A Method and System for Regional and Country Hotspot Mining and Briefing Generation Based on Complex Networks and Large Language Models

Abstract

This invention discloses a method and system for regional and national hotspot mining and briefing generation based on complex networks and large language models, belonging to the field of natural language processing and public opinion analysis technology. The method includes: collecting, cleaning, deduplicating, and sorting multilingual public opinion data; constructing a news dissemination network based on temporal semantic similarity and using connected component algorithms to discover hotspot communities; performing quality assessment and topic consistency optimization on the communities, and selecting representative texts using an improved PageRank algorithm; and automatically generating structured briefings including event summaries, knowledge graphs, news rewriting and commentary, and hotspot extensions by combining preset prompt templates with a generative large language model. This invention can effectively capture the dynamic evolution of multilingual public opinion, improve the accuracy and completeness of hotspot event mining, achieve full automation from data collection to report generation, and significantly improve the efficiency and depth of regional and national public opinion analysis.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for regional and national hotspot mining and briefing generation based on complex networks and large language models. Background Technology

[0002] In today's era of deep globalization and digitalization, the internet has become the core carrier of information dissemination and public opinion fermentation. Timely and accurate mining and analysis of multi-dimensional hot topics in politics, economy, and society across different regions and countries are of paramount importance for national strategic decision-making, corporate risk assessment, academic research, and insights into international relations. With the widespread adoption of social media, news portals, and various new media platforms, massive amounts of heterogeneous and multilingual text data are generated and disseminated rapidly every day, making traditional manual monitoring and analysis methods unsustainable. Automated and intelligent public opinion analysis systems have become an inevitable requirement.

[0003] Currently, the field of automated public opinion hotspot mining and analysis has accumulated numerous technical solutions, with its mainstream technical approach typically following the paradigm of "feature extraction - event clustering - information extraction." Specifically:

[0004] First, in the feature extraction stage, word vector models or pre-trained language models such as Word2Vec, GloVe or BERT are commonly used to convert text into numerical vector representations in order to capture the semantic information of words and text.

[0005] Secondly, in the event clustering and discovery phase, two main types of algorithms are relied upon: First, traditional clustering algorithms, such as K-Means and DBSCAN, group texts based on the distance between their vectors in the feature space, classifying similar texts as the same potential event. Second, network community discovery algorithms, such as Louvain and Infomap. However, these methods are traditionally mainly used to analyze the hyperlink structure (link network) between web pages, dividing communities based on link relationships, reflecting the propagation or citation relationships of information sources.

[0006] Finally, in the stage of summarizing hot topics, topic models (such as LDA) are often used to extract keywords for hot topics, or retrieval models such as BM25 are used to extract key sentences to generate a summary description of the hot content.

[0007] However, the aforementioned existing technical solutions exhibit the following significant limitations when faced with complex and dynamically evolving regional and national multilingual public opinion:

[0008] 1) Difficulty in capturing the dynamic evolution of public opinion: Traditional clustering algorithms (such as K-Means) are usually based on static data snapshots for analysis, and have a weak ability to model the temporal relationship of text. It cannot effectively reveal how a hot topic develops, ferments, transforms or fades over time, that is, it lacks the characterization of the event's life cycle and evolution path, resulting in static and fragmented analysis results with insufficient interpretability.

[0009] 2) Disconnect between community detection and semantic themes: Traditional community detection algorithms based on webpage link structures divide communities based on referencing or navigational relationships between pages, rather than the semantic connections between text content. This results in identified "communities" potentially being merely collections of websites referencing each other, rather than collections of texts discussing the same core event or theme. In other words, this method cannot ensure that texts within the same community semantically belong to the same hot topic, thus limiting its accuracy.

[0010] 3) Incomplete and one-sided information summarization ability: Hot topic summaries generated based on extraction methods (such as keyword and key sentence extraction) are essentially just splicing together fragments from the original text. This method relies heavily on the original text's expression, making it difficult to integrate different perspectives from multiple reports and generate a coherent and complete narrative that transcends the original text. It usually only reflects a partial aspect or scattered facts of an event, failing to summarize the whole picture or systematically present the diverse viewpoints and positions of various stakeholders, resulting in insufficient information integration and depth. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides a method and system for regional and national hotspot mining and report generation based on complex networks and large language models. This method overcomes the deficiencies of existing technologies, effectively captures the dynamic evolution of multilingual public opinion, improves the accuracy and completeness of hotspot event mining, and automates the entire process from data collection to report generation, significantly improving the efficiency and depth of regional and national public opinion analysis.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] A method for regional and national hotspot mining and briefing generation based on complex networks and large language models includes:

[0014] S1: Multilingual public opinion data collection and processing: Collect multilingual text data from multiple information sources in a specified region or country, clean and deduplicate the text data, and sort it according to the publication time;

[0015] S2: Temporal Semantic Hotspot Network Construction and Community Discovery: Based on the text data processed in step S1, a directed network is constructed with text as nodes and semantic similarity between texts under temporal constraints as edge weights; based on the directed network, a network community discovery algorithm is used to identify potential hotspot event communities.

[0016] S3: Hotspot Community Optimization and Representative Content Screening: Conduct quality assessment and theme consistency judgment on the identified hotspot event communities, and segment or filter communities that do not meet the requirements; for each optimized hotspot community, screen out several pieces of text content that can represent the theme of the community;

[0017] S4: Structured Hotspot Briefing Generation: Based on representative text content selected for each hotspot community, combined with preset prompt templates, the generative large language model is called to automatically generate a structured hotspot briefing containing at least one of the following elements: event overview, key entity relationships, content interpretation, and background extension.

[0018] Preferably, the specific process of constructing the temporal semantic hotspot network in step S2 includes:

[0019] S211: Text Vectorization: Convert the title and / or body content of each text into semantic vectors using a multilingual pre-trained model;

[0020] S212: Semantic similarity calculation: Calculate the cosine similarity between the semantic vectors to form an initial semantic similarity matrix;

[0021] S213: Temporal masking application: Construct a temporal masking matrix, where, for the i-th text sorted by time, only its association with the j-th text (j < i) sorted before it is retained; multiply the temporal masking matrix element-wise with the initial semantic similarity matrix to obtain a semantic association matrix with temporal directionality;

[0022] S214: Network Construction: Using text as nodes, if the value of the corresponding element in the semantic association matrix is ​​greater than or equal to a preset similarity threshold, a directed edge is established between the corresponding text nodes to construct the directed network.

[0023] Preferably, in step S2, the network community discovery algorithm is used to identify hot event communities. Specifically, based on the connection relationship between nodes in the directed network, a connected component-based algorithm is used to discover connected subgraphs in the network, and each connected subgraph is initially identified as a hot event community.

[0024] Preferably, the quality assessment and topic consistency judgment of hotspot communities in step S3 include:

[0025] S311: Community Size Filtering: Filter out communities containing fewer than a preset threshold number of texts;

[0026] S312: Topic Distribution Consistency Check: Statistically analyze the topic tag distribution of all texts in the community. If the cumulative proportion of the top N high-frequency topic tags is lower than the preset consistency threshold, the community is determined to be inconsistent in terms of topic.

[0027] Preferably, for communities determined to have inconsistent themes, step S3 further includes:

[0028] S313: Construct a weighted graph within the community: Using texts within the community as nodes, the edge weights between nodes are calculated by weighting the similarity of the topic distribution and the proximity of the publication time between the texts.

[0029] S314: Community Segmentation: Based on the weighted graph, the community is re-divided using a modularity optimization-type community discovery algorithm until the resulting sub-communities meet the topic consistency requirement.

[0030] Preferably, the specific method for selecting representative content in step S3 is as follows:

[0031] S321: Based on the directed network or its derivative network, calculate the importance score of each text node in the community; wherein, the calculation of the importance score comprehensively considers the semantic centrality, topic representativeness and time freshness of the text node;

[0032] S322: Sort the text nodes in the community according to the importance score, and select a preset number of texts with the highest ranking as representative content.

[0033] Preferably, the importance score is calculated using an improved PageRank algorithm, where the transfer weights between nodes are determined by a weighted sum of semantic similarity, topic similarity, and temporal proximity.

[0034] Preferably, the preset prompt template in step S4 includes at least one of the following:

[0035] Hot Topic Guide Template: Used to guide generative large language models to summarize and generate titles and abstracts of hot topics based on multiple representative texts;

[0036] Knowledge graph generation prompt template: used to guide generative large language models to extract entities and relationships from text and form structured knowledge triples;

[0037] News rewriting and brief commentary template: used to guide generative large language models to refine the content and rewrite the language of a single representative text, and generate commentary with a specific perspective;

[0038] Hot Topic Extension Generation Prompt Template: This template guides the generative large language model to combine external knowledge base information to generate in-depth background analysis or trend forecasts for hot topics. In the Hot Topic Extension Generation Prompt Template, the generative large language model must mark the unique identifier of the referenced information in the generated text when referencing external knowledge base information.

[0039] This invention also discloses a regional and national hotspot mining and briefing generation system based on complex networks and large language models, used to implement the above-mentioned method. The system includes:

[0040] The data acquisition and processing module is configured to periodically or as needed collect multilingual raw public opinion data from news websites and government portals of multiple regions or countries specified on the Internet through a preset web crawler program; and to clean, deduplicate, standardize the format, and sort the raw public opinion data in ascending order by publication timestamp, and output a structured time-series text data set.

[0041] The hotspot mining and analysis module is communicatively connected to the data acquisition and processing module, receives the time-series text data set, and is configured as follows:

[0042] (a) Based on the text content in the aforementioned time-series text dataset, construct a news dissemination network that integrates time-series constraints and semantic similarity;

[0043] (b) Execute a complex network community detection algorithm on the news dissemination network to initially identify text communities corresponding to potential hot topics;

[0044] (c) Perform quality assessment, topic consistency verification and optimized segmentation on the initially identified text communities, and merge semantically similar communities across data batches to form the final set of hot events;

[0045] (d) For each final hot topic event, calculate the importance score of each text node in the community, and select a preset number of representative texts based on the scores;

[0046] The briefing intelligent generation module is communicatively connected to the hotspot mining and analysis module, receives the set of hotspot events and their corresponding representative texts, and is configured as follows:

[0047] (e) Integrate at least one generative large language model;

[0048] (f) A variety of structured prompt templates are pre-stored, each template corresponding to a component of the presentation;

[0049] (g) Based on the structural requirements of the target briefing, combine representative texts with corresponding prompt templates to form a standardized prompt input;

[0050] (h) Call the generative large language model to process the prompt input, automatically generate hot topic briefing components that meet the specified style and format requirements, and integrate them into a complete structured hot topic analysis report.

[0051] Preferably, the hotspot mining and analysis module further includes:

[0052] The network construction unit includes a text vectorization subunit, a similarity calculation subunit, a temporal mask application subunit, and a network generation subunit, wherein:

[0053] The text vectorization subunit utilizes a multilingual natural language processing model to convert the title and / or core content of each text into a high-dimensional semantic vector.

[0054] The similarity calculation subunit is used to calculate the cosine similarity between all pairs of text semantic vectors and generate the initial semantic similarity matrix;

[0055] The temporal mask application subunit is used to generate a lower triangular temporal mask matrix based on the publication time order of the text, which is used to mask the semantic association between texts that are published later and texts that are published earlier.

[0056] The network generation subunit is used to perform element-wise multiplication of the initial semantic similarity matrix and the temporal mask matrix to obtain a weighted similarity matrix with temporal directionality; with text as nodes and values ​​in the weighted similarity matrix that exceed a preset threshold as edge weights, a directed or weighted news semantic network graph is constructed.

[0057] The community discovery unit, connected to the network construction unit, is configured to execute an algorithm based on connected component discovery on the news semantic network graph to identify the set of all connected nodes in the graph as an initial hot text community.

[0058] A community optimization unit, connected to the community discovery unit, includes a quality filtering subunit, a topic consistency assessment subunit, a community segmentation subunit, and a cross-block merging subunit; wherein:

[0059] The quality filtering subunit is used to remove initial communities with fewer nodes than the preset minimum community size;

[0060] The topic consistency assessment subunit is used to count the distribution of topic tags in the text within each community. If the sum of the percentages of the top two topic tags is lower than a preset threshold, the community is judged to be topic inconsistent.

[0061] The community segmentation subunit is used to construct an internal subgraph with edge weights of topic similarity and time proximity for communities with inconsistent topics. A dynamic community detection algorithm based on modularity optimization is then used to recursively segment the community until all sub-communities meet the topic consistency requirement.

[0062] The cross-block merging subunit is used to calculate the feature vectors of communities generated by different data batches. If the semantic similarity between the feature vectors exceeds the merging threshold, the communities are merged into the same cross-cycle hotspot event.

[0063] The content filtering unit, connected to the community optimization unit, is configured to perform the following operations for each optimized trending community:

[0064] Importance calculation subunit: Based on the network structure within the community, an improved PageRank algorithm is used to calculate the importance score of each text node, where the weight of the edges between nodes is composed of a linear weighted sum of semantic similarity, topic distribution similarity, and publication time proximity;

[0065] Representative text selection subunit: Based on the calculated importance scores, all text nodes in the community are sorted in descending order, and the top N texts are selected as representative content of the hot topic community and output to the briefing intelligent generation module.

[0066] This invention provides a method and system for regional and national hotspot mining and briefing generation based on complex networks and large language models, which has the following beneficial effects:

[0067] By constructing a temporal mask matrix and combining it with a semantic similarity matrix, the network is forced to retain only temporally forward semantic associations, thus constructing a propagation network with temporal directionality. This allows the system to naturally capture the origin, development, and spread of hot topics, identifying persistent hot topics that evolve over time and are difficult to detect using traditional methods. Furthermore, community identification based on a complex network connectivity component discovery algorithm, which divides nodes according to their actual semantic connections, more accurately reflects the inherent semantic aggregation of texts than clustering algorithms based on preset distance metrics, avoiding the "semantic drift" or "event fragmentation" problems caused by improper parameter settings.

[0068] By using word vector models to vectorize text, the semantics of the same event reported in different languages ​​are mapped to a unified vector space. In the community optimization stage, edge weights are defined by combining semantic similarity, topic distribution similarity, and time proximity, which realizes a multi-dimensional and refined measurement of news relevance, making the final event community more cohesive and the theme purer.

[0069] By employing a strategy of "block processing, local discovery, and global merging," the system can efficiently handle massive, streaming public opinion data with controllable resource consumption and support for incremental updates. The entire hotspot mining process is based on a graph network model, with clear mathematical definitions and visual representations of the formation, merging, and segmentation of communities, offering strong interpretability and facilitating manual review and algorithm optimization. From data collection and cleaning to hotspot mining and report generation, the entire process is highly automated, significantly reducing manual intervention and substantially improving the efficiency and scale of regional and national public opinion monitoring and analysis. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of this invention or the prior art will be briefly introduced below.

[0071] Figure 1 This is a flowchart illustrating the overall steps of the present invention;

[0072] Figure 2 This is a schematic diagram of the multilingual public opinion data collection and processing flow in this invention;

[0073] Figure 3 This is a schematic diagram of the hotspot mining process in this invention;

[0074] Figure 4 This is a detailed flowchart of the structured hotspot briefing generation process in this invention;

[0075] Figure 5 This is an example diagram of semantic similarity matrix calculation in this invention;

[0076] Figure 6 This is an example diagram of semantic association matrix calculation in this invention;

[0077] Figure 7 This is an example diagram of news hotspot community mining in this invention;

[0078] Figure 8 This is an example diagram of cross-block merging communities in this invention;

[0079] Figure 9 Example diagrams for generating hotspot guides in this invention;

[0080] Figure 10 This is an example diagram of knowledge graph generation in this invention;

[0081] Figure 11 This is an example diagram of news rewriting and brief commentary generation in this invention;

[0082] Figure 12 Example diagrams are generated to illustrate the hotspot extension in this invention;

[0083] Figure 13 This is a schematic diagram of the system architecture of the regional and national hotspot mining and briefing generation system of the present invention;

[0084] Figure 14 This is a diagram showing the internal structure of the hotspot mining and analysis module in this invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0086] Example 1, as Figure 1-12 As shown, this invention discloses a method for regional and national hotspot mining and briefing generation based on complex networks and large language models, including:

[0087] S1: Multilingual Public Opinion Data Collection and Processing: Collecting multilingual text data from multiple information sources in a specified region or country, cleaning and deduplicating the text data, and sorting it according to the publication time; specifically, including:

[0088] S101: Multilingual Public Opinion Data Acquisition; This invention integrates information sources such as official news websites, mainstream media platforms, and government portals from multiple regions and countries to construct a configurable multilingual public opinion data acquisition system. Users first select the target geographical area (which can be refined to the country or region level) and the monitoring time range. Based on the user configuration, the system launches a distributed web crawler, automatically capturing web page content from the selected information sources according to a preset collection frequency. The crawler system uses intelligent parsing technology to extract the following structured fields from the web pages: article title, main text content, publication time, source website name, unique article identifier (URL or hash value), and the country or region to which the article belongs (as shown in Table 1 below). Simultaneously, the system calls a pre-trained multilingual topic classification model to automatically label each news text, categorizing it into a preset topic category system, such as politics, economy, science and technology, society, military, culture, environment, energy, healthcare, education, sports, and transportation.

[0089] Table 1. Field Information for Public Opinion Data

[0090] field name type describe category String The subject categories include politics, culture, economics, energy and environment, finance, healthcare, industry, society, sports, science and technology, transportation, and education. title String title siteName String Website or source name publishTime Number Release time pageId String Unique Identifier content String text href String Link country String News site country or region

[0091] S102: Multilingual public opinion data processing; the collected raw data needs to undergo standardized preprocessing before it can be used for subsequent analysis. First, data cleaning is performed to remove irrelevant information such as HTML tags and advertising content. Then, a hash algorithm based on "title + site name" is used for deduplication, effectively reducing data redundancy caused by the repeated publication of the same news text on different channels. Finally, all data is globally sorted in ascending order by publication time to form a time-series text data stream, providing a foundation for the subsequent construction of a time-series semantic network.

[0092] S2: Temporal Semantic Hotspot Network Construction and Community Discovery: Based on the text data processed in step S1, a directed network is constructed with text as nodes and semantic similarity between texts under temporal constraints as edge weights. Based on the directed network, a network community discovery algorithm is used to identify potential hotspot event communities. Specifically, this includes:

[0093] S201: Segmented Processing of Public Opinion Data

[0094] Considering the computational resource consumption of large-scale data processing, the system adopts a block processing strategy. All sorted public opinion data is divided into multiple consecutive data blocks according to a preset batch size, with articles within each data block maintaining chronological order. The system independently executes the hotspot discovery process for each data block, and finally merges hotspots across data blocks.

[0095] S202: Text Vectorization Representation

[0096] For each data block containing n news articles, they first need to be converted into machine-understandable semantic vectors. This embodiment selects the title as the core semantic carrier. WordPiece and other word segmentation algorithms are used to segment multilingual titles. To adapt to multilingual scenarios, the paraphrase multilingual-MiniLM-L12-v2 word vector model is chosen, which supports over 100 languages. The segmented title is input into this model to obtain the vector representation of each word element. For text T, after word segmentation, a sequence of word elements is obtained. The vector representation of each word element is obtained through a pre-trained model. Finally, the text vector V is obtained by averaging these word vectors.

[0097]

[0098] in, It is the title text vector. It is the first Word vectors, This is the number of tokens in the title text after word segmentation. This vector can effectively capture the deep semantic information of the title.

[0099] S203: Constructing a temporally constrained semantic similarity network

[0100] To calculate the semantic similarity between title text vectors, all text title vectors within the data block are concatenated into a matrix. , matrix Transpose to obtain the transpose matrix Perform matrix multiplication on two matrices to obtain the dot product of the vectors, and then calculate the matrix. The L2 norm (row norm) of each vector forms a column vector. Transpose the norm column vector to obtain The two norm vectors are multiplied using a matrix multiplication operation to obtain the norm product of the vectors. Finally, the dot product of the vectors is divided by the norm product to obtain the adjacency matrix of the title text based on semantic similarity. .like Figure 5 As shown, the text title vectors 1, 2, 3, and 4 are concatenated into a matrix. ,matrix and transpose matrix Perform matrix multiplication to obtain a symmetric semantic similarity matrix.

[0101]

[0102]

[0103] in, It is a text title vector matrix. It is the norm of the text title vector. It is a semantic similarity matrix, and this formula calculates the cosine similarity matrix between the row vectors of matrix A.

[0104] S204: Constructing the temporal mask matrix

[0105] Construct a lower triangular time sequence mask matrix based on the publication time order of the text. Similarity matrix with title text Perform element-wise multiplication to obtain the updated semantic association matrix. This ensures that each piece of public opinion data focuses only on the public opinion data preceding it. Since the public opinion data is sorted in ascending order of publication time, this mask matrix... Defined as: for the i-th text after sorting by time, (i.e., the publication time of the j-th text is earlier than that of the i-th text); otherwise The purpose of the mask matrix is ​​to ensure that for each piece of public opinion data after sorting, only public opinion data published earlier than that data are considered "time neighbors" and become candidate similar texts. For example... Figure 6 As shown, based on Construct the semantic similarity matrix. The mask matrix is ​​used, where the black part represents a value of 1 and the gray part represents a value of 0. Node 2 only focuses on node 1, node 3 only focuses on nodes 1 and 2, and node 4 only focuses on nodes 1, 2, and 3. The symmetric semantic similarity matrix and the current mask matrix are then multiplied element-wise (Hadamard product) to obtain the updated semantic association matrix with temporal directionality. .

[0106]

[0107]

[0108] in, It is a time-based mask matrix. It is a semantic similarity matrix. It is the updated semantic association matrix.

[0109] After this operation, the matrix becomes an asymmetric matrix, which means that for the i-th news item, only its semantic association with the j-th (j < i) news item published earlier is retained. This is more in line with the temporal logic of information dissemination in the real world.

[0110] S205: Hotspot Community Detection Based on Connected Components

[0111] Based on the updated semantic association matrix Building a news dissemination network Among them, the set of network nodes This corresponds to n news articles. The network edge set E is determined by a similarity threshold. In this embodiment, based on multiple experiments (calculating the cosine similarity of different news reports on the same event), the similarity threshold is set to 0.7. If the similarity between any two public opinion data nodes is greater than or equal to 0.7, it is recorded as an edge. ,Right now:

[0112]

[0113] In this network, for connected component discovery, the edge directions introduced by the temporal mask are ignored, and the network is treated as an undirected graph. Then, a connected component discovery algorithm (such as Depth-First Search (DFS) or Union-Find) is used. This algorithm can find all maximally connected subgraphs in the network, where nodes within each subgraph are directly or indirectly connected by edges, but not connected to nodes outside the subgraph. By finding naturally connected groups of nodes, these groups of nodes represent the initially identified hotspot communities. Reachability between nodes refers to the existence of a shortest path (set of edges) that connects one node directly or indirectly to another.

[0114] Hotspot Community It is a connected component that satisfies:

[0115] (a) Hot Topics Any two nodes within the same node have a reachability relationship, that is: ;

[0116] (b) Hotspot Communities Any node and community within There is no reachability relationship between any node outside the range, that is: .

[0117]

[0118]

[0119] in, It is a news dissemination network. It is a set of network nodes. It is a set of network edges. It is the first The node reaches the first The shortest path length of each node. It is the number of edges in the shortest reachable path. It is the first The node reaches the first The reachability of each node, if the shortest reachable path If the value is greater than 1, it is reachable; otherwise, it is not reachable.

[0120] For example, as shown in Table 2 below, six Indonesian media news articles are selected, and a time-series-based semantic similarity matrix is ​​calculated to construct a news dissemination network. Figure 7 As shown, based on network reachability, news nodes 1, 2, 3, and 4 are classified as hotspot communities. The semantic similarity between nodes 1 and 2, 1 and 3, 2 and 3, 2 and 4, and 3 and 4 is greater than 0.7, indicating direct reachability between these nodes (path length 1). News nodes 1 and 4 have a semantic similarity less than 0.7, meaning they are not directly reachable, but are indirectly reachable through nodes 2 or 3 (path length 2). Nodes 5 and 6 have no reachable paths to any other node in this community, and there are no reachable paths between any two nodes; therefore, they are considered isolated nodes. From the perspective of news content, the hotspot community formed by news nodes 1, 2, 3, and 4 is related to the "Indonesian AI boom," while nodes 5 and 6 are news about Indonesian handicrafts and Indonesian criminal cases, respectively, and are unrelated to the hotspot community.

[0121] Table 2 News Examples

[0122] Serial Number News Headline 1 Lewat Program Presight Accelerator, Startup AI Asal Indonesia Masuk 10 Startup AITerbaik Dunia (Through the Presight Accelerator Program, an Indonesian AI startup has entered the ranks of the world's top ten AI startups) 2 BRIN pacu kemajuan AI Indonesia lewat ajang InaRI Expo (BRIN accelerates the development of artificial intelligence in Indonesia through InaRI Expo) 3 Cross-sectoral collaboration is key to accelerating the application of artificial intelligence in Indonesia. 4 World AI Show ke-45 di Jakarta: Indonesian Chamber of Commerce and Industry Emphasizes the Importance of Digital Transformation Through Artificial Intelligence (The 45th World Artificial Intelligence Exhibition was held in Jakarta: The Indonesian Chamber of Commerce and Industry stressed the importance of digital transformation through artificial intelligence) 5 Authentic West Javanese handicrafts will be exhibited in Bandung this July. 6 Top 3 News: Notary Public Murder in Bekasi, Police Reveal Victim's Driver Suspected of Involvement

[0123] S206: Cross-block merging news hotspots community

[0124] Related hotspots may exist in different data blocks. Since some hotspots have long durations and different stages of development are distributed across different data blocks, it is necessary to merge the initial communities generated from all data blocks. For each community, its "community vector" is calculated: the sum and mean of the headline text vectors of the top three news items with the highest PageRank scores within that community are used as the community vector. Then, the cosine similarity between community vectors from different data blocks is calculated. If the similarity is greater than or equal to a merging threshold (e.g., 0.7), the two communities are considered to belong to the same hotspot event and are merged.

[0125] S3: Hotspot Community Optimization and Representative Content Selection: Quality assessment and thematic consistency judgment are performed on the identified hotspot event communities; communities that do not meet the requirements are segmented or filtered. For each optimized hotspot community, several textual pieces representing the community's theme are selected; specifically including:

[0126] S301: Quality Assessment and Thematic Consistency Judgment of News Hotspot Communities

[0127] Not all identified communities are high-quality hotspots. This embodiment implements a two-step quality assessment:

[0128] (1) Community size filtering: Filter out communities with fewer than a preset threshold (e.g., 3) of text content. If the size is too small, it will not be able to form an effective hotspot, thus avoiding noise interference.

[0129] (2) Topic Consistency Check: Semantic similarity can easily lead to the same person participating in different events being considered semantically similar, resulting in a mix of trending topics. This is addressed by statistically analyzing the distribution of topic tags across all news items within the community. The cumulative percentage of the top N (usually N=2) high-frequency topic tags is calculated. If this percentage is lower than a preset threshold (70% in this embodiment after evaluation), the community is considered to have mixed topics, potentially including multiple themes, and needs to be segmented. If the percentage of the top two topic tags exceeds the preset threshold, the community is considered to have concentrated topics and belong to a high-quality trending community.

[0130] S302: Secondary Segmentation of News Hotspot Community

[0131] For communities with inconsistent themes, they need to be segmented. The implementation process is as follows:

[0132] (1) Construct a weighted graph based on topic and time similarity: For communities with inconsistent topics, calculate the edge weight of any adjacent node within the community. Edge weights are composed of the topic similarity and time proximity between nodes. Node topic similarity is represented by the JS divergence of the probability distribution of topic tags, ranging from 0 to 1; the closer to 1, the more similar the topics. Node time proximity refers to the difference in publication time between nodes, after normalization, with a value ranging from 0 to 1; the closer to 1, the closer the times. Specifically:

[0133]

[0134]

[0135]

[0136]

[0137] in, It is the first The node and the first The edge weights of each node are composed of semantic similarity, topic distribution similarity, and temporal proximity. The weight for topic similarity is set to 0.8. The time proximity is set to 0.2. It is the first The probability of topic distribution at each node and the The probability of topic distribution at each node JS divergence, It refers to the number of hashtags. It is the first The node The probability of each topic, It is the first The node The probability of each topic It is a node and nodes The The average probability of each topic It is the first The node and the first The time proximity of the nodes Yes, yes, the first The node and the first Time difference between nodes It is the minimum time difference. That is the maximum time difference.

[0138] (2) Modularity-based segmentation: On the constructed weighted graph, a modularity-based community detection algorithm (such as the Louvain algorithm) is used to recursively segment the community. The algorithm continuously moves nodes to maximize the modularity Q value until convergence, thereby dividing the original community into multiple sub-communities with the same theme.

[0139] Specifically, each node is viewed as an independent community, and each node... Place it in the community of each of its neighboring nodes and calculate the gain of modularity. , will node Move to enable The community with the largest gain. Iterate through all nodes until no node movement can improve modularity. Aggregate the moved nodes into new nodes; the edge weights between the new nodes are equal to the sum of the weights of all edges between the corresponding communities in the original network. Use the resulting new network as input to begin a new round of local optimization until the modularity score converges, resulting in a new community partition.

[0140] The formula for calculating the modularity Q is:

[0141]

[0142]

[0143]

[0144]

[0145]

[0146] in, Refers to nodes and nodes The weight of the edges between them. It is a node The sum of the weights of all adjacent edges, It is the sum of the weights of all edges in the network. It is the Kronecker function, if the node and nodes Within the same community, the value is 1; otherwise, it is 0.

[0147] S303: Filtering typical news from trending communities

[0148] To select the most representative news from the community, this embodiment uses an improved PageRank algorithm to calculate node importance. The higher a node's PageRank score, the more important the node and the more representative the news.

[0149] First, construct a weighted directed graph within the community, with edge weights. It is determined by a combination of three dimensions: semantic similarity, topic distribution similarity, and temporal proximity.

[0150]

[0151] in, The weight for semantic similarity is set to 0.4. The weight for topic similarity is set to 0.4. The time proximity is set to 0.2. It is the first The node and the first The semantic similarity of each node. It is the first The probability of topic distribution at each node and the The probability of topic distribution at each node JS divergence, It is the first The node and the first The time proximity of the nodes.

[0152] Then, based on this weight matrix, the PageRank formula is run to calculate the inbound weights of each node and update the score of each node until the PageRank scores converge, at which point the iteration stops.

[0153]

[0154] in, It is the first The PageRank score of each node. It is the damping coefficient, a constant between 0 and 1, usually with a value of 0.85.

[0155] After the algorithm converges, the nodes are sorted in descending order of their PageRank scores, and the top three news items are selected as representative content of the hot topic community.

[0156] S4: Structured Hotspot Briefing Generation: Based on representative text content selected for each hotspot community, and combined with preset prompt templates, a generative large language model is used to automatically generate a structured hotspot briefing containing at least one element from the following: event overview, key entity relationships, content interpretation, and background extension. Specifically, this includes:

[0157] S401: Generate trending titles and introductions

[0158] The top 3 news items by PageRank score for each community are selected. Each item includes a news ID, publication time, title, body text, website name, and topic category. This news information is then used to construct a news context. This context is then combined with a predefined "hotspot lead-in template" and the large language model API is invoked. The template explicitly requires the model to generate a title and lead-in field in a formal political and economic writing style. The model output is parsed into a specified JSON format, containing the `hotspot_title` and `hotspot_lead_in` fields.

[0159] For example, such as Figure 9 As shown, the top three news items with the highest PageRank scores from the news hotspot community are selected: News 1: Lewat Program Presight Accelerator, Startup AI Asal... (Through the Presight Accelerator program, an AI startup from...), News 2: BRIN pacu kemajuan AIIndonesia lewat ajang InaRI Expo... (BRIN promotes the development of artificial intelligence in Indonesia through the InaRI Expo...), and News 3: World AI Show ke-45 di Jakarta: Kadin Tegaskan PentingnyaTransformasi Digital Lewat AI... (The 45th World AI Show in Jakarta: The Indonesian Chamber of Commerce and Industry emphasizes the importance of achieving digital transformation through artificial intelligence...). The titles and text of these three news items, along with the hotspot guide template, are input into the large language model, which outputs the hotspot title: "Indonesia Achieves Significant Progress in Artificial Intelligence, Demonstrating Global Competitiveness" and the hotspot guide: "Recently, Indonesia has achieved several breakthroughs in the field of artificial intelligence (AI), demonstrating the country's global competitiveness in the digital economy era...".

[0160] S402: Generate a Hotspot Knowledge Graph

[0161] Construct a context for trending information, including trending titles, the top 3 news headlines by PageRank score, and their body text. Create a knowledge graph to generate a prompt template: by parsing the trending titles and news content, extract core entities and relationships, and construct a hierarchical network structure. The central nodes are core entities such as people or organizations, while non-central nodes are news details. The network hierarchy should not exceed three levels to maintain simplicity. The template uses predefined semantic tags (such as trending events, countries, people, etc.) and concise Chinese verbs to define relationships, ensuring the accuracy and consistency of entities and relationships. The output conforms to the specified JSON format specification, with fields including head entity, tail entity, type, and a list of relationships. Submit the input trending information context and the knowledge graph-generated prompt template to a large language model. The prompt template requires the model to extract core entities (such as countries, organizations, people, events) and their relationships (described using concise Chinese verbs such as "layout," "promote," "participate") from the text, constructing a hierarchical knowledge network. The model outputs a series of triples (head entity, relationship, tail entity) in JSON format. The system then calls a drawing interface (such as Graphviz or ECharts) to visualize these triples as a knowledge graph, intuitively showing the sequence of events.

[0162] For example, such as Figure 10 As shown, the three news items on the hot topic "Indonesia's AI wave" and the knowledge graph prompt template are input into the large model, which outputs a triple of the knowledge graph: {'head':'Indonesia', 'tail'; 'AI strategy', 'relationship'; 'layout'}... The drawing interface is then called to convert the triple into a knowledge graph.

[0163] S403: News Rewriting and Brief Commentary Generation

[0164] For each selected representative news article, construct a news context including the title, publication time, website name, and body content. Create a news rewriting and commentary template: based on the news article context and the current date, rewrite the news title, generate a 60-70 word news summary, rewrite the body content, and generate a news commentary. The template should adopt a formal political and economic writing style, use standard Chinese terminology, highlight the strategic significance and policy reference value for China, and output a JSON format conforming to a specified format, with fields including the rewritten news title, news summary, news content, and news commentary. Combine the news context with the news rewriting and commentary template. The template guides the large language model to complete the following tasks: 1) Rewrite the original title to conform to Chinese media conventions (e.g., "Indian media:..."); 2) Generate a 60-70 word news summary, noting the original publication time; 3) Refine, reorganize, and localize the news body, extracting key points and adding Chinese subheadings; 4) Generate a "news commentary" focusing on analyzing the strategic significance, policy reference value, or potential impact of the event on China. Generate the rewritten title, abstract, structured text, and perspective briefs.

[0165] For example, such as Figure 11 As shown, the headline and body of News 1: Lewat Program Presight Accelerator, StartupAI Asal... (Artificial intelligence startups from... through the Presight Accelerator program) and news rewriting prompt template are input into the large model, and the output is the rewritten headline: "Indonesian media: Indonesian AI education startup AJARI selected as one of the top ten global AI startups", news summary: "On July 8, 2025, Indonesian media MediaIndonesia reported that Indonesian AI education technology company AJARI was successfully selected as one of the top ten global startups in the first Presight AI-StartupAccelerator program...", news content: "Innovation highlights: AI empowers education... Policy support: A model of Indonesia's digital transformation... International cooperation: Accelerating the development of the AI ​​ecosystem...", news commentary: "AJARI's success not only demonstrates Indonesia's innovation capabilities in the field of AI, but also provides a reference for Chinese AI companies in international cooperation and market expansion...".

[0166] S404: Generate Hotspot Extension

[0167] To supplement the background depth, the system performs the following steps: First, using the hot topic title as the query item, it searches the internet and knowledge bases, internal knowledge bases, or authorized authoritative external network information to obtain relevant background information (such as historical data, expert opinions, and statistical reports), and assigns a unique identifier to each piece of content. Then, the hot topic title, introduction, and this background information together constitute the input context. A specially designed hot topic extension generation prompt template is used, which requires writing a hot topic extension of no more than 200 words based on the hot topic context and background information, adopting a formal political and economic writing style, imitating the Xinhua News Agency's external propaganda style, emphasizing policy, authority, and strategic perspective, requiring objective expression and the use of standardized citation methods, and requiring the content ID to be marked when citing background information, with the output conforming to the specified JSON pattern, and the fields including the hot topic extension.

[0168] The input context is combined with the hot topic extension prompt template, and the large language model is called to output the hot topic extension text that conforms to the standards, ensuring that the content is authoritative and objective, the citation is standardized and meets the requirements of Chinese expression.

[0169] For example, such as Figure 12 As shown, the hot topic title "Indonesia Makes Significant Progress in Artificial Intelligence, Demonstrating Global Competitiveness" and the hot topic summary "Recently, Indonesia has made several breakthroughs in the field of artificial intelligence (AI), demonstrating the country's global competitiveness in the digital economy era..." are combined to form the hot topic information context. The hot topic title is used to search the network and local knowledge base to obtain background information context, such as the challenges Indonesia faces in developing AI. The hot topic information context, background information context, and hot topic extension are used to generate a prompt template. The template is then input into a large model, and the hot topic extension is output: "AI will contribute 2.3% to 3.1% to the GDP of ASEAN countries, with Indonesia benefiting the most, highlighting its market potential... However, Indonesia's AI development still faces challenges such as insufficient infrastructure and talent shortage. Its AI ecosystem maturity still lags behind neighboring countries such as Malaysia. Priority needs to be given to solving the shortcomings in chip infrastructure and cultivating local AI talent through education system reform [Content ID: 1]...".

[0170] Through the above steps, the system will eventually integrate the automatically generated hot topic titles, introductions, knowledge graphs, rewritten and brief comments on three news articles, and hot topic extensions into a standardized, comprehensive, and professional regional and national hot topic analysis report according to the predetermined report template.

[0171] This invention constructs a temporal mask matrix and combines it with a semantic similarity matrix to force the network to retain only temporally forward semantic associations, thereby building a propagation network with temporal directionality. This design enables the system to naturally capture the origin, development, and spread of hot topics, identifying persistent hot topics that evolve over time and are difficult to detect using traditional methods. Simultaneously, community identification based on a complex network connected component discovery algorithm, which divides nodes according to their actual semantic connections, more accurately reflects the inherent semantic aggregation of texts than clustering algorithms based on preset distance metrics, avoiding the problems of "semantic drift" or "event fragmentation" caused by improper parameter settings.

[0172] By using word vector models to vectorize text, the semantics of the same event reported in different languages ​​are mapped to a unified vector space. In the community optimization stage, edge weights are defined by combining semantic similarity, topic distribution similarity (JS divergence), and time proximity, which realizes a multi-dimensional and refined measurement of news relevance, making the final event community more cohesive and the theme purer.

[0173] This invention employs a strategy of "block processing, local discovery, and global merging," enabling efficient processing of massive, streaming public opinion data with controllable system resource consumption and support for incremental updates. The entire hotspot mining process is based on a graph network model; the formation, merging, and segmentation of communities have clear mathematical definitions and visual representations, offering strong interpretability and facilitating manual review and algorithm optimization. From data collection, cleaning, and hotspot mining to report generation, the entire process is highly automated, significantly reducing manual intervention and substantially improving the efficiency and scale of regional and national public opinion monitoring and analysis.

[0174] Example 2: To make the objectives, technical solutions, and advantages of the present invention clearer, this example will provide a detailed description of a regional and national hotspot mining and briefing generation system based on complex networks and large language models that implements the above method.

[0175] See Figure 13 and Figure 14 This invention provides a regional and national hotspot mining and briefing generation system based on complex networks and large language models, including a data acquisition and processing module 100, a hotspot mining and analysis module 200, and a briefing intelligent generation module 300. The modules exchange and communicate data through defined application programming interfaces (APIs) or message queues, collaboratively completing a fully automated process from raw data to structured reports. Wherein:

[0176] The data acquisition and processing module, as the system's data input terminal, is responsible for acquiring and initially organizing raw public opinion information. Specifically, it includes:

[0177] Configuration Management Unit 101: Used to maintain a list of news websites and government portal URLs for the target country / region, as well as parameters such as collection task cycle and proxy settings.

[0178] Web crawler engine 102: Depending on the configuration, it starts periodically or on demand, concurrently accessing target websites and downloading the original webpage content. The engine has capabilities such as deduplication, compliance with the robots.txt protocol, and simulating browser access.

[0179] Content parsing and extraction unit 103: Performs structured parsing of downloaded web pages, and uses XPath, CSS selectors or machine learning models to accurately extract metadata such as article title, body text, publication time, author, source site, etc., and generates a unique pageId.

[0180] Data Cleaning and Standardization Unit 104: Performs operations such as deduplication (based on the hash value of "title + site name"), encoding conversion, removal of irrelevant advertisements and navigation text, and timestamp formatting.

[0181] Topic Classification Unit 105: Integrates a multilingual topic classification model (such as a classifier finely tuned based on XLM-RoBERTa) to automatically classify each cleaned news article and label it with topic tags such as "politics", "economics", and "technology".

[0182] Time-series sorting and output unit 106: Sorts all processed data entries strictly in ascending order according to the publishTime field, encapsulates them into structured JSON or database records, outputs them as a time-series text data set, and passes them to the hotspot mining and analysis module 200 through message bus or API call.

[0183] The hotspot mining and analysis module 200, as the core analysis engine of the system, communicates with the data acquisition and processing module, receives data streams from the data acquisition and processing module, and performs complex network construction and community discovery tasks. Internally, it includes a network construction unit 210, a community discovery unit 220, a community optimization unit 230, and a content filtering unit 240; such as... Figure 14 As shown, where,

[0184] Network building block 210 includes:

[0185] Text Vectorization Subunit 211: Integrates pre-trained word vector models (such as paraphrase multilingualMiniLM L12-v2). This subunit receives text, performs word segmentation and model forward propagation, and outputs a high-dimensional semantic vector for each news headline (or headline and summary).

[0186] Similarity calculation subunit 212: It is used to receive a batch of high-dimensional semantic vectors output by text vectorization subunit 211, and efficiently calculate the cosine similarity between all pairs of vectors through matrix operations to generate an initial nxn-dimensional semantic similarity matrix.

[0187] Subunit 213, the temporal mask application, dynamically generates a temporal mask matrix with all 1s in the lower triangular section and all 0s in the upper triangular section based on the publication time order of news items within this batch of data. This subunit performs element-wise multiplication (Hadamard product) between the semantic similarity matrix and this temporal mask matrix to obtain a semantic association matrix with temporal directionality.

[0188] Network generation subunit 214: Apply a preset similarity threshold (such as 0.7) to binarize the semantic association matrix or retain the weights to construct a directed weighted graph with news items as nodes and temporal semantic associations as edges, i.e., the news semantic network graph.

[0189] The community discovery unit 220 receives the news semantic network graph output by the network construction unit 210. The community discovery unit 220 operates based on graph algorithms for connected component discovery (such as depth-first search or disjoint-set data structure), traversing the entire news semantic network graph and identifying all sets of nodes directly or indirectly connected by edges. Each such maximally connected subgraph is marked as an initial hotspot text community.

[0190] Community optimization unit 230 includes:

[0191] Quality filtering subunit 231: Used to filter all initial communities output by community discovery unit 220, and remove communities that contain fewer than a preset minimum number of nodes (e.g., 3).

[0192] Sub-unit 232 of the topic consistency assessment: For each community, statistically analyze the distribution of topic tags in all member news. Calculate the cumulative frequency of the top two topic tags. If this frequency is lower than a preset threshold (e.g., 70%), the community is judged to have "topic inconsistency".

[0193] Community Segmentation Subunit 233: For communities with inconsistent themes, this subunit initiates a secondary processing flow. It constructs a new subgraph based on the nodes within the community, with edge weights calculated by weighting the theme distribution similarity (JS divergence) and temporal proximity (e.g., weight = 0.8 × JS + 0.2 × temporal proximity). Subsequently, the Louvain algorithm based on modularity optimization is run on this weighted graph to recursively segment the original community into multiple sub-communities with consistent themes and stronger cohesion.

[0194] Cross-block merging subunit 234: To handle long-tail hotspots spanning multiple time periods, this subunit maintains a global community pool. For newly generated communities, its "community feature vector" (e.g., the average of the title vectors of the top-3 news items within the community) is calculated, and its cosine similarity is calculated with the feature vectors of existing communities in the pool. If the similarity exceeds a merging threshold (e.g., 0.7), they are merged into the same cross-period hotspot event, and their feature vectors are updated. Finally, a purified and merged final set of hotspot events is output.

[0195] Content filtering unit 240 includes:

[0196] Importance calculation subunit 241: For each optimized hotspot community, an improved PageRank algorithm is executed on its internal network structure. The transition weights of edges between nodes in the algorithm are linearly weighted by semantic similarity, topic similarity, and temporal proximity (e.g., weight = 0.4 × semantic + 0.4 × topic + 0.2 × time). Through iterative calculation, the importance score of each text node within the community is obtained.

[0197] Representative text selection subunit 242: Based on the calculated importance score, all news nodes in the community are sorted in descending order. The top-ranked news items (e.g., 3) are selected as representative texts of the hot topic, packaged together with the metadata of the hot topic they belong to, and output to the briefing intelligent generation module 300.

[0198] The intelligent report generation module 300 is the system's output, responsible for converting the analysis results into easily readable professional reports. It includes: a large language model integration interface 301, a prompt template library 302, a task scheduling and input assembly unit 303, and a content generation and synthesis unit 304.

[0199] The large language model integration interface 301 encapsulates API communication with one or more generative large language models (in this invention, generative large language models refer to pre-trained artificial intelligence models with powerful text understanding and generation capabilities, such as, but not limited to, the GPT series, LLaMA series, DeepSeek series, etc. Those skilled in the art can select the appropriate specific model according to the actual situation.), and is responsible for authentication, request sending, response receiving, and error handling.

[0200] The prompt template library 302 stores a variety of structured prompt templates, which are key to guiding large models to generate specific content. The template library includes at least: hot topic guidance prompt templates; knowledge graph generation prompt templates; news rewriting and brief commentary templates; and hot topic extension generation prompt templates.

[0201] The task scheduling and input assembly unit 303 receives a set of hot events and representative text from module 200. Based on a preset briefing structure, it creates a generation task for each item of content (such as a guide, knowledge graph, etc.) of each hot event. This unit selects the corresponding template from the template library 302 and fills the template with specific representative text, hot event titles, and other contextual information into the predetermined positions of the template, forming a neat and specific prompt input.

[0202] The content generation and synthesis unit 304 calls the large language model integration interface 301 to process each generation task sequentially. It receives the JSON-formatted results returned by the large model and parses out the title, introduction, knowledge graph triples, rewritten news, brief commentary, extended text, and other content. Finally, following a standard report template (such as Word, HTML, or Markdown format), it integrates all generated parts with metadata (such as time range and involved countries) to generate a complete, well-formatted, structured hot topic analysis report, available for preview, download, or distribution.

[0203] Through the collaborative work of the above modules, this system achieves fully intelligent processing of the entire process of automatically, efficiently, and accurately mining regional and national hotspots from multilingual Internet information and generating in-depth analysis reports.

[0204] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for regional and national hotspot mining and briefing generation based on complex networks and large language models, characterized in that, include: S1: Multilingual public opinion data collection and processing: Collect multilingual text data from multiple information sources in a specified region or country, clean and deduplicate the text data, and sort it according to the publication time; S2: Temporal Semantic Hotspot Network Construction and Community Discovery: Based on the text data processed in step S1, a directed network is constructed with text as nodes and semantic similarity between texts under temporal constraints as edge weights; based on the directed network, a network community discovery algorithm is used to identify potential hotspot event communities. S3: Hotspot Community Optimization and Representative Content Screening: Conduct quality assessment and theme consistency judgment on the identified hotspot event communities, and segment or filter communities that do not meet the requirements; for each optimized hotspot community, screen out several pieces of text content that can represent the theme of the community; S4: Structured Hotspot Briefing Generation: Based on representative text content selected for each hotspot community, combined with preset prompt templates, the generative large language model is called to automatically generate a structured hotspot briefing containing at least one of the following elements: event overview, key entity relationships, content interpretation, and background extension.

2. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 1, characterized in that, The specific process of constructing the temporal semantic hotspot network in step S2 includes: S211: Text Vectorization: Convert the title and / or body content of each text into semantic vectors using a multilingual pre-trained model; S212: Semantic similarity calculation: Calculate the cosine similarity between the semantic vectors to form an initial semantic similarity matrix; S213: Temporal masking application: Construct a temporal masking matrix, where, for the i-th text sorted by time, only its association with the j-th text (j < i) sorted before it is retained; multiply the temporal masking matrix element-wise with the initial semantic similarity matrix to obtain a semantic association matrix with temporal directionality; S214: Network Construction: Using text as nodes, if the value of the corresponding element in the semantic association matrix is ​​greater than or equal to a preset similarity threshold, a directed edge is established between the corresponding text nodes to construct the directed network.

3. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 1, characterized in that, In step S2, the network community discovery algorithm is used to identify hot event communities. Specifically, based on the connection relationship between nodes in the directed network, a connected component-based algorithm is used to discover connected subgraphs in the network, and each connected subgraph is initially identified as a hot event community.

4. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 1, characterized in that, The quality assessment and topic consistency judgment of hotspot communities in step S3 include: S311: Community Size Filtering: Filter out communities containing fewer than a preset threshold number of texts; S312: Topic Distribution Consistency Check: Statistically analyze the topic tag distribution of all texts in the community. If the cumulative proportion of the top N high-frequency topic tags is lower than the preset consistency threshold, the community is determined to be inconsistent in terms of topic.

5. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 4, characterized in that, For communities determined to be thematically inconsistent, step S3 further includes: S313: Construct a weighted graph within the community: Using texts within the community as nodes, the edge weights between nodes are calculated by weighting the similarity of the topic distribution and the proximity of the publication time between the texts. S314: Community Segmentation: Based on the weighted graph, the community is re-divided using a modularity optimization-type community discovery algorithm until the resulting sub-communities meet the topic consistency requirement.

6. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 1, characterized in that, The specific method for selecting representative content in step S3 is as follows: S321: Based on the directed network or its derivative network, calculate the importance score of each text node in the community; wherein, the calculation of the importance score comprehensively considers the semantic centrality, topic representativeness and time freshness of the text node; S322: Sort the text nodes in the community according to the importance score, and select a preset number of texts with the highest ranking as representative content.

7. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 6, characterized in that, The importance score is calculated using an improved PageRank algorithm, where the transition weights between nodes are determined by a weighted sum of semantic similarity, topic similarity, and temporal proximity.

8. The method for regional and national hotspot mining and briefing generation based on complex networks and large language models according to claim 1, characterized in that, The preset prompt template in step S4 includes at least one of the following: Hot Topic Guide Template: Used to guide generative large language models to summarize and generate titles and abstracts of hot topics based on multiple representative texts; Knowledge graph generation prompt template: used to guide generative large language models to extract entities and relationships from text and form structured knowledge triples; News rewriting and brief commentary template: used to guide generative large language models to refine the content and rewrite the language of a single representative text, and generate commentary with a specific perspective; Hot Topic Extension Generation Prompt Template: This template guides the generative large language model to combine external knowledge base information to generate in-depth background analysis or trend forecasts for hot topics. In the Hot Topic Extension Generation Prompt Template, the generative large language model must mark the unique identifier of the referenced information in the generated text when referencing external knowledge base information.

9. A system for regional and national hotspot mining and briefing generation based on complex networks and large language models, characterized in that, The system for implementing the method as described in any one of claims 1 to 8 comprises: The data acquisition and processing module is configured to periodically or as needed collect multilingual raw public opinion data from news websites and government portals of multiple regions or countries specified on the Internet through a preset web crawler program; and to clean, deduplicate, standardize the format, and sort the raw public opinion data in ascending order by publication timestamp, and output a structured time-series text data set. The hotspot mining and analysis module is communicatively connected to the data acquisition and processing module, receives the time-series text data set, and is configured as follows: (a) Based on the text content in the aforementioned time-series text dataset, construct a news dissemination network that integrates time-series constraints and semantic similarity; (b) Execute a complex network community detection algorithm on the news dissemination network to initially identify text communities corresponding to potential hot topics; (c) Perform quality assessment, topic consistency verification and optimized segmentation on the initially identified text communities, and merge semantically similar communities across data batches to form the final set of hot events; (d) For each final hot topic event, calculate the importance score of each text node in the community, and select a preset number of representative texts based on the scores; The briefing intelligent generation module is communicatively connected to the hotspot mining and analysis module, receives the set of hotspot events and their corresponding representative texts, and is configured as follows: (e) Integrate at least one generative large language model; (f) A variety of structured prompt templates are pre-stored, each template corresponding to a component of the presentation; (g) Based on the structural requirements of the target briefing, combine representative texts with corresponding prompt templates to form a standardized prompt input; (h) Call the generative large language model to process the prompt input, automatically generate hot topic briefing components that meet the specified style and format requirements, and integrate them into a complete structured hot topic analysis report.

10. A regional and national hotspot mining and briefing generation system based on complex networks and large language models according to claim 9, characterized in that, The hotspot mining and analysis module further includes: The network construction unit includes a text vectorization subunit, a similarity calculation subunit, a temporal mask application subunit, and a network generation subunit, wherein: The text vectorization subunit utilizes a multilingual natural language processing model to convert the title and / or core content of each text into a high-dimensional semantic vector. The similarity calculation subunit is used to calculate the cosine similarity between all pairs of text semantic vectors and generate the initial semantic similarity matrix; The temporal mask application subunit is used to generate a lower triangular temporal mask matrix based on the publication time order of the text, which is used to mask the semantic association between texts that are published later and texts that are published earlier. The network generation subunit is used to perform element-wise multiplication of the initial semantic similarity matrix and the temporal mask matrix to obtain a weighted similarity matrix with temporal directionality; with text as nodes and values ​​in the weighted similarity matrix that exceed a preset threshold as edge weights, a directed or weighted news semantic network graph is constructed. The community discovery unit, connected to the network construction unit, is configured to execute an algorithm based on connected component discovery on the news semantic network graph to identify the set of all connected nodes in the graph as an initial hot text community. A community optimization unit, connected to the community discovery unit, includes a quality filtering subunit, a topic consistency assessment subunit, a community segmentation subunit, and a cross-block merging subunit; wherein: The quality filtering subunit is used to remove initial communities with fewer nodes than the preset minimum community size; The topic consistency assessment subunit is used to count the distribution of topic tags in the text within each community. If the sum of the percentages of the top two topic tags is lower than a preset threshold, the community is judged to be topic inconsistent. The community segmentation subunit is used to construct an internal subgraph with edge weights of topic similarity and time proximity for communities with inconsistent topics. A dynamic community detection algorithm based on modularity optimization is then used to recursively segment the community until all sub-communities meet the topic consistency requirement. The cross-block merging subunit is used to calculate the feature vectors of communities generated by different data batches. If the semantic similarity between the feature vectors exceeds the merging threshold, the communities are merged into the same cross-cycle hotspot event. The content filtering unit, connected to the community optimization unit, is configured to perform the following operations for each optimized trending community: Importance calculation subunit: Based on the network structure within the community, an improved PageRank algorithm is used to calculate the importance score of each text node, where the weight of the edges between nodes is composed of a linear weighted sum of semantic similarity, topic distribution similarity, and publication time proximity; Representative text selection subunit: Based on the calculated importance scores, all text nodes in the community are sorted in descending order, and the top N texts are selected as representative content of the hot topic community and output to the briefing intelligent generation module.

Images