CN-122021622-A - False news intelligent detection system based on multidimensional fusion

CN122021622ACN 122021622 ACN122021622 ACN 122021622ACN-122021622-A

Abstract

The invention belongs to the technical field of artificial intelligence information processing, and particularly relates to a false news intelligent detection system based on multidimensional fusion. The system comprises a data acquisition processing module, a rapid detection module, a multi-dimensional slow auditing module, a multi-mode image-text consistency detection module, a multi-agent collaborative analysis module and a four-level classification conclusion output module, wherein the data acquisition processing module is used for acquiring news data and preprocessing, the rapid detection module is used for performing real-false two-class rapid detection on news by adopting a ERNIE 3.0.0 model, the multi-dimensional slow auditing module is used for performing deep analysis from six dimensions of news source confidence evaluation, text consistency detection, AI generation content detection, multi-mode image-text consistency analysis, user propagation confidence evaluation and logic judgment based on a large language model, and the multi-agent collaborative analysis module is used for constructing seven specialized agents to perform parallel analysis and collaborative decision. According to the invention, by constructing a rapid detection and slow audit double-layer detection framework and combining a multi-agent cooperative mechanism, the omnibearing and high-precision detection of false news is realized, the content threat is effectively generated on AI, and the user media literacy is improved.

Inventors

LUO FANG
XIAO JUNHUA
WANG RUI
WANG BOZHAO
HUANG TIANCI
Song tengfei

Assignees

武汉理工大学

Dates

Publication Date: 20260512
Application Date: 20251209

Claims (10)

1. A false news intelligent detection system based on multidimensional fusion is characterized by comprising The data acquisition processing module is used for acquiring and preprocessing news data from a plurality of news platforms, wherein the news data comprises news headlines, news texts, news configuration drawings, news videos, news audios, release platform information, propagation time records, propagation geographic position data and propagation user portrait data; The rapid detection module adopts ERNIE 3.0.0 model which is finely adjusted based on word segmentation-enhancement-pretraining three-level architecture, carries out true and false two-class rapid detection on the preprocessed news data, and outputs true and false two-class results; the system comprises a multi-dimensional slow auditing module, a real-false two-classification quick detection module and a real-false two-classification quick detection module, wherein the news data is subjected to deep analysis from six evaluation dimensions, the six evaluation dimensions comprise news source confidence assessment, text consistency detection, AI generation content detection, multi-mode image-text consistency analysis, user transmission confidence assessment and logic judgment based on a large language model, and comprehensive confidence scores are calculated through a weighted fusion algorithm; The multi-agent collaborative analysis module is used for constructing seven agents comprising a title analysis expert, a grammar verification expert, a common sense verification expert, a credibility evaluation expert, a multi-source retrieval expert, a user attribute analysis expert and a consistency detection expert, wherein the seven agents carry out parallel analysis and collaborative decision on the output results of the rapid detection module and the multi-dimensional slow auditing module, and output a final detection conclusion.
2. The multi-dimensional fusion-based false news intelligent detection system of claim 1, wherein the preprocessing includes generating a digest using a TextRank algorithm and extracting keywords using a TF-IDF algorithm.
3. The multi-dimensional fusion-based false news intelligent detection system according to claim 1, wherein the news source confidence evaluation uses as input data release platform information collected by the data collection processing module, the release platform information including a platform URL and a platform identifier, and the method comprises: Constructing a news platform credit database, dividing the platform into five grades of high authority, authoritative, neutral, minor and minor according to the platform identification, and outputting the platform authority grade identification; calculating basic reputation scores according to official backgrounds, auditing mechanisms, historical accuracy rates and user base numbers of the platforms corresponding to the release platform information, and outputting the basic reputation scores; and dynamically adjusting credit scores according to recent false news release rate, correction rate and complaint rate of the release platform information corresponding platform, and outputting news source confidence scores.
4. The multi-dimensional fusion-based false news intelligent detection system of claim 1, wherein the topic consistency detection comprises: converting the news headlines and news texts into high-dimensional vector representations by using a pre-trained text2vec-base-chinese model, calculating cosine similarity between the two vectors, and outputting semantic consistency scores; Extracting keywords of the news headline by using a TF-IDF algorithm, calculating the coverage rate and weight matching degree of the keywords in the news headline, and outputting the coverage rate and weight correlation score of the keywords; A final topic consistency score is calculated based on the semantic consistency score, the keyword coverage and the weight relevance score.
5. The multi-dimensional fusion-based false news intelligent detection system of claim 1, wherein the AI-generated content detection comprises: The news text is taken as input, a bidirectional encoder structure based on a Transformer is adopted, text content generated by AI is identified by analyzing the confusion, the burstiness and the vocabulary diversity indexes of the text, and the text AI is output to generate probability scores; Taking the news map as input, extracting texture features, color distribution features and edge detail features of the image by using a convolutional neural network, judging whether the image is generated by an AI through a support vector machine classifier, and outputting an AI generation probability score of the image; Taking the news video as input, extracting inter-frame difference features, motion vector features and time sequence consistency features of the video, adopting a cyclic neural network to perform AI (analog input) generation video identification, and outputting video AI generation probability scores; The news audio is taken as input, the mel frequency cepstrum coefficient, the fundamental frequency outline and the formant characteristics of the audio are analyzed, the depth neural network is used for identifying the AI to synthesize the audio, and the audio AI is output to generate a probability score; And the AI generated content detection outputs the AI generated content comprehensive detection score according to the detection result of the news form comprehensive text, picture, video and audio.
6. The intelligent detection system for false news based on multidimensional fusion according to claim 1, wherein the multi-modal image-text consistency analysis comprises: Using the news text as input, extracting semantic features of the text by using a BERT model, and outputting 768-dimensional text feature vectors; Taking the news map as input, extracting visual features of the image by using ResNet-50 models, and outputting 2048-dimensional image feature vectors; Calculating a correlation matrix between the text feature vector and the image feature vector, searching for consistency features in a high correlation area, exploring inconsistent marks in a low correlation area, and outputting consistency scores and inconsistent scores; and splicing the cross-modal attention features with the original features, and outputting multi-modal image-text consistency scores through the full-connection layer.
7. The multi-dimensional fusion-based false news intelligent detection system of claim 1, wherein the user propagation confidence assessment comprises: taking the propagation time record as input, calculating the time span and propagation speed of news propagation, and outputting a time line dimension score; the spread geographic position data is used as input to evaluate the spread coverage rate and geographic diversity of news in different provinces, and the geographic distribution dimension score is output; analyzing age distribution degree and geographic diversity of the users participating in the transmission by taking the transmission user portrait data as input, and outputting user diversity scores; and the user transmission confidence assessment outputs a user transmission confidence score by weighting and fusing the timeline dimension score, the geographic distribution dimension score and the user diversity score.
8. The intelligent detection system of false news based on multidimensional fusion according to claim 1, wherein the logic judgment based on the large language model adopts DeepSeek model to perform three-dimensional logic evaluation: Analyzing the news text, detecting whether paradoxical fact description, time line conflict and detail incompatibility exist in news, and outputting content logic consistency scores; analyzing the events described in the news text, evaluating whether the event causal link accords with common sense and basic logic, judging whether the causal relationship strength accords with the actual influence degree, and outputting a causal relationship rationality score; analyzing the description mode of the news text, identifying emotion tendency words in the text, evaluating the balance of different views and whether a guiding language is used or not, and outputting the objectivity score of the narrative manipulation; the three-dimensional logic assessment outputs a logic quality score by averaging the content logic consistency score, the causal relationship rationality score, and the narrative maneuver objectivity score.
9. The multi-dimensional fusion-based false news intelligent detection system according to claim 1, wherein the multi-agent collaborative analysis is implemented using a Swarm framework, comprising: the seven agents independently analyze the true and false classification result and the six-dimensional evaluation result at the same time, and each agent outputs the evaluation result and the confidence coefficient of a specific dimension; integrating analysis results of all the agents through a weighted voting mechanism, dynamically adjusting weights according to the historical accuracy, and outputting weighted comprehensive scores; When the opinion of the intelligent agent diverges beyond a threshold value, a negotiation mechanism is started, consensus is achieved through evidence exchange and reasoning verification, and a fourth-level classification conclusion including false news, suspicious content and possibly real and trusted content is output; And in the knowledge updating stage, the verified new knowledge is stored in a knowledge base in a structured manner, the subsequent analysis capability is optimized, and updated knowledge base data is output.
10. The multi-dimensional fusion-based false news intelligent detection system of claim 1, further comprising a user interaction and education module to detect process data, user submitted supplemental evidence, user misjudgment records, and user contribution records as input data: the evidence chain visualization unit takes the detection process data as input, visually displays the detection process and the judgment basis through a semantic tracing graph and a logic thermodynamic diagram, and outputs a visual display interface; The community collaborative verification unit takes the supplemental evidence submitted by the user as input, allows the user to submit the supplemental evidence, and receives the supplemental evidence library after being checked by an administrator, and outputs the verified supplemental evidence to the evidence chain signboard; the media literacy education unit takes the user error judgment record as input, generates a personalized learning path and a personalized test question according to the user error judgment record, and outputs the personalized learning path and the personalized test question; and the contribution degree excitation unit takes the user contribution record as input, establishes a user contribution ranking list, gives point rewards and grade promotion to high-contribution users, and outputs user point and grade data.

Description

False news intelligent detection system based on multidimensional fusion Technical Field The invention belongs to the technical field of artificial intelligence information processing, and particularly relates to a false news intelligent detection system based on multidimensional fusion. Background In recent years, the explosive development of large model technology brings revolutionary breakthrough to the field of artificial intelligence and provides fertile soil for the generation and propagation of false news. The powerful text generation capability of these large models makes it easier to make high quality, indistinguishable spurious content, providing unprecedented technical support for spurious news generation and dissemination. Therefore, the innovation and exploration of false news detection systems are not separated from the application of artificial intelligence technology. In the existing false news detection technology, the natural language processing technology covers multiple dimensions such as text word segmentation, word vector representation, semantic analysis, emotion recognition and the like, can deeply understand the internal structure and semantic features of false news texts, and extracts key language modes and expression characteristics for distinguishing true news from false news. The machine learning technology comprises various forms such as supervised learning, unsupervised learning, deep learning and the like, and can automatically learn deep features of false news texts. The multi-mode technology breaks the limitation of single text analysis, and realizes the comprehensive analysis of various information carriers such as texts, images, audios, videos and the like, thereby capturing the characteristic difference of false news in multiple dimensions. However, the current false news detection system has the following technical problems: First, the multi-dimensional automatic analysis capability is insufficient. Most of the existing platforms rely on single dimension or manual auditing, lack the capability of comprehensively evaluating the authenticity of news from multiple angles, and cannot effectively process complex false information forms. Second, AI generates a content detection capability loss. With the development of large model technology, false news generated by AI is more and more difficult to identify, but most platforms have not developed a specific detection mechanism for such content. Third, multi-modal analysis is weak. The existing platform mainly focuses on text analysis, and lacks effective consistency detection on multi-mode content combined with graphics and texts, and the content has larger influence in transmission. Fourth, intelligent interaction and decision-making assistance capabilities are limited. The lack of application of advanced AI technology, especially the lack of application of large models and multimodal analysis in false news identification. Fifth, user engagement and educational functions are inadequate. Most of the existing platforms are provided with unidirectional information, lack of community interaction and user education functions, and cannot improve the media literacy and judgment capability of users. Disclosure of Invention The invention aims to solve the defects of the background technology and provides a false news intelligent detection system based on multi-dimensional fusion. By constructing a dynamic multi-mode joint detection system, the omnibearing and high-precision detection of false news is realized. The technical scheme adopted by the invention is that the false news intelligent detection system based on multidimensional fusion comprises The data acquisition processing module is used for acquiring and preprocessing news data from a plurality of news platforms, wherein the news data comprises news headlines, news texts, news configuration drawings, news videos, news audios, release platform information, propagation time records, propagation geographic position data and propagation user portrait data; The rapid detection module is used for performing real and false two-class rapid detection on the preprocessed news data by adopting a ERNIE 3.0.0 model which is finely adjusted based on a word segmentation-enhancement-pretraining three-level framework, and outputting a real and false two-class result, wherein the real and false two-class result comprises a real or false class label and a corresponding confidence score; the system comprises a multi-dimensional slow auditing module, a real-false two-classification quick detection module and a real-false two-classification quick detection module, wherein the news data is subjected to deep analysis from six evaluation dimensions, the six evaluation dimensions comprise news source confidence assessment, text consistency detection, AI generation content detection, multi-mode image-text consistency analysis, user transmission confidence assessment and logic judgment based on a large la