False news identification method, system and equipment based on evidence perception and medium

By constructing semantic graphs and syntax graphs and using dual-channel semantics to enhance network and evidence-aware three-layer attention network, the problem of insufficient information interaction in fake news detection in the existing technology is solved, and the accuracy of fake news detection is improved and The effect of combining information.

CN118916487APending Publication Date: 2024-11-08UNIV OF JINAN
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202410944063.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing fake news identification methods based on evidence perception are difficult to capture long-distance semantic dependencies in text information and fine-grained semantic associations between news and evidence, affecting the accuracy of fake news detection.

Method used

By converting news text and evidence text into semantic graphs and grammatical graphs, a dual-channel semantic enhancement network is built, and a three-layer attention network of evidence perception can be combined with an evidence-awareness to achieve information interaction and semantic similarity capture between news and evidence.

Benefits of technology

The accuracy of false news detection is improved, and by effectively combining news and evidence information to capture rich text context information, the interpretability of the detection results is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916487A_ABST
    Figure CN118916487A_ABST
Patent Text Reader

Abstract

The invention discloses a false news identification method, system and device based on evidence perception and a medium, and the method comprises the steps: carrying out the information transmission of a news text semantic graph, a news text syntax graph, an evidence text semantic graph and an evidence text syntax graph, obtaining a news text semantic representation, a news text grammatical representation, an evidence text semantic representation and an evidence text grammatical representation; splicing the news text semantic representation and the news text grammatical representation to obtain a news context representation; splicing the evidence text semantic representation and the evidence text grammar representation to obtain an evidence context representation; performing information interaction on the news context representation and the evidence context representation to obtain a first intermediate variable and a second intermediate variable, and splicing the two intermediate variables to obtain an evidence-based text representation; and classifying the evidence-based text representation to obtain an identification result of whether the to-be-identified text is false news or not.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a method, system, device and medium for identifying false news based on evidence perception. Background Art

[0002] The Internet provides an effective way for individuals and organizations to disseminate information to a wide audience. With the rise of social media platforms, anyone can be a publisher of Internet information, which makes the scope of information dissemination more extensive, especially false narratives and misleading content can quickly attract attention. The widespread dissemination of false information in different fields has caused great and lasting damage to public interests and social welfare. The harmful effects of false information will intensify over time, so it is crucial to verify the authenticity of news content to prevent the spread of false news and maintain a credible environment on social media.

[0003] The method of false news identification based on evidence perception is to collect verified factual information as evidence from credible fact-checking websites, and use the evidence to verify the authenticity of a given news. To verify the news facts, the semantic representation of the text is first extracted from the news text and related evidence, and then the semantic relevance between the news text and the evidence text is captured, and the authenticity of the news is inferred by combining multiple pieces of evidence information. False news detection can be regarded as a binary classification problem. With the development of deep learning, various deep neural network models have been widely used in false news detection, including recurrent neural networks (RNN), convolutional neural networks (CNN), attention networks, and graph neural networks (GNN). Traditional deep text classification models such as CNN and RNN construct text as a sequence, while GNN constructs text as a collection of edge-connected word networks. Compared with CNN and RNN, graph neural networks can more effectively model the complex relationship between data, capture long-distance semantic association information, and have stronger reasoning ability than sequence models.

[0004] Evidence-aware fake news detection can provide users with evidence to judge the authenticity of news and improve the interpretability of fake news detection. The existing evidence-based fake news identification method extracts the semantic information of news and evidence respectively, then interacts with the news and evidence, and verifies the authenticity of news by semantically comparing the news and evidence. However, the current methods mainly have the following problems: First, sequence neural networks are widely used in fake news detection, which identify fake news by encoding news content and social context information. When obtaining contextual semantic information from news or evidence, sequence neural networks have difficulty capturing long-range semantic dependencies in text information and cannot integrate relevant information scattered in evidence, resulting in failure to learn complete text context information, affecting the accuracy of fake news detection. Second, when verifying news, it is necessary to capture the fine-grained semantic association between news and evidence, and there is also a strong semantic correlation between evidence. In the current method, the interaction between news and evidence is not fully captured. Since these useful information cannot be obtained, the final judgment of fake news detection is affected. Summary of the invention

[0005] In order to address the deficiencies of the prior art, the present invention provides a false news identification method, system, device and medium based on evidence perception; converts news text and evidence text into semantic graphs and grammatical graphs, constructs a dual-channel semantic enhancement network, calculates the semantic representation and grammatical representation of news text and evidence text, and realizes information interaction between long-distance nodes and non-continuous nodes; constructs an evidence-aware three-layer attention network to perform information interaction between news and evidence, captures the semantic similarity between news and evidence, effectively combines and utilizes news information and evidence information, and improves the accuracy of false news detection.

[0006] On the one hand, a fake news identification method based on evidence perception is provided, including:

[0007] Acquire a text to be identified; the text to be identified includes: news text and evidence text; perform data preprocessing on the text to be identified; extract text semantic features and text grammatical features of the preprocessed text to be identified, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features;

[0008] Based on the semantic features of news text, the grammatical features of news text, the semantic features of evidence text and the grammatical features of evidence text, a news text semantic graph, a news text grammatical graph, an evidence text semantic graph and an evidence text grammatical graph are constructed respectively;

[0009] Inputting the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained fake news recognition model to obtain fake news recognition results;

[0010] Among them, the trained fake news recognition model is used to: perform information propagation on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation; splice the news text semantic representation and the news text grammatical representation to obtain the news context representation; splice the evidence text semantic representation and the evidence text grammatical representation to obtain the evidence context representation; perform information interaction between the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable, and splice the two intermediate variables to obtain the evidence-based text representation; classify the evidence-based text representation to obtain the recognition result of whether the text to be identified is fake news.

[0011] On the other hand, a fake news identification system based on evidence perception is provided, including:

[0012] The preprocessing module is configured to: obtain a text to be recognized, wherein the text to be recognized includes: a news text and an evidence text; perform data preprocessing on the text to be recognized; extract text semantic features and text grammatical features of the preprocessed text to be recognized, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features;

[0013] A graph construction module is configured to: construct a news text semantic graph, a news text grammatical graph, an evidence text semantic graph, and an evidence text grammatical graph, respectively, based on news text semantic features, news text grammatical features, evidence text semantic features, and evidence text grammatical features;

[0014] A recognition module configured to input the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained false news recognition model to obtain a false news recognition result;

[0015] Among them, the trained fake news recognition model is used to: perform information propagation on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation; splice the news text semantic representation and the news text grammatical representation to obtain the news context representation; splice the evidence text semantic representation and the evidence text grammatical representation to obtain the evidence context representation; perform information interaction between the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable, and splice the two intermediate variables to obtain the evidence-based text representation; classify the evidence-based text representation to obtain the recognition result of whether the text to be identified is fake news.

[0016] On the other hand, there is also provided an electronic device, comprising:

[0017] a memory for non-transitory storage of computer-readable instructions; and

[0018] a processor for executing the computer readable instructions,

[0019] When the computer-readable instructions are executed by the processor, the method described in the first aspect is executed.

[0020] On the other hand, a storage medium is provided, which non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the instructions of the method described in the first aspect are executed.

[0021] On the other hand, a computer program product is provided, comprising a computer program, wherein the computer program is used to implement the method described in the first aspect when running on one or more processors.

[0022] The above technical solution has the following advantages or beneficial effects:

[0023] The present invention provides an evidence-aware fake news identification method, which generates text context representation by using a dual-channel semantic enhancement network, realizes information interaction between long-distance nodes and non-continuous nodes, captures rich text context information, and improves the accuracy of text representation.

[0024] The present invention provides an evidence-aware fake news identification method, which uses an evidence-aware three-layer attention network to perform information interaction between news and evidence, mines useful information in the evidence, obtains evidence-based text context representation, effectively combines and utilizes news information and evidence information, and improves the accuracy of fake news detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0026] Figure 1 is a flow chart of an evidence-aware false news identification method described in an embodiment of the present invention;

[0027] Figure 2 4 is a flow chart of the dual-channel semantic enhancement network described in an embodiment of the present invention.

[0028] Figure 3 is a flowchart of the evidence-aware three-layer attention network described in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0030] Embodiment 1

[0031] This embodiment provides a fake news identification method based on evidence perception;

[0032] The fake news identification method based on evidence perception includes:

[0033] S101: Acquire a text to be recognized; the text to be recognized includes: a news text and an evidence text; perform data preprocessing on the text to be recognized; extract text semantic features and text grammatical features of the preprocessed text to be recognized, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features, and evidence text grammatical features;

[0034] S102: Based on the news text semantic features, the news text grammatical features, the evidence text semantic features and the evidence text grammatical features, respectively construct a news text semantic graph, a news text grammatical graph, an evidence text semantic graph and an evidence text grammatical graph;

[0035] S103: inputting the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained false news recognition model to obtain a false news recognition result;

[0036] The trained fake news recognition model is used to: propagate information on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation;

[0037] The news text semantic representation and the news text grammatical representation are concatenated to obtain a news context representation; the evidence text semantic representation and the evidence text grammatical representation are concatenated to obtain an evidence context representation; the news context representation and the evidence context representation are interacted with each other to obtain a first intermediate variable and a second intermediate variable, and the two intermediate variables are concatenated to obtain an evidence-based text representation; the evidence-based text representation is classified to obtain an identification result of whether the text to be identified is false news.

[0038] Furthermore, the S101: obtaining the text to be recognized; the text to be recognized includes: news text and evidence text; performing data preprocessing on the text to be recognized, including: data cleaning processing, stop word deletion processing and word form restoration processing.

[0039] Data cleaning processing refers to: removing useless information from the text, deleting punctuation marks and other special characters in news text and evidence text, and retaining only meaningful text information;

[0040] Stop word removal means: after word segmentation of news text and evidence text, stop words that appear frequently but have little practical significance are removed. These words may interfere with subsequent operations.

[0041] Lemmatization means that different forms of the same word may represent the same meaning. Lemmatization is performed using the NLTK tool to extract the stems of the words and convert them to lowercase. NLTK is the abbreviation of Natural Language Toolkit.

[0042] Furthermore, the S101: extracts text semantic features and text grammatical features of the preprocessed text to be identified, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features, uses the BERT model to implement semantic feature extraction, and uses StanfordNLP to implement grammatical feature extraction.

[0043] Further, the S102: based on the news text semantic features, the news text grammatical features, the evidence text semantic features and the evidence text grammatical features, respectively construct a news text semantic graph, a news text grammatical graph, an evidence text semantic graph and an evidence text grammatical graph, wherein the construction process of the news text semantic graph and the evidence text semantic graph is the same, and the construction process of the news text semantic graph includes:

[0044] Set the size of the sliding window and the moving step of the sliding window;

[0045] Slide the sliding window on the news text. Each time it slides, the connectivity between any two words in the current sliding window is judged. The words are taken as nodes. Based on whether there is a co-occurrence relationship between the words, it is determined whether there are interconnected edges between the nodes. If there is a co-occurrence relationship, there are interconnected edges. If there is no co-occurrence relationship, there are no interconnected edges. The semantic features of the news text are used as the attributes of the corresponding nodes to obtain the news text semantic graph.

[0046] The co-occurrence relationship means that if two words appear at the same time in a given sliding window, they are considered to have a co-occurrence relationship.

[0047] For example, a fixed-size sliding window mechanism is used to describe the connectivity between any two words, with nodes representing words and edges representing the co-occurrence relationship between words. In order to obtain the contextual information of the text and capture the relevant information between non-continuous nodes, each repeated word is merged into one node to construct a text semantic graph.

[0048] Among them, the construction process of the news text grammar graph and the evidence text grammar graph is consistent. The construction process of the news text grammar graph includes: extracting the grammatical dependency relationship between words, obtaining the grammatical weight of words, and treating words as nodes. If there is a grammatical dependency relationship between words, it means that there is an edge connecting the corresponding two nodes. Otherwise, it means that there is no edge connecting the corresponding two nodes. The grammatical features of the news text are used as attributes of the node, and the news text grammar graph is obtained.

[0049] Exemplarily, the word grammar weight is obtained according to the grammatical dependency between words extracted by Stanford CoreNLP. The nodes represent words, the edges represent the syntactic dependency between words, and the text grammar graph is constructed.

[0050] It should be understood that the Stanford CoreNLP refers to a natural language processing toolkit that has rich functions, including word segmentation, part-of-speech tagging, named entity recognition, syntactic analysis, grammatical dependency extraction, sentiment analysis, etc.

[0051] Furthermore, the S103: inputting the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph into the trained false news recognition model to obtain a false news recognition result. The training process of the trained false news recognition model includes:

[0052] Constructing a training set, wherein the training set is a news text semantic graph, a news text grammatical graph, an evidence text semantic graph, and an evidence text grammatical graph of news texts with known labels of whether the news is false or not;

[0053] The training set is input into the fake news recognition model to train the fake news recognition model. When the total loss function value of the model no longer decreases, or the number of iterations exceeds the set number, the training is stopped to obtain the trained fake news recognition model.

[0054] Furthermore, if Figure 1 , Figure 2 and Figure 3 As shown, the trained fake news identification model includes:

[0055] A first dual-channel semantic enhancement network and a second dual-channel semantic enhancement network in parallel;

[0056] The internal structures of the first dual-channel semantic enhancement network and the second dual-channel semantic enhancement network are consistent. The first dual-channel semantic enhancement network includes: a first branch and a second branch in parallel; the input end of the first branch is used to input a news text semantic graph; the input end of the second branch is used to input a grammatical graph of the news text; the output end of the first branch and the output end of the second branch are both connected to the input end of the first series splicing unit, and the output end of the first series splicing unit outputs a news context representation;

[0057] The second dual-channel semantic enhancement network includes: a third branch and a fourth branch in parallel; the input end of the third branch is used to input the semantic graph of the evidence text; the input end of the fourth branch is used to input the syntax graph of the evidence text; the output end of the third branch and the output end of the fourth branch are both connected to the input end of the second serial splicing unit, and the output end of the second serial splicing unit outputs the evidence context representation;

[0058] An evidence-aware three-layer attention network is used to interact the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable. The two intermediate variables are concatenated to obtain the evidence-based text representation.

[0059] The evidence-based text representation is input into the classifier to obtain the recognition result of whether the text to be identified is fake news.

[0060] Furthermore, the internal structures of the first branch and the third branch are consistent, and the first branch includes:

[0061] A gated graph neural network is used to propagate information on the news text semantic graph. Each node in the news text semantic graph receives information from its neighboring nodes and merges it into its own node information to update its own node representation and obtain the text semantic representation.

[0062] Furthermore, the gated graph neural network is used to propagate information on the news text semantic graph, and each node in the news text semantic graph receives information from its neighboring nodes and merges it into its own node information to update its own node representation, thereby obtaining a text semantic representation, including:

[0063] (11): A gated graph neural network is used to construct a semantic encoder, propagate context information on the news text semantic graph, and update node representations; each node merges the information received from its neighbor nodes into its own information to update its own feature representation, and stacks multiple layers of semantic encoders to achieve high-order feature information interaction; the learning process of the gated graph neural network is:

[0064] a t =∑A ij h j Wa (1)

[0065] z t = σ(W z a t +U z h t +b z ) (2)

[0066] r t = σ(W r a t +U r h t +b r ) (3)

[0067] l t =tanh(W h a t +U h (r t ⊙h t )+b h ) (4)

[0068] h t = l t ⊙z t +h t ⊙(1-z t ) (5)

[0069] Among them, a t represents the information that node i aggregates from its neighboring nodes, h t represents the embedding representation of the t-th layer of the gated graph neural network, l t represents the newly generated information of the tth layer. W, U and b are all trainable parameters. A ij is the adjacency matrix corresponding to the semantic graph, σ is the sigmoid function, z and r are the update gate and reset gate respectively.

[0070] The gated graph neural network is used to propagate information on the semantic graph of news text to obtain rich semantic information and obtain the semantic representation of news text.

[0071] (12): Use graph learning to remove redundant information in the evidence sentence:

[0072]

[0073] p=top k (score) (9)

[0074] Among them, W ij represents the correlation score between evidence node i and news node j, represents the i-th evidence node, represents the jth news node, represents the redundancy score of evidence node i, score represents the redundancy score list of evidence nodes, Gem represents a layer of semantic encoder, represents the adjacency matrix corresponding to the evidence semantic graph, and p represents the index of the first k evidence nodes.

[0075] The news text semantic graph and the evidence text semantic graph are respectively passed through the semantic channel to obtain the updated news semantic representation. and evidence semantic representation

[0076] It should be understood that the beneficial effect of the above step (12) is that there is usually redundant information in the evidence sentence, which may affect the model results. In order to reduce the impact of redundant information on model performance, graph learning is used to remove the redundant information in the evidence sentence.

[0077] Redundant information is related to the information contained in each node and the contextual information involved in its neighboring nodes. The redundancy score is calculated based on the information of each node and the contextual information received from its neighboring nodes. First, the correlation score between the evidence node and the news node is calculated, and the redundancy score of each evidence node is calculated based on the correlation score. The final redundancy score is calculated through a layer of semantic encoder, and the nodes are sorted according to these redundancy scores, and the k nodes with the highest scores are discarded. The redundant information in the evidence is filtered through the graph learning module to obtain a refined semantic representation.

[0078] Furthermore, the internal structure of the second branch is consistent with that of the fourth branch, and the second branch includes: a multi-head attention mechanism layer, a dense connection layer and a linear combination layer connected in sequence.

[0079] (21) The node representations in the grammar graph are updated through the multi-head attention mechanism layer to generate a grammar weight graph;

[0080] (22) The local and non-local information of the grammatical weight graph is captured through a densely connected layer to obtain an updated node representation;

[0081] (23) The updated node representations are connected through a linear combination layer to obtain the text grammatical representation.

[0082] Furthermore, (21) updates the node representation in the grammar graph through the multi-head attention mechanism layer to generate a grammar weight graph, including:

[0083] The graph adjacency matrix corresponding to the text grammar graph is based on Generate N graph adjacency matrices using multi-head attention mechanism represents the adjacency matrix of the tth fully connected grammar weight graph, The calculation formula is:

[0084]

[0085] Where Q and K are equal to The syntax diagram at h (t-1) , and They are all parameter matrices, and d represents the input feature dimension.

[0086] get Afterwards, use Calculate the corresponding syntax graph node i representation The calculation formula is as follows:

[0087]

[0088] Among them, W (t) represents the weight matrix, represents the syntax graph node representation of the t-1th layer, b (t) represents the bias vector and ρ is the activation function.

[0089] The original grammar graph G is transformed into (0) Converted into N weight graphs (G (1) ,…,G (N) ), the corresponding syntax graph node is represented as (h (1) ,…,h (N) ).

[0090] Furthermore, the (22) captures the local and non-local information of the grammatical weight graph through a densely connected layer to obtain an updated node representation, including:

[0091] A dense connection layer is used to capture the rich local and non-local information in the text syntax graph. The dense connection layer consists of M sub-layers, each of which is connected to any subsequent sub-layer. The output of the dense connection layer Defined as:

[0092]

[0093] Among them, x represents the input representation of the densely connected layer, is the output of the ith sub-layer.

[0094] Based on the generated N weight graphs (G (1) ,…,G (N) ), construct N densely connected layers to connect N weight graphs respectively, and update the node representation in each weight graph

[0095]

[0096] Among them, t∈(1,2,…,N), W (t) is the weight matrix, b (t) is the bias term.

[0097] Furthermore, the updated node representations in (23) are connected through a linear combination layer to obtain a text grammar representation, including:

[0098] Construct a linear combination layer to represent the nodes generated by N densely connected layers Connect them together to get the text grammar representation H based on context information syn =W syn h out +b syn .

[0099] in, W syn is the weight matrix, b syn is the bias matrix.

[0100] The news grammar graph and evidence grammar graph generate news grammar representation through grammar channel and evidence grammar representation

[0101] The obtained news text semantic representation and news text grammar representation Concatenate to get the news context representation H c , the obtained evidence text semantic representation and evidence text grammar representation Concatenate to get contextual representation of evidence

[0102] Furthermore, the evidence-aware three-layer attention network includes:

[0103] The word-level attention layer, evidence-level attention layer, and sentence-level attention layer are connected in sequence.

[0104] The word-level attention layer calculates the attention score of each word in the news and evidence to obtain the attention representation of the text word level;

[0105] The evidence-level attention layer constructs a text graph with each piece of evidence as a node to perform information interaction between evidence sentences;

[0106] The sentence-level attention layer performs information interaction between the news text and the corresponding multiple evidences.

[0107] Furthermore, the word-level attention layer calculates the attention score of each word in the news and evidence to obtain the attention representation of the text word level, including:

[0108] Construct a word-level attention layer. Since the words in a sentence contribute differently to the entire sentence, a word-level attention mechanism is used to calculate the word-level attention representation. c Calculate the kth word in each evidence The attention score is calculated by using the attention weight to perform weighted summation on each word in the evidence graph to obtain the updated evidence representation. The calculation formula is as follows:

[0109]

[0110] in, represents the news node embedding representation, n is the length of the news, W represents the embedding representation of the i-th news node. c and W w is a trainable parameter, l e is the length of the evidence, p k represents the attention score of the kth word in the evidence, a k represents the attention score of the kth word in the normalized evidence, Indicates evidence.

[0111] The publisher information of news and evidence is also very important for verifying the authenticity of news. e Connecting with evidence representation There are n pieces of evidence related to a piece of news. Each piece of evidence updates its own feature representation through the word-level attention layer to obtain the updated n pieces of evidence representation (h e1 ;…;h en ).

[0112] Furthermore, the evidence-level attention layer constructs a text graph with each piece of evidence as a node to perform information interaction between evidence sentences, including:

[0113] Construct an evidence-level attention layer. Since there is semantic correlation between evidences, in order to propagate information between evidences, an evidence-level attention layer is used to interact with evidences. Take each evidence as a node and construct an evidence text graph. Use MLP to calculate the attention weight p between evidence node i and its neighbor node j. ij :

[0114]

[0115] in, and is the weight matrix, represents the evidence node i at the tth layer, Represents the evidence node j at the tth layer.

[0116] Then, the attention weights are normalized using the softmax function, and the weighted sum of all neighbor nodes is performed to obtain the feature representation of evidence node i at the tth layer:

[0117] a ij =softmax(p ij ) (19)

[0118]

[0119] Among them, a ij represents the attention weight between evidence node i and its neighbor node j, Represents the feature representation of evidence node j at the t-1th layer.

[0120] By stacking T layers of evidence-level attention layers, each evidence node interacts with other evidence nodes to obtain related semantic information, and the updated evidence representations are connected:

[0121] Furthermore, the sentence-level attention layer performs information interaction between the news text and the corresponding multiple evidences, including:

[0122] Construct a sentence-level attention layer. Since each evidence sentence has a different impact on the authenticity of the news, sentence-level attention is used to interact between the news and each piece of evidence. c Connect with the news context representation and update the news context representation as In the sentence-level attention layer, the news context representation h c The information is exchanged between the corresponding n pieces of evidence. The evidence representation is calculated according to the following formula

[0123]

[0124] in, represents the qth piece of evidence, W c , W s is a trainable parameter, n is the number of evidences corresponding to each news, and p q represents the attention score of the qth evidence sentence, a q represents the normalized attention score of the qth evidence sentence.

[0125] Finally, the evidence is represented by connecting With news context representation h c Put together to get evidence-based text representation

[0126] It should be understood that the three-layer attention network module of evidence perception: Figure 3 As shown in Figure 2, news and evidence are passed through a dual-channel semantic enhancement network to obtain the contextual representation of news H. c and contextual representation of evidence In order to exchange information between news and evidence, an evidence-aware three-layer attention network is used. This module consists of three parts. First, the word-level attention layer is used to calculate the attention weight of each word in the news and evidence to obtain the word-level attention representation; second, a text graph is constructed with each evidence as a node, and information exchange between evidence is carried out through the evidence-level attention layer; finally, information exchange is carried out between the news text and the corresponding multiple evidences at the sentence-level attention layer.

[0127] Furthermore, the news text semantic graph, news text grammatical graph, evidence text semantic graph and evidence text grammatical graph are respectively subjected to information dissemination to obtain news text semantic representation, news text grammatical representation, evidence text semantic representation and evidence text grammatical representation; wherein the process of performing information dissemination on the news text semantic graph is consistent with the process of performing information dissemination on the main text semantic graph, and the process of performing information dissemination on the news text semantic graph is realized through the first branch.

[0128] The process of information dissemination on the news text grammar graph is consistent with the process of information dissemination on the evidence text grammar graph, and the process of information dissemination on the news text grammar graph is implemented through the second branch.

[0129] Furthermore, the news text semantic representation and the news text grammatical representation are spliced ​​to obtain the news context representation; the evidence text semantic representation and the evidence text grammatical representation are spliced ​​to obtain the evidence context representation; the splicing is serial splicing.

[0130] Furthermore, the news context representation and the evidence context representation are subjected to information interaction to obtain a first intermediate variable and a second intermediate variable, and the two intermediate variables are concatenated to obtain an evidence-based text representation, and the information interaction is implemented based on an evidence-aware three-layer attention network.

[0131] Furthermore, the classification of the evidence-based text representation to obtain an identification result of whether the text to be identified is fake news includes:

[0132] The evidence-based text representation uses a classifier to make the final classification judgment to determine whether it is fake news, and presents the results to the user in a visual way.

[0133] An evidence-aware fake news identification method and system includes: obtaining news and related evidence as text to be identified, and performing data preprocessing on the text to be identified; extracting text semantic features and text grammatical features based on the preprocessed text to be identified, and then constructing corresponding text semantic graphs and text grammatical graphs respectively; based on the text semantic graph and the text grammatical graph, using a dual-channel semantic enhancement network to propagate semantic graph information and grammatical graph information, generating text semantic representation and text grammatical representation, and splicing the two representations to generate news context representation and evidence context representation; based on the news context representation and the evidence context representation, using an evidence-aware three-layer attention network module to perform information interaction between news and evidence, and obtaining evidence-based text representation; splicing the evidence-based text representation and the news context representation to generate the final representation, and using a classifier to obtain the result of whether the text to be identified is fake news. The purpose of enriching text context information, capturing the information interaction between news and evidence, and improving the accuracy of verifying the authenticity of news is achieved.

[0134] Embodiment 2

[0135] This embodiment provides a fake news identification system based on evidence perception;

[0136] The fake news identification system based on evidence perception includes:

[0137] The preprocessing module is configured to: obtain a text to be recognized, wherein the text to be recognized includes: a news text and an evidence text; perform data preprocessing on the text to be recognized; extract text semantic features and text grammatical features of the preprocessed text to be recognized, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features;

[0138] A graph construction module is configured to: construct a news text semantic graph, a news text grammatical graph, an evidence text semantic graph, and an evidence text grammatical graph, respectively, based on news text semantic features, news text grammatical features, evidence text semantic features, and evidence text grammatical features;

[0139] A recognition module configured to input the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained false news recognition model to obtain a false news recognition result;

[0140] Among them, the trained fake news recognition model is used to: perform information propagation on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation; splice the news text semantic representation and the news text grammatical representation to obtain the news context representation; splice the evidence text semantic representation and the evidence text grammatical representation to obtain the evidence context representation; perform information interaction between the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable, and splice the two intermediate variables to obtain the evidence-based text representation; classify the evidence-based text representation to obtain the recognition result of whether the text to be identified is fake news.

[0141] It should be noted that the above-mentioned preprocessing module, graph construction module and recognition module correspond to steps S101 to S103 in Embodiment 1, and the examples and application scenarios implemented by the above-mentioned modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0142] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0143] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0144] Embodiment 3

[0145] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory so that the electronic device executes the method described in the above embodiment one.

[0146] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0147] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0148] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

[0149] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0150] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0151] Embodiment 4

[0152] This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in the first embodiment is completed.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. The fake news identification method based on evidence perception is characterized by: include: Get the text to be recognized; The text to be identified includes: news text and evidence text; data preprocessing is performed on the text to be identified; text semantic features and text grammatical features are extracted from the preprocessed text to be identified, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features; Based on the semantic features of news text, the grammatical features of news text, the semantic features of evidence text and the grammatical features of evidence text, a news text semantic graph, a news text grammatical graph, an evidence text semantic graph and an evidence text grammatical graph are constructed respectively; Inputting the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained fake news recognition model to obtain fake news recognition results; Among them, the trained fake news recognition model is used to: perform information propagation on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation; splice the news text semantic representation and the news text grammatical representation to obtain the news context representation; splice the evidence text semantic representation and the evidence text grammatical representation to obtain the evidence context representation; perform information interaction between the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable, and splice the two intermediate variables to obtain the evidence-based text representation; classify the evidence-based text representation to obtain the recognition result of whether the text to be identified is fake news.

2. The method for identifying fake news based on evidence perception as claimed in claim 1, characterized in that: Based on the news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features, a news text semantic graph, a news text grammatical graph, an evidence text semantic graph and an evidence text grammatical graph are constructed respectively, wherein the construction process of the news text semantic graph and the evidence text semantic graph is the same, and the construction process of the news text semantic graph includes: Set the size of the sliding window and the moving step of the sliding window; Slide the sliding window on the news text. Each time the sliding window slides, the connectivity between any two words in the current sliding window is determined. The words are taken as nodes. Based on whether there is a co-occurrence relationship between the words, it is determined whether there are interconnected edges between the nodes. If there is a co-occurrence relationship, there are interconnected edges. If there is no co-occurrence relationship, there are no interconnected edges. The semantic features of the news text are taken as the attributes of the corresponding nodes, and then a semantic graph of the news text is obtained. Among them, the construction process of the news text grammar graph and the evidence text grammar graph is consistent. The construction process of the news text grammar graph includes: extracting the grammatical dependency relationship between words, obtaining the grammatical weight of words, and treating words as nodes. If there is a grammatical dependency relationship between words, it means that there is an edge connecting the corresponding two nodes. Otherwise, it means that there is no edge connecting the corresponding two nodes. The grammatical features of the news text are used as attributes of the node, and the news text grammar graph is obtained.

3. The method for identifying fake news based on evidence perception as claimed in claim 1, characterized in that: The trained fake news identification model includes: A first dual-channel semantic enhancement network and a second dual-channel semantic enhancement network in parallel; The internal structures of the first dual-channel semantic enhancement network and the second dual-channel semantic enhancement network are consistent. The first dual-channel semantic enhancement network includes: a first branch and a second branch in parallel; the input end of the first branch is used to input a news text semantic graph; the input end of the second branch is used to input a grammatical graph of the news text; the output end of the first branch and the output end of the second branch are both connected to the input end of the first series splicing unit, and the output end of the first series splicing unit outputs a news context representation; The second dual-channel semantic enhancement network includes: a third branch and a fourth branch in parallel; the input end of the third branch is used to input the semantic graph of the evidence text; the input end of the fourth branch is used to input the syntax graph of the evidence text; the output end of the third branch and the output end of the fourth branch are both connected to the input end of the second serial splicing unit, and the output end of the second serial splicing unit outputs the evidence context representation; An evidence-aware three-layer attention network is used to interact the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable. The two intermediate variables are concatenated to obtain the evidence-based text representation. The evidence-based text representation is input into the classifier to obtain the recognition result of whether the text to be identified is fake news.

4. The method for identifying fake news based on evidence perception as claimed in claim 3, characterized in that: The internal structure of the first branch is consistent with that of the third branch, and the first branch includes: A gated graph neural network is used to propagate information on the news text semantic graph. Each node in the news text semantic graph receives information from its neighboring nodes and merges it into its own node information to update its own node representation and obtain the text semantic representation.

5. The method for identifying fake news based on evidence perception as claimed in claim 3, characterized in that: The internal structure of the second branch is consistent with that of the fourth branch. The second branch includes: a multi-head attention mechanism layer, a dense connection layer and a linear combination layer connected in sequence; (21) The node representations in the grammar graph are updated through the multi-head attention mechanism layer to generate a grammar weight graph; (22) The local and non-local information of the grammatical weight graph is captured through a densely connected layer to obtain an updated node representation; (23) The updated node representations are connected through a linear combination layer to obtain the text grammatical representation.

6. The method for identifying fake news based on evidence perception as claimed in claim 5, characterized in that: After the multi-head attention mechanism layer updates the node representation in the grammar graph, a grammar weight graph is generated, including: The graph adjacency matrix corresponding to the text grammar graph is based on Generate N graph adjacency matrices using multi-head attention mechanism represents the adjacency matrix of the tth fully connected grammar weight graph, The calculation formula is: Where Q and K are equal to The syntax diagram at h (t-1) , and They are all parameter matrices, and d represents the input feature dimension; get Afterwards, use Calculate the corresponding syntax graph node i representation The calculation formula is as follows: Among them, W (t) represents the weight matrix, represents the syntax graph node representation of the t-1th layer, b (t) represents the bias vector, ρ is the activation function; The original grammar graph G is transformed into (0) Convert to N weight graphs G (1) ,…,G (N) , the corresponding syntax graph node is represented by h (1) ,…,h (N) .

7. The method for identifying fake news based on evidence perception as claimed in claim 3, characterized in that: The evidence-aware three-layer attention network includes: The word-level attention layer, evidence-level attention layer, and sentence-level attention layer are connected in sequence; The word-level attention layer calculates the attention score of each word in the news and evidence to obtain the attention representation of the text word level; The evidence-level attention layer constructs a text graph with each piece of evidence as a node to perform information interaction between evidence sentences; The sentence-level attention layer performs information interaction between the news text and the corresponding multiple evidences.

8. The fake news identification system based on evidence perception is characterized by: include: A preprocessing module is configured to: obtain text to be recognized; The text to be identified includes: news text and evidence text; data preprocessing is performed on the text to be identified; text semantic features and text grammatical features are extracted from the preprocessed text to be identified, respectively, to obtain news text semantic features, news text grammatical features, evidence text semantic features and evidence text grammatical features; A graph construction module is configured to: construct a news text semantic graph, a news text grammatical graph, an evidence text semantic graph, and an evidence text grammatical graph, respectively, based on news text semantic features, news text grammatical features, evidence text semantic features, and evidence text grammatical features; A recognition module configured to input the news text semantic graph, the news text grammatical graph, the evidence text semantic graph, and the evidence text grammatical graph into the trained false news recognition model to obtain a false news recognition result; Among them, the trained fake news recognition model is used to: perform information propagation on the news text semantic graph, the news text grammatical graph, the evidence text semantic graph and the evidence text grammatical graph, respectively, to obtain the news text semantic representation, the news text grammatical representation, the evidence text semantic representation and the evidence text grammatical representation; splice the news text semantic representation and the news text grammatical representation to obtain the news context representation; splice the evidence text semantic representation and the evidence text grammatical representation to obtain the evidence context representation; perform information interaction between the news context representation and the evidence context representation to obtain the first intermediate variable and the second intermediate variable, and splice the two intermediate variables to obtain the evidence-based text representation; classify the evidence-based text representation to obtain the recognition result of whether the text to be identified is fake news.

9. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; as well as a processor for executing the computer readable instructions, When the computer-readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: The computer-readable instructions are non-transitory stored, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1 to 7 are executed.

Citation Information

Patent Citations

  • Double-filtering evidence perception false news detection method based on graph neural network

    CN117349501A