False news detection method based on breadth view angle

Abstract

The invention provides a false news detection method based on a breadth view angle. The method aims at solving the problem that news backgrounds and propagation time dynamics are neglected in an existing detection method. According to the method, macroscopic and microscopic news backgrounds are constructed, the publishing background of an original post is observed, popularity-oriented and novelty-oriented features are extracted respectively, and semantic features of news are extracted by using a pre-training model BERT. The method of combining the time graph snapshot and the graph convolutional neural network is also introduced, the time dynamic feature and the structural feature of the false news are extracted, the multi-stage graph snapshot is formed, and the diffusion process of the false news is captured. And finally, fusing the breadth view angle features, the propagation structure features and the text semantic features, and realizing news true and false classification through a SoftMax classifier. According to the method, the news background, the propagation rule and the dynamic characteristics of the news background can be more effectively analyzed, so that the accuracy and timeliness of false news detection are improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and specifically relates to technologies such as graph neural networks, natural language processing, and deep learning, and in particular to a fake news detection method based on a breadth perspective. Background Art

[0002] With the popularity of mobile devices and the rapid development of social media, it has become increasingly convenient to publish and spread fake news on social platforms. A large amount of fake news may cause public panic and even have adverse effects on individuals. Therefore, automated fake news detection is particularly important.

[0003] In addition to directly analyzing the content patterns of posts, existing fake news detection technologies also verify the authenticity of information by checking users' responses to posts and using external knowledge to obtain more relevant data. However, these technologies often ignore the overall and local news context of fake news in its generation and dissemination process from a broad perspective.

[0004] The creation of fake news is usually influenced by recent mainstream media opinion and public attention. This is the context in which the news is located and is also an important reference factor for the creation of contemporary fake news. Only when fake news is widely disseminated and spread rapidly can it achieve its intended effect. Therefore, the makers of fake news often deliberately design content to increase the exposure of posts in the current news context and attract more audience attention. This careful planning makes fake news closely related to the current news context that is widely disseminated. In turn, researchers can also extract valuable clues from this context to more effectively identify and detect fake news.

[0005] In addition, some recent fake news detection methods based on deep learning, such as the use of bidirectional graph convolutional network models, analyze the completion stage of the fake news propagation process to form a graph structure, and try to learn the pattern of news propagation from it. However, these methods usually simply represent the spread of fake news as a static graph, and fail to fully consider the temporal dynamics of the spread of fake news. Therefore, there is still room for improvement when capturing the dynamic information of fake news.

[0006] In summary, combining the news context of the news and the dynamic characteristics of the news dissemination process can provide a more comprehensive and accurate solution for fake news detection. Summary of the invention

[0007] The purpose of the present invention is to solve the problem that the existing false news detection methods do not observe the background of the news from the perspective of breadth and fail to fully consider the temporal dynamics of false news propagation, and to provide a false news detection method based on breadth perspective.

[0008] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0009] 1) Given an original post p, we first observe its news context from a breadth perspective and construct the macro context and micro context (macEnv and micEnv) using the news data closest to the time when p was published.

[0010] 2) Model the macro background and micro background respectively and generate the macro background vector v p,mac and microscopic background vector v p,mic .

[0011] 3) The macro background vector v p,mac and microscopic background vector v p,mic Splicing, we get the wide view feature vector V(p).

[0012] 4) Create a time graph snapshot containing text content and comment content for each post, which contains graph snapshots at each time step.

[0013] 5) For the graph snapshot at each time step, use Graph Convolutional Networks (GCN) to extract the propagation structure features Gn of each time step.

[0014] 6) For the propagation structure feature Gn of each time step, the attention mechanism is used to learn the attention weights and aggregate the weighted inputs to obtain the propagation structure feature G(p) that fully considers the temporal dynamics.

[0015] 7) Use the pre-trained model BERT to extract the semantic feature vector of the text content of the original post p and generate the text embedding representation BERT(p).

[0016] 8) Concatenate V(P) with broad perspective features, G(p) with propagation structure features, and BERT(p) with news text semantic features to obtain the final feature vector F.

[0017] 9) Input the feature vector F into SoftMax to classify the news as true or false.

[0018] Specifically, the step 1) includes:

[0019] Let E be the set of all collected news items published earlier than the original post p, the macro context macEnv is the set of news in E published T days before p was published, and the micro context micEnv is the set of the top k news items in the macro context macEnv that are most similar to p.

[0020] The macro context macEnv obtained by time restriction provides a macro perspective of how the general public reads and pays attention to recent events, while the micro context micEnv obtained by further relevance restriction describes the distribution of news about similar events. The pre-trained language model BERT is used to obtain news representation. For each news in the macro context and micro context, the [CLS] output of BERT is used as its initial representation.

[0021] Further, step 2) comprises:

[0022] Fake news often chooses to piggyback on hot news for greater exposure, and at the same time adds some novel information to attract readers' attention. Therefore, popularity and novelty are two very important indicators in the creation of fake news. From the perspective of macro context, if many items in macEnv are similar to p, then p may also be popular in such an environment. That is to say, in macEnv, it is necessary to extract popularity-oriented macro context features as the perceived macro context vector v p,mac .

[0023] Unlike macEnv, micEnv contains mainstream news items close to p, which indicates that they are likely to point to similar events. However, even in a hot event, a post will not attract people's attention if it is too similar to other news. What this type of news needs to do is to be different from other news, that is, novelty. If the content of a post is novel, it will be considered an outlier in the micro-context. Therefore, it is necessary to extract novelty-oriented micro-context information as the perceived micro-context vector v p,mic .

[0024] Further, step 4) includes:

[0025] In order to capture the evolution law of fake news diffusion, the present invention constructs a series of graph snapshots to simultaneously consider the temporal information and spatial structure information of the graph structure.

[0026] Specifically, in order to construct a time graph snapshot, the present invention divides the post into several stages at fixed time intervals and generates n graph snapshot sequences {G p1 ,G p2 ,...,G pn}, where G pi = <V pi ,E pi > represents the graph snapshot at the i-th time stage. At each time graph snapshot G pi In the example, posts are connected to their corresponding comment nodes according to the time order of the comments, thus generating the adjacency matrix A. pi ∈{0,1} m×mAt the time graph snapshot G pi For each node, word embedding is used as its initial feature vector to generate the node’s feature matrix.

[0027] After generating the graph snapshot, the present invention uses the graph convolutional network GCN to perform representation learning on the graph snapshot to obtain the propagation structure feature Gn of each time step.

[0028] After the above steps 1) to 9), for a certain news post, we can get its prediction label whether it is false or not. The specific model framework is shown in the figure Figure 1 As shown in Figure 2, the wide view feature extractor framework is as follows: Figure 2 As shown in Figure 2, the propagation structure feature extractor framework is as follows: Figure 3 shown.

[0029] The beneficial effects of the present invention are as follows:

[0030] 1) It proposes a method of analyzing news from a broad perspective, combining the macro and micro backgrounds of the news for observation.

[0031] 2) Extract popularity-oriented background features from the macro level, extract novelty-oriented background features from the micro level, and then derive breadth perspective features.

[0032] 3) Using the time graph snapshot method, the temporal dynamics of fake news propagation are taken into account, and then combined with the graph convolutional neural network and attention mechanism to extract the temporal information features and structural information features of fake news. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0034] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0035] Figure 2 It is a detailed schematic diagram of the wide viewing angle feature extractor of the present invention.

[0036] Figure 3 It is a detailed schematic diagram of the propagation structure feature extractor of the present invention.

[0037] Figure 4 These are the statistical data of the two data sets used in the experiments of the present invention.

[0038] Figure 5The performance comparison chart of the present invention and other methods on PHEME and WEIBO datasets.

[0039] Figure 6 It is a performance comparison chart of different variants of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The present invention proposes a false news detection method based on a breadth perspective. The main process of the method is as follows: Figure 1 The specific implementation process is as follows:

[0042] 1) Given an original post p, we first observe its news context from a breadth perspective and construct the macro context and micro context (macEnv and micEnv) using the news data closest to the time when p was published.

[0043] 2) Model the macro background and micro background respectively and generate the macro background vector v p,mac and microscopic background vector v p,mic .

[0044] 3) The macro background vector v p,mac and microscopic background vector v p,mic Splicing, we get the wide view feature vector V(p).

[0045] 4) Create a time graph snapshot containing text content and comment content for each post, which contains graph snapshots at each time step.

[0046] 5) For the graph snapshot at each time step, a graph convolutional neural network is used to extract the propagation structure feature Gn of each time step.

[0047] 6) For the propagation structure feature Gn of each time step, the attention mechanism is used to learn the attention weights and aggregate the weighted inputs to obtain the propagation structure feature G(p) that fully considers the temporal dynamics.

[0048] 7) Use the pre-trained model BERT to extract the semantic feature vector of the text content of the original post p and generate the text embedding representation BERT(p).

[0049] 8) Concatenate V(P) with broad perspective features, G(p) with propagation structure features, and BERT(p) with news text semantic features to obtain the final feature vector F.

[0050] 9) Input the feature vector F into SoftMax to classify the news as true or false.

[0051] Specifically, the step 1) includes:

[0052] Let E be the set of all collected news that were published earlier than the original post p, and the macro context macEnv is the set of news in E that were published T days before p was published:

[0053] ε mac ={e:e∈ε,0 <t p -t e ≤T}

[0054] Among them, t p and t e Represent the release dates of p and news e respectively.

[0055] Microscopic background micEnv is ε mac The set of news items related to p in . Specifically, use p and ε mac Perform similarity query on each news item in the , and take the first k most similar news as the set of micro-background:

[0056] ε mic ={e:e∈Topk(p,ε mac )}

[0057] in r∈(0,1), r determines the scale.

[0058] The macro context macEnv obtained by time restriction provides a macro perspective of how the general public reads and pays attention to recent events, while the micro context micEnv obtained by further relevance restriction describes the news distribution of similar events. The pre-trained language model BERT is used to obtain news representation. For each news e in the macro context, micro context, and original post p, the [CLS] output of BERT is used as its initial representation:

[0059] p=Bert(p),e=Bert(e)

[0060] Further, step 2) comprises:

[0061] Fake news often chooses to take advantage of hot news for greater exposure, and at the same time adds some novel information to attract readers' attention. Therefore, popularity and novelty are two very important indicators in the creation of fake news. From the perspective of macro context, if many items in macEnv are similar to p, then p may also be popular in such an environment. In other words, in macEnv, it is necessary to extract popularity-oriented macro context features.

[0062] First calculate p and ε mac The cosine similarity of each news in :

[0063]

[0064] After that, considering ε mac The length of the calculated similarity list is variable. The effect will not be good in the network with fixed-dimensional vector as input. Therefore, after calculating the similarity, Gaussian Kernel Pooling method is used to obtain soft count values ​​based on the cosine similarity range. Assuming that C kernels are used Then the output of the kth kernel is:

[0065]

[0066] where μ k and σ k are the mean and width of the kth kernel.

[0067] Finally, the outputs of all kernels are concatenated and the results are normalized to obtain the C-dimensional similarity feature of macEnv:

[0068]

[0069] Where ⊕ is the concatenation operator and Norm(·) represents normalization.

[0070] By calculating K(p,E mac ), the soft distribution of similarity between p and macEnv is obtained as the perceptual information of popularity. In order to enrich the perceptual information, the macEnv macro background vector is generated for p by fusing similarity and semantic information. Specifically, a multi-layer perceptron (MLP) is used to aggregate the central vector Avg(ε mac ) (obtained by averaging all vectors) and similarity features to obtain the final macroscopic background perception vector:

[0071]

[0072] Unlike macEnv, micEnv includes mainstream news items close to p, which indicates that they are likely to point to similar events. However, even in a hot event, a post will not attract people's attention if it is too similar to other news. What this type of news needs to do is to be different from other news, that is, novelty. If the content of a post is novel, it will be considered an outlier in the micro context. Therefore, it is necessary to extract micro context information guided by novelty.

[0073] Use the center vector Avg(ε of micEnv) mic ) as a reference, calculate the similarity feature K(p,ε mic ) and K(Avg(ε mic ),ε mic ), the latter serves as a reference for the former, which facilitates the model to calibrate its perception, and finally obtains the microscopic background perception vector:

[0074]

[0075] u sim =MLP(g(K(p,ε mic ),K(Avg(ε mic ),ε mic )))

[0076]

[0077] The comparison function ⊙ is the Hadamard product operator. sem and u sim Aggregate semantic and similarity information separately. MLPs are parameterized separately.

[0078] Further, step 4) includes:

[0079] In order to capture the evolution law of fake news diffusion, the present invention constructs a series of graph snapshots to simultaneously consider the temporal information and spatial structure information of the graph structure.

[0080] Specifically, in order to construct a time graph snapshot, the present invention divides the post into several stages at fixed time intervals and generates n graph snapshot sequences {G p1 ,G p2 ,...,G pn}, where G pi = <V pi ,E pi > represents the graph snapshot at the i-th time stage. At each time graph snapshot G pi In the example, posts are connected to their corresponding comment nodes according to the time order of the comments, thus generating the adjacency matrix A.pi ∈{0,1} m×m At the time graph snapshot G pi For each node, word embedding is used as its initial feature vector to generate the node’s feature matrix.

[0081] After generating the graph snapshot, the present invention uses the graph convolutional network GCN to perform representation learning on the graph snapshot to obtain the propagation structure feature Gn of each time step.

[0082] After the above steps 1) to 9), for a certain news post, we can get its prediction label whether it is false or not. The specific model framework is shown in the figure Figure 1 As shown in Figure 2, the wide view feature extractor framework is as follows: Figure 2 As shown in Figure 2, the propagation structure feature extractor framework is as follows: Figure 3 shown.

[0083] Experimental verification

[0084] 1. Dataset

[0085] According to the fake news detection method proposed above, the present invention uses two Chinese and English public data sets, PHEME and WEIBO, to conduct relevant experiments to verify the performance of the model. These two data sets are from Twitter and Weibo respectively. The PHEME data set is collected based on 5 breaking news, each of which contains a set of statements. The WEIBO data set is collected based on news reported on www.weibo.com, where each news contains text, image URL, and comments or forwarding, etc. Since the present invention does not use pictures, only text, comments or forwarding are extracted as the data set of the present invention. Figure 4 Statistics for both datasets are shown.

[0086] 2. Implementation details

[0087] The present invention uses a split ratio of 7:1:2 to divide the data set into a training set, a validation set, and a test set. For the BERT model used in step 8), for the English data set PHEME, the present invention uses bert-base-uncased, and for the Chinese data set WEIBO, the present invention uses bert-base-chinese. Such a choice can enable the model to have the ability to effectively process data in different languages. For the evaluation of model performance, the present invention uses accuracy, precision, recall rate, and F1 value as evaluation indicators. The initial learning rate of the experiment is 0.001, 5 epochs are used, batch_size is set to 32, and the Adam algorithm is used to optimize the parameters of the model.

[0088] 3. Comparative experiment

[0089] The present invention compares the proposed method with some representative models and names the proposed method as BP.

[0090] (1) Transformer: It uses an encoder-decoder structure, which can process the relationship between different positions in the input sequence and is suitable for processing various sequence data processing tasks. In the comparative experiment, the present invention uses its encoder to encode the post text content and then input it into the fully connected layer for classification prediction.

[0091] (2) BERT: It is a pre-trained natural language processing (NLP) model based on the Transformer architecture. It also considers the contextual information in the text and can better understand the context of the text. In the comparative experiment, the present invention inputs the post text content into BERT and encodes it into a rich word embedding representation, which is then passed to the fully connected layer for classification prediction.

[0092] (3) STS-NN: A spatiotemporal structured neural network for rumor detection that considers spatial and temporal structures as a whole to model message propagation.

[0093] (4) BiGCN: It uses a bidirectional graph convolutional network to explore the propagation and diffusion characteristics of rumors through top-down and bottom-up directed graphs.

[0094] (5) RumorGAN: It adopts the idea of ​​adversarial learning and uses the generator to generate uncertain or conflicting content to complicate the original posts and comments to improve the learning ability of the discriminator.

[0095] (6) KMGCN-v: Combine text information, knowledge concepts and image information to establish a semantic representation model. In the comparative experiment, since the model of the present invention does not use image information, a variant of KMGCN, KMGCN-v, is used, which removes the image information part.

[0096] For fair comparison, this paper uses commonly used evaluation indicators: Accuracy, Precision, Recall and F1 value as unified evaluation indicators to obtain different performances of different models on PHEME and WEIBO datasets. Figure 5 Observation Figure 5It can be found from the data that compared with the models with good classification effects in previous studies, the BP proposed in this invention outperforms other models in all indicators. Its performance on the PHEME dataset and WEIBO dataset was improved by 5.1% and 5% on average, respectively, which proves the effectiveness of BP in fake news detection and can more comprehensively capture a variety of semantic information, including macro and micro background information, time information, structural information, post text features, etc., so as to better improve the detection effect.

[0097] 4. Ablation experiment

[0098] Since the BP proposed in this invention contains multiple key components, in this section, the inventors will compare the variants of BP from the following aspects to demonstrate the effectiveness of BP:

[0099] (1) BP-mac: remove the macro background vector v p,max .

[0100] (2) BP-mic: Remove the microscopic background vector v p,mic .

[0101] Figure 6 The effects of various variants of BP on the PHEME and WEIBO datasets are shown.

[0102] from Figure 6 It is not difficult to see that the performance of BP is significantly better than BP-mac and BP-mic, which shows the effectiveness of each sub-module on the overall BP model. In addition, it can be observed that when any of the environments is removed, the performance of F1 decreases, indicating that both environments are necessary in detection and play a complementary role.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the embodiments of the present invention have been shown and described, it is understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A false news detection method based on breadth perspective, characterized in that: The following steps are involved: 1) Given an original post p, first observe its news background from a breadth perspective, and use the news data closest to the time when p was published to construct the macro background and micro background (macEnv and micEnv). 2) Model the macro background and micro background respectively to generate the macro background vector v p,mac and microscopic background vector v p,mic . 3) The macro background vector v p,mac and microscopic background vector v p,mic Splicing to obtain the breadth perspective feature vector V(p). 4) For each post, establish a time graph snapshot containing text content and comment content, which contains graph snapshots of each time step. 5) For each time step graph snapshot, use Graph Convolutional Networks (GCN) to extract the propagation structure feature Gn of each time step. 6) For the propagation structure feature Gn of each time step, use the attention mechanism to learn the attention weight, and aggregate the weighted input to obtain the propagation structure feature G(p) that fully considers the temporal dynamics. 7) Use the pre-trained model BERT to extract the semantic feature vector of the text content of the original post p, and generate the text embedding representation BERT(p). 8) Splice V(P) with breadth perspective features, G(p) with propagation structure features, and BERT(p) with news text semantic features to obtain the final feature vector F. 9) Input the feature vector F into SoftMax to classify the news as true or false.

2. The false news detection method based on breadth perspective according to claim 1, characterized in that: The step 1) includes: the macro context macEnv is a set of news items published within T days before the original post p is published. The micro context micEnv is a set of news items related to p in the macro context. Specifically, a similarity query is performed using p and each news item in , and the top k most similar news items are taken as the set of micro contexts.

3. The false news detection method based on breadth perspective according to claim 1, characterized in that: The step 2) includes: in the macro background macEnv, it is necessary to extract the macro background features guided by popularity. First, the cosine similarity between the original post p and each news in macEnv is calculated. After calculating the similarity, the Gaussian kernel pooling method is used to obtain the soft count value based on the cosine similarity range. Finally, the outputs of all kernels are connected in series, and the results are normalized to obtain the soft distribution of the similarity between p and macEnv as the perception information of popularity. In the micro background micEnv, it is necessary to extract the micro background information guided by novelty. Using the central vector of micEnv as a reference, the similarity features of the central vector and other news in micEnv, as well as the similarity features of the original post p and other news in micEnv are calculated. The latter is used as a reference for the former to facilitate the model to calibrate its perception, and finally the micro background perception vector is obtained.

4. The false news detection method based on breadth perspective according to claim 1, characterized in that: The step 4) includes: the present invention divides the posts into several stages at fixed time intervals, generates n graph snapshot sequences, uses the graph convolutional network GCN to perform representation learning on the time graph snapshot of each time step, and obtains the propagation structure feature Gn of each time step.

Images