Comprehensive analysis method for social media hot events

Abstract

The invention discloses a comprehensive analysis method for social media hot events, belongs to the technical field of hot event analysis, and can solve the problem that an analysis result has deviation due to the fact that an existing hot event analysis method is difficult to comprehensively analyze hot events. The method comprises the steps that S1, an analysis framework of a hot event is constructed from multiple analysis dimensions, each analysis dimension comprises multiple analysis directions, and each analysis direction corresponds to an analysis method; s2, analyzing the hot events on the target social media platform by using the analysis framework to obtain an analysis result corresponding to each analysis direction; and S3, according to the analysis dimension, the analysis direction and the analysis method corresponding to each analysis result, utilizing a large language model to generate an overall analysis report of the hot event. The method is used for analyzing the hot events on the social media.

Description

Technical Field

[0001] The invention relates to a comprehensive analysis method for hot events in social media, and belongs to the technical field of hot event analysis. Background Art

[0002] At present, the trend of analyzing hot events on social media is to analyze the content and user behavior related to hot events on social media. However, the existing hot event analysis methods can only analyze in a single analysis direction during one analysis process, such as only analyzing the content or only analyzing the user behavior; at the same time, the existing analysis methods often only analyze from one analysis dimension during one analysis process, such as only analyzing from the macro dimension or only analyzing from the micro dimension.

[0003] Since the analysis process of existing analysis methods does not take into account multiple analysis directions and multiple analysis dimensions, it is difficult to conduct a comprehensive analysis of hot events. As a result, the analysis results are unable to reflect the interrelationships and mutual influences between multiple analysis directions and multiple analysis dimensions, resulting in a large deviation between the analysis results and the actual situation, and the accuracy and reliability of the analysis results are low. Summary of the invention

[0004] The present invention provides a comprehensive analysis method for hot events in social media, which can solve the problem that the existing hot event analysis methods are difficult to comprehensively analyze hot events, resulting in large deviations in analysis results.

[0005] The present invention provides a comprehensive analysis method for hot events in social media, the method comprising:

[0006] S1. Construct an analytical framework for hot events from multiple analytical dimensions. Each analytical dimension includes multiple analytical directions, and each analytical direction corresponds to an analytical method.

[0007] S2. Analyze hot events on the target social media platform using the analysis framework to obtain analysis results corresponding to each analysis direction;

[0008] S3. Generate an overall analysis report of the hot event based on the analysis dimension, analysis direction and analysis method corresponding to each analysis result and using a large language model.

[0009] Optionally, the multiple analysis dimensions include a structure-macro analysis dimension, a structure-micro analysis dimension, a content-macro analysis dimension, and a content-micro analysis dimension.

[0010] Optionally, S3 is specifically:

[0011] S31. Based on each analysis result and its corresponding analysis method, generate an explanation text of each analysis method using a large language model, and based on each analysis result and its corresponding analysis method and the explanation text of the analysis method, generate an analysis report for each analysis direction using a large language model;

[0012] S32. Based on the analysis report of each analysis direction, a large language model is used to generate an overall analysis report of the hot event.

[0013] Optionally, the S32 is specifically:

[0014] S321, using a large language model to generate explanatory text for each analysis dimension;

[0015] S322. Generate an overall analysis report of the hot event using a large language model based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension.

[0016] Optionally, the S322 is specifically:

[0017] Based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension, the analysis report of each dimension is generated using the large language model;

[0018] Based on the analysis report of each dimension, a large language model is used to generate an overall analysis report of the hot event.

[0019] Optionally, the analysis report of each dimension is used to generate an overall analysis report of the hot event using a large language model, specifically:

[0020] Generate an explanation text of the analysis framework using a large language model;

[0021] Based on the explanatory text of the analysis framework and the analysis report of each dimension, a large language model is used to generate an overall analysis report of the hot event.

[0022] Optionally, the analysis directions of the content-macro analysis dimension include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text theme analysis.

[0023] Optionally, the analysis method corresponding to the change analysis of the community terms is:

[0024] Determine a word occurrence probability vector for each time period in a continuous time period based on the text data of the target social media platform;

[0025] The cross entropy of the word occurrence probability vectors in multiple time periods is calculated, and the change of community terms is determined according to the cross entropy.

[0026] Optionally, the analysis directions of the structure-macro analysis dimension include: indicator change analysis of the social media platform, long-term relationship analysis between multiple indicators and short-term status analysis of each indicator.

[0027] Optionally, before S2, the method further includes:

[0028] Calculate the importance index of each social media platform based on the number of users and data volume of each social media platform among multiple social media platforms;

[0029] A target social media platform is determined according to the importance index.

[0030] The beneficial effects that the present invention can produce include:

[0031] The present invention constructs an analysis framework including multiple analysis dimensions and multiple analysis directions, and uses the analysis framework to analyze hot events to obtain analysis results of multiple analysis directions, and then uses a large language model to generate an overall analysis report of hot events based on the analysis results. In this way, the interrelationships and mutual influences between multiple analysis directions and multiple analysis dimensions are fully considered in one analysis process, and a comprehensive and integrated analysis of hot events is achieved, which is conducive to improving the accuracy and reliability of the analysis results. At the same time, the present invention uses a large language model and combines the analysis framework to integrate the analysis results to form an overall analysis report of hot events, realizing intelligent analysis, which is conducive to improving analysis efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a comprehensive analysis method for hot social media events provided by an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of selecting keywords for data screening of ChatGPT events provided in an embodiment of the present invention;

[0034] Figure 3 The analysis results corresponding to the structure-macro analysis dimension of the ChatGPT event provided by the embodiment of the present invention;

[0035] Figure 4 The analysis results corresponding to the structure-micro analysis dimension of the ChatGPT event provided by the embodiment of the present invention;

[0036] Figure 5 A schematic diagram of the structure of the prompt engineering structure provided by an embodiment of the present invention;

[0037] Figure 6 A schematic diagram of the process of generating an analysis report of an analysis direction provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention is described in detail below in conjunction with embodiments, but the present invention is not limited to these embodiments.

[0039] The embodiment of the present invention provides a comprehensive analysis method for hot events in social media. Figure 1 As shown, the method includes:

[0040] S1. Construct an analytical framework for hot events from multiple analytical dimensions. Each analytical dimension includes multiple analytical directions, and each analytical direction corresponds to an analytical method.

[0041] S2. Use the analysis framework to analyze hot events on the target social media platform and obtain the analysis results corresponding to each analysis direction;

[0042] S3. Based on the analysis dimension, analysis direction and analysis method corresponding to each analysis result, and using the large language model, an overall analysis report of hot events is generated.

[0043] Specifically, the multiple analysis dimensions may include a structure-macro analysis dimension, a structure-micro analysis dimension, a content-macro analysis dimension, and a content-micro analysis dimension.

[0044] The definition of the structure-macro analysis dimension in this embodiment is:

[0045] Focus on the macro-structure analysis of social media platforms. For the overall data set of social media platforms, multiple indicators are analyzed, including the number of posts, comments, users, new users, topics, etc. The analysis methods used include time series prediction algorithm, piecewise regression algorithm, vector error correction model (VECM) algorithm, etc., to analyze the changes of various indicators on the timeline and the possible impact on each other.

[0046] The definition of the structure-micro analysis dimension in this embodiment is:

[0047] Focus on the segmented user groups of social media platforms and analyze the differences in indicators of different segmented user groups. For example, the overall data set of the social media platform can be divided from the perspective of user activity and user active time, and then the data can be statistically analyzed from the perspective of multiple indicators to obtain the differences between different segmented user groups.

[0048] In this embodiment, the content-macro analysis dimension is defined as:

[0049] Focus on the overall text content analysis of social media platforms. For example, we can use the unigram language model to analyze the changes in the language habits of user groups in the overall data set of social media platforms; use text sentiment analysis algorithms, text topic analysis algorithms, etc. to label the text and observe the changes in the number of labels on the timeline.

[0050] In this embodiment, the content-micro analysis dimension is defined as:

[0051] Focus on the segmented user groups and segmented text content of social media platforms, and analyze the differences in text content of different segmented user groups. For example, text sentiment analysis algorithms, text topic analysis algorithms, etc. can be used to label the text in the overall data set of the social media platform, and at the same time, the text can be divided from the perspectives of new users, active users, and topics to analyze the changes in text sentiment under new and old user groups and different topics.

[0052] In summary, the analysis directions of the structural-macro analysis dimension can include: indicator change analysis of social media platforms, long-term relationship analysis between multiple indicators, and short-term status analysis of each indicator.

[0053] The analysis direction of the structural-micro analysis dimension may include: indicator change analysis for segmented user groups, long-term relationship analysis between multiple indicators for segmented user groups, and short-term status analysis of each indicator. For example, user groups can be divided according to the number of user comments, and the changes in the number of segmented user groups and the changes in the proportion of indicators such as the number of posts and the number of comments can be analyzed; user groups can also be divided according to the user's ability to post continuously, and the changes in the number of user groups that post continuously can be observed.

[0054] The analysis directions of the content-macro analysis dimension can include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text theme analysis.

[0055] The analysis direction of the content-micro analysis dimension may include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text topic analysis for segmented user groups. For example, user groups may be divided according to the time sequence of user participation in topics, and the text sentiment of different segmented user groups may be analyzed; text data may also be divided by topic, and the text sentiment in different topics may be analyzed.

[0056] In practice, before conducting multi-dimensional analysis, it is often necessary to first determine the target social media platforms and hot events to be analyzed.

[0057] Specifically, before S2, the comprehensive analysis method of the present invention may further include:

[0058] Calculate the importance index of each social media platform based on the number of users and data volume of each social media platform among multiple social media platforms;

[0059] Identify target social media platforms based on their importance index.

[0060] Specifically, this embodiment measures the importance of each social media platform based on the number of users and the amount of data on the platform. The higher the number of users of a social media platform, the more important the social media platform is; the more data a social media platform has, the more reliable the data on the social media platform is. This embodiment determines the importance index of each social media platform by calculating the weighted average of the number of users and the amount of data, and then determines the importance of each social media platform.

[0061] Specifically, the calculation formula of the importance index is:

[0062]

[0063] In formula (1), I i is the importance index of social media platform i; J i is the number of users of social media platform i; M i is the amount of data on social media platform i; Ns is the total number of social media platforms; j = 1, 2, ... N s ;ω u and ω d are the weight parameters of the number of users and the amount of data, respectively, and ω u +ω d =1.

[0064] In this embodiment, a social media platform with the highest importance index is selected from multiple social media platforms as a target social media platform.

[0065] Furthermore, this embodiment measures the hotness of an event based on its widespread dissemination and social influence, and then determines whether it is a hot event. The wider the dissemination and social influence of an event, the higher the popularity of the event. Specifically, the keywords of the event are searched on the target social media platform, and the hotness of each event is determined by calculating the weighted average of the number of posts, the number of post comments, and the number of post tags.

[0066] Specifically, the calculation formula for the hotspot degree is:

[0067]

[0068] In formula (2), A i is the hotness of event i; W i is the number of posts for event i; B i is the number of comments on the post of event i;i is the number of post tags for event i; Ne is the total number of events; j = 1, 2, ... Ne; ω p ,ω c ,ω k are the weight parameters of the number of posts, the number of post comments, and the number of post tags, respectively, and ω p +ω c +ω k =1.

[0069] In this embodiment, the event with the highest popularity among multiple events on the target social media platform is selected as the hot event.

[0070] After determining the target social media platform and hot events, this embodiment collects relevant data of the hot events from the target social media platform, and screens and cleans the data.

[0071] Specifically, this embodiment uses crawlers, data downloading and other methods to collect data. For example, crawler tools such as BeautifulSoup can be used to collect relevant data of hot events on the target social media platform, or the data interface officially provided by the target social media platform can be used to directly download relevant data.

[0072] Specifically, this embodiment uses keywords to filter data to obtain data that is strongly related to hot events. The selection of keywords can be determined based on hot events, including direct keywords of the hot events themselves, and indirect keywords related to the hot events. Figure 2 As shown, taking the ChatGPT event as an example, according to the field of ChatGPT technology and its related fields, keywords such as artificial intelligence, natural language processing, deep learning, and large language models can be selected for data screening.

[0073] After obtaining the relevant data of hot events on the target social media platform, the hot events can be analyzed according to the above analysis framework.

[0074] The following describes in detail the analysis directions that can be included in each analysis dimension and the analysis methods corresponding to each analysis direction.

[0075] (I) Analysis directions of the structural-macro analysis dimension can include: analysis of indicator changes of social media platforms, analysis of long-term relationships between multiple indicators, and analysis of the short-term status of each indicator.

[0076] Specifically, the segmented regression algorithm can be used to determine the mutation point of the indicator on the timeline, and the Prophet algorithm can be used to analyze the specific differences before and after the key nodes of the mutation point, so as to realize the indicator change analysis of the social media platform. At the same time, the VECM algorithm can be used to analyze the long-term relationship between indicators and the short-term status analysis of each indicator.

[0077] 1. Use the segmented regression algorithm to determine the mutation point of the indicator on the timeline.

[0078] 1) Select the indicators that need to be analyzed for mutation points, such as the number of comments, the number of posts, the number of users, etc., and organize these indicators into a data set in chronological order: (x 1 ,y 1 ), (x 2 ,y 2 ),...,(x n ,y n ). Among them, x i is the time corresponding to time period i, y i is the indicator data corresponding to time period i, x i and i Form a data set (x i ,y i ), n is the total number of data groups contained in the data set, i = 1, 2, ... n.

[0079] 2) Clean up missing values ​​and outliers in the dataset.

[0080] 3) Using the segmented regression algorithm, assuming that there is a mutation point τ, the segmented regression formula is:

[0081]

[0082] In formula (3), x is time; y is the indicator data corresponding to time x; τ is the mutation point; α 1 and α 2 are the intercepts before and after the mutation point, respectively; β 1 and β 2 are the slopes before and after the mutation point, respectively.

[0083] 4) Use the residual sum of squares to estimate the parameter α in formula (3) 1 , α 2 , β 1 , β 2 , τ, the residual sum of squares is calculated as:

[0084]

[0085] In formula (4), SSE is the residual sum of squares; u is the index position of the mutation point τ in the data set, x u <τ≤x u+1 ; n is the total number of data groups contained in the data set; x i is the time corresponding to time period i; y i is the indicator data corresponding to time period i; α 1 and α 2are the intercepts before and after the mutation point, respectively; β 1 and β 2 are the slopes before and after the mutation point, respectively.

[0086] 5) Combined with hot events, analyze the changing trends of indicators at the mutation point and before and after it to obtain the analysis results of the mutation point.

[0087] 2. Use the Prophet algorithm to analyze the specific differences of indicators before and after the key nodes in the mutation point.

[0088] 1) Select the more obvious mutation point obtained in the previous step as the key node t, and use the Prophet algorithm to fit the data before the key node t. The fitting formula is:

[0089] y(t)=g(t)+s(t)+h(t)+∈ t (5)

[0090] In formula (5), y(t) is the predicted value at the key node t; g(t) is the trend predicted value; s(t) is the seasonal part predicted value; h(t) is the holiday effect predicted value; ∈ t is the error term.

[0091] Among them, the calculation formulas of g(t), s(t) and h(t) are as follows;

[0092]

[0093] In formula (6), g(t) is the trend prediction value; γ is the maximum asymptotic value of the data curve; k g is the growth rate of the data curve; ε is the midpoint of the data curve.

[0094]

[0095] In formula (7), s(t) is the seasonal part prediction value; a i and b i are two parameters of the Fourier series; t is the time; Period is the period of the time series; U is the number of periods, i = 1, 2,…, U.

[0096] Z(t)=[1(t∈D 1 ),...,1(t∈D θ )],K=(K 1 , ..., K θ ) (8)

[0097] h(t)=Z(t)K (9)

[0098] In formula (8) and formula (9), h(t) is the predicted value of holiday effect; K~Normal(0,v 2 ), v is the holiday index, that is, the normal distribution of K is affected by the holiday index v, K i is the impact range of holidays; θ is the total number of holidays; D i is the period before and after the holiday; Z(t) is D i Vector composed of.

[0099] According to the above formula, the predicted value y(t) at the key node t can be calculated.

[0100] 2) Calculate the square of the difference between the predicted value y(t) and the actual value at the key node t to reflect the degree of deviation between the predicted result and the actual situation.

[0101] 3) Select key nodes with larger differences, analyze them in combination with the development of hot events, and obtain the analysis results of the specific differences before and after the key nodes in the mutation point.

[0102] 3. Use the VECM algorithm to analyze the long-term relationship between indicators and the short-term status analysis of each indicator.

[0103] 1) Establish the VECM equation. Assume that the number of indicators is δ and the time series data of multiple indicators are , then the time series matrix composed of the data of multiple indicators is:

[0104]

[0105] In formula (10), X t It is a time series matrix composed of data of multiple indicators; is the time series data of the ith indicator; t is the time.

[0106] Then, the formula of the VECM equation is:

[0107]

[0108] In formula (11), ΔX t is the variable vector after difference at time t; ΔX t-i is the variable vector after difference at time ti; Π is the cointegration matrix; represents the long-term equilibrium relationship; Γ i is the short-term adjustment coefficient matrix; ∈ t is the error term; ζ is the lag order.

[0109] 2) Use the Johansen cointegration test. Estimating equation (11) yields the estimated value of ∏ right Conduct a trace test or a maximum eigenvalue test. If the test passes, it indicates that there is a long-term relationship between the indicators. The value of is the long-term equilibrium coefficient, which indicates the degree to which the current indicator is affected by other indicators in the long term, thereby obtaining the long-term relationship analysis results between the indicators.

[0110] 3) According to Γ i Analyze the short-term adjustments made by indicators to fluctuations that deviate from the long-term equilibrium state and obtain the short-term state analysis results of each indicator.

[0111] (II) Analysis directions of the structural-micro-analysis dimension may include: indicator change analysis for segmented user groups, long-term relationship analysis between multiple indicators for segmented user groups, and short-term status analysis of each indicator. For example, user groups may be divided according to the number of user comments, and the changes in the number of segmented user groups, as well as the changes in the proportion of indicators such as the number of posts and the number of comments, may be analyzed; user groups may also be divided according to the user's ability to post continuously, and the changes in the number of user groups that post continuously may be observed.

[0112] (III) The analysis directions of the content-macro analysis dimension may include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text theme analysis.

[0113] Specifically, you can use the unigram language model to obtain the cross entropy before and after different time nodes to analyze the changes in community terms. Observe the changes in the number of keywords through word frequency statistics. Use text sentiment analysis and text topic analysis to label the text with sentiment and topic tags, count the changes in sentiment and topic comments on the timeline, and observe the overall user sentiment trend and topic trend.

[0114] The analysis method corresponding to the change analysis of the above-mentioned community terms can be:

[0115] Determine the word occurrence probability vector of each period in the continuous period based on the text data of the target social media platform; calculate the cross entropy of the word occurrence probability vectors of multiple periods, and determine the change of community terms based on the cross entropy.

[0116] 1. Use the unigram language model to obtain the cross entropy before and after different time nodes to analyze the changes in community terms.

[0117] 1) Divide the text data of the target social media platform into certain time intervals to obtain the text library Text of each time period in the continuous period t .

[0118] 2) Use the unigram language model to count the text library Text t The probability of occurrence of words in the text library Textt The vocabulary is V 1 ={w 1 , w 2 , ..., w m}, where each word w i The probability of occurrence of words (i=1,2…m) is p(w i ), then the text library Text t The word occurrence probability vector is P t =[p(w 1 ), p(w 2 ), ..., p(w m )].

[0119] 3) Calculate the word occurrence probability vector P for multiple time periods t The cross entropy between them, the higher the cross entropy value, the greater the difference in terms, and the cross entropy matrix is ​​obtained. The calculation formula of the cross entropy matrix is:

[0120]

[0121] In formula (12), H(P i , P j ) ij is the word occurrence probability vector P corresponding to time i i The word occurrence probability vector P corresponding to time j j The cross entropy between i (w k ) is the kth word in P i The probability of the corresponding words appearing in j (w k ) is the kth word in P j The probability of the corresponding words appearing in .

[0122] 4) Use a heat map to display the calculated cross entropy matrix, analyze the changes in community terms based on the color depth of the heat map, and obtain the analysis results of the changes in community terms.

[0123] 2. Use NRC sentiment dictionary to perform text sentiment analysis.

[0124] 1) Divide the text data of the target social media platform into certain time intervals to obtain the text library Text of each time period in the continuous period t .

[0125] 2) Import the NRC sentiment dictionary. Sentiment classification is E = {e 1 , e 2 , ..., e μ}, the vocabulary of the NRC dictionary is V 2 ={w 1, w 2 , ...}, each word w i Has a sentiment vector representing its sentiment category score, word w i The sentiment vector is e(w i )=[e 1 (w i ), e 2 (w i ), ..., e μ (w i )].

[0126] 3) Text Library t Traverse, assuming that the text library Text t contains L words, then the text library Text t The sentiment vector e(Text t ) is expressed as the average sum of the sentiment vectors of all words. That is:

[0127]

[0128] In formula (13), e(Text t ) is the text library Text t The sentiment vector of i ) is the i-th word w i Emotion vector.

[0129] 4) According to the text library Text t The sentiment vector e(Text t )Draw the text sentiment change curve and obtain the analysis results of the text sentiment change trend.

[0130] 3. Use the BERTopic topic model to perform text topic analysis.

[0131] 1) Text Library t Perform data cleaning, including removing stop words, special characters, etc., and obtain the text set T = {T 1 , T 2 , ..., T h}.

[0132] 2) Use the pre-trained sentence vector model Sentence-BERT to convert the text set T into a high-dimensional vector representation S. The conversion formula is:

[0133] S=BERT(T) (14)

[0134] In formula (14), T is the text set; S is the high-dimensional vector of the text set T.

[0135] 3) Use HDBSCAN algorithm for clustering. Calculate Ti The distances between (i=1,2…h) are used by the HDBSCAN algorithm to cluster the T i (i=1,2…h) is assigned to cluster C k , the distance calculation formula is:

[0136]

[0137] In formula (15), s i and j are the i-th and j-th vectors in S respectively, cos(s i ,s j ) is the vector s i With vector s j The cosine value between i ,s j ) vector s i With vector s j The direct distance between.

[0138] 4) Use TF-IDF weighted method to extract cluster C k Calculate the number of words w in cluster C. k The word frequency tf(w,C k ), and the document frequency df(w) in all clusters, and then calculate the TF-IDF value corresponding to the word. The calculation formula of TF-IDF value is:

[0139]

[0140] In formula (16), TFIDF(w, C k ) is the TF-IDF value corresponding to word w; tf(w, C k ) is the word w in cluster C k The word frequency in ; df(w) is the document frequency of word w in all clusters; θ is the total number of documents.

[0141] In this embodiment, the l words with the highest TF-IDF value are taken as cluster C k Keywords.

[0142] 5) Count the clusters C in the continuous time period k The number and proportion of the text topics are analyzed to obtain the analysis results of the text topic change trend.

[0143] (IV) Content-Analysis of the micro-analysis dimension may include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text theme analysis for segmented user groups. For example, user groups may be divided according to the chronological order of user participation in topics, and the text sentiment of different segmented user groups may be analyzed; text data may also be divided according to text themes, and the text sentiment in different text themes may be analyzed.

[0144] Furthermore, this embodiment can use line graphs, heat maps, scatter plots, etc. to display the analysis results of multiple analysis directions corresponding to the above dimensions. Taking the ChatGPT event as an example, the analysis results of the structure-macro analysis dimension are as follows: Figure 3 As shown in the figure, the analysis results of the corresponding structure-micro analysis dimension are as follows Figure 4 shown.

[0145] Specifically, S3 may be:

[0146] S31. Based on each analysis result and its corresponding analysis method, generate an explanation text of each analysis method using a large language model, and based on each analysis result and its corresponding analysis method and the explanation text of the analysis method, generate an analysis report for each analysis direction using a large language model;

[0147] S32. Based on the analysis report of each analysis direction, use the large language model to generate an overall analysis report of the hot events.

[0148] Further, S32 may specifically be:

[0149] S321, using a large language model to generate explanatory text for each analysis dimension;

[0150] S322. Generate an overall analysis report of the hot events using a large language model based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension.

[0151] Further, S322 may specifically be:

[0152] Based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension, the analysis report of each dimension is generated using the large language model;

[0153] Based on the analysis report of each dimension, a large language model is used to generate an overall analysis report of hot events.

[0154] Furthermore, based on the analysis report of each dimension, the large language model is used to generate an overall analysis report of the hot events, which can be specifically:

[0155] Generate explanatory text of the analysis framework using a large language model;

[0156] Based on the explanatory text of the analysis framework and the analysis report of each dimension, a large language model is used to generate an overall analysis report on hot events.

[0157] Specifically, this embodiment constructs Figure 5 The prompt engineering structure shown is based on deep traversal search. Through the tree-like prompt engineering structure, each level uses a large language model to strengthen and expand the conceptual knowledge of this level, and then combines the analysis results of this level with the analysis results of the next level to finally generate an overall analysis report of hot events.

[0158] In this embodiment, the specific steps of using the prompt project structure to generate an overall analysis report of hot events are as follows:

[0159] 1) The analysis methods and analysis results corresponding to each analysis direction are expressed as (f 1 , r 1 ), (f 2 , r 2 ), ..., where f i is the analysis method corresponding to the analysis direction, r i The analysis method f is the analysis result corresponding to the analysis direction. i Input into the large language model, the question format is "Please explain the analysis method f i ", and get the analysis method f i Explanatory text, i.e. methodological knowledge i , and (f i , o i , r i ) is input into the large language model, and the question format is "Use analysis method: f i , Method knowledge: o i , analysis results: r i , Analysis report: ", get the analysis report F corresponding to the analysis direction i ,like Figure 6 shown.

[0160] 2) Input the name and definition of the analysis dimension into the large language model to obtain the interpretation text of the analysis dimension, i.e., the dimension knowledge g i ; The analysis reports of multiple analysis directions corresponding to the analysis dimensions are F i , ..., F j , will (g i , F i , ..., F j ) is input into the large language model, and the questioning method is similar to the previous step to obtain the analysis report G of the analysis dimension i .

[0161] 3) Input the name and definition of the analysis framework into the large language model to obtain the interpretation text of the analysis framework, namely, the framework knowledge q; the analysis reports of the multiple analysis dimensions corresponding to the analysis framework are G i , ..., G j , (q, G i , ..., G j ) is input into the large language model. The questioning method is similar to the previous step, and an overall analysis report Q of the hot events is obtained.

[0162] The present invention constructs an analysis framework including multiple analysis dimensions and multiple analysis directions, and uses the analysis framework to analyze hot events to obtain analysis results of multiple analysis directions, and then uses a large language model to generate an overall analysis report of hot events based on the analysis results. In this way, the interrelationships and mutual influences between multiple analysis directions and multiple analysis dimensions are fully considered in one analysis process, and a comprehensive and integrated analysis of hot events is achieved, which is conducive to improving the accuracy and reliability of the analysis results. At the same time, the present invention uses a large language model and combines the analysis framework to integrate the analysis results to form an overall analysis report of hot events, realizing intelligent analysis, which is conducive to improving analysis efficiency.

[0163] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A comprehensive analysis method for hot social media events, characterized in that: The method comprises: S1. Construct an analytical framework for hot events from multiple analytical dimensions. Each analytical dimension includes multiple analytical directions, and each analytical direction corresponds to an analytical method. S2. Analyze hot events on the target social media platform using the analysis framework to obtain analysis results corresponding to each analysis direction; S3. Generate an overall analysis report of the hot event based on the analysis dimension, analysis direction and analysis method corresponding to each analysis result and using a large language model.

2. The method according to claim 1, characterized in that The multiple analysis dimensions include structure-macro analysis dimension, structure-micro analysis dimension, content-macro analysis dimension and content-micro analysis dimension.

3. The method according to claim 1 or 2, characterized in that: The S3 is specifically: S31. Based on each analysis result and its corresponding analysis method, generate an explanation text of each analysis method using a large language model, and based on each analysis result and its corresponding analysis method and the explanation text of the analysis method, generate an analysis report for each analysis direction using a large language model; S32. Based on the analysis report of each analysis direction, a large language model is used to generate an overall analysis report of the hot event.

4. The method according to claim 3, characterized in that The S32 is specifically: S321, using a large language model to generate explanatory text for each analysis dimension; S322. Generate an overall analysis report of the hot event using a large language model based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension.

5. The method according to claim 4, characterized in that The S322 is specifically: Based on the explanatory text of each analysis dimension and the analysis reports of multiple analysis directions corresponding to each dimension, the analysis report of each dimension is generated using the large language model; Based on the analysis report of each dimension, a large language model is used to generate an overall analysis report of the hot event.

6. The method according to claim 5, characterized in that According to the analysis report of each dimension, the overall analysis report of the hot event is generated by using the large language model, specifically: Generate an explanation text of the analysis framework using a large language model; Based on the explanatory text of the analysis framework and the analysis report of each dimension, a large language model is used to generate an overall analysis report of the hot event.

7. The method according to claim 2, characterized in that: The analysis directions of the content-macro analysis dimension include: analysis of changes in community terms, analysis of changes in text keywords, text sentiment analysis, and text theme analysis.

8. The method according to claim 7, characterized in that The corresponding analysis method for analyzing the changes in the community terms is: Determine a word occurrence probability vector for each time period in a continuous time period based on the text data of the target social media platform; The cross entropy of the word occurrence probability vectors in multiple time periods is calculated, and the change of community terms is determined according to the cross entropy.

9. The method according to claim 2, characterized in that: The analysis directions of the structural-macro analysis dimension include: indicator change analysis of social media platforms, long-term relationship analysis between multiple indicators, and short-term status analysis of each indicator.

10. The method according to claim 1, characterized in that Before S2, the method further includes: Calculate the importance index of each social media platform based on the number of users and data volume of each social media platform among multiple social media platforms; A target social media platform is determined according to the importance index.

Images