False news detection method and device, electronic equipment and storage medium

Abstract

The application provides a false news detection method and device, electronic equipment and a storage medium. The method comprises the following steps: obtaining an event from news data based on a text feature corresponding to the news data, and determining whether the news data is first type news or second type news; when the news data is the first type news, determining whether the news data is false news based on a first event discriminator and a historical event corresponding to the news data; when the news data is the second type news, determining whether the news data is false news based on a second event discriminator; the second event discriminator comprises a false event detector and an event feature extractor; the false event detector is used for identifying the text feature to obtain a probability that the corresponding event is a false event; and the event feature extractor is used for classifying the news data based on the probability to determine whether the news data is false news. The application can solve the technical problem that false news cannot be quickly and efficiently identified in the prior art.

Description

Technical Field

[0001] This invention relates to the field of network information dissemination technology, specifically to a method, device, electronic device, and storage medium for detecting fake news. Background Technology

[0002] The widespread adoption of social networking services has led to a rapid expansion of their user base, resulting in a surge in the amount of information. Social networking platforms allow users to freely publish information. These massive user groups release a vast amount of information daily, but much of it is false or misleading. False information can spread rapidly through social networks; therefore, identifying fake news quickly and efficiently is a crucial and significant challenge. Summary of the Invention

[0003] In view of this, it is necessary to provide a method, device, electronic device and storage medium for detecting fake news, so as to solve the technical problem that the existing technology cannot quickly and efficiently identify fake news.

[0004] To achieve the above objectives, the present invention provides a method for detecting fake news, comprising:

[0005] Obtain the news data to be tested, extract features from the news data, and obtain the corresponding text features;

[0006] Based on the text features, events are obtained from the news data, and based on the events, the news data is determined to be either a first type of news or a second type of news;

[0007] If the news data is determined to be of the first type, the news data is input into the first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and to determine whether the news data is fake news;

[0008] If the news data is determined to be of the second type, the news data is input into the second event discriminator to determine whether the news data is fake news based on the second event discriminator;

[0009] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0010] Furthermore, the fake event detector is obtained by transfer learning of fake events based on adversarial generative networks.

[0011] Furthermore, the step of extracting features from the news data to obtain corresponding text features includes:

[0012] The news data is segmented and tagged with parts of speech to obtain tagged words;

[0013] The word vectors corresponding to the labeled words are learned based on the pre-trained word embedding model;

[0014] Based on the word vectors, the sentence vectors corresponding to the news data are reduced in dimensionality to obtain word embedding vectors;

[0015] The text features are obtained based on the word embedding vectors.

[0016] Further, obtaining the text features based on the word embedding vectors includes:

[0017] The word embedding vector is input into a convolutional filter to obtain the feature vector corresponding to each sentence in the news data;

[0018] The feature vector is subjected to max pooling to obtain the text features.

[0019] Further, the step of obtaining events from the news data based on the text features, and determining whether the news data is a first type of news or a second type of news based on the events, includes:

[0020] Search for keywords in the news data based on the text features;

[0021] Based on the keywords, retrieve similar news sets from the keyword news inverted index table;

[0022] Determine the cosine similarity of different news items in the similar news set, and based on the cosine similarity of different news items in the similar news set, cluster the news items in the similar news set to obtain the event set corresponding to the keyword;

[0023] Based on the event set, the news data will be determined to be either the first type of news or the second type of news;

[0024] The keyword news inverted index table includes multiple preset keywords and a news inverted table corresponding to each preset keyword.

[0025] Furthermore, the step of determining whether news data is a first type of news or a second type of news based on the event set includes:

[0026] Entropy filtering is performed on the event set to obtain the filtered events;

[0027] The filtered events are then filtered using the LCS algorithm, and the filtered events are determined to be either type 1 news or type 2 news based on the LCS algorithm filtering results.

[0028] Furthermore, the step of retrieving similar news sets from the keyword news inverted index table based on the keywords includes:

[0029] Determine the cosine similarity of the keyword to each news item in the keyword news inverted index table;

[0030] The similar news set is constructed based on news items whose cosine similarity is greater than a preset threshold.

[0031] The present invention also provides a fake news detection device, comprising:

[0032] The extraction module is used to acquire the news data to be tested, extract features from the news data, and obtain the corresponding text features.

[0033] The first judgment module is used to obtain events from the news data based on the text features, and to judge the news data as a first type of news or a second type of news based on the events;

[0034] The second judgment module is used to input the news data into the first event discriminator when it is determined that the news data is a first type of news, so as to retrieve the historical events corresponding to the news data based on the first event discriminator and determine whether the news data is fake news.

[0035] The third judgment module is used to input the news data into the second event discriminator when it is determined that the news data is the second type of news, so as to determine whether the news data is fake news based on the second event discriminator;

[0036] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0037] The present invention also provides an electronic device, including a memory and a processor, wherein,

[0038] The memory is used to store programs;

[0039] The processor, coupled to the memory, is configured to execute the program stored in the memory to implement the steps in the fake news detection method as described in any of the preceding claims.

[0040] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fake news detection method as described in any of the preceding claims.

[0041] The beneficial effects of the above implementation are as follows: The fake news detection method, device, electronic device, and storage medium provided by this invention obtain events from news data by using the text features corresponding to the news data to be tested, and determine whether the news data is a first type of news or a second type of news based on the events; if the news data is determined to be a first type of news, historical events corresponding to the news data are retrieved based on a first event discriminator to determine whether the news data is fake news; if the news data is determined to be a second type of news, a second event discriminator is used to determine whether the news data is fake news; wherein, the second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify text features to obtain the probability that the corresponding event is fake; the event feature extractor is used to classify the news data based on the probability to determine whether the news data is fake news. This invention divides news data into first type of news and second type of news, and inputs them into different event discriminators for fake news discrimination, achieving fast and efficient detection of fake news based on the event correlation between different news items. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating an embodiment of the fake news detection method provided by the present invention;

[0044] Figure 2 This is a schematic diagram of the keyword news inverted index provided by the present invention;

[0045] Figure 3 This is a schematic diagram of the inverted index of keyword event IDs provided by the present invention;

[0046] Figure 4 A schematic diagram of the structure of the fake news detection device provided by the present invention;

[0047] Figure 5 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] In the description of the embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0050] In this embodiment of the invention, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, apparatus, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product or device.

[0051] The naming or numbering of steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical order indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effect can be achieved.

[0052] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0053] This invention provides a method, apparatus, electronic device, and storage medium for detecting fake news, which will be described below.

[0054] like Figure 1 As shown, the present invention provides a method for detecting fake news, comprising:

[0055] Step 110: Obtain the news data to be tested, extract features from the news data, and obtain the corresponding text features.

[0056] It is understood that the method provided by this invention can be implemented based on the ECNNet architecture, which integrates word vector models, convolutional neural networks, generative adversarial networks and incremental clustering algorithms. It achieves fake news detection for sudden events on social networks through four modules: feature extractor, event mapper, familiar event judger and unknown event judger.

[0057] Define a time frame T, and use web crawlers to obtain sufficient raw news data from mainstream social networks such as Sina Weibo, Toutiao, and Tencent News. Filter and define the research scope S of the events. Input the news data into a feature extractor, which extracts information features from the text content of the news data. A convolutional neural network (CNN) can be used as the core module of the feature extractor.

[0058] Step 120: Obtain events from the news data based on the text features, and determine whether the news data is a first type of news or a second type of news based on the events.

[0059] Understandably, text features are input into the event mapper, which combines the text features extracted by the feature extractor to collect events from the news data and classify them into familiar news (i.e., type 1 news) and unknown news (i.e., type 2 news).

[0060] Step 130: If the news data is determined to be of the first type of news, the news data is input into the first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and to determine whether the news data is fake news.

[0061] Understandably, the first event discriminator, also known as the familiar event discriminator, receives news from the event mapper and predicts whether the news is true or false based on the relevant historical events.

[0062] For each news event identified as "familiar news," a corresponding set of historical events is searched, and then the event correlation between the familiar news and the retrieved event set is calculated. The definition of event correlation in this invention is as follows:

[0063] Suppose there exists a historical event 1 and an event 2. Event 2 consists of n news items, namely news1, news2, ..., news n The cosine similarities of these news items to historical event 1 are sim1, sim2, ..., sim1, sim2, ..., sim3, sim4, sim5, sim6, sim7, sim8, sim9, sim1, sim1, sim2, ..., sim1, sim2, sim3, sim4, sim5, sim6, sim7, sim8, sim9, sim1, sim9, sim1, n The correlation S between historical event 1 and event 2 is defined as follows:

[0064]

[0065] For each historical event in the retrieval set, a label of 1 can be used to indicate that the event is fake, and a label of 0 can be used to indicate that the event is real. The news authenticity R is defined as follows:

[0066]

[0067] Among them l i This represents the tag value of the i-th news item. If the final calculated news authenticity R is greater than 0, it indicates that the predicted familiar news is fake news; otherwise, it is considered that the familiar news is real.

[0068] Step 140: If the news data is determined to be the second type of news, the news data is input into the second event discriminator to determine whether the news data is fake news based on the second event discriminator;

[0069] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0070] Understandably, the second event discriminator is also the unknown event discriminator. This unknown event discriminator uses a generative adversarial model to perform transfer learning on fake news, and then uses this to determine whether unknown news is fake news. In addition, a convolutional neural network layer is integrated for event feature learning.

[0071] An event is typically a collection of news reports; that is, news reports about the same event are related to each other, and there is also a relationship between news reports and events. Therefore, generative adversarial networks (GANs) can be used to construct the unknown event determiner for ECNNet. The unknown event determiner consists of two modules: a fake event detector and an event feature extractor.

[0072] In some embodiments, the fake event detector is obtained by transfer learning of fake events based on an adversarial generative network.

[0073] Understandably, the fake event detector uses generative adversarial networks for transfer learning in fake events, and the event feature extractor uses a two-layer fully connected neural network to extract event features.

[0074] The fake event detector uses a generative adversarial network (GAN) for transfer learning in fake events. It uses the text feature representation R extracted by the feature extractor. T As input, output the probability that the i-th event is a false event, and represent the i-th event as m. i The fake news detector uses a fully connected layer with softmax to predict whether a news story is fake. This invention denotes the fake news detector as G. d (·;θ d ), where θd This represents all the parameters included. The text features R of the i-th news item... T As input, the output is denoted as P. θ (m i The calculation formula is as follows:

[0075] P θ (m i ) = G d (G f (m i ;θ f );θ d )

[0076] The goal of the event detector is to identify whether unknown news is fake news, using Y... d Let represent the manually labeled training set, and use cross-entropy to calculate the loss function, as shown in the following formula:

[0077]

[0078] In the following process, by finding the optimal parameter θ f and θ d To make the loss function L d (θ f θ d The process of reaching the minimum value can be represented as:

[0079]

[0080] However, fake event detectors can only learn event-specific features and cannot generalize, which is detrimental to detecting events not included in the training dataset. Therefore, it is necessary to learn transferable feature representations of newly emerging events. Consider models that can learn more general feature representations that can represent news items related to the same event. Therefore, to learn common features of events, consider adding an event feature extractor to improve the model.

[0081] Event feature extractor is essentially a two-way adversarial transfer learning process that makes full use of data features from both the source and target domains.

[0082] The event feature extractor is represented as G. e (R F ,θ e ), where represents all the parameters G contained therein. e It deploys a convolutional neural network layer to correctly classify the input domain, including news and events. Define Y... e Given a set of domain labels, the loss of the event discriminator can be expressed as:

[0083]

[0084] Since the discriminator seeks to identify the input domain, from the discriminator's perspective, the loss function should be minimized to find the optimal parameters:

[0085]

[0086] However, for an event feature extractor, the goal is to fool a fake event detector into learning common features of events. Increasing the loss function means it learns more common features, as a larger loss function indicates greater difficulty in correctly classifying events into clusters. Therefore, maximizing the aforementioned loss is necessary to find the optimal parameters.

[0087] Before the training phase of the unknown event determiner, all historical events are extracted as the training set. During the training phase, a generative adversarial network G is used. f (·;θ f To improve the performance of fake event detection, the event feature extractor collaborates with the fake event detector to minimize the loss L on the training dataset. d (θ f ,θ d On the other hand, there is a minimax game between the event feature extractor and the fake event detector. Therefore, the final loss can be expressed as:

[0088] L final (θ f ,θ d ,θ e ) = L d (θ f ,θ d ) = L e (θ f ,θ e )

[0089] To find the optimal parameters Need to minimize L final And considering minimax games. Therefore, the optimal parameters are related to the time it takes for the process to reach equilibrium, and they can be expressed as:

[0090]

[0091]

[0092] In some embodiments, the feature extraction of the news data to obtain corresponding text features includes:

[0093] The news data is segmented and tagged with parts of speech to obtain tagged words;

[0094] The word vectors corresponding to the labeled words are learned based on the pre-trained word embedding model;

[0095] Based on the word vectors, the sentence vectors corresponding to the news data are reduced in dimensionality to obtain word embedding vectors;

[0096] The text features are obtained based on the word embedding vectors.

[0097] Understandably, to extract rich textual features from news articles, the input news text (i.e., news data) undergoes word segmentation and part-of-speech tagging sequentially. Then, a pre-trained word embedding model is used to learn the word vector for each word. Furthermore, a PCA (Principal Component Analysis) model is used to reduce the dimensionality of the sentence vectors for each news article. The resulting ordered list of words (i.e., word embedding vectors) serves as the input to the text feature extractor.

[0098] To better extract relevant information features from news texts, a convolutional neural network (CNN) was used as the core module of the text feature extractor. A modified CNN model, Text-CNN, was incorporated into the text feature extractor. The Text-CNN architecture uses multiple filters with windows of different sizes to select text features at different granularities.

[0099] In some embodiments, obtaining the text features based on the word embedding vectors includes:

[0100] The word embedding vector is input into a convolutional filter to obtain the feature vector corresponding to each sentence in the news data;

[0101] The feature vector is subjected to max pooling to obtain the text features.

[0102] Understandably, each word in the news text is vectorized into a word embedding vector. The word embedding vector for each word or phrase is initialized using a pre-trained word embedding model on a given dataset. For the i-th word in a sentence, the corresponding k-dimensional word embedding vector is denoted as T. i ∈R k Therefore, a sentence with n words can be represented as:

[0103]

[0104] in, This represents the concatenation operator. A convolutional filter with a window size of h takes a continuous sequence of h words from the text as input and outputs a feature. To clearly illustrate the process, taking a continuous sequence of h words starting from the i-th word as an example, the filtering operation can be represented as:

[0105] t i =σ(W c ·T i:i+h-1 )

[0106] Here, σ(·) is a ReLU activation function, W c This represents the weights of the convolutional filter. This filter can also be applied to the remaining words, resulting in a feature vector for the sentence as shown below:

[0107] t = [t1, t2, ..., t n-h+1 ]

[0108] For each feature vector, max pooling is used to find its maximum value, thus extracting the most important information from the text. This yields the characteristics of a specific filter. This process is repeated until the characteristics of all filters are obtained. To extract text features with different granularities, different window sizes are applied. For a given window size, n... h A different filter.

[0109] Therefore, assuming there are c possible window sizes, there are a total of c·n h A filter. The text features after max pooling are denoted as... After the max pooling operation, a fully connected layer is used to ensure the final text feature representation (denoted as ). It has p-dimensional features:

[0110]

[0111] Among them, W tf This is the weight matrix of the fully connected layer.

[0112] In some embodiments, the step of obtaining events from the news data based on the text features and determining whether the news data is a first type of news or a second type of news based on the events includes:

[0113] Search for keywords in the news data based on the text features;

[0114] Based on the keywords, retrieve similar news sets from the keyword news inverted index table;

[0115] Determine the cosine similarity of different news items in the similar news set, and based on the cosine similarity of different news items in the similar news set, cluster the news items in the similar news set to obtain the event set corresponding to the keyword;

[0116] Based on the event set, the news data will be determined to be either the first type of news or the second type of news;

[0117] The keyword news inverted index table includes multiple preset keywords and a news inverted table corresponding to each preset keyword.

[0118] It is understood that the steps in this embodiment are based on an event mapper, which consists of three parts: a keyword-news mapper, a keyword-event mapper, and a filter.

[0119] A keyword-news mapper is used to quickly retrieve whether there are similar news items among existing news items for an input news item. Essentially, it is a dynamically updated inverted index table. To reduce the time required to search for similar news items encountered previously, while maintaining constant time and space requirements, a keyword-news inverted index table maintained within a time window t can be used, such as... Figure 2 As shown. The set M is continuously updated by replacing the oldest news with the most recent news to keep the memory requirements of the keyword-news inverted index constant. Due to the unlimited use of words in the news stream, the number of keywords can become very large. Each entry in the keyword-news inverted index contains a keyword and a finite set Q. This set Q is the latest news in which the keyword appears. However, when the number of news exceeds the limit of Q, the oldest news is replaced by the latest news containing the keyword.

[0120] The top k keywords are selected from the news articles based on the TF-IDF (term frequency–inverse document frequency) method. Then, the potential similar news sets are retrieved by calculating the cosine similarity between each of these k keywords and each corresponding news article in the keyword-news inverted index table.

[0121] For example, inputting the news article "A City Press Conference Announcement: From 16:00 to 22:00 on April 22, A City reported 4 new cases of dynamic C-type individuals (2 in Zone E and 2 in Zone F). There is also 1 case of dynamic Y-type individual." The first three TF-IDF weighted keywords are "A City," "Dynamic C-type," and "Dynamic Y-type." Each keyword is searched in the keyword-news inverted index table, and news articles with IDs 3, 5, 7, 15, 18, 21, and 25 are retrieved. Finally, cosine distance is used to calculate the similarity between the two news vectors. The formula for calculating cosine distance is as follows:

[0122]

[0123] If no news item has a cosine similarity higher than the TSI value (threshold), it means that no similar news has occurred before. Therefore, a new event is created, news item d is assigned to the event, and then sent to the keyword-event mapper. If a similarity exists in the TSI above, news item d will be sent directly to the keyword-event mapper. Finally, the keyword-news mapper is updated by adding the news item to the corresponding news set for the k terms.

[0124] The keyword-event mapper is used to detect events and categorize them into familiar and unknown events. Initially, the keyword-event mapper contains only a dynamically updated inverted index table of keyword-historical events, such as... Figure 3 As shown. Similar to the keyword-news mapper, each row has a keyword and a finite set where the number of historical events does not exceed Q. When the number exceeds the limit, the oldest event is replaced by the newest event.

[0125] In some embodiments, determining whether news data is a first type of news or a second type of news based on the event set includes:

[0126] Entropy filtering is performed on the event set to obtain the filtered events;

[0127] The filtered events are filtered using the LCS (Longest Common Subsequence) algorithm, and the filtered events are determined to be either the first type of news or the second type of news based on the LCS algorithm filtering results.

[0128] Understandably, the keyword-event mapper uses clustering to group news items into events. This clustering process generates smaller events, known as fragmented events. The existence of fragmented events can negatively impact the model's accuracy and speed.

[0129] To improve the model's accuracy and speed, filters are used to remove fragmented events and meaningless candidate events. The filters mainly consist of two parts: entropy filtering and LCS algorithm filtering.

[0130] Entropy filtering uses entropy information from candidate event clusters. By calculating the entropy of each candidate event cluster and comparing it with a pre-set entropy threshold (tent), if the entropy value of a candidate event cluster is less than the entropy threshold (tent), it is considered that the information content of the candidate event cluster has not reached the set minimum information content, and it is judged as a fragmented cluster and discarded.

[0131] The LCS (Local Sequence of Criteria) filtering algorithm takes into account that news items within an event often have similar sentence structures. For each candidate event, the LCS algorithm is applied, and the length of the maximum LCS is recorded. Events with a maximum LCS below the threshold tlcs are then discarded; events with a maximum LCS above tlcs are considered news items with the maximum LCS. Finally, the remaining news items belonging to the familiar event category are sent to the familiar event determiner, while another category of unknown news is sent to the unknown event determiner.

[0132] In some embodiments, retrieving similar news sets from a keyword news inverted index table based on the keywords includes:

[0133] Determine the cosine similarity of the keyword to each news item in the keyword news inverted index table;

[0134] The similar news set is constructed based on news items whose cosine similarity is greater than a preset threshold.

[0135] Understandably, based on the above example, we can find the first three keywords "City A," "C Dynamics," and "Y Dynamics," and use incremental clustering, with the cosine similarity between news vectors as the metric, to perform clustering. Therefore, news related to the keyword "City A" is clustered into events with event IDs 2, 5, 10, and 11, as shown below. Figure 3 As shown in the diagram. Subsequently, when the input news contains the keyword "City A", the keyword-event mapper quickly retrieves the event set with event IDs 2, 5, 10, and 11, and then calculates the cosine similarity between the input news and each event in the event set. If the cosine similarity is higher than tes (the threshold), the news is added to the corresponding event. Otherwise, a new event is created, the news is added to the newly created event, and then it is inserted into the keyword-event mapper.

[0136] In summary, the ECNNet architecture provided by this invention includes four key components: a feature extractor, an event mapper, a familiar event judge, and an unknown event judge. During the model training phase, a training set X is first provided. 1 The Text-CNN model is trained in the feature extractor to obtain text feature representations of the news. Subsequently, text feature representation The news is input into the event mapper, which then categorizes the news into familiar news items (N). f With unknown news N u There are two main categories of news. Familiar news is then processed into a historical event set in the familiar event judge, which is used to quickly predict the truth or falsehood of subsequent familiar news. Unknown news is input into the unknown event judge for generative adversarial training to extract event feature representations.

[0137] During the model testing phase, a test set X is provided. u The text features of the unlabeled news set are trained in the Text-CNN model. Next, the text features of the test set are represented. The information is input into the event mapper to categorize news into familiar and unfamiliar news. In the familiar event judge, historical event sets are retrieved to calculate the news's veracity rating R. θ (u i This facilitates the determination of whether news is fake. In the unknown event judge, the fake news detector predicts whether a news item is fake and outputs the calculated tag set Y. u .

[0138] In the field of fake news detection on social networks, there is currently no internationally recognized standard test corpus or a corpus that is similar to it. In order to fairly evaluate the performance of the ECNNet model, the data source for this experiment is major social media platforms and portal websites in Simplified Chinese, such as Weibo, Zhihu, Xiaohongshu, Tencent News, and Toutiao. A total of 666 news articles from December 1, 2021 to December 31, 2021 were obtained through web crawling.

[0139] Subsequently, these 666 news items were manually tagged: 0 for real news and 1 for fake news. Of these, 446 were real news and 220 were fake news.

[0140] The statistics on the dataset are shown in Table 1:

[0141] Table 1

[0142]

[0143] In traditional fake news detection and evaluation, precision, recall, and the F-score are three crucial metrics. Precision refers to the predicted outcome; it indicates how many samples predicted as positive are actually positive. Predicting a positive class as positive is denoted as TP, and predicting a negative class as positive is denoted as FP. Recall refers to the original sample; it indicates how many positive instances in the sample were correctly predicted. Predicting a positive class as positive is denoted as TP, and predicting a positive class as negative is denoted as FN. The F1 score is the harmonic mean of precision and recall.

[0144] The evaluation indicators used in this paper are shown in the following formula:

[0145]

[0146]

[0147]

[0148] To verify the effectiveness of the proposed model, the selection of benchmark methods is mainly considered from the following two aspects: traditional machine learning models and neural network deep learning models.

[0149] This invention primarily selects the following four benchmark methods:

[0150] 1. Support Vector Machine (SVM). SVM models are trained using standardized text feature representations and a set of true labels. For this competition, C is set to 50, and the kernel function is set to RF.

[0151] 2. Random Forest (RF). Random Forest uses normalized text feature representations and a set of ground truth labels to train a random forest model. The parameter n_estimators is set to 50.

[0152] 3. Linear Regression (LR). Linear regression uses normalized text feature representations and a set of true labels to train a logistic regression model. The parameter solver is set to lbfgs.

[0153] 4. Long Short-Term Memory (LSTM) Neural Network. LSTM uses fully connected layers as text feature extractors, with text vector representations derived from the text feature estimator. This model has a 256-dimensional hidden size; the input to the fully connected layers is text features, and the output is the probability of real news.

[0154] Experimental steps:

[0155] This experiment was conducted using the machine learning libraries Scikit-Learn and PyTorch. Python version 3.7.2, Scikit-Learn version 0.21.2, and PyTorch version 1.10.0 were used. The training and test sets were divided in an 8:2 ratio. The training set was used to optimize the parameters of ECNNet, and the test set was used to evaluate the model's performance. Recommended parameter settings for the ECNNet model are shown in Table 2.

[0156] Table 2

[0157]

[0158] In the feature extractor, the word embedding dimension is set to 512, the time window size varies from 1 to 4, and the fully connected layer size of the feature extractor is 32. For the event detector, its fully connected layer size is set to 64. The remaining detailed parameter settings are shown in Table 2. For all basic comparison methods and the proposed ECNNet model, the same batch size of 100 and 100 training epochs are used during the training phase.

[0159] For all basic control groups, recommended or optimal parameter settings were used. To ensure fair evaluation, all experimental methods underwent the same pretreatment procedures.

[0160] Results and Analysis:

[0161] The above method was validated on the dataset provided in this invention and compared with several of the most popular fake news detection methods. The results are shown in Table 3.

[0162] Table 3

[0163]

[0164]

[0165] Compared to other experimental methods, the proposed algorithm ECNNet achieves excellent results in terms of accuracy, recall, and F1 score. Specifically, ECNNet's accuracy is significantly higher than other baseline methods, improving by 0.02 compared to the previous state-of-the-art baseline method RF. In terms of recall, ECNNet also outperforms most existing baseline methods, only differing by 0.06 from the current best baseline method LSTM. Considering both recall and accuracy, ECNNet achieves the best F1 score, demonstrating its unique superiority in fake news detection.

[0166] The comprehensive experimental results show that the event-related fake news detection algorithm proposed in this paper can effectively improve the recall rate and achieve accurate detection of fake news in social networks.

[0167] This invention proposes ECNNet, an event-correlation-based fake news detector for breaking news on social networks. It can quickly and efficiently detect fake news based on the event correlation between different news items.

[0168] The ECNNet model proposed in this invention uses an event clusterer to measure the differences between different events and further realizes event clustering. The keyword-event ID inverted index table maintained by the event clusterer can quickly detect familiar news.

[0169] The ECNNet model proposed in this invention is a general framework for fake news detection, with high cohesion and low coupling among its various modules. Users can modify or extend it according to their actual needs.

[0170] Experiments show that the ECNNet model proposed in this invention can effectively detect fake news and demonstrates good performance in terms of accuracy and recall.

[0171] The fake news detection method provided by this invention extracts events from news data by identifying the text features corresponding to the news data to be tested, and determines whether the news data is of a first type or a second type based on the events. If the news data is determined to be of the first type, a first event discriminator is used to retrieve historical events corresponding to the news data to determine whether the news data is fake news. If the news data is determined to be of the second type, a second event discriminator is used to determine whether the news data is fake news. The second event discriminator includes a fake event detector and an event feature extractor. The fake event detector is used to identify text features and obtain the probability that the corresponding event is fake. The event feature extractor is used to classify the news data based on the probability to determine whether the news data is fake news. This invention divides news data into first type and second type news, and inputs them into different event discriminators for fake news discrimination, achieving rapid and efficient detection of fake news based on the event correlation between different types of news.

[0172] like Figure 4 As shown, the present invention also provides a fake news detection device 400, comprising:

[0173] The extraction module 410 is used to acquire the news data to be tested, extract features from the news data, and obtain the corresponding text features.

[0174] The first judgment module 420 is used to obtain events from the news data based on the text features, and to judge the news data as a first type of news or a second type of news based on the events.

[0175] The second judgment module 430 is used to input the news data into the first event discriminator when it is determined that the news data is a first type of news, so as to retrieve the historical events corresponding to the news data based on the first event discriminator and determine whether the news data is fake news.

[0176] The third judgment module 440 is used to input the news data into the second event discriminator when it is determined that the news data is the second type of news, so as to judge whether the news data is fake news based on the second event discriminator;

[0177] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0178] The fake news detection device provided in the above embodiments can implement the technical solutions described in the above fake news detection method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above fake news detection method embodiments, and will not be repeated here.

[0179] like Figure 5 As shown, the present invention also provides an electronic device 500. The electronic device 500 includes a processor 501, a memory 502, and a display 503. Figure 5 Only some components of the electronic device 500 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0180] In some embodiments, memory 502 may be an internal storage unit of electronic device 500, such as a hard disk or memory of electronic device 500. In other embodiments, memory 502 may also be an external storage device of electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 500.

[0181] Furthermore, the memory 502 may include both internal storage units of the electronic device 500 and external storage devices. The memory 502 is used to store application software and various types of data installed on the electronic device 500.

[0182] In some embodiments, processor 501 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 502 or process data, such as the fake news detection method of the present invention.

[0183] In some embodiments, display 503 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 503 is used to display information from electronic device 500 and to display a visual user interface. Components 501-503 of electronic device 500 communicate with each other via a system bus.

[0184] In some embodiments of the present invention, when the processor 501 executes the fake news detection program in the memory 502, the following steps may be performed:

[0185] Obtain the news data to be tested, extract features from the news data, and obtain the corresponding text features;

[0186] Based on the text features, events are obtained from the news data, and based on the events, the news data is determined to be either a first type of news or a second type of news;

[0187] If the news data is determined to be of the first type, the news data is input into the first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and to determine whether the news data is fake news;

[0188] If the news data is determined to be of the second type, the news data is input into the second event discriminator to determine whether the news data is fake news based on the second event discriminator;

[0189] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0190] It should be understood that when the processor 501 executes the fake news detection program in the memory 502, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0191] Furthermore, this embodiment of the invention does not specifically limit the type of electronic device 500 mentioned. Electronic device 500 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the invention, electronic device 500 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0192] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the fake news detection methods provided by the methods described above, the method comprising:

[0193] Obtain the news data to be tested, extract features from the news data, and obtain the corresponding text features;

[0194] Based on the text features, events are obtained from the news data, and based on the events, the news data is determined to be either a first type of news or a second type of news;

[0195] If the news data is determined to be of the first type, the news data is input into the first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and to determine whether the news data is fake news;

[0196] If the news data is determined to be of the second type, the news data is input into the second event discriminator to determine whether the news data is fake news based on the second event discriminator;

[0197] The second event discriminator includes a fake event detector and an event feature extractor; the fake event detector is used to identify the text features and obtain the probability that the corresponding event is a fake event; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

[0198] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0199] The above provides a detailed description of the fake news detection method, apparatus, electronic device, and storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting fake news, characterized in that, include: Obtain the news data to be tested, extract features from the news data, and obtain the corresponding text features; Based on the text features, events are obtained from the news data, and based on the events, the news data is determined to be either a first type of news or a second type of news; If the news data is determined to be of the first type, the news data is input into the first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and to determine whether the news data is fake news; If the news data is determined to be of the second type, the news data is input into the second event discriminator to determine whether the news data is fake news based on the second event discriminator; The second event discriminator includes a fake event detector and an event feature extractor; The fake event detector is used to identify the text features and obtain the probability that the corresponding event is fake; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

2. The method for detecting fake news according to claim 1, characterized in that, The fake event detector is obtained by transfer learning of fake events based on generative adversarial networks.

3. The method for detecting fake news according to claim 1, characterized in that, The step of extracting features from the news data to obtain corresponding text features includes: The news data is segmented and tagged with parts of speech to obtain tagged words; The word vectors corresponding to the labeled words are learned based on the pre-trained word embedding model; Based on the word vectors, the sentence vectors corresponding to the news data are reduced in dimensionality to obtain word embedding vectors; The text features are obtained based on the word embedding vectors.

4. The method for detecting fake news according to claim 3, characterized in that, The process of obtaining the text features based on the word embedding vectors includes: The word embedding vector is input into a convolutional filter to obtain the feature vector corresponding to each sentence in the news data; The feature vector is subjected to max pooling to obtain the text features.

5. The method for detecting fake news according to any one of claims 1-4, characterized in that, The step of obtaining events from the news data based on the text features and determining whether the news data is a first type of news or a second type of news based on the events includes: Search for keywords in the news data based on the text features; Based on the keywords, retrieve similar news sets from the keyword news inverted index table; Determine the cosine similarity of different news items in the similar news set, and based on the cosine similarity of different news items in the similar news set, cluster the news items in the similar news set to obtain the event set corresponding to the keyword; Based on the event set, the news data will be determined to be either the first type of news or the second type of news; The keyword news inverted index table includes multiple preset keywords and a news inverted table corresponding to each preset keyword.

6. The method for detecting fake news according to claim 5, characterized in that, The step of determining whether news data is of type one or type two based on the event set includes: Entropy filtering is performed on the event set to obtain the filtered events; The filtered events are then filtered using the LCS algorithm, and the filtered events are determined to be either type 1 news or type 2 news based on the LCS algorithm filtering results.

7. The method for detecting fake news according to claim 5, characterized in that, The process of retrieving similar news sets from the keyword news inverted index table based on the keywords includes: Determine the cosine similarity of the keyword to each news item in the keyword news inverted index table; The similar news set is constructed based on news items whose cosine similarity is greater than a preset threshold.

8. A fake news detection device, characterized in that, include: The extraction module is used to acquire the news data to be tested, extract features from the news data, and obtain the corresponding text features. The first judgment module is used to obtain events from the news data based on the text features, and to judge the news data as a first type of news or a second type of news based on the events; The second judgment module is used to input the news data into the first event discriminator when it is determined that the news data is a first type of news, so as to retrieve the historical events corresponding to the news data based on the first event discriminator and determine whether the news data is fake news. The third judgment module is used to input the news data into the second event discriminator when it is determined that the news data is the second type of news, so as to determine whether the news data is fake news based on the second event discriminator; The second event discriminator includes a fake event detector and an event feature extractor; The fake event detector is used to identify the text features and obtain the probability that the corresponding event is fake; the event feature extractor is used to classify the news data based on the probability and determine whether the news data is fake news.

9. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is configured to execute the program stored in the memory to implement the steps in the fake news detection method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the fake news detection method as described in any one of claims 1 to 7.

Images