A multi-domain false information detection method and device based on large model enhancement

Abstract

A multi-domain false information detection method and device based on large model enhancement, the method comprising: inputting an original news text and a corresponding domain label into a large language model to generate a structured analysis text; obtaining word-level and sentence-level representations of the original news text and the structured analysis text, fusing the word-level and sentence-level representations to determine a joint semantic feature representation; inputting the joint semantic feature representation into a hybrid expert module, the hybrid expert module comprising a plurality of expert networks, and determining a domain discriminative feature for each expert network; weighting and aggregating the domain discriminative features of the expert networks to generate a domain-adaptive deep discriminative feature; obtaining a usefulness score of the deep discriminative feature; based on the usefulness score, dynamically adjusting the fusion weight of the original news text feature and the deep discriminative feature through an adaptive fusion module, inputting the final fusion feature into a classifier, and outputting a false information detection result. The method can improve the accuracy of false news information detection.

Description

Technical Field

[0001] This invention relates to the field of information monitoring technology, and in particular to a method and apparatus for detecting multi-domain false information based on large model enhancement. Background Technology

[0002] With the widespread adoption of the internet and the popularity of social media, fake news has become a serious problem. Fake news not only misleads people but also has a severe impact on society. Therefore, fake news detection has become a popular research area. Currently, mainstream fake news detection methods include those based on content semantic analysis and those based on external knowledge verification. Deep learning-based content detection methods extract multi-granular semantic features from news texts using pre-trained language models and then use classifiers to determine their authenticity. However, these methods often rely solely on the surface semantic information of the news itself, lacking the ability to deeply reason about the logical facts, making it difficult to cope with increasingly complex and concealed false narratives. Furthermore, some studies have attempted to use large language models to generate reasoning paths to assist detection, but these still have limitations in practical applications. For example, the reasoning content generated by large language models may exhibit illusions, generating seemingly reasonable but actually erroneous reasoning bases. Direct use of these bases introduces noise, reducing the accuracy of fake news detection. Summary of the Invention

[0003] To address the problems existing in the prior art, embodiments of this application provide a method, apparatus, computing device, computer storage medium, and product containing a computer program for multi-domain fake information detection based on large model enhancement, which can effectively suppress large model illusion interference and improve the accuracy of fake news information detection.

[0004] In a first aspect, embodiments of this application provide a multi-domain misinformation detection method based on large model enhancement, comprising: inputting original news text and corresponding domain labels into a large language model to generate structured analysis text; generating an enhanced training sample set based on the structured analysis text; obtaining word-level representations and sentence-level representations of the original news text and the structured analysis text in the enhanced training sample set, respectively; fusing the word-level representations and the sentence-level representations through a multi-granularity attention interaction mechanism to determine a joint semantic feature representation; inputting the joint semantic feature representation into a hybrid expert module, the hybrid expert module comprising multiple expert networks, and determining the domain discriminant features of each expert network; weighting and aggregating the weights of each expert network with the domain discriminant features to generate a domain-adaptive deep discriminant feature; obtaining a usefulness score of the deep discriminant feature; dynamically adjusting the fusion weights of the original news text features and the deep discriminant feature based on the usefulness score through an adaptive fusion module; inputting the final fused feature into a classifier and outputting a misinformation detection result.

[0005] In some possible implementations, the original news text and its corresponding domain labels are input into a large language model to generate structured analysis text. This includes: constructing a domain-ID mapping table and converting domain names into numerical codes; adopting a domain-adaptive prompting strategy, dynamically calling predefined expert prompt word templates based on domain labels, and guiding the large language model to gradually generate structured analysis text containing logical reasoning processes, fact-checking points, and truth-or-false judgments according to the thought chain.

[0006] In some possible implementations, generating an enhanced training sample set based on the structured analysis text includes: performing format and consistency checks, comparing the discrimination results output by the large language model with manually annotated real labels to determine useful labels; and integrating the original news content, manually annotated real labels, structured analysis text, useful labels, and domain labels to obtain the enhanced training sample set.

[0007] In some possible implementations, determining the joint semantic feature representation includes: using a pre-trained language model to encode the original news text and the structured analysis text respectively, extracting the vectors corresponding to the [CLS] markers in their respective sequences as sentence-level semantic representations, and retaining the complete sequence output as word-level semantic representations; calculating the semantic relevance weights between the original news text and the structured analysis text at the word level through a cross-attention mechanism to generate word-level fusion representations; concatenating and linearly transforming the sentence-level semantic representations of the original news text and the structured analysis text to generate sentence-level fusion representations; and weighting and superimposing the word-level fusion representations and sentence-level fusion representations to determine the joint semantic feature representation.

[0008] In some possible implementations, generating domain-adaptive deep discriminative features includes: initializing a learnable domain embedding matrix and learning a domain-specific embedding vector for each domain; setting up multiple parallel expert networks, each employing a text convolutional neural network structure, and simultaneously inputting the joint semantic feature representation into all expert networks to obtain the domain discriminative features of each expert network; concatenating the domain embedding vector with the globally pooled representation of the joint semantic feature representation, and inputting it into a gated generative network composed of multilayer perceptrons to calculate the expert weight distribution; the formula is expressed as:

[0009] In the formula, This represents the weight of the i-th expert. This represents the vector concatenation operation. Indicates learnable parameters, Represents the domain embedding vector. This represents the joint semantic feature representation. The gating generative network is represented; based on the expert weight distribution, the domain discrimination features output by all expert networks are weighted and summed to determine the deep discrimination features.

[0010] In some possible implementations, obtaining the usefulness score of the depth discriminative features includes: inputting the depth discriminative features into a usefulness evaluation module composed of multilayer perceptrons, and outputting the usefulness score through a sigmoid function, expressed by the formula: In the formula, The depth discriminative features are represented; the usefulness score is used to characterize the quality of the information contribution of the structured analysis text to the detection task, so as to suppress the noise interference caused by the illusion of large language models.

[0011] In some possible implementations, based on the usefulness score, the fusion weights of the original news text features and deep discriminative features are dynamically adjusted by an adaptive fusion module. The final fused features are then input into a classifier, and the false information detection result is output. This includes: merging the sentence-level semantics of the original news text... The original news text features are represented as features; fusion weights are calculated based on the usefulness score, and the original news text features are then combined. The weights are scaled to adjust the depth discrimination features. The weights are scaled, and the two are concatenated to obtain the final fused feature, expressed by the formula: The final fused features are input into the classifier, which outputs the false information detection results.

[0012] In some possible implementations, the method also includes an auxiliary task: constructing a large language model to determine the prediction task, inputting structured analysis text into the pre-trained language model, extracting sentence-level semantic representations, and outputting the probability of truth prediction through an auxiliary predictor; the auxiliary task only takes effect during the training phase and does not participate in forward computation during the inference phase.

[0013] Secondly, embodiments of this application provide a multi-domain misinformation detection device based on large model enhancement, comprising: a generation module, used to input original news text and corresponding domain labels into a large language model to generate structured analysis text, and generate an enhanced training sample set based on the structured analysis text; a processing module, used to obtain word-level representations and sentence-level representations of the original news text and the structured analysis text in the enhanced training sample set respectively, and fuse the word-level representations and sentence-level representations through a multi-granularity attention interaction mechanism to determine a joint semantic feature representation; the processing module is further used to input the joint semantic feature representation into a hybrid expert module, the hybrid expert module including multiple expert networks, and determine the domain discriminant features of each expert network; weighting and aggregating the weights of each expert network with the domain discriminant features to generate a domain-adaptive deep discriminant feature; the processing module is further used to obtain a usefulness score of the deep discriminant feature; the processing module is further used to dynamically adjust the fusion weights of the original news text features and the deep discriminant feature based on the usefulness score through an adaptive fusion module, input the final fused feature into a classifier, and output a misinformation detection result.

[0014] Thirdly, embodiments of this application provide a computer-readable storage medium including computer-readable instructions that, when read and executed by a computer, cause the computer to perform the method as described in any of the first aspects.

[0015] Fourthly, embodiments of this application provide a computing device, including a processor and a memory, wherein the memory stores computer program instructions, which, when executed by the processor, perform the method as described in any of the first aspects.

[0016] Fifthly, embodiments of this application provide a product comprising a computer program that, when the computer program product is run on a processor, causes the processor to perform the method as described in any of the first aspects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a multi-domain misinformation detection method based on large model enhancement provided in an embodiment of this application;

[0019] Figure 2 This is a schematic diagram illustrating the construction process of an augmented dataset provided in an embodiment of this application;

[0020] Figure 3 This is a schematic diagram of a detection model training process provided in an embodiment of this application;

[0021] Figure 4 This is a schematic diagram of the structure of a multi-domain fake information detection device based on large model enhancement provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0023] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0024] The terms "first" and "second," etc., used in the specification and claims herein are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0025] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0026] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0027] To facilitate understanding of the embodiments of this application, the following will provide further explanation and description with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0028] In multi-domain fake news detection, a known technique is based on multi-view collaboration. The input news text is encoded into word embedding vectors using a BERT model, and then its sequential features are extracted using a bidirectional LSTM. Subsequently, four sub-modules are constructed: a semantic network, a domain network, a sentiment network, and a style network. The semantic network and the domain network directly operate on the text representation output by the BERT-LSTM to obtain semantic and domain-specific features, while the sentiment and style features rely on predefined external indicators (such as sentiment lexicon, readability, logicality, and eight other style dimensions) and are extracted by a hybrid expert system. Next, these four types of features are fed into a cross-view fusion module, where preliminary fusion is achieved by assigning weights to each view (such as weighted summation of semantic, sentiment, style, and domain local representations). Furthermore, a multi-head mechanism is used to generate multiple adaptive cross-view representations. Finally, the contributions of each view are dynamically adjusted by combining the domain label input domain gating function (softmax(MLP(g))), and the aggregated data is fed into a classifier with a softmax output layer to complete the real / fake news discrimination. This method, based on multi-perspective collaboration and expert network feature extraction, effectively utilizes domain features by introducing domain gate networks to learn the relationship between the domain and perspectives. However, this approach relies entirely on BERT+BiLSTM to extract text sequence features and combines them with manually predefined sentiment dictionaries, style dimensions, and other external metrics to construct features. It lacks external reasoning capabilities and cannot uncover the essential characteristics of fake news, such as internal logical contradictions and common sense conflicts, thus limiting its ability to detect complex fake news. Furthermore, this method directly weights and fuses manually designed multi-perspective features without evaluating their effectiveness. If the manually predefined metrics do not match the current detection scenario, or if the features contain irrelevant information, it will directly affect the detection accuracy.

[0029] In another known technical solution, based on large language model analysis guidance, the knowledge contained in the analysis of the large language model is selectively injected into a small language model, achieving collaborative detection between the large and small language models. Input news text is fed into a pre-trained small language model (such as BERT) for encoding, resulting in a news semantic representation. Simultaneously, the news is also input into the large language model, which, through specific prompting engineering techniques, generates corresponding analytical reasons from both textual description and common-sense reasoning perspectives. These reasons are also vectorized by another BERT encoder. Next, the algorithm enters the core interaction and selection phase: the news representation and each reason representation interact deeply through a bidirectional cross-attention mechanism, generating reason-enhanced news features and news-perceived reason features. Then, the system uses a reason usefulness evaluation module to dynamically judge the credibility and relevance of each reason based on the news content and weights its enhanced features. Simultaneously, a large model judgment and prediction module is used to understand the initial conclusions made by the large model based on each reason, to better utilize its logic. Finally, the weighted multi-perspective reason-enhanced features are adaptively fused with the original news features to form the final classification representation, which is then fed into a classifier for truth / falsehood judgment. However, the analytical perspectives used in this scheme when reasoning with a large language model are fixed, such as common sense, writing style, and writing intention. This method assumes that the analytical patterns of these perspectives are equally effective and consistent across all fields (health, technology, entertainment, etc.). It does not adopt different reasoning methods based on different news, and it only provides the reasoning content without making a judgment. The subsequent auxiliary tasks of the small model cannot fully learn and simulate the coherent decision-making process from multi-faceted evidence to the final judgment.

[0030] In view of this, embodiments of this application provide a multi-domain misinformation detection method based on large model enhancement.

[0031] For example, Figure 1 This illustration shows a flowchart of a multi-domain misinformation detection method based on large model enhancement provided in an embodiment of this application. Figure 1 As shown, the method may include the following steps:

[0032] S11: Input the original news text and its corresponding domain labels into the large language model to generate structured analysis text, and generate an enhanced training sample set based on the structured analysis text.

[0033] In this embodiment, the original news text and its corresponding domain labels are used as joint input. Through a domain-aware prompt routing mechanism, the large language model is driven to automatically call the corresponding professional prompt templates (such as medical verification templates, financial audit templates, etc.) based on different domain labels, generating a structured analysis text that is domain-specific and contains logical reasoning chains and fact-checking points, thereby constructing an enhanced training sample set that integrates original semantics and deep reasoning signals.

[0034] For details, please refer to Figure 2 First, we build an API interface for interaction between the raw news text data and the large model. For each piece of news text data to be processed, we obtain the raw text content of that data. The system takes the corresponding domain label d and the manually labeled authenticity label y. A domain-ID mapping table is constructed, for example, domain={'Technology':0,'Entertainment':2}, converting domain names into numerical codes that the model can process. Based on the domain label d of the input news text, a specialized prompt template matching that domain is automatically selected and invoked. The original news text is combined with the selected prompt template to construct a complete model input, driving the large language model to perform deep semantic reasoning, generating a structured analysis text R containing the logical reasoning process, fact-checking points, and the final truth / falseness judgment. This structured analysis text not only contains the large model's judgment result on the news's authenticity but also... It also includes detailed logical basis to support this judgment.

[0035] In some possible embodiments, a quality verification and cleaning mechanism is also set up to ensure the reliability and validity of the data generated by the large language model. To address the potential illusion problem that may arise during the inference process of the large model and to ensure the reliability of the enhanced data, the system performs a dual verification logic during the quality verification phase: first, format verification is performed to remove output text that does not conform to the preset JSON structure; second, consistency verification is performed to compare the judgment results of the large model with the manually labeled real tags. The comparison results are represented using usefulness tags. Usefulness tags are represented as follows:

[0036]

[0037] In the formula, Characterizing the authenticity of structured analytical text generated by a large language model, when the large model's discrimination results... Compared with manually labeled real labels Under the same conditions, it means that the judgment of the large model is completely consistent with the true label, denoted as . 0. When the large model makes a misclassification, and the large model's judgment is inconsistent with the true label, it is recorded as . 1.

[0038] After the above cleaning process, the original news content will be... 1. Manually labeled real-world tags (y); 2. Structured analysis text (R) generated by the large model; 3. Useful tags. By fusing the domain label d, a final enhanced dataset was constructed. Unless otherwise specified, all data described in the following description of the methods can be obtained from the enhanced dataset.

[0039] S12: Obtain word-level and sentence-level representations of the original news text and the structured analysis text respectively, and fuse word-level and sentence-level representations through a multi-granularity attention interaction mechanism to determine the joint semantic feature representation.

[0040] In this embodiment, multi-granularity fusion encoding is performed on the news content and the reasoning text generated by the large model. Word-level embeddings of the original news text and the structured analysis text are extracted separately using a pre-trained language encoder, and then sentence-level semantic representations are captured using a sequence encoding layer. Subsequently, attention interaction and feature fusion are performed on the semantic representations of the two types of text at both the word and sentence levels. Finally, the fused representations from the two levels are weighted and superimposed to obtain a joint semantic feature representation that combines the original semantic information with the external reasoning logic.

[0041] Specifically, since the original news texts mostly use declarative factual expressions, while the structured analysis texts generated by the large language model use logical analytical language, the two differ in linguistic style and semantic density. Directly concatenating the two and inputting them into a single encoder can easily lead to interference between semantic information from different modalities. Therefore, a dual-tower semantic encoding strategy is adopted, encoding the two types of text independently before deep fusion. Please refer to [reference needed]. Figure 3 For both the original news text and the structured analysis text generated by the large language model, the pre-trained language model BERT is used for tokenization. A [CLS] tag is added to the beginning of the input sequence to aggregate sentence-level semantics, and a [SEP] tag is added at sentence boundaries to distinguish different paragraphs. After passing through multiple Transformer encoders, deep context embeddings are extracted, represented as follows:

[0042]

[0043]

[0044] exist and Semantic representations at two different granularities are extracted from each sequence. The output vector corresponding to the [CLS] marker in each sequence is then taken. and Each of these serves as a sentence-level semantic representation of its respective text, representing the overall global semantic information of the text. The complete sequence output by the encoder is preserved. and As a word-level semantic representation, it represents the fine-grained semantic information of each local unit (e.g., word or subword) in the text. Attention interaction and feature fusion are performed on the semantic representations of the original news text and the structured analysis text at both the word and sentence levels. At the word level, through a cross-attention mechanism, each word representation of the original news text interacts with the word sequence of the structured analysis text, calculating the semantic relevance weights between them, and generating a weighted word-level fusion representation accordingly. This allows keywords in the original news text to obtain corresponding logical explanations and factual support from the analysis text. At the sentence level, the sentence-level vectors of the original news text are... Sentence-level vectors of the analyzed text The two text segments are concatenated and subjected to linear transformation and nonlinear activation to obtain a sentence-level fusion representation, which is used to capture the complementary relationship between the two segments in overall semantics. After completing the fusion at both granularities, the word-level fusion representation and the sentence-level fusion representation are weighted and superimposed to obtain a joint semantic feature representation, denoted as... Since word-level fusion representations preserve fine-grained information about local semantic interactions, while sentence-level fusion representations provide a macroscopic perspective of global semantics, combining the two with learnable weight coefficients yields a joint semantic feature representation. It includes the original semantic information of the original news, the external reasoning logic of the large language model, and the semantic structure from local to multi-level.

[0045] S13: Input the joint semantic feature representation into the hybrid expert module, which includes multiple expert networks, and determine the domain discriminative features of each expert network; weight the weights of each expert network and the domain discriminative features to generate domain-adaptive deep discriminative features.

[0046] In this embodiment, the joint semantic feature representation is input into a hybrid system composed of multiple domain expert networks, and each expert network extracts its preferred domain discrimination features in parallel. At the same time, the joint semantic features are compressed by attention pooling and then input into a gating generation network together with the domain label embedding vector. The gating network calculates the weight distribution of the outputs of each expert network, and finally dynamically weights and aggregates the features extracted by the expert networks based on the weights to generate a deep feature representation with domain adaptability.

[0047] Specifically, due to significant differences in language style, terminology, semantic emphasis, and the presentation patterns of misinformation across different fields, using a single feature extraction network to process news from all fields would be insufficient to fully capture the unique discriminative patterns of each field. Therefore, this embodiment designs multiple expert networks to extract various representations of the news. The i-th expert network is represented as... Where 1 ≤ i ≤ T. And W represents the fused features of the input. Let T represent the learnable parameters of the expert network, where T is a hyperparameter representing the total number of experts. Each expert network employs a TextCNN architecture to effectively capture local semantic features. The joint semantic features are represented as... Simultaneously, all expert networks are input, and each expert network independently extracts features from them, outputting the corresponding domain-discriminating features. The calculation formula is expressed as follows: .

[0048] To achieve optimal performance in multi-domain fake news detection tasks, the key lies in how to adaptively integrate the outputs of different experts based on the characteristics of the input news. While a simple averaging operation is intuitive, it ignores the differences in expertise among experts in specific domains, leading to the loss of crucial domain-specific information. Therefore, this embodiment employs a domain gating mechanism to dynamically allocate expert weights based on the domain background and specific content of the news. First, a learnable domain embedding matrix is ​​initialized for each domain k. This vector is continuously optimized during training, ultimately capturing the personalized distribution characteristics of the domain in the feature space. For the current input news, based on its domain label d, the corresponding domain embedding vector is retrieved from the domain embedding matrix. Simultaneously, the joint semantic features will be represented. The global representation vector is obtained by compression using an attention pooling layer. The domain embedding vector is then used. The global representation vector is concatenated to obtain the input vector for the gating network. This concatenated vector is then input into the gating generator network. A gated network is a feedforward network that includes a multilayer perceptron, containing several fully connected layers and nonlinear activation functions. The gated network outputs a T-dimensional original score vector (the number of expert networks), which is then normalized using the Softmax function to obtain the expert weight distribution. The formula is expressed as:

[0049]

[0050] In the formula, This represents the weight of the i-th expert. This represents the vector concatenation operation. This represents the learnable parameters.

[0051] Domain discrimination features of all expert network outputs using the calculated weight distribution 'a' Weighted summation is performed to obtain the final depth discriminative features. The formula is expressed as:

[0052]

[0053] S14: Obtain the usefulness score of the deep discriminative features.

[0054] In this embodiment, although the final depth discriminative features have been obtained, not all extracted features are useful. Large models may produce illusions, and direct fusion could introduce noise and mislead the detection model. Therefore, it is necessary to assess the usefulness of the depth discriminative features. This embodiment uses a usefulness evaluator to assess the usefulness of the depth discriminative features. The usefulness evaluator consists of a lightweight multilayer perceptron. The depth discriminative features... The input to this multilayer perceptron, after passing through several fully connected layers and nonlinear activation functions, outputs a scalar value. Subsequently, the scalar value is mapped to the [0,1] interval using the sigmoid activation function to obtain a usefulness score, denoted as . The formula is as follows:

[0055]

[0056] when When the value approaches 1, it indicates that the deep discriminative feature has a high information contribution quality, and the reasoning logic and fact-checking points in the corresponding structured analysis text have a positive effect on the detection of false information. When the value approaches 0, it indicates that the quality of the deep discriminative feature is low, which may be due to the illusion or reasoning bias of the large language model. Its contribution to the detection task is relatively limited, and it may even introduce interference.

[0057] S15: Based on the usefulness score, the fusion weights of the original news text features and deep discriminative features are dynamically adjusted through the adaptive fusion module. The final fused features are then input into the classifier, and the false information detection results are output.

[0058] In this embodiment, the system receives a usefulness score from a usefulness evaluation module, domain discriminative features from a domain-gated hybrid expert network, and semantic features from the original news text. The fusion weights of the inference enhancement features and the original semantic features are dynamically calculated based on the usefulness score, achieving adaptive feature selection and complementarity. The final concatenated features are then input into a classifier for final prediction.

[0059] Specifically, after obtaining the usefulness score, if the usefulness score is too low, the accuracy of the judgment based on it will also be low. Therefore, the original text content, the extracted fusion features, and the usefulness score can be integrated. The specific formula is as follows:

[0060]

[0061] The usefulness score and the extracted fusion features are integrated. Then, the original text content and the weights (calculated by subtracting the usefulness score from 1) are integrated, and the two are concatenated. This way, when the confidence of the fusion features is low, the model can focus on the original text and avoid continuously learning features with low confidence. After fusion, the final fusion features can be input into the classifier to obtain the final false information detection result. The classifier consists of a fully connected layer and a softmax function.

[0062] In some possible embodiments, in addition to the main detection process, this method introduces an auxiliary task module to further enhance the detection model's ability to understand the analytical text generated by the large language model, thereby improving the overall detection performance. It should be noted that this auxiliary task only takes effect during the model training phase and does not participate in forward computation during the inference phase, thus not increasing the inference overhead during actual deployment. As mentioned earlier, although the large language model possesses strong logical reasoning and fact-checking capabilities when generating structured analytical text, its final output of truth / falsehood judgment results... Possibly related to real labels Inconsistencies exist. That is, a large language model might provide a reasonable logical reasoning chain and correct fact-checking points in the analyzed text, but ultimately make an incorrect judgment. Conversely, there are also cases where the reasoning process is flawed but the judgment is correct by chance. If the detection model only learns the semantic representation of the analyzed text and ignores the logical mapping relationship between it and the judgment result, it cannot fully mine the deep information in the analyzed text. Therefore, this embodiment also uses an auxiliary task to enable the detection model to predict the authenticity of news based solely on the structured analyzed text R generated by the large language model, without relying on the original news text content. Specifically, the structured analyzed text R is input into the BERT encoder to extract the sentence-level semantic representation corresponding to its [CLS] tag. This representation is then input into an auxiliary predictor composed of a multilayer perceptron, and the authenticity prediction probability is output through the Sigmoid function, expressed as: This auxiliary prediction task is jointly trained with the main detection task, and its loss function, along with the main classification loss and the usefulness assessment loss, constitutes a multi-task learning objective. Through this auxiliary task, the detection model is forced to learn the ability to deduce corresponding conclusions from the logical arguments of the analyzed text, understanding which reasoning chains and verification points should lead to true conclusions and which should lead to false conclusions. After joint training with this auxiliary task, the deep discriminative features used by the detection model... This will incorporate richer logical reasoning semantics, thereby improving the accuracy and interpretability of the final misinformation detection. Furthermore, the parameters of this auxiliary task module can be discarded or retained as auxiliary output after training, without affecting the detection efficiency during the inference phase.

[0063] The above describes the multi-domain misinformation detection method based on large model enhancement provided in this application. It automatically calls corresponding templates through domain labels, guiding the large language model to generate domain-specific and logically complete explanatory analyses. Furthermore, it introduces a final judgment, enabling subsequent auxiliary tasks to fully learn and simulate the coherent decision-making process from multi-faceted evidence to the final judgment. A domain-gated hybrid expert network is constructed, and multiple parallel expert sub-networks are designed to learn specific feature patterns from different domains. A domain gating mechanism is introduced, using the embedding vectors generated from the domain labels of the news article and the fusion features of the news and reasoning to dynamically calculate the weight distribution of each expert network. A usefulness evaluation module and an adaptive fusion module are designed to perform quality evaluation and confidence quantification on the reasoning content generated by the large language model. A lightweight neural network outputs a scalar usefulness score. The usefulness score is received from the usefulness evaluation module, the domain discriminative features from the domain-gated hybrid expert network, and the semantic features of the original news text. Based on the usefulness score, the fusion weights of the two types of features (reasoning enhancement features and original semantic features) are dynamically calculated to achieve adaptive feature selection and complementarity. Through auxiliary modules, the detection model no longer merely perceives the semantics of external analytical text, but is forced to learn how to deduce corresponding conclusions from given logical arguments. This method generates reasoning evidence from perspectives such as logical consistency and common-sense fact-checking through large language model thought chain prompts, achieving an upgrade from static feature splicing to dynamic reasoning enhancement, effectively capturing deep logical loopholes in fake news. The usefulness evaluator can construct supervision signals based on the consistency between the large model's zero-sample predictions and the true labels, scoring the quality of reasoning evidence and dynamically weighting it, effectively suppressing interference from invalid information. For news in different fields, this method uses corresponding domain-customized prompts, guiding the LLM to act as experts in various corresponding fields, conducting in-depth and professional analysis. This ensures that high-quality, highly relevant domain-specific reasoning evidence is obtained from the information source. In addition, a final judgment is given, allowing the auxiliary task of the detection model to fully learn and simulate the coherent decision-making process from multi-angle evidence to the final judgment. Meanwhile, the domain-specific gating mechanism will also perform personalized extraction based on various domain features of the input. The domain-specific design from beginning to end ensures the high robustness and high accuracy of the model in multiple domains and scenarios.

[0064] Therefore, this method designs prompt templates for news in different fields, incorporating specific expert roles (such as medical experts and financial auditors) and thought chain guidance. These prompts instruct the large language model to perform in-depth analysis and simultaneously output the reasoning process and final judgment, fundamentally changing the nature of the enhanced signal. This method does not obtain static statistical features or general descriptive text, but rather expert analysis reports rich in domain knowledge, with complete logical chains and clear conclusions. It solves the problem of existing technologies being unable to perform deep logical analysis and compensates for the shortcomings of shallow reasoning and broken decision chains. Thus, downstream models can access and learn complete and interpretable decision-making basis, significantly improving the ability to detect complex fake news. Furthermore, the domain-gated hybrid expert network and dynamic feature fusion enable refined adaptation throughout the feature extraction and fusion process. This allows the model to dynamically adjust the feature fusion strategy based on specific news content and domain labels, retaining more key discriminative information through domain-specific feature extraction and splicing fusion. Thus, while maintaining its cross-domain generalization capabilities, the system has gained a keen ability to identify complex deception techniques in professional fields, providing a solution for multi-scenario fake news detection that is both adaptable, accurate, and stable.

[0065] It is understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. Furthermore, in some possible implementations, each step in the above embodiments may be selectively executed according to actual circumstances; it may be partially or fully executed, without limitation here. All or part of any feature of any embodiment of this application can be freely and arbitrarily combined without contradiction. The combined technical solutions are also within the scope of this application.

[0066] Based on the methods in the above embodiments, this application also provides a multi-domain fake information detection device based on large model enhancement. For example, Figure 4 A schematic diagram of a multi-domain fake information detection device based on large model enhancement is shown. Figure 4 As shown, the multi-domain fake information detection device 400 includes a generation module 401 and a processing module 402.

[0067] Among them, the generation module 401 is used to input the original news text and the corresponding domain labels into the large language model, generate structured analysis text, and generate an enhanced training sample set based on the structured analysis text;

[0068] Processing module 402 is used to acquire word-level and sentence-level representations of the original news text and the structured analysis text in the enhanced training sample set, respectively, and to fuse word-level and sentence-level representations through a multi-granularity attention interaction mechanism to determine the joint semantic feature representation;

[0069] Processing module 402 is further configured to input the joint semantic feature representation into the hybrid expert module, which includes multiple expert networks, determine the domain discriminative features of each expert network, and weight and aggregate the weights of each expert network with the domain discriminative features to generate domain-adaptive deep discriminative features;

[0070] The processing module 402 is also used to obtain the usefulness score of the depth discrimination feature;

[0071] The processing module 402 is also used to dynamically adjust the fusion weights of the original news text features and the deep discriminative features based on the usefulness score through the adaptive fusion module, input the final fused features into the classifier, and output the false information detection result.

[0072] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0073] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0074] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0075] It is understood that the processor in the embodiments of this application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor may be a microprocessor or any conventional processor.

[0076] The method steps in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0077] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0078] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application.

Claims

1. A multi-domain fake information detection method based on large model enhancement, characterized in that, The method includes: The original news text and its corresponding domain labels are input into a large language model to generate structured analysis text. This includes: constructing a domain-ID mapping table and converting domain names into numerical codes; employing a domain-adaptive prompting strategy, dynamically calling predefined expert prompt word templates based on domain labels to guide the large language model to gradually generate structured analysis text containing logical reasoning processes, fact-checking points, and truth / falseness judgments according to a thought chain approach; generating an enhanced training sample set based on the structured analysis text, including: performing format and consistency checks, comparing the discrimination results output by the large language model with manually annotated real labels to determine useful labels; and integrating the original news content, manually annotated real labels, structured analysis text, useful labels, and domain labels to obtain the enhanced training sample set. A pre-trained language model is used to encode the original news text and the structured analysis text respectively. The vectors corresponding to the [CLS] markers in each sequence are extracted as sentence-level semantic representations, and the complete sequence output is retained as word-level semantic representations. The semantic relevance weights between the original news text and the structured analysis text are calculated at the word level through a cross-attention mechanism to generate word-level fusion representations. The sentence-level semantic representations of the original news text and the structured analysis text are concatenated and linearly transformed to generate sentence-level fusion representations. The word-level fusion representations and sentence-level fusion representations are weighted and superimposed to determine the joint semantic feature representation. The joint semantic feature representation is input into a hybrid expert module, which includes multiple expert networks, and the domain discriminative features of each expert network are determined. The weights of each expert network are weighted and aggregated with the domain discriminative features to generate a domain-adaptive deep discriminative feature. Obtaining the usefulness score of the depth discriminative features includes: inputting the depth discriminative features into a usefulness evaluation module composed of multilayer perceptrons, and outputting the usefulness score through the Sigmoid function, expressed by the formula: In the formula, The depth discriminative features are represented; the usefulness score is used to characterize the quality of the information contribution of the structured analysis text to the detection task, so as to suppress the noise interference caused by the illusion of large language models; Based on the usefulness score, the fusion weights of the original news text features and the deep discriminative features are dynamically adjusted by the adaptive fusion module. The final fused features are then input into the classifier, and the false information detection result is output.

2. The method according to claim 1, characterized in that, The generated domain-adaptive depth discriminative features include: Initialize a learnable domain embedding matrix, and learn a domain-specific embedding vector for each domain; Multiple parallel expert networks are set up, each of which adopts a text convolutional neural network structure. The joint semantic feature representation is simultaneously input into all expert networks to obtain the domain discrimination features of each expert network. The domain embedding vector is concatenated with the global pooled representation of the joint semantic feature representation, and input into a gated generative network composed of multilayer perceptrons to calculate the expert weight distribution; the formula is expressed as: In the formula, This represents the weight of the i-th expert. This represents the vector concatenation operation. Indicates learnable parameters, Represents the domain embedding vector. This represents the joint semantic feature representation. This represents a gated generative network; Based on the aforementioned expert weight distribution, the domain discrimination features output by all expert networks are weighted and summed to determine the deep discrimination features.

3. The method according to claim 1, characterized in that, Based on the usefulness score, the fusion weights of the original news text features and deep discriminative features are dynamically adjusted through an adaptive fusion module. The final fused features are then input into the classifier, and the false information detection result is output, including: Sentence-level semantics of the original news text This is indicated as a feature of the original news text; The fusion weights are calculated based on the usefulness scores, and the features of the original news text are then combined. The weights are scaled to distribute the depth discriminative features. The weights are scaled, and the two are concatenated to obtain the final fused feature, expressed by the formula: ; The final fused features are input into the classifier, which outputs the false information detection results.

4. The method according to claim 1, characterized in that, The method also includes auxiliary tasks: To construct a large language model for the prediction task, the structured analysis text is input into the pre-trained language model, sentence-level semantic representations are extracted, and the probability of true prediction is output through an auxiliary predictor. The auxiliary task is only effective during the training phase and does not participate in forward computation during the inference phase.

5. A multi-domain fake information detection device based on large model enhancement, characterized in that, The device includes: The generation module is used to input the original news text and its corresponding domain tags into the large language model to generate structured analysis text. This includes: constructing a domain-ID mapping table and converting domain names into numerical codes; employing a domain-adaptive prompting strategy, dynamically calling predefined expert prompt templates based on domain tags to guide the large language model to gradually generate structured analysis text containing logical reasoning processes, fact-checking points, and truth / falseness judgments according to a thought chain approach; and generating an enhanced training sample set based on the structured analysis text, including: performing format and consistency checks, comparing the discrimination results output by the large language model with manually annotated real tags to determine useful tags; and integrating the original news content, manually annotated real tags, structured analysis text, useful tags, and domain tags to obtain the enhanced training sample set. The processing module is used to encode the original news text and the structured analysis text using a pre-trained language model, extract the vectors corresponding to the [CLS] markers in their respective sequences as sentence-level semantic representations, and retain the complete sequence output as word-level semantic representations; it calculates the semantic relevance weights between the original news text and the structured analysis text at the word level through a cross-attention mechanism to generate word-level fusion representations; it concatenates and linearly transforms the sentence-level semantic representations of the original news text and the structured analysis text to generate sentence-level fusion representations; and it weights and superimposes the word-level fusion representations and sentence-level fusion representations to determine the joint semantic feature representation. The processing module is further configured to input the joint semantic feature representation into a hybrid expert module, the hybrid expert module comprising multiple expert networks, determine the domain discriminative features of each expert network, and weight and aggregate the weights of each expert network with the domain discriminative features to generate a domain-adaptive deep discriminative feature; The processing module is further configured to obtain a usefulness score for the depth discriminative features, including: inputting the depth discriminative features into a usefulness evaluation module composed of a multilayer perceptron, and outputting a usefulness score through a Sigmoid function, expressed by the formula: In the formula, The depth discriminative features are represented; the usefulness score is used to characterize the quality of the information contribution of the structured analysis text to the detection task, so as to suppress the noise interference caused by the illusion of large language models; The processing module is also used to dynamically adjust the fusion weights of the original news text features and deep discriminative features based on the usefulness score through the adaptive fusion module, input the final fused features into the classifier, and output the false information detection result.

6. A computer-readable storage medium comprising computer-readable instructions that, when read and executed by a computer, cause the computer to perform the method as described in any one of claims 1-4.

Images