Misinformation AI complaint and multimedia clarification publishing system

Abstract

The application discloses a false information AI complaint and fusion media clarification publishing system, which comprises a propagation space-time graph construction module, an interactive data collection module, a propagation graph construction module, a forwarding level module, a sentiment polarity score module, a stay time length combination weight module, a weighted summation calculation node susceptible index module, a core fermentation node division module, an edge propagation node division module, a susceptible latent node division module, a rumor-busting targeting problem solving module, a cross-modal clarification content adaptive generation module, a dominant propagation mode analysis module, a structured evidence chain combination AI complaint module, a cross-modal clarification content package generation module, a content form and channel mismatch problem solving module, a targeted clarification publishing and dynamic hedging module, a different modal content package accurate pushing module, a user group with different degrees of misinformation, a feedback monitoring module and a secondary hedging module. The application belongs to the field of internet information processing and artificial intelligence technology.

Description

Technical Field

[0001] This invention relates to the field of network data mining technology. Background Technology

[0002] With the rapid development of converged media technologies, information dissemination is characterized by multi-source, cross-modal, and explosive growth, and the spread of misinformation has also generated serious social impacts. Existing systems for handling complaints and clarifications regarding misinformation have the following significant shortcomings: 1. Rumor debunking distribution lacks targeting. Existing systems typically use full broadcast to release debunking information, which fails to identify which users are deeply affected by rumors. This results in debunking information failing to effectively counteract false information in the same time and space, leading to low debunking efficiency and potentially disturbing uninformed users.

[0003] 2. Mismatch between the format of the clarification content and the channels of dissemination. The clarification content generated by the existing system is mostly in the form of single text. When rumors spread in the form of short videos, the reach and acceptance of text-based debunking are extremely low, and it is difficult to adapt to the format specifications of different integrated media platforms. This results in a serious mismatch between the format of the debunking content and the channels of dissemination, and poor audience acceptance.

[0004] Therefore, there is an urgent need to develop an AI-powered appeal and integrated media clarification system for disseminating false information to address the problems in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-powered appeal and integrated media clarification system for disseminating false information. This system can achieve targeted hedging distribution by constructing a dissemination graph and calculating node susceptibility indices. It can also generate cross-modal content adapted to different platforms by analyzing the dominant modality and combining evidence chains to achieve multiple forms of a single draft. The system is simple in structure and easy to use, thus solving the problems of lack of targeting in the dissemination of false information and the mismatch between the form of the clarification content and the dissemination channels in the aforementioned background technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A system for handling false information using AI and for clarifying information through integrated media platforms, comprising: The propagation spatiotemporal graph construction module collects interactive data of target false information from multi-source converged media platforms, constructs a propagation graph, and calculates the node susceptibility index by combining forwarding level, sentiment polarity score and dwell time with their respective weights based on the weighted summation principle. The propagation nodes are divided into core fermentation nodes, peripheral propagation nodes and susceptible latent nodes, which solves the problem of lack of targeting in the distribution of debunking. The cross-modal clarification content adaptive generation module analyzes the dominant dissemination modality of the target's false information, combines it with the structured evidence chain passed by AI appeals, and generates a cross-modal clarification content package that adapts to the release specification parameters of the target's converged media platform, thus solving the problem of mismatch between the form of clarification content and the dissemination channel; The targeted clarification and dynamic hedging module matches the converged media content package with nodes in the dissemination map. Based on the node susceptibility index, it accurately pushes converged media content packages of different modalities to user groups with different degrees of misleading. It also uses a sentiment analysis model to calculate the negative sentiment decline rate and monitor feedback effects for secondary hedging.

[0007] By adopting the above technical solutions, the system can accurately target and refute false information and provide multimodal adaptation and clarification. It can construct a propagation map to identify susceptible nodes, adaptively generate cross-modal content, and accurately push and dynamically offset based on node matching, thus comprehensively solving the problems of lack of targeting in the distribution of false information and mismatch in content format.

[0008] As a further aspect of the present invention, the propagation spatiotemporal map construction module performs the following steps: S1: Data collection and preprocessing. Through the application programming interfaces (APIs) of various converged media platforms, collect forwarding records, comment records, and like records of false information, and perform deduplication and removal of invalid special characters on the collected text content. S2: Propagation graph construction, using users and false information as nodes, and forwarding, commenting and liking behaviors as directed edges, constructing a directed acyclic graph, with the direction of the edges pointing from the behavior initiator to the behavior receiver, and assigning a timestamp attribute to each node, forming a propagation graph containing spatiotemporal attributes; S3: Node Susceptibility Index Calculation. Based on the user's behavioral characteristics in the propagation graph, the node susceptibility index is calculated. Based on the calculated node susceptibility index value, the propagation nodes are divided into core fermentation nodes, edge propagation nodes, and susceptible latent nodes.

[0009] By adopting the above technical solutions, the construction of a spatiotemporal propagation map and the function of node classification are realized. By collecting interactive data, a directed acyclic graph is constructed and a susceptibility index is calculated, providing a precise data foundation and a structured node division basis for the targeted distribution of subsequent clarification content.

[0010] As a further aspect of the present invention: In S3, the node susceptibility index is calculated based on the weighted summation principle by combining forwarding level, sentiment polarity score, dwell time, forwarding level weight, sentiment polarity weight, and dwell time weight. The forwarding level is the shortest path length from the current node to the source node of the false information, the sentiment polarity score is the sentiment polarity score of the user's comment, the dwell time is the dwell time of the user on the false information page, and the sum of the forwarding level weight, sentiment polarity weight, and dwell time weight is a preset constant.

[0011] By adopting the above technical solution, the node susceptibility index can be quantitatively calculated. By integrating forwarding level, emotional polarity, dwell time and their respective weights, a weighted sum is performed to scientifically and accurately measure the depth to which users are misled by false information.

[0012] As a further aspect of the present invention: the cross-modal clarification content adaptive generation module performs the following steps: S4: Dominant modality parsing, extracts multimodal features from the target false information, inputs it into a pre-trained modality classification model, and outputs the dominant propagation modality of the false information. The dominant propagation modality includes video-dominant, image-text-dominant, and plain text-dominant. S5: Evidence chain injection and cross-modal content generation. Obtain the structured evidence chain output by the AI ​​appeal module. This structured evidence chain contains factual data and authoritative source tracing links. Based on the dominant dissemination modality, call the corresponding generation interface to generate the initial draft of the clarification content. S6: Platform adaptation and trimming. Obtain the publishing specification parameters of the target integrated media platform, trim and format the initial draft of the clarification content according to the publishing specification parameters, and generate an integrated media content package.

[0013] By adopting the above technical solution, the adaptive generation function of cross-modal clarification content is realized. By parsing the dominant modality, injecting evidence chain to generate the initial draft, and performing platform adaptation and trimming, the final generated integrated media content package is ensured to both conform to the form of rumor dissemination and comply with platform specifications.

[0014] As a further aspect of the present invention: In S4, the training process of the modality classification model includes: collecting online posts containing videos, images, and plain text as training samples; labeling the dominant propagation modality of each sample; extracting image features using a residual network, extracting video features using a three-dimensional convolutional network, and extracting text features using a bidirectional encoding representation model; concatenating the above features and inputting them into a fully connected layer for classification; using a cross-entropy loss function and an optimizer for backpropagation training until the loss function converges to obtain the modality classification model.

[0015] By adopting the above technical solution, the training function of the modality classification model is realized. Through multi-dimensional feature extraction and splicing classification training, the model is able to accurately identify the dominant mode of false information dissemination.

[0016] As a further aspect of the present invention: In S5, the fine-tuning training process of the large language model includes: constructing a debunking-specific dataset, which contains pairs of false information and corresponding professional debunking articles; using a structured evidence chain as the conditional input and professional debunking articles as the target output; and using a causal language model loss function for instruction fine-tuning training, so that the large language model generates clarification text based on the evidence chain.

[0017] By adopting the above technical solution, the debunking fine-tuning function of the large language model is realized. By constructing a dedicated dataset and fine-tuning the instructions based on the evidence chain, the model can generate logically rigorous clarification text based on factual data.

[0018] As a further aspect of the present invention: In S5, if the dominant dissemination mode is video-dominated, the digital human video generation interface is called to convert factual data into a digital human broadcast script and generate a clarification short video containing an authoritative source link overlay; if the dominant dissemination mode is text-based, the large language model is called to generate a clarification long text, and the data visualization interface is called to render the factual data into charts and generate a clarification long image; if the dominant dissemination mode is plain text-dominated, the large language model is called to generate structured clarification text.

[0019] By adopting the above technical solutions, differentiated content generation functions for different modalities can be realized. Based on the dominant modalities such as video, images and text, and plain text, digital human, visualization or text generation interfaces are called respectively to ensure that the form of the clarification content is highly matched with the original form of rumor dissemination.

[0020] As a further aspect of the present invention: in S6, the published specification parameters include the upper limit of video duration, the upper limit of text and image length, and the upper limit of the number of images; for short videos exceeding the duration limit, keyframe extraction and audio acceleration processing are performed; for text and image length exceeding the length limit, segmented summary extraction is performed.

[0021] By adopting the above technical solution, the platform-adaptive cropping function of the clarification content is realized. The initial draft is cropped and extracted according to the standard parameters such as video length and text and image length, so as to ensure that the clarification content is published and displayed in a compliant manner on various media convergence platforms.

[0022] As a further aspect of the present invention, the targeted clarification release and dynamic hedging module performs the following steps: S7: Targeted matching establishes a mapping relationship between node susceptibility index and converged media content package modality. If a node belongs to the core fermentation node, it is matched with a deep text and image content package. If a node belongs to the peripheral propagation node or a susceptible latent node, it is matched with a short video content package. S8: Precise push, through the application programming interface of the converged media platform, pushes the matched converged media content package to the corresponding user node; S9: Hedging feedback and secondary triggering. Within a preset time window after the release of the converged media content package, real-time data on the interaction of target user nodes with the clarification content is collected. The negative sentiment decline rate is calculated through the sentiment analysis model. If the negative sentiment decline rate is lower than the preset decline rate threshold, the cross-modal clarification content adaptive generation module is triggered to change the modality and regenerate supplementary clarification content, and a secondary targeted push is performed.

[0023] By adopting the above technical solutions, the system can achieve targeted push and dynamic hedging of clarification content. It can accurately push content by establishing a mapping relationship between susceptibility index and content modality, and trigger secondary generation based on feedback effects to form a closed loop for debunking rumors.

[0024] As a further aspect of the present invention: in S9, the negative emotion reduction rate is obtained based on the difference calculation principle by comparing the emotional polarity score of the user before receiving the clarification information with the emotional polarity score after receiving the clarification information.

[0025] By adopting the above technical solution, the function of quantitatively calculating the rate of decline in negative emotions can be realized. By comparing the difference in emotional polarity scores before and after receiving clarification information, an accurate numerical judgment standard can be provided for evaluating the clarification hedging effect and triggering secondary push.

[0026] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a spatiotemporal propagation map and calculates node susceptibility index to accurately push multimedia content packages of different modalities to user groups with different degrees of misleadingness. This enables targeted distribution of debunking information, significantly improves clarification efficiency, and reduces the disturbance of invalid information.

[0027] 2. This invention generates cross-modal clarification content adapted to the target platform by analyzing the dominant dissemination modality of false information and combining it with the evidence chain. It can achieve accurate matching of one draft to multiple modalities, solve the problem of content form mismatch, and greatly improve audience acceptance and readability of debunking.

[0028] 3. This invention calculates the negative sentiment reduction rate by real-time monitoring and feedback and triggers secondary generation and push, which can form a dynamic closed loop for debunking rumors, effectively respond to rumor mutations and continuously suppress false information.

[0029] 4. By specifically training and fine-tuning the modality classification model and the large language model, this invention can ensure the accuracy of rumor modality recognition and the professional rigor of clarification text generation.

[0030] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall structure of an AI-powered appeal and integrated media clarification system for false information, as described in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] In this embodiment of the invention, an AI-powered system for appealing and clarifying false information through integrated media is provided. (See also...) Figure 1 As shown, it includes: a propagation spatiotemporal map construction module, a cross-modal clarification content adaptive generation module, and a targeted clarification release and dynamic hedging module.

[0034] S1: Collect interactive data from multi-source converged media platforms through the propagation spatiotemporal graph construction module, construct a directed acyclic graph and quantify the node susceptibility index, and accurately divide the propagation nodes into core fermentation nodes, edge propagation nodes and susceptible latent nodes; S2: The cross-modal clarification content adaptive generation module analyzes the dominant dissemination modality of false information, combines the structured evidence chain passed by AI appeal, calls the corresponding generation interface and the fine-tuned large language model to generate the initial draft of the clarification content, and trims and adapts it according to the release specification parameters of the target media convergence platform to generate a media convergence content package. S3: By establishing a mapping relationship between node susceptibility index and converged media content package modality through targeted clarification release and dynamic hedging modules, converged media content packages of different modalities are accurately pushed to user groups with different degrees of misleading. At the same time, the feedback effect is monitored and the negative sentiment reduction rate is calculated. If it is lower than the threshold, secondary generation and push are triggered to form a dynamic hedging closed loop.

[0035] The system also includes an AI appeal module, which receives appeal requests from users or administrators, automatically verifies the authenticity of information, and generates a structured chain of evidence, specifically including: Appeal Receiving Unit: Used to receive appeal requests submitted by users or administrators. The appeal requests are submitted through a preset appeal entry point, such as the "Appeal" button in the integrated media platform interface, the API interface in the administrator backend, or the interface for connecting to a third-party platform. The request format is structured data, such as JSON, containing the following fields: Appeal Identifier, a unique ID; Identifier of the target false information, such as the URL or ID of the rumor; Reason for Appeal, a description of the user's objection to the rumor; Evidence materials submitted by the user, such as screenshots or text descriptions. Authenticity Verification Unit: Used to automatically verify the authenticity of appeal information. The authenticity verification unit connects to a third-party fact-checking database, such as government open data platforms, databases of authoritative news organizations, and databases of industry regulatory agencies. It verifies the appeal information through a cross-verification mechanism, which includes: extracting key elements from the appeal information, such as the time, location, involved parties, and nature of the event, and comparing them with corresponding facts in the third-party database, such as whether the time matches, whether the location is consistent, and whether the parties are real. If the data from multiple sources (at least 2 independent sources) are consistent, the appeal information is determined to be authentic. Evidence Chain Generation Unit: Used to generate a structured evidence chain containing factual data and authoritative source links. The structured evidence chain is in JSON format and includes the following fields: factual statement, verified true information, such as "a certain beverage ingredient was tested by the State Administration for Quality Supervision, Inspection and Quarantine and no carcinogenic substances were detected"; authoritative source link, access link to a third-party database, such as a link to the test report on the official website of the State Administration for Quality Supervision, Inspection and Quarantine; verification result, such as "verified and passed". The AI ​​appeal module and the cross-modal clarification content adaptive generation module interact through an API interface. The cross-modal clarification content adaptive generation module calls the interface of the AI ​​appeal module, such as getEvidenceChain(appeal ID), to obtain a structured evidence chain for generating a draft of the clarification content.

[0036] Example 1 This embodiment provides a specific execution flow for an AI-based appeal and integrated media clarification system for disseminating false information. The system includes the following specific execution steps: S1: Data Acquisition and Preprocessing; The Spatiotemporal Map Construction Module collects forwarding records, comment records, and like records of false information from various converged media platforms through the application programming interfaces (APIs); it also performs deduplication and cleaning operations to remove invalid special characters from the collected text content.

[0037] For example, in response to a false information circulating on a short video platform about a major chemical explosion in a certain place, the system collected 5,000 related forwarding records, 12,000 comment records, and 80,000 like records, and cleaned and filtered out emoticons and garbled characters in the comments.

[0038] S2: Construction of the propagation graph; using users and false information as nodes, and forwarding, commenting, and liking behaviors as directed edges, construct a directed acyclic graph.

[0039] The edges point from the initiator of the action to the receiver of the action, and each node is assigned a timestamp attribute, forming a propagation graph that includes spatiotemporal attributes.

[0040] For example, if user A forwards false information from the source node, a directed edge is constructed from user A to the source node, and the time of the forwarding action is recorded as the timestamp attribute of user A's node.

[0041] S3: Node Susceptibility Index Calculation; Calculate the node susceptibility index based on the user's behavioral characteristics in the propagation graph.

[0042] The formula for calculating the node susceptibility index is:

[0043] in, Indicates the node susceptibility index; This represents the shortest path length from the current node to the node containing the false information, i.e., the forwarding level. This represents the sentiment polarity score of the user's comment. The score is obtained through a sentiment analysis model fine-tuned based on RoBERTa (a pre-trained language model), and ranges from 0 to 1. A higher score indicates a higher degree of agreement with the false information. The sentiment analysis model is fine-tuned using a dataset that includes stance detection, and categorizes user comments into three types: supporting the rumor, refuting the rumor, and neutral. Specifically, the model outputs probability values ​​that support rumor categories, thereby accurately quantifying the degree to which users are misled; This indicates the length of time a user stays on the page displaying the false information. This indicates the maximum duration of stay within the statistical period; , , These are weights for forwarding level, sentiment polarity, and dwell time, respectively, and satisfy the following conditions: .

[0044] In this embodiment, the following settings are provided: , , .

[0045] If a user directly forwards the message from the source node, then The comments, calculated using a sentiment analysis model, showed a very high level of agreement. Duration of stay Seconds, within the statistical period The node susceptibility index for that user is calculated in seconds. .

[0046] Based on calculations The value is used to divide the propagation nodes into core fermentation nodes. Edge propagation nodes and susceptible latent nodes The above Users with a score of 0.83 were designated as core fermentation nodes.

[0047] This also includes AI-generated appeals, specifically the following: The AI-powered appeal module receives appeal requests and generates a chain of evidence. The AI ​​appeal module's appeal receiving unit receives appeal requests submitted by users. These requests include the ID of the target false information (such as the rumor ID of "a well-known beverage contains carcinogenic industrial raw materials") and preliminary evidence provided by the user (such as a screenshot of the beverage's ingredient list). The authenticity verification unit accesses the database of the State Administration for Quality Supervision, Inspection and Quarantine (AQSIQ) to extract key elements ("carcinogenic industrial raw materials") from the complaint information and compares them with AQSIQ's "List of Prohibited Raw Materials" to confirm that the raw material is not on the prohibited list. At the same time, it accesses the database of authoritative media (such as Xinhua News Agency) to verify the beverage's test report and confirm that the report conclusion is "qualified". The evidence chain generation unit generates a structured evidence chain in JSON format. { Statement of Facts: "The ingredients of this beverage were tested by the State Administration for Quality Supervision, Inspection and Quarantine, and no carcinogenic substances were detected." "Authoritative source link": ["https:.......html", "https:.......htm"], Verification result: "Verification passed" } The cross-modal clarification content adaptive generation module calls the interface of the AI ​​appeal module, getEvidenceChain(appealID), to obtain the above-mentioned structured evidence chain, which is used to generate the initial draft of the clarification content.

[0048] S4: Dominant Modality Analysis; The cross-modal clarification content adaptive generation module extracts multimodal features from the target misinformation, inputs them into a pre-trained modality classification model, and outputs the dominant propagation modality of the misinformation. The dominant propagation modality includes video-dominant, image-text-dominant, and plain text-dominant modes.

[0049] The training process of the modality classification model includes: collecting 500,000 online posts containing videos, images, and plain text as training samples; manually labeling the dominant propagation modality of each sample; extracting image features using a ResNet-50 residual network, extracting video features using a 3D convolutional network (3D ResNet), and extracting text features using a bidirectional encoding representation model (BERT); concatenating the above three types of features and inputting them into a fully connected layer for classification; using the cross-entropy loss function and the Adam optimizer (an optimizer algorithm) for backpropagation training, setting the initial learning rate to 0.001, the batch size to 64, and iterating 100 times until the loss function converges, thus obtaining the modality classification model. For short video rumors in this embodiment, the modality classification model outputs that its dominant propagation modality is video-dominated.

[0050] S5: Evidence chain injection and cross-modal content generation; Obtain the structured evidence chain output by the AI ​​appeal module, which contains factual data and authoritative source tracing links; Based on the dominant dissemination modality, call the corresponding generation interface to generate the initial draft of the clarification content.

[0051] The fine-tuning training process of the large language model includes: constructing a dedicated debunking dataset containing 100,000 pairs of false information and corresponding professional debunking articles; using structured evidence chains as input and professional debunking articles as output; and employing low-rank adaptive fine-tuning techniques based on Llama-3-8B (a large language model), using a causal language model loss function for instruction fine-tuning training, setting the learning rate to 0.00005, so that the large language model can generate logically rigorous and easy-to-understand clarification texts based on the evidence chains.

[0052] For video-driven rumors in this embodiment, the digital human video generation interface is invoked. This interface has a built-in speech synthesis model based on VITS (a speech synthesis model) and a lip-syncing model based on Wav2Lip (a lip-syncing model). The factual data generated by the large language model is converted into a digital human broadcast script, which is then converted into audio through speech synthesis. The digital human face is then driven to generate a clarifying short video containing an authoritative source traceability link overlay.

[0053] If the dominant communication modality is text-based, a large language model is used to generate a clarification text, and a data visualization interface based on Echarts (a data visualization library) is used to render the factual data into a chart, generating a clarification chart. If the dominant communication modality is plain text-based, a large language model is used to generate a structured clarification text.

[0054] S6: Platform adaptation and cropping; obtain the publishing specifications parameters of the target converged media platform, which include the upper limit of video duration, the upper limit of text and image length, and the upper limit of the number of images.

[0055] The initial draft of the clarification content was trimmed and formatted according to the release specifications to generate a multimedia content package.

[0056] In this embodiment, the maximum video length of a certain short video platform is 60 seconds. If the generated clarification short video draft is 90 seconds, then the short video will be processed by extracting keyframes and speeding up the audio by 1.5 times to make it conform to the 60-second maximum length. For images and texts that exceed the maximum length limit, segmented summary extraction will be performed.

[0057] S7: Targeted matching; the targeted clarification release and dynamic hedging module establishes a mapping relationship between node susceptibility index and converged media content package modality.

[0058] If the node belongs to the core fermentation node, then a deep text and image content package is matched; if the node belongs to the edge propagation node or the susceptible latent node, then a short video content package is matched.

[0059] In this embodiment, for the core fermentation node users calculated above, the system matches them with a deep graphic content package containing detailed data charts; for the edge propagation node users, it matches them with a 60-second clarification short video content package that has been accelerated.

[0060] S8: Precise push; Through the application programming interface of the converged media platform, the matched converged media content package is pushed to the corresponding user node.

[0061] For core fermentation nodes, a strategy combining targeted private messages and replies to the original post is adopted; for peripheral propagation nodes and susceptible latent nodes, a strategy of inserting information streams with the same tag is adopted, inserting clarification short videos into the video streams with the same tag that the user subsequently browses.

[0062] S9: Hedging feedback and secondary triggering; within a preset time window after the release of the converged media content package, such as 24 hours after release, real-time collection of interaction data of target user nodes on the clarification content.

[0063] The negative sentiment decline rate is calculated using a sentiment analysis model. The formula is as follows:

[0064] in, Indicates the rate of decrease in negative emotions. This represents the user's sentiment polarity score before receiving clarification information. This represents the user's emotional polarity score after receiving clarification information.

[0065] like If the rate of decline is lower than the preset threshold (set to 0.2 in this embodiment), the hedging is deemed to have failed, triggering the cross-modal clarification content adaptive generation module to change the modality and regenerate supplementary clarification content, and then perform a second targeted push.

[0066] Example 2 The technical feature that distinguishes this embodiment from Embodiment 1 is that this embodiment uses a piece of false information dominated by images and text, such as "a well-known beverage contains carcinogenic industrial raw materials," as an example to illustrate the specific execution logic of the system.

[0067] In S3, the system calculates that a user forwarded the image and text rumor at a relatively deep level. The comments expressed slight concern. The stay is relatively short. Second, Seconds, weight settings are the same as in Example 1 Then its node susceptibility index The user was classified as an edge propagation node.

[0068] In S4, the modality classification model identifies that the rumor contains a large number of out-of-context images and inflammatory text, and outputs that the dominant dissemination modality is image-text dominant.

[0069] In S5, the system obtains the evidence chain for the successful AI appeal, including the test report data and official statement link from the State Administration for Quality Supervision, Inspection and Quarantine. It calls the large language model to generate a clarification text and calls the data visualization interface to render the various indicator data in the test report into bar charts, generating a draft of a clarification text with illustrations.

[0070] In S6, the system retrieves the posting specifications of a social media platform, requiring that the maximum number of images posted at one time be 9, and that the text in each image be clearly visible.

[0071] The system automatically slices and lays out the generated long image draft to create a multimedia content package that meets the limit of 9 images.

[0072] In S7 and S8, the node susceptibility index was used. For users at the edge propagation node with a score of 0.39, the system matches short video content packages, converts the core data in the text and image content packages into short videos through the digital human interface, and pushes them using a strategy of inserting information streams with the same tags to avoid excessively disturbing users.

[0073] In S9, the system detected a decrease in negative emotions among some core fermentation node users after receiving in-depth graphic content packages. With a value of only 0.15, below the threshold of 0.2, it indicates that the clarification of the text and images failed to completely eliminate their doubts.

[0074] The system triggers the cross-modal clarification content adaptive generation module, which regenerates supplementary clarification content, including expert interpretation videos, for these core node users, and then pushes it to them in a second targeted manner through targeted private messages until the negative sentiment reduction rate meets the threshold requirement.

[0075] This invention provides an AI-powered appeal and integrated media clarification system for disseminating false information. It can achieve targeted hedging distribution by constructing a dissemination graph and calculating node susceptibility indices. It can also generate cross-modal content adapted to different platforms by analyzing the dominant modality and combining evidence chains to achieve multiple forms of a single draft.

[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0077] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0078] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A system for handling false information using AI and for clarifying and releasing information through integrated media, characterized in that: include: The propagation spatiotemporal graph construction module collects interactive data of target false information from multi-source converged media platforms, constructs a propagation graph, and calculates the node susceptibility index by combining forwarding level, sentiment polarity score and dwell time with their respective weights based on the weighted summation principle. The propagation nodes are divided into core fermentation nodes, peripheral propagation nodes and susceptible latent nodes, which solves the problem of lack of targeting in the distribution of debunking. The cross-modal clarification content adaptive generation module analyzes the dominant dissemination modality of the target's false information, combines it with the structured evidence chain passed by AI appeals, and generates a cross-modal clarification content package that adapts to the release specification parameters of the target's converged media platform, thus solving the problem of mismatch between the form of clarification content and the dissemination channel; The targeted clarification and dynamic hedging module matches the converged media content package with nodes in the dissemination map. Based on the node susceptibility index, it accurately pushes converged media content packages of different modalities to user groups with different degrees of misleading. It also uses a sentiment analysis model to calculate the negative sentiment decline rate and monitor feedback effects for secondary hedging.

2. The AI-based complaint and integrated media clarification system for false information as described in claim 1, characterized in that, The propagation spatiotemporal map construction module performs the following steps: S1: Data collection and preprocessing. Through the application programming interfaces (APIs) of various converged media platforms, collect forwarding records, comment records, and like records of false information, and perform deduplication and removal of invalid special characters on the collected text content. S2: Propagation graph construction, using users and false information as nodes, and forwarding, commenting and liking behaviors as directed edges, constructing a directed acyclic graph, with the direction of the edges pointing from the behavior initiator to the behavior receiver, and assigning a timestamp attribute to each node, forming a propagation graph containing spatiotemporal attributes; S3: Node Susceptibility Index Calculation. Based on the user's behavioral characteristics in the propagation graph, the node susceptibility index is calculated. Based on the calculated node susceptibility index value, the propagation nodes are divided into core fermentation nodes, edge propagation nodes, and susceptible latent nodes.

3. The AI-based complaint and integrated media clarification system for false information as described in claim 2, characterized in that, In S3, the node susceptibility index is calculated based on the weighted summation principle by combining forwarding level, sentiment polarity score, dwell time, forwarding level weight, sentiment polarity weight, and dwell time weight. The forwarding level is the shortest path length from the current node to the source node of the false information, the sentiment polarity score is the sentiment polarity score of the user's comment, and the dwell time is the duration of the user's stay on the false information page. The sum of the forwarding level weight, sentiment polarity weight, and dwell time weight is a preset constant.

4. The AI-based complaint and integrated media clarification system for false information as described in claim 1, characterized in that, The cross-modal clarification content adaptive generation module performs the following steps: S4: Dominant modality parsing, extracts multimodal features from the target false information, inputs it into a pre-trained modality classification model, and outputs the dominant propagation modality of the false information. The dominant propagation modality includes video-dominant, image-text-dominant, and plain text-dominant. S5: Evidence chain injection and cross-modal content generation. Obtain the structured evidence chain output by the AI ​​appeal module. This structured evidence chain contains factual data and authoritative source tracing links. Based on the dominant dissemination modality, call the corresponding generation interface to generate the initial draft of the clarification content. S6: Platform adaptation and trimming. Obtain the publishing specification parameters of the target integrated media platform, trim and format the initial draft of the clarification content according to the publishing specification parameters, and generate an integrated media content package.

5. The AI-based complaint and integrated media clarification system for false information as described in claim 4, characterized in that, S4 includes the training process of the modality classification model, which includes: collecting online posts containing videos, images, and plain text as training samples; labeling each sample with the dominant propagation modality label; extracting image features using a residual network, extracting video features using a three-dimensional convolutional network, and extracting text features using a bidirectional encoding representation model; concatenating the above features and inputting them into a fully connected layer for classification; using the cross-entropy loss function and an optimizer for backpropagation training until the loss function converges to obtain the modality classification model.

6. The AI-based complaint and integrated media clarification system for false information as described in claim 4, characterized in that, The S5 also includes a fine-tuning training process for the large language model, including: constructing a debunking dataset containing pairs of false information and corresponding professional debunking articles; using a structured evidence chain as input and professional debunking articles as output; and using a causal language model loss function for instruction fine-tuning training, so that the large language model generates clarification text based on the evidence chain.

7. The AI-based complaint and integrated media clarification system for false information as described in claim 4, characterized in that, In S5, if the dominant propagation mode is video-dominated, the digital human video generation interface is called to convert factual data into a digital human broadcast script and generate a clarification short video containing an authoritative source link overlay. If the dominant communication modality is text-based, a large language model is invoked to generate a clarification text, and a data visualization interface is used to render the factual data into charts to generate a clarification chart. If the dominant communication modality is plain text-based, a large language model is invoked to generate a structured clarification text.

8. The AI-based complaint and integrated media clarification system for false information as described in claim 4, characterized in that, In S6, the publishing specification parameters include the upper limit of video duration, the upper limit of text and image length, and the upper limit of the number of images; Keyframe extraction and audio acceleration processing are performed on short videos that exceed the time limit; For text and images exceeding the length limit, segmented summaries are extracted.

9. The AI-based complaint and integrated media clarification system for false information as described in claim 1, characterized in that, The targeted clarification release and dynamic hedging module performs the following steps: S7: Targeted matching establishes a mapping relationship between node susceptibility index and converged media content package modality. If a node belongs to the core fermentation node, it is matched with a deep text and image content package. If a node belongs to the peripheral propagation node or a susceptible latent node, it is matched with a short video content package. S8: Precise push, through the application programming interface of the converged media platform, pushes the matched converged media content package to the corresponding user node; S9: Hedging feedback and secondary triggering. Within a preset time window after the release of the converged media content package, real-time data on the interaction of target user nodes with the clarification content is collected. The negative sentiment decline rate is calculated through the sentiment analysis model. If the negative sentiment decline rate is lower than the preset decline rate threshold, the cross-modal clarification content adaptive generation module is triggered to change the modality and regenerate supplementary clarification content, and a secondary targeted push is performed.

10. The AI-based complaint and integrated media clarification system for false information as described in claim 9, characterized in that, In S9, the negative emotion reduction rate is obtained by calculating the difference between the user's emotional polarity score before receiving the clarification information and the emotional polarity score after receiving the clarification information.

Images