False news detection method and system based on partition filtering network

Abstract

The invention discloses a false news detection method and system based on a partition filtering network, and the method comprises the steps: carrying out the fusion of news text features and comment text features, and obtaining event features; inputting the news text and the comment text into a dual-emotion acquisition module to obtain dual-emotion features; inputting the word vector of the comment text into a partition filtering network to obtain an initial vertical feature and an initial rumor feature; fusing the event feature with the initial vertical feature to obtain a corrected vertical feature; classifying the corrected vertical features to obtain a first classification result; fusing the initial rumor feature, the dual emotion feature and the corrected standing feature to obtain a corrected rumor feature; classifying the corrected rumor features to obtain a second classification result; and fusing the first classification result and the second classification result to obtain a final classification result.

Description

Technical Field

[0001] The present invention relates to the field of social media and deep learning technology, and in particular to a false news detection method and system based on a partition filtering network. Background Art

[0002] With the advent of the Web 2.0 era, social media has become popular in social life. People express their personal opinions on social media platforms, and at the same time, a large amount of false news and unverified information are spread wantonly in human society. As a type of social media, Weibo has become an important source of information in today's society with its huge user base and strong user stickiness.

[0003] Fake news detection based on social context usually uses interactive information of news in social networks for identification, such as comments, publishers, and forwarding. In recent years, researchers have improved the accuracy of fake news detection by modeling user preferences, user credibility, and sentiment information flows in articles to assist in fake news detection. According to research, rumor posts often evoke various, mainly controversial positions among participating users. Therefore, determining the stance on relevant posts may be related to the successful detection of rumors. Unlike sentiment analysis, which can be applied to a piece of text without a specific target, stance detection requires a given target for analysis to measure users' views on it, which are usually divided into support, opposition, and others. Compared with the joint sentiment fake news detection method, stance detection obtains people's views on news by extracting comment stances, which is more convincing.

[0004] In recent years, people have applied the idea of ​​multi-task learning to the field of fake news detection. For example, researchers focus on building a multi-task fake news detection model by combining comment sentiment, using LSTM, ESIM, and BERT models to improve the accuracy of fake news detection. However, in multi-task work, fake news detection and other tasks (such as sentiment analysis, topic classification, etc.) may use different features, and sharing features may lead to feature mismatch, thus affecting the effect of false information detection.

[0005] Some researchers used naive Bayes and SVM to build generative and discriminative models, and established a user stance detection model based on vocabulary selection. With the emergence of pre-trained models, traditional learning methods pale in comparison. Later, some scholars used the propagation structure information to extract and construct comment branches from the tree structure of tweets and their comments and enter the LSTM model. They performed stance detection on the output representation of each LSTM unit in the Branch-LSTM model, and used the overall output representation of the LSTM model to identify false information. However, the modeling of conversation threads and tree structures requires a lot of time and memory, and its complex design increases the amount of calculation. Summary of the invention

[0006] In order to solve the shortcomings of the prior art, the present invention provides a false news detection method and system based on a partition filtering network; by using the existing stance detection technology to mark the stance of the long text of the comment of the target news, the overall stance label of the comment text is obtained; the target news and the comment text are respectively input into the BERT model to extract word vectors, and the Bi-LSTM is used to extract the context features of the news and the comments, and the extracted news and comment context features are input into the multi-head attention mechanism to obtain event features about the news; the news text and the comment text are input into the dual emotion acquisition module to obtain the dual emotion features; and then the representation vector of the comment is passed through the filtering partition network to obtain the task-specific comment stance features and rumor features; the event features and the comment stance features are fused to obtain the final stance features, which are input into the stance classifier to obtain the stance results; finally, the final stance features, the rumor features of the comments and the dual emotion features are fused through low-rank tensors to obtain the rumor features which are input into the news classifier.

[0007] On the one hand, a false news detection method based on a partition filtering network is provided, including:

[0008] S101: Obtaining news to be detected and several comments corresponding to the news to be detected;

[0009] S102: Input the news to be detected and a number of comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false;

[0010] The trained fake news detection model is used to: vectorize the news to be detected to obtain the word vector of the news to be detected; perform feature extraction on the word vector of the news to be detected to obtain the news text features; vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; perform feature extraction on the word vector of the comment text to obtain the comment text features;

[0011] The news text features and comment text features are integrated to obtain event features. The news text and comment text are input into the dual emotion acquisition module to obtain dual emotion features. The word vector of the comment text is input into the partition filtering network to obtain the initial position features and initial rumor features.

[0012] The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result; the initial rumor features, the dual emotion features and the revised stance features are integrated to obtain the revised rumor features; the revised rumor features are classified to obtain the second classification result; the first classification result and the second classification result are integrated to obtain the final classification result.

[0013] On the other hand, a fake news detection system based on partition filtering network is provided, including:

[0014] An acquisition module is configured to: acquire the news to be detected and a number of comments corresponding to the news to be detected;

[0015] A classification module is configured to: input the news to be detected and a number of comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false;

[0016] Among them, the trained fake news detection model is used for:

[0017] Vectorize the news to be detected to obtain the word vector of the news to be detected; extract features from the word vector of the news to be detected to obtain the news text features;

[0018] Vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; extract features from the word vector of the comment text to obtain the features of the comment text;

[0019] The news text features and comment text features are integrated to obtain event features;

[0020] Input the news text and comment text into the dual emotion acquisition module to obtain the dual emotion features;

[0021] Input the word vector of the comment text into the partition filtering network to obtain the initial position feature and initial rumor feature;

[0022] The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result;

[0023] The initial rumor features, the dual emotion features and the revised stance features are integrated to obtain revised rumor features; the revised rumor features are classified to obtain the second classification result;

[0024] The first classification result and the second classification result are fused to obtain the final classification result.

[0025] On the other hand, there is also provided an electronic device, comprising:

[0026] a memory for non-transitory storage of computer-readable instructions; and

[0027] a processor for executing the computer readable instructions,

[0028] When the computer-readable instructions are executed by the processor, the method described in the first aspect is executed.

[0029] On the other hand, a storage medium is provided, which non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the instructions of the method described in the first aspect are executed.

[0030] On the other hand, a computer program product is provided, comprising a computer program, wherein the computer program is used to implement the method described in the first aspect when running on one or more processors.

[0031] The above technical solution has the following advantages or beneficial effects:

[0032] 1. The technical method provided by the present invention is based on the correlation between rumor detection and joint stance detection. The method proposes a joint framework to make full use of the role of user comment stance in false news detection. In addition, in view of the problem that traditional conversation threads and tree-structured models consume a lot of time and memory, the method extracts the overall stance of the comments by inputting long text comments into the model to assist in judging false news. Under the premise of ensuring accuracy, it saves memory and shortens running time.

[0033] 2. The technical solution provided by the present invention is based on the dual emotion modeling of news and comments. The method obtains the dual emotion by modeling news emotion, comment emotion and the similarities and differences between the two emotions, and adds it to the rumor detection model based on the partition filtering network as an enhancement.

[0034] 3. The technical solution provided by the present invention is based on the method of partition filtering network, which divides neurons into three partitions through a gating mechanism: position partition, rumor partition and shared partition, and then combines the single task partition and the shared partition to generate task-specific features, filter out irrelevant information stored in the opposite task partition, and solve the problem of insufficient feature interaction between tasks.

[0035] 4. The technical solution provided by the present invention adopts a low-rank tensor fusion mechanism to address the problems of insufficient feature fusion and latent splicing. It can fully integrate rumor features and stance features, explore the deep correlation between different features, and solve the problem of insufficient feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 A schematic diagram of a process flow provided by an embodiment of the present invention;

[0038] Figure 2A schematic diagram of the overall network architecture provided for an embodiment of the present invention.

[0039] Figure 3 A schematic diagram of the dual emotion acquisition module structure provided in an embodiment of the present invention.

[0040] Figure 4 A schematic diagram of a partition filtering network structure provided in an embodiment of the present invention.

[0041] Figure 5 A schematic diagram of a low-rank tensor fusion structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0043] Terminology explanation:

[0044] Target news: the text content in the news. Comment text: a long text composed of several comments on the target news.

[0045] Bidirectional Long Short-Term Memory Network (Bi-LSTM): LSTM is a special type of RNN that is used to solve the gradient vanishing or gradient exploding problems faced by traditional RNNs when dealing with long-term dependencies.

[0046] Position feature: a feature used to classify positions. Rumor feature: a feature used to identify fake news.

[0047] Attention mechanism: is a mechanism that allows the model to "focus on" different parts of the input as needed.

[0048] Vectorization: vectorized representation of text (which can be understood as converting text into digital representation for computer processing).

[0049] Concatenate: concatenate two or more vectors into one vector.

[0050] Low-rank tensor fusion: A feature fusion mechanism applied in multimodal fusion.

[0051] Embodiment 1

[0052] This embodiment provides a false news detection method based on a partition filtering network;

[0053] like Figure 1 As shown, the false news detection method based on the partition filtering network includes:

[0054] S101: Obtaining news to be detected and several comments corresponding to the news to be detected;

[0055] S102: Input the news to be detected and a number of comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false;

[0056] Among them, the trained fake news detection model is used for:

[0057] Vectorize the news to be detected to obtain the word vector of the news to be detected; extract features from the word vector of the news to be detected to obtain the news text features;

[0058] Vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; extract features from the word vector of the comment text to obtain the features of the comment text;

[0059] The news text features and comment text features are integrated to obtain event features;

[0060] Input the news text and comment text into the dual emotion acquisition module to obtain the dual emotion features;

[0061] Input the word vector of the comment text into the partition filtering network to obtain the initial position feature and initial rumor feature;

[0062] The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result;

[0063] The initial rumor features, the dual emotion features and the revised stance features are integrated to obtain revised rumor features; the revised rumor features are classified to obtain the second classification result;

[0064] The first classification result and the second classification result are fused to obtain the final classification result.

[0065] Furthermore, the S101: obtaining the news to be detected and several comments corresponding to the news to be detected includes: using a web crawler to obtain the news to be detected and several comments corresponding to the news to be detected, preprocessing the comment text to remove noise data, taking the first M characters of the news text as the news title, M=16, the news title is the target of stance detection, splicing the target of stance detection and the denoised comments into a long text, inputting the long text into the trained false news detection model, and obtaining the stance label of the news.

[0066] Exemplarily, the noise data includes: special characters, emoticons, repeated symbols, such as emoji and facial expressions; in addition to Chinese, the data also contains a small amount of Japanese, Korean and English.

[0067] Further, in S102, the news to be detected and a number of comments corresponding to the news to be detected are input into the trained false news detection model to obtain a classification result of whether the news to be detected is false, wherein the training process of the trained false news detection model includes:

[0068] Constructing a training set, wherein the training set is a target news with a known label of whether the news is fake news, and a number of comments corresponding to the target news;

[0069] The training set is input into the fake news detection model to train the fake news detection model. When the loss function value of the fake news detection model no longer decreases, or the number of iterations exceeds the set number, the training is stopped to obtain the trained fake news detection model.

[0070] Furthermore, if Figure 3 As shown, the trained fake news detection model includes:

[0071] A first BERT model and a second BERT model, wherein the input value of the first BERT model is a news text, and the input value of the second BERT model is a comment text corresponding to the news text; the output end of the first BERT model is connected to the input end of the first Bi-LSTM model; the output end of the second BERT model is connected to the input end of the second Bi-LSTM model; the output end of the second BERT model is also connected to the input end of the partition filtering network;

[0072] The first Bi-LSTM model is used to extract features from the word vectors of the news to obtain news text features; the second Bi-LSTM model is used to extract features from the word vectors of the comment text to obtain comment text features;

[0073] The output end of the first Bi-LSTM model and the output end of the second Bi-LSTM model are both connected to the input end of the multi-head attention mechanism layer, and the output end of the multi-head attention mechanism layer outputs event features;

[0074] The trained false news detection model further includes a dual emotion acquisition module, wherein the input end of the dual emotion acquisition module inputs the news text and the comment text respectively, and the output end of the dual emotion acquisition module outputs the dual emotion feature;

[0075] The output end of the partition filtering network outputs the initial position features and initial rumor features;

[0076] The trained false news detection model further includes an adder and a first classifier, wherein the adder fuses the event feature and the initial stance feature to obtain a revised stance feature; and the first classifier classifies the revised stance feature to obtain a stance classification result;

[0077] The trained false news detection model also includes a low-rank tensor fusion network and a second classifier, wherein the low-rank tensor fusion network fuses the initial rumor features, the dual emotion features and the revised stance features to obtain revised rumor features; the second classifier classifies the revised rumor features to obtain a classification result of whether the news is true or false;

[0078] The trained fake news detection model also includes: a weighted fusion unit, which performs weighted fusion on the first classification result and the second classification result to obtain a final detection result.

[0079] Furthermore, if Figure 2 As shown, the dual emotion acquisition module includes:

[0080] The dual emotion acquisition module consists of two parts: text emotion fine-grained extraction sub-module and dual emotion extraction sub-module.

[0081] The text sentiment fine-grained extraction submodule extracts sentiment category features, sentiment vocabulary features, sentiment intensity features, sentiment polarity features and additional sentiment features of news texts, and concatenates all extracted features to obtain news text fine-grained sentiment features;

[0082] The text sentiment fine-grained extraction submodule extracts sentiment category features, sentiment vocabulary features, sentiment intensity features, sentiment polarity features and additional sentiment features of the comment text, and concatenates all the extracted features to obtain the comment text fine-grained sentiment features;

[0083] The dual sentiment extraction submodule calculates the offset sentiment features based on the fine-grained sentiment features of the news text and the fine-grained sentiment features of the comment text, and then concatenates the fine-grained sentiment features of the news text, the fine-grained sentiment features of the comment text and the offset sentiment features to obtain the dual sentiment features.

[0084] Furthermore, if Figure 2 As shown, the partition filtering network includes:

[0085] The partition filter network consists of a partition filter encoder (PFE). The partition filter encoder is used to generate task-specific features. After being processed by the partition filter encoder, the task-specific features are assigned to the stance classification task unit and the rumor detection unit as inputs for stance detection and rumor prediction.

[0086] Furthermore, if Figure 4 As shown, the low-rank tensor fusion network includes:

[0087] Sequentially connected tensor representation layers and low-order weight decomposition layers;

[0088] The tensor representation layer performs outer product on the input corrected stance feature, initial rumor feature and dual sentiment feature to obtain a representation tensor;

[0089] Tensor Representations Layer: constructs multiple features into tensor representations in the form of outer products. Assume there are two input features Then their tensors can be expressed as This is the tensor product of two eigenvectors, forming a d a ×d b dimensional matrix.

[0090] Low-rank Weight Decomposition Layer: In order to reduce the complexity of the constructed tensor, low-rank tensor decomposition technology is used. Low-rank tensor decomposition includes CP decomposition (CANDECOMP / PARAFAC decomposition) and Tucker decomposition. Through these methods, the tensor is decomposed into several low-dimensional matrices, which greatly reduces the amount of calculation and storage requirements.

[0091] Furthermore, the news to be detected is vectorized to obtain the word vector of the news to be detected, including: using the BERT model to realize the vectorization of the news to be detected.

[0092] Furthermore, the BERT model is used to vectorize the news to be detected, including:

[0093] The word embedding vector, sentence block vector and position encoding vector are extracted from the news to be detected, and then the three extracted vectors are fused by adding them together to obtain the word vector of the news to be detected.

[0094] It should be understood that the word embedding vector (Token Embedding) maps each word or subword (token) in the input text to a vector space of fixed dimension. In the BERT model, this process is achieved by looking up the pre-trained word embedding matrix. For each token in the text, its corresponding subscript is found in the pre-established index dictionary, and the vector obtained by looking up this subscript in the lookup table is the token embedding of the token.

[0095] It should be understood that the sentence segment vector (Segment Embedding) is used to distinguish different sentences. When processing tokens, it is necessary to add the [CLS] tag at the beginning of the sentence and the [SEP] tag at the end of the sentence. And this embedding layer has only two vector representations, namely 0 and 1: the tokens of the first sentence in the sentence pair are all assigned 0, and the tokens of the second sentence are all assigned 1. If there is only one sentence in the input, then its segment embedding is all 0.

[0096] Since Transformer does not have the ability to obtain the sequence of the entire sentence like RNN (recurrent neural network), the BERT model adds a position embedding to the word at each position to better understand the sequence order.

[0097] For BERT, the key part is the Transformer encoder based on the self-attention mechanism. It mainly adjusts the weight coefficient matrix by the degree of association between words in the sentence to obtain the word representation vector, that is:

[0098]

[0099] Among them, Q, K, V are word vector matrices, d k is the embedding dimension. The multi-head attention mechanism projects Q, K, and V through multiple different linear transformations, and finally concatenates different attention results to obtain information in multiple spaces.

[0100] It should be understood that the position encoding vector (Position Embedding) is used to provide information about the position of the token in the sentence, because the Transformer architecture itself does not have the ability to process sequence order. BERT uses position encoding to capture the position information of the word in the sentence.

[0101] Use BERT's tokenizer to preprocess the text, including tokenization, adding special tags (such as [CLS] and [SEP], and truncating or padding to meet BERT's input length requirements; input the preprocessed text into the BERT model to obtain the representation vectors of news and comments.

[0102] Furthermore, the feature extraction of the word vectors of the news to be detected to obtain news text features includes: using a first Bi-LSTM model to implement feature extraction of the word vectors of the news to be detected.

[0103] Long short term memory (LSTM) is a type of recurrent neural network (RNN). LSTM adds three gate structures in the hidden layer h, namely forget gate, input gate, and output gate, and also adds a cell state. The cell state is used for information storage; the input gate indicates how much of the network input is saved to the cell state at the current moment; the forget gate needs to forget or discard some information, and its task is to accept a long-term memory (the output from the previous cell module) and decide which part to keep and forget; the output gate determines the output value based on the cell state.

[0104] LSTM selectively extracts features through a gating mechanism. The calculation process is:

[0105] f t =σ(W f x t +U f h t-1 +b f )

[0106] i t =σ(W o x t +U i h t-1 +b i )

[0107] o t =σ(W o x t +U o h t-1 +b o )

[0108]

[0109] The Bi-LSTM neural network structure model is divided into two independent LSTMs. The input sequence is input into the two LSTM neural networks in forward and reverse order for feature extraction. The word vector formed by concatenating the two output vectors (i.e. the extracted feature vectors) is used as the final feature expression of the word. The design concept of the Bi-LSTM model is to make the feature data obtained at time t have information between the past and the future at the same time, so as to obtain context information.

[0110] Furthermore, the comment text corresponding to the news to be detected is vectorized to obtain the word vector of the comment text, including: using the BERT model to vectorize the comment text corresponding to the news to be detected.

[0111] Furthermore, the extracting features from the word vectors of the comment text to obtain the comment text features includes: using a second Bi-LSTM model to extract features from the word vectors of the comment text to obtain the comment text features.

[0112] Furthermore, the news text features and the comment text features are fused to obtain event features, and a multi-head attention mechanism layer is used to achieve the fusion of the news text features and the comment text features to obtain event features.

[0113] It should be understood that the attention mechanism is a method that imitates the human visual and cognitive system, which allows the neural network to focus on relevant parts when processing input data. By introducing the attention mechanism, the neural network can automatically learn and selectively focus on important information in the input, improving the performance and generalization ability of the model. The specific steps of the attention mechanism include:

[0114] Given a query vector (query), a set of key vectors (keys), and a set of value vectors (values).

[0115] (1) Dot-Product similarity calculation: By calculating the dot product between the query vector query and the key vector keys, we get the similarity score between each query and all the keys. Then we scale these scores by dividing them by the square root of d. k , to prevent the dot product value from being too large, which would cause the gradient to disappear or explode.

[0116] (2) Mask optionality: The purpose is to fill the padding part with negative infinity, so that when calculating softmax, the attention here is 0, thus avoiding the impact of padding.

[0117] (3) Softmax normalization: The similarity scores are softmax normalized to obtain the weight of each key, which represents the importance of the corresponding value vector.

[0118]

[0119] Weighted sum: Use these weights to perform a weighted sum on the value vectors to get the final attention output.

[0120] The Multi-Head Attention Mechanism is an extension of the attention mechanism. It uses multiple "heads" to perform attention calculations in parallel to better capture the relationships and features of different aspects of the input data. We define three weight matrices w_query, w_key, and w_value, and use these three matrices to perform three linear transformations on all word vectors. As a result, all word vectors derive three new vectors q, k, and v. Then, q, k, and v are cut into multiple attention heads, and the attention scores are calculated separately. Then, we merge all the attention heads to get the output.

[0121] The multi-head attention mechanism is as follows: calculate the relevance between comments and news, perform softmax normalization on the similarity scores, obtain the weight of each part of the comments, use these weights to perform weighted summation on the comments, and obtain the final event features of the comments about the news.

[0122] Furthermore, the news text and the comment text are input into a dual emotion acquisition module to obtain dual emotion features. The dual emotion acquisition module includes:

[0123] The dual emotion acquisition module consists of two parts: text emotion fine-grained extraction sub-module and dual emotion extraction sub-module.

[0124] The text sentiment fine-grained extraction submodule extracts sentiment category features, sentiment vocabulary features, sentiment intensity features, sentiment polarity features and auxiliary sentiment features, and concatenates these features to obtain text fine-grained sentiment features.

[0125] The emotion category feature uses the pre-trained model of NVIDIA and Baidu AI as the emotion classifier f, where the emotion categories include "anger, disgust, fear, happiness, like, neutral, positive, negative, and sadness". Let the output emotion vector dimension be d f , so the prediction for the text is f(T). Therefore, we get the sentiment category feature in The calculation formula for sentiment category is:

[0126]

[0127] The emotional vocabulary features are obtained by processing English and Chinese text information with the help of NRC emotional dictionary and emotional vocabulary ontology. Assume that the text T contains m kinds of emotional categories, that is, E = {e 1 ,e 2 ,e 3 ,…,e m}, where any e∈E, the set of sentiment words of this category in the sentiment dictionary is wei represents the i-th sentiment word of sentiment category e, and n is the number of sentiment words of category e;

[0128] For a specific sentiment category e, the i-th word in the text T is t i , calculate word t i The sentiment score corresponding to the sentiment category e is calculated by calculating the frequency of occurrence of the word and the negative word neg(t i ,s), degree words deg(t i ,s) and other context words to get the sentiment score of the word, and then get the sentiment score of the entire text. The sentiment score of word ti corresponding to sentiment category e is s(t i ,e) is calculated as follows:

[0129]

[0130] Among them, neg(t i ,s) means that when the left window size is w, word t i The value of the negated word in the context; deg(t i ,s) means that when the left window size is w, word t i The value of the degree adverb in the context of . σ takes the value of 1, which means that word t i Belongs to the sentiment category e in the sentiment dictionary, if not, let σ be 0.

[0131] After calculating the sentiment scores of all words in text T, we get the overall sentiment score s(T,e) of the entire text for sentiment category e. Then, we calculate the text-level sentiment scores of all categories in the sentiment dictionary to get the final sentiment vocabulary features of text T. The calculation formula is:

[0132]

[0133] Among them, ⊕ vector concatenation operation, s(t i ,e) represents word t i The sentiment score corresponding to sentiment category e, s(T,e) represents the sentiment score of the entire text corresponding to sentiment category e.

[0134] The emotion intensity feature extraction process is similar to the emotion word feature extraction process, but the difference is that the intensity score is added. For text T, the corresponding emotion word intensity score is s′(T,e), int(t i ) represents the word t i The intensity score of the text T can be obtained by calculating the text-level sentiment scores of all categories in the sentiment dictionary. The calculation formula is:

[0135]

[0136] Among them, int(t i ) represents the word t i The intensity fraction, s(t i ,e) represents the word t corresponding to the sentiment category e i The sentiment score, s ′ (T,e) represents the sentiment intensity score of text T corresponding to sentiment category e. It represents the emotion intensity characteristics of all emotion categories corresponding to T.

[0137] The sentiment polarity feature is a positive or negative value, which represents the degree of positive or negative polarity of the entire article. The sentiment polarity value of Chinese and English text information is calculated by using HowNet and NLTK toolkit as the sentiment polarity feature of text information. This feature reflects the sentiment polarity of text T ("positive" means positive sentiment score, "negative" means negative sentiment score, and "neutral" means 0 sentiment score). Assume that the dimension of the sentiment score is d s (usually d s =1), we can get the sentiment score feature

[0138] The emotional additional feature is calculated through the bilingual emotional dictionary of HowNet and the emoticon list of Wikipedia, including elements with emotional colors such as emoticons, punctuation marks, and capital letters, and the frequency of use of different non-lexical information is calculated as an auxiliary emotional feature. Assume that the dimension of the emotional additional feature is d a (usually d a =1), and get the emotion score feature

[0139] Combine all the above five features to get the sentiment feature of text T. Calculation method:

[0140]

[0141] Among them, ⊕ represents vector concatenation, represents the sentiment classification feature, Represents the characteristics of sentiment words, Indicates the emotional intensity characteristics, represents the sentiment score, Indicates additional emotional characteristics.

[0142] Input the news text into the text sentiment fine-grained extraction submodule to obtain the news sentiment E N,Different from news text, after the comment text is input into the ,text sentiment fine-grained extraction sub-module, it passes through two aggregators, namely mean pooling and maximum pooling, to capture the ,average sentiment signal and extreme sentiment signal of the comment respectively.

[0143] Comment text input text sentiment fine-grained extraction submodule to obtain initial comment sentiment The comment sentiment E is obtained by concatenating the features after the maximum pooling and the mean pooling. C . Calculation formula:

[0144]

[0145] Among them, mean represents the mean pooling aggregator, max represents the maximum pooling aggregator, represents the initial comment sentiment output by the text sentiment fine-grained extraction submodule, represents the average sentiment characteristic of the comments, Indicates the extreme sentiment characteristics of the comment.

[0146] The dual sentiment extraction submodule obtains the fine-grained sentiment features E of the news text. N And the fine-grained sentiment features of the comment text E C Then calculate the sentiment offset E between the two gap .

[0147]

[0148] Among them, E N Represents the fine-grained sentiment features of news text, represents the average sentiment characteristic of the comments, Indicates the extreme sentiment characteristics of the comment, E gap Indicates biased emotion.

[0149] After obtaining the offset emotion, the news emotion, comment emotion and emotion offset are concatenated to obtain the dual emotion feature. The calculation formula is:

[0150] E dual =E N ⊕E C ⊕E gap

[0151] Among them, E C Indicates the comment sentiment, E gap Indicates biased emotion, E dual Indicates dual emotional characteristics.

[0152] Furthermore, the word vector of the comment text is input into the partition filtering network to obtain the initial stance feature and the initial rumor feature. The partition filtering network includes:

[0153] Partition filtering network is divided into two processes: partition and filter;

[0154] During the partitioning process, the gating mechanism is calculated by the cumulative maximum activation function cummax(·), which is an activation function whose output is a vector of [0,...,0,1,...,1]:

[0155] cummax(·)=cumsum(softmax(·))

[0156] cumsum(x 1 ,x 2 ···x n-1 ,x n )=(x 1 ,x 1 +x 2 ,x 1 +x 2 +x 3 ,···,x 1 +x 2 +···x n-1 ,

[0157] x 1 +x 2 +···x n-1 +x n )

[0158] To select features, features irrelevant to each task are filtered out. The partition filtering network uses two gating units, the stance gate and rumor gate The calculation methods are:

[0159]

[0160] Among them, cummax represents the activation function, Linear represents the linear layer, and x t Indicates the current input, h t-1 Represents the output of the previous time step.

[0161] At each time step, the cell neurons are divided into three partitions, two task partitions storing the internal information of the task, namely the position partition ρ s and rumor partition ρ f , and a shared partition ρ that stores information between tasks sh The cells to be divided are candidate memory cells representing the current moment. and the cell state c at the previous moment t-1 That is, the cell state at time t-1. The cell state c at the previous time t-1For example, the partition of the stand gate and rumor gate For the cell state c at the previous moment t-1 The partition formula is:

[0162]

[0163] in, represents the element-wise product, represents the cell state of the shared partition at time t-1; It indicates the cell state at the last moment after partitioning by the position gate, which contains the shared partition and position partition information. It represents the cell state at the last moment after the rumor gate partition, which contains the shared partition and rumor partition information. Indicates the cell state at the last moment that only contains position information. Represents the cell state at the last moment that only contains rumor information, candidate memory cell The partitioning method is related to the cell state c at the previous moment t-1 The calculation is the same, and this process is similar to the forgetting mechanism of LSTM.

[0164] Then, the information of the current candidate memory cell is aggregated with the information of the memory cell at the previous moment to obtain the status information of the three partitions:

[0165]

[0166] Among them, ρ s Represents the information of the position partition, ρ f represents the information of rumor partition, ρ sh Indicates the information of the shared partition, c t-1 represents the cell state at time t-1, represents the cell state of the position partition at time t-1, represents the cell state of the rumor partition at time t-1, represents the cell state of the shared partition at time t-1; represents the candidate cell state of the position partition at time t, represents the candidate cell state of the rumor partition at time t-1, Represents the candidate cell state of the shared partition at time t-1. This process is similar to the update mechanism of LSTM.

[0167] During the filtering process, three memory modules are used: position memory module μ s , rumor memory module μ f and shared memory module μ sh, the stance memory module selects information from the stance partition and the shared partition, filtering out information irrelevant to the stance detection task, and the rumor memory module selects information from the rumor detection and shared partitions:

[0168] μ s =ρ s +ρ sh ;

[0169] μ f =ρ f +ρ sh ;

[0170] μ sh =ρ sh ;

[0171] Among them, ρ s Represents the information of the position partition, ρ f represents the information of rumor partition, ρ sh Indicates the shared partition information.

[0172] After updating the information in each memory module, the initial stance feature h is generated s , initial rumor feature h f and shared features h sh :

[0173] h s =tanh(μ s )

[0174] h f =tanh(μ f )

[0175] h sh =tanh(μ sh )

[0176] Among them, h s represents the initial position feature of the output at the current moment, h f represents the initial rumor feature of the output at the current moment, h sh Represents the shared features of the output at the current moment.

[0177] After partitioning and filtering at time t, the information in all three memory modules is used to generate the cell state c of the next neuron input. t and the hidden state h t :

[0178] c t =Linear([μ s,t ;μ f,t ;μ sh,t ])

[0179] ht =tanh(c t )

[0180] Among them, Linear represents the calculation of linear transformation, [μ s,t ;μ f,t ;μ sh,t ] means that all the information of the three memory modules is contained at time t, c t Represents the generated cell state, which is used to input the next neuron, h t Represents the hidden layer output, the same as the cell state c t Input into the next time step together;

[0181] Finally, after partition filtering, the maximum pooling method is used to extract the stance features and rumor features of the entire sequence.

[0182] h gs,t =tanh([Linear[h s,t ;h sh,t ])

[0183] h gf,t =tanh([Linear[h f,t ;h sh,t ])

[0184] h gs =maxpool(h gs,1 ,···,h gs,L )

[0185] h gf =maxpool(h gf,1 ,···,h gf,L )

[0186] Among them, Linear represents the calculation of linear transformation, maxpool represents the calculation of maximum pooling, and h gs,t ,h gf,t They represent the stance features and rumor features obtained after partition filtering at time t, respectively. gs ,h gf It is the global stance characteristic and rumor characteristic of the entire time series.

[0187] Furthermore, the fusing the event feature with the stance feature to obtain the revised stance feature includes: using an adder to fuse the event feature with the stance feature to obtain the revised stance feature.

[0188] Furthermore, the classifying the modified stance features to obtain the first classification result includes: using a classifier to classify the modified stance features to obtain the first classification result. The classifier includes: a support vector machine or a random forest model.

[0189] Furthermore, the fusing of the initial rumor features with the revised stance features to obtain the revised rumor features includes: using a low-rank tensor fusion method to fuse the initial rumor features with the revised stance features to obtain the revised rumor features.

[0190] Furthermore, the low-rank tensor fusion is a mechanism for multi-modal feature fusion. It is mainly used to effectively combine features from different modalities while maintaining computational efficiency and information integrity. The core idea of ​​low-rank tensor fusion is to reduce the computational complexity of high-dimensional features through the low-rank representation of tensors, while improving the performance and generalization ability of the model.

[0191] The tensor representation layer first converts multiple inputs into high-dimensional tensors and then maps them back to a low-dimensional output vector space. In order to simulate the interaction between multiple features with a tensor, multiple features are outer-producted to obtain the tensor fusion layer.

[0192] The input tensor Z is calculated by expressing the outer product of different features:

[0193]

[0194] in, z m are input vectors from different modalities.

[0195] Input Tensor Passing it through a linear layer g(Z; W; b) produces a fused tensor:

[0196] h=g(Z;W;b)=W·Z+b

[0197] in, W is the weight and b is the bias.

[0198] The low-order weight decomposition layer decomposes W into a set of modal low-rank factors and regards W as d h M-order tensors, each of which can be expressed as There is a pattern that decomposes into vectors: The smallest R that makes the decomposition effective is called the rank of the tensor, the set of vectors It is called the rank R factorization factor of the original tensor.

[0199] Then use the factorization To reconstruct a low-rank version of

[0200] These vectors can be reassembled into M mode-specific low-rank factors, let

[0201]

[0202] Then the low-rank factor corresponding to mode m is Then the low-rank weight tensor is Reconstruction obtained.

[0203] Based on the decomposition of W, according to Parallel decomposition of Z into the original input z m , then the formula for calculating h is as follows:

[0204]

[0205] in, represents the element-wise product of a series of tensors,

[0206] like Figure 5 As shown, the input is the stance feature z s 、Rumor characteristics f and dual emotional characteristics dual Three vectors, so h is expressed as:

[0207]

[0208] Among them, Z represents the tensor layer obtained by doing the outer product of three features, z s Indicates the characteristics of the position, z f Indicates rumor characteristics z f , z dual Indicates dual emotional characteristics, represents the weight of the decomposed stance feature, represents the weight of the decomposed rumor feature, Represents the weight of the decomposed stance feature.

[0209] Furthermore, the classifying the revised rumor features to obtain the second classification result includes: using a classifier to classify the revised rumor features to obtain the second classification result. The classifier includes: a support vector machine or a random forest model.

[0210] Furthermore, the fusing the first classification result and the second classification result to obtain the final classification result includes: fusing the first classification result and the second classification result in a weighted fusion manner to obtain the final classification result.

[0211] Embodiment 2

[0212] This embodiment provides a fake news detection system based on a partition filtering network, including:

[0213] An acquisition module is configured to: acquire the news to be detected and a number of comments corresponding to the news to be detected;

[0214] A classification module is configured to: input the news to be detected and a number of comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false;

[0215] Among them, the trained fake news detection model is used for:

[0216] Vectorize the news to be detected to obtain the word vector of the news to be detected; extract features from the word vector of the news to be detected to obtain the news text features;

[0217] Vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; extract features from the word vector of the comment text to obtain the features of the comment text;

[0218] The news text features and comment text features are integrated to obtain event features;

[0219] Input the news text and comment text into the dual emotion acquisition module to obtain the dual emotion features;

[0220] Input the word vector of the comment text into the partition filtering network to obtain the initial position feature and initial rumor feature;

[0221] The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result;

[0222] The initial rumor features, the dual emotion features and the revised stance features are integrated to obtain revised rumor features; the revised rumor features are classified to obtain the second classification result;

[0223] The first classification result and the second classification result are fused to obtain the final classification result.

[0224] It should be noted that the acquisition module and the classification module correspond to steps S101 to S102 in Embodiment 1, and the examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0225] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0226] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0227] Embodiment 3

[0228] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory so that the electronic device executes the method described in the above embodiment one.

[0229] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0230] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0231] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

[0232] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0233] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0234] Embodiment 4

[0235] This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in the first embodiment is completed.

[0236] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A false news detection method based on partition filtering network, characterized by: include: Obtaining the news to be detected and several comments corresponding to the news to be detected; inputting the news to be detected and several comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false; The trained fake news detection model is used to: vectorize the news to be detected to obtain the word vector of the news to be detected; perform feature extraction on the word vector of the news to be detected to obtain the news text features; vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; perform feature extraction on the word vector of the comment text to obtain the comment text features; The news text features and comment text features are integrated to obtain event features. The news text and comment text are input into the dual emotion acquisition module to obtain dual emotion features. The word vector of the comment text is input into the partition filtering network to obtain the initial position features and initial rumor features. The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result; the initial rumor features, the dual emotion features and the revised stance features are integrated to obtain the revised rumor features; the revised rumor features are classified to obtain the second classification result; the first classification result and the second classification result are integrated to obtain the final classification result.

2. The false news detection method based on partition filtering network as claimed in claim 1 is characterized in that: The trained fake news detection model includes: A first BERT model and a second BERT model, wherein the input value of the first BERT model is a news text, and the input value of the second BERT model is a comment text corresponding to the news text; the output end of the first BERT model is connected to the input end of the first Bi-LSTM model; the output end of the second BERT model is connected to the input end of the second Bi-LSTM model; the output end of the second BERT model is also connected to the input end of the partition filtering network; The first Bi-LSTM model is used to extract features from the word vectors of the news to obtain news text features; the second Bi-LSTM model is used to extract features from the word vectors of the comment text to obtain comment text features; The output end of the first Bi-LSTM model and the output end of the second Bi-LSTM model are both connected to the input end of the multi-head attention mechanism layer, and the output end of the multi-head attention mechanism layer outputs event features; The trained false news detection model further includes a dual emotion acquisition module, wherein the input end of the dual emotion acquisition module inputs the news text and the comment text respectively, and the output end of the dual emotion acquisition module outputs the dual emotion feature; The output end of the partition filtering network outputs the initial position features and initial rumor features; The trained false news detection model further includes an adder and a first classifier, wherein the adder fuses the event feature and the initial stance feature to obtain a revised stance feature; and the first classifier classifies the revised stance feature to obtain a stance classification result; The trained false news detection model also includes a low-rank tensor fusion network and a second classifier, wherein the low-rank tensor fusion network fuses the initial rumor features, the dual emotion features and the revised stance features to obtain revised rumor features; the second classifier classifies the revised rumor features to obtain a classification result of whether the news is true or false; The trained fake news detection model also includes: a weighted fusion unit, which performs weighted fusion on the first classification result and the second classification result to obtain a final detection result.

3. The false news detection method based on partition filtering network as claimed in claim 1 or 2, characterized in that: The dual emotion acquisition module comprises: The dual emotion acquisition module consists of two parts: the text emotion fine-grained extraction submodule and the dual emotion extraction submodule; The text sentiment fine-grained extraction submodule extracts sentiment category features, sentiment vocabulary features, sentiment intensity features, sentiment polarity features and additional sentiment features of news texts, and concatenates all extracted features to obtain news text fine-grained sentiment features; The text sentiment fine-grained extraction submodule extracts sentiment category features, sentiment vocabulary features, sentiment intensity features, sentiment polarity features and additional sentiment features of the comment text, and concatenates all the extracted features to obtain the comment text fine-grained sentiment features; The dual sentiment extraction submodule calculates the offset sentiment features based on the fine-grained sentiment features of the news text and the fine-grained sentiment features of the comment text; Then the fine-grained sentiment features of news text, the fine-grained sentiment features of comment text and the offset sentiment features are concatenated to obtain the dual sentiment features.

4. The false news detection method based on partition filtering network as claimed in claim 2 is characterized in that: The low-rank tensor fusion network comprises: Sequentially connected tensor representation layers and low-order weight decomposition layers; The tensor representation layer performs outer product on the input corrected stance feature, initial rumor feature and dual sentiment feature to obtain a representation tensor; The tensor representation layer constructs multiple features into tensor representation in the form of outer products; assuming there are two input features Then their tensors are expressed as This is the tensor product of two eigenvectors, forming a d a ×d b Dimensional matrix; Low-order weight decomposition layer: In order to reduce the complexity of the constructed tensor, low-rank tensor decomposition is used; low-rank tensor decomposition includes CP decomposition or Tucker decomposition.

5. The false news detection method based on partition filtering network as claimed in claim 1 is characterized in that: The word vector of the comment text is input into the partition filtering network to obtain the initial position feature and the initial rumor feature. The partition filtering network includes: the partition filtering network is divided into two processes: partitioning and filtering; in the partitioning process, the gating mechanism is calculated by the cumulative maximum activation function cummax(·), cummax is an activation function, and the output of the activation function is a vector of [0,...,0,1,...,1]: cummax(·)=cumsum(softmax(·)) cumsum(x1,x2···x n-1 ,x n )=(x1,x1+x2,x1+x2+x3,···,x1+x2+···x n-1 , x1+x2+···x n-1 +x n ) In order to select features, features irrelevant to each task are filtered out; the partition filtering network uses two gating units, the stance gate and rumor gate The calculation methods are: Among them, cummax represents the activation function, Linear represents the linear layer, and x t Indicates the current input, h t-1 Represents the output of the previous time step; At each time step, the cell neurons are divided into three partitions, two task partitions storing the internal information of the task, namely the position partition ρ s and rumor partition ρ f , and a shared partition ρ that stores information between tasks sh ; The cells to be divided are candidate memory cells representing the current moment and the cell state c at the previous moment t-1 That is, the cell state at time t-1; the cell state c at the previous time t-1 For example, the partition of the stand gate and rumor gate For the cell state c at the previous moment t-1 The partition formula is: in, represents the element-wise product, represents the cell state of the shared partition at time t-1; It indicates the cell state at the last moment after partitioning by the position gate, which contains the shared partition and position partition information. It represents the cell state at the last moment after the rumor gate partition, which contains the shared partition and rumor partition information. Indicates the cell state at the last moment that only contains position information. Represents the cell state at the last moment that only contains rumor information, candidate memory cell The partitioning method is related to the cell state c at the previous moment t-1 The calculations are the same; Then, the information of the current candidate memory cell is aggregated with the information of the memory cell at the previous moment to obtain the status information of the three partitions: Among them, ρ s Represents the information of the position partition, ρ f represents the information of rumor partition, ρ sh Indicates the information of the shared partition, c t-1 represents the cell state at time t-1, represents the cell state of the position partition at time t-1, represents the cell state of the rumor partition at time t-1, represents the cell state of the shared partition at time t-1; represents the candidate cell state of the position partition at time t, represents the candidate cell state of the rumor partition at time t-1, Represents the candidate cell state of the shared partition at time t-1.

6. The false news detection method based on partition filtering network as claimed in claim 5 is characterized in that: During the filtering process, three memory modules are used: position memory module μ s , rumor memory module μ f and shared memory module μ sh , the stance memory module selects information from the stance partition and the shared partition, filtering out information irrelevant to the stance detection task, and the rumor memory module selects information from the rumor detection and shared partitions: m s =ρ s +r sh ;m f =ρ f +r sh ;m sh =ρ sh ; Among them, ρ s Represents the information of the position partition, ρ f represents the information of rumor partition, ρ sh Indicates the information of the shared partition; After updating the information in each memory module, the initial stance feature h is generated s , initial rumor feature h f and shared features h sh : h s =tanh(μ s );h f =tanh(μ f );h sh =tanh(μ sh ); Among them, h s represents the initial position feature of the output at the current moment, h f represents the initial rumor feature of the output at the current moment, h sh Represents the shared features of the output at the current moment; After partitioning and filtering at time t, the information in all three memory modules is used to generate the cell state c of the next neuron input. t and the hidden state h t : c t =Linear([μ s,t ;μ f，t ;μ sh,t ]);h t =tanh(c t ) Among them, Linear represents the calculation of linear transformation, [μ s，t ;μ f，t ;μ sh,t ] means that all the information of the three memory modules is contained at time t, c t Represents the generated cell state, which is used to input the next neuron, h t Represents the hidden layer output, the same as the cell state c t Input into the next time step together; Finally, after partition filtering, the position features and rumor features of the entire sequence are extracted using the maximum pooling method; h gs，t =tanh([Linear[h s，t ;h sh,t ]);h gf，t =tanh([Linear[h f，t ;h sh,t ]); h gs =maxpool(h gs，1 ,…,h gs，L );h gf =maxpool(h gf，1 ,…,h gf，L ); Among them, Linear represents the calculation of linear transformation, maxpool represents the calculation of maximum pooling, and h gs，t ,h gf，t They represent the stance features and rumor features obtained after partition filtering at time t, respectively. gs ,h gf It is the global stance characteristic and rumor characteristic of the entire time series.

7. The false news detection method based on partition filtering network as claimed in claim 3 is characterized in that: The emotion category feature uses the pre-trained model of NVIDIA and AI as the emotion classifier f, where the emotion categories include: anger, disgust, fear, happiness, like, neutral, positive, negative, and sadness; let the output emotion vector dimension be d f , so the prediction of the text is f(T); therefore, the sentiment category feature is obtained in The emotion intensity feature extraction process is similar to the emotion vocabulary feature extraction process, but the difference from the emotion vocabulary feature extraction is that the intensity score is added; The sentiment polarity feature is a positive or negative value, which represents the degree of positive or negative polarity of the entire article. The sentiment polarity value of Chinese and English text information is calculated by using HowNet and NLTK toolkit as the sentiment polarity feature of the text information; The emotional additional features are calculated by using the bilingual emotional dictionary of HowNet and the emoticon list of Wikipedia to calculate the elements with emotional colors, and by calculating the usage frequency of different non-lexical information as auxiliary emotional features.

8. A false news detection system based on partition filtering network, characterized by: include: An acquisition module is configured to: acquire the news to be detected and a number of comments corresponding to the news to be detected; A classification module is configured to: input the news to be detected and a number of comments corresponding to the news to be detected into the trained false news detection model to obtain a classification result of whether the news to be detected is false; Among them, the trained fake news detection model is used for: Vectorize the news to be detected to obtain the word vector of the news to be detected; extract features from the word vector of the news to be detected to obtain the news text features; Vectorize the comment text corresponding to the news to be detected to obtain the word vector of the comment text; extract features from the word vector of the comment text to obtain the features of the comment text; The news text features and comment text features are integrated to obtain event features; Input the news text and comment text into the dual emotion acquisition module to obtain the dual emotion features; Input the word vector of the comment text into the partition filtering network to obtain the initial position feature and initial rumor feature; The event features are integrated with the initial stance features to obtain the revised stance features; the revised stance features are classified to obtain the first classification result; The initial rumor features, the dual emotion features and the revised stance features are integrated to obtain revised rumor features; the revised rumor features are classified to obtain the second classification result; The first classification result and the second classification result are fused to obtain the final classification result.

9. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; as well as a processor for executing the computer readable instructions, When the computer-readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: The computer-readable instructions are non-transitory stored, wherein when the non-transitory computer-readable instructions are executed by a computer, the method according to any one of claims 1 to 7 is performed.

Images