Multi-field false information detection method and device based on multi-view triple-alignment multi-expert architecture

Abstract

The invention belongs to the technical field of nuclear power, and particularly relates to a multi-field false information detection method and device based on a multi-view triple-alignment multi-expert architecture. According to the multi-field false information detection method based on the multi-view triple-alignment multi-expert architecture, provided by the invention, a new multi-field FEND framework, namely multi-view Triple Alignment Multi-Experts (MTM), is constructed, so that the reliability and the safety of a social media platform are improved. According to the method, multi-view extraction and fusion are adopted, by combining comments, besides texts, auxiliary information such as emotion and writing styles is combined, and richer information view angles are provided for news posts. Furthermore, the MTM applies three types of alignment, better characterizes the posts by aligning the multi-view representations of the posts, focuses the model on a task inference representation by spatially aligning all examples, and allows experts to collectively make more discriminative results in specific areas.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power technology and specifically relates to a method based on multi-view Figure 3 Multi-domain false information detection method and device based on realigned multi-expert architecture. Background Art

[0002] In recent years, with the development of the Internet and the popularity of smart devices, social media websites have rapidly gained popularity and become an important platform for people to obtain social information. Inevitably, these websites have also led to the rapid spread of unreliable data (such as fake news). With the exponential growth of social software and users, this problem has become increasingly serious.

[0003] In the related art, attempts are made to discover and filter false information through the Fake News Detection (FEND) automated model. In the real world, there are thousands of news posts published on different platforms in various fields, such as health, politics, and society, but early FEND models generally do not consider domain information. News posts from different fields usually have different writing methods, emotional expressions, and target users. It is difficult to achieve high efficiency by building a simple model or testing on a dataset different from the training dataset. At the same time, due to the natural overlap and intersection between topics or related objects, the topics of news reports often cover multiple fields. It is difficult to accurately use a single label to establish an efficient FEND framework for multiple fields.

[0004] In addition, there is no precise fusion between domain differences and invariant features in related technologies, so early methods tend to degrade model performance as the number of domains increases. Among these methods, some methods tend to focus only on domain differences, while others only consider invariant features, which often leads to poor model performance. In the process of model training and reasoning, domain differences often lead to domain bias. This bias will intensify as the number of domains increases. At the same time, the false correlation brought by domain bias will cause the model to overfit, and its excessive reliance on domain differences will conflict with human inductive reasoning ability. For example, just because you know that the subject area of ​​a piece of news is military, it cannot be considered correct. But in some ways, domain differences are also mixed with task reasoning information in a specific field, so relying only on invariant features will also lead to valuable information loss. It is usually difficult for FEND systems to accurately obtain the fusion between domain differences and invariant features.

[0005] Existing methods do not consider solving the imbalance problem under multi-domain relations with cross-domain noise. This domain imbalance with cross-domain noise significantly affects the effectiveness of model training. For example, in the dataset of multi-domain fake news detection, the frequency of data in different domains follows a long-tail distribution. The domains with higher frequencies in the long-tail distribution are usually called head domains, and conversely, the domains with lower frequencies are called tail domains, where the frequency difference between the head domain and the tail domain is even greater. In different domains, the frequency of occurrence of positive and negative examples also roughly follows a long-tail distribution. Since the models are more exposed to the head domain during training, they tend to learn the unique features of the head domain in the multi-domain case. This phenomenon hinders model learning in the tail domain, which can also be regarded as noise caused by cross-domains.

[0006] Therefore, how to improve the accuracy and efficiency of false information detection has become an urgent problem to be solved. Summary of the invention

[0007] In order to overcome the problems existing in the related art, a multi-view Figure 3 Multi-domain false information detection method and device based on realigned multi-expert architecture.

[0008] According to one aspect of an embodiment of the present disclosure, a multi-view Figure 3 A multi-domain false information detection method for realigning a multi-expert architecture, the method comprising:

[0009] Step 1: For the collected news data in different fields, each news post is split into post text content and comment information, and style information and sentiment information are extracted from the post text content. The text content of each news post, the style information and sentiment information in the text content, and the comment information are used as model input data;

[0010] Step 2: Establish an end-to-end model MTM, use batch data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and use intra-sample view alignment, inter-sample feature alignment and domain alignment. Through supervised learning, use labels to optimize parameters and obtain multi-expert feature fusion features;

[0011] Step 3: After obtaining the multi-expert feature fusion features, the target domain expert model is used to obtain the discrimination result using the softmax classifier.

[0012] In a possible implementation, step 2 includes step 21, performing multi-view representation extraction;

[0013] Using style encoder E sty For style embedding {s1, s2, ..., s M} is encoded into the original style representation vector v sty; Using emotion encoder E emo For sentiment embedding {e1, e2, ..., e K} is encoded into the original sentiment representation vector v emo , as shown in Equation 1 and Equation 2;

[0014] v sty =E sty ({s1, s2, ..., s M}), Formula 1

[0015] v emo =E emo ({e1, e2, ..., e K}), Formula 2

[0016] Using content encoder E cnt Embed {cn1, cn2, ..., cn T} is encoded into the original content feature vector v cnt , using the comment encoder E cmm Embedding {cm1, cm2, ..., cm T} is encoded into the original comment feature vector v cmm .

[0017] v cnt =E cnt ({cn1, cn2, ..., cn T}), Formula 3

[0018] v cmm =E cmm ({cm1, cm2, ..., cm T}), Formula 4

[0019] In a possible implementation, step 2 includes step 22, performing multi-view information fusion;

[0020] In the additional information extraction part, multi-channel E sty 、E emo Get diverse information of style and sentiment information respectively and Using projection to represent a unified view of sentiment and writing style extra Projected into an additional feature subspace; the multi-channel representations of the additional views are concatenated and merged into an auxiliary space, as shown in Equation 10:

[0021]

[0022] Where I represents E sty The number of channels, J represents E emoThe number of channels, is the concatenation operator;

[0023] Get a representation of the news post v extra and v cnt 、v cmm After that, the independent features are mapped to the mixed space to represent each sample. The representation Z with various separation features can obtain the sample-level representation information through formula 11:

[0024]

[0025] Among them, H extra , H cnt , H cmm , H emo It is the number of channels that obtain different view information in each feature. Then, through the preliminary fusion projection, the fusion representation ability is further enhanced through the feature-level ST.

[0026] In a possible implementation, step 2 includes step 23, performing multi-domain perception expert learning, and the expert group is represented by ε i ={ε1,ε2,…,ε G}, the i-th expert ε i Corresponding to the i-th domain D i ; ε i The probability prediction of is shown in formula 13;

[0027]

[0028] where ε i (·)=FC(SA(·)) and R ε (·) is the domain-aware router algorithm for the expert ε i Select the corresponding domain in the randomly sampled mini-batch;

[0029] Use cross entropy loss to learn the discriminative ability of labels, as shown in Equation 14;

[0030]

[0031] Since each expert cannot see other experts, even if the encoder performs shared parameter operation, one expert can only learn from the fused features within its domain.

[0032] In a possible implementation, step 2 includes step 24, a triple alignment strategy, using triple alignment as intra-sample view alignment, inter-sample feature alignment, and domain alignment respectively;

[0033] In-sample perspective alignment for V cnt and V cmmThe view representation extraction focuses on the core views and contents of the posts through the following objective function:

[0034]

[0035] Where N is the batch size.

[0036] Feature alignment between samples In multi-view fusion, an alignment method is used to constrain the model, so that the representation is transformed in the direction of the representation of the real task and the cross-domain noise is filtered out; by spatially aligning all representations of the fused features to focus on the accuracy discrimination task, this can pull samples of different labels together spatially in each domain, as shown below:

[0037]

[0038] Where p is the norm used to set the feature distance, is the learnable cluster center corresponding to sample i with label y.

[0039] Application Loss DE To provide domain-based judgment, the expert network is allowed to make predictions based on its corresponding domain embedding as shown in Equation 17:

[0040]

[0041] in It is the fusion feature of real posts and fake posts; is the true sample and false sample y corresponding to the label y tr ,y fa The learnable cluster centers of p (x, y) is ||xy|| p 2 Norm.

[0042] In general, the objective function can be expressed as shown in Equation 18 and Equation 19:

[0043]

[0044] loss triple =λ1Loss IV +λ2Loss IF +λ3Loss DE , Formula 19

[0045] where {λ1, λ2, λ3} are learnable parameters.

[0046] In one possible implementation, a multi-layer perceptron is used as a style encoder and a sentiment encoder to model the sentiment dimension.

[0047] In a possible implementation, a word-level self-attention ST module is used to set E cnt and E cmm ;

[0048] In the word-unit dimension of the ST module, the attention is focused on the text and the comment respectively, and the ST is stacked so that the ST module consists of multi-head attention and two feed-forward neural networks with two residual connections and two normalization layers;

[0049] In E cnt 、E cmm In the above example, the word-level input matrix I of the ST module is used as the query Q, key K and value V; each head attention h of the multi-head attention i As shown in formula 5;

[0050]

[0051] where d k is the dimension of each head in the multi-head attention; the output matrix H of the multi-head attention block is concatenated with all the head attentions as shown in Equation 6;

[0052]

[0053] Where M is the number of heads in the multi-head attention, and Π is set as the continuous connection operator on multiple variables; after the multi-head attention structure, H is put into the subsequent network, as shown in Equation 7 and Equation 8;

[0054] X=LayerNorm(I+H), Equation 7

[0055] X=LayerNorm(X+FC(FC(X))) Formula 8

[0056] The above structure from input I to output X is uniformly set to SA(·); therefore, the overall stacked structure is represented by three rounds of SA(·) modules and one norm layer, as shown in Equation 9;

[0057] O=LayerNorm(SA(SA(SA(I)))). Formula 9

[0058] According to another aspect of the present disclosure, a multi-view Figure 3 A multi-domain false information detection device for realigning a multi-expert architecture, the device comprising:

[0059] The data processing module is used to split each news post into post text content and comment information for the collected news data in different fields, and extract style information and sentiment information from the post text content, and use the text content of each news post, the style information and sentiment information in the text content, and the comment information as model input data;

[0060] The feature processing module is used to build an end-to-end model MTM, which uses batched data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and uses intra-sample view alignment, inter-sample feature alignment, and domain alignment. Through supervised learning, the label is used for parameter optimization to obtain multi-expert feature fusion features;

[0061] The prediction module is used to obtain the discrimination result by using the target domain expert model and the softmax classifier after obtaining the multi-expert feature fusion features.

[0062] According to another aspect of the present disclosure, a multi-view Figure 3 A multi-domain false information detection device for realigning a multi-expert architecture, the device comprising:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] Wherein, the processor is configured to execute the above method.

[0066] According to another aspect of an embodiment of the present disclosure, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor.

[0067] The beneficial effects of the present disclosure are: Figure 3 The multi-domain false information detection method based on the realigned multi-expert architecture constructs a new multi-domain FEND framework, namely the multi-view Figure 3Alignment Multi-expert (Multi-viewTriple Alignment Multi-expert, MTM), is used to improve the reliability and security of social media platforms. Multi-view extraction and fusion are adopted, which provides a richer information perspective for news posts by combining comments, in addition to text, and incorporating auxiliary information such as sentiment and writing style. In addition, MTM applies three types of alignment to better characterize them by aligning the multi-view representations of individual posts, focusing the model on task reasoning representations by spatially aligning all examples, and allowing experts to jointly make more discriminative results in specific fields. At the same time, the framework utilizes a multi-expert system, which mitigates cross-domain interference between long-tail distributions in different domains by separating domains after learning shared core feature representations in the backbone network. The combination of these techniques better addresses various long-tail domain interferences, eliminates cross-domain noise, and integrates domain differences and invariant features. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a multi-view based Figure 3 Flowchart of the multi-domain false information detection method with realigned multi-expert architecture.

[0069] Figure 2 This is a multi-view based Figure 3 Flowchart of the multi-domain false information detection method with realigned multi-expert architecture.

[0070] Figure 3 This is a multi-view based Figure 3 Block diagram of the multi-domain false information detection device with realigned multi-expert architecture. DETAILED DESCRIPTION

[0071] The present disclosure is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0072] Unless otherwise defined, the technical and scientific terms used in the present disclosure have the same meanings as those generally understood by those skilled in the art to which the present disclosure belongs; the terms used in the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure; the term "including" and any variations thereof in the present disclosure are intended to cover non-exclusive inclusions. Obviously, the embodiments described in the present disclosure are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative efforts are within the scope of protection of the present disclosure.

[0073] Reference to "embodiments" in this disclosure means that a particular feature, structure, or characteristic described in conjunction with the embodiments can be included in at least one embodiment of the disclosure. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0074] Figure 1 This is a multi-view based Figure 3 Flow chart of a multi-domain false information detection method for realigning a multi-expert architecture. The method can be executed by a terminal device, wherein the terminal device can be a server, a desktop computer, a laptop computer, etc. The embodiment of the present disclosure does not limit the type of the terminal device. Figure 1 As shown, the method includes:

[0075] Step 1: For the news data collected from different fields, split each news post into post text content and comment information, and extract style information and sentiment information from the post text content. The text content of each news post, the style information and sentiment information in the text content, and the comment information are used as model input data.

[0076] For example, using {(X i ,y i ,D i )} as a multi-domain dataset. i For news posts, X i It consists of comments and content, where each comment and content contains T tokens. i Extract style features and sentiment features from the dataset and pre-train the comments and content tags. i There are M style embeddings {s1,s2,…,s M}, K sentiment embeddings {e1,e2,…,e K}, T comments are embedded in {cm1,cm2,…,cm T} and content embedding {cn1,cn2,…,cn T}. For the true label y∈{0,1}, 0 indicates that the news post is false and 1 indicates that the news post is true. The value range of domain d is set to D={D1,D2,…,D G}, where G is the number of domains in these datasets. N is the batch size in the training set.

[0077] Step 2: Establish an end-to-end model MTM, use batch data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and use intra-sample view alignment, inter-sample feature alignment and domain alignment. Through supervised learning, use labels to optimize parameters and obtain multi-expert feature fusion features.

[0078] As an example of this embodiment, Figure 2 As shown, an end-to-end model MTM is established, and MTM performs view representation extraction, multi-view information fusion, domain-aware expert learning and triple alignment on the input multi-domain dataset. Step 2 also includes steps 21 to 23.

[0079] Step 21, extract multi-view representation.

[0080] Multilayer Perception (MLP) is used as the style encoder and emotion encoder to model the emotion dimension. sty For style embedding {s1, s2, ..., s M} is encoded into the original style representation vector v sty . Using the emotion encoder E emo For sentiment embedding {e1, e2, ..., e K} is encoded into the original sentiment representation vector v emo , as shown in Equation 1 and Equation 2.

[0081] v sty =E sty ({s1, s2, ..., s M}), Formula 1

[0082] v emo =E emo ({e1, e2, ..., e K}), Formula 2

[0083] Content and comments can reflect the core ideas of news posts, and require deeper and more comprehensive processing to encode content and comments. cnt Embed {cn1, cn2, ..., cn T} is encoded into the original content feature vector v cnt , using the comment encoder E cmm Embedding {cm1, cm2, ..., cm T} is encoded into the original comment feature vector v cmm , as shown in Equation 3 and Equation 4.

[0084] v cnt =E cnt({cn1, cn2, ..., cn T}), Formula 3

[0085] v cmm =E cmm ({cm1, cm2, ..., cm T}), Formula 4

[0086] Where E cnt and E cmm It is set up with a word-level self-attention (ST) module. In order to better preserve the original feature information, the present disclosure focuses on the text and comments in the word-level dimension of the ST module. The ST is stacked so that the ST module consists of a multi-head attention and two feedforward neural networks with two residual connections and two normalization layers.

[0087] In E cnt 、E cmm In the above example, the word-level input matrix I of the ST module is used as the query Q, key K, and value V. Each head attention h of the multi-head attention i As shown in formula 5.

[0088]

[0089] where d k is the dimension of each head in the multi-head attention. The output matrix H of the multi-head attention block is concatenated with all the attention heads as shown in Equation 6.

[0090]

[0091] Where M is the number of heads in the multi-head attention, and Π is set as the continuous connection operator on multiple variables. After the multi-head attention structure, H is put into the subsequent network as shown in Equation 7 and Equation 8.

[0092] X=LayerNorm(I+H), Formula 7

[0093] X=LayerNorm(X+FC(FC(X))) Formula 8

[0094] The above structure from input I to output X is uniformly set to SA(·). Therefore, the overall stacked structure is represented by three rounds of SA(·) modules and a norm layer, as shown in Equation 9.

[0095] O=LayerNorm(SA(SA(SA(I)))). Formula 9

[0096] Step 22: perform multi-view information fusion.

[0097] In the additional information extraction part, multi-channel E sty 、E emo Get diverse information of style and sentiment information respectively and Using projection to represent a unified view of sentiment and writing style extra Projected into an additional feature subspace. The multi-channel representations of the additional views are concatenated and merged into an auxiliary space to more comprehensively represent sentiment and style, as shown below:

[0098]

[0099] Where I represents E sty The number of channels, J represents E emo The number of channels, is the concatenation operator.

[0100] Get a rich representation of the news post v extra and v cnt 、v cmm After that, the independent features are mapped to the mixed space to represent each sample, so that the representation Z with various separated features can obtain more detailed and sample-level representation information in the following way:

[0101]

[0102] Among them, H extra , H cnt , H cmm , H emo is the number of channels in each feature that obtain information from different views.

[0103] Finally, through the preliminary fusion projection, the fusion representation ability is further enhanced through the feature-level ST, as shown in Equation 12.

[0104]

[0105] Step 23: multi-domain perception expert learning.

[0106] Due to the cross-domain interference under the long-tail distribution, the present disclosure learns the features of the model in a larger domain, and the lack of exposure to a small number of domains in model learning will lead to poor performance. Multiple experts focus on fusion-based features. Learn specific domains to reduce the mutual interference between domains during reasoning. According to the characteristics of the domain, this module enables the system to achieve better results in its unique domain. Multi-domain-aware experts allow each domain expert to focus on its own domain characteristics after learning a shared core representation, which isolates the influence of irrelevant domains in the long-tail distribution. The expert group is denoted as ε i ={ε1,v2,…,εG}, the i-th expert ε i Corresponding to the i-th domain D i ε i The probability prediction of is shown in Equation 13.

[0107]

[0108] where ε i (·)=FC(SA(·)). ε (·) is the domain-aware router algorithm for the expert ε i Select the corresponding domains in the randomly sampled mini-batches.

[0109] The cross entropy loss is used to learn the discriminative ability of labels, as shown in Equation 14.

[0110]

[0111] Since each expert cannot see other experts, even if the encoder performs shared parameter operation, one expert can only learn from the fused features within its domain.

[0112] Step 24: Triple alignment strategy. In different parts of the model, triple alignment is used as intra-sample view alignment, inter-sample feature alignment, and domain alignment, respectively.

[0113] In-sample perspective alignment for v cnt and v cmm The view representation extraction is to bring different view representations of the same sample closer and push representations of different samples apart. This is because the representations of comments and text about the sample can complement each other in the representation process, highlighting the core features while weakening other noise features. The different views of each sample as a joint whole can be more evenly distributed in the feature space, thereby obtaining more accurate and comprehensive comment and content representation information.

[0114] However, for the same post, due to the different feature dimensions, the text and comments also have certain differences, which makes the expression of the post more sparse. Using contrastive learning and cross-similarity calculation methods, we match the features of comments and content views, shorten the spatial distance between these views, increase uniformity, and better focus on the core ideas and content of the post through the following objective function:

[0115]

[0116] Where N is the batch size.

[0117] Feature alignment between samples In multi-view fusion, an alignment method is used to constrain the model to transform the representation towards the representation of the real task and filter cross-domain noise. In this module, the model focuses on the accuracy discrimination task by spatially aligning all representations of the fused features, which can pull samples of different labels together spatially in each domain, as shown below:

[0118]

[0119] Where p is the norm used to set the feature distance, is the learnable cluster center corresponding to sample i with label y.

[0120] Domain alignment is applied to domain-aware expert learning. In order to prevent experts from having different understandings of the same task and to add regularization to unify experts and domains, Loss is applied. DE To provide domain-based judgment, allowing the expert network to make predictions based on its corresponding domain embedding. This confers domain specificity to mitigate cross-domain interference, which can be expressed as:

[0121]

[0122] in It is the fusion feature of real posts and fake posts; is the true sample and false sample y corresponding to the label y tr ,y fa The learnable cluster centers of p (x, y) is ||xy|| p 2 This is to adjust the ability of the expert module to judge the truth, rather than just adjusting the domain information in the feature. In this way, the expert set ε can complement each other in the field and better complete the reasoning task.

[0123] In general, the objective function can be expressed as shown in Equation 18 and Equation 19.

[0124]

[0125] loss triple =λ1Loss IV +λ2Loss IF +λ3loss DE , Formula 19

[0126] where {λ1, λ2, λ3} are learnable parameters.

[0127] Step 3, predict the score. After the multi-expert feature fusion features obtained by the model, the target domain experts use the softmax classifier to obtain the discrimination results.

[0128] The present invention provides a multi-view based Figure 3 The multi-domain false information detection method based on the realigned multi-expert architecture constructs a new multi-domain FEND framework, namely the multi-view Figure 3 Multi-view Triple Alignment Multi-expert (MTM) is used to improve the reliability and security of social media platforms. Multi-view extraction and fusion are adopted to provide a richer information perspective for news posts by combining comments, in addition to text, with auxiliary information such as sentiment and writing style. In addition, MTM applies three types of alignment to better characterize individual posts by aligning their multi-view representations, focusing the model on task reasoning representations by spatially aligning all examples, and allowing experts to jointly make more discriminative results in specific fields. At the same time, the framework utilizes a multi-expert system, which mitigates cross-domain interference between long-tail distributions in different domains by separating domains after learning shared core feature representations in the backbone network. The combination of these techniques better addresses various long-tail domain interferences, eliminates cross-domain noise, and integrates domain differences and invariant features.

[0129] According to another aspect of the present disclosure, a multi-view Figure 3 A multi-domain false information detection device for realigning a multi-expert architecture, the device comprising:

[0130] The data processing module is used to split each news post into post text content and comment information for the collected news data in different fields, and extract style information and sentiment information from the post text content, and use the text content of each news post, the style information and sentiment information in the text content, and the comment information as model input data;

[0131] The feature processing module is used to build an end-to-end model MTM, which uses batched data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and uses intra-sample view alignment, inter-sample feature alignment, and domain alignment. Through supervised learning, the label is used for parameter optimization to obtain multi-expert feature fusion features;

[0132] The prediction module is used to obtain the discrimination result by using the target domain expert model and the softmax classifier after obtaining the multi-expert feature fusion features.

[0133] The description of the above-mentioned device has been explained in detail in the description of the above-mentioned method, and will not be repeated here.

[0134] Figure 3 This is a multi-view based Figure 3Block diagram of a multi-domain false information detection device for realigning a multi-expert architecture For example, the device 1900 can be provided as a server. Figure 3 , the device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions, such as applications, that can be executed by the processing component 1922. The application stored in the memory 1932 can include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0135] The device 1900 can also include a power supply component 1926 configured to perform power management of the device 1900, a wired or wireless network interface 1950 configured to connect the device 1900 to a network, and an input / output (I / O) interface 1958. The device 1900 can operate based on an operating system stored in the memory 1932, such as Windows Server™, MacOS X™, Unix™, Linux™, FreeBSD™, or the like.

[0136] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the device 1900 to perform the above method.

[0137] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0138] A computer-readable storage medium can be a tangible device capable of holding and storing instructions used by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the above. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0139] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0140] The computer program instructions for performing the operation of the present disclosure can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions can be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA) or a programmable logic array (PLA), is personalized by using the state information of the computer-readable program instructions, and the electronic circuit can execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0141] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0142] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0143] Computer-readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0144] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0145] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A multi-domain false information detection method based on multi-view triple alignment multi-expert architecture, characterized in that: The method comprises: Step 1: For the collected news data in different fields, each news post is split into post text content and comment information, and style information and sentiment information are extracted from the post text content. The text content of each news post, the style information and sentiment information in the text content, and the comment information are used as model input data; Step 2: Establish an end-to-end model MTM, use batch data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and use intra-sample view alignment, inter-sample feature alignment and domain alignment. Through supervised learning, use labels to optimize parameters and obtain multi-expert feature fusion features; Step 3: After obtaining the multi-expert feature fusion features, the target domain expert model is used to obtain the discrimination result using the softmax classifier.

2. The method according to claim 1, characterized in that Step 2 includes step 21, performing multi-view representation extraction; Using style encoder E sty For style embedding {s1,s2,…,s M } is encoded into the original style representation vector v sty ; Using emotion encoder E emo For sentiment embedding {e1,e2,…,e K } is encoded into the original emotion representation vector v emo ; Using content encoder E cnt Embed {cn1, cn2, ..., cn T } is encoded into the original content feature vector V cnt , using the comment encoder E cmm Embed the comment {cm1, cm2, …, cm T } is encoded into the original comment feature vector V cmm .

3. The method according to claim 1, characterized in that Step 2 includes step 22, performing multi-view information fusion; In the additional information extraction part, multi-channel E sty 、E emo Get diverse information of style and sentiment information respectively and Using projection to represent a unified view of sentiment and writing style extra Projection into additional feature subspace; The multi-channel representations of the additional views are concatenated and merged into an auxiliary space as shown in Equation 10: Where I represents E sty The number of channels, J represents E emo The number of channels, is the concatenation operator; In getting a representation of a news post V extra and V cnt 、V cmm After that, the independent features are mapped to the mixed space to represent each sample. The representation Z with various separation features can obtain the sample-level representation information through formula 11: Among them, H extra , H cnt , H cmm , H emo It is the number of channels that obtain different view information in each feature. Then, through the preliminary fusion projection, the fusion representation ability is further enhanced through the feature-level ST.

4. The method according to claim 1, characterized in that: Step 2 includes step 23, performing multi-domain perception expert learning, for training data D = {D1, D2, ..., D G }, the expert group is denoted as ε i ={ε1,ε2,…,ε G }, the i-th expert ε i Corresponding to the i-th domain D i ; ε i The probability prediction of is shown in formula 13; where ε i (·)=FC(SA(·)) and R ε (·) is the domain-aware router algorithm for the expert ε i Select the corresponding domain in the randomly sampled mini-batch; Use cross entropy loss to learn the discriminative ability of labels, as shown in Equation 14; Since each expert cannot see other experts, even if the encoder performs shared parameter operation, one expert can only learn from the fused features within its domain.

5. The method according to claim 1, characterized in that Step 2 includes step 24, a triple alignment strategy, using triple alignment as intra-sample view alignment, inter-sample feature alignment, and domain alignment respectively; In-sample perspective alignment for V cnt and V cmm The view representation extraction focuses on the core views and contents of the posts through the following objective function: Where N is the batch size; Feature alignment between samples In multi-view fusion, an alignment method is used to constrain the model, so that the representation is transformed in the direction of the representation of the real task and the cross-domain noise is filtered out; by spatially aligning all representations of the fused features to focus on the accuracy discrimination task, this can pull samples of different labels together spatially in each domain, as shown below: Where p is the norm used to set the feature distance, is the learnable cluster center corresponding to sample i with label y; Application Loss DE To provide domain-based judgment, the expert network is allowed to make predictions based on its corresponding domain embedding as shown in Equation 17: in It is the fusion feature of real posts and fake posts; is the true sample and false sample y corresponding to the label y tr ,y fa Learnable cluster centers of f p (x, y) is ||xy|| o 2 norm; In general, the objective function can be expressed as shown in Equation 18 and Equation 19: loss triple = λ1Loss IV + λ2Loss IF + λ3loss DE , Equation 19 where {λ1, λ2, λ3} are learnable parameters.

6. The method according to claim 1, characterized in that Multilayer perceptron is used as style encoder and sentiment encoder to model the sentiment dimension.

7. The method according to claim 1, characterized in that The word-level self-attention ST module is used to set E cnt and E cmm ; In the word-unit dimension of the ST module, the attention is focused on the text and the comment respectively, and the ST is stacked so that the ST module consists of multi-head attention and two feed-forward neural networks with two residual connections and two normalization layers; In E cnt 、E xmm In the above example, the word-level input matrix I of the ST module is used as the query Q, key K and value V; each head attention h of the multi-head attention i As shown in formula 5; where d k is the dimension of each head in the multi-head attention; the output matrix H of the multi-head attention block is concatenated with all the head attentions as shown in Equation 6; Where M is the number of heads in the multi-head attention, and Π is set as the continuous connection operator on multiple variables; after the multi-head attention structure, H is put into the subsequent network, as shown in Equation 7 and Equation 8; X=LayerNorm(I+H), Equation 7 X=LayerNorm(X+FC(FC(X))) Formula 8 The above structure from input I to output X is uniformly set to SA(·); therefore, the overall stacked structure is represented by three rounds of SA(·) modules and one norm layer, as shown in Equation 9; O=LayerNorm(SA(SA(SA(I)))). Formula 9.

8. A multi-domain false information detection device based on a multi-view triple alignment multi-expert architecture, characterized in that: The device comprises: The data processing module is used to split each news post into post text content and comment information for the collected news data in different fields, and extract style information and sentiment information from the post text content, and use the text content of each news post, the style information and sentiment information in the text content, and the comment information as model input data; The feature processing module is used to build an end-to-end model MTM, which uses batched data to pass into the fusion multi-view encoder and multi-expert module for classification and discrimination, and uses intra-sample view alignment, inter-sample feature alignment, and domain alignment; through supervised learning, parameters are optimized using labels to obtain multi-expert feature fusion features; The prediction module is used to obtain the discrimination result by using the target domain expert model and the softmax classifier after obtaining the multi-expert feature fusion features.

9. A multi-domain false information detection device based on a multi-view triple alignment multi-expert architecture, characterized in that: The device comprises: processor; a memory for storing processor-executable instructions; The processor is configured to execute the method according to any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images