CN116579337B - False news detection method integrating evidence credibility - Google Patents
False news detection method integrating evidence credibility Download PDFInfo
- Publication number
- CN116579337B CN116579337B CN202310825631.6A CN202310825631A CN116579337B CN 116579337 B CN116579337 B CN 116579337B CN 202310825631 A CN202310825631 A CN 202310825631A CN 116579337 B CN116579337 B CN 116579337B
- Authority
- CN
- China
- Prior art keywords
- evidence
- news
- false
- features
- preliminary
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 40
- 238000013528 artificial neural network Methods 0.000 claims abstract description 22
- 238000012795 verification Methods 0.000 claims abstract description 20
- 230000009193 crawling Effects 0.000 claims abstract description 12
- 239000011159 matrix material Substances 0.000 claims description 27
- 230000006870 function Effects 0.000 claims description 18
- 238000010586 diagram Methods 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000009466 transformation Effects 0.000 claims description 8
- 238000002790 cross-validation Methods 0.000 claims description 7
- 230000003993 interaction Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 238000004140 cleaning Methods 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims 6
- 238000000034 method Methods 0.000 abstract description 7
- 238000012360 testing method Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 239000013598 vector Substances 0.000 description 2
- 241000237519 Bivalvia Species 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 235000020639 clam Nutrition 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- 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
- G06F16/951—Indexing; Web crawling techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention relates to the technical field of false news detection, and provides a false news detection method integrating evidence credibility. The method comprises the following steps: obtaining news to be tested, and crawling to obtain the true evidence of the news to be tested; generating false evidence corresponding to the news to be tested through an artificial intelligence program, and constructing a data set according to the news to be tested, the true evidence and the false evidence; learning the data set through a neural network to obtain a preliminary feature representation of the data set, wherein the preliminary feature representation comprises news features and preliminary evidence features; performing iterative cross verification on the preliminary evidence features, and introducing a multi-head input layer to obtain multi-head credibility scores corresponding to each evidence in the data set; modulating the attention weight between the news feature and the preliminary evidence feature according to the multi-head credibility score to obtain an integral evidence feature; and carrying out convolution learning on the integral evidence characteristics to detect the true and false of the news to be detected. According to the method, the evidence scene which is less trusted is simulated, different credibility is simulated through multi-head credibility, and the accuracy of false news detection is improved.
Description
Technical Field
The invention relates to the technical field of false news detection, in particular to a false news detection method integrating evidence credibility.
Background
Many studies are made in the false news detection direction based on evidence nowadays, and many models obtain good effects on common data sets such as Snopes and politifect, wherein news and labels are obtained on a fact checking website, evidence is obtained by searching news in a search engine, and one piece of news corresponds to a plurality of pieces of evidence. The existing work takes news and evidence as input, encodes the input through a gating circulation unit (GRU), a long-term short-term memory network (LSTM) or a Gating Graph Neural Network (GGNN) and the like, establishes the connection between the news and the evidence through an attention mechanism to obtain the characteristic representation of the evidence, finally obtains the probability that the news is true or false through a full connection layer and a softmax function together by the characteristics of the news and the evidence, and selects the probability with the maximum probability as a prediction result.
First, the snops and polifact extract the most similar segments of crawling evidence when data cleaning is performed after evidence acquisition, and the similarity between word embedding of the segments and claims is at least 0.5. This data processing makes the data sets too clean and there is only 1 evidence of much news in the data sets, which is not realistic because there is not only one search result for a piece of news, which cannot reflect the reality, possibly resulting in significant performance degradation of the models that perform well on these data sets in real world applications.
The existing false news detection method based on evidence models the relation between the clams and the evidence through methods such as an attention mechanism and the like, improves the model performance, but ignores a critical problem, namely whether the evidence used for false news detection is reliable or not, the evidence used for false news detection is retrieved by a search engine, and can be thought that some unreliable evidence exists in the evidence, and even some false evidence with model aggressiveness scattered by malicious attackers interfere with the prediction result of the false news detection model, so that the existing model uses the false evidence as the basis of news detection due to lack of judgment on whether the evidence is reliable or not, thereby giving false judgment to the false of news.
Still other studies include decare, MAC and GET using sources of evidence as a complement to the textual features of evidence, but model understanding of the degree of evidence trustworthiness is far from relying on sources alone, and new sources are layered, once evidence sources made by an attacker are not seen by the model, the model cannot judge whether evidence is trustworthy at all.
Disclosure of Invention
The present invention is directed to solving at least one of the technical problems existing in the related art. Therefore, the invention provides a false news detection method for fusing evidence credibility.
The invention provides a false news detection method integrating evidence credibility, which comprises the following steps:
s100: obtaining news to be detected, and crawling the true evidence of the news to be detected;
s200: generating false evidence corresponding to the news to be tested through an artificial intelligence program, and constructing a data set according to the news to be tested, the true evidence and the false evidence;
s300: learning the data set through a neural network to obtain a preliminary feature representation of the data set, wherein the preliminary feature representation comprises news features and preliminary evidence features;
s400: performing iterative cross-validation on the preliminary evidence features and introducing a multi-head output layer to obtain multi-head credibility scores corresponding to each piece of true evidence and false evidence in the data set;
s500: modulating the attention weight between the news feature and the preliminary evidence feature according to the multi-head credibility score to obtain the integral evidence feature of the data set;
s600: and performing convolution learning on the integral evidence characteristics to obtain true and false detection results of the news to be detected.
According to the false news detection method for fusing evidence credibility provided by the invention, the step S100 comprises the following steps:
s110: acquiring news to be tested based on the existing database;
s120: crawling the news text to be tested again through a crawler tool;
s130: locating the position of the news related keywords in the news text;
s140: and intercepting the evidence text within a preset length according to the position, and cleaning the evidence text to obtain the true evidence of the news to be tested.
According to the false news detection method for fusing evidence credibility provided by the invention, the step of obtaining the news characteristic in the step S300 comprises the following steps:
s311: setting a sliding window for the news to be tested;
s312: identifying words through the sliding window as nodes, and establishing a news text graph corresponding to the news to be tested;
s313: establishing an adjacency matrix according to the news text graph, carrying out Laplace standardization, and obtaining news text word characteristics through a graph neural network layer;
s314: and taking the average value of all the news text word characteristics to obtain the news characteristics.
According to the false news detection method for fusing evidence credibility provided by the invention, the step of obtaining the preliminary evidence features in the step S300 comprises the following steps:
s321: setting a sliding window for the true evidence and the false evidence;
s322: identifying words through the sliding window as nodes, and establishing a pre-evidence text diagram corresponding to the true evidence and the false evidence;
s323: establishing a pre-adjacency matrix according to the pre-evidence text graph, carrying out Laplacian standardization, and obtaining evidence text word characteristics through a graph neural network layer;
s324: reducing the dimensions of all the evidence text word features through a full-connection layer, and respectively calculating the reduced dimensions of the evidence text word features through a graph neural network layer to obtain redundancy scores corresponding to each evidence text word feature;
s325: and carrying out edge removing operation on the pre-evidence text graph according to the redundancy score, obtaining the evidence text graph, establishing an adjacency matrix, and obtaining the preliminary evidence features through a graph neural network layer.
According to the false news detection method for fusing evidence credibility provided by the invention, the step S400 comprises the following steps:
s410: respectively calculating and obtaining interaction results of the current preliminary evidence features and other preliminary evidence features according to the scaling dot product attention scoring function and the trainable parameters;
s420: taking the average value of each interaction result to obtain an average value result;
s430: performing nonlinear transformation on the mean value result to obtain a verification score of the current preliminary evidence feature;
s440: step S410 to step S430 are respectively carried out on each preliminary evidence feature, and verification scores corresponding to each preliminary evidence feature are obtained;
s450: and carrying out multiple rounds of operations from step S410 to step S440 on each preliminary evidence feature, introducing multiple attention to the obtained corresponding verification score, and obtaining multiple credibility scores corresponding to each piece of true evidence and false evidence in the data set.
According to the false news detection method for fusing evidence credibility provided by the invention, the step S450 comprises the following steps:
s451: performing full-connection layer calculation mapping on verification scores corresponding to the initial evidence features to a two-dimensional space;
s452: and carrying out nonlinear transformation of an exponential function and an activation function on the verification score corresponding to each preliminary evidence feature in the two-dimensional space to obtain a multi-head credibility score corresponding to each evidence in the data set.
According to the false news detection method for fusing evidence credibility provided by the invention, the step S500 comprises the following steps:
s510: performing multi-head attention calculation on the news features and the preliminary evidence features to obtain an attention weight matrix;
s520: element multiplication is carried out on the attention weight matrix and the multi-head credibility score corresponding to each evidence, so that a credibility modulation attention weight matrix is obtained;
s530: and carrying out matrix multiplication on the credibility modulation attention weight matrix and the preliminary evidence features to obtain the integral evidence features.
In the false news detection method integrating the evidence credibility, in the false news detection task based on the evidence, the credibility score of the retrieval evidence is modeled under the condition of no credibility label, a new data set is constructed to simulate a less credible evidence scene, a new method framework is provided, different credibility of different evidences is definitely simulated through the calculation of the credibility of multiple evidence, and the accuracy of the false news detection based on the evidence is improved through training and testing on the data set.
Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the invention, and other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a false news detection method for fusing evidence credibility, which is provided by the embodiment of the invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention. The following examples are illustrative of the invention but are not intended to limit the scope of the invention.
In the description of the embodiments of the present invention, it should be noted that the terms "center", "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, are merely for convenience in describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the apparatus or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and thus should not be construed as limiting the embodiments of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
In describing embodiments of the present invention, it should be noted that, unless explicitly stated and limited otherwise, the terms "coupled," "coupled," and "connected" should be construed broadly, and may be either a fixed connection, a removable connection, or an integral connection, for example; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium. The specific meaning of the above terms in embodiments of the present invention will be understood in detail by those of ordinary skill in the art.
In embodiments of the invention, unless expressly specified and limited otherwise, a first feature "up" or "down" on a second feature may be that the first and second features are in direct contact, or that the first and second features are in indirect contact via an intervening medium. Moreover, a first feature being "above," "over" and "on" a second feature may be a first feature being directly above or obliquely above the second feature, or simply indicating that the first feature is level higher than the second feature. The first feature being "under", "below" and "beneath" the second feature may be the first feature being directly under or obliquely below the second feature, or simply indicating that the first feature is less level than the second feature.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the embodiments of the present invention. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, the different embodiments or examples described in this specification and the features of the different embodiments or examples may be combined and combined by those skilled in the art without contradiction.
An embodiment provided by the present invention is described below with reference to fig. 1.
The invention provides a false news detection method integrating evidence credibility, which comprises the following steps:
s100: obtaining news to be detected, and crawling to obtain the true evidence of the news to be detected;
wherein, step S100 includes:
s110: acquiring news to be tested based on the existing database;
s120: crawling the news text to be tested again through a crawler tool;
s130: locating the position of the news related keywords in the news text;
s140: and intercepting the evidence text within a preset length according to the position, and cleaning the evidence text to obtain the true evidence of the news to be tested.
In some embodiments, based on the website links of news and evidence thereof in the existing snoes dataset, the evidence is re-crawled using a Python crawler tool; and positioning the position of the keyword in the news for the crawled text, intercepting the text according to the position, and controlling the text length to be within 100 to obtain a crawled evidence collection after simple cleaning.
S200: generating false evidence corresponding to the news to be tested through an artificial intelligence program, and constructing a data set according to the news to be tested, the true evidence and the false evidence;
in some embodiments, 5 pieces of false evidence are generated for each piece of news by constructing a Prompt and using ChatGPT, the generated false evidence is added into a news evidence set with the crawling evidence data number being more than 5, and the total evidence number is controlled to be not more than 30, so that a final data set snoese is obtained.
Further, in the step of generating the false evidence, because of the strong text generating function of the ChatGPT, the constructed Prompt is input into the "gpt-3.5-turbo" API of the ChatGPT to generate 5 false evidence, and then the generated evidence is combined with the previous crawling evidence, and the source of the generated evidence is randomly selected from the original evidence sources, so that a possibly-occurring malicious evidence attack scene can be simulated. For the combined evidence set, the number of the evidence sets needs to be controlled to be not more than 30, so that the generated evidence of the part exceeding 30 pieces needs to be removed by judging whether the number of the evidence sets is more than 30, otherwise, the evidence sets are not processed, then the evidence crawling step is continued if the next news exists, and the evidence crawling step is not ended.
S300: learning the data set through a neural network to obtain a preliminary feature representation of the data set, wherein the preliminary feature representation comprises news features and preliminary evidence features;
the step of obtaining the news feature in step S300 includes:
s311: setting a sliding window for the news to be tested;
s312: identifying words through the sliding window as nodes, and establishing a news text graph corresponding to the news to be tested;
s313: establishing an adjacency matrix according to the news text graph, carrying out Laplace standardization, and obtaining news text word characteristics through a graph neural network layer;
s314: and taking the average value of all the news text word characteristics to obtain the news characteristics.
In some embodiments, the use of the gatekeeper neural network to learn feature representations of news and evidence continues in the news and evidence encoding step due to its effectiveness in evidence redundancy word removal and feature encoding in the GET method.
Further, in the news feature representation coding, a sliding window size is set to be 3 for each word in news, each word is a node, edges are established among all words in the sliding window, the same word nodes are combined to form a node, a text diagram of the news is obtained, then an adjacency matrix is established according to the text diagram, laplace standardization is carried out, features of each word of the news are obtained through a standard gating diagram neural network layer, initial word node features are Glove pretraining word embedding vectors, each feature dimension is 300, and finally average values of all word features are obtained to obtain the news feature representation.
Wherein, the step of obtaining the preliminary evidence feature in step S300 includes:
s321: setting a sliding window for the true evidence and the false evidence;
s322: identifying words through the sliding window as nodes, and establishing a pre-evidence text diagram corresponding to the true evidence and the false evidence;
s323: establishing a pre-adjacency matrix according to the pre-evidence text graph, carrying out Laplacian standardization, and obtaining evidence text word characteristics through a graph neural network layer;
s324: reducing the dimensions of all the evidence text word features through a full-connection layer, and respectively calculating the reduced dimensions of the evidence text word features through a graph neural network layer to obtain redundancy scores corresponding to each evidence text word feature;
s325: and carrying out edge removing operation on the pre-evidence text graph according to the redundancy score, obtaining the evidence text graph, establishing an adjacency matrix, and obtaining the preliminary evidence features through a graph neural network layer.
In some embodiments, for each evidence, a sliding window is set to be 3 for each word, each word is a node, edges are established among all words in the sliding window, the same word nodes are combined to form a node, a text diagram of the evidence is obtained, then an adjacency matrix is established according to the text diagram and Laplacian standardization is carried out, preliminary characteristics of each word of the evidence are obtained through a gating diagram neural network layer, wherein the node characteristics of the initial word are Glove pretrained word embedding vectors, each characteristic dimension is 300, the preliminary characteristics of each word are mapped from 300 to 1 through a full connection layer, then the dimension of each word is calculated through a gating diagram neural network layer, redundancy score of each word is obtained, then the words 40% higher than the redundancy score are obtained according to the number of words in the evidence, edges of each word are removed from the text diagram, a new evidence text diagram and corresponding adjacency are obtained, finally the preliminary characteristics of each word are taken as initial node characteristics, and finally the characteristic matrix representing all words in the final evidence is obtained through calculation of the gating diagram neural network layer.
S400: performing iterative cross-validation on the preliminary evidence features and introducing a multi-head output layer to obtain multi-head credibility scores corresponding to each piece of true evidence and false evidence in the data set;
in some embodiments, to address the problem of lack of evidence credibility labels, we assume that most of the retrieved evidence is credible, since it is almost impossible to use false evidence to occupy most of the top-ranked positions on the internet, so the present invention chooses to cross-verify multiple pieces of evidence for a piece of news, which enables them to evaluate each other's credibility, and in view that only one cross-validation may be unreliable, it is necessary to iterate this process to ensure that the output credibility score has stabilized.
Wherein, step S400 includes:
s410: respectively calculating and obtaining interaction results of the current preliminary evidence features and other preliminary evidence features according to the scaling dot product attention scoring function and the trainable parameters;
s420: taking the average value of each interaction result to obtain an average value result;
s430: performing nonlinear transformation on the mean value result to obtain a verification score of the current preliminary evidence feature;
s440: step S410 to step S430 are respectively carried out on each preliminary evidence feature, and verification scores corresponding to each preliminary evidence feature are obtained;
s450: and (3) carrying out multiple rounds of operations from step S410 to step S440 on each preliminary evidence feature, introducing multi-head attention to the obtained verification score, and obtaining multi-head credibility scores corresponding to each piece of true evidence and false evidence in the data set.
In some embodiments, in the iterative cross-validation step, firstly, an id input embedding layer of an evidence source is calculated to obtain 128-dimensional source characteristics, the 128-dimensional source characteristics and the preliminary evidence characteristics are connected to obtain 1628-dimensional evidence characteristic representations for a subsequent step, a plurality of evidence characteristic representations are iteratively validated with each other, the first round of cross-validation is based on the current evidence characteristic and other evidence characteristics, a trainable parameter is multiplied by a scoring function of the attention of a scaling dot product, an interaction result is obtained through calculation, and is used as a validation score of other evidence on the current evidence, and the other evidence also carries out the step; the interactive operation is carried out on the current evidence feature and all other evidences to obtain verification scores of the current evidence feature, the scores are averaged and subjected to nonlinear transformation of a full-connection layer, an exponential function and a tanh activation function to obtain an overall verification score of the current evidence, and other evidences also carry out the step; and performing the same operation on each iteration, wherein the operation is still based on the evidence characteristics, but the operation is performed on the product of the total verification score of the iteration of the previous iteration of other evidences and the original characteristic representation thereof, the total iteration number is 6, and the total verification score of each evidence after the final iteration of the I-th round is the preliminary result reflecting the credibility of each evidence.
Wherein, step S450 includes:
s451: performing full-connection layer calculation mapping on verification scores corresponding to the initial evidence features to a two-dimensional space;
s452: and carrying out nonlinear transformation of an exponential function and an activation function on the verification score corresponding to each preliminary evidence feature in the two-dimensional space to obtain a multi-head credibility score corresponding to each evidence in the data set.
In some embodiments, a full-join layer calculation is performed on the score to map it to 2 dimensions, as is other evidence; and obtaining the multi-head reliability score with the head number of 2 through nonlinear transformation of an exponential function and a tanh activation function, and performing the step by other evidences.
S500: modulating the attention weight between the news feature and the preliminary evidence feature according to the multi-head credibility score to obtain the integral evidence feature of the data set;
wherein, step S500 includes:
s510: performing multi-head attention calculation on the news features and the preliminary evidence features to obtain an attention weight matrix;
s520: element multiplication is carried out on the attention weight matrix and the multi-head credibility score corresponding to each evidence, so that a credibility modulation attention weight matrix is obtained;
s530: and carrying out matrix multiplication on the credibility modulation attention weight matrix and the preliminary evidence features to obtain the integral evidence features.
S600: and carrying out convolution learning on the integral evidence features and the news features to obtain true and false detection results of the news to be detected.
In some embodiments, in the news authenticity prediction step, based on the news features and all evidence integral features obtained before, the news features and all evidence integral features are subjected to concat connection operation and then enter a full-connection layer and a softmax function to calculate to obtain probability of true or false news, labels with high probability are used as predicted values, a data set snoese is divided into five-fold cross-validation data sets, cross entropy is used as a loss function to train parameters of the architecture model, and finally test results are obtained in a test set.
According to the false news detection method fusing the evidence credibility, the acquired evidence of news in the data set is sufficient, objective rules can be met and real conditions can be reflected, in addition, the false evidence is introduced after the evidence is crawled, a new data set is built for detection of false news, a more real and less credible evidence scene is simulated, after a new method frame is built, different credibility of different evidences is simulated through multiple evidence credibility, influences of the false evidence on news detection are considered, and accuracy of false news detection based on evidence is improved.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (7)
1. A false news detection method integrating evidence credibility is characterized by comprising the following steps:
s100: obtaining news to be detected, and crawling the true evidence of the news to be detected;
s200: generating false evidence corresponding to the news to be tested through an artificial intelligence program, and constructing a data set according to the news to be tested, the true evidence and the false evidence;
s300: learning the data set through a neural network to obtain a preliminary feature representation of the data set, wherein the preliminary feature representation comprises news features and preliminary evidence features;
s400: performing iterative cross-validation on the preliminary evidence features and introducing a multi-head output layer to obtain multi-head credibility scores corresponding to each piece of true evidence and false evidence in the data set;
s500: modulating the attention weight between the news feature and the preliminary evidence feature according to the multi-head credibility score to obtain the integral evidence feature of the data set;
s600: and carrying out convolution learning on the integral evidence features and the news features to obtain true and false detection results of the news to be detected.
2. The false news detection method with evidence reliability fusion according to claim 1, wherein step S100 includes:
s110: acquiring news to be tested based on the existing database;
s120: crawling the news text to be tested again through a crawler tool;
s130: positioning the position of a keyword related to news in the news text;
s140: and intercepting the evidence text within a preset length according to the position, and cleaning the evidence text to obtain the true evidence of the news to be tested.
3. The false news detection method with evidence reliability fusion according to claim 1, wherein the step of obtaining the news feature in step S300 includes:
s311: setting a sliding window for the news to be tested;
s312: identifying words through the sliding window as nodes, and establishing a news text graph corresponding to the news to be tested;
s313: establishing an adjacency matrix according to the news text graph, carrying out Laplace standardization, and obtaining news text word characteristics through a graph neural network layer;
s314: and taking the average value of all the news text word characteristics to obtain the news characteristics.
4. The false news detection method with evidence reliability fusion according to claim 1, wherein the step of obtaining the preliminary evidence features in step S300 includes:
s321: setting a sliding window for the true evidence and the false evidence;
s322: identifying words through the sliding window as nodes, and establishing a pre-evidence text diagram corresponding to the true evidence and the false evidence;
s323: establishing a pre-adjacency matrix according to the pre-evidence text graph, carrying out Laplacian standardization, and obtaining evidence text word characteristics through a graph neural network layer;
s324: reducing the dimensions of all the evidence text word features through a full-connection layer, and respectively calculating the reduced dimensions of the evidence text word features through a graph neural network layer to obtain redundancy scores corresponding to each evidence text word feature;
s325: and carrying out edge removing operation on the pre-evidence text graph according to the redundancy score, obtaining the evidence text graph, establishing an adjacency matrix, and obtaining the preliminary evidence features through a graph neural network layer.
5. The false news detection method with evidence reliability fusion according to claim 1, wherein step S400 includes:
s410: respectively calculating and obtaining interaction results of the current preliminary evidence features and other preliminary evidence features according to the scaling dot product attention scoring function and the trainable parameters;
s420: taking the average value of each interaction result to obtain an average value result;
s430: performing nonlinear transformation on the mean value result to obtain a verification score of the current preliminary evidence feature;
s440: step S410 to step S430 are respectively carried out on each preliminary evidence feature, and verification scores corresponding to each preliminary evidence feature are obtained;
s450: and (3) carrying out multiple rounds of operations from step S410 to step S440 on each preliminary evidence feature, introducing multi-head attention to the obtained verification score, and obtaining multi-head credibility scores corresponding to each piece of true evidence and false evidence in the data set.
6. The false news detection method with evidence reliability fusion according to claim 5, wherein step S450 includes:
s451: performing full-connection layer calculation mapping on verification scores corresponding to the initial evidence features to a two-dimensional space;
s452: and carrying out nonlinear transformation of an exponential function and an activation function on the verification score corresponding to each preliminary evidence feature in the two-dimensional space to obtain a multi-head credibility score corresponding to each evidence in the data set.
7. The false news detection method with evidence reliability fusion according to claim 1, wherein step S500 includes:
s510: performing multi-head attention calculation on the news features and the preliminary evidence features to obtain an attention weight matrix;
s520: element multiplication is carried out on the attention weight matrix and the multi-head credibility score corresponding to each evidence, so that a credibility modulation attention weight matrix is obtained;
s530: and carrying out matrix multiplication on the credibility modulation attention weight matrix and the preliminary evidence features to obtain the integral evidence features.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310825631.6A CN116579337B (en) | 2023-07-07 | 2023-07-07 | False news detection method integrating evidence credibility |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310825631.6A CN116579337B (en) | 2023-07-07 | 2023-07-07 | False news detection method integrating evidence credibility |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116579337A CN116579337A (en) | 2023-08-11 |
| CN116579337B true CN116579337B (en) | 2023-10-10 |
Family
ID=87536103
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310825631.6A Active CN116579337B (en) | 2023-07-07 | 2023-07-07 | False news detection method integrating evidence credibility |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116579337B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781679A (en) * | 2019-10-15 | 2020-02-11 | 上海大学 | News event keyword mining method based on associated semantic chain network |
| CN111177554A (en) * | 2019-12-27 | 2020-05-19 | 西安交通大学 | False news identification system and method capable of explaining exploration based on generation of confrontation learning |
| CN112988959A (en) * | 2021-01-13 | 2021-06-18 | 西安交通大学 | False news interpretability detection system and method based on evidence inference network |
| CN114462420A (en) * | 2022-01-28 | 2022-05-10 | 河海大学 | False news detection method based on feature fusion model |
| CN114579878A (en) * | 2022-03-15 | 2022-06-03 | 中国科学院空天信息创新研究院 | Training method of false news discrimination model, false news discrimination method and device |
| CN115130613A (en) * | 2022-07-26 | 2022-09-30 | 西北工业大学 | False news identification model construction method, false news identification method and device |
| WO2023093574A1 (en) * | 2021-11-25 | 2023-06-01 | 北京邮电大学 | News event search method and system based on multi-level image-text semantic alignment model |
-
2023
- 2023-07-07 CN CN202310825631.6A patent/CN116579337B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781679A (en) * | 2019-10-15 | 2020-02-11 | 上海大学 | News event keyword mining method based on associated semantic chain network |
| CN111177554A (en) * | 2019-12-27 | 2020-05-19 | 西安交通大学 | False news identification system and method capable of explaining exploration based on generation of confrontation learning |
| CN112988959A (en) * | 2021-01-13 | 2021-06-18 | 西安交通大学 | False news interpretability detection system and method based on evidence inference network |
| WO2023093574A1 (en) * | 2021-11-25 | 2023-06-01 | 北京邮电大学 | News event search method and system based on multi-level image-text semantic alignment model |
| CN114462420A (en) * | 2022-01-28 | 2022-05-10 | 河海大学 | False news detection method based on feature fusion model |
| CN114579878A (en) * | 2022-03-15 | 2022-06-03 | 中国科学院空天信息创新研究院 | Training method of false news discrimination model, false news discrimination method and device |
| CN115130613A (en) * | 2022-07-26 | 2022-09-30 | 西北工业大学 | False news identification model construction method, false news identification method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116579337A (en) | 2023-08-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022068196A1 (en) | Cross-modal data processing method and device, storage medium, and electronic device | |
| Qi et al. | Attentive relational networks for mapping images to scene graphs | |
| CN110728209B (en) | Gesture recognition method and device, electronic equipment and storage medium | |
| CN109446927B (en) | Double-person interaction behavior identification method based on priori knowledge | |
| CN109413023B (en) | Training of machine recognition model, machine recognition method and device, and electronic equipment | |
| CN106060008B (en) | A kind of network intrusions method for detecting abnormality | |
| CN109902018B (en) | Method for acquiring test case of intelligent driving system | |
| CN110213244A (en) | A kind of network inbreak detection method based on space-time characteristic fusion | |
| CN110929840A (en) | Continuous learning neural network system using rolling window | |
| CN112217787B (en) | Method and system for generating mock domain name training data based on ED-GAN | |
| CN112528035A (en) | Knowledge graph reasoning method and device based on relational attention and computer equipment | |
| CN111241338B (en) | Depth feature fusion video copy detection method based on attention mechanism | |
| CN113628059A (en) | Associated user identification method and device based on multilayer graph attention network | |
| CN113343901A (en) | Human behavior identification method based on multi-scale attention-driven graph convolutional network | |
| CN115712740B (en) | Method and system for multi-modal implication enhanced image text retrieval | |
| CN113065601A (en) | Deep learning forest fire abnormity detection method based on genetic algorithm optimization | |
| CN113792594B (en) | Method and device for locating language fragments in video based on contrast learning | |
| EP4030355A1 (en) | Neural reasoning path retrieval for multi-hop text comprehension | |
| CN116579337B (en) | False news detection method integrating evidence credibility | |
| CN112163493A (en) | Video false face detection method and electronic device | |
| CN115277065B (en) | Anti-attack method and device in abnormal traffic detection of Internet of things | |
| CN116304341A (en) | Fraud discrimination method and system based on user network big data | |
| CN114881103A (en) | Countermeasure sample detection method and device based on universal disturbance sticker | |
| CN112487394A (en) | Method, system, device and medium for identifying graph reasoning verification code | |
| CN116155626B (en) | Complex network attack detection method based on cross-host abnormal behavior recognition |
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 |