A method for predicting the development of public opinion events based on theme evolution

Abstract

The present invention provides a method for predicting the development of public opinion events based on topic evolution, which relates to the field of public opinion prediction in natural language processing. The method includes the following specific steps: Step 1, collect and preprocess the data of public opinion events, divide the initial development stage P' of the public opinion event, and extract keywords; Step 2, divide the topic evolution stage of the public opinion event, reconstruct the development stage P” according to the persistence and popularity of keywords in adjacent initial development stages, update the keyword set, and cluster the keywords to form topics; Step 3, extract the attributes of each stage and topic attributes, and construct a topic relationship graph; Step 4, construct a feature extraction model to capture stage features and topic dependence features; Step 5, construct a time series model to predict the topics of public opinion events. Through the method for predicting the development of public opinion events based on topic evolution, the prediction of sudden online public opinion events can be realized, and the focus public opinion and development direction of public opinion events can be grasped.

Description

Technical Field

[0001] The present invention relates to the field of public opinion prediction using natural language processing, and specifically to a method for predicting the development of public opinion events based on topic evolution. Background Art

[0002] With the development of mobile internet technology and emerging social media, netizens are using emerging social platforms like Weibo, Douyin, Kuaishou, and WeChat as platforms to interact and express their opinions, views, and attitudes on social emergencies through video and text. This public opinion is known as online public opinion. Due to the ease and speed of these platforms, netizens often express their opinions online without careful consideration after a social emergency occurs. Through forwarding and commenting, public opinion spreads rapidly across social platforms, forming online public opinion. The generation of public opinion topics is not accidental; rather, it corresponds to real-world events or phenomena. In other words, real-life social life provides the social context for the emergence of hot topics, which are of vital concern to the general public. The inherent openness and sharing nature of the internet allows netizens to freely choose whether to participate in or post on hot topics without having to ensure the authenticity of their words. This makes it easy for extreme or inappropriate speech to deviate from the original focus of online public opinion, leading to adverse consequences. The evolution of online public opinion topics typically consists of six stages: incubation, outbreak, spread, recurrence, remission, and a long tail. The latent period is mainly the generation of the topic, the outbreak period is mainly manifested as the increase in topic attention due to the attention of mainstream media, the spreading period is manifested as the continuous increase in topic attention, the recurrence period is manifested as the attention increases again, the relief period is manifested as the topic attention continues to decrease, and the long tail period is manifested as the attention drops to the lowest value and does not recur for a long time.

[0003] In order to ensure the reduction of the negative impact of certain negative or improper online public opinion topics and strengthen the timeliness and accuracy of online public opinion event topic warnings, it is possible to guide the evolution direction of online public opinion topics. By building a scientific and reasonable online public opinion event topic prediction mechanism, relevant departments can pay attention to and control the development of public opinion topics, and reserve more time for adopting response strategies. This has important practical significance for reducing the adverse effects of online public opinion topics. With the development and application of big data technology, real-time attention can be paid to the topics of online public opinion events. With the help of semantic parsing, keyword analysis and other technologies, the topics of public opinion events can be identified, the evolution of topics can be analyzed, and prior knowledge of the development laws of public opinion events can be obtained. Neural network models can be used for training and prediction, which will help to discover potential online public opinion topics in the future.

[0004] The present invention provides a method for predicting the development of public opinion events based on theme evolution, which solves the defect of the traditional method that only relies on statistical data such as reposts, likes, and comments of public opinion events while ignoring the event themes. First, due to behaviors such as posting and liking fraud on social network platforms, the statistical data has characteristics such as inaccuracy and large differences, making it impossible to accurately know the development status of online public opinion events; second, online public opinion events have a large amount of text data, and existing methods have not fully mined the text information for application; finally, traditional theme prediction methods rely on known theme information while ignoring the evolution of themes during the development of events, making it impossible to know the specific content driving the development of public opinion events and not fully deeply mining the themes leading to the development of public opinion events. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a method for predicting the development of public opinion events based on theme evolution, which solves the above problems.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for predicting the development of public opinion events based on theme evolution, including the following specific steps:

[0007] Step 1, collect data of public opinion events, divide the initial development stage P' of public opinion events with 15 minutes as the time unit, and extract keywords for each initial development stage;

[0008] Step 2, divide the theme evolution stage of public opinion events, and this step includes: judge the similarity by comparing the persistence and heat of the keyword sets of adjacent initial development stages. If it is judged to be similar, merge the adjacent initial development stages to reconstruct a new development stage and update the keyword set; then use the edit distance clustering algorithm to cluster the keywords under the new development stage to form a theme set;

[0009] Step 3, extract stage attributes and theme attributes, and construct a theme relationship graph according to the co-occurrence relationship of themes;

[0010] Step 4, construct a feature extraction model, and extract stage features and theme dependence features;

[0011] Step 5, construct a time series model to predict the theme of public opinion events, and use the stage features and theme dependence features as inputs to predict the theme of the next stage.

[0012] Preferably, in the above Step 1, the public opinion event is a network emergency event, and the initial development stage P' of the public opinion event data is divided with 15 minutes as the time unit, P' = {p'1, p'2, …, p' n}, where n is the number of stages, and this time unit is an empirical value. This method can reasonably divide the evolution stages of the theme according to the characteristics of the theme, analyze the development mechanism of the public opinion event, and can predict the number of themes in the next development stage based on the historical development stages of the public opinion event, so as to predict the development trend of the public opinion event, which helps relevant departments to timely understand the development trend of network public opinion and the voices of the media and netizens, and provides auxiliary decision-making for public opinion guidance and management.

[0013] Preferably, step 2 includes:

[0014] Step 21, use keyword extraction technology to extract the keywords of the text under each initial development stage P', and obtain the keyword set Keywords' = {keywords'1, keywords'2,..., keywords' i ,..., keywords' n} where n is the number of stages and also the number of keyword sets in the initial stage. keywords′ i ∈Keywords' is the keyword set of the i-th initial development stage, kw is the keyword, is the number of keywords in the i-th development stage;

[0015] Step 22, compare the persistence and popularity of the keywords of adjacent initial development stages p' i ∈P' and p' i+1 ∈P'. If the adjacent initial development stages are similar, then merge the two initial stages to reconstruct a new development stage p” = {p″1, p″2,..., p″ m}, where m is the number of stages after reconstructing the development stage, and update the keyword set Keywords” = {keywords″1, keywords″2,..., keywords″ τ ,..., keywords″ m}, keywords″ τ ∈Keywords″, keywords″ τ = {kw″1, kw″2,..., kw″ ρ}. Hereinafter, the development stages mentioned are all the new development stage P”.

[0016] Step 23, use the edit distance clustering algorithm to cluster the keyword sets Keywords” under each development stage to form a theme set Topics = {Topics1, Topics2,..., Topics p ,..., Topics m}, Topics p= {t1, t2, …, t v}, where v is the number of topics in the p-th development stage.

[0017] Preferably, step 3 includes:

[0018] Step 31, extract the stage attributes under each development stage, including the number of texts, the number of topics, the number of publishers, the number of influential people, and the duration;

[0019] Step 32, extract the topic attributes under each development stage, including topic frequency, the number of publishers, the number of influential people, topic sentiment, topic novelty, and topic popularity;

[0020] Step 33, in the same development stage, construct a topic relationship graph according to the co-occurrence relationship of topics. The co-occurrence of topics comes from the co-occurrence relationship between the keywords that make up the topics, and there is no intersection between the keyword sets of the topics formed. Then, if the keywords under different topics belong to the same text sentence, the keyword has a co-occurrence relationship, and the topic has a co-occurrence relationship. Construct a topic relationship graph according to the topic co-occurrence dependence relationship, where the nodes represent topics and the edges represent co-occurrence relationships. The construction of the topic relationship graph is for further analysis.

[0021] Preferably, step 4 includes:

[0022] Step 41, use the normalization method to map the stage features of each development stage attribute;

[0023] Step 42, use the graph neural network model to capture the topic dependence features according to the topic attributes and the topic relationship graph, and output the final topic dependence features according to the average method or the attention mechanism.

[0024] Preferably, step 5 includes:

[0025] Step 51, splice the stage features of step 41 and the topic dependence features of step 42 to form the prior knowledge of the prediction task;

[0026] Step 52, use the prior knowledge spliced in step 51 as the input of the time series model, use the time series model to learn the time series dependence features, and perform the topic prediction of the next development stage. When it is necessary to predict the topic of the next part of the public opinion event, upload the things of the public opinion event that need to be predicted for predicting the next event.

[0027] The present invention provides a method for predicting the development of public opinion events based on topic evolution. It has the following beneficial effects:

[0028] Through the research on the data of five public opinion events, the present invention reasonably divides the theme evolution stages of public opinion events by using the proposed method for dividing theme evolution stages, obtains the themes in each stage, analyzes the characteristics of each stage, the characteristics of the themes, constructs a theme relationship diagram based on the theme co-occurrence relationship in each stage, and conducts theme trend prediction through the construction of a neural network feature extraction model and a sequence model. In terms of method, the present invention can solve the defect of traditional methods that only rely on statistical data such as the number of reposts, likes, and comments of public opinion events while ignoring the themes of public opinion events, make up for the shortcomings of inaccurate and large differences in statistical data, and can understand the information and specific content of public opinion events, as well as the evolution of themes during the development process of public opinion events by mining a large amount of text data of online public opinion events. In practical applications, it can reasonably divide the evolution stages of themes according to the characteristics of themes, analyze the development mechanism of public opinion events, and predict the number of themes in the next development stage based on the historical development stages of public opinion events, so as to predict the development trend of public opinion events, which helps relevant departments to timely understand the development trend of online public opinion and the voices of the media and netizens, and provides auxiliary decision-making for public opinion guidance and processing management. Description of the Drawings

[0029] Figure 1 is a flowchart of the present invention;

[0030] Figure 2 is a flowchart for dividing the theme evolution stages of the present invention;

[0031] Figure 3 is a diagram of the stage feature and theme dependence feature extraction model of the present invention;

[0032] Figure 4 is a diagram of the theme prediction model of the present invention;

[0033] Figure 5 is a theme relationship diagram of the present invention;

[0034] Figure 6 is a predicted trend diagram of the public opinion event "Ding Zhen became popular" of the present invention;

[0035] Figure 7 is a predicted trend diagram of the public opinion event "Eight-child incident in Feng County" of the present invention;

[0036] Figure 8 is a predicted trend diagram of the public opinion event "Hongxing Erke donated materials to support Henan" of the present invention;

[0037] Figure 9 is a predicted trend diagram of the public opinion event "Liu Genghong became popular" of the present invention;

[0038] Figure 10 is a predicted trend diagram of the public opinion event "Cyndi Wang made a comeback" of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0040] Example 1:

[0041] like Figure 1 As shown, the embodiment of the present invention provides a method for predicting the development of public opinion events based on topic evolution, including the following specific steps:

[0042] A method for predicting the development of public opinion events based on topic evolution includes the following specific steps:

[0043] Step 1: Collect data on public opinion events, divide the initial development stage P' of public opinion events into 15-minute time units, and extract keywords for each initial development stage;

[0044] Step 2: Divide the evolution stages of public opinion events. This step includes: comparing the persistence and popularity of keyword sets in adjacent initial development stages to determine their similarity. If similarity is determined, the adjacent initial development stages are merged to reconstruct a new development stage and the keyword set is updated. The keywords in the new development stage are then clustered using an edit distance clustering algorithm to form a topic set.

[0045] Step 3: Extract stage attributes and topic attributes, and construct a topic relationship graph based on the co-occurrence relationship of the topics;

[0046] Step 4: Build a feature extraction model to extract stage features and topic-dependent features;

[0047] Step 5: Build a time series model to predict the topic of public opinion events. Use stage characteristics and topic dependency characteristics as input to predict the topic of the next stage. This can solve the defect of traditional methods that only rely on statistical data such as reposts, likes, and comments of public opinion events while ignoring the topic of public opinion events, and make up for the shortcomings of inaccurate statistical data and large differences. In addition, by mining a large amount of text data of online public opinion events, it can understand the information and specific content of public opinion events, as well as the evolution of the topic of public opinion events during their development.

[0048] The public opinion event in step 1 is an online sudden public opinion event. The initial development stage of the public opinion event data is divided into 15-minute time units P'={p'1,p'2,…,p' n}, n is the number of stages, and the time unit is experience value;

[0049] Step 2 includes:

[0050] Step 21, using keyword extraction technology to extract the keywords of the text under each initial development stage P' = {P'1, P'2, …, P' n}, obtaining the keyword set keywords' = {keywords'1, keywords'2, …, keywords' i , …, keywords' n}, where n is the number of stages and also the number of keyword sets in the initial stage. keywords' i ∈ keywords' is the keyword set of the i-th initial development stage, kw is the keyword, is the number of keywords in the i-th development stage;

[0051] Step 22, comparing the persistence and popularity of the keywords of adjacent initial development stages p' i ∈ P' and p' i+1 ∈ P'. If the adjacent initial development stages are similar, then merge the two initial stages to reconstruct a new development stage P” = {p″1, p″2, …, p″ m}, where m is the number of stages after reconstructing the development stage, and update the keyword set keywords” = {keywords″1, keywords″2, …, keywords″ τ , …, keywords″ m}, keywords″ τ ∈ Keywords”, keywords″ τ = {kw″1, kw″2, …, kw″ ρ}. All the following development stages refer to the new development stage P”;

[0052] Step 23, using the edit distance clustering algorithm to cluster the keyword sets Keywords” under each development stage to form a theme set Topics = {Topics1, Topics2, …, Topics p , …, Topics m}, Topics p = {t1, t2, …, t v}, where v is the number of themes in the p-th development stage;

[0053] Step 3 includes:

[0054] Step 31: Extract the stage attributes for each development stage, including the number of texts, the number of topics, the number of publishers, the number of influential people, and the duration;

[0055] Step 32: Extract the topic attributes for each development stage, including the topic frequency, the number of publishers, the number of influential people, the topic sentiment, the topic novelty, and the topic popularity;

[0056] Step 33: In the same development stage, construct a topic relationship graph based on the co-occurrence relationship of topics. The co-occurrence of topics comes from the co-occurrence relationship between the keywords that make up the topics, and there is no intersection between the keyword sets of the topics formed. If the keywords under different topics belong to the same text sentence, the keyword has a co-occurrence relationship, then the topic has a co-occurrence relationship. Construct a topic relationship graph according to the topic co-occurrence dependency relationship, where the nodes represent topics and the edges represent co-occurrence relationships;

[0057] Step 4 includes:

[0058] Step 41: Use the normalization method to map the stage features of each development stage attribute;

[0059] Step 42: Use a graph neural network model to capture topic dependency features based on topic attributes and the topic relationship graph, and output the final topic dependency features according to the average method or the attention mechanism;

[0060] Step 5 includes:

[0061] Step 51: Concatenate the stage features of Step 41 and the topic dependency features of Step 42 to form the prior knowledge for the prediction task;

[0062] Step 52: Use the prior knowledge concatenated in Step 51 as the input of the time series model, and use the time series model to learn the time series dependency features for the topic prediction of the next development stage.

[0063] Embodiment 2:

[0064] As Figure 1 shown, the embodiment of the present invention provides a method for predicting the development of public opinion events based on topic evolution, including the following specific steps:

[0065] To achieve the above object, the present invention is realized through the following technical solutions:

[0066] Step 1: Collect and preprocess public opinion event data;

[0067] Step 2: Divide the topic evolution stages of public opinion events;

[0068] Step 3: Extract stage attributes, topic attributes, and construct a topic relationship graph;

[0069] Step 4, design a feature extraction and topic-dependent feature capture model for the design phase;

[0070] Step 5, construct a time series model to predict the topic of the public opinion event.

[0071] Preferably, the above-mentioned step 1 includes:

[0072] Step 11, design a crawler program to capture data of public opinion events from the Weibo platform, including text, text information, publisher, and publisher information;

[0073] Step 12, taking 15 minutes as a time unit, divide the data of the public opinion event to form the initial development stage of the event P' = {P'1, P'2, …, P' n}, where n is the number of stages; then use keyword technology to extract keywords in the text of each initial development stage to form the keyword set of the initial development stage Keywords' = {keywords'1, keywords'2, …, keywords' i , …, keywords' n}, and keywords' i ∈ Keywords' is the keyword set in the i-th initial stage.

[0074] Preferably, the above-mentioned step 2 includes:

[0075] Step 21, count the frequency of keywords in each development stage and sort them, and reconstruct the development stage P” according to two conditions set by the persistence and popularity of the keywords. As Figure 2 shown, where condition 1, if the intersection of the first 5 keywords in two adjacent initial stages is greater than 3, it is considered that the two initial stages belong to the same development stage and are merged; condition 2, on the basis of the newly generated development stage in condition 1, if the frequency of the keyword in the last initial stage of the newly generated development stage is greater than the threshold 5, and the frequency of the keyword in the first initial stage of the next newly generated development stage is still greater than the threshold 5, it is considered that the two newly generated development stages belong to the same development stage and are merged and updated to form the final development stage P” = {P″1, P″2, …, P″ p , …, P” m}, where m is the newly generated development stage after update. All the following stages are newly generated development stages, and P″ p is the p-th development stage;

[0076] Step 22: According to the similarity of keywords in each development stage, use the edit distance algorithm to cluster the keywords in each development stage. Each set of keywords forms a theme, and a set of themes Topics = {Topics1, Topics2, …, Topics} is formed in the development stage. p , …, Topics m}, Topics p = {t1, t2, …, t v} where v is the number of themes in the p-th development stage.

[0077] Preferably, step 3 includes:

[0078] Step 31: Take the statistics in each development stage as stage attributes, including the number of texts, the number of themes, the number of publishers, the number of influential people, and the duration, and formalize them as K p = {k text_num , k topic_num , k poster , k inf_poster , k duration} where p is one of the development stages. Among them, the number of influential people is the number of people with Weibo certifications among the publishers, and the duration is the total time of the initial stage included in the new development stage;

[0079] Step 32: As shown in Figure 3 , take the attributes of the keywords forming the theme categories in each development stage as theme attributes, including theme frequency, number of publishers, number of influential people, theme sentiment, theme novelty, and theme popularity, and formalize them as x q = {x frequency , x poster , x inf_poster , x Sentiment , x Novelty , x Popularity}, X p = {x1, x2, …, x q , …, x v}, x q ∈ X p where q is the attribute of the q-th theme in the p-th development stage. Among them, the theme frequency is the sum of the frequencies of the keywords forming the theme, the number of publishers is the sum of the number of publishers of the keywords forming the theme, the number of influential people is the number of people with Weibo certifications among the publishers, the theme sentiment is the sentiment of the keywords forming the theme, the sentiment of the keywords is the sentiment of the text sentence where the keywords are located, the theme novelty is the reciprocal of the difference between the stage when the theme first appears and the current stage, as shown in formula (1), P″ pFor the current development stage, P″1 is the development stage when the theme first appears. The popularity of the theme is determined by the frequency of the theme and the adjustable frequency, as shown in Formulas (2) and (3), where TF is the theme frequency, ATF is the adjustable theme frequency, and ρ is the attenuation coefficient;

[0080]

[0081] ATF p =TF p +ρ×ATF p-1 (2)

[0082]

[0083] Step 33, as shown, the construction of the theme relationship graph depends on the co-occurrence relationship of themes. If the keywords under two different themes belong to the same text sentence, then the two themes have a co-occurrence relationship, and the theme relationship graph G = {G1, G2,..., G p ,…,G m}, G p =(V p , E p ), G is the set of theme relationship graphs for all development stages, G p is the theme relationship graph for the pth development stage, V p is the set of theme nodes, |V p | = v, E p is the set of theme relationship edges, and the adjacency matrix A p ∈{0, 1} v×v . If there is a co-occurrence relationship between theme i and theme j, then A ij = 1; otherwise, A ij = 0.

[0084] Preferably, the said Step 4 includes:

[0085] Step 41, for all development stages, use the softmax method to extract the features of each development stage attribute, formalized as The calculation is as shown in Formula (4);

[0086]

[0087] Step 42, under each development stage, use a feedforward neural network to extract the theme attribute features, formalized as F = {F1, F2,..., F P ,…,F m}, F p ={f1, f2,…,f v}∈F and the calculation is as shown in Formula (5), with the input being X p ={x1, x2,…,xv}, W p is the weight, b p is the bias;

[0088] F p = FNN(W p X p + b p )(5)

[0089] Step 43, Use the graph attention network to capture the co-occurrence dependencies between topics at each development stage as shown in formulas (6)-(8), formalized as Then use the average method or the attention mechanism method to quantify the importance of the topic as the topic dependency feature, formalized as where can be calculated in two ways. Way 1, the average method is shown in formula (9). Way 2, the attention mechanism method is shown in formulas (10)-(12);

[0090]

[0091] ɑ i = softmax(u i )(11)

[0092]

[0093] Preferably, the said step 5 includes:

[0094] Step 51, Concatenate the stage feature in step 41 and the topic dependency feature in step 43 as the final development stage feature, and then divide it by the input step length as the input feature of the time series model, formalized as N is the step length, where is calculated as shown in formula (13);

[0095]

[0096] Step 52, As Figure 4 shown, use the divided sequence features as the input, and use the long short-term neural network to predict the number of topics in the next development stage. The calculation process is shown in formulas (14)-(20), where is the input of the p-th development stage, is the output of the (p-1)-th development stage, which is the previous development stage of the p-th development stage. s p , t p , o p are the gate vectors, C p is the cell state vector, W s , W t , Wo ,W C ,W y is the weight vector, b s ,b t ,b o ,b C ,b y is the bias vector, h p+1 is the prediction target, that is, the number of topics in the next stage p + 1 to be predicted.

[0097]

[0098] Step 52, use the proposed method to predict the topics of five public opinion events. As Figure 6 shown, it is the prediction trend chart of the public opinion event "The Event of Ding Zhen Becoming Popular" of the present invention; Figure 7 is the prediction trend chart of the public opinion event "The Event of the Eight Children in Feng County" of the present invention; Figure 8 is the prediction trend chart of the public opinion event "The Event of Hongxing Erke Donating Supplies to Support Henan" of the present invention; Figure 9 is the prediction trend chart of the public opinion event "The Event of Liu Genghong Becoming Popular" of the present invention; Figure 10 is the prediction trend chart of the public opinion event "The Event of Cyndi Wang Making a Comeback" of the present invention. In the prediction of the topic trends of the five public opinion events, the abscissa is the development stage, the ordinate is the number of topics in each stage, and the number of topics in each development stage changes with the evolution and development of the public opinion events. The solid line is the actual trend of the public opinion event topics, and the dotted line is the prediction trend. In the prediction of the topic trends of the five public opinion events, the prediction results are consistent with the actual results.

[0099] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the development of public opinion events based on theme evolution, characterized in that: It includes the following specific steps: Step 1: Collect data on public opinion events and divide the initial development stage of public opinion events into 15-minute time units and extract the keywords for each initial development stage; Step 2: Divide the theme evolution stages of public opinion events, and reconstruct the development stages according to the persistence and popularity of keywords in adjacent initial development stages , and update the keyword set, and cluster the keywords in the new development stage to form a theme set; The said step 2 includes: Step 21, use keyword extraction technology to extract the keywords of each initial development stage of the following text, and obtain the keyword set of the initial development stage , is the number of stages, and also the number of keyword sets in the initial stage. is the keyword set of the th initial development stage, is the keyword, and of the adjacent initial development stages, if the adjacent initial development stages are similar, then merge the two initial stages to reconstruct a new development stage , is the number of stages after reconstructing the development stage, and update the keyword set . All the following development stages mentioned are new development stages ; Step 23, use the edit distance clustering algorithm to cluster the keyword sets of each development stage to form a theme set is the number of themes of the th development stage; Step 3, extract the stage attributes and theme attributes, and construct a theme relationship graph according to the co-occurrence relationship of themes; Step 4, construct a feature extraction model to extract stage features and theme dependence features; Step 5, construct a time series model to conduct theme prediction of public opinion events. Using the stage features and theme dependence features as inputs, predict the theme of the next stage.

2. The method for predicting the development of public opinion events based on theme evolution according to claim 1, characterized in that: In step 1, the public opinion event is a network emergency, and the initial development stage of the public opinion event data is divided with 15 minutes as the time unit. , is the number of stages, and this time unit is an empirical value.

3. The method for predicting the development of public opinion events based on theme evolution according to claim 1, wherein: The said step 3 includes: Step 31, extract the stage attributes under each development stage, including the number of texts, the number of themes, the number of publishers, the number of influential people, and the duration; Step 32, extract the theme attributes under each development stage, including theme frequency, the number of publishers, the number of influential people, theme sentiment, theme novelty, and theme popularity; Step 33, under the same development stage, construct a theme relationship graph according to the co-occurrence relationship of themes. The co-occurrence of themes comes from the co-occurrence relationship between the keywords that make up the themes, and there is no intersection between the keyword sets of the themes formed. Then, if the keywords under different themes belong to the same text sentence and the keyword has a co-occurrence relationship, then the theme has a co-occurrence relationship. Construct a theme relationship graph according to the theme co-occurrence dependence relationship, where the nodes represent themes and the edges represent co-occurrence relationships.

4. The method for predicting the development of public opinion events based on theme evolution according to claim 1, characterized in that: The said step 4 includes: Step 41, use the normalization method to map the stage features of each development stage attribute; Step 42, use the graph neural network model to capture theme dependence features according to the theme attributes and the theme relationship graph, and output the final theme dependence features according to the average value method or the attention mechanism.

5. The method for predicting the development of public opinion events based on theme evolution according to claim 1, wherein: The said step 5 includes: Step 51, splice the stage features of step 41 and the theme dependence features of step 42 to form the prior knowledge of the prediction task; Step 52, use the spliced prior knowledge in step 51 as the input of the time series model, and use the time series model to learn the time series dependence features to conduct theme prediction of the next development stage.

Images