Rumor detection method based on user cognition deviation mining

Abstract

The invention relates to the technical field of natural language processing and deep learning, in particular to a rumor detection method based on user cognitive deviation mining. The method comprises the following steps: performing cognitive information extraction on user comments based on a big language model, and obtaining a standing judgment for news contents and a corresponding explanatory basis; respectively extracting embedded features of the news text, the news image and the user cognitive information by means of a CLIP multi-mode model, and carrying out feature fusion; the initial standing site label is optimized and calibrated through a knowledge distillation mechanism, so that the reliability of standing site identification is improved; a triple loss function is constructed, and the deviation degree between user cognition and news authenticity is mined and measured; and realizing end-to-end rumor classification and discrimination by fusing multi-modal features and introducing cognitive deviation. According to the method, the sensitivity of the model to cognitive deviation can be enhanced, the rumor detection accuracy and robustness are improved, good interpretability is achieved, and the method is suitable for a network public opinion analysis and fact verification system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, cognitive computing and deep learning, and particularly relates to a rumor detection method based on user cognitive bias mining, which aims to model and quantify the cognitive information contained in user comments to assist in determining the truthfulness of news. BACKGROUND

[0002] In recent years, with the continuous progress of Internet technology and the vigorous development of digital economy, the production mode, work form, learning path and lifestyle of human society have undergone an unprecedented profound transformation. Internet, as the core medium for the publication and dissemination of opinions, has shown novel characteristics of information production and dissemination under the shaping of platform characteristics. Social media platforms, as virtual spaces for users to publish information, exchange views and achieve global interaction, have accelerated the spread of information while becoming a breeding ground for the rapid spread of rumors. Under this trend, the effective governance of online rumors has become a hot topic of widespread concern in academia and practice.

[0003] Existing rumor detection methods mainly develop in three directions: first, based on factual content, language features and semantic features of text are extracted; second, based on writing style and emotional features, the emotional tendency and rhetorical style of information are analyzed; third, based on user feedback and social context, user comments, forwarding behavior and social network structure are used to assist in judgment. The above methods have achieved considerable performance on several benchmark datasets, but still face the problems of insufficient robustness and explainability in actual complex network environment.

[0004] From the perspective of psychology and cognitive science, users' judgment of information is deeply influenced by cognitive biases, including heuristic bias, confirmation bias, overconfidence, etc. Cognitive bias can lead individuals to make different or even opposite judgments when facing the same information, and this phenomenon is particularly evident in the comment section of social media.

[0005] Although existing research uses user stance as a feature, it mainly uses labeled support / oppose / neutral hard labels, ignoring the explanatory evidence and cognitive reasons behind the user's given stance, as well as the collective effect of cognitive differences among different users. Lack of systematic modeling of cognitive bias limits the in-depth understanding of rumor transmission mechanism and audience response, and also makes the detection model perform poorly when facing deceptive or highly influential false information.

[0006] Therefore, there is an urgent need for a method that can not only determine the user's position, but also extract the explanatory content behind the position, quantify the user's cognitive bias and integrate it into rumor detection. The present application is based on the above-mentioned needs, by combining the explanatory output of large language model, multi-modal feature fusion and triple constraint mechanism, realizing the deep mining of user cognition and its bias, thereby revealing the deep mechanism of rumor affecting different audience psychology, improving the accuracy, robustness and explainability of rumor detection, which has far-reaching significance for improving rumor detection. SUMMARY

[0007] In order to solve the problems in the background art, the purpose of the present application is to provide a rumor detection method based on user cognitive bias mining, by obtaining and utilizing the cognitive information expressed explicitly or implicitly by users in comments, combining the multi-modal features of news text and images, and constructing an interpretable, robust and efficient rumor detection model.

[0008] In order to achieve the above-mentioned purpose, the technical scheme of the present application is as follows:

[0009] S1: User cognitive information acquisition

[0010] The text and picture resources of the news to be detected and the corresponding several user comments are taken as input, and a pre-trained large language model (LLM) is used to automatically generate the position judgment (support, oppose, cannot determine) and its brief explanatory content of each comment by combining the customized prompt template, which is called user cognitive information. The prompt template example is: "Given a news content and its corresponding comments: [news text], user comments: [comment content], please judge the position of the comment on the news content [support / oppose / cannot determine], and explain the basis."

[0011] Through the processing of the large language model, the structured cognitive information tuple (s i ,e i ) is obtained, wherein s i is the position label, and e i is the explanatory text.

[0012] S2: Multi-modal feature extraction

[0013] The news text, news image and user cognitive information are input into a unified multi-modal encoding CLIP for feature extraction to obtain vector representation in a unified semantic space. The vector representation of the news is E_news∈R^D, and the cognitive vector representation of the i-th comment is E_cog i ∈R^D. For subsequent processing, the text of each comment and its explanatory basis are spliced and encoded.

[0014] News text characteristics: E_text = CLIP_text(news_text)

[0015] News image features: E_image = CLIP_image(news_image)

[0016] User cognitive characteristics: E_cog i = CLIP_text(cog_info) i )

[0017] By concatenating news text and image features, the overall news features are obtained: E_news = [E_text; E_image]

[0018] S3: Comment Stance Detection

[0019] To improve the consistency and accuracy of comment stance tags, this step uses the multimodal representation of news, E_news, and the user cognition representation, E_cog. i Perform distillation-style classification: separate E_news from E_cog i After concatenation, the data is input into a multilayer perceptron (MLP), which outputs a class probability distribution:

[0020] p i =softmax(MLP([E news E_cog i ])) (6)

[0021] Training using cross-entropy loss:

[0022] L stance =-∑y i log(p i (1)

[0023] Where y i Soft labels are given for the initial large model.

[0024] The category with the highest predicted probability is taken as the corrected position label. The prediction results are compared with the original labels: if they match, the original labels are retained; otherwise, they are replaced.

[0025] S4: Cognitive Bias Detection and Quantification

[0026] Based on the Pred_stances obtained in step S3, a set of triples for mining cognitive biases is constructed. The specific method is as follows: For each news sample, a set of comment vectors S_support that are labeled (or predicted) as support and a set of comment vectors S_oppose that are labeled as opposition are selected. The principle for constructing the triples is as follows:

[0027] 1) For fake news, the anchor is E_news. For positive news, choose the opposing comment vector (which should be closer to a), and for negative news, choose the supporting comment vector (which should be further away from a). This is the sample of fake news:

[0028] (E_news_fake,info_neg,info_pos).

[0029] 2) For true news, the anchor is E_news, positive comments are selected to support the news, and negative comments are selected to oppose the news. That is, the true news sample is: (E_news_true, info_pos, info_neg).

[0030] The above vectors are constrained using triplet loss:

[0031]

[0032] Where d(·) is usually taken as cosine distance or 1-cosine similarity, and α is the interval hyperparameter.

[0033] The following methods are used to quantify and utilize cognitive biases:

[0034] bias vector =[μ s ,σ s ,min s ,max s ,μ o ,σ o ,min o ,max o ,r support ,r oppose (3)

[0035] Among them are distance statistics supporting sets: mean μ s Standard deviation σ s Minimum value min s Maximum value (max) s The same applies to the distance statistics for opposing sets; the proportion statistics for set size: the percentage of supporting comments r support The percentage of opposing comments is r oppose .

[0036] bias vector As additional features, they are input into the final classifier along with news features and cognitive features to determine the authenticity of news.

[0037] S5: Rumor Detection

[0038] Aggregate the news vector E_news with the mean vector of several selected user cognition vectors E_cog:

[0039]

[0040] Here, mean_pooling(·) represents average pooling.

[0041] The aggregated vector and the cognitive bias vector from step S4 are used together as input to the classifier, and the final true / false probability distribution is output through the MLP.

[0042] Training is performed using binary cross-entropy loss:

[0043]

[0044] The total loss function for the entire process is:

[0045] L total =λ1L stance +λ2L triplet +λ3L rumor (5)

[0046] The specific beneficial effects of this invention include:

[0047] 1) This invention enhances the interpretability of the detection model by incorporating user cognitive information into the rumor detection task;

[0048] 2) By employing a stance distillation and label correction mechanism, the consistency and reliability of comment stance labels have been improved;

[0049] 3) By using triplet loss to constrain the relative relationship between news and supporting / opposing comments, the model becomes more sensitive to abnormal cognitive phenomena;

[0050] 4) Cognitive bias indicators provide the model with additional judgment criteria at the level of group behavior, thereby improving its robustness in complex propagation scenarios;

[0051] 5) This method is compatible with multimodal inputs (text, images) and pre-trained encoders, and is easy to engineer.

[0052] Signature and expansion. Attached Figure Description

[0053] Figure 1 This is an overall flowchart of a rumor detection method based on user cognitive bias mining according to the present invention;

[0054] Figure 2 This is a schematic diagram of the user cognitive information acquisition process in this invention;

[0055] Figure 3This is a schematic diagram of the multimodal feature extraction and position detection process in this invention;

[0056] Figure 4 This is a schematic diagram of the cognitive bias mining and quantification process in this invention; Detailed Implementation

[0057] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. It should be noted that those skilled in the art can make various modifications and improvements based on the concept of the present invention without departing from its spirit and essence, and all such modifications and improvements should fall within the scope of protection of the present invention.

[0058] according to Figure 1 As shown, this invention provides a rumor detection method based on user cognitive bias mining, the overall process of which mainly includes the following steps:

[0059] Step 1: Obtaining User Cognitive Information

[0060] By inputting the news text and user comment texts into a pre-trained large language model and combining them with preset prompt templates, the model obtains each commenter's stance label on the news content (including support, opposition, or uncertainty) and corresponding explanatory content. This stance label and explanatory content together constitute the user's cognitive information.

[0061] Step 2: Feature Extraction

[0062] The news text, news images, and user cognitive information are input into the CLIP model to obtain news text vector representations, news image vector representations, and user cognitive vector representations, respectively. These are then concatenated to obtain the overall news feature vector. The text and explanation portions of the user cognitive information can be concatenated before being input into the model to ensure the integrity of the representation.

[0063] Step 3: Predicting the Commentary Stance

[0064] The overall feature vector of the news article and the user's cognitive feature vector are input into a multilayer perceptron (MLP), which outputs a new stance prediction result. If the predicted stance matches the original stance obtained in step one, the original label is retained; otherwise, a new predicted label, denoted as Pred_stances, is updated. This process is trained using a cross-entropy loss function to optimize the accuracy of stance discrimination.

[0065] Step 4: Extraction and Quantification of User Cognitive Biases

[0066] In this step, triplet samples are constructed based on the modified Pred_stances. The overall feature vector of the news is used as the anchor. Positive and negative samples are selected according to the combination of news authenticity and commentary stance. The triplet loss function is then used for optimization to constrain the relative distance between news features and user cognitive features.

[0067] Define a cognitive bias quantification index: Divide comments into support and opposition sets according to Pred_stances, calculate the distance distribution between the comment vector and the news in each set, extract statistics such as mean, variance, extreme values, and set size proportion, and concatenate them into a cognitive bias feature vector bias_vector to comprehensively represent the antagonistic relationship between user cognition and news content.

[0068] Step 5: Rumor Detection

[0069] The overall feature vector of the news article is concatenated with multiple user-perceived feature vectors, and then input into a classifier for binary classification, outputting a true / false label for the news article. This step also uses the cross-entropy loss function for training.

[0070] like Figure 2 As shown, the user cognitive information acquisition process provided by this invention specifically includes the following steps:

[0071] 1) Receive the input news text content and user comments related to the news text content.

[0072] 2) Pre-design a prompt template adapted to the large language model. This template guides the model to output each commenter's stance on the news content and its corresponding explanatory content. For example, the stance may include three options: support, opposition, or uncertainty. The explanatory content explains the reasons why the commenter held that stance.

[0073] 3) Input the news text and comment text together into the large language model, and under the constraints of the prompt template,

[0074] The model output includes position labels and explanatory notes.

[0075] The above process enables the unified modeling of user comments' stances and reasons while ensuring interpretability, thereby providing reliable cognitive information support for subsequent feature extraction and rumor detection.

[0076] like Figure 3 As shown, the multimodal feature extraction and position detection process provided by this invention specifically includes the following steps:

[0077] 1) Input the news text content and news image content into the pre-trained multimodal feature extraction model to obtain the news text vector representation and the news image vector representation respectively; at the same time, input the user cognitive information obtained in step one (composed of comment text and explanatory content) into the same model to obtain the user cognitive vector representation.

[0078] 2) Concatenate or weightedly fuse the news text vector and the news image vector to obtain the overall news feature vector; then map it one-to-one with the user's cognitive vector representation to form an input pair.

[0079] 3) Input the overall feature vector of the news article and the user's cognitive vector into a multilayer perceptron (MLP) to output the predicted stance label corresponding to the comment. If the predicted stance label matches the original stance label obtained in step one, the original label is retained; otherwise, the original label is replaced with the predicted stance label, and the corrected stance label Pred_stances is updated. The prediction result is optimized using the cross-entropy loss function to ensure the accuracy of stance label discrimination.

[0080] like Figure 4 As shown, the cognitive bias mining and quantification process provided by this invention specifically includes the following steps:

[0081] 1) Based on the position tags Pred_stances obtained in step 3, user comments are divided into a support set and an opposition set.

[0082] 2) Using the overall feature vector of the news as an anchor, combined with the news's truthfulness / falseness labels, select user cognitive vectors that align with the ideal stance as positive samples, and select user cognitive vectors that contradict the ideal stance as negative samples, constructing triplet samples:

[0083] 3) Optimize the above samples using the triplet loss function to make the anchor closer to the positive one.

[0084] By moving away from negatives, the representation of cognitive biases can be explicitly constrained in the model.

[0085] Finally, the bias_vector is input into the subsequent classification module as an important reference feature for rumor detection, thereby explicitly introducing user cognitive bias information into the overall model and improving the accuracy and robustness of rumor detection.

[0086] The above descriptions are specific embodiments of the present invention and should be understood as illustrating the technical solutions of the present invention, rather than limiting the scope of protection of the present invention. Those skilled in the art can make various modifications and variations to the technical solutions of the present invention without departing from the spirit and substance of the present invention, and all such modifications and variations should be included within the scope of protection of the present invention.

Claims

1. A rumor detection method based on user cognitive bias mining, characterized in that, Includes the following steps: S1: User cognitive information acquisition: Input news content and user comments into the large language model to generate user cognitive information that includes stance tags and explanatory content; S2: Multimodal feature extraction: The CLIP model is used to extract feature vectors from news text, news images and user cognitive information respectively, and the news text and images are concatenated to obtain the overall features of the news. S3: Comment Stance Detection: Input the overall features of the news and the user's cognitive features into the MLP model, and correct the stance labels through knowledge distillation to obtain the optimized stance prediction results; S4: Cognitive Bias Mining and Quantification: Based on the corrected stance labels, triples are constructed, and the degree of deviation between user cognition and news authenticity is quantified through the triple loss function; S5: Rumor Detection: It integrates news features with aggregated user cognitive features and outputs rumor detection results through a classifier.

2. The rumor detection method based on user cognitive bias mining according to claim 1, characterized in that: In S1, user cognitive information is obtained through a large language model. The input news text and comment text are used to obtain the stance label (support, oppose, uncertain) and corresponding explanatory content using a prompt template. The prompt template format is: "Given a news article and its corresponding comment: [news text], user comment: [comment content], please determine the comment's stance on the news article [support / oppose / uncertain], and explain the basis." 3. The rumor detection method based on user cognitive bias mining according to claim 1, characterized in that: In S2, the CLIP model is used to extract features from news text, news images and user cognitive information to obtain vector representations E_news and E_cog of the same dimension. The overall news feature vector E_news is formed by concatenating news text features and news image features, i.e., E_news = [E_text; E_image].

4. The rumor detection method based on user cognitive bias mining according to claim 1, characterized in that, In step S3, E_news and E_cog are concatenated and input into a multilayer perceptron (MLP) for distillation classification, outputting predicted stance labels Pred_stances; the classifier is optimized using cross-entropy loss, and the stance prediction loss is calculated. The cross-entropy loss function is expressed as: L stance =-∑y i log(p i ) (1) Where y i As the initial position label, p i The position distribution predicted by MLP.

5. The rumor detection method based on user cognitive bias mining according to claim 1, characterized in that, In step S4, triples are constructed based on Pred_stances, and the triple loss function is used to optimize the triples, guiding the model to learn feature representations that can keenly capture the contradictory relationship between user cognition and news facts. The loss function is defined as: Where a i As a news feature, p i For cognitive features that are similar to expectations, n i Let α be the cognitive feature that is desired to be avoided, and let α be the interval hyperparameter.

6. The method according to claim 5, characterized in that, The cognitive bias quantification in step S4 further includes: constructing a set of multi-dimensional cognitive bias feature vectors to comprehensively characterize the antagonistic relationship and distribution characteristics between user cognition and news content. Cognitive bias feature vector bias vector Represented as: bias vector =[μ s ,s s ,min s ,max s ,m o ,s o ,min o ,max o ,r support ,r oppose ] (3) Among them are distance statistics supporting sets: mean μ s Standard deviation σ s Minimum value min s Maximum value (max) s The mean μ of the distance statistic of the opposing set o Standard deviation σ o minimum value min o maximum value o ; Proportional statistics of collection size: percentage of supported comments r support The percentage of opposing comments is r oppose .

7. The rumor detection method based on user cognitive bias mining according to claim 1, characterized in that, In S5, E_news is concatenated with multiple E_cog entries and input into the classifier, and news authenticity labels are output. During model training, the main classification cross-entropy loss and triplet loss are jointly optimized. The classification cross-entropy loss function is expressed as: Where y∈{0,1}, is the true label. The model is used to predict the probability that the news is a rumor. The total loss function is the weighted sum of the losses from each step: L total =λ1L stance +λ2L triplet +λ3L rumor (5)。

Images