Event analysis report generation method

Abstract

The invention discloses an event analysis report generation method, and belongs to the technical field of information processing. According to the scheme, a traditional NLP machine learning algorithm and a plurality of AI agents are integrated, an event analysis report generation task is split, the NLP machine learning algorithm is used for early-stage data preprocessing, and the length limitation of large language model input is avoided from the source; prompt language packaging and agent combination design are carried out based on large language models with different capabilities, each type of AI agents process independent tasks, the capability boundary of a single AI agent is broken through, meanwhile, the defect that the understanding capability of an NLP deep learning model is insufficient is avoided, the accuracy of semantic understanding and event analysis is improved, and the semantic understanding and event analysis efficiency is improved. And the depth and precision of the report content are ensured, so that more efficient, more complex and more intelligent report content generation is realized, and the comprehensiveness and precision of the content are improved. In addition, a flexible report template and a customization function are provided, and the report content can be adjusted according to specific requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing, and particularly to a method for generating an event analysis report. Background Art

[0002] With the rapid development of artificial intelligence technology, especially the wide application of NLP (Natural Language Processing) technology, the automatic generation of event analysis reports has gradually become a research and practice hotspot. The current automatic generation technical solutions for event analysis reports mainly include methods based on traditional NLP deep learning models, methods based on large language models, and the way of manually writing reports. These methods have their own advantages and disadvantages, but there are still obvious deficiencies in practical applications and cannot fully meet the user's requirements for efficient, intelligent, and accurate report generation.

[0003] 1. Traditional NLP Deep Learning Models

[0004] Traditional NLP deep learning models, such as LSTM (Long Short-Term Memory networks), GRU (Gated Recurrent Unit), etc., mainly generate reports through word frequencies, context relevance, and simple semantic rules. Such models perform poorly in processing semantic understanding and logical reasoning of complex events, and the generated content usually has problems such as lack of depth and limited coverage, and it is difficult to generate reports with rich content and clear organization.

[0005] 2. Large Language Models

[0006] In recent years, large language models, such as GPT (Generative Pre-trained Transformer), BERT (Bidirectional Encoder Representations from Transformers), etc., have made remarkable progress in semantic understanding and text generation. Such models can better capture semantic relationships in text and generate natural and fluent content. However, large language models still have two major limitations:

[0007] First, length limitation: When dealing with extremely long texts, large language models are prone to semantic loss or omission of important information. Especially when integrating a large amount of event information, it is difficult to comprehensively cover all aspects. For the application scenario of generating event analysis reports, the idea is to input all news and posts, and let the large model understand and then output the report chapters and content that users want. However, in actual implementation, it will be found that the large model is limited by the length limitation and cannot achieve the above tasks. For example, if the character count of news and posts related to an event is 500,000 or more than 1 million, but the large model may only output a maximum of 128,000 characters, which will cause an output error.

[0008] Second, ability boundary: Although a single large language model performs well in a certain semantic generation aspect, for the understanding of domain-specific knowledge and logical reasoning, it is still restricted by the training data and model structure, and the generated content may have misunderstandings or errors. For example, DeepSeek-R1 and its fine-tuned models have strong reasoning capabilities, but their performance in outputting JSON format is not good; another example is Qwen and its fine-tuned models, whose reasoning ability is not as good as DeepSeek, but they have better performance in outputting the standard JSON format.

[0009] 3. Handwritten reports

[0010] Manual handwritten reports, relying on experts' comprehensive understanding and in-depth analysis of events, can often provide the most accurate and logical content. However, this method requires a large amount of human resources and takes a long time, making it difficult to meet the need for quickly generating analysis reports. In the face of emergencies, real-time public opinion and other scenarios, the efficiency problem of handwritten reports is particularly prominent, restricting its application in high-timeliness scenarios. Summary of the Invention

[0011] The technical problem to be solved by the present invention is: In view of the above defects of the prior art, to provide an event analysis report generation method that integrates NLP machine learning algorithms and multiple AI agents, significantly improving the intelligence, accuracy, and real-time performance of event analysis reports to meet the efficient requirements of complex event analysis and report generation in multiple fields.

[0012] To achieve the above object, the present invention provides an event analysis report generation method, and the method includes the following steps:

[0013] Step S1, real-time collect news data from the target news platform and post data from the target social media platform;

[0014] Step S2: Integrate and clean the news data and post data, and use NLP machine learning algorithms to perform preliminary classification, keyword extraction, entity extraction, and event clustering on the news data, and perform keyword extraction and entity extraction on the post data to form structured data and store it in the source database;

[0015] Step S3: According to the input configuration conditions, screen out the target data set related to the target event from the source database; based on a preset event analysis report template, introduce NLP machine learning algorithms as needed to perform preliminary feature engineering processing on the target data set; call multiple AI agents to generate the content of each part in the event analysis report template; the multiple AI agents are multiple AI agents with independent functions;

[0016] Step S4: Visualize and display the content of the event analysis report generated in Step S3.

[0017] In the event analysis report generation method of the present invention, the configuration conditions include associated event configuration, and the associated event configuration includes the event name, event keyword phrase, excluded keyword phrase, and analysis scope of the associated event. The analysis scope includes the site source, and the online and offline times for continuous collection and monitoring of the event.

[0018] In the event analysis report generation method of the present invention, the event analysis report template includes an event summary;

[0019] The method for generating the event summary is as follows:

[0020] According to the event keyword phrase, excluded keyword phrase, and analysis scope, screen out multiple target sub-events from the source database. Each target sub-event includes a sub-event name, occurrence time, and news content;

[0021] Call the event content summary AI agent, use the sub-event name as a precondition, select the news content with the highest view count in the sub-event as the input, and generate a content summary of a single target sub-event through context semantic analysis;

[0022] Use the content summaries of multiple target sub-events as the input, use the event name as a precondition, call the event summary text generation AI agent, and generate an event summary through context semantic analysis; the event summary includes the time, location, participating entities, description of the event development process, results, or subsequent impacts of the event.

[0023] In the event analysis report generation method of the present invention, the event analysis report template includes an event context;

[0024] The method for generating the event context is as follows:

[0025] Obtain multiple target sub - event information obtained during the event summary generation process, including sub - event name, occurrence time, news content, and content summary; count the popularity of each target sub - event, sort the target sub - events in chronological order to form structured time - series data;

[0026] Call the event context summarization generation AI agent. Using the event name as a pre - condition and the content summaries of multiple target sub - events as input, perform semantic parsing on each target sub - event and summarize the overall development context of the event in stages;

[0027] Taking the target sub - events as nodes, the names and popularities of the target sub - events as node labels, arranging the nodes in chronological order, draw an event context diagram, and display the content summaries of each target sub - event in a floating window form.

[0028] In the event analysis report generation method of the present invention, the event analysis report template includes the volume trend;

[0029] The generation method of the volume trend is as follows:

[0030] According to the event keyword phrases, excluded keyword phrases, and event analysis time range, extract the full - volume data of the target event from the source database. Obtain the volume raw data and target sub - events based on the full - volume data. The volume raw data includes the number of posts, interaction volume, and news report volume; the target sub - events include sub - event name and occurrence time;

[0031] Unify the time format of the volume raw data and perform segmented processing according to the event occurrence date to form multiple date - segmented volume raw data;

[0032] Calculate the volume value for each of the multiple date - segmented volume raw data. The volume value = number of posts+interaction volume+news report volume;

[0033] Generate a volume trend chart based on the date segmentation and volume value;

[0034] Based on the statistical analysis of the news report volume, identify the target sub - event with the highest daily popularity as the key sub - event, and mark the key sub - event as the peak or turning point in the volume trend chart.

[0035] In the event analysis report generation method of the present invention, the event analysis report template includes the event influence distribution;

[0036] The generation method of the event influence distribution is as follows:

[0037] Extract all the data of the target event from the source database according to the event keyword phrases, the excluded keyword phrases, and the event analysis time range. Obtain the original influence data based on the all data, where the original influence data includes the number of posts and the number of news reports, and classify and summarize them by the dissemination platform.

[0038] Calculate the platform influence and the influence proportion of each dissemination platform; Platform influence = number of posts + number of news reports; Influence proportion = influence of this platform / influence of all platforms, and display the influence proportion of each dissemination platform in graphical form.

[0039] Sort the influence proportions of each dissemination platform to identify the main dissemination platforms; Invoke the event platform influence analysis AI agent, using the event name as a precondition, take the main dissemination platforms, the number of posts / reports, and the influence percentage as inputs, and summarize the dissemination characteristics of the event through semantic analysis, and output an analysis result with fluent and professional language.

[0040] In the event analysis report generation method of the present invention, the event analysis report template includes the analysis of the dissemination network and dissemination nodes.

[0041] The generation method of the analysis of the dissemination network and dissemination nodes is as follows:

[0042] Extract the original data of the social media platform of the target event from the source database according to the event keyword phrases, the excluded keyword phrases, and the event analysis time range. Obtain the original dissemination data based on the original data of the social media platform, where the original dissemination data includes social accounts, the number of posts, and the number of interactions.

[0043] Clean the original dissemination data to construct a relationship matrix of the nodes and edges of the dissemination network.

[0044] Use the community discovery algorithm to partition the dissemination network to identify the main community structures in the dissemination process.

[0045] Rank the nodes within each community according to the dissemination metrics to identify the core dissemination nodes; The dissemination metrics include the number of interactions.

[0046] Calculate the key metrics of the dissemination network, and further filter the communities based on the key metrics to obtain the important communities of the event dissemination.

[0047] Further screen out the top-ranked key dissemination nodes from the core dissemination nodes based on the key metrics, and extract the node information of each key dissemination node, where the node information includes the post content and the number of interactions.

[0048] Use the NLP clustering algorithm to cluster all the post contents corresponding to each selected key dissemination node, so that the clustering results are grouped according to similar contents.

[0049] After clustering is completed, the top M categories are obtained according to the interaction volume, and the most representative N posts in each category are extracted to form a list of posts to be analyzed.

[0050] Taking the event name and event summary as the preconditions for analysis, the posts in the list of posts to be analyzed for each key communication node are input one by one into the post view generation AI intelligent agent for semantic understanding, generating a clear and logically rigorous view summary, and associating the generated view with the corresponding key communication node to form a complete view description of each key communication node.

[0051] In the method for generating an event analysis report of the present invention, the event analysis report template includes news view analysis;

[0052] The generation method of the news view analysis is as follows:

[0053] According to the event keyword phrases, excluded keyword phrases, and analysis scope, multiple target sub-events are screened out from the source database, and each target sub-event includes a sub-event name, an occurrence time, and news content;

[0054] Count the number of news corresponding to each target sub-event, sort them in descending order of the number of news, and select the top K target sub-events as the analysis objects to form the original data set for generating news views;

[0055] Taking the event name and event summary as the input preconditions, for each target sub-event, the first P news texts are extracted in order of the view volume as the input content of the news view generation AI intelligent agent; the news view generation AI intelligent agent performs semantic parsing on the news content, identifies the core view, and generates a news comment.

[0056] In the method for generating an event analysis report of the present invention, the event analysis report template includes netizen view analysis;

[0057] The generation method of the netizen view analysis is as follows:

[0058] According to the event keyword phrases, excluded keyword phrases, and event analysis time range, the original post data set of the social media platform of the target event is extracted from the source database;

[0059] Randomly select L original post contents from the deduplicated original post data set to obtain a random post data set;

[0060] Use the NLP sentiment analysis algorithm to perform sentiment analysis on the original post contents in the random post data set, classify the emotions of the posts into three categories: positive, negative, and neutral, and display the proportion of each type of emotion through a graph;

[0061] Group the random post dataset based on the clustering algorithm to form multiple clustered datasets; display the proportion of each clustered dataset through a graph.

[0062] Select the Q clustered datasets with the largest number of posts. For each clustered dataset, use all its original post content as the input content, call the post view generation AI agent to generate the description of the netizen's view, and select the most prevalent emotion in the clustered dataset as the main emotion of the clustered dataset.

[0063] In the method for generating an event analysis report of the present invention, after the step S3, the following is further included:

[0064] Regenerate the event analysis report according to the user's feedback and configuration.

[0065] The present invention has the following beneficial effects: The method for generating an event analysis report of the present invention integrates traditional NLP machine learning algorithms and multiple AI agents, splits the event analysis report generation task, uses NLP machine learning algorithms for pre - processing data in the early stage, reduces the length of the input data for AI agents, and avoids the length limitation of the input of large language models from the source; designs prompt encapsulation based on large language models with different capabilities to form multiple types of AI agents, each type of AI agent processes independent tasks, breaks through the ability boundary of a single AI agent, and at the same time avoids the shortcoming of insufficient understanding ability of NLP deep learning models, improves the accuracy of semantic understanding and event analysis, ensures the depth and precision of the report content, thereby realizing more efficient, more complex, and more intelligent report content generation, and improving the comprehensiveness and accuracy of the content. In addition, flexible report templates and customization functions are provided, which can adjust the report content according to specific requirements and improve the practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The drawings described herein are used to provide a further understanding of the present invention, form a part of the present invention, and the schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0067] Figure 1 It is a schematic diagram of the steps of the method for generating an event analysis report provided by an embodiment of the present invention.

[0068] Figure 2 It is a schematic diagram of the architecture of the method for generating an event analysis report provided by an embodiment of the present invention.

[0069] Figure 3 It is a schematic diagram of the configuration interface of the event analysis report provided by an embodiment of the present invention.

[0070] Figure 4 It is an example of the event summary and event context generated by an embodiment of the present invention.

[0071] Figure 5 An example of the volume trend generated for the embodiments of the present invention.

[0072] Figure 6 An example of the influence distribution generated for the embodiments of the present invention.

[0073] Figure 7 An example of the propagation network generated for the embodiments of the present invention.

[0074] Figure 8 An example of the key propagation nodes generated for the embodiments of the present invention.

[0075] Figure 9 An example of the news viewpoints generated for the embodiments of the present invention.

[0076] Figures 10 - 11 An example of the viewpoints of Internet users generated for the embodiments of the present invention. Detailed implementation manners

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0078] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings of the specification. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0079] The present invention is applicable to application scenarios that require quickly and automatically generating high-quality event analysis reports based on the information disclosed by news platforms and social media platforms, and has significant advantages especially in the fields of emergencies and rapid responses.

[0080] As Figure 1 and Figure 2 shown, the embodiments of the present invention provide an event analysis report generation method, and the method includes the following steps:

[0081] Step S1, collecting news data of a target news platform and post data of a target social media platform in real time.

[0082] In the embodiments of the present invention, hot news and social account data are captured in real time, supporting the collection and integration of unstructured data in multiple sites and formats, generating high-quality and timely raw data, laying a foundation for subsequent event analysis. Specifically, through web crawler technology, data from specified news platforms is captured in real time, and the data sources cover multiple mainstream sites globally, extracting key information such as news titles, texts, and release times. The target accounts on the target social media platform are monitored in real time, and the post information of the target accounts is collected, including original posts, forwarded posts, comment posts, etc. At the same time, the interaction data of the post information is extracted, including the number of likes, comments, and forwards.

[0083] Step S2: Integrate and clean the news data and post data, and use NLP machine learning algorithms to perform preliminary classification, keyword extraction, entity extraction, and event clustering on the news data, and perform keyword extraction and entity extraction on the post data to form structured data and store it in the source database.

[0084] The integration and cleaning include deduplication, format conversion, and word segmentation processing of multi-source data to facilitate subsequent processing using NLP machine learning algorithms. At the same time, integrating and cleaning the raw data and event clustering are also beneficial for reducing the subsequent data processing volume and reducing the input length of the large language model.

[0085] The entities include personal names, place names, organization names, etc. The event clustering refers to news clustering based on keywords and graph models, and automatically generating event names based on NLP machine learning algorithms.

[0086] Step S3: According to the input configuration conditions, filter out the target data set related to the target event from the source database; based on a preset event analysis report template, introduce NLP machine learning algorithms as needed to perform preliminary feature engineering processing on the target data set; call multiple AI agents to generate the content of each part in the event analysis report template; the multiple AI agents are multiple AI agents with independent functions.

[0087] In the embodiments of the present invention, as Figure 3 shown, the configuration conditions include associated event configuration, and the associated event configuration includes the event name, event keyword phrases, excluded keyword phrases, and analysis scope of the associated event. The analysis scope includes site sources, the online and offline times for continuous collection and monitoring of the event. Site sources include news, forums, blogs, social media, etc. In addition, the report name and the event scope of the event analysis can also be configured. In the embodiments of the present invention, by setting event keyword phrases and excluded keyword phrases simultaneously, the retrieval efficiency is improved; by matching the theme with keywords + time + source, it is more accurate and efficient compared to the method of gradually identifying the user's intention through AI conversations in the prior art.

[0088] The preset event analysis report template defines the components of the event analysis report. Embodiments of the present invention provide a user operation interface, enabling users to adjust the report content, structure, and style as needed to generate a personalized event analysis report template.

[0089] In some embodiments of the present invention, the event analysis report template and the descriptions of its components are as follows.

[0090] 1) Event summary, including the time, location, participating entities, description of the development process, results, or subsequent impacts of the event. Through the event summary, the basic situation of the event can be understood.

[0091] 2) Event trend, including two parts: the event context and the volume trend. Among them, the event context is the timeline of the event from outbreak to development, presented in two forms: a text summary and a context diagram. The context diagram consists of the descriptions of sub-events in the event, the occurrence time, and the event popularity. The text summary is a written summary of the context diagram; the volume trend is the change in the popularity of the event in social media and news reports, presented in the form of a line chart, with the key moments of the event development on the x-axis and the event popularity statistics on the y-axis.

[0092] 3) Influence distribution, presented in the form of a pie chart to show the influence distribution of the event on major communication platforms to identify the most active communication platforms.

[0093] 4) Communication analysis, including two parts: the communication network and the key communication nodes. Among them, the communication network visually presents the network structure of the main communication nodes and information diffusion on all social media platforms; the key communication nodes list the information of key figures / institutions in the communication network, including the account name, social media platform, main viewpoints, the link to the post corresponding to the viewpoint, and the post content.

[0094] 5) Summary analysis, including two parts: news viewpoints and netizen viewpoints. Among them, the news viewpoints are the summary of the viewpoints of the media with the highest interaction volume in the news event, including the news content and news link corresponding to the viewpoint; the netizen viewpoints are the summary of the viewpoints of netizens who voice their opinions on the target event, composed of three forms: emotion statistics, viewpoint statistics, and display of viewpoint content. Among them, the emotion statistics are a pie chart statistics of the netizen emotions, the viewpoint statistics show the proportion of the main netizen viewpoints in the form of a percentage bar chart, and the viewpoint content displays the link to the post and the post content corresponding to the main viewpoints.

[0095] In the embodiments of the present invention, the event analysis report is split into several different parts through an event analysis report template. For each part of the report, prompts for the large language model are designed separately, and an AI agent is constructed. According to specific needs, the input of the AI agent is provided, and the content of each part is generated in sequence. Then, each part is concatenated to form an overall event analysis report. By concatenating multiple AI agents, the problem of the length limit of a single large language model is solved. For each AI agent, its implemented function is relatively single. Based on different combinations of input data formats, large models, and prompts, the large model and prompt with the most compliant output data format are selected to construct the AI agent, thus solving the problem of the ability boundary of a single large language model.

[0096] The solution of the embodiments of the present invention combines traditional NLP machine learning algorithms and multiple independent AI agents, giving full play to the advantages of traditional NLP machine learning algorithms and large language models. Specifically, considering the problems of the ability boundary and length limit of a single AI agent, in the embodiments of the present invention, the task of generating an event analysis report is split, reducing the length of the input data of the AI agent and avoiding the length limit of the input of the large language model from the source. Based on large language models with different capabilities, prompt encapsulation and agent combination design are carried out to form multiple types of AI agents. Each type of AI agent processes independent tasks, breaking through the ability boundary of a single AI agent, thereby realizing the generation of more complex and intelligent report content and improving the comprehensiveness and accuracy of the generated content. The AI agents in the embodiments of the present invention can be directly implemented by invoking existing large language models through prompt engineering, or can be implemented by fine-tuning existing large language models and then invoking prompt engineering. Considering that traditional NLP deep learning models often cannot accurately capture the internal relationships between events when processing complex semantics and long texts, resulting in one-sided or illogical report content, therefore, in the embodiments of the present invention, traditional NLP machine learning algorithms are used for data processing such as clustering, community discovery, and sentiment analysis, avoiding the deficiencies of NLP deep learning models in understanding ability.

[0097] Step S4, visually display the content of the event analysis report generated in step S3.

[0098] In some embodiments of the present invention, the event analysis report template includes an event summary, and the generated content is as Figure 4 shown;

[0099] The method for generating the event summary is as follows:

[0100] Based on the event keyword phrases, excluded keyword phrases, and analysis scope, multiple target sub-events are screened from the source database. Each target sub-event includes a sub-event name, occurrence time, and news content, where the news content includes a news title, body content, and link. Specifically, the target sub-events can be obtained from the source database in real time through an API interface.

[0101] Call the event content summary AI agent. Using the sub-event name as a precondition, select the news content with the highest view count in the sub-event as the input, and generate a content summary of a single target sub-event through context semantic analysis. Specifically, based on a large language model, event content summary can be achieved through prompt engineering. The following is an example of prompt design:

[0102] "Please act as an event content summary expert. Based on {{sub-event name}}, analyze in combination with the news context semantics to generate a summary of this event. Ensure that the language logic is smooth, the key points are prominent, and it closely adheres to the core information of the sub-event. The output format is as follows:

[0103] {"Summary": "The generated summary content, with concise and condensed language, highlighting the key points"}

[0104] Note:

[0105] 1. If the news content contains comparative data or significant facts, give priority to including them in the summary content;

[0106] 2. Avoid redundant background information and only extract information core to the sub-event;

[0107] 3. The language style is neutral to ensure suitability for different readers.

[0108] The user input is as follows: ".

[0109] Use the content summaries of multiple target sub-events as the input. With the event name as a precondition, call the event summary text generation AI agent to generate an event summary in combination with context semantic analysis; the event summary includes the time, location, participating entities, description of the event development process, results, or subsequent impacts of the event. Specifically, based on a large language model, the summary text can be generated through prompt engineering. The following is an example of prompt design:

[0110] "Please act as a summary text generation expert. Based on {{event name}} and the provided input of multiple sub-event summaries, analyze the core information and generate a concise summary text. Ensure that the summary language is smooth, the logic is clear, and the key content is prominent. The output format is as follows:

[0111] {"Abstract": "The generated concise abstract is compact and highly condensed, ensuring complete information and highlighting key points. The abstract can extract time, place of occurrence, participating organizations or people, viewpoints or actions of all parties, the sequence of events, and subsequent impacts."}

[0112] Note:

[0113] 1. Prioritize extracting factual information and key data, and ignore lengthy backgrounds;

[0114] 2. Ensure the content is suitable for different readers and avoid using tendentious words;

[0115] 3. Keep the language concise, highlight the main information, and avoid redundant descriptions.

[0116] The user input is as follows: "

[0117] In some embodiments of the present invention, the event analysis report template further includes an event context; the event context includes two presentation forms: text summary and context diagram, and the generated content is as Figure 4 shown.

[0118] The method for generating the event context is as follows:

[0119] Obtain multiple target sub-event information obtained during the generation process of the event summary, including sub-event names, occurrence times, news content, and content summaries; count the popularity of each target sub-event, sort the target sub-events in chronological order, and form structured time-series data. Among them, the content summary is generated by the event content summarization AI agent according to the method above. The popularity of the sub-event is statistically obtained based on the number of news reports of the event.

[0120] Call the event context summarization AI agent, use the event name as a precondition, take the content summaries of multiple target sub-events as input, perform semantic parsing on each target sub-event, and summarize the overall development context of the event in stages to form a text summary of the event context. It is necessary to ensure that the generated content summary of the event context is logically clear, the language is fluent, and it meets the requirements of the target report. Specifically, based on the large language model, the content summary of the event context can be generated through prompt engineering. The following is an example of prompt design:

[0121] "Please act as an event context summarization expert. According to the multiple sub-event context contents corresponding to the provided {{event name}}, analyze its development context and generate a clear summary. Ensure logical clarity and key points are highlighted. The output format is as follows:

[0122] {

[0123] "Overview of the event context": "Summarize the development of the event in stages, list the occurrence time of each stage, describe the process of the event development in each stage, and summarize the position of the event in this stage (such as outbreak, end, etc.).",

[0124] }

[0125] Note:

[0126] 1. Organize the context according to the development order of the event, and clearly present the main nodes;

[0127] 2. Highlight key facts or data, and avoid redundant information;

[0128] 3. Ensure that the language is neutral, concise, and suitable for different readers.

[0129] The user input is as follows: ".

[0130] Taking the target sub - event as a node, with the name and popularity of the target sub - event as node labels, arrange the nodes in chronological order, draw an event context diagram, and display a brief description of each target sub - event in a floating window form.

[0131] In some embodiments of the present invention, the event analysis report template further includes the volume trend, and the generated content is as Figure 5 shown;

[0132] The method for generating the volume trend is as follows:

[0133] Extract the full - volume data of the target event from the source database according to the event keyword phrase, the excluded keyword phrase, and the event analysis time range. Obtain the raw volume data and the target sub - event from the full - volume data. The raw volume data includes the number of posts, the number of interactions, and the number of news reports; the target sub - event includes the sub - event name and the occurrence time. The target sub - event is used for identifying key sub - events and can directly use the results screened out when generating the event context.

[0134] Unify the time format of the raw volume data and perform segmented processing according to the event occurrence date to form multiple date - segmented raw volume data;

[0135] Calculate the volume value for each of the multiple date - segmented raw volume data. The volume value = the number of posts + the number of interactions + the number of news reports;

[0136] Generate a volume trend chart according to the date segmentation and the volume value; where the horizontal axis is time and the vertical axis is the volume value.

[0137] Based on the statistical analysis of the volume of news reports, the target sub-events with the highest daily heat are identified as key sub-events, and the key sub-events are marked as peaks or turning points in the volume trend chart to visually display the volume changes at key nodes during the development of the event.

[0138] In some embodiments of the present invention, the event analysis report template further includes the distribution of event influence, and the generated content is as Figure 6 shown;

[0139] The method for generating the distribution of event influence is as follows:

[0140] According to the event keyword phrases, the excluded keyword phrases, and the time range of event analysis, the full amount of data of the target event is extracted from the source database, and the original influence data is obtained according to the full amount of data. The original influence data includes the number of posts and the volume of news reports, and is classified and summarized according to the communication platform;

[0141] Calculate the platform influence and influence ratio of each communication platform; Platform influence = number of posts + volume of news reports; Influence ratio = influence of this platform / influence of all platforms; The influence ratio of each communication platform is displayed in graphical form, usually in the form of a pie chart.

[0142] Sort the influence ratios of each communication platform to identify the main communication platforms; Call the event platform influence analysis AI agent, with the event name as the precondition, take the main communication platforms, the number of posts / reports, and the influence percentage as inputs, and summarize the communication characteristics of the event through semantic analysis, and output an analysis result with fluent and professional language. Specifically, based on the large language model, the event platform influence analysis result can be generated through prompt engineering. The following is an example of prompt design:

[0143] "Please act as an event platform influence analysis expert, based on the data of the main communication platforms of the provided {{event name}}, identify the main communication platforms and conduct intelligent analysis on their influence. Generate a clear analysis result. The output format is as follows:

[0144] {

[0145] "Analysis conclusion": "Comprehensive analysis of the influence of the main communication platforms, highlighting the event communication characteristics and the role of the platform"

[0146] }

[0147] Note:

[0148] 1. The analysis conclusion should closely follow the data and event communication characteristics, and the language should be concise, concise, and fluent.

[0149] The user input is as follows: ".

[0150] In some embodiments of the present invention, the event analysis report template includes the analysis of the propagation network and propagation nodes; the propagation network visually presents the network structure of the main propagation nodes and information diffusion on all social media platforms; the key nodes list the post view information of key persons / organizations in the propagation network, and generate content such as Figure 7 and Figure 8 as shown.

[0151] The generation method of the above-mentioned analysis of the propagation network and propagation nodes is as follows:

[0152] According to the event keyword phrases, the excluded keyword phrases, and the event analysis time range, extract the original data of the social media platform of the target event from the source database, and obtain the propagation original data according to the original data of the social media platform. The propagation original data includes social accounts, the number of posts, and the number of interactions. The above-mentioned original data of the social media platform can also use the data extracted in the voice volume trend analysis.

[0153] Clean the propagation original data to construct the relationship matrix of the nodes and edges of the propagation network. The purpose of cleaning is to remove invalid nodes and noise data, such as accounts without propagation behavior or interaction behavior, invalid links, advertising content, etc.

[0154] Use the community discovery algorithm to partition the propagation network and identify the main community structure in the propagation process. In practical applications, community discovery algorithms such as Louvain and Girvan - Newman can be selected.

[0155] Rank the nodes within each community according to the propagation metrics to identify the core propagation nodes; the propagation metrics include the number of interactions, and the identification of the core propagation nodes means selecting users or organizations with higher propagation metrics as the core propagation nodes.

[0156] Calculate the key metrics of the propagation network, and further filter the communities based on the key metrics to obtain the important communities for event propagation. The key metrics include but are not limited to degree centrality, betweenness centrality, and clustering coefficient. Among them, degree centrality is used to measure the direct connection number of nodes, indicating the activity of the propagation subject; betweenness centrality is used to evaluate the role of nodes as bridges in the network, reflecting their control over the information propagation path; clustering coefficient is used to describe the tightness of the local network of nodes, reflecting the collaborative characteristics within the community. By setting thresholds for the key metrics, the communities can be further filtered.

[0157] Further screen out the top-ranked key dissemination nodes from the core dissemination nodes based on key indicators, and extract the node information of each key dissemination node. The node information includes the posting content and the interaction volume. The number of key dissemination nodes is at most 20 to ensure the accuracy and effectiveness of the analysis. The node information is screened out based on event keyword phrases and excluded keyword phrases.

[0158] Use the NLP clustering algorithm to cluster all the posting contents corresponding to each screened key dissemination node, so that the clustering results are grouped according to similar contents;

[0159] After clustering, obtain the top M categories according to the interaction volume, and extract the most representative N posts from each category to form a list of posts to be analyzed. In practical applications, the values of M and N can be determined according to the actual situation. For example, both are set to 3, and they can be adjusted to appropriate values during subsequent application.

[0160] Take the event name and event summary as the preconditions for analysis. Input the posts in the list of posts to be analyzed for each key dissemination node into the post view generation AI agent one by one for semantic understanding, generate a clear and logically rigorous view summary, and associate the generated view with the corresponding key dissemination node to form a complete view description for each key dissemination node. Specifically, based on the large language model, the post view can be generated through prompt engineering. The following is an example of prompt design:

[0161] "Please act as an expert in generating post views. Given that the basic information of {{event name}} is {{event summary}}, based on the post information corresponding to the event, analyze the core information and generate a post with a clear and logical view. The view expressed by the poster should be combined with the event background, ensuring that the language is distinct and persuasive, and suitable for social media posting. The output format is as follows:

[0162] {"View": "The generated post view, with distinct, vivid, and logical language, can trigger interaction and discussion"}

[0163] Note:

[0164] 1. The view should closely follow the core content of the event and avoid general statements;

[0165] 2. Facts or data can be combined to strengthen the persuasiveness of the view;

[0166] 3. The language style can be flexibly adjusted. If humor, solemnity, or neutrality is required, please specify;

[0167] 4. Keep the language concise but not lacking in appeal, suitable for dissemination and interaction.

[0168] The user input is as follows: ".

[0169] In some embodiments of the present invention, the event analysis report template includes news opinion analysis, and the generated content is as follows Figure 9 as shown;

[0170] The generation method of the news opinion analysis is as follows:

[0171] According to the event keyword phrases, excluded keyword phrases, and analysis scope, multiple target sub-events are screened out from the source database. Each target sub-event includes a sub-event name, occurrence time, and news content;

[0172] Count the number of news corresponding to each target sub-event, sort them in descending order of the number of news, and select the top K target sub-events as the analysis objects to form the original data set for news opinion generation;

[0173] Taking the event name and event summary as input preconditions, for each target sub-event, extract the first P news texts in order of view count as the input content for the news opinion generation AI agent; the news opinion generation AI agent performs semantic parsing on the news content, identifies the core views, and generates news comments. The value of P can be set according to the actual situation, such as set to 5, or can be continuously adjusted during the application process. Specifically, based on the large language model, news opinions can be generated through prompt engineering. The following is an example of prompt design:

[0174] "Please act as an expert in generating news opinions. Given that the basic information of {{event name}} is {{event summary}}, according to the provided news content, extract the opinion information in the news content that evaluates this event. The opinion should reflect a profound analysis and insight into the event, be persuasive and able to trigger discussions. The output format is as follows:

[0175] {"opinion": "The view and stance of the generated news report on the event, with clear language and rigorous logic"}

[0176] Note:

[0177] 1. The comment should closely follow the core content of the news and avoid general discussions;

[0178] 2. Data, historical background, or similar events can be combined as evidence;

[0179] 3. The language style can be neutral, critical, or humorous, and can be adjusted according to specific needs;

[0180] 4. The comment should be concise and powerful, suitable for dissemination and communication.

[0181] The user input is as follows: ".

[0182] In some embodiments of the present invention, the event analysis report template further includes netizen opinion analysis, and the generated content is as followsFigure 10 and Figure 11 as shown;

[0183] The method for generating the analysis of the views of Internet users is as follows:

[0184] According to the event keyword phrases, the excluded keyword phrases, and the event analysis time range, extract the original post dataset of the target event from the source database; the original post data in the original post dataset of the social media platform is the post content of Internet users under the event-related topics.

[0185] Randomly select L original post contents from the deduplicated original post dataset to obtain a random post dataset; the value of L is set according to the actual situation. For example, it is set to 200. It is necessary to ensure that the sample selection is random and representative, so that the sampling result is close to the normal distribution and reduce the bias caused by factors such as the interaction volume.

[0186] Use the NLP sentiment analysis algorithm to perform sentiment analysis on the original post contents in the random post dataset, classify the emotions of the posts into three categories: positive, negative, and neutral, and display the proportion of each type of emotion through a graph. Specifically, the proportion of the number of posts of each type of emotion can be displayed in the form of a bar chart or a pie chart.

[0187] Group the random post dataset based on the clustering algorithm to form multiple clustering datasets; display the proportion of each clustering dataset through a graph. The clustering dataset is also the main view dataset, and the proportion of the clustering dataset is the proportion of each main view. Specifically, the proportion of each clustering dataset can be displayed in the form of a bar chart or a pie chart.

[0188] Select the Q clustering datasets with the largest number of posts. For each clustering dataset, use all its original post contents as the input content, call the post view generation AI agent to generate the description of the views of Internet users, and select the emotion with the largest proportion in the clustering dataset as the main emotion of the clustering dataset. The number of Q can be set according to the actual situation, such as set to 5, or it can be adjusted continuously during the application process. The clustering results of emotions include three types of emotions: positive, negative, and neutral.

[0189] The above text introduced the generation methods of several modules including the event summary, event context, volume trend, influence distribution, communication network and communication node analysis, news views, and the views of Internet users. In actual application, the specific modules in the event analysis report can be configured according to needs. You can only select a part of the above modules, or add new modules not mentioned in this article according to needs. For the new modules, the corresponding content can also be automatically generated by combining the traditional NLP machine learning algorithm and the independent AI agent.

[0190] In the embodiments of the present invention, after each part of the event analysis report is generated according to a preset analysis report template, the generated event analysis report content can be displayed in a format with both pictures and texts as per the preset format, and a content editing function is provided to support operations such as adding, deleting, and modifying the report content, including adjusting the wording, supplementing information, or deleting irrelevant data. In addition, a distribution function for the event analysis report is provided, and the report can be pushed to specific personnel or teams through a messaging system or the like.

[0191] In the embodiments of the present invention, after step S3, the following steps are further included:

[0192] Regenerate the event analysis report according to the feedback and configuration of the user.

[0193] In addition to directly modifying the generated event analysis report, the embodiments of the present invention also support regenerating the event analysis report according to the feedback of the user. Specifically, the user can put forward feedback on each module in the event analysis report, such as the event summary, event trend, etc., such as insufficient data, inaccurate analysis, etc. The system analyzes the event analysis report according to the feedback of the user, adjusts and optimizes the internal parameters according to the analysis results, such as increasing the data volume of the original news or posts participating in generating the content, and then regenerates the event analysis report to optimize the report quality.

[0194] The above is only the specific implementation manner of the present invention, and the scope of the present invention cannot be limited thereby. Equivalent changes made by those of ordinary skill in the art according to this creation, as well as changes well-known to those skilled in the art, should still fall within the scope covered by the present invention.

Claims

1. A method for generating an event analysis report, characterized in that, The method includes the following steps: Step S1, collect news data of the target news platform and post data of the target social media platform in real time; Step S2, integrate and clean the news data and post data, and use NLP machine learning algorithms to conduct preliminary classification, keyword extraction, entity extraction, and event clustering on the news data, and conduct keyword extraction and entity extraction on the post data to form structured data and store it in the source database; Step S3, according to the input configuration conditions, screen out the target data set related to the target event from the source database; based on the preset event analysis report template, introduce NLP machine learning algorithms as needed to conduct preliminary feature engineering processing on the target data set; call multiple AI agents to generate the content of each part in the event analysis report template; the multiple AI agents are multiple AI agents with independent functions; Step S4, visually display the content of the event analysis report generated in Step S3.

2. The method for generating an event analysis report according to claim 1, wherein The configuration conditions include associated event configuration, and the associated event configuration includes the event name, event keyword phrases, excluded keyword phrases, and analysis scope of the associated event. The analysis scope includes site source, the online and offline times for continuous collection and monitoring of the event.

3. The method for generating an event analysis report according to claim 2, wherein The event analysis report template includes an event summary; The generation method of the event summary is as follows: According to the event keyword phrases, excluded keyword phrases, and analysis scope, screen out multiple target sub-events from the source database. Each target sub-event includes a sub-event name, occurrence time, and news content; Call the event content summary AI agent, use the sub-event name as a precondition, select the news content with the highest view volume in the sub-event as the input, and generate a content summary of a single target sub-event through context semantic analysis; Use the content summaries of multiple target sub-events as the input, use the event name as a precondition, call the event summary text generation AI agent, and generate an event summary through context semantic analysis. The event summary includes the time, place, participating entities, description of the event development process, results, or subsequent impacts of the event.

4. The method for generating an event analysis report according to claim 3, wherein The event analysis report template includes an event context; The generation method of the event context is as follows: Obtain the information of multiple target sub-events obtained during the generation process of the event summary, including sub-event name, occurrence time, news content, and content summary; count the popularity of each target sub-event, sort the target sub-events in chronological order to form structured time series data; Call the event context summary generation AI agent, use the event name as a precondition, use the content summaries of multiple target sub-events as the input, conduct semantic parsing on each target sub-event, and summarize the overall development context of the event in stages; Use the target sub-events as nodes, the names and popularities of the target sub-events as node labels, arrange the nodes in chronological order, draw an event context diagram, and display the content summaries of each target sub-event in a floating window form.

5. The method for generating an event analysis report according to claim 2, wherein, The event analysis report template includes a volume trend; The generation method of the volume trend is as follows: Extract all the data of the target event from the source database according to the event keyword phrases, excluded keyword phrases, and the event analysis time range, and obtain the raw volume data and target sub-events based on the all data. The raw volume data includes the number of posts, the number of interactions, and the number of news reports; the target sub-events include the sub-event name and the occurrence time. Unify the time format of the raw volume data and segment it according to the event occurrence date to form multiple date-segmented raw volume data. Calculate the volume value for each of the multiple date-segmented raw volume data. The volume value = the number of posts + the number of interactions + the number of news reports. Generate a volume trend chart based on the date segmentation and the volume value. Based on the statistical analysis of the number of news reports, identify the target sub-event with the highest daily heat as the key sub-event, and mark the key sub-event as the peak or turning point in the volume trend chart.

6. The method for generating an event analysis report according to claim 2, wherein The event analysis report template includes the event influence distribution. The generation method of the event influence distribution is as follows: Extract all the data of the target event from the source database according to the event keyword phrases, excluded keyword phrases, and the event analysis time range, and obtain the raw influence data based on the all data. The raw influence data includes the number of posts and the number of news reports, and classify and summarize them by the communication platform. Calculate the platform influence and influence proportion of each communication platform. The platform influence = the number of posts + the number of news reports. The influence proportion = the influence of this platform / the influence of all platforms, and display the influence proportion of each communication platform in graphical form. Sort the influence proportions of each communication platform, and identify the main communication platforms; call the event platform influence analysis AI agent, use the event name as the precondition, take the main communication platforms, the number of posts / reports, and the influence percentage as the input, summarize the communication characteristics of the event through semantic analysis, and output the analysis result with fluent and professional language.

7. The method for generating an event analysis report according to claim 3, wherein The event analysis report template includes the analysis of the communication network and communication nodes. The generation method of the analysis of the communication network and communication nodes is as follows: Extract the raw data of the social media platform of the target event from the source database according to the event keyword phrases, excluded keyword phrases, and the event analysis time range, and obtain the raw communication data based on the raw data of the social media platform. The raw communication data includes social accounts, the number of posts, and the number of interactions. Clean the raw communication data and construct the relationship matrix of the nodes and edges of the communication network. Use the community discovery algorithm to partition the communication network and identify the main community structure in the communication process. Rank the nodes within each community according to the communication metrics, and identify the core communication nodes; the communication metrics include the number of interactions. Calculate the key metrics of the communication network, and further filter the communities based on the key metrics to obtain the important communities for event communication. Further screen out the top-ranked key communication nodes from the core communication nodes based on the key metrics, and extract the node information of each key communication node. The node information includes the post content and the number of interactions. Use the NLP clustering algorithm to cluster all the posting contents corresponding to each key dissemination node selected, so that the clustering results are grouped according to similar contents; After the clustering is completed, obtain the top M categories according to the interaction volume, extract the most representative N posts from each category, and form a list of posts to be analyzed. Take the event name and event summary as the preconditions for analysis. Input the posts in the list of posts to be analyzed for each key dissemination node into the post view generation AI agent one by one for semantic understanding, generate a clear and logically rigorous view summary, and associate the generated view with the corresponding key dissemination node to form a complete view description of each key dissemination node.

8. The method for generating an event analysis report according to claim 3, wherein The event analysis report template includes news view analysis; The generation method of the news view analysis is as follows: According to the event keyword phrases, excluding keyword phrases, and analysis scope, screen out multiple target sub-events from the source database. Each target sub-event includes a sub-event name, occurrence time, and news content; Count the number of news corresponding to each target sub-event, sort them in descending order of the number of news, and select the top K target sub-events as the analysis objects to form the original data set for news view generation; Take the event name and event summary as the input preconditions. For each target sub-event, extract the top P news texts in order of page views as the input content of the news view generation AI agent; the news view generation AI agent performs semantic parsing on the news content, identifies the core view, and generates a news comment.

9. The method for generating an event analysis report according to claim 3, wherein The event analysis report template includes netizen view analysis; The generation method of the netizen view analysis is as follows: According to the event keyword phrases, excluding keyword phrases, and the event analysis time range, extract the original post data set of the target event's social media platform from the source database; Randomly select L original post contents from the deduplicated original post data set to obtain a random post data set; Use the NLP sentiment analysis algorithm to perform sentiment analysis on the original post contents in the random post data set, classify the emotions of the posts into three categories: positive, negative, and neutral, and display the proportion of each type of emotion through a graph; Group the random post data set based on the clustering algorithm to form multiple clustering data sets; display the proportion of each clustering data set through a graph; Select the Q clustering data sets with the largest number of posts. For each clustering data set, use all its original post contents as the input content, call the post view generation AI agent to generate a netizen view description, and select the emotion with the largest proportion in this clustering data set as the main emotion of this clustering data set.

10. The method for generating an event analysis report according to claim 1, wherein After step S3, it further includes: Regenerate the event analysis report according to the user's feedback and configuration.

Images