Dynamic adaptive event development prediction method and device, equipment and storage medium

Abstract

The embodiment of the application relates to the technical field of event early warning, and discloses a dynamic self-adaptive event development prediction method, device, equipment and storage medium, the method comprises the following steps: multi-level keyword matching is carried out on original data, a special subject is determined, a sentiment tendency label and a negative viewpoint quantity of the original data are determined through a large language model, whether an event occurring in the original data is a hot event is determined through the special subject, the sentiment tendency label and the negative viewpoint quantity, if yes, the event is stored and tracked, network information of the hot event is crawled when tracking, a deterioration score is dynamically calculated according to network interaction information, a development dynamic factor is determined according to a network transmission platform and a network transmission address, a development stage of the hot event is determined according to the deterioration score and the development dynamic factor, and early warning is carried out, through the above mode, the development stage of the hot event is dynamically self-adaptively judged through the deterioration score and the development dynamic factor, and real-time performance and accuracy are better.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of event early warning, in particular to a dynamic adaptive event development prediction method, device, equipment and storage medium. BACKGROUND

[0002] The core challenge of event early warning is that the current network events show explosive growth, especially in social media, news platforms and vertical fields (such as fire, medical, government WeChat groups), and traditional early warning methods face the following problems:

[0003] 1. Single early warning rule: Most event early warning systems rely on a single static threshold (such as keyword frequency, fixed sentiment score) to trigger early warning. Even if multiple factors are considered, the weights of the factors are fixed and cannot adapt to the dynamic development of events, and the real-time performance is poor.

[0004] 2. High false negative and false positive rates: Traditional event early warning systems are difficult to predict the evolution trend of events and lack the ability to identify "hot events", and the accuracy is low. SUMMARY

[0005] In view of the above problems, the embodiment of the present application provides a dynamic adaptive event development prediction method, device, equipment and storage medium to solve the problems in the prior art. According to one aspect of the embodiment of the present application, a dynamic adaptive event development prediction method is provided, the method comprising:

[0006] Multi-level keyword matching is performed on the preprocessed original data, and after successful matching, the subject of the original data is determined; the original data is analyzed based on a preset first large language model to obtain the sentiment tendency label corresponding to the original data output by the first large language model;

[0007] The original data is extracted based on a preset second large language model to obtain the number of negative views corresponding to the original data output by the second large language model; and whether the event occurring in the original data is a hot event is determined according to the subject, sentiment tendency label and negative view number;

[0008] If the event occurring in the original data is a hot event, the original data is saved as event data in an event database for event tracking; the network information of the hot event in the event database is crawled, and the network information includes network interaction information, network transmission platform and network transmission address;

[0009] Determine a deterioration score of the hot event according to the network interaction information, determine a development dynamic factor of the hot event according to the network transmission platform and the network transmission address, determine a development stage of the hot event and perform event early warning according to the deterioration score and the development dynamic factor.

[0010] In an optional manner, the preprocessed original data is subjected to multi-level keyword matching, and after successful matching, the subject of the original data is determined, comprising:

[0011] Load a general keyword in a preset dynamic general keyword library, and match the preprocessed original data with the general keyword;

[0012] If the preprocessed original data matches the general keyword successfully, load a subject keyword in a preset dynamic subject keyword library, and match the preprocessed original data with the subject keyword;

[0013] If the preprocessed original data matches the subject keyword successfully, the subject of the original data is determined.

[0014] In an optional manner, the determination of whether the event occurring in the original data is a hot event according to the subject, the sentiment tendency label and the number of negative opinions comprises:

[0015] According to a preset subject risk dimension relationship, the subject and a corresponding risk category are determined, and a risk point to which the subject belongs under the risk category is determined, and a risk value of the subject is determined according to a risk level corresponding to the risk point;

[0016] According to a preset sentiment score relationship, a sentiment score corresponding to the sentiment tendency label is determined;

[0017] According to a preset opinion score relationship, an opinion score corresponding to the number of negative opinions is determined;

[0018] According to the risk value, the sentiment score and the opinion score, a sensitive value corresponding to the original data is determined, and whether the event occurring in the original data is a hot event is determined according to the sensitive value.

[0019] In an optional manner, the network interaction information comprises a browsing volume, a like volume, a comment volume and a forwarding volume, and the determination of the deterioration score of the hot event according to the network interaction information and the determination of the development dynamic factor of the hot event according to the network transmission platform and the network transmission address comprise:

[0020] statistic the first browse amount, the first like amount, the first comment amount and the first forwarding amount of the current moment, statistic the second browse amount, the second like amount, the second comment amount and the second forwarding amount of the last moment of the current moment, determine the browse amount difference value, the like amount difference value, the comment amount difference value and the forwarding amount difference value according to the first browse amount, the first like amount, the first comment amount, the first forwarding amount, the second browse amount, the second like amount, the second comment amount and the second forwarding amount, determine the deterioration score based on the predetermined weight and the browse amount difference value, the like amount difference value, the comment amount difference value and the forwarding amount difference value;

[0021] statistic the network communication platform quantity and the network communication address quantity of the hotspot event added in the current moment, determine the development dynamic factor based on the network communication platform quantity and the network communication address quantity added.

[0022] In an optional manner, the development dynamic factor is greater than one, and the determining the development stage of the hotspot event and the event early warning according to the deterioration score and the development dynamic factor comprises:

[0023] if the deterioration score is less than a preset first threshold, it is determined that the hotspot event is in a latent stage;

[0024] if the deterioration score is greater than or equal to the first threshold and less than the development dynamic factor, it is determined that the hotspot event is in a fermentation stage;

[0025] if the deterioration score is greater than or equal to the development dynamic factor, it is determined that the hotspot event is in an outbreak stage;

[0026] if the hotspot event has entered the outbreak stage and the deterioration score is less than or equal to half of the development dynamic factor, it is determined that the hotspot event is in a dissipation stage.

[0027] In an optional manner, the network interaction information further comprises interaction behavior, interaction time information and interaction user information, and before the determining the deterioration score of the hotspot event according to the network interaction information, further comprising:

[0028] input the network interaction information into a predetermined double-module collaborative network, identify the false interaction information in the network interaction information based on the double-module collaborative network, and eliminate the false interaction information, the double-module collaborative network comprises a generator and a discriminator, the generator comprises a long short-term memory network and a deconvolution network, and the discriminator comprises a convolutional neural network and an attention mechanism network;

[0029] the determining the deterioration score of the hotspot event according to the network interaction information comprises:

[0030] determine a deterioration score of the hot event according to the network interaction information after the false interaction information is removed.

[0031] In an optional manner, if the event occurred in the original data is a hot event, the original data is saved as event data into an event database for event tracking, and the method further comprises:

[0032] the title and the body of the original data in the event database are respectively vectorized, and it is determined whether there is a same hot event according to the vectorized title and body, and if there is a same hot event, the same hot event is merged into a same hot event.

[0033] According to another aspect of the embodiment of the present application, a dynamic adaptive event development prediction device is provided, and the device comprises:

[0034] The matching module is configured to perform multi-level keyword matching on the preprocessed original data, and determine a special subject to which the original data belongs after successful matching.

[0035] The analysis module is configured to perform sentiment tendency analysis on the original data based on a preset first large language model, and obtain a sentiment tendency label corresponding to the original data output by the first large language model.

[0036] The refining module is configured to refine negative views of the original data based on a preset second large language model, and obtain a number of negative views corresponding to the original data output by the second large language model.

[0037] The first determination module is configured to determine whether an event occurred in the original data is a hot event according to the special subject, the sentiment tendency label, and the number of negative views.

[0038] The saving module is configured to save the original data as event data into an event database for event tracking if the event occurred in the original data is a hot event.

[0039] The crawling module is configured to crawl network information of the hot event in the event database, wherein the network information comprises network interaction information, a network transmission platform, and a network transmission address.

[0040] The second determination module is configured to determine a deterioration score of the hot event according to the network interaction information, and determine a development dynamic factor of the hot event according to the network transmission platform and the network transmission address.

[0041] The third determination module is configured to determine a development stage of the hot event and perform event early warning according to the deterioration score and the development dynamic factor.

[0042] According to a further aspect of the embodiments of the present application, a computer device is provided, comprising a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface performing communication with each other through the communication bus; the memory is used to store at least one executable instruction, the executable instruction causes the processor to execute the method as described above.

[0043] According to a further aspect of the embodiments of the present application, a computer readable storage medium is provided, the storage medium stores at least one executable instruction, the executable instruction, when running on a computer device, causes the computer device to execute the method as described above.

[0044] The event development prediction method dynamically and adaptively judges the development stage of the hot event through the deterioration score and the development dynamic factor, and is better in real-time performance and accuracy.

[0045] The above description is only a summary of the technical solutions of the embodiments of the present application, in order to more clearly understand the technical means of the embodiments of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the embodiments of the present application more obvious and easy to understand, the specific embodiments of the present application are described below. BRIEF DESCRIPTION OF DRAWINGS

[0046] The accompanying drawings are only used to show the embodiments and are not considered as limitations of the present application. Moreover, the same reference signs are used to represent the same parts throughout the drawings. In the drawings:

[0047] Figure 1 The flowchart of the dynamic adaptive event development prediction method provided by the embodiments of the present application is shown;

[0048] Figure 2 The structure diagram of the dynamic adaptive event development prediction device provided by the embodiments of the present application is shown;

[0049] Figure 3 The structure diagram of the computer device provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0050] Exemplary embodiments of the present application will be described in detail below with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it is to be understood that the present application can be embodied in various forms without being limited by the embodiments set forth herein.

[0051] Figure 1 A flowchart of a dynamic adaptive event development prediction method provided by an embodiment of the present application is shown, as shown in FIG. 1, the method comprises the following steps: Figure 1

[0052] Step 10, multi-level keyword matching is performed on the preprocessed original data, and after successful matching, a subject of the original data is determined.

[0053] In this embodiment, original data is collected through multiple network channels such as social media (e.g., Weibo, WeChat, forums, etc.), news websites, blogs, and video platforms, and can further be purchased from various platforms as original data. The original data is an article, which includes title, body, publishing time, publishing user, publishing region, and publishing platform, etc. After the original data is collected, it is preprocessed, including de-duplication, denoising, and word segmentation, etc., and finally the word segmentation of the original data is obtained.

[0054] Further, multi-level keyword matching is performed on the preprocessed original data, and after successful matching, a subject of the original data is determined, comprising:

[0055] The general keywords in the preset dynamic general keyword library are loaded, and the preprocessed original data is matched with the general keywords; if the preprocessed original data is successfully matched with the general keywords, the subject keywords in the preset dynamic subject keyword library are loaded, and the preprocessed original data is matched with the subject keywords; if the preprocessed original data is successfully matched with the subject keywords, a subject of the original data is determined.

[0056] ​In this embodiment, the general keyword library and the special topic keyword library are pre-set and can be dynamically added or deleted through entity linking technology association network, for example, adding the latest hot words or keywords with high frequency in a short time, deleting keywords that have not been matched for a long time, etc. Among them, the general keywords in the general keyword library are the corresponding words of events and actions related to all industries. For example, in the medical industry, general keywords include "registration, medical insurance", etc. The special topic keywords in the special topic keyword library are individual events or related to individual subjects. For example, in the medical industry, special topic keywords can be "**hospital, **patient", etc. For example, in the mobile phone industry, special topic keywords can be "**mobile phone, screen failure", etc.

[0057] In this embodiment, the segmentation of the original data is first roughly matched with the general keywords, and then accurately matched with the special topic keywords through the special topic keywords. Through multi-dimensional matching, the matching granularity can be refined, the matching success rate can be improved, and the ability to detect hot events can be improved. After successful matching, the special topic subject to which the original data belongs can be determined. In the matching process, the segmentation of the original data is expanded by synonym, and the specific method is to train the word vector through the Word2Vec model, calculate the cosine similarity between two words, and the cosine similarity greater than 0.7 is the synonym. In the matching, if it is determined that the segmentation of the original data is the synonym of the general keyword, the matching with the general keyword is successful, and if it is determined that the segmentation of the original data is the synonym of the special topic keyword, the matching with the special topic keyword is successful.

[0058] In other embodiments, the general keyword library and the special topic keyword library corresponding to different fields or industries can be set first, and the pre-processed original data is first confirmed to be in a field or industry, and then matched with the general keywords in the general keyword library of the corresponding field or industry. After successful matching of the general keywords, the special topic keywords in the corresponding special topic keyword library are matched.

[0059] Step 20, based on the first large language model, the sentiment of the original data is analyzed to obtain the sentiment label corresponding to the original data output by the first large language model.

[0060] In this embodiment, the pre-set first large language model can be Bidirectional Encoder Representations from Transformers (BERT) or Generative Pre-trained Transformer (GPT). Preferably, the first large language model of this embodiment adopts BERT.

[0061] BERT is a pre-trained natural language model. In the pre-training stage, BERT includes a mask language model and a next sentence prediction model. In the mask language model, BERT randomly masks some words in a sentence and predicts the masked words, so that BERT can understand the relationship between words through the context; in the next sentence prediction model, BERT inputs two sentences and predicts whether the two sentences are continuous, so that BERT can learn the relevance and reasoning ability between sentences. Through the mask language model and the next sentence prediction model, BERT can learn rich semantic representations at the sentence level and the word level, so as to classify the original data.

[0062] Among them, the sentiment tendency label includes three kinds, positive, neutral and negative. In the BERT loading and configuration stage, a fully connected layer is added at the top of the BERT model, and the output dimension is three, corresponding to the three sentiment labels of positive, neutral and negative. When fine-tuning the BERT model, the text is converted into a tensor format (for example, input_ids, attention_mask, etc.), and the maximum length of text segmentation or truncation is 256-512, so as to alleviate the limitation of processing long sequences. When the BERT model outputs, the logits are output, and the probability of each label is calculated through the Softmax activation function, and finally the sentiment tendency label with the maximum probability is taken as the predicted label. The first large language model is used to analyze the sentiment tendency of the original data in this embodiment. Compared with the form of neural network in the prior art, the generalization ability of this embodiment is stronger, which helps to improve the ability to detect hot events.

[0063] Further, BiLSTM or BiGRU can be fused on the basis of the BERT model to realize further improvement of the prediction label efficiency. Specifically, BiLSTM or BiGRU layers are added at the top of the BERT model, the [CLS] vector or the last layer hidden state (768 dimensions) of the pre-trained BERT is used as the context representation of the text, the sequence vector (such as [batch_size, seq_len, 768]) output by BERT is input into the bidirectional LSTM / GRU layer to capture long-distance context dependence, the hidden state at the last moment of BiLSTM / BiGRU (spliced in both directions for [batch_size, 1024]) or global average pooling is taken, and finally the sentiment tendency label probability distribution is output through the fully connected layer and the Softmax activation function.

[0064] Step 30, based on the second large language model, the negative views of the original data are extracted, and the number of negative views corresponding to the original data output by the second large language model is obtained.

[0065] In this embodiment, the preset second large language model can be a DeepSeek-R1 model or a generative pre-training transformer GPT. Preferably, the second large language model of this embodiment adopts the DeepSeek-R1 model. The DeepSeek-R1 model first performs vectorization and coding processing on the preprocessed text, then inputs it into the negative semantic detection layer to detect negative semantics, and then processes it through the opinion structured distillation layer, and finally outputs the text target entity or sentiment polarity field to obtain negative opinions.

[0066] In the negative semantic detection layer, the BERT and BiLSTM parallel architecture can be used to capture global semantics and local timing features, respectively. In the negative semantic detection layer, the gradient reversal layer (GRL) can also be used to eliminate field bias and improve cross-scene generalization capability. In the opinion structured distillation layer, a double-layer architecture of a teacher model (large model) and a student model (lightweight model) is established in advance, the teacher model generates structured knowledge, and the student model receives distillation signals. The text with detected negative semantics is input into the teacher model, the teacher model processes the input text, and outputs a triple label (such as entity-opinion-sentiment probability distribution) with associated probability. For example, the label is {"property": "battery life", "opinion word": "bad", "sentiment probability": 0.92}. Further, the graph attention network can be used to capture the dependency relationship between opinion elements (such as strong binding of "photograph - amazing"), and the correlation matrix of the teacher model is transferred to the student model. The student model dynamically adjusts the knowledge distillation temperature parameter (T = 0.7~1.2) to control the discrete degree of structured elements, and further, the student model learns the JSON / XML structured format specification in the output layer to ensure that the triple can be directly parsed. Finally, low-confidence triple labels (for example, triple labels with sentiment probability <0.8 are filtered), the number of remaining triple labels is counted, and the number of negative opinions is obtained.

[0067] Step 40, determining whether the event occurred in the original data is a hot event according to the topic subject, sentiment tendency label and number of negative opinions.

[0068] In which, the topic subject involves different risk levels, and the low risk level is not easy to trigger a hot event; the articles with general positive and neutral sentiment tendency labels are not easy to trigger a hot event; the articles with fewer negative opinions are not easy to trigger a hot event. In this embodiment, the topic subject, sentiment tendency label and number of negative opinions are considered to determine whether the event occurred in the original data is a hot event, which is more accurate than the single sentiment label or negative opinion form in the prior art, and helps to improve the ability to detect hot events.

[0069] Further, the method further comprises determining whether the event occurred in the original data is a hot event according to the topic subject, the sentiment tendency label, and the number of negative opinions.

[0070] According to a preset topic risk dimension relationship, the topic subject and a corresponding risk category are determined, and a risk point to which the topic subject belongs under the risk category is determined, and a risk value of the topic subject is determined according to a risk level corresponding to the risk point.

[0071] According to a preset sentiment score relationship, a sentiment score corresponding to the sentiment tendency label is determined.

[0072] According to a preset opinion score relationship, an opinion score corresponding to the number of negative opinions is determined.

[0073] According to the risk value, the sentiment score, and the opinion score, a sensitive value corresponding to the original data is determined, and whether the event occurred in the original data is a hot event is determined according to the sensitive value.

[0074] In the topic risk dimension relationship, a risk category, a risk point, a risk level, and a risk value are included, one risk category includes multiple risk points, the risk level includes low risk, medium risk, high risk, and major risk, and the higher the risk level is, the higher the risk value is. Taking the risk categories “medical service quality” and “medical charge” as examples, the topic risk dimension relationship is shown in Table 1 as follows:

[0075] Table 1

[0076] For the topic subject “** hospital service”, the corresponding risk category is “medical service quality”, the corresponding risk point is “service attitude” under “medical service quality”, the corresponding risk level is “medium risk”, and the corresponding risk value is 3.

[0077] In the sentiment score relationship, the sentiment scores corresponding to the positive, neutral, and negative sentiment tendency labels are -1, 0, and 1 respectively, as shown in Table 2 as follows:

[0078] Table 2

[0079] For the opinion score relationship, the opinion scores corresponding to the number of negative opinions of 0, 1, 2, and more than 2 are -0.5, 0.5, 1, and 2 respectively, as shown in Table 3 as follows:

[0080] Table 3

[0081] According to the above score, a sensitivity value is calculated, and the sensitivity value Sensitivity = ET + NV + CV, wherein ET is the emotional score, NV is the opinion score, and CV is the risk value. When the sensitivity value is in [-1.5, 5), it belongs to a general event, and when the sensitivity value is greater than or equal to 5, it belongs to a hot event.

[0082] Step 50, if the event occurring in the original data is a hot event, the original data is saved as event data in the event database for event tracking.

[0083] In this embodiment, an event database is established, and the original data with a sensitivity value greater than or equal to 5 is saved as event data in the event database for tracking all hot events in the event database. In this embodiment, after the original data is judged to be a hot event, it is judged to be likely to develop into a hot event, that is, an event that is "not hot yet", which triggers the event monitoring and tracking system to track it.

[0084] Further, if the event occurring in the original data is a hot event, the original data is saved as event data in the event database for event tracking, and further includes:

[0085] The title and body of the original data in the event database are respectively vectorized, and it is determined whether there is a same hot event according to the vectorized title and body, and if there is a same hot event, the same hot event is merged into the same hot event.

[0086] In this embodiment, the title and body of the original data can be respectively vectorized by a predetermined vectorization model. Preferably, the predetermined vectorization model is a bge-large-zh model, which can convert the title and body into high-dimensional vectors, so as to determine whether two hot events are the same or repeated hot events through the high-dimensional vectors. In this embodiment, the title similarity is calculated through the high-dimensional vectors of the titles of two hot events, and the body similarity is calculated through the high-dimensional vectors of the bodies of two hot events. When the title similarity and the body similarity are both greater than or equal to a corresponding threshold value, for example, the title similarity is greater than or equal to 0.7, and the body similarity is greater than or equal to 0.63, it is judged that the two hot events are the same or repeated hot events. This embodiment can reduce the repeated operation of event tracking by merging the same hot events in the event database into the same hot event, thereby improving the system performance.

[0087] Step 60, crawling network information of the hot events in the event database, the network information including network interaction information, network transmission platform and network transmission address.

[0088] In the event tracking, the network information of the hot event is crawled in real time or at a fixed time. The network interaction information includes the number of views, the number of likes, the number of comments, and the number of forwards. The network heat of the hot event can be determined through the network interaction information. If the number of views, the number of likes, the number of comments, and the number of forwards are all high, it indicates that the current heat is high. The propagation range of the hot event can be determined through the network propagation platform. If the network propagation platform is more, it indicates that the propagation range is larger. The region of the hot event propagation can be determined through the network propagation address. If the network propagation address is more concentrated, it indicates that the propagation region is more concentrated, otherwise, if it is more dispersed, it indicates that the propagation region is more extensive.

[0089] In step 70, the deterioration score of the hot event is determined according to the network interaction information, and the development dynamic factor of the hot event is determined according to the network propagation platform and the network propagation address.

[0090] In the embodiment, the deterioration score represents the deterioration degree of the development of the hot event under the network interaction, and the development dynamic factor represents the range and region of the development and propagation of the hot event.

[0091] In the embodiment, the deterioration score represents the deterioration degree of the development of the hot event under the network interaction, and the development dynamic factor represents the range and region of the development and propagation of the hot event.

[0092] The first number of views, the first number of likes, the first number of comments, and the first number of forwards at the current time are counted, the second number of views, the second number of likes, the second number of comments, and the second number of forwards at the previous time of the current time are counted, the number of view differences, the number of like differences, the number of comment differences, and the number of forwarding differences are determined according to the first number of views, the first number of likes, the first number of comments, the first number of forwards, the second number of views, the second number of likes, the second number of comments, and the second number of forwards, and the deterioration score is determined based on the predetermined weight and the number of view differences, the number of like differences, the number of comment differences, and the number of forwarding differences.

[0093] The number of newly added network propagation platforms and the number of newly added network propagation addresses of the hot event at the current time are counted, and the development dynamic factor is determined based on the number of newly added network propagation platforms and the number of newly added network propagation addresses.

[0094] The first number of views, the first number of likes, the first number of comments, and the first number of forwards at the current time are respectively V new , L new , C new , and F new , the second number of views, the second number of likes, the second number of comments, and the second number of forwards at the previous time of the current time are respectively V old , L old , Cold and F old The deterioration score M event The calculation formula is:

[0095] ,

[0096] Wherein, 0.3, 0.2, 0.2 and 0.3 are the weights corresponding to the number of views, the number of likes, the number of comments and the number of forwards respectively.

[0097] The embodiment based on the dynamic change rate of network interaction information weighted calculation of deterioration score, realizes the real-time quantification of the evolution of hot events, and solves the problem of early warning lag caused by the fixed threshold of the prior art.

[0098] Further, the network interaction information further includes interaction behavior, interaction time information and interaction user information, and before determining the deterioration score of the hot event according to the network interaction information, it further includes:

[0099] The network interaction information is input into a predetermined double module collaborative network, false interaction information in the network interaction information is identified based on the double module collaborative network, and the false interaction information is removed, the double module collaborative network includes a generator and a discriminator, the generator includes a long short-term memory network and a deconvolution network, and the discriminator includes a convolutional neural network and an attention mechanism network;

[0100] Determining the deterioration score of the hot event according to the network interaction information includes determining the deterioration score of the hot event according to the network interaction information after removing the false interaction information.

[0101] The interaction time information refers to the occurrence time of browsing, liking, commenting or forwarding interaction behaviors. The hot event can be disturbed by false interaction information, including water army interaction information, which affects the accuracy of event early warning. In order to remove the disturbance of false interaction information and improve the accuracy of event early warning, the embodiment identifies false interaction information through a double-module cooperative network. The generator inputs interaction behaviors, interaction time information, interaction user information, browsing volume, like volume, comment volume and forwarding volume. The generator includes 3 layers of long short-term memory network and 1 layer of deconvolution network. The 3 layers of long short-term memory network capture the time sequence features in the network interaction information, and the 1 layer of deconvolution network generates the behavior (including browsing volume, like volume, comment volume and forwarding volume) sequence. The generator outputs the dense behavior sequence (such as 50 likes in 30 seconds) simulated by the water army. The discriminator inputs the above data input by the generator and the data output by the generator. The convolutional neural network of the discriminator captures the local mode, and the attention mechanism network identifies the abnormal time aggregation behavior. The discriminator outputs the probability value (0, -1) of the authenticity of the interaction behavior. If it is 0, it means that the probability value of authenticity is 0, and if it is 1, it means that the probability value of authenticity is 1. Through the double-module cooperative network, the embodiment can identify the coordinated forwarding of a large number of accounts to the hot event in a short time, so as to remove false interaction information.

[0102] In the embodiment, the number of network transmission platforms and the number of network transmission addresses of the current hot event are counted. The development dynamic factor dynamic_factor increases by 3% for every 1 unit increase of the network transmission platform (Reach, R) and the network transmission address (Affected_Area, A): Wherein, the development dynamic factor dynamic_factor is a positive number greater than 1. In step 80, the development stage of the hot event is determined according to the deterioration score and the development dynamic factor, and event early warning is performed.

[0103] In the embodiment, the development stage of the hot event includes the incubation stage, the fermentation stage, the outbreak stage and the dissipation stage. In the incubation stage, it belongs to a small range of discussion (such as a local post in a forum), and has not formed a diffusion trend. In the fermentation stage, it can be a single platform hot search (such as the top 50 of Weibo), regional hot discussion, diffusion acceleration but not overall outbreak. In the outbreak stage, it exists in multiple platform hot search Top10 (such as Weibo + TikTok...), and presents cross-provincial transmission, triggering nationwide discussion. In the dissipation stage, the network discussion cools down (such as the rapid decline of the malignant value after the press conference).

[0104] The development stage of the hot event is determined according to the deterioration score and the development dynamic factor, and event early warning is performed, including:

[0105] If the deterioration score is less than a preset first threshold value, it is determined that the hotspot event is in a latent stage;

[0106] If the deterioration score is greater than or equal to the first threshold value and less than the development dynamic factor, it is determined that the hotspot event is in a fermentation stage;

[0107] If the deterioration score is greater than or equal to the development dynamic factor, it is determined that the hotspot event is in an outbreak stage.

[0108] If the hotspot event has entered the outbreak stage and the deterioration score is less than or equal to half of the development dynamic factor, it is determined that the hotspot event is in a dissipation stage.

[0109] Preferably, the first threshold value is 0.5. When the deterioration score M event is less than 0.5, it belongs to an event discussed in a small range, and it is determined that the hotspot event is in a latent stage. The embodiment can track the event that is not hot yet in real time, improve real-time performance, and avoid early warning lag based on the dynamically changing deterioration score to predict the latent period. event is greater than or equal to 0.5 and less than the development dynamic factor dynamic_factor, it is in a single platform hot search, and it is determined that the hotspot event is in a fermentation stage. If the deterioration score M event is greater than or equal to the development dynamic factor dynamic_factor, there is a multi-platform hot search Top10, and it is determined that the hotspot event is in an outbreak stage. If the hotspot event has entered the outbreak stage and the deterioration score M event is less than or equal to half of the development dynamic factor dynamic_factor, it is determined that the hotspot event is in a dissipation stage.

[0110] The dynamic adaptive event development prediction method of the embodiment of the application determines the subject of the topic to which the original data belongs through multi-level keyword matching of the original data, determines the sentiment tendency label and the number of negative views of the original data through a large language model, comprehensively judges whether the event occurring in the original data is a hotspot event by combining the subject of the topic, the sentiment tendency label and the number of negative views, and if it is a hotspot event, it is stored and tracked. When tracking the event, the network information of the hotspot event is crawled, the deterioration score is dynamically calculated according to the network interaction information, the development dynamic factor is determined according to the network transmission platform and the network transmission address, and finally the development stage of the hotspot event is determined according to the deterioration score and the development dynamic factor and an early warning is performed. The embodiment dynamically and adaptively judges the development stage of the hotspot event through the deterioration score and the development dynamic factor, and the real-time performance and accuracy are better.

[0111] Figure 2 A structure schematic diagram of a dynamic adaptive event development prediction device of the embodiment of the application is shown. As Figure 2As shown, the device 200 comprises:

[0112] The matching module 201 is configured to perform multi-level keyword matching on the preprocessed original data, and determine a special subject to which the original data belongs after successful matching.

[0113] The analysis module 202 is configured to perform sentiment tendency analysis on the original data based on a preset first large language model, and obtain a sentiment tendency label corresponding to the original data output by the first large language model.

[0114] The refining module 203 is configured to refine negative views of the original data based on a preset second large language model, and obtain a number of negative views corresponding to the original data output by the second large language model.

[0115] The first determination module 204 is configured to determine whether an event occurring in the original data is a hot event according to the special subject, the sentiment tendency label and the number of negative views.

[0116] The saving module 205 is configured to save the original data as event data in an event database for event tracking if the event occurring in the original data is a hot event.

[0117] The crawling module 206 is configured to crawl network information of the hot event in the event database, wherein the network information comprises network interaction information, a network transmission platform and a network transmission address.

[0118] The second determination module 207 is configured to determine a deterioration score of the hot event according to the network interaction information, and determine a development dynamic factor of the hot event according to the network transmission platform and the network transmission address.

[0119] The third determination module 208 is configured to determine a development stage of the hot event and perform event early warning according to the deterioration score and the development dynamic factor.

[0120] The dynamic adaptive event development prediction device of the embodiment is basically the same as the dynamic adaptive event development prediction method described above, and the embodiments of the dynamic adaptive event development prediction method can be referred to.

[0121] Figure 3 The structure schematic diagram of the computer device embodiment of the present application is shown, and the specific implementation of the computer device is not limited in the specific embodiment of the present application.

[0122] As Figure 3As shown, the computer device can include a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0123] The processor 402, the communications interface 404, and the memory 406 can communicate with each other through the communications bus 408. The communications interface 404 is configured to communicate with network elements such as clients or other servers, etc. The processor 402 is configured to execute the program 410, and specifically can execute the related steps in the above embodiments for the computer device.

[0124] Specifically, the program 410 can include program codes including computer executable instructions.

[0125] The processor 402 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present application. The one or more processors included in the computer device can be the same type of processor, such as one or more CPUs; or can be different types of processors, such as one or more CPUs and one or more ASICs.

[0126] The memory 406 is configured to store the program 410. The memory 406 can include a high-speed RAM memory, and can also include a non-volatile memory such as at least one disk memory.

[0127] The program 410 can be specifically invoked by the processor 402 to cause the computer device to perform the following operations:

[0128] Multi-level keyword matching is performed on the preprocessed original data, and after successful matching, a subject matter to which the original data belongs is determined;

[0129] Based on a preset first large language model, sentiment tendency analysis is performed on the original data, and a sentiment tendency label corresponding to the original data output by the first large language model is obtained;

[0130] Based on a preset second large language model, negative viewpoints are extracted from the original data, and a number of negative viewpoints corresponding to the original data output by the second large language model is obtained;

[0131] determine whether the event occurred in the original data is a hot event according to the topic subject, the sentiment tendency label and the number of negative opinions; if the event occurred in the original data is a hot event, save the original data as event data into an event database for event tracking;

[0132] crawl network information of the hot event in the event database, the network information including network interaction information, a network transmission platform and a network transmission address;

[0133] determine a deterioration score of the hot event according to the network interaction information, and determine a development dynamic factor of the hot event according to the network transmission platform and the network transmission address;

[0134] determine a development stage of the hot event and perform event early warning according to the deterioration score and the development dynamic factor.

[0135] In an optional manner, the preprocessed original data is subjected to multi-level keyword matching, and after successful matching, the topic subject to which the original data belongs is determined, including:

[0136] loading general keywords in a preset dynamic general keyword library, and matching the preprocessed original data with the general keywords;

[0137] if the preprocessed original data matches the general keywords successfully, loading special topic keywords in a preset dynamic special topic keyword library, and matching the preprocessed original data with the special topic keywords;

[0138] if the preprocessed original data matches the special topic keywords successfully, determining the topic subject to which the original data belongs.

[0139] In an optional manner, the determination of whether the event occurred in the original data is a hot event according to the topic subject, the sentiment tendency label and the number of negative opinions includes:

[0140] determining the topic subject and a corresponding risk category according to a preset topic risk dimension relationship, and determining a risk point to which the topic subject belongs under the risk category, and determining a risk value of the topic subject according to a risk level corresponding to the risk point;

[0141] determining a sentiment score corresponding to the sentiment tendency label according to a preset sentiment score relationship;

[0142] determining an opinion score corresponding to the number of negative opinions according to a preset opinion score relationship;

[0143] determine a sensitive value corresponding to the original data according to the risk value, the sentiment score and the opinion score, and determine whether an event occurring in the original data is a hot event according to the sensitive value.

[0144] In an optional manner, the network interaction information includes a browsing quantity, a like quantity, a comment quantity and a forwarding quantity, the deterioration score of the hot event is determined according to the network interaction information, and the development dynamic factor of the hot event is determined according to the network transmission platform and the network transmission address, including:

[0145] The first browsing quantity, the first like quantity, the first comment quantity and the first forwarding quantity of the current moment are counted, the second browsing quantity, the second like quantity, the second comment quantity and the second forwarding quantity of the last moment of the current moment are counted, the browsing quantity difference value, the like quantity difference value, the comment quantity difference value and the forwarding quantity difference value are respectively determined according to the first browsing quantity, the first like quantity, the first comment quantity, the first forwarding quantity, the second browsing quantity, the second like quantity, the second comment quantity and the second forwarding quantity, and the deterioration score is determined based on the predetermined weight and the browsing quantity difference value, the like quantity difference value, the comment quantity difference value and the forwarding quantity difference value;

[0146] The number of network transmission platforms and the number of network transmission addresses newly added to the hot event at the current moment are counted, and the development dynamic factor is determined based on the number of newly added network transmission platforms and the number of newly added network transmission addresses.

[0147] In an optional manner, the development dynamic factor is greater than one, and the development stage of the hot event and the event warning are determined according to the deterioration score and the development dynamic factor, including:

[0148] If the deterioration score is less than a preset first threshold value, it is determined that the hot event is in a latent stage;

[0149] If the deterioration score is greater than or equal to the first threshold value and less than the development dynamic factor, it is determined that the hot event is in a fermentation stage;

[0150] If the deterioration score is greater than or equal to the development dynamic factor, it is determined that the hot event is in an outbreak stage;

[0151] If the hot event has entered the outbreak stage and the deterioration score is less than or equal to half of the development dynamic factor, it is determined that the hot event is in a dissipation stage.

[0152] In an optional manner, the network interaction information further includes interaction behavior, interaction time information and interaction user information, and before determining the deterioration score of the hot event according to the network interaction information, further including:

[0153] input the network interaction information into a predetermined dual-module collaborative network, identify false interaction information in the network interaction information based on the dual-module collaborative network, and eliminate the false interaction information, the dual-module collaborative network comprising a generator and a discriminator, the generator comprising a long short-term memory network and a deconvolution network, and the discriminator comprising a convolutional neural network and an attention mechanism network;

[0154] The deterioration score of the hot event is determined according to the network interaction information.

[0155] The deterioration score of the hot event is determined according to the network interaction information after the false interaction information is eliminated.

[0156] In an optional manner, if the event occurred in the original data is a hot event, the original data is saved as event data into an event database for event tracking, and the method further comprises:

[0157] The title and the body of the original data in the event database are respectively vectorized, and it is determined whether there is a same hot event according to the vectorized title and body, and if there is a same hot event, the same hot event is merged into a same hot event.

[0158] The embodiment of the present application provides a computer readable storage medium, the storage medium stores at least one executable instruction, the executable instruction runs on the computer equipment, and the computer equipment executes any method embodiment.

[0159] The embodiment of the present application provides a computer program, the computer program can be called by a processor to make a computer equipment execute any method embodiment.

[0160] The embodiment of the present application provides a computer program product, the computer program product comprises a computer program stored on a computer readable storage medium, the computer program comprises program instructions, and when the program instructions run on a computer, the computer executes any method embodiment.

[0161] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with these teachings, or with variations thereof, depending on the implementation. The structure required to be present on such systems to implement the present application is apparent from the above description. In addition, the present application is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages can be used to implement the present application as described herein, and any references below to specific languages are provided for disclosure of enablement only.

[0162] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application can be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been described in detail in order not to obscure the understanding of this description.

[0163] Similarly, it is to be understood that the embodiments of the application can be adapted to other applications and that their features can be interchanged in any way. In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application can be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been described in detail in order not to obscure the understanding of this description.

[0164] Those skilled in the art will appreciate that the modules in the computer device in the embodiments can be adapted and arranged in one or more computer devices different from the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and can be divided into multiple sub-modules or sub-units or sub-components. All the features disclosed in the specification (including the claims, abstract and drawings) and all the processes or units of any method or computer device disclosed in the specification can be combined in any combination, except that at least some of such features and / or processes or units are mutually exclusive. Each feature disclosed in the specification (including the claims, abstract and drawings) can be replaced by alternative features that serve the same, equivalent or similar purpose, unless otherwise expressly stated.

[0165] It should be noted that the above-mentioned embodiments illustrate rather than limit the application, and that one skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of elements or steps other than those listed in a claim. The word 'a' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In a unit claim, several devices can be listed, even though they are not, in fact, incoporated into a single physical device. The use of the word 'at least' followed by a list of one or more items does not exclude additional such items. The use of the terms 'first','second' and 'third', etc. does not limit the scope of the application, but merely identifies a name of an element. Steps in the methods provided herein do not have to be performed in the exact order disclosed, unless a specific order is required (for example, a step cannot be performed or understood unless a previous step has already been completed).

Claims

1. A dynamic adaptive event development prediction method, characterized in that, The method includes: Multi-level keyword matching is performed on the preprocessed raw data. After a successful match, the subject matter to which the raw data belongs is determined. Based on a preset first language model, sentiment analysis is performed on the raw data to obtain the sentiment labels corresponding to the raw data output by the first language model. Based on a preset second language model, negative opinions are extracted from the original data to obtain the number of negative opinions corresponding to the original data output by the second language model. Based on the topic subject, sentiment tags, and the number of negative opinions, determine whether the events occurring in the raw data are hot topics; If the events occurring in the raw data are hot events, the raw data will be saved as event data to the event database for event tracking. The network information of the hot events in the event database is crawled, including network interaction information, network dissemination platforms, and network dissemination addresses. The deterioration score of the hot topic event is determined based on the network interaction information, and the development dynamic factor of the hot topic event is determined based on the network propagation platform and the network propagation address. Based on the deterioration score and the development dynamic factor, the development stage of the hotspot event is determined, and event early warning is issued, including: If the deterioration score is less than a preset first threshold, then the hotspot event is determined to be in a latent stage; If the deterioration score is greater than or equal to the first threshold and less than the development dynamic factor, then the hot topic event is determined to be in the fermentation stage. If the deterioration score is greater than or equal to the development dynamic factor, then the hotspot event is determined to be in the outbreak stage; If the hot topic event has entered the outbreak stage and the deterioration score is less than or equal to half of the development dynamic factor, then the hot topic event is determined to be in the dissipation stage.

2. The method according to claim 1, characterized in that, The process of performing multi-level keyword matching on the preprocessed raw data, and determining the topic subject to which the raw data belongs after a successful match, includes: Load common keywords from a preset dynamic common keyword library, and match the preprocessed raw data with the common keywords; If the preprocessed original data matches the general keyword, then the topic keywords in the preset dynamic topic keyword library are loaded, and the preprocessed original data is matched with the topic keywords. If the preprocessed raw data successfully matches the topic keywords, then the topic to which the raw data belongs is determined.

3. The method according to claim 1 or 2, characterized in that, The process of determining whether an event in the raw data is a trending event based on the topic subject, sentiment tags, and the number of negative opinions includes: Based on the preset risk dimension relationship, the subject of the topic and its corresponding risk category are determined, and the risk point to which the subject of the topic belongs under the risk category is determined. The risk value of the subject of the topic is determined according to the risk level corresponding to the risk point. Based on the preset emotional score relationship, determine the emotional score corresponding to the emotional tendency label; Based on a preset opinion score relationship, the opinion score corresponding to the number of negative opinions is determined; Based on the risk value, the sentiment score, and the opinion score, a sensitivity value corresponding to the original data is determined, and based on the sensitivity value, it is determined whether the event occurring in the original data is a hot topic.

4. The method according to claim 1, characterized in that, The online interaction information includes page views, likes, comments, and shares. The process of determining the deterioration score of the trending event based on the online interaction information, and determining the dynamic factors of the trending event's development based on the online dissemination platform and the online dissemination address, includes: The system calculates the first number of views, the first number of likes, the first number of comments, and the first number of shares at the current moment. It also calculates the second number of views, the second number of likes, the second number of comments, and the second number of shares at the previous moment. Based on the first number of views, the first number of likes, the first number of comments, the first number of shares, the second number of views, the second number of likes, the second number of comments, and the second number of shares, the system determines the difference in number of views, the difference in number of likes, the difference in number of comments, and the difference in number of shares. Based on predetermined weights and the differences in number of views, the difference in number of likes, the difference in number of comments, and the difference in number of shares, the system determines the deterioration score. The number of newly added online dissemination platforms and addresses for the current hot topic event is counted, and the development dynamic factor is determined based on the number of newly added online dissemination platforms and addresses.

5. The method according to claim 4, characterized in that, The network interaction information also includes interaction behavior, interaction time information, and interaction user information. Before determining the deterioration score of the hot topic event based on the network interaction information, the process further includes: The network interaction information is input into a predetermined dual-module collaborative network. Based on the dual-module collaborative network, false interaction information in the network interaction information is identified and removed. The dual-module collaborative network includes a generator and a discriminator. The generator includes a long short-term memory network and a deconvolutional network. The discriminator includes a convolutional neural network and an attention mechanism network. The step of determining the deterioration score of the hot topic event based on the network interaction information includes: The deterioration score of the hot topic event is determined based on the network interaction information after removing the false interaction information.

6. The method according to claim 1, characterized in that, If the events occurring in the original data are hot events, then after saving the original data as event data to the event database for event tracking, the method further includes: The titles and texts of the original data in the event database are vectorized respectively. Based on the vectorized titles and texts, it is determined whether there are identical hot events. If identical hot events exist, the identical hot events are merged into the same hot event.

7. A dynamic adaptive event development prediction device, characterized in that, The device includes: The matching module is used to perform multi-level keyword matching on the preprocessed raw data, and after a successful match, determine the topic subject to which the raw data belongs; The analysis module is used to perform sentiment analysis on the raw data based on a preset first language model, and to obtain the sentiment labels corresponding to the raw data output by the first language model. The extraction module is used to extract negative opinions from the original data based on a preset second language model, and to obtain the number of negative opinions corresponding to the original data output by the second language model. The first determining module is used to determine whether an event occurring in the raw data is a hot topic based on the topic subject, sentiment tags, and the number of negative opinions. The storage module is used to save the original data as event data to the event database for event tracking if the events occurring in the original data are hot events. The crawling module is used to crawl the network information of the hot events in the event database, including network interaction information, network dissemination platforms, and network dissemination addresses. The second determining module is used to determine the deterioration score of the hot topic event based on the network interaction information, and to determine the development dynamic factor of the hot topic event based on the network propagation platform and the network propagation address. The third determining module is used to determine the development stage of the hotspot event and to issue an event warning based on the deterioration score and the development dynamic factor, wherein: If the deterioration score is less than a preset first threshold, then the hotspot event is determined to be in a latent stage; If the deterioration score is greater than or equal to the first threshold and less than the development dynamic factor, then the hot topic event is determined to be in the fermentation stage. If the deterioration score is greater than or equal to the development dynamic factor, then the hotspot event is determined to be in the outbreak stage; If the hot topic event has entered the outbreak stage and the deterioration score is less than or equal to half of the development dynamic factor, then the hot topic event is determined to be in the dissipation stage.

8. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the method as described in any one of claims 1-6.

Images