Search

CN-122021920-A - Social network rumor tracing method based on multi-mode fusion and thinking chain reasoning

CN122021920ACN 122021920 ACN122021920 ACN 122021920ACN-122021920-A

Abstract

The invention relates to the field of computer network application, and particularly discloses a social network rumor tracing method based on multi-mode fusion and thinking chain reasoning. The method comprises the steps of 1, obtaining and preprocessing multi-mode information, obtaining public opinion data, extracting multi-mode multi-dimensional feature space, 2, fusing multi-mode feature perception and knowledge, learning and measuring feature isomerism among different modes, training sub-models layer by layer, 3, carrying out thinking chain reasoning, constructing a multi-level reasoning chain, extracting feature vectors, establishing fact association according to the feature vectors, synchronously constructing a fact relation knowledge graph, and reasoning out conclusions, 4, carrying out dynamic verification, mapping the reasoning step text into vectors, carrying out similarity calculation, setting a starting threshold, and judging that the current reasoning has illusion when the similarity is smaller than the starting threshold. The social network rumor tracing method based on multi-mode fusion and thinking chain reasoning can improve the reasoning accuracy and the credibility of the conclusion.

Inventors

  • GUO ZHIWEI
  • WANG YUTONG
  • XIAO PENG
  • ZHANG KECHENG
  • SHI ZE

Assignees

  • 重庆工商大学

Dates

Publication Date
20260512
Application Date
20260211

Claims (10)

  1. 1. The social network rumor tracing method based on multi-mode fusion and thinking chain reasoning is characterized by comprising the following steps of: Step 1, multi-modal information acquisition and preprocessing, namely acquiring public opinion data, respectively pre-training a text model and a visual model, and deeply extracting multi-modal multi-dimensional feature space; Step 2, multi-modal feature perception and knowledge fusion, wherein multi-modal neural network feature learning is adopted to measure feature heterogeneity among different modalities, and a layer-by-layer training algorithm is adopted to train the sub-models layer by layer; Step 3, thinking chain reasoning, namely constructing a multi-level reasoning chain according to the large language model, extracting feature vectors, establishing fact association according to the feature vectors, synchronously constructing a fact relation knowledge graph, and reasoning out a conclusion based on the large language model; step 4, dynamically checking, namely mapping the reasoning step text into a vector by using a comparison learning model, searching a Top-K node set related to the generation step in a rumor propagation subgraph, and carrying out similarity calculation by using the following formula 1: Equation 1 Wherein, the Representing nodes in a subgraph Is a vector of (2); representing a vector obtained by mapping the text of the reasoning step through a comparison learning model; Representing a complete rumor spread subgraph; And setting a starting threshold, and judging that the current reasoning has illusion when the similarity is smaller than the starting threshold.
  2. 2. The social network rumor tracing method based on multi-modal fusion and thinking chain reasoning as claimed in claim 1, comprising step 5 of self criticizing, starting when the similarity calculation result is smaller than the starting threshold, wherein step 5 comprises judging error type mapping and attribution, the error judgment type comprises fact absence, logic contradiction and expression ambiguity, Normalized entropy is calculated by the following equation 2: Equation 2 Wherein N is the number of key Token, V is the vocabulary space, an entropy threshold is set, and the fact absence is judged when the value of normalized entropy is smaller than the entropy threshold; The final score is determined by the following equation 3: Equation 3 And when the similarity is smaller than the starting threshold, judging that the expression is fuzzy.
  3. 3. The social network rumor tracing method based on multimodal fusion and thinking chain reasoning of claim 2, comprising the steps of 6, dynamically correcting, wherein the error judgment type is dynamically assembled into a diagnosis report, the dynamic correction comprises in-situ correction, and the diagnosis report is added to a context window to construct a correction prompt when in-situ correction.
  4. 4. The social network rumor tracing method based on multi-mode fusion and thinking chain reasoning according to claim 3, wherein step 6 is to dynamically assemble the error judgment type into an anti-thinking instruction, then output the anti-thinking instruction into a diagnosis report in a JSON format, wherein the diagnosis report comprises error positioning coordinates, error cause analysis and suggested correction direction, and the diagnosis report is used as an in-situ correction signal.
  5. 5. The social network rumor tracing method based on multi-mode fusion and thinking chain reasoning of claim 3, wherein the dynamic correction further comprises a trace-back operation, wherein a super parameter is preset as a serious error threshold, and when any one of M gnd < serious error threshold and the error type is a logic contradiction is met, serious logic error is determined to be found; after finishing the dynamic correction of the step 6, repeating the step 4 and the step 5, and marking the last dynamic correction as in-situ correction failure when the similarity calculation result is smaller than the starting threshold value; triggering a backtracking operation when serious logic errors and continuous K times of in-situ correction failures are found, wherein the backtracking operation comprises the following steps: s1, pruning, namely pruning the current reasoning branches in a search tree; S2, state rollback, namely presetting a high confidence threshold, marking the node as a high confidence node when M gnd is more than or equal to the high confidence threshold, pruning, returning to the last high confidence node, and returning to a root node when no last high confidence node exists; s3, resampling temperature disturbance, namely temporarily lifting decoding temperature parameters and exploring an inference path which is not pruned.
  6. 6. The social network rumor tracing method based on multimodal fusion and mental chain reasoning as claimed in claim 5, wherein step S3 outputs a path with highest accumulated confidence as a final rumor tracing evidence chain, wherein the path is successfully traced to a root node.
  7. 7. The social network rumor tracing method based on multimodal fusion and thinking chain reasoning of claim 6 further comprising step 7 of visually tracing, forming a tree diagram by the reasoning process of step 3 and step 4, setting a display window, and linking the display window with a rumor tracing evidence chain.
  8. 8. The social network rumor tracing method based on multi-modal fusion and thinking chain reasoning of claim 7 further comprising step 8 of man-machine reflux, receiving manual feedback and refluxing the feedback result to step 2.
  9. 9. The social network rumor tracing method based on multimodal fusion and mental chain reasoning as set forth in claim 1, wherein the feature vector of step 3 comprises one or more of an entity, an event, a time, and a place.
  10. 10. The social network rumor tracing method based on multimodal fusion and mind chain reasoning of claim 1, wherein step 1 obtains public opinion data through a crawler technology and retains dynamic interaction information.

Description

Social network rumor tracing method based on multi-mode fusion and thinking chain reasoning Technical Field The invention relates to the field of computer network application, in particular to a social network rumor tracing method based on multi-modal fusion and thinking chain reasoning. Background The rumor propagation speed in the current social network is high, the influence range is wide, the traditional rumor discovery and refute a rumour are completed by combining keyword matching and manual reporting and auditing, but the workload of the mode is large under the condition that the current social network information is more. Later, rumor discovery and refute a rumour mode combined with AI occurred, but AI refute a rumour had the following problems: 1) The accuracy rate is insufficient, namely rumor recognition is dependent on simple keyword matching, a shallow neural network or a single-mode classification model, the model lacks understanding capability of language deep semantics such as metaphors, irony, context association and the like, complex association between multi-mode information including misleading characters of picture configuration and deep counterfeit video cannot be effectively processed, and rumor recognition is regarded as a simple classification task, and the back complex logic and fact reasoning process is ignored; 2) Black box decision, namely, a high-efficiency deep learning model such as CNN and DNN of some complex structures is a 'black box', and the decision process is difficult to trace; 3) The interactive rigidification is that the design paradigm of the traditional system is 'one-time question-answering', the interactive logic is closed, and a sustainable traceable and inquired reasoning memory system is absent in the system. The above problems result in relatively poor recognition accuracy of the network rumors, and the reasoning process cannot trace and inquire, so that the user cannot trust the reasoning result. Disclosure of Invention The invention aims to provide a social network rumor tracing method based on multi-mode fusion and thinking chain reasoning so as to improve the reasoning accuracy and the credibility of conclusions. In order to achieve the aim, the social network rumor tracing method based on multi-mode fusion and thinking chain reasoning comprises the following steps: Step 1, multi-modal information acquisition and preprocessing, namely acquiring public opinion data, respectively pre-training a text model and a visual model, and deeply extracting multi-modal multi-dimensional feature space; Step 2, multi-modal feature perception and knowledge fusion, wherein multi-modal neural network feature learning is adopted to measure feature heterogeneity among different modalities, and a layer-by-layer training algorithm is adopted to train the sub-models layer by layer; Step 3, thinking chain reasoning, namely constructing a multi-level reasoning chain according to the large language model, extracting feature vectors, establishing fact association according to the feature vectors, synchronously constructing a fact relation knowledge graph, and reasoning out a conclusion based on the large language model; step 4, dynamically checking, namely mapping the reasoning step text into a vector by using a comparison learning model, searching a Top-K node set related to the generation step in a rumor propagation subgraph, and carrying out similarity calculation by using the following formula 1: Equation 1 And setting a starting threshold, and judging that the current reasoning has illusion when the similarity is smaller than the starting threshold. The beneficial effect of this scheme is: 1. In the scheme, deep analysis can be performed when reasoning is performed on a conclusion through a large language model, multi-mode information is collected in combination with the step 1, the conclusion is deduced, meanwhile, the extraction objective entity- > establishment of causal/emotion association- > establishment of a fact relation map- > obtaining of a cognition process of the conclusion of a human expert is simulated, complex causal relation can be processed through the model, and the problem of a logic shallow layer is solved. And step 1 obtains multi-mode information and preprocesses the data, and the chaotic multi-mode display world data can be digested, cleaned, understood and encoded into a high-quality characteristic vector diagram by combining step 2, so that a quality basis is provided for the completion of subsequent high-level cognitive tasks. 2. The operation of the step 2 measures and fuses the characteristic isomerism among different modes through an attention mechanism and a layer-by-layer training algorithm, can filter redundant, repeated and nonsensical multi-mode network public opinion big data, and strengthens the characteristic recognition effect, thereby being beneficial to improving the reasoning accuracy. 3. Public opinion with "illusion" is not considere