Event report generation method, device and equipment based on large model

Abstract

The invention relates to the field of network public opinions, and discloses an event report generation method, device and equipment based on a large model. According to the method, semantic comprehension and networking retrieval capability of a large model are combined, and multi-modal acquisition, semantic extraction and intelligent generation of event information are realized. The method specifically comprises the steps of obtaining a title and content of a target event, calling a large model to automatically retrieve and summarize related text, picture and video information, and generating an abstract reflecting a core fact of the event; according to the abstract, extracting the place, the subject and the behavior keyword, generating a logic expression to execute targeted retrieval, and obtaining multi-source event information; high-correlation data is screened through semantic similarity calculation to form a precise data set, and an event report with comprehensive content and reliable facts is generated accordingly. Through semantic control and similarity analysis driven by a large model, the accuracy, richness and credibility of event report generation are improved.

Description

Technical Field

[0001] This invention relates to the field of online public opinion, and in particular to a method, apparatus, and device for generating event reports based on a large model. Background Technology

[0002] In the field of public opinion analysis and event assessment, one of the important applications of Large Language Models (LLMs) is to assist in the generation of event reports, situation summaries, and public opinion analysis documents. By aggregating and summarizing information from multiple sources, LLMs can generate structured text content in a short time, thereby significantly improving the efficiency and automation of information assessment.

[0003] However, current event report generation based on large models still has significant shortcomings. On the one hand, large models are limited by the length of the context window during the generation process, making it impossible to process large-scale, multi-source event data simultaneously, resulting in incomplete coverage of factual elements. On the other hand, the raw data on the input side often suffers from problems such as information redundancy, mixed sources, large amounts of repetitive information, and ambiguous expressions, leading to quality defects in the generated content such as "fabricated facts," "biased viewpoints," "incomplete evidence chains," and "weak element dimensions." In addition, reports from different sources differ significantly in terms of authority and factual granularity, and the lack of effective credibility weighting and semantic aggregation mechanisms further reduces the accuracy and interpretability of the generated results.

[0004] Existing technologies mostly employ keyword retrieval, time filtering, and coarse-grained deduplication methods based on title or body text similarity, which cannot simultaneously ensure information richness, semantic deduplication accuracy, and suppression of interfering information. Summary of the Invention

[0005] In view of this, this application provides an event report generation method based on a large model, which solves the technical problems of existing technologies being unable to effectively achieve accurate filtering, semantic deduplication, and interference suppression of event data while ensuring information richness, resulting in reports generated by large models having insufficient descriptive authenticity, comprehensive event perception, and reliability of results.

[0006] According to a first aspect of this application, a method for generating event reports based on a large model is provided, comprising:

[0007] Retrieve the title and content of the target event;

[0008] Based on the title and content of the target event and preset large model prompts, the large model is invoked to obtain the first information of the target event and generate the first summary of the target event, wherein the first information includes text, images and videos;

[0009] According to the first summary output by the large model, extract keywords, wherein the keywords include a place keyword, a subject keyword, and a behavior keyword representing the target event;

[0010] According to the keywords, call a large model to generate a logical expression containing the keywords;

[0011] Based on the logical expression, retrieve at least two pieces of second information related to the target event;

[0012] Calculate the similarity between the first summary and the second information, and filter out a data set with a similarity that meets a preset condition as a precise data set of the target event;

[0013] According to the precise data set of the target event, call a large model to generate an event report of the target event.

[0014] In one possible implementation, the step of calling a large model to obtain first information of the target event based on the title and content of the target event and a preset large model prompt word, and generating a first summary of the target event includes:

[0015] Using a large model with network search capability, taking the title, content, and preset prompt word of the target event as input, automatically retrieving and summarizing text, picture, and video information related to the target event from network public information sources, and generating a first summary reflecting the main facts and transmission profile of the event based on the first information.

[0016] In another possible implementation, the step of calling a large model to generate a logical expression containing the keywords based on the keywords includes using a large model to perform semantic combination on the extracted place keyword, subject keyword, and behavior keyword based on a preset prompt word template, to generate a logical expression with the structure of “(place keyword) AND (subject keyword group) AND (behavior keyword group)”, which is used for targeted event data retrieval in network information sources.

[0017] In another possible implementation, the step of calculating the similarity between the first summary and the second information, and filtering out a data set with a similarity that meets a preset condition as a precise data set of the target event includes:

[0018] Calculate the similarity score between the first summary and the second information using a semantic similarity model;

[0019] Aggregate and merge data with a similarity score exceeding a preset threshold value, and when the number of merged data exceeds a preset upper limit, increase the preset threshold value by a preset step and re-aggregate and merge, until the number of data is not greater than the preset upper limit, to generate an aggregated data set;

[0020] The large model is called to identify and remove interference information containing aggregated news or repeated content in the aggregated data set according to a prompt word, and generate a precise data set of the target event.

[0021] In another possible implementation, the step of calculating the similarity of the first summary and the second information and screening out a data set with a similarity meeting a preset condition as the precise data set of the target event comprises:

[0022] The second information is divided into different sub-data sets according to data source types, wherein data of a publisher type of a central media, a government media or a provincial media constitutes a first data set, and data of other sources constitutes a second data set;

[0023] The semantic similarity model is used to calculate the similarity score between the first summary and the first data set; data with a similarity score exceeding a first preset threshold is aggregated and merged, when the number of merged data exceeds a first preset upper limit, the first preset threshold is increased by a first preset step and the aggregation and merging are performed again until the number of data is not greater than the first preset upper limit, and a first aggregated data set is generated;

[0024] The semantic similarity model is used to calculate the similarity score between the first summary and the second data set; data with a similarity score exceeding a second preset threshold is aggregated and merged, when the number of merged data exceeds a second preset upper limit, the second preset threshold is increased by a second preset step and the aggregation and merging are performed again until the number of data is not greater than the second preset upper limit, and a second aggregated data set is generated;

[0025] After the first aggregated data set and the second aggregated data set are combined, the large model is called to identify and remove interference information containing aggregated news or repeated content in the data set according to a prompt word, and generate a precise data set of the target event.

[0026] In another possible implementation, the step of calculating the similarity score comprises:

[0027] The summary to be calculated and the information to be calculated are respectively input into a sentence vector embedding model based on a BERT architecture to generate corresponding text embedding vectors;

[0028] The semantic similarity between the two is calculated through the embedding vectors, and a cosine similarity is calculated as a similarity score reflecting the degree of semantic association.

[0029] In another possible implementation, the step of calculating the semantic similarity between the two through the embedding vectors and calculating a cosine similarity as a similarity score reflecting the degree of semantic association comprises:

[0030] construct a similarity network based on the similarity scores, wherein nodes represent text data corresponding to the to-be-calculated summary and the to-be-calculated information, edges represent semantic association relationships between texts, and weights of the edges are the corresponding similarity scores;

[0031] reserve edges with similarity scores higher than a preset threshold to filter low-correlation text pairs, and identify a text cluster highly relevant to the to-be-calculated summary by using the similarity network.

[0032] According to a second aspect of the present application, an event report generation device based on a large model is provided, comprising:

[0033] an acquisition module configured to acquire a title and content of a target event;

[0034] a calling module configured to call a large model to obtain first information of the target event based on the title and content of the target event and a preset large model prompt word, and generate a first summary of the target event, wherein the first information includes text, pictures, and videos;

[0035] an extraction module configured to extract keywords from the first summary output by the large model, wherein the keywords include a location keyword, a subject keyword, and a behavior keyword representing the target event;

[0036] a generation module configured to call the large model to generate a logical expression containing the keywords based on the keywords;

[0037] a query module configured to retrieve at least two second information related to the target event based on the logical expression;

[0038] a calculation module configured to calculate a similarity between the first summary and the second information, and filter a data set with a similarity meeting a preset condition as a precise data set of the target event;

[0039] a generation module configured to call the large model to generate an event report of the target event based on the precise data set of the target event.

[0040] According to a third aspect of the present application, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the event report generation method based on a large model when executing the computer program.

[0041] By the technical scheme, the application provides an event report generation method based on a large model, which realizes multi-modal collection, semantic extraction and intelligent generation of event information by combining semantic understanding and network search capability of the large model. The method first acquires a title and content of a target event, calls the large model to automatically search and summarize text, picture and video information related to the event in a network information source, and generates a first summary reflecting core facts of the event. Then, keywords containing location, subject and behavior elements are extracted according to the summary, and the large model is used to generate corresponding logical expressions to perform targeted search and acquire second information from multiple sources. Through semantic similarity calculation, data highly related to the event summary is screened out to form a high-quality, non-redundant and accurate data set, and finally an event report with comprehensive content, clear structure and reliable facts is generated based on the data set. The application introduces semantic control and similarity analysis driven by the large model in the whole process of information acquisition, screening and generation, which significantly improves the accuracy, richness and verifiability of the event report generation.

[0042] The above description is only a summary of the technical scheme of the application. In order to more clearly understand the technical means of the application and can be implemented according to the content of the description, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0043] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The schematic embodiments of the application and their descriptions are used to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0044] Figure 1 An application scenario diagram of an event report generation method based on a large model provided in an embodiment of the application is shown;

[0045] Figure 2 A processing flow diagram of an event report generation method based on a large model provided in an embodiment of the application is shown;

[0046] Figure 3 Another aggregation and merging flow diagram of an accurate data set of a target event provided in an embodiment of the application is shown;

[0047] Figure 4 Another aggregation and merging flow diagram of an accurate data set of a target event provided in an embodiment of the application is shown;

[0048] Figure 5 A structure diagram of an event report generation device based on a large model provided in an embodiment of the application is shown. DETAILED DESCRIPTION

[0049] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0050] The event report generation method based on a large model provided by the embodiment of the present application can be applied to the scene as shown in Figure 1 The application scenario mainly includes four parts of a terminal, a large model, the Internet and an event library. The terminal executes the event report generation method based on a large model, and realizes multi-modal collection, semantic extraction and intelligent generation of event information by combining semantic understanding and network search capability of the large model. The terminal obtains the title and content of the target event from the event library, calls the large model to automatically search and summarize text, picture and video information related to the event in the Internet, and generates a first summary reflecting the core facts of the event. Then, according to the summary, keywords containing place, subject and behavior elements are extracted, and a logical expression is generated by the large model to perform targeted search and obtain second information from multiple sources.

[0051] The terminal further utilizes the semantic similarity analysis function of the large model to perform semantic calculation and aggregation filtering on the obtained multi-source information, filters out accurate data highly related to the event summary, forms a high-quality and non-redundant data set, and generates an event report with clear structure, comprehensive content and reliable facts based on the data set by the large model. Through the scheme, the terminal can realize efficient acquisition and automatic generation of event information in the scene of public opinion monitoring, emergency management and media analysis, and significantly improve the accuracy and intelligent level of event research and judgment.

[0052] The present application will be described in detail below through specific embodiments.

[0053] Embodiment one:

[0054] As shown in Figure 2 , the event report generation method based on a large model provided by the embodiment of the present application includes:

[0055] Step 201, obtaining the title and content of the target event;

[0056] Step 202, based on the title and content of the target event and the preset large model prompt word, calling the large model to obtain the first information of the target event, and generating the first summary of the target event;

[0057] The first information includes text, pictures and videos, wherein the text generation large model mainly processes text. The multi-modal large model can process pictures and videos. In this step, a large model with network search capability is used to automatically search and summarize text, pictures and video information related to the target event from network public information sources based on the title, content and preset prompt words of the target event as input, and generate a first summary reflecting the main facts and transmission profile of the event based on the first information.

[0058] Step 203, extracting keywords according to the first summary output by the large model;

[0059] The keywords include location keywords, subject keywords and behavior keywords representing the target event.

[0060] Step 204, calling a large model to generate a logical expression containing the keywords according to the keywords;

[0061] The extracted location keywords, subject keywords and behavior keywords are semantically combined based on a preset prompt word template to generate a logical expression with the structure of "location keywords) AND (subject keyword group) AND (behavior keyword group)", which is used for targeted event data retrieval in network information sources.

[0062] Step 205, retrieving at least two second information related to the target event based on the logical expression;

[0063] Step 206, calculating the similarity between the first summary and the second information, and screening out a data set with a similarity meeting a preset condition as a precise data set of the target event;

[0064] Step 207, calling a large model to generate an event report of the target event according to the precise data set of the target event.

[0065] Wherein, the accurate data set is input into a large model with text generation and semantic understanding capability (for example, DeepSeek-V3.1-Terminus, DeepSeek-V3.2-Exp, DeepSeek-R1, DeepSeek-V3, Qwen3-Next-80B-A3B-Instruct, Qwen3-30B-A3B-Thinking-2507, MiniMax-M1-80k, ERNIE-4.5-300B), combined with a preset prompt word template (for example, "Please generate a report containing event profile, timeline, impact range, disposal measures and public opinion trend analysis based on the following facts"), to guide the large model to perform semantic induction, logical reorganization and content extraction on the accurate data set. The large model performs fact consistency verification, information integration and hierarchical organization on the input data during the generation process, automatically identifies core elements and generates structured text output. The finally generated event report includes event background, key time nodes, main participants, social impact analysis and subsequent development prediction, etc., ensuring that the report content is real and traceable, logically clear and comprehensive, thereby significantly improving the intelligent degree and reliability of event report generation.

[0066] Embodiment one provides a large model-based event report generation method, which realizes multi-modal collection, semantic extraction and intelligent generation of event information by combining the semantic understanding and network search capabilities of the large model. The method first acquires the title and content of the target event, calls the large model to automatically search and summarize text, picture and video information related to the event in the network information source, and generates a first summary reflecting the core facts of the event. Then, according to the summary, key words containing location, subject and behavior elements are extracted, and the large model is used to generate corresponding logical expressions to perform targeted search and obtain multi-source second information. Through semantic similarity calculation, data highly related to the event summary is selected to form a high-quality, non-redundant accurate data set, and finally an event report with comprehensive content, clear structure and reliable facts is generated based on the data set. The present application introduces large model-driven semantic control and similarity analysis in the whole process of information acquisition, filtering and generation, significantly improving the accuracy, richness and verifiability of event report generation.

[0067] Embodiment two:

[0068] As shown in Figure 3 As a refinement of step 206, the aggregation and merging process of the accurate data set of the target event provided in the embodiments of the present application includes:

[0069] Step 301, calculating the similarity score between the first summary and the second information using a semantic similarity model;

[0070] Step 302: Aggregate and merge data whose similarity scores exceed a preset threshold. When the number of merged data exceeds a preset upper limit, increase the preset threshold by a preset step size and re-aggregate and merge until the number of data is no greater than the preset upper limit, and generate an aggregated dataset.

[0071] Step 303: Call the large model to identify and remove interfering information in the aggregated dataset that contains aggregated news or duplicate content based on the prompt words, and generate an accurate dataset of the target event.

[0072] Through the above steps, this invention achieves semantic association filtering and quality optimization of multi-source event information. By using a semantic similarity model to calculate the similarity score between the first summary and the second piece of information, text related to the core content of the event can be accurately identified at the semantic level. Through dynamic threshold adjustment and aggregation merging mechanisms, the data scale is effectively controlled and redundant information is removed, ensuring the refinement and diversity of the dataset. Furthermore, by leveraging the semantic understanding capabilities of a large model, aggregated news and duplicate content are eliminated, significantly improving the authenticity, relevance, and information density of the target event's accurate dataset.

[0073] Example 3:

[0074] like Figure 4 As shown, as a refinement of step 206, another embodiment of the present invention provides a process for aggregating and merging precise datasets of target events, including:

[0075] Step 401: Divide the second information into different subsets according to the data source type;

[0076] Among them, data from publishers that are central media, official media or provincial media constitute the first dataset, and data from other sources constitute the second dataset.

[0077] Step 402: Calculate the similarity score between the first summary and the first dataset using a semantic similarity model; aggregate and merge data whose similarity scores exceed the first preset threshold. When the number of merged data exceeds the first preset upper limit, increase the first preset threshold by the first preset step size and re-aggregate and merge until the number of data is no greater than the first preset upper limit, and generate the first aggregated dataset.

[0078] Step 403: Calculate the similarity score between the first summary and the second dataset using the semantic similarity model; aggregate and merge data whose similarity scores exceed the second preset threshold. When the number of merged data exceeds the second preset upper limit, increase the second preset threshold by the second preset step size and re-aggregate and merge until the number of data is no greater than the second preset upper limit, and generate the second aggregated dataset.

[0079] Step 404, after merging the first aggregated data set with the second aggregated data set, the large model is called to identify and remove the interference information containing aggregated news or repeated content in the data set according to the prompt word, and a precise data set of the target event is generated.

[0080] Through the above steps, the application realizes multi-source event data refinement and semantic aggregation based on source layering. By dividing the second information into different sub-data sets according to the publisher type, the authoritative source and the general source can be distinguished in the data processing process, thereby ensuring the credibility and coverage of the data; using a semantic similarity model to calculate the similarity score of different levels of data and perform dynamic threshold aggregation, the data size can be effectively controlled while ensuring information diversity, and redundant and noisy content can be removed; finally, through semantic recognition of the large model, aggregated news and repeated content are removed to generate a precise data set with rich information, reliable source and no interference, thereby significantly improving the authenticity, comprehensiveness and analysis value of the event report generation.

[0081] Embodiment four:

[0082] The similarity score calculation step in steps 301, 402 and 403 includes:

[0083] Step S1: text input and vector generation.

[0084] Among them, the target event corresponding to the to-be-calculated summary and the plurality of to-be-calculated information retrieved from the network information source are respectively input into a sentence vector embedding model based on the BERT architecture, such as a SimBERT model. The model encodes the semantic features of the input text through a bidirectional Transformer structure, extracts the deep representation of the text in the semantic space, and outputs the corresponding high-dimensional embedding vector. Each embedding vector represents the feature distribution of the input text in the semantic layer.

[0085] Step S2: semantic similarity calculation.

[0086] Among them, the generated summary embedding vector and each to-be-calculated information embedding vector are matched with each other, and the cosine similarity formula is used to calculate the semantic similarity. The cosine similarity reflects the similarity of the angle between two vectors, and the calculation formula is Where A is the summary embedding vector and B is the embedding vector of the to-be-calculated information. The similarity value obtained by calculation ranges from 0 to 1, and the value closer to 1 indicates a higher semantic similarity. All similarity results are recorded as a similarity score list for subsequent screening analysis.

[0087] Step S3: similarity network construction.

[0088] The similarity network structure is constructed based on the similarity scores, wherein nodes are used to represent the summary text and each to-be-calculated information text, edges are used to represent the semantic association relationship between the texts, and the weight of the edge is the corresponding similarity score. The network is used to depict the semantic connection strength between the summary and each text as a whole.

[0089] Step S4: high-correlation text recognition.

[0090] The similarity network is filtered by an edge threshold, and only edges with a similarity score higher than a preset threshold are retained to remove low-correlation or irrelevant text relationships. Subsequently, a text cluster that is highly semantically related to the summary is identified by using a clustering or connectivity analysis method, and the information in the text cluster is marked as a high-correlation dataset for subsequent accurate data screening and report generation.

[0091] Through the above steps, the present application introduces sentence vector embedding and similarity network structure on the basis of traditional text similarity calculation, and realizes the transition from "local semantic comparison" to "global semantic association recognition". The method of the fourth embodiment of the present application can accurately capture the deep semantic relationship between the event summary and the multi-source information, effectively identify the related content that is semantically similar but different in expression, filter redundant and noise data, and thus significantly improve the accuracy of event data screening and the reliability of report generation.

[0092] Further, as Figures 2 to 4 A specific implementation of the method, the present application provides an event report generation device based on a large model, as shown in Figure 5 The device comprises:

[0093] The acquisition module 510 is configured to acquire the title and content of the target event.

[0094] The calling module 520 is configured to call the large model based on the title and content of the target event and the preset large model prompt word to obtain first information of the target event and generate a first summary of the target event, wherein the first information includes text, pictures, and videos.

[0095] The extraction module 530 is configured to extract keywords from the first summary output by the large model, wherein the keywords include location keywords, subject keywords, and behavior keywords representing the target event.

[0096] The generation module 540 is configured to call the large model to generate a logical expression containing the keywords according to the keywords.

[0097] The query module 550 is configured to retrieve at least two second information related to the target event based on the logical expression.

[0098] The computing module 560 is configured to calculate the similarity between the first summary and the second information, and screen a data set with a similarity meeting a preset condition as a precise data set of the target event.

[0099] The generating module 570 is configured to call a large model to generate an event report of the target event according to the precise data set of the target event.

[0100] In the embodiment of the present application, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the event report generation method based on a large model are implemented, including:

[0101] The title and content of the target event are obtained.

[0102] Based on the title and content of the target event and a preset large model prompt word, a large model is called to obtain first information of the target event, and a first summary of the target event is generated, wherein the first information includes text, pictures and videos.

[0103] According to the first summary output by the large model, a keyword is extracted, wherein the keyword includes a location keyword, a subject keyword and a behavior keyword representing the target event.

[0104] According to the keyword, a large model is called to generate a logical expression containing the keyword.

[0105] At least two pieces of second information related to the target event are retrieved based on the logical expression.

[0106] The similarity between the first summary and the second information is calculated, and a data set with a similarity meeting a preset condition is screened as a precise data set of the target event.

[0107] According to the precise data set of the target event, a large model is called to generate an event report of the target event.

[0108] It should be noted that the principles and implementation steps of the present application are only illustrated by taking public opinion event analysis and report generation as an example in the above embodiments, and the specific application scenarios are not limited. The present application is also applicable to other scenarios that need to identify events, filter semantics, and generate reports based on multi-source information, such as emergency information summarization in government emergency management, news clue aggregation and fact verification in media agencies, enterprise brand or product public opinion monitoring, public security event research and judgment, and case information intelligent induction of judicial or regulatory agencies. By embedding the method of the present application into different types of intelligent analysis systems, it is possible to automatically extract high-value semantic content from complex information, improve event processing efficiency and report generation quality. For the functions or steps that can be implemented by a computer readable storage medium or a computer device, please refer to the description in the foregoing method embodiments, which will not be repeated here.

[0109] Those skilled in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above embodiments. Any reference to memory, storage, database or other medium in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0110] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is taken as an example for illustration. In actual applications, the above functions can be completed by different functional units or modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the above described functions.

[0111] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for generating event reports based on a large model, characterized in that, include: Retrieve the title and content of the target event; Based on the title and content of the target event and preset large model prompts, the large model is invoked to obtain the first information of the target event and generate the first summary of the target event, wherein the first information includes text, images and videos; Based on the first summary output by the large model, keywords are extracted, wherein the keywords include location keywords, subject keywords, and behavioral keywords that characterize the target event; Based on the keywords, the large model is invoked to generate a logical expression containing the keywords; Based on the logical expression, at least two pieces of second information related to the target event are retrieved; Calculate the similarity between the first summary and the second information, and select the datasets whose similarity meets the preset conditions as the accurate datasets of the target event; Based on the precise dataset of the target event, a large model is invoked to generate an event report for the target event.

2. The method for generating event reports based on a large model according to claim 1, characterized in that, The step of generating a first summary of the target event by calling a large model to obtain the first information of the target event based on the title and content of the target event and preset large model prompt words includes: Using a large model with network retrieval capabilities, the title, content, and preset prompts of the target event are used as input to automatically retrieve and summarize text, image, and video information related to the target event from publicly available online information sources, and a first summary reflecting the main facts and dissemination overview of the event is generated based on the first information.

3. The method for generating event reports based on a large model according to claim 1 or 2, characterized in that, The step of calling the large model to generate a logical expression containing the keywords includes using the large model based on a preset prompt word template to semantically combine the extracted location keywords, subject keywords and behavioral keywords to generate a logical expression with the structure "location keyword AND (subject keyword group) AND (behavioral keyword group)", which is used to achieve targeted event data retrieval in network information sources.

4. The method for generating event reports based on a large model according to claim 1, characterized in that, The step of calculating the similarity between the first summary and the second information, and selecting the datasets whose similarity meets preset conditions as the accurate datasets for the target event, includes: The similarity score between the first summary and the second information is calculated using a semantic similarity model; Data with similarity scores exceeding a preset threshold are aggregated and merged. When the number of merged data exceeds a preset upper limit, the preset threshold is increased by a preset step size and the aggregation and merging are repeated until the number of data is no greater than the preset upper limit, thus generating an aggregated dataset. The large model is invoked to identify and remove interfering information containing aggregated news or duplicate content in the aggregated dataset based on prompt words, thereby generating an accurate dataset of the target event.

5. The method for generating event reports based on a large model according to claim 1, characterized in that, The step of calculating the similarity between the first summary and the second information, and selecting the datasets whose similarity meets preset conditions as the accurate datasets for the target event, includes: The second information is divided into different subsets according to the data source type. Data from publishers such as central media, official media, or provincial media constitutes the first dataset, while data from other sources constitutes the second dataset. The similarity score between the first summary and the first dataset is calculated using a semantic similarity model; data with similarity scores exceeding a first preset threshold are aggregated and merged; when the number of merged data exceeds a first preset upper limit, the first preset threshold is increased by a first preset step size and the aggregation and merging are performed again until the number of data is no greater than the first preset upper limit, thereby generating the first aggregated dataset. The similarity score between the first summary and the second dataset is calculated using a semantic similarity model; data with similarity scores exceeding a second preset threshold are aggregated and merged; when the number of merged data exceeds a second preset upper limit, the second preset threshold is increased by a second preset step size and the aggregation and merging are performed again until the number of data is no greater than the second preset upper limit, thus generating a second aggregated dataset; After merging the first aggregated dataset and the second aggregated dataset, the large model is invoked to identify and remove interfering information containing aggregated news or duplicate content in the dataset based on prompt words, thereby generating a precise dataset of the target event.

6. The method for generating event reports based on a large model according to claim 4 or 5, characterized in that, The steps for calculating similarity scores include: The summary to be calculated and the information to be calculated are respectively input into the sentence vector embedding model based on the BERT architecture to generate the corresponding text embedding vectors; The semantic similarity between the two is calculated using the embedded vectors, and the cosine similarity is calculated as a similarity score that reflects the degree of semantic association.

7. The method for generating event reports based on a large model according to claim 6, characterized in that, The step of calculating the semantic similarity between the two through the embedding vector and calculating the cosine similarity as a similarity score reflecting the degree of semantic association includes: A similarity network is constructed based on the similarity scores, where nodes represent text data corresponding to the summary to be calculated and the information to be calculated, edges represent semantic relationships between texts, and the weight of an edge is the corresponding similarity score. Edges with similarity scores higher than a preset threshold are retained to filter low-relevance text pairs, and the similarity network is used to identify text clusters that are semantically highly related to the summary to be calculated.

8. An event report generation device based on a large model, characterized in that, include: The acquisition module is used to acquire the title and content of the target event; The calling module is used to call the large model to obtain the first information of the target event based on the title and content of the target event and the preset large model prompt words, and to generate the first summary of the target event, wherein the first information includes text, images and videos; The extraction module is used to extract keywords based on the first summary output by the large model, wherein the keywords include location keywords, subject keywords, and behavioral keywords that characterize the target event; The generation module is used to call the large model to generate a logical expression containing the keywords based on the keywords; The query module is used to retrieve at least two pieces of second information related to the target event based on the logical expression; The calculation module is used to calculate the similarity between the first summary and the second information, and to select the dataset with similarity that meets the preset conditions as the accurate dataset of the target event; The generation module is used to generate an event report for the target event by calling a large model based on the accurate dataset of the target event.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the step of generating event reports based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of generating event reports based on a large model as described in any one of claims 1 to 7.

Images