Real-time false news event detection method and system based on multi-source social data

Abstract

The real-time false news event detection method and system based on multi-source social data disclosed in the present invention include: obtaining news and comments associated with news events; determining the reliability of each news according to the relevance of each news to the news event; determining the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of the news comment, splicing the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature to obtain the comment feature of the news event; determining the truth or falsity of each news according to the content of each news, taking the reliability of each news as the weight to perform weighted summation on the truth or falsity of each news, and obtaining the content feature of the news event; splicing the content feature and comment feature of the news event to obtain the splicing feature of the news event; determining the truth or falsity judgment result of the news event according to the splicing feature of the news event. Accurate judgment of news events is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of fake news detection, and in particular to a real-time fake news event detection method and system based on multi-source social data. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rise of the Internet and social media, the way news information is disseminated has changed dramatically. The monopoly of traditional media channels has been broken, and everyone can publish and share information through social media platforms, making the dissemination of news information faster and more widespread. However, it is precisely because of the rapid dissemination and widespread sharing of information that fake news has also emerged. The emergence of fake news undermines the credibility and reliability of information, making it difficult for the public to distinguish between truth and lies. This may lead to public distrust of the news media and even question the stability of the entire information ecosystem. Given the potential harm of fake news, the demand for fake news detection and combating is increasing.

[0004] Currently, the Internet is flooded with a large number of fake news of all kinds. It is difficult to cope with such a huge amount of news by relying solely on manual detection. Therefore, it is inevitable to use deep learning algorithm models to replace manual detection. The current mainstream fake news detection algorithms face the following problems:

[0005] 1. Existing fake news detection models only focus on the content of the news itself, but lack due attention to the relevant comments during its dissemination process. This results in the judgment of the truth or falsehood of a news article being largely limited by the writing skills of the news writer, which makes it difficult for the algorithm model to perform well when encountering "fake masters". Therefore, in order to improve the accuracy of the algorithm model in complex situations, it is urgent to pay attention to news comments.

[0006] 2. When training, the existing fake detection algorithms often do not dynamically update their training data. Using past training data will lead to a decrease in the recognition rate of currently popular news and a lack of basic understanding of Internet hot words. Therefore, a dynamic and real-time background training data is the only way to solve this problem.

[0007] 3. The current fake news detection model can only judge the truth of a single piece of news, but cannot judge the truth of the entire news event as a whole. The lack of overall control makes the usability of its judgment greatly reduced. At the same time, when judging news, it also lacks the idea of ​​news classification, and often only studies a certain category, which is not universal. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a real-time false news event detection method and system based on multi-source social data, which combines all news comments and news content in the news time to achieve accurate judgment of news events.

[0009] To achieve the above object, the present invention adopts the following technical solution:

[0010] First, a real-time fake news event detection method based on multi-source social data is proposed, including:

[0011] Get news and comments related to news events;

[0012] Determine the credibility of each piece of news based on its relevance to the news event;

[0013] Determine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of news comments, and combine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature to obtain the comment feature of the news event;

[0014] According to the content of each piece of news, the truth or falsity of each piece of news is determined, and the reliability of each piece of news is used as a weight to perform weighted summation on the truth or falsity of each piece of news to obtain the content characteristics of the news event;

[0015] Splicing the content features and comment features of the news event to obtain the splicing features of the news event;

[0016] According to the splicing features of the news event, the truth or falsity of the news event is determined.

[0017] Secondly, a real-time fake news event detection system based on multi-source social data is proposed, including:

[0018] A data set acquisition module is used to acquire news and comments associated with news events;

[0019] A reliability acquisition module is used to determine the reliability of each piece of news based on the relevance of each piece of news to the news event;

[0020] The comment feature acquisition module is used to determine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of the news comment, and to obtain the comment feature of the news event by combining the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature;

[0021] The content feature acquisition module is used to determine the truth or falsity of each piece of news based on its content, and to perform weighted summation of the truth or falsity of each piece of news using the reliability of each piece of news as a weight to obtain the content feature of the news event;

[0022] A feature splicing module is used to splice the content features and comment features of the news event to obtain the splicing features of the news event;

[0023] The news event truth or falsity judgment module is used to determine the truth or falsity judgment result of the news event based on the splicing features of the news event.

[0024] In a third aspect, an electronic device is proposed, comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps described in the real-time fake news event detection method based on multi-source social data are completed.

[0025] In a fourth aspect, a computer-readable storage medium is proposed for storing computer instructions. When the computer instructions are executed by a processor, the steps described in the real-time false news event detection method based on multi-source social data are completed.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. Compared with the currently popular fake news detection technology, it only focuses on the content of the news content itself, lacks attention to the comments generated during the dissemination of news, resulting in fake news detection technology being limited by the knowledge reserves of news writers and algorithm developers, and unable to accurately identify some special news. The present invention combines news comments and content for comprehensive analysis, which greatly improves the stability of the algorithm and the accuracy of judging the truth of news events.

[0028] 2. Existing fake news detection technologies can usually only obtain news from a single data source, and cannot effectively integrate news reports from different websites or channels, and cannot achieve dynamic update and expansion of data. This leads to biased judgments on emerging events because existing algorithms do not update the latest relevant news information. The present invention uses web crawler technology to expand original news data from multiple large news websites every time an actual access request is made, and uses the latest relevant data as the basis for judgment. After judgment, the original news data set will be dynamically updated, thereby achieving real-time false news event detection.

[0029] 3. When determining the truth or falsity of news based on the content of the news, the present invention first determines the category to which the news event belongs, and then identifies the news content through a trained news truth or falsity judgment model corresponding to the category to which the news event belongs, and determines the truth or falsity of the news, thereby making the truth or falsity judgment of the news more targeted, ensuring the accuracy of the truth or falsity judgment of the news based on the content of the news, and further ensuring the accuracy of the content features of the extracted news, thereby ensuring the accuracy of the truth or falsity identification of the news event.

[0030] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0032] Figure 1 A flow chart of the method disclosed for the embodiment;

[0033] Figure 2 It is a schematic diagram of the overall framework of the method disclosed in the embodiment;

[0034] Figure 3 This is a feature fusion model diagram disclosed in the embodiment. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. 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 application belongs.

[0037] Terminology explanation:

[0038] Multi-source social data: Multi-source social data refers to the news content of an event obtained from multiple mainstream news media and the social data generated during the dissemination process, namely comments.

[0039] Real-time fake news: Based on the hot events happening at the moment, some people with ulterior motives will forge untrue news with the intention of confusing the public.

[0040] False news event detection based on social data fusion: Based on the text description of the news content and the netizens’ comments attached to the news, joint identification is performed to determine the truth or falsity of the entire event.

[0041] Vectorized representation: Convert text data into numerical representation, use numerical values ​​to characterize the dimensional features of the text, and provide basic data for subsequent models and algorithms.

[0042] Word2Vec: is a technology that vectorizes words.

[0043] Bidirectional LSTM (Bidirectional Long Short-Term Memory): A special recurrent neural network that can examine the overall data from a time span and obtain relational features on the sequence from both forward and backward dimensions.

[0044] BERT (Bidirectional Encoder Representations from Transformers): A pre-trained model based on the Transformer model, which can achieve various tasks through fine-tuning and achieve word segmentation effect by processing the input sequence.

[0045] Posseg: A submodule of the jieba module that provides Chinese word segmentation and part-of-speech tagging functions.

[0046] SnowNLP: SnowNLP provides the function of sentiment analysis on Chinese texts. It can determine the sentiment tendency of a text, such as whether the text is positive, negative or neutral.

[0047] Gensim: A Python library mainly used for natural language processing related tasks. It extracts text features through TF-IDF to build a bag-of-words model, and then calculates cosine similarity to obtain text similarity.

[0048] LDA (Latent Dirichlet Allocation): It is a text topic modeling method based on a probability model, which can discover potential topic structures in text datasets. By adjusting the parameters, the topic word set in the text dataset can be obtained.

[0049] Feature fusion: combining features from different sources or types to improve the performance and accuracy of the model. The present invention adopts the feature weighting form for fusion, that is, the content features and comment features are fused in a weighted manner.

[0050] News credibility: The obtained news set is used to calculate the similarity with the event description, and the value obtained after processing is the credibility of the news and the event.

[0051] Binary classification model: is a type of machine learning model used to solve classification problems involving only two categories. In a binary classification task, the goal is to classify samples into two mutually exclusive categories based on given input features, such as whether news is true or false.

[0052] Example 1

[0053] At present, the Internet is full of a large number of fake news of all kinds. It is difficult to cope with such a huge amount of news by relying solely on manual detection. Therefore, it is an inevitable trend to use deep learning algorithm models to replace manual detection. The current mainstream fake news detection algorithm is regarded as a text classification problem. It extracts information such as text and pictures from the news content, analyzes their respective feature representations, and performs feature fusion to judge the authenticity of the news. After development in recent years, the algorithms related to fake news detection have experienced a transition from the shallow level of simple statistical models to the deep level of deep learning models. With the improvement of deep learning models such as recurrent neural networks, convolutional neural networks, and attention mechanisms, the performance of different text classification tasks has been continuously improved. It is worth noting that compared with the shallow learning method based on statistical learning, it can directly learn feature representations from input data without manually designing rules and features, which greatly alleviates the pressure of manual work.

[0054] However, in the existing fake news detection algorithms, the training data of the models is often fixed and cannot interact well with current hot events to make accurate judgments. At the same time, the current algorithm models are heavily dependent on the news content itself, which is subject to the knowledge reserves of the news authors and the reviewers at the beginning of the algorithm design. They only focus on the news content itself and ignore the importance of related comments during the dissemination of news.

[0055] To this end, in this embodiment, a real-time false news event detection method based on multi-source social data is disclosed. First, the existing news data set + crawler technology is used to obtain a set of relevant news about a certain event and its related comments. The LDA method is used to extract the keywords of the news event and then judge the category to which it belongs. Subsequently, tf-idf is used to extract the news content keywords for each news in this event in turn and compare the similarity with the news event description, and the news relevance is obtained as a weight assignment called news credibility. Then, feature vectors are extracted for the content of the news and its comments respectively, and content features and comment features are output. Feature fusion is performed in combination with their corresponding weights to finally obtain a judgment on the true or false of the corresponding news. Finally, the data obtained by the crawler is saved in the background database as new training data. After cosine similarity comparison, news with too high repetition is deleted to achieve dynamic expansion of training data.

[0056] Specifically, the real-time false news event detection method based on multi-source social data disclosed in this embodiment is as follows: Figure 1-Figure 3 As shown, including:

[0057] S1: Obtain the original news data set in the database, and use the news belonging to the same news event category in the original news data set to train the news truth judgment model. After the training is completed, the trained news truth judgment models corresponding to different news event categories are obtained.

[0058] In addition, the news and its comments belonging to the same news event category in the original news dataset are used to train the news sentiment prediction model and the news comment sentiment prediction model respectively. After the training is completed, the trained news sentiment prediction model and the news comment sentiment prediction model are obtained.

[0059] S2: Obtain news and comments related to the news event; determine the credibility of each piece of news based on the relevance of each piece of news to the news event.

[0060] This embodiment uses crawler technology to crawl various news and comments related to news events from the network platform.

[0061] This embodiment obtains the event description of the news event and the keyword vector of each news; calculates the similarity between the keyword vector of each news and the event description to obtain the relevance between each news and the news event; normalizes the relevance between each news and the news event to obtain the reliability of each news.

[0062] Preferably, news and comments related to news events are crawled mainly from mainstream media platforms such as Tencent News, NetEase News and Weibo.

[0063] The process of crawling news and its comments is:

[0064] First, get the event description of the news event, enter the event description into the requests library, use the requests library to crawl the target website for the first time, get the page source code, and then use BeautifulSoup, Xpath, etc. to extract data according to the website loading situation, extract the URL list of detailed news, and perform the above operations on each URL in turn to extract the title, content, comment, publication time, and source of the news, corresponding to title, content, comment, get-time, and source respectively.

[0065] Then, LDA is used to segment the news titles, and compared with the subject vocabulary in the original news dataset. The number of times the words appear is compared to determine the category to which the event belongs, which is the label attribute, so as to prepare for the subsequent selection of the corresponding training model.

[0066] Finally, for each piece of news obtained, we first use the word segmentation tool jieba to segment the news, extract the key words in the news, and set it to return only common nouns, names and place names, save them in list form, then build a bag of words model to filter out words with more occurrences, and then use gensim to build a tfidf object to obtain the subject words of each piece of news. Finally, we calculate the similarity between the subject words of each news and the event description, and normalize the results as the dependency of each piece of news.

[0067] Preferably, cosine similarity is used to calculate the similarity between news keywords and event descriptions.

[0068] Regarding the cosine similarity, the event description is represented as M, and the vector of each news keyword is represented as S i , they are all represented in vector form, then the cosine similarity between news keywords and event descriptions (M,S i ) can be expressed as:

[0069]

[0070] Among them, M and S i are the two vectors to be compared, · represents the inner product of the vectors, ||M|| and ||S i || represents the norm of a vector.

[0071] Use the MMS formula to normalize the similarity for subsequent processing operations and scale the data to a fixed range. Let X represent the original data, X_scaled represent the scaled data, X_min and X_max represent the minimum and maximum values ​​of the original data, min_value and max_value represent the target range of the scaled data, and the formula is as follows:

[0072]

[0073] By scaling the similarity to the range of [0, 1] through MMS, the scale difference of the original data can be eliminated, making different features comparable.

[0074] S3: Determine the adjective proportion features, similarity features, characteristic word features and sentiment analysis features of news comments, and concatenate the adjective proportion features, similarity features, characteristic word features and sentiment analysis features to obtain the comment features of the news event.

[0075] Among them, the number of all words and the number of adjectives in all news comments in the news event are determined; the number of adjectives is divided by the number of all words to obtain the adjective proportion feature of the news comments;

[0076] Calculate the similarity between all news comments in the news event, divide the number of comments with similarity greater than the set value by the number of all comments, and obtain the similarity feature of the news comments;

[0077] Determine the most frequently appearing words in all news comments in the news event, and determine the characteristic word features of the news comments according to the number of times the most frequently appearing words appear in all news comments;

[0078] Sentiment prediction is performed on the content and comments of each news item to obtain the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents in the news event; based on the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents, the sentiment analysis characteristics of the news comments are determined.

[0079] In this embodiment, news comments are vectorized, and then four features of comments are extracted in sequence. Finally, the four generated features are spliced ​​to obtain the comment features of the news event. The specific process is as follows:

[0080] Step 3-1: Encode all news comments in the news event, input the comment text into the BERT model, perform vector representation, generate a 32-dimensional vector, and then obtain the word vector matrix of the comment text as input for subsequent steps;

[0081] This embodiment uses BERT as a text encoder, which accepts a word sequence input with a fixed length, expressed as: t = [CLS] + tokens + [SEP], where tokens is a word sequence, and [CLS] and [SEP] are special tags. The BERT model used in this embodiment is a pre-trained model that can convert a word sequence into a word embedding vector, where the number of word embedding vectors is equal to the length of the input word sequence.

[0082] e=BERT(t)

[0083] Where BERT(·) represents encoding the input sequence t to generate the corresponding word embedding vector e.

[0084] Step 3-2: Analyze the proportion of adjectives. After being processed by the possess module in the jieba library, the specific proportion of adjectives in all news comments in the event can be obtained and finally output in numerical form.

[0085] This embodiment uses the possess submodule in the jieba library to extract the number of all adjectives in all comments in the event, and also counts the number of all words in all comments, which are defined as n1 and all1 respectively. The adjective proportion adj-socre is calculated and generated by the two, and is expressed as:

[0086]

[0087] Step 3-3: For the word vector matrix in 3-1, use the similarities submodule in the gensim library to loop through the calculation to obtain the template sentences in all comments in the event. Determine the similarity between the comments through the template sentences in the comments. Divide the number of comments with similarity greater than the set value by the number of all news comments to obtain the similarity feature of the news comments. The set value is preferably 0.7.

[0088] Among them, n2 is the number of similarities greater than 0.7, all2 is the number of all news comments, and sim-socre is the similarity feature of news comments, expressed as:

[0089]

[0090] Step 3-4: For the word vector matrix of 3-1, first generate keywords for the description of the news event as the theme words of the event, then extract words from all news comments of the entire event, and input the theme words as stop words into the LDA algorithm. Then, we can get the most frequently appearing words in all the known false news comments in the news event. According to the number of times these words appear in all the news comments of the event, determine the characteristic word features of the news comments and output them;

[0091] Preferably, LatentDirichletAllocation in the sklearn library is used to extract words from news comments, and the topic is set to 1.

[0092] The number of most frequent words is limited to twice the number of comments.

[0093] The number of times these words appear in all news comments on the event is divided by ten to obtain the characteristic word features of the news comments.

[0094] Among them, all known false news comments are news comments in the original news dataset.

[0095] Step 3-5: First, for the word vector matrix of 3-1, perform sentiment prediction on the content and comments of each news item to obtain the number of positive comments, the number of negative comments, the number of positive contents, and the number of negative contents in the news event; based on the number of positive comments, the number of negative comments, the number of positive contents, and the number of negative contents, determine the sentiment analysis features of the news comments.

[0096] This embodiment determines the category of news events, selects the trained news sentiment prediction model and news comment sentiment prediction model corresponding to the category to which the news event belongs, performs sentiment prediction on the news content and news comment respectively, obtains news content sentiment prediction results and news comment sentiment prediction results, and determines the number of positive comments c, the number of negative comments d, the number of positive contents a, and the number of negative contents b in the news event according to the prediction results; and determines the sentiment analysis feature emo-socre of the news comment according to the number of positive comments c, the number of negative comments d, the number of positive contents a, and the number of negative contents b:

[0097]

[0098] Step 3-6: Concatenate the feature vectors output from step 3-2 to step 3-6, and output the above four feature vectors as a vector group to obtain the comment feature com-socre of the news event for subsequent use.

[0099]

[0100] S4: Determine the truth or falsity of each piece of news based on its content, and use the reliability of each piece of news as a weight to perform weighted summation on the truth or falsity of each piece of news to obtain the content features of the news event.

[0101] After determining the category of a news event, this embodiment selects a trained news truth-or-fake judgment model corresponding to the category to which the news event belongs, identifies the content of the news, and obtains the truth or falsity of the news.

[0102] Specific:

[0103] Step 4-1: Encode the content of the news, input the content of the news into the Word2Vec model, vectorize it, generate a 32-dimensional vector, and then obtain the word vector matrix of the content text as input for subsequent steps.

[0104] Step 4-2: The word vector matrix of step 4-1 is input into the trained news truth judgment model corresponding to the category to which the news event belongs, so as to obtain the truthfulness of each piece of news.

[0105] The news truth judgment model of this embodiment takes news content as input and the truth of news as output, and is constructed using a bidirectional LSTM.

[0106] Use bidirectional LSTM to vectorize the news content text and obtain the representation of each word and its relationship:

[0107] S for =BiLSTM for (t,w)

[0108] S bac =BiLSTM bac (t,w)

[0109] Among them, S represents the hidden state, for represents the forward direction, and bac represents the backward direction. t is the input sequence to be used in BiLSTM, and w is the parameter.

[0110] The input sequence t is passed to the bidirectional LSTM network and processed from the forward and backward directions respectively, which will produce two sets of hidden state sequences: one is the hidden state sequence generated from the forward pass, and the other is the hidden state sequence generated from the backward pass.

[0111] After calculation, the truth or falsity of each piece of news can be obtained. Preferably, 0 represents true and 1 represents false.

[0112] Step 4-3: After obtaining the truth or falsity of each piece of news in step 4-2, perform weighted summation of features in combination with the news credibility to obtain the content features of the news.

[0113] The news credibility is used as the weight of each news when weighting, and the truthfulness of each news is weighted summed up. If the sum is greater than 0.5, the event is considered to be false at the content level, and if it is less than 0.5, the event is considered to be true at the content level. The truthfulness at the content level is used as the content feature of the news.

[0114] S5: Concatenate the content features and comment features of the news event to obtain the concatenated features of the news event, feat-socre; and determine the truth or falsehood judgment result of the news event based on the concatenated features of the news event.

[0115] in,

[0116] The final result of judging whether a news event is true or false can be obtained by inputting the spliced ​​features of the news event into a binary classifier model. The binary classifier model used here is the logistic regression model, which is a basic linear classification model. It performs binary classification by mapping the linear combination samples of the input features to the probability space and limiting the probability to the range of 0-1 with the logistic function. The logistic function is shown below, and the Sigmoid function is used:

[0117]

[0118] The method disclosed in this embodiment comprehensively considers the content of all news in the event and its comments when commenting on the truth or falsity of a news event, thereby ensuring the accuracy of false detection of news events.

[0119] Afterwards, the news and comments obtained in this embodiment are added to the original news data set to update the original news data set.

[0120] Preferably, the acquired news and its comments are compared with the news and comments in the original news data set. After deduplication, the remaining news and its comments are stored in the original news data set, and the original news data set is updated to achieve dynamic update and expansion of data, thereby ensuring the accuracy of model training using the dynamically updated data set and improving the accuracy of judgment of real-time news events.

[0121] Since different categories of news events have extremely significant characteristics, the common methods and fake techniques of false news in different categories of news events are also different. In order to improve the accuracy of false judgments of news events, we first collected a set of mainstream keywords in the six major fields of politics, economy, society, technology, entertainment, and international affairs. For the acquired news, LDA is used to generate the subject words of the news collection, and similarity comparison is performed to determine its category affiliation. This helps the algorithm to be more targeted, and the algorithm can be more targeted for detection and identification.

[0122] Example 2

[0123] In this embodiment, a real-time fake news event detection system based on multi-source social data is disclosed, including:

[0124] A data set acquisition module is used to acquire news and comments associated with news events;

[0125] A reliability acquisition module is used to determine the reliability of each piece of news based on the relevance of each piece of news to the news event;

[0126] The comment feature acquisition module is used to determine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of the news comment, and to obtain the comment feature of the news event by combining the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature;

[0127] The content feature acquisition module is used to determine the truth or falsity of each piece of news based on its content, and to perform weighted summation of the truth or falsity of each piece of news using the reliability of each piece of news as a weight to obtain the content feature of the news event;

[0128] A feature splicing module is used to splice the content features and comment features of the news event to obtain the splicing features of the news event;

[0129] The news event truth or falsity judgment module is used to determine the truth or falsity judgment result of the news event based on the splicing features of the news event.

[0130] Example 3

[0131] In this embodiment, an electronic device is disclosed, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps described in the real-time false news event detection method based on multi-source social data disclosed in Example 1 are completed.

[0132] Example 4

[0133] In this embodiment, a computer-readable storage medium is disclosed for storing computer instructions. When the computer instructions are executed by a processor, the steps described in the real-time false news event detection method based on multi-source social data disclosed in Example 1 are completed.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A real-time fake news event detection method based on multi-source social data, characterized in that: include: Get news and comments related to news events; Determine the credibility of each piece of news based on its relevance to the news event; Determine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of news comments, and combine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature to obtain the comment feature of the news event; Specifically, determine the number of all words and adjectives in all news comments in the news event; divide the number of adjectives by the number of all words to obtain the adjective proportion feature of the news comments; Calculate the similarity between all news comments in the news event, divide the number of comments with similarity greater than the set value by the number of all comments, and obtain the similarity feature of the news comments; Determine the most frequently appearing words in the known false news comments in the news event, and determine the characteristic word features of the news comments according to the number of times the most frequently appearing words appear in all the news comments; The sentiment prediction is performed on the content and comments of each news item to obtain the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents in the news event; the sentiment analysis features of the news comments are determined based on the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents; According to the content of each piece of news, the truth or falsity of each piece of news is determined, and the reliability of each piece of news is used as a weight to perform weighted summation on the truth or falsity of each piece of news to obtain the content characteristics of the news event; Splicing the content features and comment features of the news event to obtain the splicing features of the news event; According to the splicing features of the news event, the truth or falsity of the news event is determined.

2. The real-time false news event detection method based on multi-source social data as claimed in claim 1, characterized in that: Obtain the event description of the news event and the keyword vector of each news; calculate the similarity between the keyword vector of each news and the event description to obtain the relevance between each news and the news event; normalize the relevance between each news and the news event to obtain the credibility of each news.

3. The real-time false news event detection method based on multi-source social data as claimed in claim 1, characterized in that: The number of positive comments c, the number of negative comments d, the number of positive content a and the number of negative content b determine the sentiment analysis feature emo-socre of news comments:

4. The real-time false news event detection method based on multi-source social data as claimed in claim 1, characterized in that: Determine the category of the news event, select the trained news truth judgment model corresponding to the category of the news event, identify the content of the news, and obtain the truthfulness of the news.

5. The real-time false news event detection method based on multi-source social data as claimed in claim 4, characterized in that: Get the original news dataset; The news belonging to the same news event category in the original news data set is used to train the news truth-and-false judgment model. After the training is completed, the trained news truth-and-false judgment models corresponding to different news event categories are obtained.

6. The real-time false news event detection method based on multi-source social data as claimed in claim 5, characterized in that: The obtained news and its comments are added to the original news dataset to update the original news dataset.

7. A real-time fake news event detection system based on multi-source social data, characterized in that: include: A data set acquisition module is used to acquire news and comments associated with news events; A reliability acquisition module is used to determine the reliability of each piece of news based on the relevance of each piece of news to the news event; The comment feature acquisition module is used to determine the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature of the news comment, and to obtain the comment feature of the news event by combining the adjective proportion feature, similarity feature, characteristic word feature and sentiment analysis feature; Specifically, determine the number of all words and adjectives in all news comments in the news event; divide the number of adjectives by the number of all words to obtain the adjective proportion feature of the news comments; Calculate the similarity between all news comments in the news event, divide the number of comments with similarity greater than the set value by the number of all comments, and obtain the similarity feature of the news comments; Determine the most frequently appearing words in the known false news comments in the news event, and determine the characteristic word features of the news comments according to the number of times the most frequently appearing words appear in all the news comments; The sentiment prediction is performed on the content and comments of each news item to obtain the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents in the news event; the sentiment analysis features of the news comments are determined based on the number of positive comments, the number of negative comments, the number of positive contents and the number of negative contents; The content feature acquisition module is used to determine the truth or falsity of each piece of news based on its content, and to perform weighted summation of the truth or falsity of each piece of news using the reliability of each piece of news as a weight to obtain the content feature of the news event; A feature splicing module is used to splice the content features and comment features of the news event to obtain the splicing features of the news event; The news event truth or falsity judgment module is used to determine the truth or falsity judgment result of the news event based on the splicing features of the news event.

8. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the real-time false news event detection method based on multi-source social data as described in any one of claims 1 to 6 are completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of the real-time false news event detection method based on multi-source social data as described in any one of claims 1 to 6.

Images