A rumor detection method integrating opinion divergence and emotional game

Abstract

The present invention belongs to the field of online public opinion analysis, and in particular relates to a rumor detection method that integrates opinion divergence and emotional game, comprising obtaining multi-type topic information including rumors, non-rumors, and comment forwarding messages, as well as relevant user information from an API interface provided by a data platform and a social platform; based on topic and user characteristics, extracting features of user relationships, forwarding relationships, user behaviors, and comment information from each topic, and fusing the features in combination with a polynomial equation; fusing constructed user and comment divergence features and topic features into message influence, constructing emotional mutual influence in combination with a three-party cognitive game theory, using the Dempster combination rule in combination with early fusion and late fusion that introduce an attention mechanism, and establishing a rumor detection model using a long short-term memory network model; the present invention can monitor and control online rumors in a more timely and accurate manner, and guide and suppress them at a reasonable time.

Description

Technical Field

[0001] The present invention belongs to the field of online public opinion analysis, relates to user comment analysis, and in particular to a rumor detection method that integrates opinion divergence and emotional game. Background Art

[0002] Rumors are a widespread social phenomenon, essentially a collective behavior characterized by the dissemination of information. In practical research and application, most researchers consider rumors to be "deliberate fabrications," classifying information confirmed by authoritative sources as fabricated or fictitious. The development of social media has exacerbated the spread of online rumors, making the identification and management of online rumors even more crucial. Rumors are now spreading faster on social platforms than at any time in history. Failure to curb online rumors will seriously jeopardize the public's right to know and social order.

[0003] Rumor detection research holds significant significance and value in social media. First, rumor detection not only provides a prerequisite for limiting its spread but also helps uncover real-life issues of concern to people, as well as difficult-to-detect social phenomena and problems. Second, the results of rumor detection have positive implications for journalists, financial markets, emergency response, and the quality of information on social media. Third, the lifeblood of rumors lies in their spread. Studying the patterns of rumor propagation is an essential component of rumor detection research, and this research contributes to the advancement of information dissemination research.

[0004] In recent years, many scholars have conducted extensive research on rumor detection on social media. Existing rumor detection methods can be divided into three main categories: manual detection, traditional machine learning-based detection, and deep learning-based detection. Early automated rumor detection was primarily based on traditional machine learning, converting rumor detection into a binary classification problem. To overcome the limitations of traditional machine learning, which requires manual feature extraction, researchers are using deep learning to automatically extract information features and perform classification.

[0005] Although many scholars have conducted extensive research on rumor prediction models and achieved quite good results, some challenges still exist:

[0006] 1. Opinion divergence in the review space. Quantifying opinion divergence in the review space requires considering many factors, among which the analysis of text sentiment and the study of review weights are major challenges, making it difficult to accurately quantify opinion divergence.

[0007] 2. The hidden nature of topic sentiment. Analysis of topic sentiment needs to consider the influence of comment sentiment. The positive and negative sentiment of comments can further reflect the true sentiment of the topic. Quantifying the influence of comment sentiment on topic sentiment is a new challenge.

[0008] 3. Complexity of rumors and the diversity of their multiple features. The feature space of rumors is influenced by many factors, requiring the extraction of different features. However, these features often differ significantly, leading to information redundancy or loss during the fusion process. Therefore, how to rationally design the weights of different features is a research topic that needs to be considered. Summary of the Invention

[0009] In order to more accurately detect rumors on social networks, this paper proposes a rumor detection method that integrates opinion divergence and emotional game.

[0010] Obtain information on various topics, including rumors, non-rumors, and comments and forwarding messages, as well as related user information, from the API interfaces provided by data platforms and social platforms;

[0011] Based on topic and user characteristics, we extract features of user relationships, forwarding relationships, user behaviors, and comment information from each topic, and fuse these features using polynomial equations.

[0012] The constructed user and comment disagreement features and topic features are integrated into message influence, and the emotional mutual influence is constructed by combining the three-party cognitive game theory. The Dempster combination rule is combined with the early and late fusion of the attention mechanism, and the long short-term memory network model is used to build a rumor detection model.

[0013] A rumor detection model is established using the long short-term memory network model to detect rumors.

[0014] The present invention can be applied to the early detection and control of rumors. Public opinion departments can monitor and control online rumors more timely and accurately, and guide and suppress them at a reasonable time. It can also be used to promote corporate products and services, helping to quickly promote and spread advertising among target groups, increase advertising exposure and brand awareness, and thus achieve good economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a rumor detection method integrating opinion divergence and emotional game in the present invention;

[0016] Figure 2 Schematic diagram of extracting topic semantic information in the present invention;

[0017] Figure 3 The mechanism of quantifying emotional influence in this invention;

[0018] Figure 4 The present invention reviews the emotional interaction influence model;

[0019] Figure 5 Schematic diagram of the DSMF model of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] The present invention proposes a rumor detection method that integrates opinion divergence and emotional game, which specifically includes the following steps:

[0022] Obtain information on various topics, including rumors, non-rumors, and comments and forwarding messages, as well as related user information, from the API interfaces provided by data platforms and social platforms;

[0023] Based on topic and user characteristics, we extract features from user relationships, forwarding relationships, user behaviors, and comment information from each topic. The extracted features include message attributes and user sentiment influencing factors. Message attributes include basic information attributes and the degree of disagreement in comments. User sentiment influencing factors include the sentiment of user messages and user credibility.

[0024] The long short-term memory network model is used to extract features for early fusion and late fusion from basic information attributes, comment disagreement, user message sentiment, and user credibility;

[0025] Use the obtained features to construct the early fusion results and the late fusion results based on the attention mechanism and classifier;

[0026] The Dempster combination rule is used to fuse the early fusion results and the late fusion results to obtain a mixed fusion result, that is, a prediction result.

[0027] This embodiment illustrates a rumor detection method of the present invention that integrates opinion divergence and emotional game from three aspects: online data acquisition, extraction of relevant attributes, and construction of a prediction model.

[0028] (1) Online data acquisition

[0029] The data can be obtained from public data websites or by using mature social network public APIs. What needs to be obtained here is the behavior records of all participants in the social network topic during its life cycle, as well as the topic participation records and message forwarding. The topic participation records need to obtain the time when the topic was forwarded and commented on, the basic information of the participating users, and the friend relationship information of the participating users (including follow and follow information); the behavior records of topic participants include the information that the user has forwarded and commented on in the past, such as Figure 1 In this embodiment, the main data extracted include: user basic attributes, user historical behavior data and topic message data. Topic message data and user historical behavior data include user relationships, forwarding relationships, user behavior, and comment information. After obtaining the three aspects of data, it is necessary to perform simple data cleaning and storage, specifically including:

[0030] S11: Simple data cleaning: The raw data usually obtained is unstructured and cannot be used directly for data analysis. Simple data cleaning can make most unstructured data structured, so that outliers or null values no longer appear, reducing the inconvenience caused by subsequent calculations;

[0031] S12: Data storage: After simple data clarification, the data needs to be stored in a database, and the data is further normalized through the table structure. The database can also greatly improve the data retrieval efficiency and the mapping of relationships between tables.

[0032] (2) Extraction of relevant attributes

[0033] Based on the acquired data, we extract message comments and forwarding information, basic user information, user follow lists, and user historical behavior. We can also divide the time period into specific periods based on the characteristics of the data.

[0034] Exploring more accurate rumor detection essentially involves extracting relevant features from both the message and user levels and modeling and analyzing these features. Based on this, this paper explores the deep characteristics of topic messages on social networks from two perspectives: message attributes and user sentiment. Specifically, the following are the key factors:

[0035] 1) Message attributes

[0036] The basic attributes of information include the basic semantic content of the information, the number of emoticons contained, the number of URLs contained, and the number of hashtags, which play a certain auxiliary role in the identification of rumors. Therefore, in this embodiment, the basic attributes of information are defined as:

[0037] MAttri(m i )={text(m i),emojiNums(m i ),urlNums(m i ),tagNums(m i )} (1)

[0038] Among them, text(m i ) represents user comment m i The basic semantic content of emojiNums(m i ) represents user comment m i The number of emojis included, urlNums(m i ) represents user comment m i The number of URLs included, tagNums(m i ) represents user comment m i The number of hashTags included.

[0039] There is a significant difference in the divergence of comments under rumors and normal information. The divergence of comments is related to the sentiment value of the comments. At the same time, the sentiment intensity factor is added, that is, the ratio of the number of likes of the comment to the total number of likes of all comments. Therefore, in this embodiment, the divergence of comments is defined as:

[0040]

[0041] Among them, W senti (m i ) is the user comment m i Senti(m i ) is the user comment m i The sentiment value of P{y=Senti(m i )|C} means that when the user's sentiment value y is Senti(m i ), the probability of the user's sentiment value being distributed in the comment set C.

[0042] 2) Factors affecting user emotions

[0043] The emotional tendencies expressed by users through comments play an important role in detecting rumor topics. By analyzing the sentiment of user comments, the internal driving factors of user emotions are defined as follows:

[0044]

[0045]

[0046] Among them, posWord(m i ) is the user comment m iThe number of positive words in the word, negWord(m i ) is the user comment m i The number of negative words in the text; represents the internal factors influencing users’ support for topics, Indicates the internal factors that influence users' opposition to a topic.

[0047] A user's emotional tendency is often influenced by external friend groups. Therefore, the external factors affecting the user's emotions are defined based on the emotional attitudes of users participating in topic discussions among the user's friends:

[0048]

[0049]

[0050] Among them, FriNum(u i ) represents user u i The number of friends who participated in the topic, FriSupNum(u i ) represents the number of the user’s friends who hold a positive attitude. represents the external factors that influence users’ support for topics, Indicates external factors that influence users' opposition to a topic.

[0051] The emotional attitudes of high-reputation users in the topic space are often more valuable for rumor topic detection than those of ordinary users. Therefore, the user reputation in the topic space is defined as:

[0052] Cre(u i )=baseCre(u i )+authCra(u i )+α×fansNum(u i ) (7)

[0053] Among them, baseCre(u i ) represents the basic credibility of all users, authCra(u i ) indicates the official account level of the user, fansNum(u i ) represents the number of followers of a user. The emotional attitudes of high-quality topic commentators are important for identifying rumor topics. High-quality users often represent official accounts or celebrity accounts. However, official accounts often have higher credibility than celebrity accounts. Therefore, the attenuation factor α∈[0,1] is used to weaken the influence of celebrity accounts on the overall credibility of users.

[0054] 3) Sentiment of user messages

[0055] A user's historical behavioral characteristics can, to a certain extent, reflect their interests and opinions on topics over different time periods. We can treat the social content posted by a user during a specific time period as an article. Based on the fluctuations in trending topics, we can segment the time period into smaller segments. The user's social content within each segment is then represented as a paragraph within the article, with each paragraph having a clear theme. Furthermore, each short text post from a user on Weibo can be considered a sentence. Using the Doc2vec algorithm, we can learn to represent a user's historical social content as a vector reflecting their social interests.

[0056] Considering the particularity of Chinese grammar, there is undoubtedly a problem if a fixed-length sliding window is used directly to obtain the context of the target word. To address this problem, this embodiment makes certain improvements to the context of the target word, including: first, performing Chinese word segmentation and part-of-speech tagging on the Weibo content posted by the user, removing useless words and stop words, and obtaining candidate keywords for the user's Weibo content; then using the TF-IDF algorithm to calculate the weight of each candidate keyword, extracting the information center keywords of the user's social content, thereby eliminating noise in the social content and retaining the user's main social interests. The idea of word frequency statistics is as follows:

[0057]

[0058] Where count(d,w) is the number of times word w appears in document d, and count(d,*) is the total number of words in document d.

[0059] The inverse document frequency is calculated as follows:

[0060]

[0061] Where N(w) is the total number of documents in the corpus and N(w) is the number of documents in which the word appears.

[0062] However, in some special cases, such as when a rare word has never appeared in the corpus, the denominator N(w) = 0. In this case, the IDF needs to be smoothed so that the word that does not appear in the corpus can also get a suitable IDF value. The IDF smoothing formula is as follows:

[0063]

[0064] After obtaining the keyword sequence representing the user's historical social content, the Doc2vec algorithm is used to learn it and output the user's historical behavior feature vector:

[0065] D=n×F b (11)

[0066] Among them, n is the number of users under the rumor topic, and F is the historical social content interest feature representation vector of the corresponding user.

[0067] Then, we use the sentiment dictionary to get the sentiment score of each text, and let the text with a positive sentiment score be supportive and the text with a negative sentiment score be opposed. The sentiment calculation formula is:

[0068]

[0069] Among them, w t Represents the words after the blog text is segmented, SentiDic represents the sentiment dictionary containing sentiment values, SentiDeg(w t ) represents w t The sentiment value of AdvDeg(w t ) indicates modification of w t The degree value of the degree adverb.

[0070] 4) Correction of the emotional level of user messages

[0071] The emotional tendency shown by users is antagonistic. The emotional influence function is affected by both internal and external factors. That is, users’ antagonistic emotions are often affected by their inner drive and external environment. Therefore, Figure 3 ,Combining the internal and external factors of users, using the multiple linear regression algorithm,,construct the user emotional influence function as follows:

[0072]

[0073]

[0074] Among them, Inf sup (u i ) represents the emotional influence of user u in supporting the topic, Inf obj (u i ) represents the emotional influence of user u against the topic; and are the internal driving factors of users’ sentiments in support and opposition to topics, and are the external influencing factors of users' emotions in support and opposition to the topic, respectively; ρ0, ρ1, and ρ2 are the partial regression coefficients of the trained multiple linear regression algorithm, and ρ1 and ρ2 reflect the proportion of internal driving factors and external influencing factors in the influence of users' emotions.

[0075] Considering that the emotional tendency of comments can more truly reflect the emotional tendency of the topic, this paper quantifies the influence of positive and negative comments on the sentiment of the topic through evolutionary game theory. Users' comments on the topic will show support or opposition, so this paper has two game strategies: "user comments support" and "user comments oppose". Let P1 and P2 represent the target topic m respectively. i The ratio of comments supporting and opposing in the comments, neutral comments on the target topic m i The influence of is not significant, so it is not considered, so P1+P2=1. The profit functions of these two strategies are:

[0076] Pro sup (u i )=P1×Inf sup (u i ) (15)

[0077] Pro obj (u i )=P2×Inf obj (u i ) (16)

[0078] Furthermore, the present invention measures the interaction between positive and negative emotions through evolutionary game theory:

[0079]

[0080]

[0081] Among them, Mut sup (u i ) and Mut obj (u i ) represents the positive and negative messages after the game for user u i Emotional impact.

[0082] Finally, considering that users are affected by other factors when expressing their emotions, the user's emotions are corrected through the emotional interaction to dig out the user's true emotions, such as Figure 4 The correction function is:

[0083]

[0084]

[0085] Among them, α and β are adjustment coefficients, N Senti (m i ) is the sentiment tendency of the user's comments before adjustment, Senti(m i ) is the adjusted emotional tendency.

[0086] (3) Construction of prediction model

[0087] In this embodiment, the processing of the text to be predicted includes the following stages:

[0088] The neural network model is used to extract features from basic information attributes, comment opinion divergence, user message sentiment, and user credibility. In this embodiment, the long short-term memory network model is used for feature extraction. The structure of the long short-term memory network model is as follows: Figure 2 As shown in the figure, Ct-1 represents the cell state at the previous moment (t-1); X represents the input data, which can be a vector or matrix containing the input features of the current time step; ft represents the forget gate, which controls which information needs to be forgotten from the cell state; it represents the input gate, which controls which information needs to be added to the cell state; Ot represents the output gate, which controls how much information in the hidden state of the current time step will be output; ht-1 represents the hidden state of the previous time step; Xt represents the input data of the current time step; σ represents an activation function used to calculate the forget gate, input gate and output gate; tanh represents a nonlinear activation function used to calculate new candidate cell states and update hidden states; | represents the connection operation, which is used to jointly process the previous cell state and hidden state with the current input data.

[0089] Based on the Dempster combination rule in DS evidence theory, a feature hybrid fusion model DSMF is constructed, and the above information is integrated to detect rumors on topics.

[0090] The applicability of single multi-feature early fusion and late fusion is poor. This paper combines the two fusion methods through the Dempster combination rule, and defines the output results of early fusion and late fusion as evidence in evidence theory. The basic probability assignment (BPA) is the category probability P E (m i ), P L (m i ), and use Dempster's combination rule to get the final rumor probability P(m i ), to determine whether the blog post is a rumor. Under the framework of evidence theory, two sets of independent mass functions can be fused using the Dempster combination rule.

[0091] like Figure 5, through the long short-term memory network model, the features R1, R2, R3, and R4 are extracted from the basic attributes of information, the degree of disagreement of comments, the sentiment of user messages, and the user credibility. R1, R2, R3, and R4 are input into the self-attention mechanism to obtain the weight corresponding to each feature. The weight is used to perform a weighted sum of R1, R2, R3, and R4 to obtain the preliminary fusion result;

[0092] R1, R2, R3, and R4 are respectively input into a classifier consisting of a fully connected layer for processing. The output of each classifier is weighted and summed using the attention weight obtained by the self-attention mechanism to obtain the late fusion result;

[0093] The Dempster combination rule is used to fuse the early fusion results and the late fusion results to obtain the mixed fusion results, including:

[0094]

[0095]

[0096] Among them, P(n,r|topic) is the probability distribution of rumor and non-rumor of information, A is the value space of P(n,r|topic), expressed as A∈{rumor, non-rumor}; m E (B) is the probability distribution of the previous fusion evidence B, that is, the previous fusion result; m L (C) is the probability distribution of the late fusion evidence C, that is, the late fusion result; B is the feature space of the early fusion, and C is the feature space of the late fusion; K is the degree of conflict between the evidences.

[0097] The output of this model is Y = P (n, r | topic), because this article discusses a binary classification problem, the probability value

[0098] The larger category is the prediction result, which can be defined as follows:

[0099]

[0100] Among them, Y represents the output of the rumor detection model. When Y = 1, it means that the topic is a rumor topic, and when Y = 0, it means that the topic is not a rumor topic; P(r|topic) is the probability that the current topic is a rumor; P(n|topic) is the probability that the current topic is not a rumor; P(n,r|topic) is the set of probabilities of the current topic being a rumor and a non-rumor.

[0101] In the above embodiment, P(n, r|topic) is a set of probabilities, including two probability values, namely the probability P(r|topic) that the current topic is a rumor and the probability P(n|topic) that the current topic is not a rumor. When the model finally determines whether the input topic is a rumor or not, the larger value of P(n, r|topic) prevails. It can be understood that: when P(n|topic) is greater than P(r|topic), the output value is P(n|topic), that is, the current topic is not a rumor; otherwise, the current topic is a rumor topic.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A rumor detection method integrating opinion divergence and emotional game, characterized by: The specific steps include: Obtain information on various topics, including rumors, non-rumors, and comments and forwarding messages, as well as related user information, from the API interfaces provided by data platforms and social platforms; Based on topic and user characteristics, we extract features from user relationships, forwarding relationships, user behaviors, and comment information from each topic. The extracted features include message attributes and user sentiment influencing factors. Message attributes include basic information attributes and the degree of disagreement in comments. User sentiment influencing factors include the sentiment of user messages and user credibility. The long short-term memory network model is used to extract features for early fusion and late fusion from basic information attributes, comment disagreement, user message sentiment, and user credibility; The differences in opinions include: Among them, W senti (m i ) is the user comment m i Senti(m i ) is the user comment m i Sentiment value; P{y=Senti(m i )|C} means that when the user's sentiment value y is Senti(m i ), the probability of the user's sentiment value being distributed in the comment set C; The calculation of the sentiment of user messages includes: Among them, N Senti (m i ) represents user comment m i The emotional degree of w t Indicates user comment m i The words after word segmentation; SentiDic represents the sentiment dictionary containing sentiment values; SentiDeg(w t ) represents the w score calculation based on the BosonNLP sentiment dictionary t The sentiment value of AdvDeg(w t ) indicates modification of w t The degree value of the degree adverb; Use the obtained features to construct the early fusion results and the late fusion results based on the attention mechanism and classifier; The Dempster combination rule is used to fuse the early fusion results and the late fusion results to obtain a mixed fusion result, that is, a prediction result.

2. A rumor detection method integrating opinion divergence and emotional game according to claim 1, characterized in that: Basic information attributes include: MAttri(m i )= {text(m i ),emojiNums(m i ),urlNums(m i ),tagNums(m i )}; Among them, text(m i ) represents user comment m i The basic semantic content of emojiNums(m i ) represents user comment m i The number of emojis included, urlNums(m i ) represents user comment m i The number of URLs included, tagNums(m i ) represents user comment m i The number of hashTags included.

3. The rumor detection method integrating opinion divergence and emotional game according to claim 1 is characterized in that: The emotional level of user messages is modified through emotional interaction, including: Combining the internal and external factors of users, using the multiple linear regression algorithm, construct the user emotional influence function, and calculate the user u according to the function. i the emotional impact of supporting and opposing the topic; Get support and opposition to the current user's comments m i The ratio of user u i Calculation of sentiment influence of support and opposition topics User support and opposition User comments m i income; Based on user support and opposition user comments m i Calculate the revenue of user u i The mutual influence of positive and negative emotions in support and opposition of topics; According to user u i The positive and negative emotional interaction of the support and opposition topics is used to modify the emotional degree of user messages, including: Among them, Senti(m i ) is the sentiment of the user message after correction; α and β are adjustment coefficients; Mut sup (u i ) is user u i The positive and negative emotional interaction of the supporting topics, Mut obj (u i ) is user u i The positive and negative emotional interactions of opposing topics.

4. The rumor detection method integrating opinion divergence and emotional game according to claim 3 is characterized in that: User i The positive and negative sentiment interaction of the supporting topic is expressed as: User i The positive and negative emotional interaction of the opposing topic is expressed as: Among them, Pro sup (u i ) represents the benefits of users supporting the topic, denoted as Pro sup (u i )=P1×Inf sup (u i ), Inf sup (u i ) is user u i The emotional influence of the supporting topic, P1 is the proportion of supporting topics; Pro obj (u i ) is user u i The emotional influence of the opposing topic, denoted as Pro obj (u i )=P2×Inf obj (u i ), Inf obj (u i ) is user u i The emotional influence of opposing topics, P2 is the proportion of opposing topics.

5. The rumor detection method integrating opinion divergence and emotional game according to claim 4 is characterized in that: User i The sentiment influence of the supporting topic is expressed as: User i The emotional influence of opposing topics is expressed as: in, For user u i The partial regression coefficients of the multivariate linear regression algorithm trained in the sentiment influence function of the supporting topic; For user u i The internal driving factors supporting topic sentiment are expressed as FriNum(u i ) represents user u i The number of friends who participated in the topic, FriSupNum(u i ) represents user u i The number of friends with positive attitudes; is the internal influencing factor of user support topic, expressed as posWord(m i ) is the user comment m i The number of positive words in the word, negWord(m i ) is the user comment m i The number of negative words in the text; For user u i The partial regression coefficients of the multivariate linear regression algorithm trained on the sentiment influence function of the opposing topic; is the internal influencing factor of users’ opposition to the topic, expressed as is the external influencing factor of users’ opposition to the topic, expressed as 6. The rumor detection method integrating opinion divergence and emotional game according to claim 1 is characterized in that: User reputation includes: Cre(u i )=baseCre(u i )+authCra(u i )+α×fansNum(u i ); Among them, baseCre(u i ) represents user u i Basic credibility, authCra(u i ) represents user u i Official account level, fansNum(u i ) represents user u i The number of fans; α∈[0,1] is the attenuation factor.

7. The rumor detection method integrating opinion divergence and emotional game according to claim 2 is characterized in that: Obtaining the early fusion features and the late fusion features and fusing the obtained early fusion features and the late fusion features using the Dempster combination rule specifically includes the following steps: The basic attributes of information, the degree of disagreement of comments, user credibility, and user sentiment values are spliced together and input into the pre-trained long short-term memory network model to extract the corresponding features; The basic attributes of information, the divergence of comments, user credibility, and user sentiment are used to obtain corresponding attention weights through the attention mechanism; The attention weights are used to perform weighted summation on the features extracted by the pre-trained long short-term memory network model to obtain the preliminary fusion results; The basic information attributes, review opinion divergence, user credibility, and user sentiment value features extracted by the pre-trained long short-term memory network model are input into a fully connected layer respectively; Use the attention weights to perform weighted summation on the output of the fully connected layer to obtain the late fusion result; The late fusion results and the late fusion results are fused based on the Dempster combination rule to obtain the rumor and non-rumor probability distribution of an information, including: Among them, P(n, r|topic) is the probability distribution of rumor and non-rumor of information, A is the value space of P(n, r|topic), expressed as A∈{rumor, non-rumor}; m E (B) is the probability distribution of the previous fusion evidence B, that is, the previous fusion result; m L (C) is the probability distribution of the late fusion evidence C, that is, the late fusion result; B is the feature space of the early fusion, and C is the feature space of the late fusion; K is the degree of conflict between features.

8. The rumor detection method integrating opinion divergence and emotional game according to claim 7 is characterized in that: When using the long short-term memory network model to build a rumor detection model, the sentiment tendency Senti(m i ), comment opinion disagreement MDiff(m i ), basic information attributes MAttri (m i ) and user credibility Cre(u i ) Four features are used as the input of the model, and the rumor detection model is expressed as: Among them, Y represents the output of the rumor detection model. When Y = 1, it means that the topic is a rumor topic, and when Y = 0, it means that the topic is not a rumor topic; P(r|topic) is the probability that the current topic is a rumor; P(n|topic) is the probability that the current topic is not a rumor; P(n, r|topic) is the set of probabilities that the current topic is a rumor and a non-rumor.

Images