template to obtain N integrated rumor texts. Set as

[0104] Subsequently, the aforementioned semantic subunits and the integrated rumor text Combined into F = [F l ;F r], as the multidimensional semantic focus set of the rumor text Each focus is represented as f. k L = M + N, used for generating subsequent semantic focus issues.

[0105] Furthermore, multidimensional focus questions are generated: based on multidimensional semantic focus, a series of questions related to these focus are generated for each rumor text, and the multidimensional focus questions are generated using a self-generated autoregressive encoder and decoder.

[0106] In some embodiments, the process of generating the multidimensional focus problem is as follows:

[0107] Given a piece of rumor text data T to be detected and its semantic focus set. For each semantic focus f k Combine the rumor text T to generate a sequence with each f k The corresponding focus issue q k This leads to the set of focus issues corresponding to the multidimensional semantic focus set F. Specifically, the text T to be detected and each semantic focus f in the multidimensional semantic focus set F mentioned above will be used. k By concatenating the text, we obtain the concatenated text sequence. The input is then fed into a focus question generation model, which employs a self-generated autoregressive encoder and decoder, including BART and T5 language models. This model calculates the conditional probability distribution of the language model corresponding to the input. This focus-point generation model will be based on the conditional probability distribution of the language model. Generate each semantic focus f k The corresponding focus issue q k Given the rumor text T to be detected and the semantic focus f k This focus question generation model generates focus question q word by word. k The formula is as follows:

[0108]

[0109] Wherein, the left side of the equation represents the text to be detected T and the semantic focus f. k Generate focus q k Language model conditional probability; the right side of the equation represents the focus problem q. k The product of the conditional probabilities generated for each character in the text; For semantic focus f k The sequence of the text T to be detected is concatenated with the input of the focus problem generation model.

[0110] Furthermore, input enhancement is performed through question integration: an integrated prompt template is designed to integrate the original rumor text input to be tested with the generated focus question, thereby semantically enhancing the rumor text input to be tested and obtaining the enhanced rumor text input.

[0111] Preferably, the process of integrating and enhancing the problem input is as follows:

[0112] In integrating the rumor text to be tested with the focus question, the prompt template used must simultaneously meet the following two key conditions: 1. Slot structure: The prompt template should include a slot [SLOT of Rumor] for inserting the rumor text data to be tested, a slot [SLOT of Focus Question] for the focus question, and a slot [SLOT of Label] for outputting candidate answers for the label to be predicted. 2. Evaluation requirements: The template must explicitly require an evaluation of the reasonableness and authenticity of the rumor text to be tested based on the focus question. This evaluation is achieved by predicting candidate answers for the label slot [SLOT of Label], thereby obtaining the final rumor detection truth / false label.

[0113] In some embodiments, multiple forms of prompts can be selected as the integrated prompt template. Here is an example of a typical embodiment: "From the semantic perspective of the question: [SLOT of FocalQ], can the veracity of the question [SLOT of Rumor] be guaranteed? Yes or no? Please answer: [SLOT of Label]."

[0114] After completing the above steps, the integrated input representation Prompt([T:q)) is obtained. k This input, used for subsequent rumor prediction model learning, ensures an effective combination of the focus issue and the test text in the rumor detection task, providing the model with targeted input for more accurate prediction of truthfulness.

[0115] Furthermore, the authenticity of rumors is predicted: a rumor prediction model based on Transformer is constructed, and this model is used to predict the rumor detection label set of the above multiple inputs. The model is then learned and optimized based on the constructed rumor detection dataset.

[0116] In some embodiments, the rumor veracity prediction process is as follows:

[0117] First, a Transformer-based rumor prediction model is established to predict rumor detection labels. This model employs a prompt learning framework based on the T5 generative model and uses backpropagation through a dataset to optimize model parameters.

[0118] This Transformer-based rumor prediction model was used to predict the integrated input representation Prompt([T:q)). k The input consists of multiple sets of rumor detection labels. The label prediction process includes two key stages: language model probability prediction and discrete answer-label alignment. The specific process is as follows:

[0119] 1. Language Model Probabilistic Prediction: The generation probability of each character in the target sequence is obtained using the T5 generative model. This process specifically includes the following steps:

[0120] First, use T5's encoding embedding layer. E The input Prompt([T; q) will be presented as a character sequence. k Embedded into a high-dimensional vector The formula for this process is as follows:

[0121]

[0122] Subsequently, through multiple layers of consecutive Transformer blocks (TransBlock) k Encode the input to obtain a multi-layered deep representation of the initial input. Where N is the number of Transformer layers in the T5 language model, which is 12, as shown in the formula below:

[0123]

[0124] Finally, based on this depth representation The vocabulary probability distribution p(w) of the j-th character in the target sequence is obtained through a fully connected layer containing weights W and biases b and a normalized exponential function SoftMax. j The formula is as follows:

[0125]

[0126] 2. Discrete Answer and Label Alignment: Given a candidate answer and label mapping table Labels of {"Real, Valid, Trustworthy, Accurate, Yes": real; "False, Untrustworthy, False, No": fake}, based on the character probability distribution of the target sequence, calculate and integrate the probability distribution P of the characters corresponding to the label slots [SLOT of Label]. LM (w [SLOT of Label] |[T;q k ], θ), calculate the average cross-entropy with candidate answers. This involves predicting the fit between the prediction and candidate answers to determine the candidate answers. Then, based on the candidate answers, a candidate answer label mapping table is searched and matched to derive the final rumor detection label L mapped to the intermediate answer. goldThe formula for the above process is as follows:

[0127]

[0128]

[0129] Where θ is the set of trainable parameters for the T5 language model, and L gold The final label is obtained through the Labels mapping table, L gold ∈{real,fake}, note L gold It is aimed at a specific focus problem q k The authenticity.

[0130] Relatedly, T5 (Text-to-Text Transfer Transformer) is a generative pre-trained model proposed by Google Research in 2019. Unlike other pre-trained models (such as BERT) that focus on downstream applications for specific tasks, T5 adopts a general text-to-text framework that can solve a variety of NLP tasks through appropriate input and output task descriptions. The core idea of ​​T5 is to treat all NLP tasks as text-to-text problems. Specifically, the input can be text in various forms, and the output is the corresponding text sequence. This allows T5 to handle diverse tasks, including classification, translation, and summarization.

[0131] Overall, T5's versatility makes it a powerful tool for solving various NLP tasks, and its performance on multiple benchmark tasks demonstrates its strong generalization ability.

[0132] Relatedly, prompt learning is a method used in Natural Language Processing (NLP) for model training and inference. Unlike traditional supervised learning and fine-tuning, prompt learning focuses on using task-descriptive prompts to guide the model to perform specific tasks. In prompt learning, the training objective of the model is to generate outputs that match the task description. Compared to fine-tuning, prompt learning is more general because it does not require predefined labels or specific downstream tasks. Instead, by changing the prompts, the model can be easily used for different tasks without additional fine-tuning. This makes prompt learning a more flexible and adaptive method, particularly suitable for multi-task learning or scenarios requiring adaptive processing.

[0133] Overall, cue learning, as an alternative learning paradigm to traditional supervised fine-tuning, emphasizes the use of concise and general task descriptions to guide model learning, thereby providing greater flexibility when applied to different tasks.

[0134] Finally, rumor authenticity verification is integrated: based on multiple sets of rumor detection labels, the confidence weights of the label results are iteratively updated through ensemble learning. In a specific embodiment, we designed and applied a rumor authenticity verification integration strategy based on ensemble learning. In each iteration, the confidence weights of the labels are updated in real time, the best prediction result is selected, and the confidence weights of the label predictions are updated according to the error rate. Finally, the label for the rumor detection task is determined by weighting the confidence weights.

[0135] The specific process of this rumor authenticity integration strategy based on ensemble learning is as follows:

[0136] Initially, the label prediction result L corresponds to each focus question t. t The confidence weight w of the result t They are initially set to be equal, and a weak rumor veracity predictor M is set for each focus issue t. t ;

[0137] In each round of iterative optimization, the error rate is calculated for each specific semantic focus problem t. And based on error rate Update its corresponding weak rumor veracity predictor M t

[0138] Simultaneously update the confidence weight w corresponding to each focus question t. t The details are as follows:

[0139] First, based on the error rate of this focal problem t Calculate the weight balance gradient α t The formula is as follows:

[0140]

[0141] Furthermore, based on the weight balancing gradient α t For the correctly predicted label result L t The confidence weight w corresponding to the focal issue t t Update the confidence weight w t The formula is as follows:

[0142]

[0143] Conversely, based on the weighted balancing gradient α t For the incorrectly predicted label result L t The confidence weight w corresponding to the focal issue t t Update the confidence weight w t The formula is as follows:

[0144]

[0145] Once the confidence weights of all semantic questions have stabilized, a weak rumor veracity predictor M is set up for each of the aforementioned focus questions t. t The confidence weights of all weak rumor predictors are summed by ∑w. t *M t To build the ultimate powerful rumor veracity predictor Right now

[0146] Finally, through Weak rumor veracity predictor M for each focal issue t t The confidence level weights are weighted and summed, and the true / false labels with the higher weighting are used as the final labels for this rumor detection task.

[0147] Preferably, the rumor detection data enhancement device based on multi-dimensional focus question generation described in this invention, in conjunction with this embodiment and appendix, Figure 2 The modular structure of the device is described in detail, including:

[0148] The rumor data acquisition module is used to legally and compliantly crawl a large amount of rumor text data to be detected from major public opinion platforms on the Internet through web crawling technology.

[0149] A rumor detection dataset module is constructed to uniformly select a subset of rumors based on rumor text data, manually label the truth or falsehood of each rumor text to be tested in the rumor subset, and manually summarize multiple semantic focuses and a series of focus issues from different semantic levels of the rumor text to obtain a rumor detection dataset for the learning and optimization of subsequent rumor prediction models.

[0150] The module for extracting multidimensional semantic focus is used to extract a series of multidimensional semantic focus F from different angles and semantic levels of a rumor text T to be detected, based on the obtained rumor detection dataset. In the process of extracting multidimensional semantic focus of the rumor text, two semantic focus exploration approaches are considered: continuous semantic span and metadata indication.

[0151] The module for generating multidimensional focus questions is used to generate a series of focus questions corresponding to the multidimensional semantic focus for each piece of rumor text data;

[0152] The input enhancement module for question integration is used to construct an integrated prompt template. Based on this prompt template, the original rumor text input to be tested and the focus question are integrated, and the semantics of the rumor text input to be tested are enhanced to obtain the enhanced rumor text input.

[0153] The rumor authenticity prediction module is used to build a Transformer-based rumor prediction model. This model predicts multiple sets of input rumor detection labels and learns and optimizes them based on the rumor detection dataset.

[0154] The rumor authenticity integration module is used to iteratively update the confidence weights of the label results based on multiple sets of rumor detection labels through ensemble learning, so as to obtain the final label for the rumor detection task.

[0155] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the rumor detection data enhancement method based on multi-dimensional focus question generation.

[0156] In some embodiments, for the computer program written in Python, mainstream deep learning frameworks such as PyTorch or Tensorflow can be selected as the training framework for the focus question generation model and the Transformer-based rumor prediction model in this method.

[0157] In some embodiments, the processor used to compute the above program can be an Intel Core i9-9900K (desktop processor, 8 cores / threads, 16 threads, clock speed: 3.6GHz, manufacturing process: 14nm), an AMD Ryzen 9 5950X (desktop processor, 16 cores / threads, 32 threads, clock speed: 3.4GHz, manufacturing process: 7nm), an Apple M1 (ARM architecture chip, 8 cores / threads, manufacturing process: 5nm), or other well-known CPU processors;

[0158] In some embodiments, the storage medium used to compute the above-described program can be a Samsung 970EVO Plus NVMe SSD, a Seagate Barracuda 3.5-inch hard drive, a Crucial MX500 SATASSD, or other well-known solid-state or mechanical hard drives.

[0159] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] To reiterate, the above embodiments are merely some examples of the methods of this invention, and not embodiments compatible with all online public opinion analysis and rumor detection scenarios. All other embodiments obtained by those skilled in the art without any inventive step, referring to the embodiments demonstrated in this invention, are also within the scope of protection of this invention.

[0164] Furthermore, the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specific embodiment.

Claims

1. A rumor detection data augmentation method based on multi-dimensional focus issues, characterized in that, Includes the following steps: Obtaining Rumor Data: Legally and compliantly crawling large amounts of rumor text data to be detected from major online public opinion platforms using web scraping technology; Constructing a rumor detection dataset: Based on rumor text data, a subset of rumors is evenly selected, and the authenticity labels of each rumor text to be tested in the subset are manually labeled. Multiple semantic focuses and a series of focus issues are manually summarized from different semantic levels of the rumor text to obtain a rumor detection dataset for the learning and optimization of subsequent rumor prediction models. Extracting multidimensional semantic focus: Based on the obtained rumor detection dataset, given a rumor text data to be tested for authenticity... Extracting a series of multidimensional semantic focuses from different angles and semantic levels of the rumor text. ; In the process of extracting multidimensional semantic focus from rumor text, two semantic focus exploration approaches were considered: continuous semantic span and metadata indication; Generate multidimensional focus questions: Based on multidimensional semantic focus, generate a series of focus questions corresponding to the multidimensional semantic focus for each rumor text data; Input enhancement for question integration: Construct an integrated prompt template, and based on this prompt template, integrate the original rumor text input to be tested and the focus question, and perform semantic enhancement on the rumor text input to be tested to obtain the enhanced rumor text input; Rumor authenticity prediction: Construct a rumor prediction model based on Transformer, predict multiple sets of input rumor detection labels using this model, and learn and optimize based on the rumor detection dataset; The process is as follows: First, a Transformer-based rumor prediction model is constructed for rumor detection and label prediction. This rumor prediction model adopts a prompt learning framework based on the T5 generative model and optimizes the model parameters through backpropagation on the dataset. This rumor prediction model is used to predict the integrated input representation Prompt([ The input consists of multiple sets of rumor detection labels. The label prediction involves two stages: language model probability prediction and discrete answer-label alignment. The specific process is as follows: Language model probabilistic prediction: The generation probability of each character in the target sequence is obtained through this T5 generative model. The specific process is as follows: First, through the T5 encoding embedding layer The input Prompt([) will be a sequence of characters. Embedded into a high-dimensional vector The formula is as follows: Subsequently, through multiple layers of consecutive Transformer blocks Encode the input to obtain a multi-layered deep representation of the initial input. Where N is the number of Transformer layers in the T5 language model, which is 12, and the formula is as follows: Finally, based on this depth representation The vocabulary probability distribution of the j-th character in the target sequence is obtained through a fully connected layer containing weights W and bias b and a normalized exponential function SoftMax. The formula is as follows: Discrete Answer and Label Alignment: Given a candidate answer and label mapping table Labels of {"True and Valid, Credible, Accurate Claim, Yes": real; "False Claim, Unreliable, False Claim, No": fake}, calculate and integrate the probability distribution of the characters corresponding to the label slots [SLOT of Label] based on the character probability distribution of the target sequence. Calculate the average cross-entropy with candidate answers. This involves predicting the fit between the intermediate answer and the candidate answer, determining the candidate answer, and then, based on the candidate answer, searching the candidate answer label mapping table and matching it to obtain the final rumor detection label mapped to the intermediate answer. The specific formula is as follows: in, It is the set of trainable parameters for the T5 language model. The final label is obtained through the Labels mapping table. ,Notice It is aimed at specific key issues. The authenticity; Rumor Authenticity Integration: Based on multiple sets of rumor detection labels, the confidence weights of the label results are iteratively updated through ensemble learning to obtain the final labels for the rumor detection task.

2. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, In the process of acquiring the text data of rumors to be detected, an adaptive dynamic crawling strategy is adopted, which includes the following features and elements: adaptive crawling frequency adjustment, real-time monitoring and hot spot tracking, content priority setting, and automated anomaly detection and response.

3. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, In the process of labeling the rumor detection dataset, a collaborative labeling mechanism was introduced, allowing different labelers to communicate and discuss during the labeling process. In addition, a feedback loop and expert review mechanism were introduced, incorporating a feedback loop into the labeling process. Labelers can review the labeled data and provide feedback on the labeling results of others, forming a positive feedback loop. After the annotations were completed, experts in the field of linguistics were invited to conduct a final review, and the annotation results were randomly checked and verified in a certain proportion to ensure the high quality of the rumor detection dataset.

4. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, In the process of extracting multidimensional semantic focus from rumor texts, two major semantic focus exploration approaches were considered: continuous semantic span and metadata indication, as detailed below: Regarding the continuous semantic span, a syntax parse tree of the rumor text was generated using a syntax parsing tool, and the parse tree was then segmented into M distinct semantic subunits. , set as ; Regarding metadata instructions, obtain N pieces of metadata about the rumor text data, including the propagator, publication date, and publication location of the rumor text data. Then, integrate the original rumor text and metadata using a template to obtain N integrated rumor texts. , set as ; Subsequently, semantic subunits and the integrated rumor text Integrate into As a set of multidimensional semantic focal points of the rumor text Each focus is represented as , This is used for generating subsequent semantic focus questions.

5. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, The process of generating multidimensional focus problems is as follows: Given a piece of rumor text data to be detected and its semantic focus set For each semantic focus Combined with the text of the rumor Generate with each Corresponding key issues This leads to the acquisition of a multidimensional semantic focus set. The corresponding set of key issues ; Specifically, the text to be detected and multidimensional semantic focus set Each semantic focus in By concatenating the text, we obtain the concatenated text sequence. The input is then fed into a focus question generation model, which employs a self-generated autoregressive encoder and decoder, including BART and T5 language models. This focus question generation model calculates the conditional probability distribution of the language model corresponding to the input. This focus-point generation model is based on the conditional probability distribution of the language model. Generate each semantic focus Corresponding key issues ; Given the rumor text to be detected Semantic focus This focus question generation model generates focus questions word by word. The formula is as follows: Wherein, the left side of the equation represents the text to be detected. and semantic focus Generate focus Language model conditional probability; the right side of the equation represents the focus problem. The product of the conditional probabilities generated for each character in the text; semantic focus and the text to be detected The sequence concatenation is the input to the focus problem generation model.

6. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, The input enhancement for question integration is as follows: First, an integrated prompt template is constructed. Based on the constructed prompt template, the original rumor text input to be tested and the focus question are integrated to obtain the semantically enhanced rumor text to be tested. The specific process is as follows: The prompt template used to integrate the rumor text to be tested with the key questions must meet the following conditions: The system includes a slot for the rumor text data to be tested (SLOT of Rumor), a slot for the focus question (SLOT of FocalQ), and a slot for the label (SLOT of Label) to output candidate answers for the label to be predicted. The prompt template assigns the following requirements: Based on the focus question, evaluate the rationality and authenticity of the rumor text to be tested, that is, predict candidate answers for the label slot [SLOT of Label], and then obtain the final rumor detection truth or falsehood label; After integrating the rumor text to be tested with the focus issue, the integrated input representation Prompt([) is obtained. This is used as input for subsequent rumor prediction model learning.

7. The rumor detection data enhancement method based on multi-dimensional focus question generation according to claim 1, characterized in that, The rumor veracity assessment process employs an ensemble learning-based strategy, the specific steps of which are as follows: Initial label prediction results for each focus question t The confidence weight of the result They are initially set to be equal, and a weak rumor veracity predictor is set for each focus issue t. ; In each round of iterative optimization, the error rate is calculated for each specific semantic focus problem t. And based on error rate Update its corresponding weak rumor veracity predictor And at the same time update the confidence weight corresponding to each focus issue t. The details are as follows: First, based on the error rate of this focal problem t Calculate the weighted equilibrium gradient The formula is as follows: Furthermore, based on the weighted gradient For correctly predicted label results The confidence weight corresponding to the focal issue t Update the confidence weight. The formula is as follows: Conversely, based on weighted gradient balancing For incorrectly predicted label results The confidence weight corresponding to the focal issue t Update the confidence weight. The formula is as follows: Once the confidence weights of all semantic questions have stabilized, a weak rumor veracity predictor is set up for each focus question t. The confidence weights of all weak rumor veracity predictors are summed. To build the ultimate powerful rumor veracity predictor ,Right now ; Finally, through Weak rumor veracity predictor for each focal issue t The confidence level weights are weighted and summed, and the true / false labels with the higher weighting are used as the final labels for this rumor detection task.

8. A rumor detection data enhancement device based on multi-dimensional focus issues, characterized in that: The device is based on the rumor detection data enhancement method based on multi-dimensional focus question generation as described in claim 1, and includes: The rumor data acquisition module is used to legally and compliantly crawl a large amount of rumor text data to be detected from major public opinion platforms on the Internet through web crawling technology. A rumor detection dataset module is constructed to uniformly select a subset of rumors based on rumor text data, manually label the truth or falsehood of each rumor text to be tested in the rumor subset, and manually summarize multiple semantic focuses and a series of focus issues from different semantic levels of the rumor text to obtain a rumor detection dataset for the learning and optimization of subsequent rumor prediction models. The multidimensional semantic focus extraction module is used to extract the true or false semantic focus of a given rumor text data to be detected, based on the obtained rumor detection dataset. Extracting a series of multidimensional semantic focuses from different angles and semantic levels of the rumor text. In the process of extracting multidimensional semantic focus from rumor texts, two semantic focus exploration approaches were considered: continuous semantic span and metadata indication. The module for generating multidimensional focus questions is used to generate a series of focus questions corresponding to the multidimensional semantic focus for each piece of rumor text data; The input enhancement module for question integration is used to construct an integrated prompt template. Based on this prompt template, the original rumor text input to be tested and the focus question are integrated, and the semantics of the rumor text input to be tested are enhanced to obtain the enhanced rumor text input. The rumor authenticity prediction module is used to build a Transformer-based rumor prediction model. This model predicts multiple sets of input rumor detection labels and learns and optimizes them based on the rumor detection dataset. The rumor authenticity integration module is used to iteratively update the confidence weights of the label results based on multiple sets of rumor detection labels through ensemble learning, so as to obtain the final label for the rumor detection task.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the rumor detection data augmentation method based on multi-dimensional focus problem generation as described in any one of claims 1-7.

Images