A plug-and-play human-machine hybrid misinformation detection method and device

Abstract

The application discloses a plug-and-play human-computer hybrid false information detection method and device, and relates to the technical field of information detection. The method comprises the following steps: obtaining a to-be-detected text, and obtaining original text representation by using a false information detection process; performing information compression and noise filtering processing on the original text representation to obtain purified features; performing feature decoupling processing based on the purified features to separate language fingerprints and deep semantics containing true-false discrimination related information; performing semantic enhancement processing based on the purified features, performing semantic compensation on the purified features through a cross-attention mechanism to obtain enhanced features, and then realizing adaptive fusion optimization of the enhanced features and the original text representation by using a word-level dynamic gating mechanism to obtain final semantic representation; and inputting the final semantic representation into a classification layer of a base detection model to output a false information detection result.

Description

Technical Field

[0001] This invention relates to the technical fields of natural language processing, misinformation detection, and human-intelligent content security, and more specifically to a plug-and-play human-machine hybrid misinformation detection method and device. Background Technology

[0002] With the rapid popularization of large language model technology, AI-generated text is appearing on a large scale in various online channels such as social media platforms, news websites, e-commerce review sections, online communication communities, government online text compliance monitoring channels, brand content authenticity monitoring platforms, content creation platforms, and teaching and research platforms. This has led to a hybrid human-machine communication model where human-written text and machine-generated text coexist. This further complicates the distribution patterns of biased content in online text, including false statements, exaggerated claims, biased viewpoints, distorted science popularization, distorted evaluations, and false comments. This places significant pressure on various fields such as online content management, verification of the authenticity of online content across the entire network, compliance verification of content on internet platforms, text review in the financial sector, verification of the authenticity of media content, and identification of scientific research and teaching content. Therefore, false information detection technology for human-machine hybrid text scenarios has become an important required technology in the fields of natural language processing and online content security.

[0003] Current mainstream misinformation detection technologies mainly rely on pre-trained language models to extract semantic representations of text, and then combine them with downstream classification networks to determine the authenticity of the text. This approach can achieve good detection results on single datasets containing only human-written text. However, when the training and inference data are mixed human-machine text, existing detection models will encounter the core technical problem of shortcut learning: the model will prioritize learning language fingerprints unique to machine text, such as sentence structure, word choice, and writing style, which are related to true / false labels, as the basis for judgment, while weakening the mining of deep semantic features shared by human and machine texts that can distinguish the authenticity of information.

[0004] This technical problem directly leads to two types of defects: First, the model's discrimination ability is unbalanced, with high accuracy in recognizing false text generated by machines but poor performance in recognizing false text written by humans, resulting in a significant gap in the detection indicators of the two types of text; Second, the overall comprehensive detection performance of the model has dropped significantly, making it unable to adapt to the real online text environment where humans and machines are mixed.

[0005] Meanwhile, existing technologies lack a supporting feature processing architecture. The original text representation is mixed with interference text features and effective discriminative semantics, and contains a lot of redundant noise. Conventional feature processing methods are prone to losing key clues for distinguishing between true and false. Moreover, most existing improvement schemes require modification of the backbone structure of the base pre-trained model. When changing the detection base, it is necessary to redevelop the adaptation module, which is not universal and has high engineering transformation costs.

[0006] In summary, existing misinformation detection models tend to rely on the unique writing characteristics of machine text to generate shortcut learning, resulting in poor judgment balance in human-machine mixed text. Furthermore, the cost of model modification and adaptation is high, and it is difficult to simultaneously meet the multiple processing requirements of feature denoising, feature decoupling, and semantic enhancement. Summary of the Invention

[0007] This invention provides a plug-and-play method and apparatus for detecting false information in human-machine hybrid systems. It addresses the problems of existing detection models relying on shortcut learning due to machine-specific language fingerprints, poor discrimination balance in human-machine hybrid text, high model adaptation and modification costs, and difficulty in simultaneously performing feature denoising, feature decoupling, and semantic enhancement.

[0008] This invention provides a plug-and-play human-machine hybrid method for detecting false information. This method is executed between the encoding step and the authenticity classification step in existing false information detection processes, and includes:

[0009] The text to be detected in the context of human-machine hybrid operation in the online content review scenario of social media and information platforms is obtained, and the text to be detected is encoded based on the encoder built into the existing false information detection process to obtain the original text representation.

[0010] The original text representation is subjected to feature aggregation and linear mapping to obtain global semantic features. The global semantic features are mapped to obtain latent variable Gaussian distribution parameters. The original KL loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterization sampling strategy.

[0011] Based on the purification features, parallel processing of source adversarial processing and semantic preservation auxiliary processing are performed to separate interfering language fingerprints and deep semantics for authenticity determination.

[0012] Based on the purified features, semantic compensation is performed using a cross-attention mechanism to obtain enhanced features; the enhanced features and the original text representation are adaptively fused and optimized using a word-level dynamic gating mechanism to obtain the final semantic representation.

[0013] The original text representation is replaced with the final semantic representation and input into the existing misinformation detection process to obtain the misinformation detection result.

[0014] This invention provides a plug-and-play human-machine hybrid false information detection device based on any human-machine hybrid false information detection method, comprising:

[0015] The acquisition unit is used to acquire the text to be detected in the context of human-machine hybrid network content review in social media and information platform scenarios, and reuse the encoder built into the existing false information detection process to encode the text to be detected to obtain the original text representation.

[0016] The first obtaining unit is used to perform feature aggregation and linear mapping on the original text representation to obtain global semantic features. The global semantic features are linearly mapped to obtain latent variable Gaussian distribution parameters. The original KL loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterized sampling strategy.

[0017] The second obtaining unit is used to perform source adversarial processing and semantic preservation auxiliary processing in parallel based on the purification features, and to separate the interfering language fingerprint and the deep semantic for true and false discrimination.

[0018] The third unit is used to obtain enhanced features by performing semantic compensation based on the purified features using a cross-attention mechanism; and to adaptively fuse and optimize the enhanced features and the original text representation through a word-level dynamic gating mechanism to obtain the final semantic representation.

[0019] The result unit is used to replace the original text representation with the final semantic representation in the existing misinformation detection process to obtain the misinformation detection result.

[0020] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform any of the above-described plug-and-play human-machine hybrid false information detection methods.

[0021] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs any of the above-described plug-and-play human-machine hybrid false information detection methods.

[0022] This invention provides a plug-and-play human-machine hybrid method and apparatus for detecting false information. The method solves for the original KL loss corresponding to the Gaussian distribution of latent variables and uses a preset threshold to truncate the loss to obtain the final effective KL loss. Combined with a reparameterized sampling strategy, it generates cleaned features that filter noise unrelated to the authenticity of the information within the original text representation. Based on the cleaned features, it performs feature decoupling and semantic preservation processing, separating the linguistic fingerprint from the deep semantics containing information related to authenticity judgment. Simultaneously, it uses the cleaned features to guide a cross-attention mechanism to supplement fine-grained semantics from the original text representation, obtaining enhanced features to compensate for missing details. A word-level dynamic gating mechanism is used to adaptively fuse and optimize the enhanced features and the original text representation, achieving dynamic fusion of enhanced features and the original representation word-by-word to obtain the final semantic representation with deep semantic enhancement. The final semantic representation is then fed into a multilayer perceptron for authenticity judgment in existing detection methods to complete false information detection, making the judgment process less reliant on linguistic fingerprints and focusing primarily on deep semantics for determination. Furthermore, during the training phase, source adversarial loss is used to extract text within the cleaned features to generate source-related features. At the same time, semantic preservation auxiliary loss is used to constrain the cleaned features to maintain the ability to distinguish between true and false information. The weighted total loss is constructed by combining the task loss of existing detection methods and the final effective KL loss to achieve multi-constraint collaborative optimization. The various losses counterbalance each other to prevent excessive feature compression and significantly improve the model's performance in detecting false information in human-machine hybrid text scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0024] Figure 1 This is a schematic flowchart of a plug-and-play human-machine hybrid false information detection method provided in an embodiment of the present invention;

[0025] Figure 2 A schematic diagram illustrating the sensitivity analysis of the joint loss weight parameters of the RoBERTa (Robustly Optimized BERT Pretraining Approach) model provided in this embodiment of the invention;

[0026] Figure 3 This is a schematic diagram of a plug-and-play human-machine hybrid false information detection device provided in an embodiment of the present invention. Detailed Implementation

[0027] 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, and 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.

[0028] Figure 1 This is a schematic flowchart of a plug-and-play human-machine hybrid false information detection method provided by an embodiment of the present invention; the following is in conjunction with... Figure 1 Taking this as an example, we will describe in detail a plug-and-play method for detecting false information involving human and machine interaction provided by an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0029] Step 101: Obtain the text to be detected in the context of human-machine hybrid operation in the online content review scenario of social media and information platforms, reuse the encoder built into the existing false information detection process to encode the text to be detected, and obtain the original text representation.

[0030] Step 102: Perform feature aggregation and linear mapping on the original text representation to obtain global semantic features. The global semantic features are linearly mapped to obtain the latent variable Gaussian distribution parameters. The original KL (Kullback-Leibler) loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterization sampling strategy.

[0031] Step 103: Based on the purification features, perform source adversarial processing and semantic preservation auxiliary processing in parallel to separate the interfering language fingerprint and the deep semantic for authenticity determination;

[0032] Step 104: Based on the cleaned features, semantic compensation is performed using a cross-attention mechanism to obtain enhanced features; the enhanced features and the original text representation are adaptively fused and optimized using a word-level dynamic gating mechanism to obtain the final semantic representation.

[0033] Step 105: Replace the original text representation with the final semantic representation in the existing misinformation detection process to obtain the misinformation detection result.

[0034] The plug-and-play human-machine hybrid misinformation detection method of this invention operates entirely between the encoding and truth / false classification steps of the existing misinformation detection process. The existing misinformation detection process is a deployed, complete misinformation detection chain, including sequentially executed encoding and truth / false classification steps. The encoding step encodes the human-machine hybrid text to be detected to obtain the original text representation, while the truth / false classification step receives feature vectors and performs a series of training tasks to obtain the misinformation detection result. Both are inherent processing logics of the existing process, and this invention does not modify its network structure or computational rules.

[0035] The entire feature processing flow of this invention is embedded as an additional computational step between two existing steps: The original text representation output from the existing encoding step is reused as input, and sequentially, the following steps are executed: human-machine hybrid text encoding, original text representation, feature cleanup, source adversarial processing and semantic preservation auxiliary processing, cross-attention semantic compensation, and gating fusion to obtain the final semantic representation. Then, the final semantic representation replaces the original text representation and is fed into the original true / false classification step to complete the discrimination. This deployment method achieves non-intrusive plug-in integration, requiring no modification to the existing false information detection process. It retains the discrimination capabilities of the existing false information detection process while separating interfering linguistic fingerprints from the text, strengthening the deep semantics of true / false discrimination, and comprehensively improving the performance of false information detection in human-machine hybrid text scenarios.

[0036] The plug-and-play human-machine hybrid misinformation detection method described in this embodiment is executed by a computer device, which includes a processor and a memory storing computer programs. The processor calls the detection program and model training program stored in the memory to automatically execute all processes, including text encoding, feature transformation, cross-attention calculation, loss calculation, feature fusion, and authenticity identification. Specifically, in online content review scenarios, the processor executes the inference program to complete the misinformation identification of the text to be detected; in offline model iteration and optimization scenarios, the processor additionally performs discriminative branch calculation and backpropagation of the total loss, updates all network parameters, and completes model training. Humans are only responsible for importing text data, configuring model hyperparameters, and starting the program; they do not participate in any feature calculations or discriminative calculations within the scheme.

[0037] In step 101, the text to be detected in the context of human-machine hybrid operation in the online content review scenario of social media and information platforms is obtained.

[0038] Specifically, the texts to be detected come from various online content channels, including social media platform images and comments, news and information website articles, e-commerce platform product reviews, online community posts, government online text compliance monitoring messages, brand content monitoring materials, texts generated by content creation platforms, teaching and research texts, etc. All texts constitute a mixed human-machine sample of human-written text and machine-generated text. At least one of the following methods is used to obtain the original text in batches: web scraping, reading from platform open content interfaces, and loading local datasets.

[0039] Furthermore, the text to be detected is encoded using the encoder built into the existing misinformation detection process to obtain the original text representation. It should be noted that the encoder architecture built into the existing misinformation detection process falls into two categories: one is a deep encoding network based on pre-trained Transformer, and the other is a traditional statistical encoder based on dictionary mapping.

[0040] Specifically, the former involves stacking multiple Transformer encoders, with backbone parameters directly reused after pre-training on a general corpus. During encoding, the text sequence is fed into each encoding layer sequentially after word embedding. Within each layer, multi-head self-attention, feedforward mapping, and layer normalization are performed sequentially, extracting shallow syntactic and deep semantic features layer by layer. The final output is a multi-dimensional feature vector sequence of the same length as the input, i.e., the original text representation. The latter relies on a pre-built dictionary (such as bag-of-words, TF-IDF, or Word2Vec static embedding), mapping lexical units to sparse or dense statistical feature vectors through dictionary lookups, without requiring parameter training.

[0041] The encoder outputs a multidimensional feature vector sequence that corresponds one-to-one with the length of the input text sequence. This feature vector sequence is denoted as the original text representation. The original text representation consists of basic sequence features that have not undergone noise reduction, feature compression, or semantic enhancement. It serves as the input data for the subsequent variational information bottleneck encoder, cross-attention module, and word-level dynamic gating fusion module.

[0042] In step 102, feature aggregation and linear mapping are performed on the original text representation to obtain global semantic features, and the global semantic features are linearly mapped to obtain the latent variable Gaussian distribution parameters.

[0043] Specifically, the original text is represented as a two-dimensional sequence feature vector of variable length. Its dimension is determined by the number of text lexical units and the feature dimension of each word. It cannot be directly used to solve the global distribution parameters. Therefore, feature aggregation and linear mapping are performed first.

[0044] Global pooling compresses the sequence features of the entire text into a single one-dimensional vector, completing sequence dimension aggregation. The pooling output is fed into a linear transformation layer for linear mapping operations, followed by ReLU (Rectified LinearUnit) and layer normalization processes, outputting a fixed-dimensional single vector. This vector is defined as the global semantic feature, specifically determined by the following formula:

[0045] (1)

[0046] in, Represents global semantic features, with dimensions of , Represents the training batch size. Represents the space of real numbers. The feature dimension representing the global semantic features. Represents the original text representation. This represents a global pooling operation, used to compress sequence features into a global vector. This indicates that the activation function of the linear unit is modified to enhance the nonlinear expression of the model. The representation layer normalization operation stabilizes the feature distribution. This represents the weights of the linear layer output by pooling within the variational information bottleneck module. Indicates the pooling output linear layer bias within the variational information bottleneck module, and identifies the prefix. This indicates that the parameter belongs to the variational information bottleneck module, and the subscript number is used to distinguish parameters of different network layers within the module.

[0047] Furthermore, using global semantic features as input, two independent linear mapping branches with no parameter sharing are constructed, and each branch independently performs the linear projection operation:

[0048] The first path is the mean branch, which calculates the latent variable mean vector through linear projection. The second path is the log-variance branch, which calculates the latent variable log-variance vector through another set of independent linear projection weights and biases.

[0049] The Gaussian distribution parameters of the latent variables include the latent variable mean vector and the latent variable log-variance vector, which are determined by the following formulas:

[0050] (2)

[0051] (3)

[0052] in, Represents the vector of latent variable means. Represents the log-variance vector of latent variables. This represents the linear projection weight of the mean branch within the variational information bottleneck module. This indicates the bias of the linear layer in the mean branch within the variational information bottleneck module. This represents the linear projection weights of the log-variance branch within the variational information bottleneck module. This indicates the bias of the linear layer in the log-variance branch within the variational information bottleneck module. Represents global semantic features.

[0053] Furthermore, the original KL loss is obtained based on the Gaussian distribution parameters of the latent variables, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterization sampling strategy.

[0054] Specifically, the Gaussian distribution parameters of the latent variables output above include the latent variable mean vector and the latent variable logarithmic variance vector, which together describe the Gaussian distribution followed by a single sample latent variable. In this embodiment of the invention, this Gaussian distribution is compared with the standard zero-mean, unit-variance Gaussian distribution. Perform a dimension-wise KL divergence calculation to obtain the divergence loss of the difference between the latent variable's Gaussian distribution and the standard Gaussian distribution, which is the original KL loss shown below:

[0055] (4)

[0056] To prevent the loss of effective information due to over-compression of features, a fixed KL divergence truncation threshold is introduced during the optimization process. The difference between the original KL loss and the truncation threshold is calculated, and the larger of the difference and 0 is used as the upper limit of the constraint on the original KL loss, i.e., the final effective KL loss. In practical applications, if the original KL loss is less than the truncation threshold, the result is 0, and no KL constraint is applied to the model; only when the original KL loss exceeds the threshold is the portion exceeding the threshold retained as the effective constraint loss, avoiding over-compression of text semantic features. The final effective KL loss is determined by the following formula:

[0057] (5)

[0058] in, This represents the original KL loss, used to measure the degree to which the feature distribution deviates from the standard Gaussian distribution; This represents the final effective KL loss. The dimension index representing the purification feature. , This represents the total dimensionality of a single sample's purified feature vector. Indicates the first The mean parameter of the latent variable. Indicates the first The variance parameter of the latent variable. This represents the KL divergence cutoff threshold. This represents the summation operation. This represents the function that takes the maximum value. Indicates the first The squared term of the mean parameter of the latent variable.

[0059] In practical applications, reparameterized sampling is used to achieve backpropagation of gradients, starting with a standard Gaussian distribution. The random sampling noise vector is used; combined with the latent variable mean vector and log-variance vector, the sample latent variables are reconstructed and used as the cleaned features. The cleaned features are determined by the following formula:

[0060] (6)

[0061] in, Indicates purification characteristics, Indicates the standard Gaussian distribution A noise vector randomly sampled from the middle. This represents the element-wise Hadamard product operation on vectors. This represents the exponential function, used to recover variance from logarithmic variance.

[0062] In step 103, source adversarial processing and semantic preservation auxiliary processing are performed in parallel based on the purification features to separate the interfering language fingerprint and the deep semantic for authenticity determination.

[0063] The method provided in this embodiment of the invention sets up a dual-path auxiliary supervision branch during the model training phase, namely a generation source adversarial supervision branch and a semantic preservation auxiliary supervision branch. Both branches take cleaned features as input and respectively realize interference feature removal and core semantic preservation constraints. The specific implementation is as follows:

[0064] First, a source discriminator is constructed, which is built by concatenating a Gradient Reversal Layer (GRL) and a Multi-Layer Perceptron (MLP). The cleaned features obtained during training are input into this source discriminator. First, the features are transferred through the Gradient Reversal Layer, then the MLP performs feature mapping and nonlinear transformation, finally outputting the predicted probability of the text's source. This forward computation process satisfies the following formula:

[0065] (7)

[0066] Furthermore, source adversarial loss is calculated based on the predicted probability of text generation source and the actual source label. The source adversarial loss adopts the form of binary classification cross-entropy loss, and the specific calculation formula is as follows:

[0067] (8)

[0068] in, Indicates purification characteristics, Indicates the first The predicted probability of text generation source output by the source discriminator for each sample. This represents the sample index within a single training batch. This indicates a gradient reversal layer, used for adversarial training to reverse the gradient propagation direction. This represents the hyperparameter indicating the gradient reversal strength. This represents a multilayer perceptron. This indicates that the source of the generation is against the loss. This indicates the total number of samples within a single training batch. Indicates the first Each sample has a binary label representing its actual source: 0 for human-written text and 1 for machine-generated text. This represents the logarithmic operation with the natural constant e as the base.

[0069] During training, the source discriminator is dedicated to accurately identifying the source of text generation based on sanitization features, while the gradient inversion layer reverses the gradient, enabling the front-end feature extraction network to continuously optimize parameters and gradually weaken the language fingerprints contained in the sanitization features, such as writing style, sentence structure, and machine-generated traces, which are falsely associated with the text's authenticity labels. This avoids the model relying on language fingerprints to generate shortcut learning and improves the model's dependence on the true semantics of the text.

[0070] Simultaneously, this method constructs an independent auxiliary discriminator, which is composed of a multilayer perceptron and is independent of the parameters of other discriminant networks. The purified features are directly input into the auxiliary discriminator, where feature transformation and semantic mining are performed by the multilayer perceptron, outputting the probability of false information to assist in prediction. This calculation process satisfies the following formula:

[0071] (9)

[0072] Furthermore, a semantic preservation auxiliary loss is constructed based on the probability prediction of inaccurate information and the true / false labels of the samples. This loss is used to constrain the purification of features to retain effective true / false discrimination semantics, avoiding excessive loss of effective semantics caused by the aforementioned feature denoising and adversarial decoupling process. The formula for calculating the semantic preservation auxiliary loss is as follows:

[0073] (10)

[0074] in, The output of the semantically preserving auxiliary discriminator represents the first... Inaccurate information from individual samples is used to assist in predicting probabilities. This represents semantic preservation auxiliary loss. This represents the total number of samples in a single training batch. Indicates the sample index within the batch; Indicates the first The true and false labels of each sample.

[0075] This branch directly uses the denoised cleaned features as the sole criterion for supervised training to distinguish between true and false information. This forces the cleaned features to retain the core semantic features that can distinguish between real and false information while removing irrelevant interference features. This achieves a two-way balance between feature denoising and semantic preservation, effectively improving the quality and robustness of the model's final semantic representation.

[0076] It should be noted that the source discriminator, auxiliary discriminator, and corresponding source adversarial loss and semantic preservation auxiliary loss only participate in the iterative optimization of network parameters during the offline training phase of the model; during the online inference phase of the model, the above two auxiliary branches are no longer used, only the main task detection inference link is retained, and no inference computation overhead is increased.

[0077] In step 104, based on the cleaned features, semantic compensation is performed using a cross-attention mechanism to obtain enhanced features.

[0078] In practical applications, the cleaned features are denoised semantic features obtained after being constrained by variational information bottlenecks. They then proceed to step 103 to perform source adversarial processing and semantic preservation auxiliary processing simultaneously. However, this process may suffer from insufficient expression of some fine-grained semantic cues and the possibility of accidental deletion of information. Therefore, this embodiment of the invention introduces a cross-attention mechanism, using the original text representation as a semantic compensation source and the cleaned features as query guidance to filter and supplement key semantic information from the original text representation.

[0079] Specifically, the cleaned features are input into the query linear transformation layer to generate a query vector; the original text representations are input into the key linear transformation layer and the value linear transformation layer respectively to generate key vectors and value vectors. Attention weights are calculated using the query vector and the key vector, and then the value vectors are weighted and summed using the attention weights to obtain the semantically compensated enhanced features.

[0080] The cross-attention calculation process can be represented as follows:

[0081] (11)

[0082] (12)

[0083] (13)

[0084] (14)

[0085] in, This represents the query matrix of the cross-attention module, obtained by projecting cleaned features. The cross-attention module key matrix is ​​obtained by projecting the original text representation. The cross-attention module value matrix is ​​obtained by projecting the original text representation. Feature dimensions representing keys and query vectors; This represents a text padding position mask matrix that masks meaningless padding characters. This represents the semantic compensation information output by cross-attention, used to supplement the semantics of cleaned-up features; This represents the normalized activation function, used to normalize the attention weights to an interval. , This indicates that the cross-attention key branch can learn the projection matrix; This indicates that the cross-attention value branch can learn the projection matrix. Represents the query vector in the cross-attention mechanism. The corresponding learnable weight matrix, Represents the key vector The transpose of the matrix, This represents the square root operation.

[0086] The semantic compensation feature carries fine-grained semantic cues weakened by the variational information bottleneck denoising process in the original text representation, as well as semantics that may be accidentally deleted in step 103, and is used to make up for the details lost by the purification feature.

[0087] In practical applications, not all samples require the same level of supplementation to the original text semantics. Therefore, an adaptive weight network is set to dynamically control the injection ratio of compensation information, and its expression is:

[0088] (15)

[0089] in, This represents the Sigmoid activation function, which constrains the output to a range between 0 and 1. This represents the upper-level linear matrix of the cross-compensation weight network within the semantic dynamic enhancement module. This represents the underlying linear matrix of the cross-compensation weight network within the semantic dynamic enhancement module. This indicates the upper-level bias of the cross-compensation weight network within the semantic dynamic enhancement module. This indicates the underlying bias of the cross-compensation weight network within the semantic dynamic enhancement module. Indicates semantic compensation adaptive injection weights, with identifier prefixes. This indicates that the parameter belongs to the semantic dynamic enhancement module, and the subscript number is used to distinguish parameters of different network layers within the module.

[0090] In this embodiment of the invention, if the purification feature itself is semantically complete and reliable, its weight approaches 0, thus weakening the fusion ratio of the semantic compensation feature; if the purification feature is severely lacking in detail, its weight approaches 1, thus amplifying the contribution of the semantic compensation feature to the overall feature.

[0091] Furthermore, the cleaned features are linearly transformed to preserve the basic discriminative semantics, while the semantic compensation features are multiplied element-wise with the adaptive injection weights and then superimposed onto the cleaned feature branch, resulting in the enhanced features shown below:

[0092] (16)

[0093] in, This represents the enhanced feature obtained by fusing cleanup features and cross-semantic compensation information; This represents the element-wise Hadamard product operation on vectors. This represents the linear weight matrix of the enhanced feature fusion layer within the semantic dynamic enhancement module. This indicates semantic compensation adaptive injection weights.

[0094] In this embodiment of the invention, the enhanced features retain the pure true / false semantics after the language fingerprint is stripped from the clean features, and adaptively supplement the key fine-grained semantics in the original text, thus balancing noise reduction effect and semantic integrity.

[0095] Furthermore, the enhanced features and the original text representation are adaptively fused and optimized through a word-level dynamic gating mechanism to obtain the final semantic representation.

[0096] Specifically, using the original text representation obtained in step 101 as input, independent fusion control weights are generated per word through a linear mapping layer combined with a Sigmoid activation function, with the corresponding calculation formula:

[0097] (17)

[0098] in, This represents word-level dynamic gating weights, which adaptively control the intensity of enhanced feature injection for each word in the text. This represents the linear weight matrix of the word-level dynamic gating network within the semantic dynamic enhancement module. This represents the bias vector of the linear layer of the lexical-level dynamic gating network within the semantic dynamic enhancement module.

[0099] The fusion control weights are generated independently based on the semantic information of each word in the original text representation, and different word positions can obtain differentiated weight values.

[0100] Furthermore, using the obtained word-level dynamic gating weights, element-wise weighted operations are performed on the enhanced features, and then the weighted enhanced features are superimposed on the original text representation. The calculation formula is as follows:

[0101] (18)

[0102] in, This represents the final semantic representation of the input base model classification layer after adaptive fusion optimization. Represents the original text representation. Indicates enhanced features, This represents the element-wise Hadamard product operation on vectors. The representation layer normalization operation standardizes the final feature distribution.

[0103] For each lexical position, the gating weights dynamically adjust the proportion of the enhanced features in the fusion result: the larger the weight value, the higher the contribution of the enhanced features at that position; the smaller the weight value, the more information of the original text representation is preserved.

[0104] After word-by-word weighted fusion and layer normalization, the complete sequence features are output, which constitute the final semantic representation. This final semantic representation integrates the stable semantic expression of the purified features after noise reduction, while also retaining the fine-grained semantic information from the original text representation. Furthermore, word-level dynamic gating reduces the interference of noisy words on the discrimination results, providing a more balanced and robust semantic basis for subsequent text authenticity classification.

[0105] In step 105, based on the final semantic representation and the existing misinformation detection process, the misinformation detection result is obtained. The specific execution process is as follows:

[0106] The final semantic representation is a sequence feature at the word level, which serves as the input to the task layer of existing misinformation detection methods.

[0107] In this embodiment of the invention, the task layer of the existing false information detection method is used to deeply mine the true and false semantic discrimination information in the text. It is not limited to a specific structure, but is flexibly constructed according to specific task requirements. It may include a variety of loss functions and network components to adapt to the optimization objectives under different business scenarios.

[0108] As an optional implementation, the task layer can adopt a multilayer perceptron structure, consisting of at least two fully connected hidden layers and an output layer in sequence. Nonlinear activation functions are introduced between layers to perform linear mapping and nonlinear transformation on the comprehensive features of the input layer by layer, so as to extract high-level discriminative features that help distinguish between true and false text. The output layer uses the sigmoid activation function to compress the output of the previous layer to the [0,1] interval, which serves as the predicted probability that the current text to be detected belongs to false information.

[0109] Furthermore, the system can pre-set a discrimination threshold, compare the output probability with the threshold, generate a final binary label, and output the false information detection result. For example, the threshold can be set to 0.5: if the predicted probability is not lower than this threshold, it is judged as false information; otherwise, it is judged as not false information. In actual deployment, the discrimination threshold can be flexibly adjusted according to the different emphases on precision and recall in specific application scenarios to balance detection performance and business needs.

[0110] The method provided in this invention aims to achieve multi-dimensional collaborative constraint optimization of the model, taking into account text feature denoising, interference feature removal, core semantic enhancement, and accurate identification of misinformation. It integrates the task loss, final effective KL loss, source adversarial loss, and semantic preservation auxiliary loss of existing misinformation detection methods to construct a joint optimization total loss for the model. Through a multi-loss weighted collaborative constraint approach, it guides the iterative update of network parameters, addressing issues such as optimization bias, feature imbalance, and insufficient discriminative balance inherent in single loss constraints. The specific formula for calculating the total model loss is as follows:

[0111] (19)

[0112] in, This represents the total loss from the joint optimization of the model. This represents the task loss of existing misinformation detection methods, which may be the sum of multiple loss functions. The weight hyperparameters representing the variational information bottleneck truncation of the original KL loss are... This represents the weight hyperparameters of the source adversarial loss. The weight hyperparameters represent the semantic preservation auxiliary loss. express , This indicates that the source of the generation is against the loss. This represents semantic preservation auxiliary loss.

[0113] In practical applications, existing methods for detecting misinformation suffer from task losses. This is the sum of all loss terms inherent to the base model itself, which are defined during the model training phase and constrained by the embedded plugin of this invention. , and Together, they participate in the training process of the misinformation detection task. This comprehensive loss relies on the final semantic representation to drive the model to complete the binary classification of misinformation in the text. By supervising the main task discrimination network through true and false labels, it learns the deep semantic features of the text, thereby ensuring the basic detection accuracy of the model.

[0114] As a feature denoising constraint, it weakens meaningless redundant features and noise information in the original text representation by normalizing the distribution of latent variables in the text. Hyperparameter The intensity of the adjustable feature denoising is greater. The larger the value, the stronger the model is at removing redundant noise features. It can be flexibly optimized according to the noise distribution of the dataset.

[0115] Source of counter-loss As a feature decoupling constraint, it works in conjunction with the gradient reversal layer to complete adversarial training, prompting the purification of linguistic fingerprints related to the source of text generation and suppressing the model's learning of linguistic fingerprints that are pseudo-correlated with true and false labels. This helps alleviate the model's shortcut learning problem. Hyperparameters This is used to adjust the constraint strength of adversarial decoupling, accurately balance the discriminator's recognition ability and the feature decoupling effect, and avoid problems of over-decoupling or under-decoupling.

[0116] Semantic Preservation Auxiliary Loss As a semantic fidelity constraint, to address the problem of excessive loss of core semantics during adversarial decoupling, it directly supervises the authenticity judgment of the purified features, ensuring that the denoised features retain sufficient effective discriminative semantics. Hyperparameters The weights used to regulate semantic retention constraints ensure that while purifying and reducing noise, the cleansing features retain the key semantic information needed to identify false information to the greatest extent possible.

[0117] The method for detecting false information provided in this invention has at least the following beneficial effects:

[0118] (1) In the human-machine hybrid scenario where human-written text and machine-generated text coexist, this invention relies on the variational information bottleneck module, the generation source adversarial processing, and the semantic preservation auxiliary processing to achieve feature decoupling, reduce the interference of text language fingerprints on model discrimination, and strengthen the deep discrimination semantics through cross-attention semantic compensation and word-level dynamic gating adaptive fusion optimization, so that the model has stable recognition ability for both types of text and improves the overall accuracy of false information detection.

[0119] (2) The present invention obtains purified features by constraining the latent variable Gaussian distribution and reparameterizing sampling, which initially weakens the redundant features and noise information in the original text representation.

[0120] (3) Based on the purification features, this invention achieves deep semantic separation of language and truth-false correlation through the generation of source adversarial loss; at the same time, it uses semantic preservation auxiliary loss to constrain the purification features to retain key semantics for discrimination, effectively improving the model's balanced discrimination performance for texts from different sources.

[0121] (4) Based on the purification features, the present invention uses the cross-attention mechanism to carry out semantic compensation to obtain enhanced features, and then uses the word-level dynamic gating mechanism to adaptively fuse and optimize the enhanced features and the original text representation, and adaptively adjusts the semantic enhancement weights word by word to generate a more complete final semantic representation, thereby further improving the effect of identifying false information.

[0122] (5) The entire feature processing, semantic compensation, adaptive fusion, and multi-loss collaborative constraint architecture of this invention can be encapsulated as a lightweight SDDE (Semantic Decoupling and Dynamic Enhancement for False Information Detection) pluggable plugin. It does not require modification of the pre-trained text encoding network and the core structure of the true and false classification steps of the existing false information detection process. It can adapt to various false detection bases, has low deployment cost and strong versatility. While controlling the inference overhead, it can improve the accuracy of human-machine hybrid text detection in a balanced way, and has high engineering practical value.

[0123] The following verifies the method for detecting false information provided in this embodiment of the invention: Tables 1, 2, and 3 contain the following annotations: HR (Human Recall), HF (Human F1-score), MR (Machine Recall), MF (Machine F1-score), AvgΔ (Average Delta), ACC (Accuracy), Real_F1 (F1 score for real text categories), Fake_F1 (F1 score for fake text categories), BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly Optimized BERT Pretraining Approach), EANN (Event Adversarial Neural Network), and MDFEND (Multi-Domain Fake News Detection). Network (Multi-domain Fake News Detection Network), SheepDog (baseline detection model code), ARG (Adaptive Rationale Guidance), DMint (Deep Multi-scale Interaction Network), LFND-AB (LLM (Large Language Model) - assisted Fake News Detection with Adaptive Boosting), w / o aux (without auxiliary loss), w / o adv (without adversarial loss), w / o vib (without variational information bottleneck), w / o adp (without adaptive fusion), SDDE (Semantic Decoupling and Dynamic Enhancement plugin).

[0124] Table 1 presents a comparison of detection performance before and after integrating SDDE into various base models. It showcases the performance comparison results of the proposed plug-and-play human-machine hybrid false information detection SDDE plugin on different base detection models. The experiments were conducted on F3 and GossipCop++. As shown in Table 1, integrating the SDDE plugin into each base detection model significantly improves overall detection performance. Specifically, after integrating the SDDE plugin, the average ACC (Accuracy) of each model improves by approximately 2.55% on the F3 dataset and by approximately 2.22% on the GossipCop++ dataset. Real_F1 and Fake_F1 scores also show stable improvements. The main reasons for the performance improvement are: (1) The SDDE plugin achieves purification feature extraction through variational information bottleneck, weakening the interference of noise information on model training; (2) The SDDE plugin extracts text language fingerprints by generating source adversarial and semantic preservation assistance, weakening the interference of language fingerprints on the model learning process and preserving information related to true and false discrimination, making the model pay more attention to the deep semantics related to true and false discrimination; (3) The SDDE plugin uses cross-attention mechanism to perform semantic compensation on the purification features to obtain enhanced features, and then uses word-level dynamic gating mechanism to adaptively fuse and optimize the enhanced features and the original text representation, realizing the enhanced expression of deep semantics in the original text, thereby improving the model's ability to identify false information mixed between humans and machines; (4) During the training phase, the model jointly optimizes the total loss by weighting the existing detection method task loss, the final effective KL loss, the source adversarial loss, and the semantic preservation assistance loss, while constraining language fingerprint suppression, deep semantic preservation, and true and false discrimination tasks, so that each module can be optimized in synergy, thereby obtaining a more robust detection effect.

[0125] Table 1: Comparison of Detection Performance Before and After Integrating SDDE into Each Base Model

[0126]

[0127] Table 2 shows the ablation experiment data of the SDDE plugin. The experiments removed the generator source discriminator, auxiliary discriminator, variational information bottleneck module, and cross-attention semantic compensation module from two base detection models: BERT (Bidirectional Encoder Representations from Transformers) and ARG. The results were compared with the complete SDDE plugin. The results show that removing any module leads to a decrease in model performance, indicating that each module has a positive effect on the final detection effect. Specifically, on the BERT model, removing the variational information bottleneck module and the cross-attention semantic compensation module reduced the accuracy by 2.67%–0.61% and 2.29%–0.21%, respectively, indicating that feature denoising based on the variational information bottleneck and semantic compensation based on cross-attention can further improve the text representation quality. For the ARG model, removing the auxiliary discriminator resulted in a 2.21% and 0.76% decrease in accuracy on the F3 and GossipCop++ datasets, respectively. Removing the generator-source discriminator further reduced accuracy by 1.87% and 1.39%, respectively. This indicates that the language fingerprint decoupling process requires using source adversarial loss to suppress text generation source-related features and using semantic preservation auxiliary loss to retain deep semantics related to authenticity discrimination. In summary, the variational information bottleneck module, cross-attention semantic compensation module, generator-source discriminator, and auxiliary discriminator work together to achieve language fingerprint suppression and deep semantic enhancement, thereby improving the performance of human-machine hybrid false information detection.

[0128] Table 2: Ablation Experiment Data of SDDE Plugin

[0129]

[0130] Table 3 shows the performance differences in human / machine text authenticity discrimination, illustrating the performance differences of different base detection models before and after integrating the SDDE plugin on human-written and machine-generated text. Real_F1 and HF_F1 represent the F1 scores for real and fake human text, respectively; MR_F1 and MF_F1 represent the F1 scores for real and fake machine-generated text, respectively; and AvgΔ measures the performance gap between human-written and machine-generated text, with a smaller value indicating a more balanced detection ability across different text sources. Experimental results show that the original base detection models generally achieve higher detection performance on machine-generated text, while their detection performance on human-written text is relatively weak, exhibiting a significant source bias. This problem is somewhat alleviated after integrating the SDDE plugin. For example, on the F3 dataset, the AvgΔ of the BERT model decreased from 8.85% to 5.60%, EANN from 8.33% to 5.90%, and MDFEND from 5.35% to 5.15%. On the GossipCop++ dataset, ARG's AvgΔ decreased from 6.92% to 6.20%, and LFND-AB's decreased from 6.79% to 5.72%. Overall, most pedestal detection models showed a significant decrease in AvgΔ after integrating the SDDE plugin, indicating that the semantic feature decoupling and dynamic enhancement designed in this invention can effectively reduce the model's dependence on language fingerprints related to the generation source, allowing the model to focus more on deep semantics related to authenticity discrimination. This narrows the detection performance gap between human-written text and machine-generated text, and improves the model's balanced discrimination ability in human-machine mixed false information detection scenarios.

[0131] Table 3: Performance Differences in Human / Machine Text Authenticity Detection

[0132]

[0133] Figure 2 This is a schematic diagram illustrating the sensitivity analysis of the joint loss weight parameters of the RoBERTa model with access to the SDDE plugin, provided in an embodiment of the present invention. To address the variational information bottleneck, the weight hyperparameters corresponding to the original KL loss are truncated. To generate the weight hyperparameters corresponding to the source adversarial loss, The weight hyperparameters corresponding to the semantic preservation auxiliary loss are defined. SheepDog represents the baseline text authenticity detection network, and Dmint represents the deep multi-scale interaction detection model.

[0134] Experiments were conducted on the F3 and GossipCop++ datasets to adjust the values ​​of the three class weight hyperparameters, using Macro-F1 as the evaluation metric. Experimental results show that the RoBERTa model with the SDDE plugin maintains stable detection performance across a wide range of parameter values, demonstrating that this method does not overly rely on parameter tuning. The model performs best when the value is around 0.01; as... As the numerical value increases, the overall detection performance gradually decreases, indicating that excessively high constraint strength corresponding to the final effective KL loss will cause excessive feature compression and loss of key semantics used for true / false discrimination. When the value is in the range of 0.1 to 0.3, better results can be obtained, indicating that an appropriate source adversarial loss constraint can effectively suppress the interference caused by text language fingerprints. However, if the adversarial weight is too high, it will interfere with the learning process of the main task of distinguishing between true and false. The model performs best when the value is around 0.3 to 0.5, indicating that the semantic preservation auxiliary loss can retain the core semantics for distinguishing between true and false information. However, if the value of this weighted hyperparameter is too large, it will weaken the model training effect. In summary, the parameter sensitivity experiment verifies that the model jointly optimizes the total loss of the embodiment of the present invention, which is composed of the task loss of the existing detection method, the final effective KL loss, the source adversarial loss, and the semantic preservation auxiliary loss. This can achieve an effective balance between suppressing the source language fingerprint of the text generation, retaining the key semantics for discrimination, and accurately identifying false information, so that the detection of false information in human-machine hybrid text can achieve stable and excellent recognition results.

[0135] Based on the same inventive concept, this invention provides a plug-and-play human-machine hybrid false information detection device. Since the principle by which this device solves the technical problem is similar to that of a plug-and-play human-machine hybrid false information detection method, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be described again.

[0136] like Figure 3 As shown, the device includes an acquisition unit 301, a first obtaining unit 302, a second obtaining unit 303, and a result unit 304.

[0137] The acquisition unit 301 is used to acquire the human-machine hybrid text to be detected in the scenario of online content review on social media and information platforms, and reuse the encoder built into the existing false information detection process to encode the text to be detected to obtain the original text representation.

[0138] The first obtaining unit 302 is used to perform feature aggregation and linear mapping on the original text representation to obtain global semantic features. The global semantic features are linearly mapped to obtain latent variable Gaussian distribution parameters. The original KL loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterized sampling strategy.

[0139] The second obtaining unit 303 is used to perform source adversarial processing and semantic preservation auxiliary processing in parallel based on the purification features, and to separate the interfering language fingerprint and the deep semantic for authenticity discrimination.

[0140] The third unit 304 is used to perform semantic compensation based on the purified features using a cross-attention mechanism to obtain enhanced features; and to perform adaptive fusion optimization of the enhanced features and the original text representation through a word-level dynamic gating mechanism to obtain the final semantic representation.

[0141] Result unit 305 is used to replace the original text representation with the final semantic representation in the existing misinformation detection process to obtain the misinformation detection result.

[0142] It should be understood that the units included in the above-described plug-and-play human-machine hybrid false information detection device are merely a logical division based on the functions implemented by the device. In practical applications, the above units can be superimposed or separated. Furthermore, the functions implemented by the plug-and-play human-machine hybrid false information detection device provided in this embodiment correspond one-to-one with the plug-and-play human-machine hybrid false information detection method provided in the above-described embodiment. The more detailed processing flow implemented by this device has been described in detail in the above-described method embodiment one, and will not be elaborated here.

[0143] Another embodiment of the present invention provides a computer device, the computer device including: a processor and a scene database; the scene database is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the electronic device executes the various steps of the plug-and-play human-machine hybrid false information detection method shown in the above method embodiment.

[0144] Another embodiment of the present invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the steps of the plug-and-play human-machine hybrid false information detection method shown in the above method embodiment.

[0145] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A plug-and-play method for detecting false information involving human and machine hybrid systems, characterized in that, This method is executed between the encoding step and the authenticity classification step in the existing misinformation detection process, and includes: The text to be detected in the context of human-machine hybrid operation in the online content review scenario of social media and information platforms is obtained, and the text to be detected is encoded based on the encoder built into the existing false information detection process to obtain the original text representation. The original text representation is subjected to feature aggregation and linear mapping to obtain global semantic features. The global semantic features are mapped to obtain latent variable Gaussian distribution parameters. The original KL loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterization sampling strategy. Based on the purification features, parallel processing of source adversarial processing and semantic preservation auxiliary processing are performed to separate interfering language fingerprints and deep semantics for authenticity determination. Based on the purification features, semantic compensation is performed using a cross-attention mechanism to obtain semantic compensation features and enhanced features; the enhanced features and the original text representation are adaptively fused and optimized using a word-level dynamic gating mechanism to obtain the final semantic representation. The original text representation is replaced with the final semantic representation and input into the existing misinformation detection process to obtain the misinformation detection result.

2. The method as described in claim 1, characterized in that, The process of feature aggregation and linear mapping on the original text representation yields global semantic features, which are then mapped to obtain Gaussian distribution parameters for latent variables. Specifically, this includes: Global pooling and linear transformation are performed on the original text representation to obtain global semantic features; The global semantic features are predicted through two independent mapping branches to obtain the latent variable mean vector and the latent variable log-variance vector. The global semantic features, the latent variable mean vector, and the latent variable log-variance vector are represented by the following formulas: in, Represents global semantic features. Represents the original text representation. This indicates a global pooling operation. This indicates a modified linear unit activation function. Presentation layer normalization operation, Represents the vector of latent variable means. Represents the log-variance vector of latent variables. This represents the weights of the linear layer output by pooling within the variational information bottleneck module. This indicates the bias of the linear layer output by pooling within the variational information bottleneck module. This represents the linear projection weight of the mean branch within the variational information bottleneck module. This indicates the bias of the linear layer in the mean branch within the variational information bottleneck module. This represents the linear projection weights of the log-variance branch within the variational information bottleneck module. Indicates the log-variance branch linear layer bias within the variational information bottleneck module, and identifies the prefix. This indicates that the parameter belongs to the variational information bottleneck module.

3. The method as described in claim 1, characterized in that, The process of obtaining the original KL loss based on the Gaussian distribution parameters of the latent variables, obtaining the final effective KL loss based on a preset threshold, and obtaining the cleaned-up features based on a reparameterized sampling strategy specifically includes: The Gaussian distribution parameters of the latent variables include the latent variable mean vector and the latent variable log-variance vector; The original KL loss is determined based on the latent variable mean vector and the latent variable logarithmic variance vector; the original KL loss is then compared with a preset threshold, and the larger value between the difference and 0 is selected as the final effective KL loss. Based on the reparameterized sampling strategy, the purification features are obtained according to the latent variable mean vector and the log-variance vector. The original KL loss, the final effective KL loss, and the purification feature are expressed by the following formulas: in, This represents the original KL loss. This represents the final effective KL loss. The dimension index representing the purification feature. , This represents the total dimensionality of a single sample's purified feature vector. Indicates the first The mean parameter of the latent variable. Indicates the first The variance parameter of the latent variable. This represents the KL divergence cutoff threshold. Indicates purification characteristics, Indicates the standard Gaussian distribution A noise vector randomly sampled from the middle. This represents the element-wise Hadamard product operation on vectors. This represents the exponential function, used to recover variance from logarithmic variance. Represents the vector of latent variable means. Represents the log-variance vector of latent variables. This represents the function that takes the maximum value. This represents the summation operation. Indicates the first The squared term of the mean parameter of the latent variable.

4. The method as described in claim 1, characterized in that, The parallel execution of source adversarial processing based on the purification features specifically includes: A source discriminant is constructed based on a gradient inversion layer and a multilayer perceptron. The sanitized features are input into the source discriminant to obtain the text generation source prediction probability. The source adversarial loss is determined based on the text generation source prediction probability and the true source label. in, Indicates purification characteristics, Indicates the first The predicted probability of text generation source output by the source discriminator for each sample. This represents the sample index within a single training batch. This indicates a gradient reversal layer, used for adversarial training to reverse the gradient propagation direction. This represents the hyperparameter indicating the gradient reversal strength. This represents a multilayer perceptron. This indicates that the source of the generation is against the loss. This indicates the total number of samples within a single training batch. Indicates the first Each sample has a binary label representing its actual source: 0 for human-written text and 1 for machine-generated text. This represents the logarithmic operation with the natural constant e as the base. This indicates a summation operation.

5. The method as described in claim 1, characterized in that, The parallel execution of semantic preservation auxiliary processing based on the purified features specifically includes: An auxiliary discriminator is constructed based on a multilayer perceptron. The purified features are input into the auxiliary discriminator to obtain the false information auxiliary prediction probability. Based on the false information auxiliary prediction probability and the true / false labels of the samples, a semantic preservation auxiliary loss is obtained. The false information-assisted prediction probability and the semantic preservation-assisted loss are determined by the following formulas: in, The output of the semantically preserving auxiliary discriminator represents the first... Inaccurate information from individual samples is used to assist in predicting probabilities. This represents semantic preservation auxiliary loss. This represents the total number of samples in a single training batch. Indicates the sample index within the batch; Indicates the first Each sample's genuine / fake label Indicates purification characteristics, This represents a multilayer perceptron. This represents the logarithmic operation with the natural constant e as the base. This indicates a summation operation.

6. The method as described in claim 1, characterized in that, The enhanced features are obtained by performing semantic compensation based on the purified features using a cross-attention mechanism, specifically including: Using the purification features as the query matrix and the original text representation as the key matrix and value matrix, a cross-attention module is constructed to obtain semantic compensation features; Weights for semantic compensation information injection are adaptively determined based on cleanup features: The enhanced feature is obtained by fusing the purification feature, the semantic compensation feature, and the semantic compensation information injection weight: in, This represents the query matrix of the cross-attention module. Represents the cross-attention module value matrix. Represents the cross-attention module key matrix. Represents the key vector The transpose of the matrix, The feature dimensions representing the key and query vector. This represents the square root operation. This represents the text padding position mask matrix. This represents semantic compensation features. Represents the normalized activation function. This represents the Sigmoid activation function. This indicates a modified linear unit activation function. Indicates purification characteristics, This represents the upper-level linear matrix of the cross-compensation weight network within the semantic dynamic enhancement module. This represents the underlying linear matrix of the cross-compensation weight network within the semantic dynamic enhancement module. This indicates the upper-level bias of the cross-compensation weight network within the semantic dynamic enhancement module. This indicates the underlying bias of the cross-compensation weight network within the semantic dynamic enhancement module. This indicates semantic compensation adaptive injection weights. Indicates enhanced features, This represents the element-wise Hadamard product operation on vectors. This represents the linear weight matrix of the enhanced feature fusion layer within the semantic dynamic enhancement module, identifying the prefix. This indicates that the parameter belongs to the semantic dynamic enhancement module.

7. The method as described in claim 1, characterized in that, The step of adaptively fusing and optimizing the enhanced features and the original text representation through a word-level dynamic gating mechanism to obtain the final semantic representation specifically includes: Based on the original text representation, the word-level dynamic gating weights are obtained: The word-level dynamic gating weights are fused element-by-element with the enhanced features, the original text representation is superimposed, and layer normalization is performed to obtain the final semantic representation: in, Represents word-level dynamic gating weights. This represents the linear weight matrix of the word-level dynamic gating network within the semantic dynamic enhancement module. This represents the bias vector of the linear layer of the word-level dynamic gating network within the semantic dynamic enhancement module. Indicates enhanced features, This represents the final semantic representation. Represents the original text representation. This represents the element-wise Hadamard product of vectors. Presentation layer normalization operation, Indicates the Sigmoid activation function, identified by the prefix. This indicates that the parameter belongs to the semantic dynamic enhancement module.

8. The method as described in claim 1, characterized in that, After obtaining the false information detection results, the process also includes: Based on the final effective KL loss, semantic preservation auxiliary loss, source adversarial loss, and the task loss of existing misinformation detection methods, the total loss is determined as follows: in, This represents the total loss from the joint optimization of the model. This indicates the task loss of existing misinformation detection methods. The weight hyperparameters representing the variational information bottleneck truncated KL loss are... This represents the weight hyperparameters of the source adversarial loss. The weight hyperparameters represent the semantic preservation auxiliary loss. This represents the final effective KL loss. This indicates that the source of the generation is against the loss. This represents semantic preservation auxiliary loss.

9. A plug-and-play human-machine hybrid false information detection device based on any one of claims 1 to 8, characterized in that, include: The acquisition unit is used to acquire the text to be detected in the context of human-machine hybrid network content review in social media and information platform scenarios, and reuse the encoder built into the existing false information detection process to encode the text to be detected to obtain the original text representation. The first obtaining unit is used to perform feature aggregation and linear mapping on the original text representation to obtain global semantic features. The global semantic features are linearly mapped to obtain latent variable Gaussian distribution parameters. The original KL loss is obtained based on the latent variable Gaussian distribution parameters, and the final effective KL loss is obtained based on a preset threshold. The cleaned features are obtained based on the reparameterized sampling strategy. The second obtaining unit is used to perform source adversarial processing and semantic preservation auxiliary processing in parallel based on the purification features, and to separate the interfering language fingerprint and the deep semantic for true and false discrimination. The third obtaining unit is used to perform semantic compensation based on the purified features using a cross-attention mechanism to obtain enhanced features; The enhanced features and the original text representation are adaptively fused and optimized using a word-level dynamic gating mechanism to obtain the final semantic representation; The result unit is used to replace the original text representation with the final semantic representation in the existing misinformation detection process to obtain the misinformation detection result.

10. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform a plug-and-play human-machine hybrid false information detection method as described in any one of claims 1 to 8.

Images