Explainable fake news detection method based on graph evidence

Abstract

The application discloses a false news detection method based on graph evidence interpretability, which comprises the following steps: collecting various news; expressing the false news detection task as a binary classification problem; constructing an evidence graph for the news in an offline manner; collecting the representation of historical news, and establishing positive and negative example evidence graphs; giving test news, and forwarding it to a multi-view feature extraction and learning module to obtain multi-view features; obtaining a matching score by referencing the evidence graph; then combining the multi-view features and the matching score and using them for prediction; searching for the predicted explanation by referencing the corresponding evidence graph; given a piece of news to be detected, comparing the news with the evidence graph to generate a matching score and related graph evidence, and predicting the news as false news or true news. The application utilizes the evidence graph constructed by historical labeled data; a new method for constructing an evidence graph from existing labeled data is proposed, which can be used to provide guidance and provide explicit explanations when new messages appear.

Description

Technical Field

[0001] This invention belongs to the field of Internet information processing technology, and in particular relates to an interpretable fake news detection method based on graph evidence. Background Technology

[0002] With the rapid development of electronic technology, multimedia platforms, and social media, people now rely more on social media applications on their smartphones to get news than on traditional television and radio. On the one hand, this shift allows people to access information faster and more efficiently. On the other hand, it also facilitates the generation and spread of fake news or rumors.

[0003] Currently established fact-checking platforms, such as PolitiFact and Snopes, primarily focus on verifying the veracity of claims. Manually labeling errors on these platforms requires significant manpower, making it difficult to meet the demands of real-time fake news detection, often resulting in widespread dissemination before the information is even identified. Therefore, in recent years, automated fake news detection systems have garnered considerable research interest and attention.

[0004] Existing fake news detection methods can be broadly categorized into two types: context-based and content-based. The former models the task as a reasoning process, where external context, such as social or textual background, helps explore the accuracy of a given statement. Social context refers to user profiles, social activities, and dissemination networks, while textual context typically consists of additional factual sources retrieved from knowledge graphs or fact-checking websites. While context-based methods offer initial interpretability due to external resources, this context is not always available. Therefore, an increasing number of methods are adopting a content-based paradigm, which focuses on using news content without assuming explicit availability of external resources. They extract useful features from news content, such as textual patterns, semantic information, and sentiment signals, and feed them into neural models to predict veracity. However, these content-based methods lack interpretability, and the results may be unreliable. Although some have attempted to mitigate this problem by using attention mechanisms to find key parts of news content that can substantiate the results, interpretability remains weak, and in some cases, the interpretation can be confusing.

[0005] Furthermore, research indicates that fake news tends to reappear over time. News previously flagged as fake often resurfaces after a period of time. However, existing fake news detection methods fail to fully leverage this recursion for predictive purposes because they employ deep neural networks, which suffer from catastrophic forgetting due to the constant influx of new data into the detection model. Summary of the Invention

[0006] In view of this, to address the aforementioned problems, this invention proposes an interpretable fake news detection framework, IKA, based on graph evidence. This invention is content-based because it does not require external resources. However, inspired by the recurring nature of fake news, it collects evidence (i.e., labeled data used to train the model) from historical news to help predict the veracity of new claims. Specifically, positive and negative example evidence graphs are constructed from the labeled data. When a news item to be detected is encountered, the news is compared with the evidence graph, and relevant evidence is searched within the graph. This not only calculates a matching vector predicting the news's veracity but, more importantly, provides a clear explanation for the prediction, thereby improving the model's interpretability. Similar to existing content-based methods, this invention extracts multi-view features from the news content, learns and combines feature representations, and predicts veracity scores based on these representations. Furthermore, this invention integrates matching vectors for prediction. Notably, the method selection for the multi-view feature learning module in this invention is flexible and can be applied to existing methods to improve their interpretability.

[0007] To achieve the above objectives, the present invention discloses a method for detecting interpretable fake news based on graph evidence, comprising the following steps:

[0008] Collect various news items;

[0009] The task of detecting fake news is formulated as a binary classification problem, which divides news on social media into fake news and real news.

[0010] Construct evidence graphs from news reports offline; collect representations of historical news reports, i.e., training data, to build positive and negative example evidence graphs;

[0011] The test news is given and forwarded to the multi-view feature extraction and learning module to obtain multi-view features; the matching score is obtained by referencing the evidence map; then the multi-view features and the matching score are combined and used for prediction; the interpretation of the prediction is retrieved by referencing the corresponding evidence map.

[0012] Given a news item to be detected, the news item is compared with an evidence graph to generate a matching score and relevant graph evidence, and the news item is predicted to be either fake or real news.

[0013] Furthermore, each news item contains content and a set of related comments; each news item is assigned a ground-truth label y∈{0,1}, where 1 or 0 represents whether the news is fake or real, respectively; given a set of training data, the task is to predict the authenticity of news items in the test set.

[0014] Furthermore, preprocessing is performed through graph construction, specifically including:

[0015] A fixed-size sliding window is used to identify the connectivity between each word in the graph. The center word of each window is connected to the remaining words in that window. If they are connected, the corresponding entry in the adjacency matrix is ​​1; otherwise, it is 0. It captures the local context around the center word.

[0016] Furthermore, all identical words are merged into a single node in the graph, which explicitly aggregates their local context; thus, several related fragments placed far apart in a news clip become closer on the graph and are then explored further using higher-order message propagation.

[0017] Furthermore, the multi-view features are extracted using the following method:

[0018] For each news item, based on the constructed graph, a graph neural network is used as an encoder to extract long-distance semantic relationships; specifically, a graph-gated neural network is used to propagate contextual information to adaptively balance node features and neighbor information; this process is represented as:

[0019]

[0020] z i =σ(W z a i +U z H i +b z ),

[0021] r i =σ(W r a i +U r H i +b r ),

[0022]

[0023]

[0024] In the formula Let C be the adjacency matrix normalized by Laplace operation, and W be the edge set. * U * and b * These are trainable parameters that control the ratio of neighborhood information to node features. σ is a non-linear activation function. After obtaining the node representations, average pooling is used to generate the semantic representation r of the news. sem :

[0025]

[0026] Style features are extracted from the news articles, denoted as S = {s1,...,s8}, including eight high-level features. Based on the stylistic features, a writing style representation r is generated using a multilayer perceptron (MLP). sty The formula is:

[0027] r sty =MLP({s1,...,s8});

[0028] Extract publisher sentiment from the content text, including: sentiment category, sentiment vocabulary, sentiment intensity, sentiment score, and auxiliary features; concatenate all of these features to obtain E. Pub On the other hand, socio-emotional features are extracted from comments on news articles, denoted as E. soc ={e1,...,e |E| Therefore, linking these features together yields a dual emotional characteristic:

[0029]

[0030] in E represents the concatenation of vectors. gap This represents the difference between the publisher's sentiment and social sentiment; then, the dual sentiment features are forwarded to another MLP to obtain the representation:

[0031] r emo =MLP(E dual ).

[0032] Furthermore, the evidence diagram is enhanced to provide interpretability and improve effectiveness, specifically including:

[0033] Extract words from three categories: verbs related to What, Why, and How, and expressions related to them.

[0034] W 3 ord=F(NER(p i ))

[0035] Where F(·) represents the filtering function, excluding words that do not belong to Who, When, and Where; NER represents an open-source named entity recognition tool;

[0036] Establishing and preserving evidence graphs of historical news data also facilitates connections between discrete news fragments; specifically, a positive example evidence graph G is constructed using authentic historical news. + A negative example evidence graph G was constructed using fake historical news. - ;

[0037] Use positive example evidence diagrams to refine negative example evidence diagrams to exclude facts that appear in the positive example evidence diagrams, and then combine these diagrams into a negative example evidence diagram.

[0038] Furthermore, the steps for constructing the positive example evidence diagram are as follows:

[0039] For each article marked as For authentic news articles, first extract the corresponding image g. i ;

[0040] These graphs are combined to form a graph of positive examples:

[0041] G + =f merg (g1,...,g i ,...g n )

[0042] Among them, (g1,...,g i ,...g n f represents the set of graphs constructed from positively labeled news items. merg (·) represents the merge function, which is implemented by using W 3 ord is used as a pivot to combine diagrams;

[0043] After obtaining the positive example evidence graph, a representation of each node is generated using a standard GCN.

[0044] Furthermore, the steps for constructing the positive example evidence diagram are as follows:

[0045] For each article marked as Extracting the corresponding images from fake news articles.

[0046] The negative example evidence diagram is refined to filter out non-error information.

[0047] Furthermore, the negative example evidence diagram is refined using the positive example evidence diagram, including the following steps:

[0048] Generate W 3 orf pairs: for each marked as The fake news, first find W 3 ord = set(w1, w2, ..., w m Then, pair adjacent words to obtain m-1 pairs: (w1, w2)(w2, w3)...(w m-1 ,w m );

[0049] Get W 3 ord pair embedding: For each negative graph, a regular GCN is used to generate a representation for each node; then, for each pair (w s ,w tThe embedding of each word is obtained from the graph representation, and then these embeddings are concatenated to obtain the embedding Sem(w). s ,w t ):

[0050]

[0051] Among them, H(w) i ) and H(w j ) represents the generated node representation; Indicates the connection of vectors;

[0052] Introducing a positive example evidence graph: for each pair of W in the negative graph 3 ord(w s ,w t Find the pair in the positive example evidence diagram, denoted as (w s ,w t ) + And it is represented by GCN embedding as Sem(w s ,w t ) + ;

[0053] Calculate similarity: Use cosine similarity to calculate W for each pair. 3 ord(w s ,w t The similarity between the pairs and their corresponding pairs in the positive example evidence diagram:

[0054] Sim(w s ,w t )=cos(Sem(w s ,w t ),Sem(w s ,w t )+)

[0055] Filtering nodes in the negative graph: Based on similarity, if Sim > α, filter relevant nodes and remove the negative graph g. j Chinese w s and w t The two-hop word nodes are then used to obtain the refined negative graph g′. j ;

[0056] Merge and refine the negative graphs to obtain negative example evidence graphs, and finally merge identical W graphs. 3 From the ord, we obtain the refined negative graph, resulting in the negative example evidence graph G. - .

[0057] Furthermore, evidence-enhanced predictions of news include:

[0058] Given a news item in the test set, create a graph;

[0059] Then, a standard GCN is used to generate the graph representation;

[0060] Extract a set of W from the news 3 ord pairs, for each pair (w m ,w n Find the positive and negative example evidence images and obtain two similarity scores:

[0061] Sim(w m ,w n ) + =cos(Sem(w m ,w n ),Sem(w m ,w n )+)

[0062] Sim(w m ,w n ) - =cos(Sem(w m ,w n ),Sem(w m ,w n )-)

[0063] Among them, Sim(w m ,w n ) + The similarity between the test news and the positive example evidence graph is represented by Sem(w). m ,w n ) - Characterize the similarity between test news and negative example evidence graphs;

[0064] Finally, collect all (w m ,w n The scores are then used to rank them based on their similarity values.

[0065] top-k Sim + Values ​​and top-k Sim - The values ​​are concatenated to form the evidence graph enhancement matching vector r. mat ;

[0066] After obtaining the multi-view representation and evidence-aware matching vectors, they are fused with learnable parameters:

[0067] r = a sem r sem +a sty r sty +a emo r emo +a mat r mat

[0068] Among them, a sem a sty a emo a mat These represent the importance of the features, and are learnable parameters; the most...

[0069] The final representation is forwarded to an MLP to determine the probability y^ that the generated news clip is false.

[0070]

[0071] Furthermore, using binary cross-entropy loss as the objective function, a loss function is designed to minimize the BCE value:

[0072]

[0073] Where y∈{0,1} is the actual label. For predicted labels.

[0074] The beneficial effects of this invention are as follows:

[0075] An interpretable method for detecting fake news is proposed by constructing an evidence graph using historical labeled data. This invention offers flexibility in the selection of feature learning modules, allowing it to be applied to existing fake news detection methods to improve their interpretability.

[0076] A novel method for constructing evidence graphs from existing labeled data is proposed, which can be used to provide guidance and explicit interpretation when new messages emerge.

[0077] This invention not only outperforms strong baselines in prediction accuracy on English and Chinese real-world datasets, but also provides a compelling explanation through evidence graphs. Attached Figure Description

[0078] Figure 1 A schematic diagram of the present invention. Detailed Implementation

[0079] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0080] This invention proposes an IKA framework for fake news detection, which fully leverages the recurring patterns of similar or related news stories. IKA constructs an evidence graph based on historically labeled news and designs efficient semantic matching to provide corresponding graph evidence for fake news detection.

[0081] Problem Definition. The fake news detection task can be formulated as a binary classification problem, dividing news on social media into fake news and real news. Each news item contains content and a set of related comments. Each news item is assigned a ground-truth label y∈{0,1}, where 1 or 0 represents whether the news is fake or real, respectively. Given a set of training data, the task aims to predict the authenticity of news items in the test set.

[0082] Model outline. (e.g.) Figure 1 As shown, IKA mainly consists of two main components: a multi-view feature extraction and learning module that obtains multi-view features from news content and generates feature representations, and an evidence graph-enhanced prediction module that predicts authenticity and provides corresponding explanations.

[0083] First, an evidence map is constructed offline. Then, a test news item is given and forwarded to the multi-view feature extraction and learning module to obtain multi-view features, which are then forwarded to the evidence map enhancement prediction module. A matching score is obtained by referencing the evidence map. The multi-view features and matching score are then combined for prediction. The interpretation of the prediction can be retrieved by referencing the corresponding evidence map.

[0084] Graph Construction Strategy. Existing research has shown that, compared to text, extracting graphs from text can capture long-distance dependencies and cross-segment, cross-domain knowledge of useful information. Therefore, a fixed-size sliding window is used to identify connectivity between each word in the graph. The center word of each window is connected to the remaining words in that window (if connected, the corresponding entry in the adjacency matrix is ​​1; otherwise, it is 0), capturing the local context around the center word. Furthermore, all identical words are merged into a single node in the graph, which explicitly aggregates their local context. Thus, several related segments placed far apart in a news segment, or even distributed across different segments, become close on the graph, allowing for higher-order message propagation on the graph. Next, IKA is described in detail.

[0085] Multi-view feature learning

[0086] Existing research has indicated significant differences between fake news and genuine news in terms of word choice, writing style, and emotional cues. To effectively utilize these features, this invention designs a multi-view feature extraction and learning framework.

[0087] Graph-based semantic encoders

[0088] For each news item, based on the constructed graph, graph neural networks (GNNs) are used as encoders to extract long-distance semantic relationships. Specifically, graph gated neural networks (GGNNs) are used to propagate contextual information because the goal of this invention is to adaptively balance node features and neighbor information. Formally, this process can be represented as:

[0089]

[0090] In the formula Let C be the adjacency matrix normalized by Laplace operation, and W be the edge set. * U * and b * σ is a trainable parameter that controls the ratio of neighborhood information to node features. σ is a non-linear activation function; here, the sigmoid function is chosen. After obtaining the node representations, average pooling is used to generate the semantic representation r of the news. sem :

[0091]

[0092] Writing Style Encoder

[0093] News writing style encompasses readability, logic, credibility, form, interactivity, interest, emotional appeal, and completeness, all of which are related to the quality and authenticity of the news. Therefore, style features are extracted from news, denoted as S = {s1,…,s8}, including eight high-level characteristics. Based on these stylistic features, a multilayer perceptron (MLP) is used to generate a writing style representation r. sty The formula is:

[0094] r sty =MLP({s1,…,s8}) (3)

[0095] Emotional Clue Encoder

[0096] Sentiment signals are important for identifying fake news. Publisher sentiment is extracted from the content text, including: sentiment category, sentiment vocabulary, sentiment intensity, sentiment score, and auxiliary features. Concatenating all these features yields E. Pub On the other hand, socio-emotional features are extracted from comments on news articles, denoted as E. soc ={e1,…,e |E| Therefore, concatenating these features yields a dual emotional characteristic:

[0097]

[0098] in E represents the concatenation of vectors. gap This represents the difference between the publisher's sentiment and social sentiment. The dual sentiment features are then forwarded to another MLP to obtain the representation:

[0099] r emo =MLP(E dual (5)

[0100] Evidence graph enhancement prediction module

[0101] After obtaining multi-view features, most methods directly input them into the predictor for prediction. However, this paradigm cannot verify the prediction results. Therefore, this invention proposes an evidence graph-enhanced prediction module to provide interpretability and improve performance.

[0102] Extract W 3 ord

[0103] News can be seen as a description or report of certain events. To judge the authenticity of news, the first step is to see if the reported events are true and accurate. The key elements of a news report are closely related to the elements describing the event, often referred to as the 5W1H: Who, What, When, Where, Why, and How. When performing fast query matching, the focus is primarily on selecting keywords related to Who, When, and Where. Verbs and expressions related to What, Why, and How often have too many variations, which can generate a large number of matches and complicate the query. Therefore, extracting these three types of words and obtaining W... 3 ord = set(w1, w2, ..., w m ):

[0104] W 3 ord=F(NER(p i (6)

[0105] Where F(·) represents a filtering function that excludes words that do not belong to Who, When, and Where. NER stands for open-source named entity recognition tools, such as NLTK and StanfordNER.

[0106] Evidence diagram construction

[0107] Establishing and preserving evidence graphs of historical news data also facilitates connections between discrete news fragments. Specifically, a positive example evidence graph G is constructed using authentic historical news. + A negative example evidence graph G- was constructed using fake historical news. First, the method for constructing a positive example evidence graph is explained in detail.

[0108] Construction of positive example evidence diagrams. For each example marked as... For authentic news articles, first extract the corresponding image g. i Then, these graphs are merged to form a graph of positive example evidence:

[0109] G + =f merg (g1,...,g i ,…g n (7)

[0110] Among them, (g1,...,g i,...g n f represents the set of graphs constructed from positively labeled news items. merg (·) represents the merge function, which is implemented by using W 3 ord is used as a pivot to combine graphs. Compared with merge graphs based on the same word, this method can reduce the size and complexity of the positive evidence graph, thus facilitating the subsequent matching process. The construction process is described in Algorithm 1. After obtaining the positive evidence graph, a representation of each node is generated using a regular GCN

[13] .

[0111] Algorithm 1: Construction of Positive Example Evidence Graph

[0112] Input: Marked as real news p +

[0113] Output: Positive example evidence graph G +

[0114] 1: Regarding

[0115] 2: Constructing subgraph g i

[0116] 3: Merge g i And the same W 3 ord to generate G + (See Formula 7)

[0117] Construction of negative example evidence diagrams. For each example labeled... To detect fake news articles, first extract their corresponding graphs using a graph construction strategy. Clearly, the knowledge in real news is correct. However, the knowledge in fake news is not entirely wrong. For example, some fake news also contains partially true information. Some people even exploit this to create misinformation, which can easily mislead readers and make it difficult for detection algorithms to identify. Therefore, considering that there is valid information in fake news, directly merging negative example evidence graphs into a negative example evidence graph using Formula 7 may introduce some real information. Therefore, it is necessary to refine the negative graph to filter out non-error information.

[0118] Therefore, this invention proposes to refine negative example evidence diagrams by drawing on positive example evidence diagrams. The main idea is to exclude facts appearing in the positive example evidence diagrams, and then combine these diagrams into a negative example evidence diagram. The negative example evidence diagram consists of the following steps:

[0119] 1) Generate W 3 ord pairs. For each tag as The fake news was first identified based on equation 6, which led to the discovery of W. 3 ord = set(w1, w2, ..., w mThen, pair adjacent words to obtain m-1 pairs: (w1, w2)(w2, w3)...(w m-1 ,w m ).

[0120] 2) Obtain W 3 ord pairs are embedded. For each negative graph, a regular GCN

[13] is used to generate a representation for each node (word). Then, for each pair (w s ,w t The embedding of each word is obtained from the graph representation, and then these embeddings are concatenated to obtain the embedding Sem(w). s ,w t ):

[0121]

[0122] Among them, H(w) i ) and H(w j ) represents the generated node representation; Represents the connection of vectors.

[0123] 3) Introduce a positive example evidence graph. For each pair of W in the negative graph... 3 ord(w s ,w t ), find the pair in the positive example evidence diagram (if it exists), denoted as (w s ,w t ) + And it is represented by GCN embedding as Sem(w s ,w t ) + .

[0124] 4) Calculate similarity. Use cosine similarity to calculate the similarity for each pair of W. 3 ord(w s ,w t The similarity between the pairs and their corresponding pairs in the positive example evidence diagram:

[0125] Sim(w s ,w t )=cos(Sem(w s ,w t ),Sem(w s ,w t ) + (9)

[0126] 5) Filter nodes in the negative graph. Based on similarity, if Sim > α, filter relevant nodes and remove the negative graph g. j Chinese w s and w tThe two-hop word nodes are then used to obtain the refined negative graph g′. j .

[0127] 6) Merge and refine the negative graphs to obtain the negative example evidence graph. Finally, according to Formula 7, merge identical W values. 3 From the ord, we obtain the refined negative graph, resulting in the negative example evidence graph G. - .

[0128] The construction process is described in Algorithm 2.

[0129] Algorithm 2: Construction of Negative Example Evidence Graph

[0130] Input: Marked as fake news p - Positive example evidence diagram G +

[0131] Output: Negative Example Evidence Graph G-

[0132] 1: Regarding

[0133] 2: Construct the negative graph g j

[0134] 3: From Extract W from 3 ord j

[0135] 4: Regarding W 3 ord j (w) s ,w t ):

[0136] 5: If (w s ,w t In G + middle:

[0137] 6: Calculate (w) s ,w t ) + With (w s ,w t Similarity (according to formula 9)

[0138] 7: If Sim(w s ,w t )>α:

[0139] 8: Filter g j Related word nodes

[0140] 9: Obtain the refined negative graph g′ j

[0141] 10: Use formula 7 to refine all negative graphs g′ j and the same W3 ord merge, generating G - .

[0142] Enhanced Evidence Prediction

[0143] Given a news item in the test set, create a graph. Then, use a standard GCN to generate the graph representation.

[0144] Next, use Formula 6 to extract a set of W from the news. 3 ord pairs, for each pair (w m ,w n Similar to steps (2) to (4) in the construction of the negative example evidence graph, both positive and negative example evidence graphs are searched simultaneously to obtain two similarity scores:

[0145] Sim(w m ,w n ) + =cos(Sem(w m ,w n ),Sem(w m ,w n ) + (10-1)

[0146] Sim(w m ,w n ) - =cos(Sem(w m ,w n ),Sem(w m ,w n ) - (10-2)

[0147] Among them, Sim(w m ,w n ) + The similarity between the test news and the positive example evidence graph is represented by Sem(w). m ,w n ) - Characterize the similarity between the test news and the negative example evidence graph. Finally, collect all (w m ,w n The scores of the top-k Sims are calculated, and then they are sorted according to their similarity scores. + Values ​​and top-k Sim - The values ​​are concatenated to form the evidence graph enhancement matching vector r. mat .

[0148] Feature fusion and prediction

[0149] After obtaining the multi-view representation and evidence-aware matching vectors, they are fused with simple learnable parameters:

[0150] r = a sem r sem +a sty r sty +a emo r emo +a mat r mat (11)

[0151] Among them, a sem a sty a emo a mat These represent the importance of the features and are learnable parameters. The final representation is then fed into an MLP to determine the probability that a generated news clip is fake.

[0152]

[0153] Training objectives

[0154] The goal of a fake news detector is to identify whether news is fake. Since it is essentially a classification task, the standard binary cross-entropy loss (BCE) is used as the objective function. The loss function is designed to minimize the BCE value:

[0155]

[0156] Where y∈{0,1} is the actual label. For predicted labels.

[0157] The interpretability of the model in this invention lies in generating matching vectors and corresponding explanatory fragments. For example... Figure 1 As shown, for a given news article, extract W. 3 The algorithm uses `ord` and then searches for highly similar pairs in the evidence graph. Therefore, after making a prediction, `W` is retrieved from the corresponding evidence graph. 3 The ord and its surrounding information are used as a graphic interpretation of the news.

[0158] The beneficial effects of this invention are as follows:

[0159] An interpretable method for detecting fake news is proposed by constructing an evidence graph using historical labeled data. This invention allows for flexible selection of the feature learning model and can be applied to other existing fake news detection methods to improve their interpretability.

[0160] A novel method for constructing evidence graphs from existing labeled data is proposed, which can be used to provide guidance and explicit interpretation when new messages emerge.

[0161] This invention not only outperforms comparative methods in prediction accuracy on real Chinese and English datasets, but also provides a convincing explanation through evidence graphs.

[0162] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0163] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component (e.g., is functionally equivalent to it), even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although specific features of this disclosure have been disclosed with respect to only one of several implementations, such features may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0164] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0165] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A method for detecting interpretable fake news based on graph evidence, characterized in that, Includes the following steps: Collect various news items; The task of detecting fake news is formulated as a binary classification problem, which divides news on social media into fake news and real news. Construct evidence graphs from news reports offline; collect representations of historical news reports, i.e., training data, to build positive and negative example evidence graphs; The test news is given and forwarded to the multi-view feature extraction and learning module to obtain multi-view features; the matching score is obtained by referencing the evidence map; then the multi-view features and the matching score are combined and used for prediction; the interpretation of the prediction is retrieved by referencing the corresponding evidence map. Given a news item to be detected, the news item is compared with an evidence graph to generate a matching score and relevant graph evidence, and the news item is predicted to be fake or real. Enhance the evidence diagram to provide interpretability and improve effectiveness, specifically including: Extract words from three categories: verbs related to What, Why, and How, and expressions related to them. in This represents a filtering function that excludes words that do not belong to Who, When, and Where. NER stands for Open Source Named Entity Recognition Tool. Establishing and preserving evidence graphs of historical news data also facilitates connections between discrete news fragments; specifically, a positive example evidence graph was constructed using authentic historical news. A negative example evidence diagram was constructed using fake historical news. ; Use positive example evidence diagrams to refine negative example evidence diagrams to exclude facts that appear in the positive example evidence diagrams, and then combine these diagrams into a negative example evidence diagram.

2. The method for detecting interpretable fake news based on graph evidence according to claim 1, characterized in that, Each news item includes content and a set of related comments; each news item is assigned a ground-truth tag. , where 1 or 0 represent that the news is fake or real, respectively; given a set of training data, the task is to predict the authenticity of news in the test set.

3. The method for detecting interpretable fake news based on graph evidence according to claim 2, characterized in that, Preprocessing is performed through graph construction, specifically including: A fixed-size sliding window is used to identify the connectivity between each word in the graph. The center word of each window is connected to the remaining words in that window. If they are connected, the corresponding entry in the adjacency matrix is ​​1; otherwise, it is 0. It captures the local context around the center word. Furthermore, all identical words are merged into a single node in the graph, which explicitly aggregates their local context; thus, several related fragments placed far apart in a news clip become closer on the graph and are then explored further using higher-order message propagation.

4. The method for detecting interpretable fake news based on graph evidence according to claim 3, characterized in that, The multi-view features are extracted using the following method: For each news item, a graph neural network is used as an encoder based on the constructed graph to extract long-distance semantic relationships; specifically, a graph gating neural network is used to propagate contextual information to adaptively balance node features and neighbor information. This process is represented as: In the formula The adjacency matrix is ​​normalized by the Laplace operation. Represents the edge set, , and These are trainable parameters that control the ratio of neighborhood information to node features. It is a non-linear activation function that, after obtaining node representations, uses average pooling to generate semantic representations of news. : ； Extracting style features from news articles, represented as... It includes eight advanced features; based on stylistic features, it uses a multilayer perceptron (MLP) to generate writing style representations. The formula is: ； Extract publisher sentiment from the content text, including: sentiment category, sentiment vocabulary, sentiment intensity, sentiment score, and auxiliary features; concatenate all of these features to obtain... On the other hand, social sentiment characteristics are extracted from the comments of news articles and represented as... Therefore, linking these features together yields a dual emotional characteristic: ； in This indicates the concatenation of vectors. This represents the difference between the publisher's sentiment and social sentiment; then, the dual sentiment features are forwarded to another MLP to obtain the representation: 。 5. The method for detecting interpretable fake news based on graph evidence according to claim 4, characterized in that, The steps to construct a positive example evidence diagram are as follows: For each article marked as For authentic news articles, first extract their corresponding images. ; These graphs are combined to form a graph of positive examples: in, This represents the set of graphs constructed from positively labeled news items. This indicates a merge function, which is implemented by using... Used as a pivot to assemble diagrams; After obtaining the positive example evidence graph, a representation of each node is generated using a standard GCN.

6. The method for detecting interpretable fake news based on graph evidence according to claim 5, characterized in that, The steps to construct a positive example evidence diagram are as follows: For each article marked as Extracting the corresponding images from fake news articles. ; The negative example evidence diagram is refined to filter out non-error information.

7. The method for detecting interpretable fake news based on graph evidence according to claim 6, characterized in that, To refine the negative example evidence diagram using the positive example evidence diagram, the following steps are included: generate Yes: For each label as Fake news, first find Then, pair adjacent words to obtain m-1 pairs: ... ; Get For embeddings: For each negative example evidence graph, a standard GCN is used to generate a representation for each node; then, for each pair... The embedding of each word is obtained from the graph representation, and then these embeddings are concatenated to obtain the embedding pairs. : in, and This represents the generated node representation; Indicates the connection of vectors; Introducing a positive example evidence diagram: For each pair in the negative example evidence diagram Find the pair in the positive example evidence diagram, and represent it as follows: And using GCN embedding representation as ; Calculate similarity: Use cosine similarity to calculate the similarity for each pair. And the similarity between the corresponding pairs in the positive example evidence diagram: Filtering nodes in the negative example evidence graph: Based on similarity, if Filter relevant nodes and remove negative example evidence graphs. middle and The two-hop word nodes are then used to obtain the refined negative example evidence graph. ; Negative Evidence Diagram After Merging and Refining Negative example evidence diagram Finally merge the same ones To the refined negative example evidence diagram Negative example evidence diagram .

8. The method for detecting interpretable fake news based on graph evidence according to claim 7, characterized in that, Evidence-enhanced predictions of news include: Given a news item in the test set, create a graph; Then, a standard GCN is used to generate the graph representation; Extract a set from the news Yes, for every pair By searching for positive and negative example evidence images, two similarity scores are obtained: in, Characterize the similarity between news articles and positive example evidence images. Characterize the similarity between test news and negative example evidence graphs; Finally, collect all The scores are then used to rank them based on their similarity values. top-k Values ​​and top-k The values ​​are concatenated to form an evidence graph-enhanced matching vector. ; After obtaining the multi-view representation and evidence graph-enhanced matching vectors, they are fused with learnable parameters: in, , , , The importance of the features are represented by the parameters, which are learnable. The final representation is then fed into an MLP to determine the probability that a generated news clip is fake. : 。 9. The method for detecting interpretable fake news based on graph evidence according to claim 8, characterized in that, Using binary cross-entropy loss as the objective function, design a loss function that minimizes the BCE value: in For actual labels, For predicted labels.

Images