CN112231562A - Network rumor identification method and system - Google Patents

Network rumor identification method and system Download PDF

Info

Publication number
CN112231562A
CN112231562A CN202011099869.8A CN202011099869A CN112231562A CN 112231562 A CN112231562 A CN 112231562A CN 202011099869 A CN202011099869 A CN 202011099869A CN 112231562 A CN112231562 A CN 112231562A
Authority
CN
China
Prior art keywords
rumor
matrix
neural network
network model
texts
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202011099869.8A
Other languages
Chinese (zh)
Other versions
CN112231562B (en
Inventor
段大高
白宸宇
韩忠明
刘文文
张翙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Technology and Business University
Original Assignee
Beijing Technology and Business University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Technology and Business University filed Critical Beijing Technology and Business University
Priority to CN202011099869.8A priority Critical patent/CN112231562B/en
Publication of CN112231562A publication Critical patent/CN112231562A/en
Application granted granted Critical
Publication of CN112231562B publication Critical patent/CN112231562B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/216—Parsing using statistical methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)

Abstract

The invention relates to a network rumor identification method, namely a system, wherein the method comprises the following steps: obtaining a text characteristic matrix according to a plurality of texts containing rumor information; constructing a propagation graph structure, wherein nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relation of rumor information among the plurality of texts; constructing a graph convolution neural network model; the input of the graph convolution neural network model is a text characteristic matrix and an adjacent matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix; training a neural network model according to the rumor characteristic matrix to obtain a rumor recognition model; network rumors are identified according to a rumor identification model. The method trains the convolutional neural network model according to the forwarding and comment relations of rumors among a plurality of texts and trains the neural network model according to the rumor feature matrix, thereby effectively capturing the spread features of wide and dispersed rumor information and effectively identifying the rumor information.

Description

Network rumor identification method and system
Technical Field
The present invention relates to the field of network rumor identification technologies, and in particular, to a network rumor identification method and system.
Background
Under a big data environment, an online social network is gradually integrated with life, entertainment and work of people. Social media become a platform for people to share information and exchange, and have the characteristics of complicated information, free and convenient transmission, large influence and the like, so that the social media become an important transmission medium for public opinion outbreak and temperature rise. Due to the lack of effective supervision, the flooding of false information such as rumors and the like can bring great threats and influences to the fields of politics, economy, culture and the like, and the method becomes one of the main bottlenecks facing the development of a plurality of applications of an online social network. The social media rumor recognition task draws strong attention of researchers in the fields of natural language processing, data mining and the like, and can be used for assisting rumor clearing work such as early warning, prevention, monitoring, management and the like, so that the social media rumor recognition is an effective means for improving the ecological environment quality of online social network information and improving user experience.
There are many methods for detecting rumors. The traditional method is mainly from the viewpoint of manually defining characteristics, mainly by constructing characteristics of related rumor microblogs, and performing event classification by using a machine learning classifier such as a decision tree or a support vector machine, and the universality is poor. Although this feature engineering based approach has achieved some success, it is resource intensive and limited by manually designed rules. Nowadays, with the rapid development of deep learning, the deep neural network model has more advantages in semantic representation and rumor detection application.
Compared with a machine learning method, the neural network model can automatically learn the event characteristics from the data, avoids a large amount of characteristic engineering, and has better expansibility in the aspect of capturing complex semantic relations between contexts. However, most current neural network rumor detection models are used for learning better event characteristics or semantic information, and information propagation of social media in real life has a structural relationship, so that the neural network rumor detection models are not ideal.
Disclosure of Invention
The present invention provides a network rumor identification method and system for effectively identifying rumor information propagated in a network.
In order to achieve the purpose, the invention provides the following scheme:
a network rumor identification method, comprising:
obtaining a text characteristic matrix according to a plurality of texts containing rumor information;
constructing a propagation graph structure, wherein nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relation of the rumor information among the texts;
constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text characteristic matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix;
training a neural network model according to the rumor characteristic matrix to obtain a rumor recognition model;
and identifying the network rumors according to the rumor identification model.
Optionally, the obtaining of the text feature matrix according to the plurality of texts containing the rumor information specifically includes:
training words in a plurality of texts containing rumor information to obtain word vectors of the words;
obtaining a plurality of first word feature matrixes of the texts according to the word vectors;
learning the dependency relationship among the words by the first word feature matrix through an attention mechanism to obtain a second word feature matrix;
and inputting the second word feature matrix into a convolution layer and a maximum pooling layer to obtain a text feature matrix.
Optionally, the method further comprises:
and cleaning and normalizing the length of a plurality of texts.
Optionally, the training may train words in a plurality of texts including rumor information to obtain word vectors of the words, specifically:
training words in a plurality of texts containing rumor information according to a Skip-Gram neural network model in Word2Vec to obtain Word vectors of the words; wherein the Skip-Gram neural network model uses a window size of 5 and the embedding dimension of the word vector is 300.
Optionally, the learning of the dependency relationship between the words by the first word feature matrix through an attention mechanism to obtain a second word feature matrix specifically includes:
learning the first word feature matrix through the same h-group multi-head attention mechanism to obtain h groups of updated first word feature matrices; h is a positive integer greater than 1;
and connecting the h groups of updated first word feature matrixes to obtain a second word feature matrix.
Optionally, the step of inputting the second word feature matrix into a convolutional layer and a maximum pooling layer to obtain a text feature matrix specifically includes:
inputting the second word feature matrix into a convolutional layer to obtain convolutional layer features;
and inputting the convolutional layer characteristics into a maximum pooling layer to obtain the text characteristic matrix.
Optionally, the training of the neural network model according to the rumor feature matrix to obtain a rumor recognition model specifically includes:
inputting the rumor feature matrix into an average pooling layer, a full-link layer and a Softmax layer to obtain a rumor prediction category;
and reversely training the neural network model according to the rumor prediction category and the real rumor category to obtain a rumor recognition model.
Optionally, the neural network model is reversely trained according to the rumor prediction category and the real rumor category to obtain a rumor recognition model, which specifically includes:
obtaining a difference between the rumor prediction category and the true rumor category;
judging whether the difference value is within a preset threshold range;
if so, determining the neural network model as the rumor recognition model;
if not, reversely training the neural network model according to the difference value to enable the difference value to be within the range of the preset threshold value.
A network rumor identification system, comprising:
the text characteristic matrix acquisition module is used for acquiring a text characteristic matrix according to a plurality of texts containing rumor information;
a first construction module, configured to construct a propagation graph structure, where nodes in the graph structure are the texts, and an adjacency matrix in the graph structure is a forwarding and comment relationship between the rumor information and the texts;
the second construction module is used for constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text characteristic matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix;
the training module is used for training a neural network model according to the rumor feature matrix to obtain a rumor recognition model;
and the identification module is used for identifying the network rumors according to the rumor identification model.
Optionally, the text feature matrix obtaining module includes:
the training unit is used for training words in a plurality of texts containing rumor information to obtain word vectors of the words;
the obtaining unit is used for obtaining a plurality of first word feature matrixes of the texts according to the word vectors;
the learning unit is used for learning the dependency relationship among the words by the first word feature matrix through an attention mechanism to obtain a second word feature matrix;
and the input unit is used for inputting the second word feature matrix into the convolutional layer and the maximum pooling layer to obtain a text feature matrix.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention discloses a network rumor identification method and a system, wherein the method comprises the following steps: obtaining a text characteristic matrix according to a plurality of texts containing rumor information; constructing a propagation graph structure, wherein nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relation of the rumor information among the texts; constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text characteristic matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix; training a neural network model according to the rumor characteristic matrix to obtain a rumor recognition model; and identifying the network rumors according to the rumor identification model. The method trains the convolutional neural network model according to the forwarding and comment relations of rumors among a plurality of texts and trains the neural network model according to the rumor feature matrix, thereby effectively capturing the spread features of wide and dispersed rumor information and effectively identifying the rumor information.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
Fig. 1 is a flowchart of a network rumor identification method according to embodiment 1 of the present invention;
FIG. 2 is a schematic diagram of an attention mechanism provided in embodiment 1 of the present invention;
fig. 3 is a schematic diagram of a convolutional layer and a pooling layer for obtaining a text feature matrix according to embodiment 1 of the present invention;
fig. 4 is a schematic diagram illustrating a network rumor identification method according to embodiment 2 of the present invention;
fig. 5 is a system block diagram of a network rumor identification system according to embodiment 3 of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The present invention provides a network rumor identification method and system for effectively identifying rumor information propagated in a network.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Example 1
Fig. 1 is a flowchart of a network rumor identification method according to embodiment 1 of the present invention, as shown in fig. 1, the method includes:
step 101: a text feature matrix is obtained from a plurality of texts containing rumor information.
Step 102: constructing a propagation graph structure, wherein nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relation of the rumor information among the texts.
Step 103: constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text feature matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor feature matrix.
Step 104: and training a neural network model according to the rumor feature matrix to obtain a rumor recognition model.
Step 105: and identifying the network rumors according to the rumor identification model.
In this embodiment, step 101 specifically includes:
step 1011: training words in a plurality of texts containing rumor information to obtain word vectors of the words. The method specifically comprises the following steps: training words in a plurality of texts containing rumor information according to a Skip-Gram neural network model in Word2Vec to obtain Word vectors of the words; wherein the Skip-Gram neural network model uses a window size of 5 and the embedding dimension of the word vector is 300.
Step 1012: and acquiring a plurality of first word feature matrixes of the texts according to the word vectors. First word feature matrix vi∈RL*dIndicating that L represents the text length and d represents the dimension of the word vector.
Step 1013: and learning the dependency relationship among the words by the first word characteristic matrix through an attention mechanism to obtain a second word characteristic matrix. The method specifically comprises the following steps: and learning the first word feature matrix through the same h-group multi-head attention mechanism to obtain h-group updated first word feature matrices. h is a positive integer greater than 1. And connecting the h groups of updated first word feature matrixes to obtain a second word feature matrix.
Fig. 2 is a schematic diagram of an attention mechanism provided in embodiment 1 of the present invention, and as shown in fig. 2, the multi-head attention mechanism is composed of the same h groups, and a calculation formula of a scaling dot product attention method in each group is as follows:
Figure BDA0002724972140000061
wherein Q, K and V are the same, and Q is belonged to RL*d，K∈RL*d，V∈RL*dAll represent the first word feature matrix vi，Wi Q，Wi K，Wi VDifferent parameter matrices representing the linear layer 1, i ∈ [1, h ]]，
Figure BDA0002724972140000062
Representing a canonical number that prevents over-inner product and is easy to train, the dimension of K, QW, is typically choseni Q，KWi K，VWi VFor Q, K, V, obtained by varying the different sets of linearity.
The self-attention mechanism can capture information of different subspaces in different groups by being divided into different groups. Then the calculated h groups Z1～ZhConnected together, a second word feature matrix O e R is output through a linear layer 2L*dThe formula is as follows:
O＝W0(Concat(Z1,Z2...Zh))
wherein W0Is a parameter matrix for the linear layer 2.
Step 1014: and inputting the second word feature matrix into a convolution layer and a maximum pooling layer to obtain a text feature matrix. The method specifically comprises the following steps: and inputting the second word feature matrix into a convolutional layer to obtain convolutional layer features. The formula is as follows:
Figure BDA0002724972140000063
where σ is a nonlinear activation function, W ∈ Rc*dIs a convolution kernel, Word is a Word,
Figure BDA0002724972140000071
is a word vector for a word. Obtaining the characteristic e of the convolution layer through the convolution layer, wherein e is [ e ═ e1,e2...eL-c+1]Where L is the number of words in the text.
And inputting the convolutional layer characteristics into a maximum pooling layer to obtain the text characteristic matrix. The method specifically comprises the following steps: for e ∈ R(L -c+1)*dPerforming maximum pooling in d dimension to obtain m ∈ RdAnd obtaining a text feature matrix through each text feature: m ═ M1,m2...mn)∈Rn*d. Fig. 3 is a schematic diagram of a convolutional layer and a pooling layer for obtaining a text feature matrix according to embodiment 1 of the present invention.
Wherein, step 101 further comprises:
step 1010: and cleaning and normalizing the length of a plurality of texts. The method specifically comprises the following steps: irregular symbols in the text are removed. Because the text length in each piece of text information is different, the length of the text is assumed to be L, when the length of the text is not L, zero is filled in the front of the text, and when the length of the text is larger than L, the part behind the length of L is cut.
In this embodiment, step 102 specifically includes:
the forwarding and commenting relations of the rumors among the texts are formed into a propagation graph structure G (M, E), nodes M in the graph represent the texts related to the rumors, an edge set E in the graph represents the forwarding and commenting relations of the rumors among the texts, and an adjacency matrix A belongs to Rn*nWherein n is the number of nodes in M, and the corresponding position element in the adjacency matrix is aijIf the text has a forwarding or comment relationship, the text is 1, otherwise the text is 0, and the corresponding relationship is as follows:
Figure BDA0002724972140000072
then the text feature matrix and the adjacency matrix are used as input to a graph convolution neural network model, and a rumor feature matrix Conv _ x E R is outputn*cWhere n is the number of rumor related texts and c is the dimension of the text feature. The calculation formula is as follows:
Figure BDA0002724972140000073
wherein
Figure BDA0002724972140000074
In order to add a contiguous matrix of self-loops,
Figure BDA0002724972140000075
Figure BDA0002724972140000076
is composed of
Figure BDA0002724972140000077
Is a degree matrix of theta ∈ Rd*cBe a learnable parameter matrix, Be is a nonlinear activation function.
In this embodiment, step 104 specifically includes:
step 1041: and inputting the rumor feature matrix into an average pooling layer, a full link layer and a Softmax layer to obtain a rumor prediction category. The method specifically comprises the following steps:
the rumor feature matrix Conv _ x ∈ Rn*cInput average pooling layer to obtain graph level output S ∈ R1*cRumor matrix, formula: s is MEAN (Conv _ X). Then inputting the rumor matrix into the full connection layer and the Softmax layer to obtain the rumor prediction category, wherein the formula is as follows:
Figure BDA0002724972140000081
wherein
Figure BDA0002724972140000082
For rumor prediction categories, W ∈ Rc*|class|Is a learnable parameter matrix, and b is a bias term.
Step 1042: and reversely training the neural network model according to the rumor prediction category and the real rumor category to obtain a rumor recognition model. The method specifically comprises the following steps:
obtaining a difference between the rumor prediction category and the true rumor category.
And judging whether the difference value is within a preset threshold range.
If so, determining the neural network model as the rumor recognition model.
If not, reversely training the neural network model according to the difference value to enable the difference value to be within the range of the preset threshold value.
The neural network iterates through a back propagation method and a stochastic gradient descent method, 64Batch sizes are selected as training sample training networks each time, prediction output is calculated, the prediction output is compared with an actual class value, the difference value between comparison results is called as an error, the loss function loss is expressed by using the error, the learnable parameter model weight in the neural network is updated through an optimization method according to the error so as to minimize the loss function loss, the above process is repeated for each Batch until the loss function is minimum for the whole sample set, and parameters in the model are updated.
Example 2
Fig. 4 is a schematic diagram of a network rumor identification method according to embodiment 2 of the present invention, as shown in fig. 4:
(1) take twitter data set as an example, which includes 1490 pieces of source microblog information, including 374 non-rumor microblogs, 370 fake rumor microblogs, 374 uncertain rumor microblogs and 372 real rumor microblogs, respectively. Dividing the data set into three parts of a training set, a verification set and a test set, randomly selecting ten percent as the verification set, taking seventy-five percent of the rest part as the training set, and taking twenty-five percent as the test set.
Rumor set { r, w1,w2,w3,w4,w5Where r denotes the source microblog, w1,w2,w3,w4,w5Representing forwarded or related microblogs. Removing special symbols without meaning in all microblog texts, shielding low-frequency words with the occurrence frequency lower than twice, setting all microblog text contents as 50 words, filling zero in front of text information when the text information is less than 50 words in length, and removing redundant parts when the text length is more than 50 words. The set of 50 words is basically longer than the length of the microblog text processed in all data sets, so that the text information of a large number of microblogs is not lost due to the length setting. For characteristics of rumor microblog categories, use [1, 0,0, 0]、[0，1，0，0]、[0，0，1，0]、[0，0，0，1]One-hot (unique thermal coding) codes are used to represent different categories of non-rumors, pseudo-rumors, true rumors and unidentified rumors. For example, the microblog r category label is [0, 1, 0]This is called pseudo-rumor.
And forming a dictionary by words in all microblog sentences, and training all microblog sentences after data cleaning by using a Skip-Gram algorithm to obtain vector expression of each word. The size of a window used in the model is 5, the embedding d of all words in the model is 300 dimensions, and the microblog word feature matrix is used for representing the text embedding information of a microblog.
(2) The dependency relationships between words are learned through a multi-head attention mechanism. The multi-head attention mechanism consists of the same h groups. The scaled dot product attention method in each group is calculated as follows:
Figure BDA0002724972140000091
wherein Q, K, V are the same, Q ∈ R50*300，K∈R50*300，V∈R50*300All represent M ═ r, w1,w2,w3,w4,w5One piece of microblog information in (W)i Q，Wi K，Wi VParameter matrix representing the ith set of linear transformation layers in FIG. 2, i ∈ [1, h]，
Figure BDA0002724972140000092
Representing a canonical number that prevents over-inner product and is easy to train, the dimension of K, QW, is typically choseni Q，KWi K，VWi VFor Q, K, V, obtained by varying the different sets of linearity. Different groups respectively carry out attention mechanism to obtain different Zi。
Output Z of h group to be finally calculated1～ZhAre connected and multiplied by a matrix W0Performing a linear transformation to obtain an output, W0Is a parameter matrix. The self-attention mechanism can capture information of different subspaces in different groups by being divided into different groups. The dimension of the input matrix is the same as that of the output microblog word feature matrix O belonging to R50*300The polymerization formula is as follows:
O＝W0(Concat(Z1,Z2...Zh))
will be provided with
Figure BDA0002724972140000093
Acting the convolutional layer on the microblog word feature matrix as the convolutional layer and maximum pooling layer input, passing through the maximum pooling layer
Figure BDA0002724972140000094
Obtaining a microblog feature representation mi∈R300. Forming a microblog text matrix M ═ M by the microblog text characteristics related to rumors1,m2...mn]∈R6*300The matrix M represents r, w per row1,w2,w3,w4,w5The text feature of (1).
(3) Constructing a graph structure from the source microblog information and the forwarding information through their forwarding or comment relationships, for example, constructing a propagation graph structure G ═ (M, E), where M ═ r, w are nodes in the graph1,w2,w3,w4,w5And the edge with comment or forwarding relation is r-w1，r-w2，r-w3，w2-w5，w1-w4Thus E ═ E01,e02,e03,e25,e14In which e01,e02,e03Is represented by e14,e25Representing the forwarding among related microblogs and using an adjacency matrix A e R6*6The corresponding position element in the adjacency matrix is represented as aijIf the forwarding relation exists, the forwarding relation is 1, otherwise, the forwarding relation is 0, and the correspondence is as follows:
Figure BDA0002724972140000101
let the adjacency matrix A be R6*6And the microblog text matrix M ═ M1,m2...mn]∈R6*300As input to the convolutional neural network, the propagation structure is captured with the convolutional neural network to update the node characteristics:
Figure BDA0002724972140000102
Figure BDA0002724972140000103
wherein the content of the first and second substances,
Figure BDA0002724972140000104
in order to add a contiguous matrix of self-loops,
Figure BDA0002724972140000105
Figure BDA0002724972140000106
is composed of
Figure BDA0002724972140000107
The degree matrix of (c) is,
Figure BDA0002724972140000108
is a normalization of the adjacency matrix, θ is a learnable parameter matrix, and σ is a nonlinear activation function. The network is arranged into two layers, and the input of the first layer is a microblog text matrix M ═ M1,m2...mn]∈R6*300The adjacency matrix A ∈ R6*6，H(0)Mapping inputs to hidden state representation of the first layer for a propagation formula, H(0)∈R6*64Each row represents the updated node feature representation, and the hidden state H is represented(0)As input feature matrix of the second layer, θ(1)Representing the parameter matrix of network learning in the second layer to obtain the output matrix H of the second layer(1)∈R6*64Each row represents an updated node signature representation.
(4) To the output characteristic matrix H(1)∈R6*64Performing average pooling in hidden dimensions to obtain graph-level output S ∈ R1*64As an expression of rumor characteristics, the probability of different rumor categories obtained by S through the full connection layer and the Softmax layer is given by the formula:
S＝MEAN(H(1))
y＝soft max(SW+b)
wherein W ∈ R6*64Is a learnable parameter matrix, and b is a bias term. The Softmax formula is:
Figure BDA0002724972140000111
after calculation in Softmax layer, the value is converted into relative probability S ═ 0.0057,0.8390,0.0418,0.1135, so the probability of the output as the second category, i.e., the pseudo-rumor, is judged to be the largest, and the event is the pseudo-rumor.
Example 3
Fig. 5 is a block diagram of a network rumor identification system according to embodiment 3 of the present invention, as shown in fig. 5, the system includes:
the text feature matrix obtaining module 201 is configured to obtain a text feature matrix according to a plurality of texts including rumor information.
A first constructing module 202, configured to construct a propagation graph structure, where nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relationship between the rumor information and the texts.
A second constructing module 203, configured to construct a graph convolution neural network model; the input of the graph convolution neural network model is the text feature matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor feature matrix.
And the training module 204 is configured to train a neural network model according to the rumor feature matrix to obtain a rumor recognition model.
An identification module 205 for identifying network rumors according to the rumor identification model.
In this embodiment, the text feature matrix obtaining module 201 includes:
the training unit 2011 is configured to train words in a plurality of texts including rumor information, and obtain word vectors of the words.
The obtaining unit 2012 is configured to obtain a plurality of first word feature matrices of the text according to the word vector.
A learning unit 2013, configured to learn, by using the first word feature matrix, a dependency relationship between the words through an attention mechanism, so as to obtain a second word feature matrix.
An input unit 2014, configured to input the second word feature matrix into the convolutional layer and the maximum pooling layer to obtain a text feature matrix.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
(1) according to the method, the modeling is directly carried out on the text information of the source microblog and the related forwarding microblog, the semantic information implied by the related microblog content is extracted, and the microblog text characteristics are captured through the attention mechanism. Compared with a traditional manual construction method of a feature place model, the method can automatically extract feature representation with higher dimensionality, reduces manual intervention, is more convenient for people to use, and is more suitable for social media complex environments.
(2) According to the method, the text features and the forwarding structure are modeled by using the graph neural network, microblog text information features related to the rumor are updated iteratively through the graph convolutional neural network, the rumor classification prediction is performed on the text features updated through the graph convolutional neural network, and the spread features with wide and dispersed rumor information are effectively captured.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (10)

1. A method for identifying network rumors, comprising:
obtaining a text characteristic matrix according to a plurality of texts containing rumor information;
constructing a propagation graph structure, wherein nodes in the graph structure are a plurality of texts, and an adjacency matrix in the graph structure is a forwarding and comment relation of the rumor information among the texts;
constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text characteristic matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix;
training a neural network model according to the rumor characteristic matrix to obtain a rumor recognition model;
and identifying the network rumors according to the rumor identification model.
2. The method of claim 1, wherein the text feature matrix is obtained from a plurality of texts containing rumor information, and comprises:
training words in a plurality of texts containing rumor information to obtain word vectors of the words;
obtaining a plurality of first word feature matrixes of the texts according to the word vectors;
learning the dependency relationship among the words by the first word feature matrix through an attention mechanism to obtain a second word feature matrix;
and inputting the second word feature matrix into a convolution layer and a maximum pooling layer to obtain a text feature matrix.
3. The method of claim 2, further comprising:
and cleaning and normalizing the length of a plurality of texts.
4. The method of claim 2, wherein the training of words in the plurality of texts containing rumor information to obtain word vectors of the words comprises:
training words in a plurality of texts containing rumor information according to a Skip-Gram neural network model in Word2Vec to obtain Word vectors of the words; wherein the Skip-Gram neural network model uses a window size of 5 and the embedding dimension of the word vector is 300.
5. The method of claim 2, wherein the learning of the first word feature matrix with the dependency relationship between the words through an attention mechanism yields a second word feature matrix, specifically:
learning the first word feature matrix through the same h-group multi-head attention mechanism to obtain h groups of updated first word feature matrices; h is a positive integer greater than 1;
and connecting the h groups of updated first word feature matrixes to obtain a second word feature matrix.
6. The method of claim 2, wherein the step of inputting the second word feature matrix into the convolutional layer and the max-pooling layer to obtain a text feature matrix comprises:
inputting the second word feature matrix into a convolutional layer to obtain convolutional layer features;
and inputting the convolutional layer characteristics into a maximum pooling layer to obtain the text characteristic matrix.
7. The method of claim 1, wherein the training of the neural network model according to the rumor feature matrix yields a rumor recognition model, specifically:
inputting the rumor feature matrix into an average pooling layer, a full-link layer and a Softmax layer to obtain a rumor prediction category;
and reversely training the neural network model according to the rumor prediction category and the real rumor category to obtain a rumor recognition model.
8. The method of claim 7, wherein the neural network model is reversely trained according to the rumor prediction category and the real rumor category to obtain a rumor recognition model, which comprises:
obtaining a difference between the rumor prediction category and the true rumor category;
judging whether the difference value is within a preset threshold range;
if so, determining the neural network model as the rumor recognition model;
if not, reversely training the neural network model according to the difference value to enable the difference value to be within the range of the preset threshold value.
9. A system for identifying network rumors, comprising:
the text characteristic matrix acquisition module is used for acquiring a text characteristic matrix according to a plurality of texts containing rumor information;
a first construction module, configured to construct a propagation graph structure, where nodes in the graph structure are the texts, and an adjacency matrix in the graph structure is a forwarding and comment relationship between the rumor information and the texts;
the second construction module is used for constructing a graph convolution neural network model; the input of the graph convolution neural network model is the text characteristic matrix and the adjacency matrix, and the output of the graph convolution neural network model is a rumor characteristic matrix;
the training module is used for training a neural network model according to the rumor feature matrix to obtain a rumor recognition model;
and the identification module is used for identifying the network rumors according to the rumor identification model.
10. The system of claim 9, wherein the text feature matrix obtaining module comprises:
the training unit is used for training words in a plurality of texts containing rumor information to obtain word vectors of the words;
the obtaining unit is used for obtaining a plurality of first word feature matrixes of the texts according to the word vectors;
the learning unit is used for learning the dependency relationship among the words by the first word feature matrix through an attention mechanism to obtain a second word feature matrix;
and the input unit is used for inputting the second word feature matrix into the convolutional layer and the maximum pooling layer to obtain a text feature matrix.
CN202011099869.8A 2020-10-15 2020-10-15 Network rumor recognition method and system Active CN112231562B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011099869.8A CN112231562B (en) 2020-10-15 2020-10-15 Network rumor recognition method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011099869.8A CN112231562B (en) 2020-10-15 2020-10-15 Network rumor recognition method and system

Publications (2)

Publication Number Publication Date
CN112231562A true CN112231562A (en) 2021-01-15
CN112231562B CN112231562B (en) 2023-07-14

Family

ID=74112967

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011099869.8A Active CN112231562B (en) 2020-10-15 2020-10-15 Network rumor recognition method and system

Country Status (1)

Country Link
CN (1) CN112231562B (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112765352A (en) * 2021-01-21 2021-05-07 东北大学秦皇岛分校 Graph convolution neural network text classification method based on self-attention mechanism
CN112925907A (en) * 2021-02-05 2021-06-08 昆明理工大学 Microblog comment viewpoint object classification method based on event graph convolutional neural network
CN113139052A (en) * 2021-04-14 2021-07-20 三峡大学 Rumor detection method and device based on graph neural network feature aggregation
CN113191144A (en) * 2021-03-19 2021-07-30 北京工商大学 Network rumor recognition system and method based on propagation influence
CN113268675A (en) * 2021-05-19 2021-08-17 湖南大学 Social media rumor detection method and system based on graph attention network
CN113469261A (en) * 2021-07-12 2021-10-01 上海交通大学 Source identification method and system based on infection map convolution network
CN113515591A (en) * 2021-04-22 2021-10-19 平安科技（深圳）有限公司 Text bad information identification method and device, electronic equipment and storage medium
CN113742604A (en) * 2021-08-24 2021-12-03 三峡大学 Rumor detection method and device, electronic equipment and storage medium
CN113919440A (en) * 2021-10-22 2022-01-11 重庆理工大学 Social network rumor detection system integrating dual attention mechanism and graph convolution
CN113946680A (en) * 2021-10-20 2022-01-18 河南师范大学 Online network rumor identification method based on graph embedding and information flow analysis
CN114444516A (en) * 2022-04-08 2022-05-06 四川大学 Cantonese rumor detection method based on deep semantic perception map convolutional network
CN117743581A (en) * 2023-12-26 2024-03-22 中国农业科学院农业信息研究所 Intervention method for agricultural product quality safety network rumors

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2008165599A (en) * 2006-12-28 2008-07-17 National Institute Of Information & Communication Technology Rumor information extraction device and rumor information extraction method
CN105354305A (en) * 2015-11-05 2016-02-24 北京邮电大学 Online-rumor identification method and apparatus
CN107797998A (en) * 2016-08-29 2018-03-13 腾讯科技（深圳）有限公司 The recognition methods of user-generated content containing rumour and device
CN108563686A (en) * 2018-03-14 2018-09-21 中国科学院自动化研究所 Social networks rumour recognition methods based on hybrid neural networks and system
WO2018237098A1 (en) * 2017-06-20 2018-12-27 Graphika, Inc. Methods and systems for identifying markers of coordinated activity in social media movements
US20180373751A1 (en) * 2017-06-21 2018-12-27 Beijing Baidu Netcom Science And Technology Co., Ltd. Method and apparatus for recognizing a low-quality news resource, computer device and readable medium
CN110362818A (en) * 2019-06-06 2019-10-22 中国科学院信息工程研究所 Microblogging rumour detection method and system based on customer relationship structure feature
CN110688484A (en) * 2019-09-24 2020-01-14 北京工商大学 Microblog sensitive event speech detection method based on unbalanced Bayesian classification
CN110795641A (en) * 2019-11-05 2020-02-14 重庆邮电大学 Network rumor propagation control method based on representation learning
CN110825948A (en) * 2019-11-05 2020-02-21 重庆邮电大学 Rumor propagation control method based on rumor-splitting message and representation learning
CN111144131A (en) * 2019-12-25 2020-05-12 北京中科研究院 Network rumor detection method based on pre-training language model
CN111159395A (en) * 2019-11-22 2020-05-15 国家计算机网络与信息安全管理中心 Chart neural network-based rumor standpoint detection method and device and electronic equipment
WO2020140633A1 (en) * 2019-01-04 2020-07-09 平安科技（深圳）有限公司 Text topic extraction method, apparatus, electronic device, and storage medium
CN111400452A (en) * 2020-03-16 2020-07-10 腾讯科技（深圳）有限公司 Text information classification processing method, electronic device and computer readable storage medium

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2008165599A (en) * 2006-12-28 2008-07-17 National Institute Of Information & Communication Technology Rumor information extraction device and rumor information extraction method
CN105354305A (en) * 2015-11-05 2016-02-24 北京邮电大学 Online-rumor identification method and apparatus
CN107797998A (en) * 2016-08-29 2018-03-13 腾讯科技（深圳）有限公司 The recognition methods of user-generated content containing rumour and device
WO2018237098A1 (en) * 2017-06-20 2018-12-27 Graphika, Inc. Methods and systems for identifying markers of coordinated activity in social media movements
US20180373751A1 (en) * 2017-06-21 2018-12-27 Beijing Baidu Netcom Science And Technology Co., Ltd. Method and apparatus for recognizing a low-quality news resource, computer device and readable medium
CN108563686A (en) * 2018-03-14 2018-09-21 中国科学院自动化研究所 Social networks rumour recognition methods based on hybrid neural networks and system
WO2020140633A1 (en) * 2019-01-04 2020-07-09 平安科技（深圳）有限公司 Text topic extraction method, apparatus, electronic device, and storage medium
CN110362818A (en) * 2019-06-06 2019-10-22 中国科学院信息工程研究所 Microblogging rumour detection method and system based on customer relationship structure feature
CN110688484A (en) * 2019-09-24 2020-01-14 北京工商大学 Microblog sensitive event speech detection method based on unbalanced Bayesian classification
CN110795641A (en) * 2019-11-05 2020-02-14 重庆邮电大学 Network rumor propagation control method based on representation learning
CN110825948A (en) * 2019-11-05 2020-02-21 重庆邮电大学 Rumor propagation control method based on rumor-splitting message and representation learning
CN111159395A (en) * 2019-11-22 2020-05-15 国家计算机网络与信息安全管理中心 Chart neural network-based rumor standpoint detection method and device and electronic equipment
CN111144131A (en) * 2019-12-25 2020-05-12 北京中科研究院 Network rumor detection method based on pre-training language model
CN111400452A (en) * 2020-03-16 2020-07-10 腾讯科技（深圳）有限公司 Text information classification processing method, electronic device and computer readable storage medium

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
夏鑫林;许亮;: "基于注意力机制的谣言检测算法研究", 现代计算机, no. 08 *
段大高,韩忠明: "基于梯度提升决策树的微博虚假消息检测", 计算机应用 *

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112765352A (en) * 2021-01-21 2021-05-07 东北大学秦皇岛分校 Graph convolution neural network text classification method based on self-attention mechanism
CN112925907A (en) * 2021-02-05 2021-06-08 昆明理工大学 Microblog comment viewpoint object classification method based on event graph convolutional neural network
CN113191144A (en) * 2021-03-19 2021-07-30 北京工商大学 Network rumor recognition system and method based on propagation influence
CN113191144B (en) * 2021-03-19 2024-04-09 北京工商大学 Network rumor recognition system and method based on propagation influence
CN113139052B (en) * 2021-04-14 2022-04-19 三峡大学 Rumor detection method and device based on graph neural network feature aggregation
CN113139052A (en) * 2021-04-14 2021-07-20 三峡大学 Rumor detection method and device based on graph neural network feature aggregation
CN113515591A (en) * 2021-04-22 2021-10-19 平安科技（深圳）有限公司 Text bad information identification method and device, electronic equipment and storage medium
CN113515591B (en) * 2021-04-22 2024-03-15 平安科技（深圳）有限公司 Text defect information identification method and device, electronic equipment and storage medium
CN113268675A (en) * 2021-05-19 2021-08-17 湖南大学 Social media rumor detection method and system based on graph attention network
CN113268675B (en) * 2021-05-19 2022-07-08 湖南大学 Social media rumor detection method and system based on graph attention network
CN113469261A (en) * 2021-07-12 2021-10-01 上海交通大学 Source identification method and system based on infection map convolution network
CN113469261B (en) * 2021-07-12 2022-07-15 上海交通大学 Source identification method and system based on infection map convolution network
CN113742604A (en) * 2021-08-24 2021-12-03 三峡大学 Rumor detection method and device, electronic equipment and storage medium
CN113742604B (en) * 2021-08-24 2024-04-16 三峡大学 Rumor detection method and device, electronic equipment and storage medium
CN113946680A (en) * 2021-10-20 2022-01-18 河南师范大学 Online network rumor identification method based on graph embedding and information flow analysis
CN113946680B (en) * 2021-10-20 2024-04-16 河南师范大学 Online network rumor identification method based on graph embedding and information flow analysis
CN113919440A (en) * 2021-10-22 2022-01-11 重庆理工大学 Social network rumor detection system integrating dual attention mechanism and graph convolution
CN114444516A (en) * 2022-04-08 2022-05-06 四川大学 Cantonese rumor detection method based on deep semantic perception map convolutional network
CN117743581A (en) * 2023-12-26 2024-03-22 中国农业科学院农业信息研究所 Intervention method for agricultural product quality safety network rumors
CN117743581B (en) * 2023-12-26 2024-06-11 中国农业科学院农业信息研究所 Intervention method for agricultural product quality safety network rumors

Also Published As

Publication number Publication date
CN112231562B (en) 2023-07-14

Similar Documents

Publication Publication Date Title
CN112231562A (en) Network rumor identification method and system
CN111125358B (en) Text classification method based on hypergraph
CN112084335B (en) Social media user account classification method based on information fusion
CN108737406B (en) Method and system for detecting abnormal flow data
CN112199608B (en) Social media rumor detection method based on network information propagation graph modeling
CN111767725B (en) Data processing method and device based on emotion polarity analysis model
CN109740655B (en) Article scoring prediction method based on matrix decomposition and neural collaborative filtering
CN111866004B (en) Security assessment method, apparatus, computer system, and medium
CN113268675B (en) Social media rumor detection method and system based on graph attention network
CN109325125B (en) Social network rumor detection method based on CNN optimization
CN113254652B (en) Social media posting authenticity detection method based on hypergraph attention network
CN112465226B (en) User behavior prediction method based on feature interaction and graph neural network
CN111641608A (en) Abnormal user identification method and device, electronic equipment and storage medium
CN113315789A (en) Web attack detection method and system based on multi-level combined network
CN112217787A (en) Method and system for generating mock domain name training data based on ED-GAN
Abinaya et al. Spam detection on social media platforms
CN113919440A (en) Social network rumor detection system integrating dual attention mechanism and graph convolution
CN113220964A (en) Opinion mining method based on short text in network communication field
CN116956289A (en) Method for dynamically adjusting potential blacklist and blacklist
CN113191144B (en) Network rumor recognition system and method based on propagation influence
CN113609294B (en) Fresh cold chain supervision method and system based on emotion analysis
CN114529762A (en) Social network abnormal user detection method based on DS evidence theory fusion
CN113157993A (en) Network water army behavior early warning model based on time sequence graph polarization analysis
CN113094504A (en) Self-adaptive text classification method and device based on automatic machine learning
CN113537272A (en) Semi-supervised social network abnormal account detection method based on deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant