Multi-platform cooperative propagation method, device and equipment for sudden news event and medium

Abstract

The invention provides a multi-platform cooperative propagation method, device and equipment for an emergency news event and a medium, and the method comprises the steps: carrying out the multi-modal feature analysis of an obtained multi-platform original data set of the emergency news event, and obtaining standardized information; performing event association and knowledge fusion on the standardized information, and integrating entities and corresponding relationships in the standardized information into a pre-constructed event knowledge graph to obtain an emergency knowledge graph; carrying out potential relation mining on the emergency knowledge graph based on a graph neural network, and identifying a development stage and an information weak region of an emergency news event to obtain a core information data set; inputting the core information data set into a pre-trained reinforcement learning model to generate a cooperative propagation instruction set; wherein the cooperative propagation instruction set is used for indicating the target platform to publish related information of the sudden news event. By adopting the method, the continuity and cognitive consistency of cross-platform information propagation can be improved.

Description

Technical Field

[0001] This invention belongs to the field of new media data analysis technology, and in particular relates to a method, device, equipment and medium for multi-platform collaborative dissemination of breaking news events. Background Technology

[0002] With the rapid development of information dissemination technology and multi-platform media ecosystems, this technology can quickly spread information about breaking news and expand audience reach. In traditional technologies, the handling of breaking news events usually relies on a model where each platform operates independently: after information is released by the source, each platform collects, edits, and pushes the information separately; information integration and collaboration between different platforms mainly rely on manual monitoring and synchronization.

[0003] However, the current approach or traditional method suffers from insufficient depth of information integration and weak dissemination synergy. The content on each platform is mostly isolated and static information points, lacking dynamic connections and semantic coherence based on the entire life cycle of an event. This results in the public obtaining fragmented information on different platforms, making it difficult to form a complete and coherent understanding of the event's context. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, device, equipment, and medium for multi-platform collaborative dissemination of breaking news events to address the aforementioned technical problems. This method can achieve dynamic, structured integration and correlation accumulation of fragmented information from multiple sources, thereby improving the coherence and consistency of information dissemination across platforms.

[0005] Firstly, this application provides a multi-platform collaborative dissemination method for breaking news events, including:

[0006] Multimodal feature analysis was performed on the acquired raw datasets of breaking news events from multiple platforms to obtain standardized information;

[0007] By performing event association and knowledge fusion on standardized information, entities and corresponding relationships in the standardized information are integrated into a pre-constructed event knowledge graph to obtain an emergency event knowledge graph.

[0008] Based on graph neural networks, potential relationships are mined from the knowledge graph of breaking events to identify the development stages and information gaps of breaking news events, and to obtain the core information dataset.

[0009] The core information dataset is input into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set; the collaborative propagation instruction set is used to instruct the target platform to release relevant information about breaking news events.

[0010] In one possible embodiment, multimodal feature parsing is performed on the acquired raw dataset of breaking news events from multiple platforms to obtain standardized information, including:

[0011] We obtain raw data from multiple platforms through news source API interfaces and official data subscription services, and perform data cleaning and format unification on the raw data from multiple platforms to obtain preprocessed heterogeneous data.

[0012] Preprocessed heterogeneous data is injected into a distributed message queue, and the message queue is used for traffic shaping and data buffering to obtain a stable data stream;

[0013] Text data is extracted from a stable data stream, and entity extraction is performed using a BERT-based named entity recognition model to obtain an entity set, which includes people, organizations, geographical locations, and time points.

[0014] Based on the text data, semantic relations are identified through a relation extraction model to obtain a set of relation triples describing the relationships between entities.

[0015] For image and / or video data in a stable data stream, target detection and scene classification are performed using a computer vision model to obtain corresponding visual description text;

[0016] The entity set, relation triplet set, and visual description text are structured and encapsulated to obtain standardized information.

[0017] In one possible embodiment, text data is extracted from a stable data stream, and entity extraction is performed using a BERT-based named entity recognition model to obtain an entity set, including:

[0018] The text data is segmented and stop words are removed to obtain a preprocessed text sequence;

[0019] The preprocessed text sequence is input into the fine-tuned BERT model, and entity boundary recognition and type classification are performed through the sequence labeling layer to obtain preliminary entity labeling results.

[0020] Based on the preliminary entity annotation results, the entity normalization algorithm is used to resolve the references, resulting in a standardized set of entities with unified names.

[0021] For the standard entity set, the reasonableness is verified by the context semantic verification module to obtain the verified entity set;

[0022] The entity set is classified and stored according to the preset entity type to obtain the entity set.

[0023] In one possible embodiment, standardized information is subjected to event association and knowledge fusion, integrating entities and corresponding relationships from the standardized information into a pre-constructed event knowledge graph to obtain an emergency event knowledge graph, including:

[0024] The association degree data between the entity set and the relation set and the existing event nodes in the event knowledge graph is calculated using an event association determination algorithm;

[0025] Based on correlation data, an event attribution determination is made by using a dynamic threshold comparator to determine the correspondence between standardized information and event nodes;

[0026] Based on the correspondence, a unified set of entity identifiers is obtained by performing referential disambiguation through an entity alignment algorithm.

[0027] For the entity identifier set, the conflict detection module performs relationship consistency checks to identify relationship pairs with semantic conflicts.

[0028] Based on the authority of the source and the confidence of the information, a set of resolved relationships is obtained by using an evidence weighting algorithm to resolve conflict relationships.

[0029] The set of resolved relationships is integrated into the event knowledge graph, and the structure is optimized through the graph update engine to obtain the event knowledge graph.

[0030] In one possible embodiment, based on source authority and information confidence, a conflict resolution algorithm is used to obtain a set of resolved relationships, including:

[0031] Based on the existing set of relations in the event knowledge graph and the new set of relations in the standardized information, relation consistency comparison is performed through semantic conflict detection to identify the set of conflicting relation pairs with different semantic representations among the same entity pairs;

[0032] Extract source authority from standardized information source information, and convert source authority into standardized authority weight through an authority quantifier to obtain the authority quantification value corresponding to each relation;

[0033] Relationship confidence scores are extracted from the information confidence scores of standardized information, and the relationship confidence scores are normalized using a confidence calibrator to obtain the standard confidence value corresponding to each relationship.

[0034] Based on authoritative metric values ​​and standard confidence values, the following evidence-weighted formula is used for relation optimization decision processing to obtain the optimal relation identifier for each conflicting relation pair:

[0035]

[0036] in, Indicates the first A relationship, Represents sets with conflicting relationships. Representing relations The authority of the source Representing relations Information confidence level, For authoritative quantification, This is the standard confidence level value;

[0037] Based on the optimal relation identifier, the optimal relation is selected from the set of conflicting relation pairs using a relation selector, resulting in a preliminary set of resolved relations;

[0038] For the relationships in the set of conflicting relationships that were not selected as the optimal relationship, metadata annotation and temporary storage are performed through a temporary storage manager to obtain a set of conflicting relationships to be re-evaluated.

[0039] Based on the standardized information subsequently received, the incremental evaluator supplements the evidence and re-scores the set of conflict relations to be re-evaluated, resulting in a supplementary set of resolved relations.

[0040] The initial set of resolved relations and the supplementary set of resolved relations are merged and deduplicated to obtain the set of resolved relations.

[0041] In one possible embodiment, latent relationship mining is performed on the knowledge graph of breaking news events based on graph neural networks to identify the development stages and information gaps of breaking news events, thereby obtaining a core information dataset, including:

[0042] Based on the topological structure of the knowledge graph of emergencies, node embedding learning is performed through GraphSAGE graph neural network to obtain low-dimensional vector representations of nodes;

[0043] Based on the low-dimensional vector representation of nodes, the cosine similarity algorithm is used to calculate the potential relationship and obtain a candidate set of potential relationships.

[0044] The set of potential relation candidates is filtered for false relations by a semantic rationality validator to obtain a set of valid potential relations.

[0045] Based on the connectivity characteristics of the knowledge graph of emergencies, the subgraph structure is analyzed and processed by the graph density analysis algorithm to obtain the subgraph density distribution data;

[0046] Based on the subgraph density distribution data, a sparse region detection process is performed using an information-weak region identification algorithm to obtain the location information of the information-weak region.

[0047] By combining time series features, an event stage determination process is performed using a development stage classifier to obtain the development stage label;

[0048] Based on the location information and development stage labels of information-weak areas, key information is extracted and processed to obtain a core information dataset.

[0049] In one possible embodiment, the core information dataset is input into a pre-trained reinforcement learning model to generate a set of cooperative propagation instructions, including:

[0050] Semantic feature extraction is performed on the core information dataset to obtain a state vector representing the event state.

[0051] The state vector is input into the policy network of the pre-built reinforcement learning model, and action decisions are made through forward propagation to obtain the original action parameters, which include the platform weight vector and the time offset.

[0052] The original motion parameters are processed by the motion constraint to verify the feasibility of the parameters, and then the feasible motion parameters that meet the platform capabilities and time constraints are obtained.

[0053] Based on the key entities and relational paths in the core information dataset, a pre-trained language generation model is used to generate text sequences to obtain semantically coherent core declarative sentences.

[0054] Based on the platform weight vector in the action parameters, a multi-platform preference decision-making process is performed through the platform selection algorithm to obtain the target platform preference list;

[0055] Based on the time offset in the action parameters, the release time window is calculated using a time-series scheduling algorithm to obtain the release time schedule.

[0056] The core statements, target platform preference list, and release schedule are structured and integrated to obtain a collaborative propagation instruction set.

[0057] Secondly, this application also provides a multi-platform collaborative dissemination device for breaking news events, including:

[0058] The feature parsing module is used to perform multimodal feature parsing on the acquired raw datasets of breaking news events from multiple platforms to obtain standardized information;

[0059] The graph fusion module is used to perform event association and knowledge fusion on standardized information, integrating entities and corresponding relationships in the standardized information into a pre-built event knowledge graph to obtain an emergency event knowledge graph.

[0060] The relation mining module is used to mine potential relationships in the knowledge graph of breaking events based on graph neural networks, identify the development stages and information gaps of breaking news events, and obtain the core information dataset.

[0061] The propagation instruction generation module is used to input the core information dataset into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set; the collaborative propagation instruction set is used to instruct the target platform to release relevant information about breaking news events.

[0062] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned multi-platform collaborative dissemination method for breaking news events.

[0063] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned multi-platform collaborative dissemination method for breaking news events.

[0064] The aforementioned multi-platform collaborative dissemination methods, devices, equipment, and media for breaking news events effectively solve the problem of dynamically integrating fragmented information from multiple sources by processing original datasets from multiple platforms through multimodal feature analysis to generate standardized information. Through event association and knowledge fusion, entities and relationships in the standardized information are integrated into a pre-constructed event knowledge graph, achieving structured information integration. Based on graph neural networks, potential relationships are mined from the breaking news knowledge graph to identify event development stages and information gaps, obtaining a core information dataset and enhancing the accumulation of information relevance. A pre-trained reinforcement learning model generates a collaborative dissemination instruction set to guide the target platforms in releasing information, thereby improving the efficiency and accuracy of public awareness of breaking news events. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 A flowchart illustrating the multi-platform collaborative dissemination method for breaking news events provided in an embodiment of the present invention;

[0067] Figure 2 A schematic diagram of the structure of a multi-platform collaborative dissemination device for breaking news events provided in an embodiment of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0069] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0070] Breaking news events refer to public information events that are time-sensitive, attract high public attention, and have a wide reach. Their core characteristics include: suddenness (occurring unplanned); news value (involving public interests or social impact); multi-source nature (information generated from dispersed sources); and evolution (the state of the event changes dynamically over time). In this technical solution, it refers to news categories that require cross-platform collaborative processing.

[0071] An event knowledge graph is a knowledge representation structure organized around events. It depicts various entities and their complex relationships within an event through nodes and edges. Unlike traditional static knowledge bases, event knowledge graphs are dynamically evolving, continuously updating and improving as events progress. In this solution, the event knowledge graph serves as the central database carrying knowledge related to emergencies, providing structured knowledge support for subsequent analysis and decision-making.

[0072] Latent relation mining refers to the technical process of discovering implicit connections in data that are not yet explicitly expressed but actually exist through computational analysis methods. This technique is based on graph computing and machine learning methods to derive new semantic connections from known explicit relationships. In this approach, latent relation mining can reveal deep connections between different information points in a sudden event, thereby discovering important associations that are easily overlooked and enhancing the understanding of the overall picture of the event.

[0073] Based on the above definitions, the implementation environment of the multi-platform collaborative dissemination method for breaking news events provided in this application embodiment will be described. Indicatively, this implementation environment includes: a terminal and a processor. The terminal includes a data acquisition terminal and a target dissemination platform; the processor includes, but is not limited to, a central processing unit, a multi-core processor, or an artificial intelligence chip.

[0074] Based on the above definitions and implementation environment, the application scenarios of the embodiments of this application are described. The multi-platform collaborative dissemination method for breaking news events provided in the embodiments of this application can be applied to scenarios including but not limited to the following:

[0075] In natural disaster emergency response scenarios, this technical solution enables intelligent integration and collaborative dissemination of disaster information. When major natural disasters such as earthquakes and floods occur, information is often scattered across multiple channels, including news media, social media platforms, and government announcements. Through multimodal feature analysis and event knowledge graph construction, this solution can automatically integrate disaster information from different sources, such as meteorological departments, rescue agencies, and on-site residents, forming a unified disaster awareness graph. The collaborative dissemination instruction set based on a reinforcement learning model ensures that key content such as authoritative early warning information, rescue guidelines, and evacuation locations are updated synchronously across platforms, effectively guiding public evacuation and the deployment of rescue forces.

[0076] In public health emergency reporting scenarios, this technical solution can significantly improve the efficiency and accuracy of epidemic information dissemination. In the face of public health emergencies, this method continuously collects information from multiple sources such as medical institutions, disease control centers, and media reports to construct a dynamically evolving epidemic knowledge graph. By utilizing potential relationship mining technology, it can promptly identify potential connections and risk areas in the epidemic transmission chain, providing support for prevention and control decisions. By generating a collaborative dissemination instruction set, it ensures that key information such as vaccination and prevention and control measures remain consistent in content and time sequence across various platforms, avoiding public confusion caused by information discrepancies.

[0077] In the context of handling social security incidents, this technical solution enables rapid response and precise delivery of authoritative information. When a major security accident or social emergency occurs, this method rapidly constructs the event's development timeline by monitoring information flow across various platforms in real time. Based on knowledge graph-based state analysis, the system accurately identifies gaps and cognitive biases in information dissemination, promptly generating targeted clarifications and situation reports. Through an optimized dissemination instruction set, it guides official media and government platforms to collaboratively release the latest developments, effectively curbing the spread of rumors and maintaining social stability.

[0078] This is merely an illustrative example; the multi-platform collaborative dissemination method for breaking news events provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.

[0079] In one exemplary embodiment, such as Figure 1 As shown, a multi-platform collaborative dissemination method for breaking news events is provided. This embodiment illustrates the application of this method to a terminal in the aforementioned implementation environment. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104:

[0080] Step 101: Perform multimodal feature analysis on the obtained multi-platform raw dataset of breaking news events to obtain standardized information.

[0081] Specifically, this method can collect various types of information constituting the original datasets of multiple platforms through news agency APIs, government data disclosure platforms, and authorized social media data interfaces. This original data includes news report texts, official announcements, social media posts, and related multimedia content such as images and videos. For example, this method uses pre-trained cross-modal encoders (such as the CLIP model for image-text alignment and the Wav2Vec model for audio) to extract multi-dimensional information such as text keywords, image semantic tags, and audio sentiment features. This information is then mapped to a unified vector space through a feature alignment layer, generating structured information units containing standardized fields such as timestamps, geographic coordinates, and event subjects.

[0082] Step 102: Perform event association and knowledge fusion on the standardized information, and integrate the entities and corresponding relationships in the standardized information into the pre-constructed event knowledge graph to obtain the emergency event knowledge graph.

[0083] This method, based on standardized information, implements event association and knowledge fusion to construct a dynamically evolving knowledge graph of emergencies. Specifically, it can also employ a hierarchical clustering algorithm based on semantic similarity to calculate the association strength between newly arriving information and existing event nodes in the knowledge graph. For example, by calculating the Jaccard similarity of entity sets, the cosine similarity of content features, and the correlation of time series, the event affiliation of information units can be comprehensively determined. For instance, the TransE graph embedding algorithm can be used to calculate entity similarity, achieving cross-platform entity disambiguation (e.g., mapping "COVID-19" and "novel coronavirus" to the same entity node), and a relational reasoning engine can be used to supplement missing spatiotemporal association edges, forming a graph structure containing the main chain of event evolution and the scope of influence. The generated emergency event knowledge graph, as the core data carrier, can realize topological storage and associated querying of fragmented information. Furthermore, this method employs a confidence-weighted knowledge fusion mechanism. When a relationship conflict is detected, dynamic arbitration is performed based on the authority of the information source and the confidence of the content, prioritizing the retention of high-confidence relationship assertions. This fusion mechanism not only maintains the consistency of the knowledge graph but also continuously enriches the semantic hierarchy of the graph through entity alignment and relationship completion.

[0084] Step 103: Based on graph neural networks, perform potential relationship mining on the knowledge graph of breaking news events to identify the development stages and information gaps of breaking news events, and obtain the core information dataset.

[0085] Specifically, this method employs graph attention networks to learn representations of knowledge graphs, capturing complex dependencies between nodes through a multi-layered message passing mechanism. For example, by calculating the similarity matrix of node embedding vectors, potential associations not explicitly expressed in the original data can be discovered, such as inferring causal relationships between events at different geographical locations. Furthermore, this method combines topological feature analysis and temporal pattern recognition, automatically dividing the development stages of events—such as the nascent, explosive, and waning phases—by monitoring changes in subgraph density and node growth rates. Simultaneously, it detects sparse information regions in the graph based on structural entropy, providing data support for subsequent precise dissemination.

[0086] Step 104: Input the core information dataset into the pre-trained reinforcement learning model to generate a collaborative propagation instruction set; wherein, the collaborative propagation instruction set is used to instruct the target platform to release relevant information about breaking news events.

[0087] Specifically, the core information dataset is input into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set. The PPO algorithm can be used to construct a platform collaborative decision-making environment model, where the state space is defined as the feature vectors of the event development stage and the weak area, and the action space includes content synchronization strategies, release time windows, and channel adaptation rules. Furthermore, the policy network is optimized with public perception consistency as the reward function, and instruction sequences are generated through Monte Carlo tree search. For example, typical instructions include requiring news platforms to release event timeline text and images at T+5 minutes, short video platforms to remove blurry videos during conflict periods, and government platforms to open a clue collection portal. This step can realize machine intelligent decision-making for cross-platform release strategies.

[0088] The aforementioned multi-platform collaborative dissemination method for breaking news events effectively solves the problem of dynamically integrating fragmented information from multiple sources by processing original datasets from multiple platforms through multimodal feature analysis to generate standardized information. It integrates entities and relationships from standardized information into a pre-constructed event knowledge graph through event association and knowledge fusion, achieving structured information integration. Based on graph neural networks, it mines potential relationships within the event knowledge graph, identifies event development stages and information gaps, and obtains a core information dataset, enhancing the accumulation of information relevance. Finally, it generates a collaborative dissemination instruction set through a pre-trained reinforcement learning model to guide the target platforms in releasing information, thereby improving the efficiency and accuracy of public awareness of breaking news events.

[0089] In one possible embodiment, multimodal feature parsing is performed on the acquired raw dataset of breaking news events from multiple platforms to obtain standardized information, including:

[0090] We obtain raw data from multiple platforms through news source API interfaces and official data subscription services, and then perform data cleaning and format unification processing on the raw data from multiple platforms to obtain preprocessed heterogeneous data.

[0091] Specifically, this method can acquire raw data from multiple platforms through news source API interfaces and official data subscription services. The news source API interfaces include RESTful API interfaces provided by mainstream media organizations, while the official data subscription service receives real-time data released by government agencies and emergency management departments via webhooks. For example, this method employs a data cleaning strategy combining rule engines and machine learning to perform deduplication, outlier detection, and format standardization operations on the acquired raw data. For instance, it uses locality-sensitive hashing (LSH) to identify nearly duplicate text, box plots to detect abnormal fluctuations in numerical data, and uniformly converts different character encodings to UTF-8 format. Furthermore, this method constructs a data quality assessment matrix to quantitatively score the preprocessed data from three dimensions: completeness, accuracy, and timeliness, ultimately outputting preprocessed heterogeneous data that meets quality requirements.

[0092] Preprocessed heterogeneous data is injected into a distributed message queue, and the message queue is used for traffic shaping and data buffering to obtain a stable data stream.

[0093] Specifically, this method employs a token bucket algorithm for traffic shaping control, dynamically adjusting the processing rate of different data streams based on data priority to ensure that urgent information with high timeliness receives priority. Furthermore, the method incorporates a multi-level caching mechanism in its message queue design, including a memory-level fast cache and a disk-level persistent cache, managing cache replacement through a Least Recently Used algorithm to effectively handle sudden traffic surges. Moreover, by monitoring message queue backlog and consumption latency metrics, the method dynamically adjusts the number of consumer instances to guarantee data stream stability and processing timeliness.

[0094] Text data is extracted from a stable data stream, and entity extraction is performed using a BERT-based named entity recognition model to obtain an entity set, which includes people, organizations, geographical locations, and time points.

[0095] Specifically, the model can employ a multi-layer bidirectional Transformer encoder architecture, fine-tuning it using news domain corpora after pre-training, such as optimizing model parameters using a breaking news annotation dataset. For example, this method introduces an attention mechanism during entity recognition, accurately identifying entity boundaries through weighted calculations of contextual semantics, with specific optimizations for common place name abbreviations, organization abbreviations, and other special expressions found in news texts. Furthermore, the method uses a conditional random field model for post-processing of the recognition results, decoding with the Viterbi algorithm to obtain optimal entity sequence annotations, ultimately outputting a set of entities including people, organizations, geographical locations, and time points.

[0096] Based on the text data, semantic relations are identified through a relation extraction model to obtain a set of relation triples describing the relationships between entities.

[0097] Specifically, this relation extraction model employs a sequence-to-sequence architecture based on pointer networks. An encoder converts the input text into a hidden-state representation, and a decoder generates relation triples. For example, during training, this method uses a negative sampling strategy, randomly replacing the head or tail entities in the triples to construct negative examples, thus improving the model's ability to detect spurious relations. Furthermore, the method introduces an adversarial training mechanism, generating adversarial examples through gradient sign attacks to enhance the model's robustness. The method also includes a relation confidence calculation module, which outputs the probability distribution for each relation type using a softmax function, retaining only relation triples with confidence levels above a threshold to ensure the reliability of the extraction results.

[0098] For image and / or video data in a stable data stream, target detection and scene classification are performed using a computer vision model to obtain corresponding visual description text.

[0099] For example, this method can employ a YOLOv5-based object detection model, pre-trained on the COCO dataset, and then trained using a news image dataset for domain-adaptive training. For example, this method introduces a multi-scale feature fusion mechanism in scene classification tasks, extracting visual features of different granularities through a feature pyramid network and highlighting key regions using a self-attention mechanism. Furthermore, this method employs a Transformer-based image description generation model, converting visual feature sequences into natural language descriptions, generating multiple candidate descriptions through a beam search algorithm, and then selecting the optimal visual description text based on a language model score.

[0100] The entity set, relation triplet set, and visual description text are structured and encapsulated to obtain standardized information.

[0101] Specifically, this method designs a unified data schema definition, uses JSON-LD format as the carrier of standardized information, and clarifies semantic relationships by defining the @context field. For example, this method adds proofance metadata to each data unit to record the data source, processing pipeline, and quality assessment results. Furthermore, this method introduces a data version management mechanism, generating data fingerprints through content addressing to ensure data immutability and traceability. This method also implements data compression optimization, using Protocol Buffers binary serialization format to reduce storage and transmission overhead, generating standardized information that is both semantically clear and processing-efficient.

[0102] In one possible embodiment, text data is extracted from a stable data stream, and entity extraction is performed using a BERT-based named entity recognition model to obtain an entity set, including:

[0103] The text data is segmented and stop words are removed to obtain a preprocessed text sequence.

[0104] Specifically, this method employs a word segmentation strategy combining a bidirectional maximum matching algorithm and a statistical language model, while simultaneously constructing a domain-specific dictionary to identify proper nouns and compound terms in news texts. For example, this method can utilize a Hidden Markov Model to probabilistically model the text sequence and solve for the optimal word segmentation path using the Viterbi algorithm, ensuring accurate segmentation of complex words such as professional names and place names. Furthermore, this method employs a multi-level stop word filtering mechanism, not only removing general stop words but also constructing a dedicated stop word list based on news domain characteristics. This effectively filters out common meaningless expressions in news reports while retaining functional words with actual semantic meaning. This processing effectively reduces data dimensionality and improves the accuracy of subsequent processing, providing clean text input for entity recognition tasks.

[0105] The preprocessed text sequence is input into the fine-tuned BERT model, and entity boundary recognition and type classification are performed through the sequence labeling layer to obtain preliminary entity labeling results.

[0106] For example, this method introduces a Conditional Random Field (CRF) layer on top of the BERT model, using context-aware word vectors output by BERT as features to model the transition probabilities between labels through the CRF. Specifically, this method employs the BIOES annotation system for fine-grained annotation of entity boundaries, where B represents the start of an entity, I represents the interior of an entity, E represents the end of an entity, and S represents a single-word entity. This annotation method can more accurately capture entity boundary information. Furthermore, this method adopts an adversarial training strategy during fine-tuning, improving the model's robustness by adding gradient perturbations, while using a focus loss function to alleviate the entity class imbalance problem. This process can accurately identify named entities and their categories in the text, providing structured annotation results for subsequent entity normalization.

[0107] Based on the preliminary entity annotation results, the entity normalization algorithm is used to resolve the referentials, resulting in a standardized set of entities with unified names.

[0108] Specifically, this method employs a multi-feature fusion approach combining edit distance and semantic similarity to calculate similarity scores between different entity references. For example, it first utilizes a rule-based cleaning strategy to handle common spelling variations and abbreviations, calculates the semantic similarity of entity references using a word vector model, and then uses a hierarchical clustering algorithm to group entity references pointing to the same real-world object into the same category. Furthermore, the method introduces a knowledge graph-based entity linking technique to align entity references in the text with standard entities in a knowledge base. The optimal link object is determined through a comprehensive calculation of prior probability and contextual relevance. This process effectively addresses the issue of different representations of the same entity, forming a unified and standardized entity representation.

[0109] For the standard entity set, the reasonableness is verified by the context semantic verification module to obtain the verified entity set.

[0110] For example, this method employs a Transformer-based pre-trained language model to construct a contextual semantic understanding module, evaluating the reasonableness of an entity's appearance by calculating the consistency score between the entity and its context. Specifically, this method trains a context validator through a masked language model task, replacing the entity to be verified with a special label before inputting it into the model, and evaluating reasonableness by comparing the consistency between the model's predictions and the actual entities. Furthermore, this method introduces a global consistency verification mechanism based on graph neural networks, utilizing the relationship network between entities to construct constraints, propagating verification information through a message passing algorithm, and identifying anomalous entities that do not conform to the overall semantic pattern. This process can effectively filter out incorrectly identified entities, improving the quality and reliability of the entity set.

[0111] The entity set is classified and stored according to the preset entity type to obtain the entity set.

[0112] Specifically, this method designs a distributed storage architecture based on NoSQL databases, establishing different data shards according to entity types, and creating an inverted index for each entity to support fast retrieval. For example, the method uses a columnar storage format to compress entity attributes, optimizes storage efficiency through dictionary encoding and bitmap indexing techniques, and maintains relationship indexes between entities to support complex graph query operations. Furthermore, the method implements an entity version management mechanism, recording the evolution history of entities through multi-version concurrency control technology, supporting time-travel queries of entity information. Further, a dynamic priority storage strategy can be designed: creating four types of storage partitions: PER (Person), ORG (Organization), LOC (Local Location), and TIME (Time Point), automatically adjusting the partition capacity ratio based on event type (e.g., allocating 60% of resources to the LOC partition for disaster events), and employing an LSM-Tree index structure to form a structured entity set that can directly serve knowledge graph construction. This embodiment realizes the transformation from raw text to a standardized entity set by constructing a complete entity recognition and processing pipeline. This method combines deep learning methods with a rule engine, which significantly improves the accuracy and recall of entity recognition while ensuring processing efficiency, and provides high-quality entity data support for subsequent knowledge graph construction and intelligent decision-making.

[0113] In one possible embodiment, standardized information is subjected to event association and knowledge fusion, integrating entities and corresponding relationships from the standardized information into a pre-constructed event knowledge graph to obtain an emergency event knowledge graph, including:

[0114] The association degree data between the entity set and the relation set and the existing event nodes in the event knowledge graph is calculated using an event association determination algorithm.

[0115] For example, this method can employ a multi-dimensional feature fusion-based relevance calculation model. Its core principle is to quantify the degree of relevance by comprehensively evaluating three dimensions: entity overlap, semantic relevance, and temporal proximity. Specifically, in calculating entity overlap, this method uses an improved Jaccard similarity coefficient, considering not only the intersection-union ratio of entity sets but also introducing a weighting factor based on entity importance. The TF-IDF algorithm is used to calculate the distinguishability of each entity in the event description, giving key entities higher weight in the relevance calculation. In calculating semantic relevance, this method uses a Sentence-BERT model to map the event description text to a semantic vector space, calculating semantic distance using cosine similarity, and introducing an attention mechanism to weight key information in the description text. In calculating temporal proximity, this method uses an exponential decay function to simulate the time decay effect, which can dynamically adjust according to the event type. These three dimensions of features are adaptively weighted and fused through a neural network to output a comprehensive relevance score.

[0116] Based on correlation data, an event attribution determination is made through a dynamic threshold comparator to establish the correspondence between standardized information and event nodes.

[0117] Specifically, the design principle of dynamic thresholds is based on the phased characteristics of event development and the changing patterns of information density. For example, a threshold adjustment model based on the event lifecycle can be established, using a lower association threshold during the event's peak to quickly absorb relevant information, and a higher threshold during the event's stable phase to ensure information quality. For instance, this method monitors changes in the knowledge graph's topology in real time, calculates node density and connectivity indices, and automatically lowers the threshold when rapid expansion of the graph structure is detected, and raises the threshold requirement otherwise. Furthermore, this method also considers the distribution characteristics of information sources; when multiple independent sources simultaneously report similar content, the association threshold is appropriately lowered to accelerate information integration. This dynamic adjustment mechanism improves the accuracy and timeliness of event classification.

[0118] Based on the correspondence, a unified set of entity identifiers is obtained by performing referential disambiguation through an entity alignment algorithm.

[0119] Specifically, the core principle of the entity alignment algorithm is to resolve referential ambiguity in natural language by using multi-source evidence for entity identity resolution. For example, this method constructs an entity disambiguation model based on graph neural networks, using entity descriptions, contextual information, and relationships in the knowledge graph as input features. For each entity referential to be disambiguated, the method first retrieves a set of candidate entities from the knowledge graph, and then comprehensively scores the evidence based on three dimensions: contextual similarity, relational network matching, and attribute consistency. Contextual similarity uses the BERT model to calculate the semantic matching degree between the referential context and the candidate entity description; relational network matching compares the similarity between the relational network of the referential and the association structure of the candidate entity in the knowledge graph; and attribute consistency verifies the degree of conformity between the attribute features of the referential and the attributes of the candidate entity. These pieces of evidence are weighted and fused through an attention mechanism to ultimately determine the optimal entity alignment scheme.

[0120] For the set of entity identifiers, the conflict detection module performs relationship consistency checks to identify relationship pairs with semantic conflicts.

[0121] Specifically, this method constructs a domain ontology constraint rule base, defining logical constraints between event types, entity categories, and relationships. For example, it defines transitivity constraints for "located" relationships and antisymmetric constraints for "belongs to" relationships. When a newly added relationship violates these logical constraints with existing relationships in the knowledge graph, it is marked as a potential conflict. Furthermore, this method can also employ a conflict propagation algorithm based on graph traversal. When a conflict in a basic relationship is detected, it automatically deduces potential derivative conflicts. For example, if it detects a contradiction between "A is located in B" and "B is located in C" and "A is located in C," the method will continue to detect other potentially affected relationships along the relationship path.

[0122] Based on the authority of the source and the confidence of the information, a set of resolved relationships is obtained by using an evidence weighting algorithm to resolve conflict relationships.

[0123] Specifically, this method constructs an evidence collection framework, treating information source authority rating, content confidence score, time freshness, and multi-source consistency as four independent evidence sources. The authority rating is dynamically calculated based on the historical accuracy of the information source; content confidence is obtained through the deterministic output of a language model; time freshness employs an exponential decay model; and multi-source consistency is calculated by statistically analyzing the support of independent sources. These pieces of evidence are fused using the DS combination rule to calculate the credibility measure of each relational hypothesis. In particular, this method introduces an evidence discount factor to handle low-quality evidence and uses a conflict coefficient to detect the degree of contradiction between pieces of evidence; when the conflict is too great, a manual review process is initiated.

[0124] The set of resolved relationships is integrated into the event knowledge graph, and the structure is optimized through the graph update engine to obtain the event knowledge graph.

[0125] Specifically, this method employs an incremental graph update algorithm to determine the optimal update position by calculating the changes in modularity and connectivity of the graph after relation insertion. For example, it uses a force-oriented algorithm-based layout optimization to simulate the attractive and repulsive forces between nodes, allowing semantically similar nodes to naturally cluster in the graph space. Simultaneously, the method implements a real-time consistency maintenance mechanism, detecting potential logical contradictions after updates using a descriptive logic inference engine and repairing them using the principle of minimum change. Furthermore, the method maintains the version history of the graph, supporting change tracing and rollback operations to ensure the interpretability of knowledge evolution.

[0126] In one possible embodiment, based on source authority and information confidence, a conflict resolution algorithm is used to obtain a set of resolved relationships, including:

[0127] Based on the existing set of relations in the event knowledge graph and the new set of relations in the standardized information, relation consistency comparison is performed through semantic conflict detection to identify the set of conflicting relation pairs with different semantic representations among the same entity pairs.

[0128] Specifically, this method first constructs a conflict rule base, defining logical constraints between different types of relationships, including but not limited to symmetry constraints, transitivity constraints, and functional constraints. For example, for the "located" relationship, it is defined as having a transitivity constraint, i.e., if A is located in B and B is located in C, then A must be located in C; for the "belongs to" relationship, it is defined as having an antisymmetry constraint. This method employs a table-aux inference algorithm, traversing entity pairs in the relationship set to check for violations of these logical constraints. When different semantic representations are detected between the same entity pair, the method records detailed information about these conflicting relationships, including the conflict type, conflict degree, and involved entity information, forming a complete set of conflicting relationship pairs. This step, based on the reasoning mechanism of descriptive logic, identifies semantically contradictory relationships by establishing domain ontology constraints.

[0129] Extract source authority from the source information of standardized information, and convert the source authority into standardized authority weights through an authority quantifier to obtain the authority quantification value corresponding to each relation.

[0130] The principle behind the authority quantifier is based on a multi-dimensional evaluation model to comprehensively quantify the credibility of information sources. Specifically, this method constructs an authority evaluation index system, including four dimensions: historical accuracy rate, institutional credibility, domain expertise, and content standardization. Historical accuracy rate is calculated by statistically analyzing the proportion of information provided by the source in the past that has been verified as correct; institutional credibility is determined based on the institution's authoritative rating within the industry; domain expertise is assessed based on the source's professional qualifications and experience in the relevant news field; and content standardization examines the format and language standardization of the information provided. This method uses the analytic hierarchy process (AHP) to determine the weights of each dimension, normalizes the index values ​​of different dimensions to the 0-1 range using feature scaling techniques, and finally obtains standardized authority weight values ​​through weighted summation.

[0131] Relationship confidence scores are extracted from the information confidence scores of standardized information. The relationship confidence scores are then normalized using a confidence calibrator to obtain the standard confidence value for each relationship.

[0132] The technical principle of confidence calibration is based on probabilistic calibration theory. It involves establishing a calibration mapping function to convert the original confidence level output by the model into a true probability estimate. Specifically, this method can employ the Platt scaling algorithm, using validation set data to learn the mapping function from the original confidence level to the calibrated confidence level. First, a validation dataset containing real labels is collected. Then, a logistic regression model is used to fit the relationship between the original confidence level and the true accuracy. The calibrated confidence level more accurately reflects the true credibility of the relationship, providing a reliable probabilistic basis for subsequent evidence weighting.

[0133] Based on authoritative metric values ​​and standard confidence values, the following evidence-weighted formula is used for relation optimization decision processing to obtain the optimal relation identifier for each conflicting relation pair:

[0134]

[0135] in, Indicates the first A relationship, Represents sets with conflicting relationships. Representing relations The authority of the source Representing relations Information confidence level, For authoritative quantification, This is the standard confidence level value.

[0136] Specifically, and The objective weighting method based on information entropy is used to determine the optimal relationship. This method calculates the information entropy of each evidence source; the smaller the entropy value, the stronger the discriminative power of the evidence source, and the higher its weight should be assigned. During the calculation, for each candidate relationship in a conflicting relationship pair, a weighted comprehensive score is calculated, and the relationship with the highest score is selected as the optimal relationship. The significance of this formula lies in providing a quantitative relationship optimization mechanism that can automatically select the most credible relationship expression based on objective evidence in the presence of conflict, ensuring the scientific rigor and reliability of knowledge fusion.

[0137] Based on the optimal relation identifier, the optimal relation is selected from the set of conflicting relation pairs using a relation selector, resulting in a preliminary set of resolved relations.

[0138] The relation selector is implemented based on an efficient data indexing and querying mechanism, quickly locating candidate relations among conflicting relation pairs by establishing a hash index. Specifically, the method uses a Bloom filter for initial screening to reduce unnecessary full comparisons, and then uses a B+ tree index to accurately retrieve the relation records to be processed. For each conflicting relation pair, the system extracts the corresponding relation triplet from the candidate relation set based on the optimal relation identifier, while recording the selection reason and decision basis. This process also includes a result verification step, which checks the logical consistency between the selected relation and existing relations in the knowledge graph to ensure that the optimal result does not introduce new semantic conflicts.

[0139] For the relationships in the set of conflicting relationships that were not selected as the optimal relationship, metadata annotation and temporary storage are performed through a temporary storage manager to obtain a set of conflicting relationships to be re-evaluated.

[0140] The temporary storage manager employs a multi-level storage architecture, including a memory cache, solid-state drives (SSDs), and hard disk drives (HDDs). Specifically, relationships that may be re-evaluated soon are stored in the memory cache for fast access; relationships with lower evaluation priority are stored on the SSD; and historical conflict relationships are archived on the HDD. This method records detailed metadata for each temporary relationship, including conflict type, confidence score, timestamp, source information, and decision history. This metadata is serialized and stored using Protocol Buffers format, ensuring both storage efficiency and facilitating subsequent incremental evaluation.

[0141] Based on the standardized information subsequently received, the incremental evaluator supplements the evidence and re-scores the set of conflict relations to be re-evaluated, resulting in a supplementary set of resolved relations.

[0142] Specifically, the core algorithm of the incremental evaluator is based on an online learning mechanism, enabling it to dynamically update relationship evaluation results using newly arriving information. For example, this method can employ a sliding window mechanism to acquire new data streams and recalculate relationship confidence using a Bayesian update algorithm. For instance, when new official hospital data is added, the temporarily stored relationship "30 casualties" can be re-evaluated and its score improved. The final supplementary set of dissolved relationships dynamically incorporates new evidence, enhancing the completeness of decision-making.

[0143] The initial set of resolved relations and the supplementary set of resolved relations are merged and deduplicated to obtain the set of resolved relations.

[0144] In this step, the merging and deduplication process is based on a unified identifier system for entities and relations, and employs a semantic-based duplicate detection algorithm. Specifically, the method first performs syntactic deduplication on the two relation sets by comparing whether the subject, predicate, and object of the relation triples are identical to identify explicit duplicates. Then, semantic deduplication is performed by calculating the semantic similarity of relation descriptions to identify relations with different expressions but the same meaning. During the merging process, the method retains the source tracking information of each relation, establishing a complete lineage record to ensure the traceability of the knowledge graph. This ensures that the resulting resolved relation set eliminates internal conflicts while maintaining the integrity and diversity of knowledge.

[0145] In one possible embodiment, latent relationship mining is performed on the knowledge graph of breaking news events based on graph neural networks to identify the development stages and information gaps of breaking news events, thereby obtaining a core information dataset, including:

[0146] Based on the topological structure of the knowledge graph of emergencies, node embedding learning is performed using GraphSAGE graph neural network to obtain low-dimensional vector representations of nodes.

[0147] Specifically, this method employs a graph representation learning framework based on neighborhood sampling. Its core principle is to generate the embedding vector of the target node by iteratively aggregating the feature information of neighboring nodes. For example, this method can construct a hierarchical sampling strategy, sampling a fixed number of neighboring nodes in the knowledge graph for each target node to form a computational subgraph. In the feature aggregation stage, the method uses a mean aggregator to perform a weighted average of the feature vectors of the neighboring nodes, where the weights are calculated through an attention mechanism, focusing on neighboring nodes with higher semantic relevance to the target node. After multiple iterations, each node obtains an embedding vector representation containing its local graph structure information and semantic features. These low-dimensional vectors not only preserve the topological characteristics of the original graph structure but also capture complex semantic relationships between nodes, laying the foundation for subsequent discovery of potential relationships.

[0148] Based on the low-dimensional vector representation of nodes, the cosine similarity algorithm is used to calculate potential relationships, resulting in a candidate set of potential relationships.

[0149] Specifically, a multi-scale cosine similarity algorithm can be used: in addition to directly calculating the cosine value of the embedded vectors of node pairs, the transitive similarity of intermediate nodes in a one-hop path is introduced. The final similarity is defined as the geometric mean of the direct similarity and the transitive similarity. When the total value exceeds a preset threshold, potential relation candidates are generated. For example, the direct similarity between the nodes "substation fault" and "hospital power outage" is 0.82, and the transitive similarity through the "power supply line" node is 0.79. The combined value of 0.81 exceeds the threshold of 0.8 and is therefore included in the candidate set.

[0150] The set of potential relation candidates is filtered for false relations by a semantic rationality validator to obtain the set of valid potential relations.

[0151] Specifically, this method filters out false relationships using a semantic rationality validator. The validator is designed based on the semantic constraints and domain logic rules of the knowledge graph, and its core principle is to evaluate the rationality of potential relationships through multi-dimensional evidence. For example, this method constructs a semantic rule base containing verification rules such as type constraints, attribute constraints, and relationship constraints. For instance, for a "person-location" type relationship, it verifies whether it conforms to geographical common sense and temporal logic; for a "organization-event" type relationship, it verifies whether it conforms to the organization's functional scope and historical behavioral patterns. This method also utilizes a pre-trained language model to calculate the semantic rationality score of the relationship description, and determines whether there is a logical connection between the entities at both ends of the relationship by analyzing their contextual information. Furthermore, this method can introduce external knowledge bases as verification references, further confirming the credibility of potential relationships by comparing relevant information from authoritative knowledge sources.

[0152] Based on the connectivity characteristics of the knowledge graph of emergencies, the subgraph structure is analyzed and processed using a graph density analysis algorithm to obtain subgraph density distribution data.

[0153] Specifically, this method utilizes graph density analysis algorithms based on connected component analysis and density calculation in graph theory to discover information distribution characteristics within knowledge graphs. It employs a community detection algorithm to divide the knowledge graph into multiple semantic subgraphs and then calculates the relative density index of each subgraph. The density calculation uses a comprehensive evaluation model based on node degree and clustering coefficients, considering not only the tightness of connections within subgraphs but also the sparsity of connections between subgraphs. Furthermore, this method can incorporate temporal dimension analysis, identifying trends in information aggregation and diffusion by comparing subgraph density changes across different time slices. These density indices collectively constitute distribution data reflecting the structural characteristics of the knowledge graph, providing a basis for identifying information-weak areas.

[0154] Based on the subgraph density distribution data, a sparse region detection process is performed using an information-weak region identification algorithm to obtain the location information of the information-weak region.

[0155] Specifically, the core principle of the identification algorithm is to discover information-deficient regions in the knowledge graph through anomaly detection and pattern recognition techniques. Specifically, this method employs a detection mechanism based on the Local Outlier Factor (LOF) to calculate the density anomaly score for each subgraph region. Regions with density significantly lower than the neighborhood average are marked as potential information-deficient areas. Simultaneously, this method combines semantic importance analysis, considering the criticality of different regions in event descriptions, and giving higher attention to semantically important but information-sparse regions. Furthermore, this method establishes a prediction model based on historical patterns, predicting potential knowledge gaps by analyzing the development patterns of similar events, thus achieving proactive identification of information-deficient areas.

[0156] By combining time series features, an event stage determination process is performed using a development stage classifier to obtain the development stage label.

[0157] Specifically, the development stage classifier is designed based on the principles of temporal pattern recognition and multi-feature fusion, dividing development stages by analyzing the dynamic characteristics of event development. For example, this method can extract multiple time-series indicators, including node growth rate, relationship density change rate, and information propagation speed, to construct a multi-dimensional feature vector. The classifier employs an attention-based recurrent neural network architecture, capable of capturing the differences in the importance of features at different time steps. During training, the method uses labeled historical event data to learn the mapping relationship from feature sequences to development stage labels. To handle the uncertainty of sudden events, the classifier also introduces an uncertainty estimation mechanism, outputting the confidence level of each prediction result; when the confidence level is low, a manual review process is initiated.

[0158] Based on the location information and development stage labels of information-weak areas, key information is extracted and processed to obtain a core information dataset.

[0159] Specifically, the core principle of this processing method is multi-objective optimization based on information value and timeliness, aiming to filter out the most valuable core content from massive amounts of information. For example, this method establishes a multi-dimensional evaluation system, assessing the value score of each information unit from four dimensions: novelty, importance, reliability, and completeness. Novelty is calculated by comparing content with existing knowledge graphs; importance is evaluated based on entity centrality and relational weights; reliability considers source authority and content consistency; and completeness analyzes the degree to which the information contributes to filling knowledge gaps. This method also dynamically adjusts the evaluation weights according to the stage of event development; for example, it emphasizes the timeliness of information during the outbreak phase and focuses on the depth and accuracy of information during the stalemate phase. Through multiple rounds of screening and optimization, an information dataset containing the core elements and development context of the event is ultimately formed, providing precise content support for subsequent collaborative dissemination.

[0160] In one possible embodiment, the core information dataset is input into a pre-trained reinforcement learning model to generate a set of cooperative propagation instructions, including:

[0161] Semantic feature extraction is performed based on the core information dataset to obtain a state vector representing the event state.

[0162] Specifically, a multi-level feature fusion architecture can be adopted. Graph convolutional networks are used to extract structural features of the knowledge graph, capturing the topological relationships between entities. Attention mechanisms are used to weight and encode text descriptions, highlighting key information. Time series analysis methods are introduced to extract dynamic features of event development. For example, this method concatenates feature vectors of different types and then performs dimensionality reduction and normalization through fully connected layers. For instance, for knowledge graph structural features, the method calculates the degree centrality, betweenness centrality, and eigenvector centrality of each node to form a structural feature vector; for text features, the BERT model is used to obtain semantic representations; and for time series features, indicators of recent event development trends are extracted. These features, after weighted fusion, form a low-dimensional dense vector that comprehensively represents the current state of the event, providing comprehensive state information for subsequent decision-making.

[0163] The state vector is input into the policy network of the pre-built reinforcement learning model, and action decisions are made through forward propagation to obtain the original action parameters, which include the platform weight vector and the time offset.

[0164] Specifically, this policy network employs a deep neural network architecture, comprising multiple fully connected layers and activation function layers. The network input layer receives the state vector, which undergoes feature transformation through two hidden layers. The final output layer uses different activation functions to generate corresponding action parameters. Specifically, for the platform weight vector, the output layer uses a softmax function to ensure the sum of all platform weights is 1; for the time offset, a sigmoid function is used to limit it to the range of 0 to 1, and then it is mapped to a specific time value according to actual needs. During training, the policy network updates parameters using a proximal policy optimization algorithm, and ensures training stability through importance sampling and gradient pruning. Furthermore, this method can also design a value network to evaluate state values, providing a baseline value for policy updates, reducing variance, and accelerating convergence.

[0165] The original motion parameters are processed by a motion constraint to verify their feasibility, resulting in feasible motion parameters that meet the platform's capabilities and time constraints.

[0166] Specifically, the action constraint mechanism establishes a parameter optimization model under multiple constraints based on the principle of solving constraint satisfaction problems. For example, this method constructs a set of platform capability constraints, including content format restrictions, posting frequency restrictions, and audience characteristics for each platform; simultaneously, it establishes a set of time constraints, considering factors such as event urgency, audience activity periods, and platform content lifecycle. For the platform weight vector, the constraint mechanism checks whether it conforms to the actual carrying capacity of each platform and solves for the optimal weight allocation using a linear programming algorithm; for the time offset, the constraint mechanism verifies whether it is within a reasonable time window and eliminates conflicting periods using a time inference algorithm. When a constraint violation is detected, the constraint mechanism uses a gradient projection-based correction algorithm to project infeasible parameters into the feasible region while preserving the original decision intent to the greatest extent possible.

[0167] Based on the key entities and relational paths in the core information dataset, a pre-trained language generation model is used to generate text sequences, resulting in semantically coherent core statements.

[0168] Specifically, this method employs a Transformer-based sequence-to-sequence model, using an encoder-decoder architecture to generate text. The encoder encodes key entities and relational paths into context vectors, while the decoder generates text sequences progressively based on an attention mechanism. Furthermore, this method introduces factual consistency constraints on top of standard language models, ensuring that the generated content aligns with known facts through knowledge graph retrieval. For example, during generation, the model retrieves relevant information from the knowledge graph in real time to verify the accuracy of the generated content; simultaneously, a diversity mechanism is introduced, using kernel sampling techniques to avoid generating overly conservative statements. The method also includes a post-processing module to optimize the fluency and control the length of the generated text, ensuring that the core statements output are both accurate and meet dissemination requirements.

[0169] Based on the platform weight vector in the action parameters, a multi-platform preference decision-making process is performed through a platform selection algorithm to obtain a target platform preference list.

[0170] Specifically, the platform selection algorithm is based on multi-objective optimization theory, comprehensively considering multiple factors such as dissemination effect, resource consumption, and risk control. This method establishes a platform evaluation index system, including dimensions such as user coverage, content matching degree, dissemination speed, and interactivity. For each candidate platform, the algorithm calculates its comprehensive utility value and then performs a weighted ranking based on the platform weight vector. For example, this method employs an objective weighting method based on entropy weighting, automatically determining weights according to the degree of data variation for each indicator to avoid subjective bias. Simultaneously, synergy analysis is introduced, considering the complementarity of different platform combinations. A greedy algorithm selects the optimal platform combination, and the output target platform preference list not only includes platform priority ranking but also indicates the advantages and applicable content types of each platform.

[0171] Based on the time offset in the action parameters, the release time window is calculated using a time-series scheduling algorithm to obtain the release time schedule.

[0172] Specifically, the time-series scheduling algorithm is based on job shop scheduling theory, treating each platform's publishing tasks as jobs to be scheduled and time resources as machine resources. This method establishes a time-series optimization model with maximizing dissemination effect as the objective function and platform characteristics, audience habits, and content relevance as constraints. For example, the algorithm can discretize continuous time into multiple time slots and then use a genetic algorithm to solve for the optimal publishing sequence. For each publishing task, the algorithm calculates its earliest start time, latest end time, and ideal publishing time, combining these with time offsets to determine the specific publishing time. The method also designs a conflict detection and resolution mechanism; when a time conflict is detected, the publishing plan is adjusted through a rescheduling algorithm to ensure the coordination of publishing times across platforms.

[0173] The core statements, target platform preference list, and release schedule are structured and integrated to obtain a collaborative propagation instruction set.

[0174] Specifically, the integration process is based on information architecture design principles, establishing a unified instruction data model. For example, this method can define a JSON schema for the instruction set, comprising four parts: metadata, content, distribution, and scheduling. The metadata area records the instruction version, generation time, and validity period; the content area stores the core statements and their variant versions; the distribution area defines the specific publishing parameters for each platform; and the scheduling area details the time requirements for each publishing task. During the integration process, this method performs consistency checks to ensure that the information in each part is coordinated. For example, it verifies whether the content format meets platform requirements and checks whether the time arrangement is reasonable. The resulting collaboratively propagated instruction set adopts a machine-readable format, which facilitates automated execution while retaining sufficient flexibility to adapt to the characteristics of different platforms.

[0175] In summary, the multi-platform collaborative dissemination method for breaking news events provided in this application transforms the original datasets from multiple platforms into standardized information containing entities, relationships, source information, and timestamps through multimodal feature analysis, establishing a unified data foundation. Based on this standardized information, it dynamically integrates it into a pre-constructed event knowledge graph through event association determination and knowledge fusion techniques, achieving structured integration of information. Furthermore, it utilizes graph neural networks to mine potential relationships within the knowledge graph, identifying event development stages and information gaps to form a core information dataset. A pre-trained reinforcement learning model generates a collaborative dissemination instruction set containing core statements, target platform preferences, and release sequence, guiding coordinated and consistent information dissemination across multiple platforms. This technical solution enables dynamic, structured integration and correlation accumulation of fragmented information from multiple sources. By constructing a semantically coherent knowledge graph of breaking news events, it eliminates cognitive ambiguity and spatiotemporal discontinuities in cross-platform information dissemination, improving the consistency of public and decision-making bodies' understanding of the evolutionary patterns of breaking news events.

[0176] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0177] Based on the same inventive concept, this application also provides a multi-platform collaborative dissemination device 10 for implementing the multi-platform collaborative dissemination method for breaking news events as described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the multi-platform collaborative dissemination device 10 for breaking news events provided below can be found in the limitations of the multi-platform collaborative dissemination method for breaking news events described above, and will not be repeated here.

[0178] In one exemplary embodiment, such as Figure 2 As shown, a multi-platform collaborative dissemination device 10 for breaking news events is provided to implement the methods in the above-described method embodiments. The device includes:

[0179] Feature parsing module 11 is used to perform multimodal feature parsing on the acquired multi-platform raw dataset of breaking news events to obtain standardized information;

[0180] The graph fusion module 12 is used to perform event association and knowledge fusion on standardized information, integrating entities and corresponding relationships in the standardized information into a pre-built event knowledge graph to obtain an emergency event knowledge graph.

[0181] Relationship mining module 13 is used to mine potential relationships in the knowledge graph of sudden events based on graph neural networks, identify the development stages and information gaps of sudden news events, and obtain core information datasets.

[0182] The propagation instruction generation module 14 is used to input the core information dataset into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set; wherein, the collaborative propagation instruction set is used to instruct the target platform to release relevant information about breaking news events.

[0183] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a multi-platform collaborative dissemination method for breaking news events as described above.

[0184] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0185] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0186] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A multi-platform collaborative dissemination method for breaking news events, characterized in that, The method includes: Multimodal feature analysis was performed on the acquired raw datasets of breaking news events from multiple platforms to obtain standardized information; The standardized information is subjected to event association and knowledge fusion, and the entities and corresponding relationships in the standardized information are integrated into a pre-constructed event knowledge graph to obtain an emergency event knowledge graph. Based on graph neural networks, potential relationship mining is performed on the knowledge graph of the sudden event to identify the development stage and information gaps of the sudden news event and obtain the core information dataset. The core information dataset is input into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set; wherein, the collaborative propagation instruction set is used to instruct the target platform to release relevant information about the breaking news event.

2. The method according to claim 1, characterized in that, The process of performing multimodal feature analysis on the acquired multi-platform raw dataset of breaking news events to obtain standardized information includes: Raw data from multiple platforms is obtained through news source API interfaces and official data subscription services. The raw data from multiple platforms is then cleaned and formatted to obtain preprocessed heterogeneous data. The preprocessed heterogeneous data is injected into a distributed message queue, and the message queue is used for traffic shaping and data buffering to obtain a stable data stream. Text data is extracted from the stable data stream, and entity extraction is performed using a BERT-based named entity recognition model to obtain an entity set, which includes people, organizations, geographical locations, and time points. Based on the text data, semantic relationship recognition is performed through a relationship extraction model to obtain a set of relation triples describing the relationships between entities. For the image data and / or video data in the stable data stream, target detection and scene classification are performed using a computer vision model to obtain the corresponding visual description text; The standardized information is obtained by structurally encapsulating the entity set, the relation triplet set, and the visual description text.

3. The method according to claim 2, characterized in that, The step involves extracting text data from the stable data stream and performing entity extraction processing using a BERT-based named entity recognition model to obtain an entity set, including: The text data is segmented and stop words are removed to obtain a preprocessed text sequence; The preprocessed text sequence is input into the fine-tuned BERT model, and entity boundary recognition and type classification are performed through the sequence labeling layer to obtain preliminary entity labeling results. Based on the preliminary entity annotation results, the entity normalization algorithm is used to resolve the references to obtain a standardized set of entities with unified names. The standard entity set is then validated using a context semantic verification module to obtain a validated entity set. The verification entity set is classified and stored according to the preset entity type to obtain the entity set.

4. The method according to claim 1, characterized in that, The process of associating and fusing events with the standardized information, integrating the entities and corresponding relationships in the standardized information into a pre-constructed event knowledge graph, yields a knowledge graph of emerging events, including: The association degree data between the entity set and the relationship set and the existing event nodes in the event knowledge graph is calculated using an event association determination algorithm. Based on the correlation data, an event attribution determination is made using a dynamic threshold comparator to determine the correspondence between standardized information and event nodes; Based on the aforementioned correspondence, a unified set of entity identifiers is obtained by performing referential disambiguation processing using an entity alignment algorithm. The entity identifier set is processed by a conflict detection module to check the relationship consistency and identify relationship pairs with semantic conflicts. Based on the authority of the source and the confidence of the information, a set of resolved relationships is obtained by using an evidence weighting algorithm to resolve conflict relationships. The set of resolved relationships is integrated into the event knowledge graph, and the structure is optimized through the graph update engine to obtain the event knowledge graph.

5. The method according to claim 4, characterized in that, The process of resolving conflicts based on source authority and information confidence, using an evidence-weighted algorithm, yields a set of resolved relationships, including: Based on the existing set of relations in the event knowledge graph and the new set of relations in the standardized information, relation consistency comparison is performed through semantic conflict detection to identify a set of conflicting relation pairs with different semantic representations among the same entity pairs. Extract the source authority from the source information of the standardized information, and convert the source authority into standardized authority weights through an authority quantifier to obtain the authority quantification value corresponding to each relation; Relationship confidence scores are extracted from the information confidence scores of the standardized information, and the relationship confidence scores are normalized using a confidence calibrator to obtain the standard confidence value corresponding to each relationship; Based on the aforementioned authoritative metric and standard confidence value, the optimal relation identifier for each conflicting relation pair is obtained through relation optimization decision processing using the following evidence-weighted formula: in, Indicates the first A relationship, This represents the set of conflicting pairs. Representing relations The authority of the source mentioned above, Representing relations The confidence level of the information mentioned above. This is the authoritative metric value. The standard confidence level value; Based on the optimal relation identifier, the optimal relation is selected from the set of conflicting relation pairs using a relation selector to obtain a preliminary set of resolved relations; For the relationships in the set of conflicting relationships that were not selected as the optimal relationship, metadata annotation and temporary storage are performed through a temporary storage manager to obtain a set of conflicting relationships to be re-evaluated. Based on the standardized information subsequently received, the incremental evaluator supplements the evidence and re-scores the set of conflict relations to be re-evaluated, resulting in a supplementary set of resolved relations. The initial set of resolution relationships and the supplementary set of resolution relationships are merged and deduplicated to obtain the set of resolution relationships.

6. The method according to claim 1, characterized in that, The method involves mining potential relationships within the knowledge graph of the breaking news event based on a graph neural network to identify the development stages and information gaps of the breaking news event, thereby obtaining a core information dataset, including: Based on the topological structure of the knowledge graph of the emergencies, node embedding learning is performed through GraphSAGE graph neural network to obtain low-dimensional vector representations of nodes; Based on the low-dimensional vector representation of the nodes, a potential relationship candidate set is obtained by performing potential relationship calculation using the cosine similarity algorithm. The set of potential relation candidates is processed by a semantic rationality validator to filter out false relations, resulting in a set of verified potential relations. Based on the connectivity characteristics of the aforementioned knowledge graph of emergencies, subgraph structure analysis is performed using a graph density analysis algorithm to obtain subgraph density distribution data; Based on the subgraph density distribution data, sparse region detection processing is performed using an information weakness region identification algorithm to obtain the location information of the information weakness region; By combining time series features, an event stage determination process is performed using a development stage classifier to obtain the label of the development stage; Based on the location information of the information-weak areas and the labels of the development stages, key information is extracted to obtain the core information dataset.

7. The method according to claim 1, characterized in that, The step of inputting the core information dataset into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set includes: Semantic feature extraction is performed on the core information dataset to obtain a state vector representing the event state. The state vector is input into the policy network of the pre-built reinforcement learning model, and action decisions are made through forward propagation to obtain the original action parameters, which include the platform weight vector and the time offset. The original action parameters are processed by an action constraint to verify their feasibility, thereby obtaining feasible action parameters that meet the platform's capabilities and time constraints. Based on the key entities and relational paths in the core information dataset, a pre-trained language generation model is used to generate text sequences to obtain semantically coherent core declarative sentences. Based on the platform weight vector in the action parameters, a multi-platform preference decision-making process is performed using a platform selection algorithm to obtain a target platform preference list; Based on the time offset in the action parameters, the release time window is calculated using a timing scheduling algorithm to obtain the release timing plan; The core statement, the target platform preference list, and the release timing plan are structurally integrated to obtain the collaborative propagation instruction set.

8. A multi-platform collaborative dissemination device for breaking news events, characterized in that, The device includes: The feature parsing module is used to perform multimodal feature parsing on the acquired raw datasets of breaking news events from multiple platforms to obtain standardized information; The graph fusion module is used to perform event association and knowledge fusion on the standardized information, integrating the entities and corresponding relationships in the standardized information into a pre-constructed event knowledge graph to obtain an emergency event knowledge graph. The relationship mining module is used to perform potential relationship mining on the knowledge graph of the sudden event based on the graph neural network, identify the development stage and information weakness area of ​​the sudden news event, and obtain the core information dataset. The propagation instruction generation module is used to input the core information dataset into a pre-trained reinforcement learning model to generate a collaborative propagation instruction set; wherein, the collaborative propagation instruction set is used to instruct the target platform to release relevant information about the breaking news event.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images