Cross-domain rapid false news detection method and system

Abstract

The invention relates to the technical field of news detection, in particular to a cross-domain rapid false news detection method and system, and the method comprises the steps: firstly, carrying out the embedding of an original news text through RoBERTa-wwm, and obtaining a corresponding word vector representation; performing feature extraction on a word vector matrix formed by word vectors by using a plurality of expert systems; performing sentiment analysis on the original news text to obtain a sentiment result, proposing an attention mechanism to perform attention distribution on word vectors, and outputting the sentiment result and the attention mechanism to assist an expert system in positioning a field; a feedforward neural network APN is used for extracting long-distance dependency information and assisting an expert system in positioning the field; and finally, performing multi-field positioning to obtain final classification features, and performing a classification task on the features by using a multi-layer perceptron to obtain a detection result that the news is true or false. According to the method, whether the news is false or not can be quickly and accurately judged in multi-field news.

Description

Technical Field

[0001] The present invention relates to the technical field of news detection, and in particular to a cross-domain rapid false news detection method and system. Background Art

[0002] The proliferation of fake news poses a serious threat to social trust and public health. With the popularity of the Internet and social media, the speed and scope of information dissemination have expanded unprecedentedly. Social media has made information sharing easier, but many news posts shared online are unreliable or of low quality. False news has also spread rapidly, causing widespread negative impacts. False news not only undermines the public's trust in the media, but may also cause social panic, mislead public opinion, and even affect major decisions. Therefore, how to effectively detect and curb fake news has become an important issue that needs to be addressed urgently. This type of false information takes advantage of the credulity of some people who lack scientific literacy or believe in biased information, thereby achieving the manipulator's intentions and causing casualties to innocent people. However, manual fact checking for this type of news is time-consuming and difficult. Therefore, the field of fake news detection research has emerged, combining multidisciplinary technologies such as natural language processing, machine learning, data mining, and psychology, aiming to develop automated tools and algorithms to identify and filter false information.

[0003] In order to solve this problem, researchers have proposed a variety of methods. Early research mainly focused on content-based detection methods: by analyzing the language characteristics, sentiment tendencies, word frequency statistics and other indicators of news texts to judge their authenticity. However, as the means of producing fake news become increasingly complex, content analysis alone can no longer cope with various disguises and misleading information. Therefore, researchers began to explore multi-dimensional detection methods, including social network analysis, user behavior analysis, and cross-platform data integration. But most of them focus on single-domain fake news detection (SFND), such as politics and health. However, for a certain field, the number of fake news may be extremely limited. Therefore, creating deep learning methods based on such insufficient single-domain data, the final performance is not satisfactory. In actual scenarios, real-world news platforms publish various news in different fields every day. Therefore, it is promising to solve the data sparsity problem and improve the performance of all domains by leveraging data from multiple fields. Summary of the invention

[0004] The present invention aims to solve the problem that the current expert system separates multiple fields separately, which is inefficient and each expert may give a larger weight to certain texts in a targeted manner, resulting in inaccurate field positioning and poor final decision results. A cross-domain rapid false news detection method and system are proposed to realize rapid and accurate judgment of whether news is false in multi-domain news.

[0005] To achieve the above purpose, the technical solution adopted is:

[0006] A cross-domain rapid fake news detection method, including:

[0007] First, use RoBERTa-wwm to embed the original news text to obtain the corresponding word vector representation;

[0008] Then, multiple expert systems are used to extract features from the word vector matrix composed of word vectors;

[0009] Perform sentiment analysis on the original news text to obtain sentiment results, and propose an attention mechanism to allocate attention to word vectors. The sentiment results and attention mechanism output assist the expert system in locating areas.

[0010] Using feed-forward neural network APN to extract long-range dependency information also assists expert system in locating areas;

[0011] Finally, after multiple field positioning, the final classification features are obtained. These features are then used to perform classification tasks using a multi-layer perceptron to obtain detection results on whether the news is true or false.

[0012] According to the cross-domain rapid fake news detection method of the present invention, further, the expert system adopts a TextCNN model containing multiple convolution kernels of different sizes, performs convolution and pooling operations on the word vector matrix to obtain the feature vector μ; and uses a residual connection after the original input to retain the original input information in the output.

[0013] According to the cross-domain fast fake news detection method of the present invention, further, an attention mechanism is introduced into the expert network to represent the weight of each expert system in the decision-making process in different news fields.

[0014] According to the cross-domain rapid false news detection method of the present invention, further, the original news text is subjected to sentiment feature extraction using a sentiment analysis model based on Chinese-Roberta-WWM-EXT-Base to obtain a sentiment feature vector V emo ; Propose an attention mechanism to allocate attention to each word vector and obtain the output β of the attention mechanism; For the fields of science and medical health, use domain BERT to embed the original news text to obtain word vectors For other fields, RoBERTa-wwm is still used; finally, after a gating, the weight coefficient g is obtained after softmax:

[0015]

[0016] Where W is a learnable parameter.

[0017] According to the cross-domain fast fake news detection method of the present invention, a feedforward neural network APN is further used to extract long-distance dependency information, and the expression is as follows:

[0018] y APN =FFN(W,{ω})

[0019] Where FFN represents a feedforward neural network, W is a learnable parameter, and w is the word vector output by RoBERTa-wwm; APN assigns different weights α to the expert system according to the long-distance text information feature connection.

[0020] According to the cross-domain rapid false news detection method of the present invention, further, after multiple domain positioning, the final classification feature representation is obtained as follows:

[0021]

[0022] Where e represents the total number of expert systems; the expression for performing classification tasks using a multilayer perceptron is as follows:

[0023]

[0024] Where MLP is a multi-layer perceptron. is the predicted label.

[0025] According to the cross-domain rapid fake news detection method of the present invention, further, the loss function used by the model is binary cross entropy loss, and the model is trained using a data set consisting of news from multiple fields.

[0026] According to the cross-domain rapid fake news detection method of the present invention, further, a dual-model neural network acceleration method based on gradient optimization is used to optimize the model training speed.

[0027] Furthermore, the present invention also proposes a cross-domain rapid false news detection system, which is used to implement the cross-domain rapid false news detection method as described above, comprising:

[0028] The text embedding module is used to embed the original news text using RoBERTa-wwm to obtain the corresponding word vector representation;

[0029] The domain feature extraction module is used to extract features from the word vector matrix composed of word vectors using multiple expert systems;

[0030] Sentiment analysis attention module, which is used to perform sentiment analysis on the original news text to obtain sentiment results, and proposes an attention mechanism to allocate attention to word vectors. The sentiment results and attention mechanism output assist the expert system in locating the field;

[0031] Auxiliary field positioning module, which is used to extract long-distance dependency information using a feed-forward neural network (APN) and also assist the expert system in locating the field;

[0032] The classification decision module is used to obtain the final classification features after multiple field positioning. These features are then classified using a multi-layer perceptron to determine whether the news is true or false.

[0033] The beneficial effects achieved by adopting the above technical solution are:

[0034] The present invention uses attention mechanism, sentiment analysis and domain BERT embedding to extract information from original news text and assist the expert system in more accurate domain positioning. For multi-domain news, the expert system in the corresponding domain has the largest weight, which can better understand the information in this domain and help improve the model's detection accuracy of news.

[0035] Because the expert system uses convolutional neural networks, which are mainly used for images, they focus on local information, while texts have longer and larger connections. Therefore, the use of feedforward neural networks (APNs) is also to make up for the shortcomings of convolutional neural networks. Capturing long-distance dependency information can assist the expert system in analysis.

[0036] In summary, the final classification features of mixed-domain news obtained by the present invention through attention mechanism, sentiment analysis, domain BERT and APN multi-domain positioning have strong feature characterization capabilities. The present invention uses multi-domain data training models and assists with other fields with more data to solve the problem that fake news data in a certain field is scarce and it is difficult to build a large-scale deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0038] Figure 1 is an architecture diagram of a cross-domain rapid false news detection method according to an embodiment of the present invention;

[0039] Figure 2 is a schematic diagram of the structure of the expert system of an embodiment of the present invention;

[0040] Figure 3 It is a model architecture diagram after speed optimization of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0042] like Figure 1 As shown, this embodiment discloses a cross-domain rapid false news detection method, comprising the following steps:

[0043] Step S101: Use RoBERTa-wwm to embed the original news text to obtain the corresponding word vector representation.

[0044] Step S102: Use multiple expert systems to perform feature extraction on the word vector matrix composed of word vectors.

[0045] Specifically, the TextCNN model uses multiple convolution kernels of different sizes to perform convolution and pooling operations on a two-dimensional matrix composed of word vectors, and then concatenates the feature vector μ for classification or prediction. This process of repeated convolution and pooling followed by concatenation can easily lead to the deviation of the original text sequence, thereby causing domain localization errors. In addition, too many layers may cause model degradation. To solve these problems, the residual concept of ResNet is introduced, and residual connections are applied after the original input to retain the original input information in the output, such as Figure 2 As shown. In addition, layer normalization is added after each layer of convolution and pooling. The residual formula is as follows:

[0046] y=σ(x i ,{W i})+W0x0

[0047] Where x0 represents the initial feature. If the dimensions are inconsistent, weighting cannot be applied directly. Therefore, a linear projection W0 is needed to adjust the dimension accordingly. W0x0 is the linear projection of the initial feature to ensure that the dimension of the residual weighting is consistent with the output feature dimension.

[0048] The final output of the expert system is as follows:

[0049] μ e =φ(x e ,{w e})

[0050] Among them, u e is the output of the expert system; Represents the application to the initial feature x e A set of convolution and pooling operations.

[0051] In addition, based on the expert system, it is hoped that the algorithm can adaptively select appropriate experts according to the news field. Therefore, the attention mechanism is introduced into the expert network:

[0052] β e =attention(ω,{W})

[0053] in, Represents the weight of each expert system in the decision-making process in the news field.

[0054] Step S103: sentiment analysis is performed on the original news text to obtain sentiment results, and an attention mechanism is proposed to allocate attention to word vectors. The sentiment results and the output of the attention mechanism assist the expert system in locating the field.

[0055] In the analysis of the data set, it was found that fake news often attracts readers' attention through exaggerated, inflammatory or extreme emotional language. Sentiment analysis can help identify these abnormal emotional tendencies and provide clues for detecting fake news. For example, fake news may use strong negative emotional words such as anger, fear or hatred, while real news usually adopts more neutral and objective language. In addition, fake news often manipulates readers' reactions through extreme emotions, arousing panic or anger to generate more comments and shares, thereby achieving rapid spread. Sentiment analysis can quantify and evaluate the emotional extremes of news content, helping to identify reports that may have exaggerated or manipulative tendencies. In addition, fake news often contains contradictory emotional expressions, such as inconsistent emotional tones in the same article. Sentiment analysis can effectively detect these contradictions and reveal inconsistencies within the article, thus providing a basis for identifying fake news.

[0056] Therefore, the sentiment analysis model based on Chinese-Roberta-WWM-EXT-Base is used to extract sentiment features from the original news text. The model is optimized using sentiment analysis datasets from eight Chinese domains, containing a total of 227,347 samples. The final sentiment analysis result is represented by V emo .

[0057] An attention mechanism is proposed to allocate attention to each word vector and obtain the output β of the attention mechanism.

[0058] Given the powerful performance of BERT, many researchers have developed optimized models for specific fields. In order to more accurately capture cross-domain semantic information and supplement the residual of the original semantics, targeted feature extraction is performed for specific fields. However, since no specific models for all fields have been found, new feature extraction is performed specifically for the science and medical health fields, while RoBERTa-wwm is still used for other fields; the newly obtained features are represented as

[0059] Finally, after a gating and softmax, the weight coefficient g is obtained, which is used to assist the expert system in locating the field.

[0060]

[0061] Where W is a learnable parameter.

[0062] Step S103: Use the feed-forward neural network APN to extract long-distance dependency information and also assist the expert system in locating the area.

[0063] Given that the expert system adopts the convolutional neural network (CNN) architecture, some inherent limitations of CNN in processing text need to be addressed. CNN usually uses local receptive fields and pooling operations to extract features, which is very effective in image processing because local features can be combined into higher-level global features. However, in text processing, important information may be distributed over a longer range in a sentence. Although CNNs can process local n-gram features, they often perform poorly in capturing long-range dependencies due to the translation invariance of convolution and pooling operations. This means that CNNs may not be able to effectively capture the precise relationships and dependency structures between words, because the order and position of words are crucial in text, especially in tasks such as sentiment analysis and semantic analysis that require precise sequence information.

[0064] To solve this problem, a feedforward neural network (APN) is used to supplement CNN and enhance the auxiliary positioning function. APN performs domain-specific adjustments by learning domain labels. This allows the network to dynamically adjust its parameters based on domain information, effectively reducing domain shift.

[0065] y APN =FFN(W,{ω})

[0066] Where FFN represents a feedforward neural network, W is a learnable parameter, and w is the word vector output by RoBERTa-wwm. In APN, the core idea is to perform classification tasks on word embeddings w. By using the SoftMax function, the domain weight α is obtained. Unlike expert systems, APN does not produce a single result, but assigns different weights α to expert systems based on long-range text information feature connections. This means that multiple fields can have significant probability values, which helps correct the results from the expert system. It should be noted that the output here is only the classification result, while the output of the expert system contains all the content information from the word embedding.

[0067] Step S104: After multiple domain positioning, the final classification features are obtained. These features are then classified using a multi-layer perceptron to obtain a detection result of whether the news is true or false.

[0068] By integrating the expert system of step S101 and the auxiliary positioning network of steps S102 and S103, the final features of the news can be expressed as:

[0069]

[0070] In the formula, e represents the total number of expert systems. The expression for performing classification tasks using a multilayer perceptron is as follows:

[0071]

[0072] Where MLP is a multi-layer perceptron. is the predicted label.

[0073] The loss function used is binary cross entropy loss, which is the most commonly used loss function for binary classification tasks.

[0074] The model is trained using a dataset consisting of news from multiple fields.

[0075]

[0076] Where N is the number of samples, y i is the true label of the i-th sample. When the news is false, y i =1; when the news is true, y i =0. is the probability that the i-th sample is predicted to be false.

[0077] Given the rapid spread of information on social media and news websites, fake news can spread rapidly in a short period of time. In order to effectively curb the spread of fake news, detection algorithms must have real-time processing capabilities. Fast algorithms can immediately identify and mark fake news when it is released, thereby reducing its spread and impact. For social media platforms and news aggregation websites, if the detection algorithm is slow, it may cause loading delays or untimely information updates, affecting users' trust and experience of the platform. Under the same hardware conditions, faster algorithms can process more data or perform more detection tasks, thereby improving the overall performance and economic efficiency of the system. This is particularly important for small and medium-sized platforms with limited resources, so speed becomes a key indicator in practical applications. Therefore, the above models can be optimized for speed. In terms of feature extraction, a streamlined version of the RoBERTa-WWM model called MiniBERT-H288 is used, and sentiment analysis and domain BERT are removed. A dual-model neural network acceleration method based on gradient optimization is also introduced to optimize the model training speed; in this method, the neural network is used for prediction, sensitivity information is calculated by backpropagation, and an additional reverse model is constructed to directly calculate sensitivity, thereby improving prediction accuracy. The speed-optimized model is referred to as Ours-fast. Figure 3 As shown; before simplification Figure 1 The model is referred to as Ours-full.

[0078] Corresponding to the above method, this embodiment also proposes a cross-domain rapid fake news detection system, including:

[0079] The text embedding module is used to embed the original news text using RoBERTa-wwm to obtain the corresponding word vector representation.

[0080] The domain feature extraction module is used to extract features from the word vector matrix composed of word vectors using multiple expert systems.

[0081] The sentiment analysis attention module is used to perform sentiment analysis on the original news text to obtain sentiment results, and proposes an attention mechanism to allocate attention to word vectors. The sentiment results and attention mechanism output assist the expert system in locating the field.

[0082] The auxiliary field positioning module is used to extract long-distance dependency information using a feed-forward neural network APN and also assist the expert system in locating the field.

[0083] The classification decision module is used to obtain the final classification features after multiple field positioning. These features are then classified using a multi-layer perceptron to determine whether the news is true or false.

[0084] In order to verify the effectiveness of the scheme in this case, further explanation is given below in combination with experimental data. The experimental results are shown in Table 1.

[0085] Table 1 Performance of different models in single domain and mixed domain

[0086]

[0087]

[0088] Type refers to the type of data set that the algorithm targets. Single refers to a single field, mixed refers to using an expert system for all fields, and multi refers to using multiple networks similar to expert systems. Then finance refers to finance, ent refers to entertainment, and society refers to social news. In addition, there are six other fields. The table is the same, and the decimals inside are the F1 values ​​for detecting the authenticity of news. The last three overall tests all the news in the nine fields together, with the highest accuracy, indicating that our model is more accurate in judging the authenticity of news and fields.

[0089] Table 2 Results of time complexity

[0090]

[0091] This table shows the time complexity experiment. Full is the complete model, and fast is the model with two modules removed and an acceleration algorithm used. The fourth and fifth rows show the change in time complexity after removing the two modules. This can be used for scenarios where accuracy is not a high requirement but speed is.

[0092] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0093] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0094] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0095] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0096] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A cross-domain rapid false news detection method, characterized in that: Include: First, use RoBERTa-wwm to embed the original news text to obtain the corresponding word vector representation; Then, multiple expert systems are used to extract features from the word vector matrix composed of word vectors; Perform sentiment analysis on the original news text to obtain sentiment results, and propose an attention mechanism to allocate attention to word vectors. The sentiment results and attention mechanism output assist the expert system in locating areas. Using feed-forward neural network APN to extract long-range dependency information also assists expert system in locating areas; Finally, after multiple field positioning, the final classification features are obtained. These features are then used to perform classification tasks using a multi-layer perceptron to obtain detection results on whether the news is true or false.

2. The cross-domain rapid false news detection method according to claim 1 is characterized in that: The expert system adopts a TextCNN model containing multiple convolution kernels of different sizes, performs convolution and pooling operations on the word vector matrix to obtain the feature vector μ; and uses a residual connection after the original input to retain the original input information in the output.

3. The cross-domain rapid false news detection method according to claim 2 is characterized in that: The attention mechanism is introduced into the expert network to represent the weight of each expert system in the decision-making process of different news fields.

4. The cross-domain rapid false news detection method according to claim 2 is characterized in that: Use the sentiment analysis model based on chinese-roberta-wwm-ext-base to extract sentiment features from the original news text and obtain the sentiment feature vector V emo ; Propose an attention mechanism to allocate attention to each word vector and obtain the output β of the attention mechanism; For the fields of science and medical health, use domain BERT to embed the original news text to obtain word vectors For other fields, RoBERTa-wwm is still used; finally, after a gating, the weight coefficient g is obtained after softmax: Where W is a learnable parameter.

5. The cross-domain rapid false news detection method according to claim 4 is characterized in that: Use the feedforward neural network APN to extract long-distance dependency information. The expression is as follows: the APN =FFN(W,{ω}) Where FFN represents a feedforward neural network, W is a learnable parameter, and w is the word vector output by RoBERTa-wwm; APN assigns different weights α to the expert system according to the long-distance text information feature connection.

6. The cross-domain rapid false news detection method according to claim 5 is characterized in that: After multiple field positioning, the final classification feature is expressed as: Where e represents the total number of expert systems; the expression for performing classification tasks using a multilayer perceptron is as follows: Where MLP is a multi-layer perceptron. is the predicted label.

7. The cross-domain rapid false news detection method according to claim 1 is characterized in that: The loss function used by the model is binary cross entropy loss, and the model is trained using a dataset consisting of news from multiple fields.

8. The cross-domain rapid false news detection method according to claim 1 is characterized in that: A dual-model neural network acceleration method based on gradient optimization is used to optimize model training speed.

9. A cross-domain rapid fake news detection system, characterized in that: A method for implementing a cross-domain rapid false news detection method as claimed in any one of claims 1 to 8, comprising: The text embedding module is used to embed the original news text using RoBERTa-wwm to obtain the corresponding word vector representation; The domain feature extraction module is used to extract features from the word vector matrix composed of word vectors using multiple expert systems; Sentiment analysis attention module, which is used to perform sentiment analysis on the original news text to obtain sentiment results, and proposes an attention mechanism to allocate attention to word vectors. The sentiment results and attention mechanism output assist the expert system in locating the field; Auxiliary field positioning module, which is used to extract long-distance dependency information using a feed-forward neural network (APN) and also assist the expert system in locating the field; The classification decision module is used to obtain the final classification features after multiple field positioning. These features are then classified using a multi-layer perceptron to determine whether the news is true or false.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images