Detection method for multi-modal data of recessive toxicity and terminal equipment
By constructing a semantic association tree of visual and textual data and establishing a cross-modal association graph, the problem that existing models cannot identify multimodal data with latent toxicity is solved. This enables effective detection and transparency assessment of latent toxicity, provides a quantitative standard for the concealment of toxicity, and fills the gap in the lack of data on highly concealed toxicity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE CHINESE UNIV OF HONG KONG (SHENZHEN)
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-12
AI Technical Summary
Existing content security detection models are unable to effectively identify hidden toxic multimodal data and lack the ability to structured reasoning about deep semantic relationships across modalities, which allows hidden harmful content to evade review and makes the decision-making process opaque.
By constructing semantic association trees of visual and textual data, cross-modal associations are established, forming a cross-modal bipartite graph. This enables structured reasoning of multimodal data, calculation of the cumulative transition probability of toxic node pairs, and provision of toxicity labels and concealment scores.
It enables effective identification of multimodal data of latent toxicity, provides quantifiable assessment of toxicity concealment, improves detection accuracy and transparency, and constructs a highly concealed toxicity dataset to support model training and evaluation.
Smart Images
Figure CN122020296A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and terminal device for detecting multimodal data of latent toxicity, belonging to the field of toxicity data detection technology. Background Technology
[0002] With the explosive growth of multimedia content on the internet, the dissemination of harmful information using multimodal data (especially combinations of images and text) has become an increasingly serious challenge. Compared to single-modal content, multimodal content can express more complex and subtle semantics. Attackers often exploit this to create "hidden toxic" content. In such content, images and text, taken individually, are completely benign or neutral; only when the recipient combines the two and uses specific cultural context, common sense, or logical reasoning to make associations can they understand the hidden malice. For example, an ordinary picture of an industrial "flange," accompanied by the text "Undefined item, private message," might imply the illegal trade of weapons components in a specific context.
[0003] Existing content security detection models are insufficient in identifying this type of highly concealed and toxic content because they primarily rely on explicit features extracted from a single modality, lacking the ability to perform structured reasoning on deep semantic relationships across modalities. This allows a large amount of hidden harmful content to evade censorship, posing potential social harm. Furthermore, the "black box" nature of existing models makes their decision-making process opaque, difficult to gain user trust, and unable to provide effective decision-making assistance to content moderators. Summary of the Invention
[0004] This invention provides a method and terminal device for detecting latent toxicity multimodal data, which can solve the problems that existing models cannot effectively detect latent toxicity content and are difficult to provide clear explanations and quantify its degree of concealment.
[0005] On one hand, the present invention provides a method for detecting multimodal data of latent toxicity, the method comprising:
[0006] S1. Obtain multiple visual nodes from the input image of the input image-text pair, and construct a visual association tree of the input image based on the multiple visual nodes;
[0007] S2. Obtain multiple text nodes from the input text of the input image-text pair, and construct a text association tree of the input text based on the multiple text nodes;
[0008] S3. Construct a cross-modal bipartite graph of the input image-text pair based on the visual association tree and the text association tree; the cross-modal bipartite graph contains all node pairs consisting of visual nodes and text nodes;
[0009] S4. Match each node pair in the cross-modal bipartite graph with all concept pairs in the toxic concept pair set, and determine the toxicity label of the input graph-text pair based on the matching results.
[0010] Optionally, S1 specifically includes:
[0011] S11. Extract entity concepts from the input image of the input image-text pair, and use the entity concepts as visual root nodes and the visual root nodes as visual parent nodes;
[0012] S12. Extend multiple visual child nodes from the visual parent node, and create edges from the visual parent node to each visual child node;
[0013] S13. Calculate the edge transition probability from the visual parent node to each visual child node, and take the visual child node as the visual parent node. Repeat S12 and S13 to obtain the visual association tree of the input image. The visual nodes include the visual root node, all visual parent nodes and all visual child nodes.
[0014] Optionally, S2 specifically includes:
[0015] S21. Extract entity concepts from the input text of the input image-text pair, use the entity concepts as the text root node, and use the text root node as the text parent node;
[0016] S22. Extend multiple text child nodes from the text parent node, and create edges from the text parent node to each text child node;
[0017] S23. Calculate the edge transition probability from the text parent node to each text child node, and take the text child node as the text parent node. Repeat S22 and S23 to obtain the text association tree of the input text; the text node includes the text root node, all text parent nodes and all text child nodes.
[0018] Optionally, determining the toxicity label of the input image-text pair based on the matching result in S4 specifically includes:
[0019] If the matching result indicates the existence of a toxic node pair, then the toxicity label of the input image-text pair is determined to be toxic; the toxic node pair is the node pair in the cross-modal bipartite graph that matches any concept pair in the set of toxic concept pairs.
[0020] If the matching result indicates that there is no toxic node pair, then the toxicity label of the input image-text pair is determined to be non-toxic.
[0021] Optionally, after S4, the method further includes:
[0022] S5. When the toxicity label of the input image-text pair is toxic, calculate the toxicity concealment score of the input image-text pair based on the visual association tree and the text association tree.
[0023] Optionally, the step S5 of calculating the toxicity concealment score of the input image-text pair based on the visual association tree and the text association tree specifically includes:
[0024] The joint transition probability of the toxic node pair is calculated based on the visual association tree and the text association tree, and the difference between 1 and the joint transition probability is used as the toxicity concealment score of the input image-text pair.
[0025] Optionally, the joint transition probability of the toxic node pair is calculated based on the visual association tree and the text association tree, specifically including:
[0026] The cumulative probability of the visual path from the visual root node to the toxic visual node of the toxic node pair is calculated based on the visual association tree, and the cumulative probability of the text path from the text root node to the toxic text node of the toxic node pair is calculated based on the text association tree.
[0027] Calculate the product of the cumulative probability of the visual path and the cumulative probability of the text path, and use the product as the joint transition probability of the toxic node pair.
[0028] Optionally, the cumulative probability of the visual path from the visual root node to the toxic visual node of the toxic node pair is calculated based on the visual association tree, specifically as follows:
[0029] Calculate the first product of all edge transition probabilities on the path from the visual root node to the toxic node pair, and use the first product as the cumulative probability of the visual path.
[0030] The cumulative probability of the text path from the root node to the toxic text node of the toxic node pair is calculated based on the text association tree, specifically as follows:
[0031] Calculate the second product of all edge transition probabilities on the path from the root text node to the toxic text node pair, and use the second product as the cumulative probability of the text path.
[0032] Optionally, after S5, the method further includes:
[0033] Based on the toxic node pairs and their inference paths in the visual association tree and the text association tree, a large language model is used to generate explanatory information for the detection process.
[0034] On the other hand, the present invention provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the detection method for latent toxicity multimodal data as described above.
[0035] The beneficial effects that this invention can produce include:
[0036] This invention provides a method for detecting latent toxicity in multimodal data. This method constructs a semantic association tree of visual and textual data and establishes cross-modal associations to organize entities and concepts in the multimodal data into a structured graph. By reasoning on this graph, latent toxic content composed of multiple benign concepts can be effectively identified, solving the technical problem that traditional methods struggle to detect deep, indirectly related toxicity.
[0037] The present invention provides a detection method for latent toxicity multimodal data. This method calculates the cumulative transition probability from the root node to the detected toxic node pair in the toxicity association graph, and for the first time provides a quantifiable and interpretable metric for the "concealment" of toxic content.
[0038] This invention systematically transforms overtly toxic content into highly concealed image-text pairs through the collaboration of multiple intelligent agents such as architects, erasers, reviewers, and gatekeepers. Finally, through manual verification, it constructs the first hidden toxicity dataset specifically for evaluating the detection of highly concealed toxicity. Attached Figure Description
[0039] Figure 1 This is a flowchart of a method for detecting multimodal data of latent toxicity provided in an embodiment of the present invention. Detailed Implementation
[0040] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.
[0041] This invention provides a method for detecting multimodal data on latent toxicity, such as... Figure 1 As shown, the method includes:
[0042] S1. Obtain multiple visual nodes from the input image of the input image-text pair, and construct a visual association tree of the input image based on the multiple visual nodes.
[0043] Specifically, it includes:
[0044] S11. Extract entity concepts from the input image of the input image-text pair, and use the entity concepts as the visual root node and the visual root node as the visual parent node.
[0045] S12. Extend multiple visual child nodes from the visual parent node and create edges from the visual parent node to each visual child node;
[0046] S13. Calculate the edge transition probability from the visual parent node to each visual child node, and take the visual child node as the visual parent node. Repeat S12 and S13 to obtain the visual association tree of the input image. The visual nodes include the visual root node, all visual parent nodes and all visual child nodes.
[0047] The above visual association tree ;
[0048] in, A set of visual nodes, representing the visual nodes from the input image. The semantic concepts or entities extracted from them; For visual edge sets, representing semantic associations between visual nodes, this tree structure represents hierarchical relationships between concepts, and child nodes are semantic extensions or associations of parent nodes; It is a visual edge transition probability set, which quantifies the probability strength of associating a parent node with its child nodes.
[0049] The specific construction process of the visual association tree is as follows:
[0050] (1) Determine the root node. Using a pre-trained multimodal large language model (MLLM), based on a preset prompt word, from the input image... Extract its main entities or core concepts and use them as the visual root node of the visual association tree. For example, for an image of "white flour," the visual root node could be "white flour." If multiple core concepts are extracted, multiple independent visual root nodes can be created.
[0051] (2) Determine the visual child nodes and edges. For the first child node in the visual association tree... Any visual node of the layer Call MLLM again to generate the maximum number of prompts based on the preset prompts. indivual( The visual child nodes (which are hyperparameters) are closely associated with the parent visual node. These associations can be semantic similarity (e.g., "flour" to "starch"), functional similarity (e.g., "shovel" to "digging tool"), appearance similarity (e.g., "white flour" to "cocaine"), cultural symbols (e.g., "rose" to "love"), etc. Edges are created from the parent visual node to these child visual nodes, with edge weights equal to the edge transition probabilities.
[0052] (3) Assign edge transition probability. Calculate the probability from the visual parent node. to each of its visual child nodes edge transition probability These probabilities represent the strength of the association, and the sum of the probabilities from a visual parent node to all its visual child nodes is 1. This probability can be estimated from the output of MLLM.
[0053] (4) Set the stopping condition. Repeat (2) and (3) to expand the visual association tree layer by layer until the preset maximum depth is reached. ( (for hyperparameters), then terminate and return the finally constructed visual association tree. .
[0054] S2. Obtain multiple text nodes from the input text of the input image-text pair, and construct a text association tree of the input text based on the multiple text nodes.
[0055] Specifically, it includes:
[0056] S21. Extract entity concepts from the input text of the input image-text pair, use the entity concepts as the root node of the text, and use the root node of the text as the parent node of the text.
[0057] S22. Extend multiple text child nodes from the text parent node, and create edges from the text parent node to each text child node;
[0058] S23. Calculate the edge transition probability from the text parent node to each text child node, and take the text child node as the text parent node. Repeat S22 and S23 to obtain the text association tree of the input text. The text nodes include the text root node, all text parent nodes and all text child nodes.
[0059] The above text association tree ;
[0060] in, A set of text nodes; For text edge sets; Let be the text edge transition probability set.
[0061] Text Association Tree Structure and definition of visual association tree Similarly, the specific construction method can be found in the visual association tree. The construction method.
[0062] S3. Construct a cross-modal bipartite graph of the input image-text pair based on the visual association tree and the text association tree; the cross-modal bipartite graph contains all node pairs consisting of visual nodes and text nodes.
[0063] The above transmodal bipartite diagram ;
[0064] This cross-modal bipartite graph connects all nodes in the visual association tree and the text association tree, and its edge set... It contains all possible node pairs consisting of visual and text nodes for subsequent cross-modal toxicity association detection.
[0065] For input image and text pairs Its toxicity association graph TAG is defined as a triple: .
[0066] The Toxicity Association Graph (TAG) is the core data structure used in this invention for modeling and inferring latent toxicity.
[0067] S4. Match each node pair in the cross-modal bipartite graph with all concept pairs in the toxic concept pair set, and determine the toxicity label of the input graph-text pair based on the matching results.
[0068] The above-mentioned determination of the toxicity label for the input image-text pair based on the matching results specifically includes:
[0069] If the matching result indicates the existence of a toxic node pair, then the toxicity label of the input image-text pair is determined to be toxic; a toxic node pair is a node pair that matches any concept pair in the set of toxic concept pairs in the cross-modal bipartite graph.
[0070] If the matching result indicates that there is no toxic node pair, then the toxicity label of the input image-text pair is determined to be non-toxic.
[0071] Given a predefined set of toxic concepts (This set can be user-defined, obtained from a knowledge base, or inferred through MLLM implicit knowledge), and then the cross-modal bipartite graph is systematically examined. Each node pair in ( , ),in From visual association tree , From text association tree .
[0072] If a node pair is found ( , ) and sets Matching a certain toxic concept in (e.g., It is "cocaine". If the input image / text pair is matched with the toxic concept "drug trade," the detection process terminates, the input image / text pair is determined to be toxic, and a toxicity label is output. To improve efficiency, the search process can start from the root node level of the tree and proceed downwards level by level.
[0073] If no match is found after traversing all possible node pairs, the input image-text pair is determined to be non-toxic, and a toxicity label is output. .
[0074] Furthermore, after S4, the method further includes:
[0075] S5. When the toxicity label of the input image-text pair is toxic, calculate the toxicity concealment score of the input image-text pair based on the visual association tree and the text association tree.
[0076] Specifically, it includes:
[0077] The joint transition probability of toxic node pairs is calculated based on the visual association tree and the text association tree, and the difference between 1 and the joint transition probability is used as the toxicity concealment score of the input image-text pair.
[0078] Specifically, the joint transition probability of toxic node pairs is calculated based on the visual association tree and the text association tree, including:
[0079] (1) Calculate the cumulative probability of the visual path from the visual root node to the toxic visual node of the toxic node pair based on the visual association tree, and calculate the cumulative probability of the text path from the text root node to the toxic text node of the toxic node pair based on the text association tree.
[0080] Specifically: calculate the first product of the edge transition probabilities on the path from the visual root node to the toxic visual node of the toxic node pair, and use the first product as the cumulative probability of the visual path; calculate the second product of the edge transition probabilities on the path from the text root node to the toxic text node of the toxic node pair, and use the second product as the cumulative probability of the text path.
[0081] (2) Calculate the product of the cumulative probability of the visual path and the cumulative probability of the text path, and use the product as the joint transition probability of the toxic node pair.
[0082] Only if the test result is toxic ( When calculating the toxicity concealment score, First, calculate the detected toxic node pairs ( , Joint transition probability This probability is equal to the probability from the visual root node of the visual association tree. To toxic visual nodes of toxic node pairs The cumulative probability of the path, relative to the text root node of the text association tree. To toxic text nodes of toxic node pairs The product of the cumulative probabilities of the paths: in, It is the product of the transition probabilities of all edges on the path from the visual root node to the toxic visual node of the toxic node pair. It is the product of the transition probabilities of all edges along the path from the root text node to the toxic text node pair. Then, the toxicity concealment score... Defined as: Its value range is within between. A higher value indicates a lower probability of joint transfer, meaning a longer and more complex reasoning process is needed to detect toxicity, i.e., higher concealment of the toxicity. If the test result is non-toxic ( ),but Set to 1 to represent the maximum concealment level.
[0083] Furthermore, following S5, the method further includes:
[0084] Based on the toxic node pairs and their inference paths in the visual association tree and text association tree, a large language model is used to generate explanatory information for the detection process.
[0085] Using the detected toxic node pairs ( , ) and its The reasoning path (i.e., the path from the root node to the toxic node pair) is used to generate a natural language explanation by calling a Large Language Model (LLM). This explanation clearly illustrates how the model infers the implicit toxic meaning from seemingly harmless text and image content through a series of semantic associations.
[0086] Table 1 below shows the pseudocode for the detection method for multimodal data of latent toxicity.
[0087] Table 1. Pseudocode of Detection Method
[0088] Table 2 below shows the pseudocode for calculating the toxicity concealment score.
[0089] Table 2 Pseudocode for Calculating Toxicity Concealment Score
[0090] To train and evaluate the above detection method, another embodiment of the present invention provides a method for systematically constructing a high-coverage toxicity dataset (CTD). This method employs an automated pipeline with multi-agent cooperation, as follows:
[0091] (1) Architect Agent: As the initiator of the process, it receives an abstract toxicity category or scenario description (such as “illegal transaction”), uses integrated Internet search tools to brainstorm, and generates a specific, explicit initial toxicity scenario description.
[0092] (2) Eraser Agent: The core function of this agent is to improve the concealment of toxicity. It receives explicit toxic scenarios generated by the architect agent and systematically replaces explicit toxic elements (such as "guns" and "drugs") with semantically neutral or deceptive alternatives (such as "industrial parts" and "flour"). The replacement is based on the commonalities between the two in terms of appearance, function, and cultural metaphor. After the replacement is completed, it decomposes the modified scenario into text prompts for generating images and corresponding text descriptions, and generates images using a text-to-image (T2I) model. Finally, it outputs a candidate image-text pair and an "erasure reason memo" recording the logic of its steganography transformation.
[0093] (3) Judge Agents: This is a judging panel composed of multiple agents with different perspectives and knowledge. They judge candidate image-text pairs using three different evaluation protocols: Blind review: Only the image-text pair is observed to determine whether its toxicity can be perceived, in order to assess whether its concealment level is appropriate. Informed review: The image-text pair and the "erasure reason memo" are observed simultaneously to assess whether its logical transformation is reasonable and explainable. Forced reasoning review: The image-text pair is informed that it contains toxicity, but no clues are provided, forcing the agents to reason out the potential malice themselves in order to test the discoverability of the toxicity.
[0094] (4) Gatekeeper Agent: As the final decision-maker in the pipeline, it integrates the feedback from all reviewer agents and makes the final decision. If the concealment, logic, and consistency of the candidate image-text pair all meet the preset standards, it is approved as a sample of the final dataset. If it does not meet the standards, the gatekeeper agent will generate targeted and constructive feedback based on the review opinions and return the task to the eraser agent for a new round of modification and iteration.
[0095] (5) Human Verification: All candidate image-text pairs approved by the agent pipeline will enter the final human verification stage. Human annotators will conduct the final review to ensure the quality, reliability, and ethical compliance of the data. Only samples that receive consistent confirmation will be included in the final latent toxicity dataset (CTD).
[0096] Table 3 below shows the pseudocode for the process of constructing the latent toxicity dataset.
[0097] Table 3 Pseudocode for constructing the latent toxicity dataset
[0098] Another embodiment of the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the detection method for latent toxicity multimodal data as described above.
[0099] The beneficial effects of this invention mainly include:
[0100] (1) Significantly improves the accuracy and recall of latent toxicity detection: The detection method of this invention, by constructing and reasoning a toxicity association graph (TAG), can discover deep toxicities that traditional methods cannot identify and that require multiple semantic associations to surface. Experiments show that, especially in highly concealed scenarios, this invention can significantly improve the F2 score of the detection model and effectively reduce false negatives.
[0101] (2) Achieving interpretability and transparency in detection decisions: Unlike existing "black box" models, this invention can output the complete reasoning path leading to its judgment. This interpretability not only enhances the credibility of the content moderation system, but also provides decision-making basis for human moderators, greatly improving the efficiency and accuracy of the moderation process.
[0102] (3) The first standard to quantify the degree of toxicity concealment: The MTC score of this invention is the first indicator that can quantitatively measure the degree of toxicity concealment in multimodal content. It makes it possible to classify, analyze and defend against toxic content with different concealment levels, providing a new dimension for content security risk assessment.
[0103] (4) Solved the problem of scarcity of highly concealed toxicity data: The multi-agent data generation pipeline proposed in this invention can automatically and scalably generate high-quality, highly concealed toxicity data samples. The constructed CTD dataset fills the gap in the evaluation benchmark of existing technologies, lays a solid foundation for training and evaluating more powerful concealed toxicity detection models, and strongly promotes the development of the entire field.
[0104] The above descriptions are merely a few embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any modifications or alterations made by those skilled in the art without departing from the scope of the technical solution of the present invention using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.
Claims
1. A method for detecting multimodal data of latent toxicity, characterized in that, The method includes: S1. Obtain multiple visual nodes from the input image of the input image-text pair, and construct a visual association tree of the input image based on the multiple visual nodes; S2. Obtain multiple text nodes from the input text of the input image-text pair, and construct a text association tree of the input text based on the multiple text nodes; S3. Construct a cross-modal bipartite graph of the input image-text pair based on the visual association tree and the text association tree; the cross-modal bipartite graph contains all node pairs consisting of visual nodes and text nodes; S4. Match each node pair in the cross-modal bipartite graph with all concept pairs in the toxic concept pair set, and determine the toxicity label of the input graph-text pair based on the matching results.
2. The method according to claim 1, characterized in that, S1 specifically includes: S11. Extract entity concepts from the input image of the input image-text pair, and use the entity concepts as visual root nodes and the visual root nodes as visual parent nodes; S12. Extend multiple visual child nodes from the visual parent node, and create edges from the visual parent node to each visual child node; S13. Calculate the edge transition probability from the visual parent node to each visual child node, and take the visual child node as the visual parent node. Repeat S12 and S13 to obtain the visual association tree of the input image. The visual nodes include the visual root node, all visual parent nodes and all visual child nodes.
3. The method according to claim 2, characterized in that, S2 specifically includes: S21. Extract entity concepts from the input text of the input image-text pair, use the entity concepts as the text root node, and use the text root node as the text parent node; S22. Extend multiple text child nodes from the text parent node, and create edges from the text parent node to each text child node; S23. Calculate the edge transition probability from the text parent node to each text child node, and take the text child node as the text parent node. Repeat S22 and S23 to obtain the text association tree of the input text; the text node includes the text root node, all text parent nodes and all text child nodes.
4. The method according to claim 1, characterized in that, Determining the toxicity label of the input image-text pair based on the matching result in S4 specifically includes: If the matching result indicates the existence of a toxic node pair, then the toxicity label of the input image-text pair is determined to be toxic; the toxic node pair is the node pair in the cross-modal bipartite graph that matches any concept pair in the set of toxic concept pairs. If the matching result indicates that there is no toxic node pair, then the toxicity label of the input image-text pair is determined to be non-toxic.
5. The method according to claim 4, characterized in that, Following S4, the method further includes: S5. When the toxicity label of the input image-text pair is toxic, calculate the toxicity concealment score of the input image-text pair based on the visual association tree and the text association tree.
6. The method according to claim 5, characterized in that, The step S5, which calculates the toxicity concealment score of the input image-text pair based on the visual association tree and the text association tree, specifically includes: The joint transition probability of the toxic node pair is calculated based on the visual association tree and the text association tree, and the difference between 1 and the joint transition probability is used as the toxicity concealment score of the input image-text pair.
7. The method according to claim 6, characterized in that, The joint transition probability of the toxic node pair is calculated based on the visual association tree and the text association tree, specifically including: The cumulative probability of the visual path from the visual root node to the toxic visual node of the toxic node pair is calculated based on the visual association tree, and the cumulative probability of the text path from the text root node to the toxic text node of the toxic node pair is calculated based on the text association tree. Calculate the product of the cumulative probability of the visual path and the cumulative probability of the text path, and use the product as the joint transition probability of the toxic node pair.
8. The method according to claim 7, characterized in that, The cumulative probability of the visual path from the visual root node to the toxic visual node of the toxic node pair is calculated based on the visual association tree, specifically as follows: Calculate the first product of all edge transition probabilities on the path from the visual root node to the toxic node pair, and use the first product as the cumulative probability of the visual path. The cumulative probability of the text path from the root node to the toxic text node of the toxic node pair is calculated based on the text association tree, specifically as follows: Calculate the second product of all edge transition probabilities on the path from the root text node to the toxic text node pair, and use the second product as the cumulative probability of the text path.
9. The method according to claim 5, characterized in that, Following S5, the method further includes: Based on the toxic node pairs and their inference paths in the visual association tree and the text association tree, a large language model is used to generate explanatory information for the detection process.
10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for detecting latent toxicity multimodal data as described in any one of claims 1 to 9.