CN115935953A - False news detection method, device, electronic device and storage medium - Google Patents

Abstract

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

Description

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

technical field

The invention relates to the technical field of network information dissemination, in particular to a false news detection method, device, electronic equipment and storage medium.

Background technique

The popularization of social networking services has led to the rapid expansion of its user groups, and at the same time, it has brought about a rapid increase in the amount of information. Social networking online platforms allow their users to freely post information on their platforms. These huge user groups publish massive amounts of information every day, but they are also full of false or false information. False information can spread quickly through the platform of social networks. Therefore, how to quickly and efficiently identify false news is a topic of great significance.

Contents of the invention

In view of this, it is necessary to provide a false news detection method, device, electronic device and storage medium to solve the technical problem that false news cannot be quickly and efficiently identified in the prior art.

In order to achieve the above object, the present invention provides a false news detection method, comprising:

Obtain news data to be tested, perform feature extraction on the news data, and obtain corresponding text features;

acquiring an event from the news data based on the text features, and judging that the news data is a first type of news or a second type of news based on the event;

In the case of determining that the news data is the first type of news, input the news data into a first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and determine Whether said news data is fake news;

In the case of determining that the news data is a second type of news, inputting the news data into a second event discriminator to determine whether the news data is false news based on the second event discriminator;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

Further, the false event detector is obtained by performing migration learning of false events based on an adversarial generation network.

Further, the feature extraction of the news data to obtain corresponding text features includes:

performing word segmentation and part-of-speech marking on the news data to obtain marked words;

The word embedding model learning based on pre-training obtains the word vector corresponding to the word with label;

Carry out dimensionality reduction processing to the sentence vector corresponding to the news data based on the word vector to obtain a word embedding vector;

Based on the word embedding vector, the text feature is obtained.

Further, the text feature is obtained based on the word embedding vector, including:

The word embedding vector is input to a convolution filter to obtain a feature vector corresponding to each sentence in the news data;

Performing maximum pooling processing on the feature vectors to obtain the text features.

Further, the acquiring events from the news data based on the text features, and judging that the news data is the first-type news or the second-type news based on the events, includes:

searching for keywords from the news data based on the text features;

Retrieving similar news sets from the keyword news inverted index table based on the keyword;

determining the cosine similarity of different news in the similar news set, and clustering the news in the similar news set based on the cosine similarity of different news in the similar news set to obtain the event set corresponding to the keyword;

Based on the set of events, it will be judged that the news data is the first type of news or the second type of news;

Wherein, the keyword news posting index table includes a plurality of preset keywords and a news posting list corresponding to each preset keyword.

Further, the judging that the news data is the first type of news or the second type of news based on the event set includes:

performing entropy filtering on the event set to obtain filtered events;

The filtered event is filtered by the LCS algorithm, and based on the LCS algorithm filtering result, it is judged that the filtered event is the first type of news or the second type of news.

Further, the retrieval of similar news sets from the keyword news inverted index table based on the keyword includes:

Determining the cosine similarity corresponding to each piece of news corresponding to the keyword in the keyword news inverted index table;

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

The present invention also provides a false news detection device, comprising:

An extraction module, configured to obtain news data to be tested, perform feature extraction on the news data, and obtain corresponding text features;

A first judging module, configured to acquire an event from the news data based on the text features, and judge that the news data is a first type of news or a second type of news based on the event;

The second judging module is configured to input the news data into the first event discriminator when it is determined that the news data is the first type of news, so as to retrieve the news related to the news based on the first event discriminator. historical events corresponding to the data, and determine whether the news data is false news;

A third judging module, configured to input the news data into a second event discriminator when it is determined that the news data is news of the second type, so as to judge the news data based on the second event discriminator Whether it is fake news;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

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

The memory is used to store programs;

The processor, coupled with the memory, is configured to execute the program stored in the memory, so as to realize the steps in the false news detection method described in any one of the above.

The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for detecting fake news as described in any one of the above is implemented.

The beneficial effect of adopting the above-mentioned implementation is: the false news detection method, device, electronic equipment and storage medium provided by the present invention obtain events from the news data through the corresponding text features of the news data to be tested, and judge the news data based on the event as the first A type of news or a second type of news; in the case of determining that the news data is the first type of news, based on the first event discriminator to retrieve the historical events corresponding to the news data, judge whether the news data is false news; determine whether the news data is In the case of the second type of news, judge whether the news data is false news based on the second event discriminator; wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify text features , to obtain the probability that the corresponding event is a false time; the event feature extractor is used to classify the news data based on the probability, and determine whether the news data is false news. The invention divides the news data into the first type of news and the second type of news, which are respectively input to different event discriminators for discriminating false news, and realizes fast and efficient detection of false news based on the event correlation between different news.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

Fig. 1 is a schematic flow chart of an embodiment of the false news detection method provided by the present invention;

Fig. 2 is the keyword news inverted index schematic diagram that the present invention provides;

Fig. 3 is the schematic diagram of keyword event ID inverted index provided by the present invention;

Fig. 4 is the structural schematic diagram of false news detection device provided by the present invention;

Fig. 5 is a schematic structural diagram of an embodiment of an electronic device provided by the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Apparently, the described embodiments are only some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts belong to the protection scope of the present invention.

In the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.

The terms "comprising" and "having" and any variations thereof in the embodiments of the present invention are intended to cover a non-exclusive inclusion, for example, a process, method, device, product or equipment that includes a series of steps or modules is not necessarily limited to expressly Instead, other steps or modules not explicitly listed or inherent to the process, method, product or apparatus may be included.

The naming or numbering of the steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time/logic order indicated by the naming or numbering, and the steps that have been named or numbered can be performed according to the requirements To achieve the technical purpose, change the execution sequence, as long as the same or similar technical effect can be achieved.

Reference herein to an "embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present invention. The occurrences of this phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is understood explicitly and implicitly by those skilled in the art that the embodiments described herein can be combined with other embodiments.

The present invention provides a false news detection method, device, electronic equipment and storage medium, which will be described respectively below.

As shown in Figure 1, the present invention provides a kind of false news detection method, comprises:

Step 110: Obtain the news data to be tested, perform feature extraction on the news data, and obtain corresponding text features.

It can be understood that the method provided by the present invention can be implemented based on the ECNNet architecture, which integrates the word vector model, convolutional neural network, generative confrontation network and incremental clustering algorithm, through the feature extractor, event mapper, familiar event The four modules of judger and unknown event judger realize false news detection for social network emergencies.

Determine a time range T, obtain enough original news data from mainstream social networks such as Sina Weibo, Toutiao, and Tencent News through web crawlers, and filter and determine the research scope S of the event. The news data is input to the feature extractor, and the feature extractor extracts information features from the text content of the news data. Here, a convolutional neural network (CNN) can be used as the core module of the feature extractor.

Step 120: Acquire events from the news data based on the text features, and determine whether the news data is news of the first type or news of the second type based on the events.

It can be understood that the text features are input to the event mapper, and the event mapper combines the text features extracted by the feature extractor to collect events from news data and divide them into familiar news (ie: first-type news) and Unknown news (ie: the second type of news).

Step 130, when it is determined that the news data is the first type of news, input the news data into the first event discriminator, so as to retrieve the history corresponding to the news data based on the first event discriminator event, judging whether the news data is fake news.

It can be understood that the first event discriminator is also the familiar event discriminator, which receives news from the event mapper and predicts whether the news is true or false by retrieving corresponding historical events.

For each news event judged as "familiar news", the corresponding historical event set is searched, and then the event correlation degree between the familiar news and the retrieved event set is calculated. The definition of event relevance among the present invention is as follows:

Suppose there is a historical event 1 and an event 2, and the event 2 is composed of n pieces of news, namely news ₁ , news ₂ , ..., news _n . The cosine similarities between these news and historical event 1 are sim ₁ , sim ₂ , ..., sim _n . The event correlation degree S of historical event 1 and event 2 is defined as follows:

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

where l _i represents the label value of the i-th news. If the final calculated news truth degree R is greater than 0, it indicates that the predicted familiar news is false news, otherwise, the familiar news is considered real.

Step 140, when it is determined that the news data is a second type of news, input the news data into a second event discriminator, so as to judge whether the news data is fake news based on the second event discriminator ;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

It can be understood that the second event discriminator is also an unknown event discriminator, and the unknown event discriminator uses a generative confrontation model to perform transfer learning on fake news, and thereby judge whether the unknown news is fake news. In addition, a convolutional neural network layer is integrated for event feature learning.

An event is usually a collection of a series of news reports, that is to say, there is an event correlation between news and news reported on the same event, and there is also an event correlation between news and events. Therefore, it can be considered to use the generative confrontation network to construct the unknown event judger of ECNNet. The unknown event judge consists of two modules: a false event detector and an event feature extractor.

In some embodiments, the false event detector is obtained by performing migration learning of false events based on an adversarial generative network.

It is understandable that the fake event detector uses the generative adversarial network to perform transfer learning in fake events, and the event feature extractor uses a two-layer fully connected neural network to extract event features.

The false event detector uses the generative confrontation network to perform migration learning in false events. It takes the text feature representation _RT extracted by the feature extractor as input, and outputs the probability that the i-th event is a false event. The i-th event denoted as m _i . The fake event detector applies a fully connected layer with softmax to predict whether news is fake or not. In the present invention, the spurious event detector is denoted as G _d (·; θ _d ), where θ _d represents all parameters involved. Taking the text feature R _T of the i-th news as input, the output is recorded as P _θ (m _i ), and the calculation formula is as follows:

P _θ (m _i ) = G _d (G _f (m _i ; θ _f ); θ _d )

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

In the next process, by finding the optimal parameters θ _f and θ _d to make the loss function L _d (θ _f , θ _d ) reach the minimum value, this process can be expressed as:

However, false event detectors can only learn event-specific features, but cannot perform generalized representations, which is not conducive to detecting events not included in the training dataset. Therefore, this requires the ability to learn transferable feature representations of emerging events. Consider the model being able to learn more general feature representations that represent news related to the same event. Therefore, in order to learn the common features of events, consider adding the following event feature extractor to improve the model.

The event feature extractor is essentially a two-way adversarial transfer learning process, which makes full use of the data features of the source domain and the target domain.

The event feature extractor, denoted _Ge ( _RF , _θe ), where denote all parameters _Ge included in it, deploys a convolutional neural network layer to correctly classify the input domain including news and events. Define _Ye as a set of domain labels, so the loss of the event discriminator can be expressed as:

Since the discriminator seeks to recognize the input domain, from the discriminator's point of view, the loss function should be minimized to find the optimal parameters:

However, for the event feature extractor, it aims to fool the fake event detector to learn common features of events. A larger loss function means that it learns more common features, since a larger loss function indicates that it is more difficult to correctly classify events into clusters. Therefore, the above loss needs to be maximized to find the optimal parameters.

Before the training phase of the unknown event discriminator, all historical events are extracted as the training set. In the training phase, on the one hand, the GAN G _f ( ; θ _f ) training data set is used. In order to improve the performance of false event detection, the event feature extractor cooperates with the false event detector to minimize the loss L _d (θ _f , θ _d ). On the other hand, there is a minimax game between the event feature extractor and the spurious event detector. Therefore, the final loss can be expressed as:

L _final (θ _f ,θ _d ,θ _e )＝L _d (θ _f ,θ _d )＝L _e (θ _f ,θ _e )

需要最小化L_final，以及考虑极小极大博弈。因此，最佳参数与过程达到平衡的时间有关，它们可以表示为：In order to find the optimal parameters

It is necessary to minimize L _final , and to consider the minimax game. Therefore, the optimal parameters are related to the time for the process to reach equilibrium, and they can be expressed as:

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

performing word segmentation and part-of-speech marking on the news data to obtain marked words;

The word embedding model learning based on pre-training obtains the word vector corresponding to the word with label;

Carry out dimensionality reduction processing to the sentence vector corresponding to the news data based on the word vector to obtain a word embedding vector;

Based on the word embedding vector, the text feature is obtained.

It can be understood that in order to extract rich text features in news, word segmentation and part-of-speech tagging are performed sequentially on the input news text (ie, news data). Then, the word vector of each word is learned through the pre-trained word embedding model. In addition, the PCA (Principal Component Analysis) model is used to reduce the dimensionality of the sentence vector of each news item. The resulting sequential list of words (i.e., word embedding vectors) is the input to the text feature extractor.

In order to better extract the corresponding information features from the news text, a convolutional neural network is used as the core module of the text feature extractor. A modified convolutional neural network model, Text-CNN, is added to the text feature extractor. The architecture of Text-CNN uses multiple filters with windows of different sizes to filter text features of different granularities.

In some embodiments, the obtaining the text features based on the word embedding vector includes:

The word embedding vector is input to a convolution filter to obtain a feature vector corresponding to each sentence in the news data;

Performing maximum pooling processing on the feature vectors to obtain the text features.

Understandably, each word in the news text is text-vectorized and represented as 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 expressed as:

表示连接操作符。窗口大小为h的卷积滤波器以文本中的h个词语连续序列作为输入，输出一个特征。为了清楚地说明过程，以第i个单词开始的h个词语的连续序列为例，过滤操作可以表示为：in,

Represents a concatenation operator. A convolutional filter with a window size h takes as input a continuous sequence of h words in the text and outputs a feature. In order to clearly illustrate the process, taking the continuous sequence of h words starting from the i-th word as an example, the filtering operation can be expressed as:

t _i =σ(W _c ·T _i:i+h-1 )

Here σ( ) is an activation function of ReLU, and W _c represents the weight of the convolution filter. This filter can also be applied to the remaining words, and then a feature vector of the sentence can be obtained as follows:

t=[t ₁ ,t ₂ ,...,t _n-h+1 ]

For each feature vector, the maximum pooling operation is used to find the maximum value, so as to extract the most important information in the text. Now, the corresponding characteristics of a particular filter are obtained. This process is repeated until all filter characteristics are obtained. To extract text features with different granularities, different window sizes are applied. For a particular window size, there are n _h different filters.

在最大池化操作之后，使用一个全连接层来确保最终的文本特征表示(记为

)具有p维特征：Thus, assuming there are c possible window sizes, there are a total of c n _h filters. The text features after the max pooling operation are denoted as

After the max pooling operation, a fully connected layer is used to ensure the final text feature representation (denoted as

) has p-dimensional features:

Among them, W _tf is the weight matrix of the fully connected layer.

In some embodiments, the acquiring events from the news data based on the text features, and judging that the news data is the first type of news or the second type of news based on the events, includes:

searching for keywords from the news data based on the text features;

Retrieving similar news sets from the keyword news inverted index table based on the keyword;

determining the cosine similarity of different news in the similar news set, and clustering the news in the similar news set based on the cosine similarity of different news in the similar news set to obtain the event set corresponding to the keyword;

Based on the set of events, it will be judged that the news data is the first type of news or the second type of news;

Wherein, the keyword news posting index table includes a plurality of preset keywords and a news posting list corresponding to each preset keyword.

It can be understood that the steps in this embodiment are implemented based on the event mapper, and the event mapper includes three parts: keyword-news mapper, keyword-event mapper and filter.

The keyword-news mapper is used to quickly retrieve whether the input news has similar news in the existing news, which is essentially a dynamically updated inverted index table. In order to reduce the time required to search for previously encountered news similar to the input news d while maintaining constant time and space requirements, a keyword-news inverted index table maintained within the time window t can be used, as shown in Figure 2 Show. The set M is constantly updated by replacing the oldest news with the newest incoming news, to keep the memory requirements of the keyword-news inverted index constant, the number of keywords may increase due to the unbounded use of vocabulary in the news stream. become very large. Each entry in the keyword-news inverted index table contains a keyword and a finite set Q. The set Q is the latest news in which the keyword appears. But when the number of news exceeds the limit of Q, the oldest news will be replaced by the newest news containing this keyword.

Based on the TF-IDF (term frequency–inverse document frequency) method, select the first k keywords from the news, and then calculate each keyword in the k keywords and each corresponding keyword in the keyword-news inverted index table Cosine similarity of a piece of news to retrieve potentially similar news sets.

For example, input the news "City A press conference notification: From 16:00 to 22:00 on April 22, 4 new cases of C dynamic personnel (2 cases in E area and 2 cases in F area) were added in city A. Another case of Y dynamic personnel was added. "The first three TF-IDF weighted keywords are "A City", "C Dynamics" and "Y Dynamics". Search each keyword in the keyword-news inverted index table, and retrieve news with IDs 3, 5, 7, 15, 18, 21 and 25. Finally, the similarity between two news vectors is calculated using cosine distance. The formula for calculating the cosine distance is shown in the following formula:

If the cosine similarity of none of the news is higher than the tsi value (threshold), it means that no similar news has occurred before. So a new event will be created and newsd assigned to the event, which will then be sent to the keyword-event mapper. If there is a similarity in the above tsi, the news d will be directly sent to the keyword-event mapper. Finally, the keyword-news mapper is updated by adding the news to the corresponding news set of k terms.

A keyword-event mapper is used to detect and classify events into familiar and unknown events. In the initial state, the keyword-event mapper only contains a dynamically updated keyword-historical event inverted index table, as shown in FIG. 3 . Similar to the keywords-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 will be replaced by the newest event.

In some embodiments, the judging that the news data is the first type of news or the second type of news based on the event set includes:

performing entropy filtering on the event set to obtain filtered events;

The filtered event is filtered by an LCS (Longest Common Subsequence) algorithm, and based on the filtering result of the LCS algorithm, it is judged that the filtered event is the first type of news or the second type of news.

It can be understood that in the keyword-event mapper, a clustering method is used to cluster news into events, and smaller events are generated during the clustering process, which are called fragmented events. The existence of fragmentation events will affect the accuracy and speed of the model.

To improve the accuracy and speed of the model, filters are used to filter out fragmented events and those nonsensical candidates. The filter is mainly composed of two parts, which are entropy filtering and LCS algorithm filtering.

Entropy filtering uses entropy information in candidate event clusters. By calculating the entropy of each candidate event cluster and comparing it with the preset entropy threshold (tent), if the entropy value of the 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 amount of information is determined to be a fragmented cluster and discarded.

LCS Algorithm Filtering Considering that the news in the event usually has a similar sentence structure, use the LCS algorithm for each candidate event, and record the length of the largest LCS. Events with maximum LCS below tlcs(threshold) are then discarded, and for events above tlcs, events representing news with maximum LCS. Finally, the remaining news belonging to familiar events is sent to the familiar event judger, while another category of unknown news is sent to the unknown event judger.

In some embodiments, the retrieval of similar news sets from the keyword news inverted index table based on the keyword includes:

Determining the cosine similarity corresponding to each piece of news corresponding to the keyword in the keyword news inverted index table;

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

It can be understood that, based on the above example, the first three keywords "A city", "C dynamic" and "Y dynamic" can be found, using the method of incremental clustering, taking the cosine similarity between news vectors as a measure Criteria for clustering. Therefore, news related to the keyword "city A" is clustered into events with event IDs 2, 5, 10, and 11, as shown in FIG. 3 . Thereafter, when the input news contains the keyword "City A", the keyword-event mapper will quickly retrieve event sets with event IDs of 2, 5, 10, and 11, and then calculate the value of each event in the input news and event sets. Cosine similarity. If the cosine similarity is higher than tes (threshold), add the news to the corresponding event. Otherwise create a new event and add the news to the newly created event and insert it into the keyword-event mapper.

其后，文本特征表示

输入到事件映射器中，事件映射器会将输入的新闻划分为熟悉新闻N_f与未知新闻N_u两大类新闻。随后，熟悉新闻在熟悉事件判断器中形成历史事件集，用于后续熟悉新闻的快速预测真假与否；未知新闻输入到未知事件判断器中，用来进行生成对抗训练，以提取得到事件特征表示。In summary, the ECNNet architecture provided by the present invention includes four key components: feature extractor, event mapper, familiar event judger and unknown event judger. In the training phase of the model, first provide the training set X ¹ to enter the feature extractor to train the Text-CNN model to obtain the text feature representation of the news

Afterwards, the text feature representation

Input into the event mapper, the event mapper will divide the input news into two types of news: familiar news N _f and unknown news N _u . Subsequently, the familiar news forms a historical event set in the familiar event judger, which is used to quickly predict whether the familiar news is true or not; the unknown news is input into the unknown event judger, which is used for generative confrontation training to extract event features express.

紧接着，将测试集的文本特征表示

输入到事件映射器中，用于划分为熟悉新闻与未知新闻。在熟悉事件判断器中，检索历史事件集用以计算新闻的真实度R_θ(u_i)以便于判断是否为虚假新闻。在未知事件判断器中，虚假新闻检测器来预测新闻是否为虚假新闻，并输出计算得到的标签集Y^u。In the testing phase of the model, the test set ^Xu is provided to enter the Text-CNN model for training to obtain the text feature representation of the unlabeled news set

Next, the text features of the test set are represented by

Input into the Event Mapper for classification into familiar versus unknown news. In the familiar event judger, retrieve the historical event set to calculate the authenticity R _θ (u _i ) of the news so as to judge whether it is fake news. In the unknown event judger, the fake news detector is used to predict whether the news is fake news, and output the calculated label set Y ^u .

In the field of fake news detection in social networks, there is no internationally recognized standard test corpus or a corpus closely related to it. In order to fairly evaluate the performance of the ECNNet model, the data sources of this experiment are major social platforms and portals in the Simplified Chinese network, such as Weibo, Zhihu, Xiaohongshu, Tencent News, Toutiao, etc. method to obtain a total of 666 news from December 1, 2021 to December 31, 2021.

Afterwards, the 666 pieces of news were labeled manually, and the news was recorded as 0 if it was real news, and 1 if it was false news. Among them, 446 were real news and 220 were fake news.

The statistical information on the dataset is shown in Table 1:

Table 1

In the traditional evaluation of false news detection, precision, recall and F-score are three very important evaluation indicators. The correct rate refers to the prediction result, which indicates how many of the samples predicted to be positive are true positive samples. The positive class is predicted as positive class as TP, and the negative class is predicted as positive class as FP. The recall rate is for the original sample, which indicates how many positive examples in the sample are predicted correctly. Predict the original positive class as positive class as TP, and predict the original positive class as negative class as FN. The F1 score is the harmonic mean of the precision and recall.

The evaluation index used in this paper is shown in the following formula:

In order to verify the effectiveness of the proposed model, the choice of benchmark methods is mainly considered from the following two aspects: traditional machine learning models and neural network deep learning models.

The present invention mainly selects following four benchmark methods:

1. Support Vector Machine (Suppoort Vector Machine, SVM). Support Vector Machines use a standardized representation of textual features and a set of ground truth labels to train a support vector machine model. The entry C is set to 50, and the kernel function is set to RF.

2. Random Forest (Random Forest, RF). Random Forest trains a random forest model using normalized text feature representations and ground truth label sets. The parameter n_estimators is set to 50.

3. Linear Regression (LR). Linear Regression trains a Logistic Regression model using normalized text feature representations and ground truth label sets. The parameter solver is set to lbfgs.

4. Long Short Term Memory neural network (Long Short Term Memory, LSTM). The LSTM neural network uses a fully connected layer as a text feature extractor, and the text vector representation comes from a text feature estimator. The model has a 256-dimensional hidden size, the input of the fully connected layer is the text feature, and the output is the probability of the real news.

Experimental steps:

This experiment is completed by calling the machine learning libraries Scikit-Learn and Pytorch. The Python version is 3.7.2, the Scikit-Leanrn version is 0.21.2, and the Pytorch version is 1.10.0. The training set and test set are divided in a ratio of 8:2. The training set is used to optimize the parameter settings of ECNNet, and the test set is used to evaluate the performance of the model. The recommended parameter settings of the ECNet model are shown in Table 2:

Table 2

In the feature extractor, the dimension of the word embedding is set to 512 dimensions, the time window size ranges from 1 to 4, and the size of the fully connected layer of the feature extractor is 32. For the event detector, set its fully connected layer size to 64. The rest of the detailed parameter settings are shown in Table 2. For all the basic control methods and the proposed ECNNet model, the same batch size of 100 is used in the training phase, and the training rounds are 100.

For all base control groups, the recommended or optimal parameter settings were used. In order to obtain a fair evaluation, the same preprocessing operation is carried out for all the experimental methods.

results and analysis:

The above method is verified on the data set provided by the present invention, and compared with the most popular false news detection methods at present, the results are shown in Table 3:

table 3

Compared with other experimental methods, the algorithm ECNNet proposed by the present invention has achieved good results in accuracy, recall and F1 value. Among them, the accuracy of ECNNet is much higher than other baseline methods, which is 0.02 higher than the previous advanced baseline method RF. In terms of recall rate, the recall rate of ECNNet is also higher than most of the previous baseline methods, and compared with the best baseline method LSTM, the difference is only 0.06. Considering the recall rate and accuracy rate of ECNNet, the F1-Score of ECNNet has achieved the best performance, indicating that the model has unique advantages in false news detection.

The comprehensive experimental results show that the fake news detection algorithm based on event correlation proposed in this paper can effectively improve the recall rate of the algorithm and realize the accurate detection of fake news in social networks.

The present invention proposes an event correlation-based false news detector ECNNet aimed at social network emergencies, which can quickly and efficiently detect false news based on the event correlation between different news.

The ECNNet model proposed by the present invention uses an event clusterer to measure the difference between different events, and further implements event clustering. The keyword-event ID inverted index table maintained by the event clusterer can quickly realize the detection of familiar news.

The ECNNet model proposed by the present invention is a general framework for false news detection, and realizes high cohesion and low coupling among various internal modules. Users can change or expand according to actual needs.

Experiments show that the ECNNet model proposed by the present invention can effectively detect false news, and shows good results in accuracy and recall.

In the method for detecting false news provided by the present invention, the event is obtained from the news data by the corresponding text features of the news data to be tested, and the news data is judged to be the first type news or the second type news based on the event; after determining that the news data is the first In the case of type news, retrieve historical events corresponding to the news data based on the first event discriminator, and judge whether the news data is false news; Whether the data is false news; wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor is used based on Probabilistically classifies news data and determines whether news data is fake news or not. The invention divides the news data into the first type of news and the second type of news, which are respectively input to different event discriminators for discriminating false news, and realizes fast and efficient detection of false news based on the event correlation between different news.

As shown in Figure 4, the present invention also provides a fake news detection device 400, including:

The extraction module 410 is used to obtain news data to be tested, perform feature extraction on the news data, and obtain corresponding text features;

The first judging module 420 is configured to acquire events from the news data based on the text features, and judge that the news data is news of the first type or news of the second type based on the events;

The second judging module 430 is configured to input the news data into the first event discriminator when it is determined that the news data is news of the first type, so as to retrieve the information related to the news based on the first event discriminator. Historical events corresponding to the news data, judging whether the news data is false news;

The third judging module 440 is configured to input the news data into the second event discriminator when it is determined that the news data is news of the second type, so as to judge the news based on the second event discriminator Whether the data is fake news;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

The false news detection device provided in the above embodiments can realize the technical solutions described in the above embodiments of the false news detection method. Let me repeat.

As shown in FIG. 5 , the present invention also provides an electronic device 500 correspondingly. The electronic device 500 includes a processor 501 , a memory 502 and a display 503 . FIG. 5 shows only some components of the electronic device 500, 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.

The storage 502 may be an internal storage unit of the electronic device 500 in some embodiments, such as a hard disk or memory of the electronic device 500 . The memory 502 may also be an external storage device of the electronic device 500 in other embodiments, such as a plug-in hard disk equipped on the electronic device 500, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.

Further, the memory 502 may also include both an internal storage unit of the electronic device 500 and an external storage device. The memory 502 is used to store application software and various data installed on the electronic device 500 .

Processor 501 can be a central processing unit (Central Processing Unit, CPU) in some embodiments, microprocessor or other data processing chip, is used for running the program code stored in memory 502 or processing data, for example among the present invention Methods for fake news detection.

In some embodiments, the display 503 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, Organic Light-Emitting Diode) touch device, and the like. The display 503 is used for displaying information on the electronic device 500 and for displaying a visualized user interface. Components 501-503 of electronic device 500 communicate with each other via a system bus.

In some embodiments of the present invention, when the processor 501 executes the fake news detection program in the memory 502, the following steps can be implemented:

Obtain news data to be tested, perform feature extraction on the news data, and obtain corresponding text features;

acquiring an event from the news data based on the text features, and judging that the news data is a first type of news or a second type of news based on the event;

In the case of determining that the news data is the first type of news, input the news data into a first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and determine Whether said news data is fake news;

In the case of determining that the news data is a second type of news, inputting the news data into a second event discriminator to determine whether the news data is false news based on the second event discriminator;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

It should be understood that, when the processor 501 executes the fake news detection program in the memory 502, other functions may be realized besides the above functions, and for details, refer to the descriptions of the above corresponding method embodiments.

Further, the embodiment of the present invention does not specifically limit the type of the mentioned electronic device 500, and the electronic device 500 may be a mobile phone, a tablet computer, a personal digital assistant (personal digital assistant, PDA), a wearable device, a laptop computer ( laptop) and other portable electronic devices. 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 may also be other portable electronic devices, such as a laptop computer (laptop) with a touch-sensitive surface (such as a touch panel). It should also be understood that, in some other embodiments of the present invention, the electronic device 500 may not be a portable electronic device, but a desktop computer with a touch-sensitive surface (such as a touch panel).

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it is implemented to perform the fake news detection method provided by the above-mentioned methods, the method includes :

Obtain news data to be tested, perform feature extraction on the news data, and obtain corresponding text features;

acquiring an event from the news data based on the text features, and judging that the news data is a first type of news or a second type of news based on the event;

In the case of determining that the news data is the first type of news, input the news data into a first event discriminator to retrieve historical events corresponding to the news data based on the first event discriminator, and determine Whether said news data is fake news;

In the case of determining that the news data is a second type of news, inputting the news data into a second event discriminator to determine whether the news data is false news based on the second event discriminator;

Wherein, the second event discriminator includes a false event detector and an event feature extractor; the false event detector is used to identify the text features to obtain the probability that the corresponding event is a false time; the event feature extractor uses After classifying the news data based on the probability, it is determined whether the news data is fake news.

Those skilled in the art can understand that all or part of the processes of the methods in the above embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium. Wherein, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, and the like.

Above, the false news detection method, device, electronic equipment and storage medium provided by the present invention have been introduced in detail. In this paper, specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only used to help Understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation and scope of application. In summary, the content of this specification should not be construed as a limitation of the invention.

Claims (10)

1. A false news detection method, comprising:

acquiring news data to be detected, and extracting characteristics of the news data to obtain corresponding text characteristics;

acquiring an event from the news data based on the text feature, and judging whether the news data is first type news or second type news based on the event;

under the condition that the news data is determined to be first type news, inputting the news data into a first event discriminator so as to retrieve a historical event corresponding to the news data based on the first event discriminator and judge whether the news data is false news or not;

under the condition that the news data is determined to be of a second type of news, inputting the news data into a second event discriminator so as to judge whether the news data is false news or not based on the second event discriminator;

wherein the second event discriminator comprises a false event detector and an event feature extractor; the false event detector is used for identifying the text characteristics to obtain the probability that the corresponding event is false time; the event feature extractor is used for classifying the news data based on the probability and determining whether the news data is false news.

2. A false news detection method according to claim 1, wherein the false event detector is based on transfer learning of false events against a generating network.

3. The false news detection method of claim 1, wherein the extracting the features of the news data to obtain corresponding text features comprises:

performing word segmentation processing and part-of-speech tagging on the news data to obtain tagged words;

learning based on a pre-trained word embedding model to obtain a word vector corresponding to the marked word;

performing dimensionality reduction processing on sentence vectors corresponding to the news data based on the word vectors to obtain word embedded vectors;

and obtaining the text features based on the word embedding vector.

4. A false news detection method according to claim 3, wherein said deriving the text feature based on the word embedding vector comprises:

inputting the word embedding vector into a convolution filter to obtain a characteristic vector corresponding to each sentence in the news data;

and performing maximum pooling on the feature vectors to obtain the text features.

5. The false news detection method of any one of claims 1-4, wherein the obtaining an event from the news data based on the text feature and determining whether the news data is a first type of news or a second type of news based on the event comprises:

searching for keywords from the news data based on the text features;

retrieving a similar news set from a keyword news inverted index table based on the keywords;

determining cosine similarity of different news in the similar news set, and clustering the news in the similar news set based on the cosine similarity of the different news in the similar news set to obtain an event set corresponding to the keyword;

judging whether the news data is first type news or second type news based on the event set;

the keyword news inverted index table comprises a plurality of preset keywords and a news inverted table corresponding to each preset keyword.

6. The false news detection method of claim 5, wherein the determining news data as either a first type of news or a second type of news based on the set of events comprises:

entropy filtering is carried out on the event set to obtain filtered events;

and performing LCS algorithm filtering on the filtered event, and judging that the filtered event is the first type news or the second type news based on the LCS algorithm filtering result.

7. The false news detection method of claim 5, wherein the retrieving similar news sets from a keyword news inverted index table based on the keyword comprises:

determining cosine similarity corresponding to each piece of news corresponding to the keywords in the keyword news inverted index table;

and constructing the similar news set based on the news with cosine similarity larger than a preset threshold.

8. A false news detection device, comprising:

the extraction module is used for acquiring news data to be detected and extracting the characteristics of the news data to obtain corresponding text characteristics;

the first judging module is used for acquiring an event from the news data based on the text characteristic and judging whether the news data is first type news or second type news based on the event;

the second judgment module is used for inputting the news data into a first event discriminator under the condition that the news data is determined to be the first type of news, so as to retrieve the historical events corresponding to the news data based on the first event discriminator and judge whether the news data is false news;

the third judging module is used for inputting the news data into a second event discriminator under the condition that the news data is determined to be the second type news, so as to judge whether the news data is false news or not based on the second event discriminator;

wherein the second event discriminator comprises a false event detector and an event feature extractor; the false event detector is used for identifying the text characteristics to obtain the probability that the corresponding event is false time; the event feature extractor is used for classifying the news data based on the probability and determining whether the news data is false news.

9. An electronic device comprising a memory and a processor, wherein,

the memory is used for storing programs;

the processor, coupled to the memory, is configured to execute the program stored in the memory to implement the steps in the false news detection method according to any one of claims 1 to 7.

10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the false news detection method according to any one of claims 1 to 7.

Images