Self-evolution multi-agent event semantic understanding system and method based on dynamic scalable model

Abstract

The application discloses a self-evolution multi-agent event semantic understanding system and method based on a dynamic scalable model, which comprises a multi-agent cooperation framework modeling module, a demand understanding and task planning module, a capability verification module, a model dynamic expansion module, a task execution module and a completion module connected in sequence, wherein the multi-agent cooperation framework modeling module comprises a central control type multi-agent, a task execution type multi-agent and a model expansion type multi-agent. The method focuses on a news event semantic understanding scene, constructs a dynamic scalable system architecture through a multi-agent cooperation mechanism, introduces a capability verification link to optimize a task execution process, and realizes on-demand iteration of model capability by using an active learning framework, so that adaptive evolution of agent capability is realized, thereby solving the problems of model rigidity and poor adaptability of a traditional event semantic understanding system in a complex news event scene, and significantly improving the dynamic adaptability and accuracy of news event semantic understanding.

Description

Technical Field

[0001] This invention relates to the field of computer artificial intelligence technology, specifically to a self-evolving multi-agent event semantic understanding system and method based on a dynamic and scalable model. Background Technology

[0002] Semantic understanding of news events is a crucial task in the field of natural language processing. It aims to accurately extract key information from massive amounts of news text, understand the core elements of events, the relationships between events, and the development of events, thereby providing users with comprehensive and accurate interpretations of news events. In the era of information overload, news data is growing exponentially, with diverse sources and formats, including text, images, and videos. How to efficiently and accurately perform semantic understanding of news events to meet users' information needs in different scenarios has become a critical problem that urgently needs to be solved.

[0003] Traditional semantic understanding methods for news events are mostly based on single models or fixed procedures, which have many limitations when dealing with complex news events. On the one hand, single models have limited capabilities and struggle to handle the rich and diverse semantic expressions and complex semantic relationships within news events. For example, different news reports may describe the same event differently, including in terms of word choice, sentence structure, and emphasis; single models often fail to accurately identify these differences and extract key information. On the other hand, fixed procedures lack flexibility and adaptability, failing to dynamically adjust according to the characteristics of news events and user needs. For instance, when processing news events from different fields, the required analytical methods and tools may vary, and fixed procedures struggle to meet these diverse needs.

[0004] Currently, multi-agent systems are increasingly being applied in the field of natural language processing. Through collaboration and division of labor among multiple agents, the efficiency and accuracy of task processing can be improved. However, most existing multi-agent news event semantic understanding systems lack capability verification mechanisms, making it impossible to evaluate and dynamically adjust the capabilities of each agent in real time. When faced with complex news events, some agents may lack sufficient capabilities, causing the entire system to fail to complete the task accurately. For example, during event extraction, if existing event extraction models cannot identify certain specific types of events, the system cannot accurately extract relevant event information, thus affecting subsequent analysis and processing.

[0005] Furthermore, existing systems often lack effective model extension mechanisms when their capabilities are insufficient to meet task requirements. While some research attempts to extend models through methods such as transfer learning and incremental learning, these methods often require large amounts of labeled data and computational resources, and the extension process lacks specificity and flexibility. For example, when introducing new news event types, it is necessary to re-collect and label a large amount of relevant data and then retrain the entire model, which is not only time-consuming and labor-intensive but may also affect the performance of the original model. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0007] 1. Technical problems to be solved:

[0008] To address the contradiction between dynamic requirements and static tools, as well as the problems of low model expansion efficiency and lack of human-machine collaboration mechanisms mentioned above, this invention is proposed.

[0009] Therefore, the purpose of this invention is to provide a self-evolving multi-agent event semantic understanding system and method based on a dynamic and scalable model. This method introduces a capability verification step to evaluate the capabilities of each agent within the multi-agent system in real time. When the existing model's capabilities cannot meet the task requirements, a dynamic model expansion module is activated. Using an active learning framework, a dedicated model is iteratively trained through human-machine collaboration, ultimately forming a dedicated model with the corresponding capabilities. This improves the accuracy and adaptability of news event semantic understanding. This invention not only provides a new approach and method for news event semantic understanding but also has significant implications for promoting the application of natural language processing technology in the news field.

[0010] 2. Technical Solution:

[0011] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0012] It includes a multi-agent collaboration framework construction module, a requirement understanding and task planning module, a capability verification module, a model dynamic expansion module, a task execution module, and a completion module, which are connected in sequence.

[0013] The multi-agent collaboration framework construction module includes central control type multi-agents, task execution type multi-agents, and model extension type multi-agents;

[0014] The model dynamic expansion module includes, in sequence, a model expansion suggestion and data acquisition unit, a data pre-labeling unit, a model training and pseudo-label generation unit, a difficult example screening and manual verification unit, a strategy optimization and model iteration unit, and a model evaluation and tool library update unit.

[0015] The task execution module includes a data retrieval unit, a data preprocessing unit, an event extraction unit, a coreference resolution unit, a summary generation unit, and an event display unit connected in sequence.

[0016] The capability verification module is used to perform model coverage verification. and capability coverage Evaluate existing models in the model library; wherein, the model coverage The calculation formula is: ,in, It is an indicator function used to determine whether a condition is true or false. This represents the i-th model needed to complete the task. This represents the collection of all models in the existing model library. This represents the total number of all models required to complete the task; the capability coverage. The calculation formula is: ,in, It is an indicator function used to determine whether a condition is true or false. express The j-th capability of the model, This represents the set of existing capabilities of the corresponding model in the model library. express The total number of all capabilities required for the model to complete the task; and If the task requirements are met, the model is dynamically expanded; otherwise, the model's dynamic expansion module is triggered.

[0017] The difficult example screening and strategy optimization iteration unit is used to manually verify difficult examples in the pseudo-label data based on a confidence threshold σ, and dynamically adjust the confidence threshold σ for the next round of difficult example screening based on the results of the manual verification; the dynamic adjustment mechanism of the confidence threshold σ is as follows: ,in, For average information value density, To average annotation cost, For the improvement in model accuracy, To adjust the step size, , , These are the weighting coefficients.

[0018] As a preferred embodiment of the self-evolving multi-agent event semantic understanding system based on a dynamic scalable model of the present invention, the multi-agent collaborative framework construction module includes a central large model, a database, a tool library, a model library, and a multi-agent architecture. The central large model is a large language model with logical reasoning capabilities, including DeepSeek-v3, Qwen-3, or ERNIE-3.5.

[0019] As a preferred embodiment of the method for a self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to the present invention, the central control type multi-agent includes:

[0020] A demand-understanding agent is used to parse user instructions and intentions;

[0021] A task planning agent is used to formulate a task execution plan and break it down into a sequence of executable steps.

[0022] Capability verification agent is used to evaluate whether existing models in the model library meet the task requirements.

[0023] As a preferred embodiment of the self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to the present invention, the task execution multi-agent includes:

[0024] The performance evaluation agent, acting as a process controller, evaluates and controls the execution results of the data retrieval, data preprocessing, event extraction, coreference resolution, summary generation, and event display agents.

[0025] As a preferred embodiment of the self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to the present invention, the model extension class of multi-agents includes:

[0026] A data acquisition agent is used to call APIs to collect task-related data.

[0027] Pre-labeled agents are used to call UIE tools to pre-label data;

[0028] The model training agent is used to call the API to train models online.

[0029] A difficult example screening agent is used to screen high-value samples based on an active learning strategy.

[0030] A strategy optimization agent is used to evaluate labeled samples and optimize the hard example selection strategy.

[0031] As a method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model according to the present invention, the self-evolving multi-agent event semantic understanding system based on the dynamic scalable model includes the following steps:

[0032] S1: Construct a multi-agent collaboration framework, including defining the responsibilities of central control agents, task execution agents, and model extension agents, and establishing a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution - task completion";

[0033] S2: Receive user requests and plan tasks to form a task execution plan;

[0034] S3: Execution capability verification, determine whether the existing model meets the task requirements; if not, proceed to S4; if it does, proceed to S5.

[0035] S4: Perform dynamic model expansion, including data collection, pre-labeling, model training, pseudo-label generation, hard example selection, manual verification, strategy optimization, and model iteration;

[0036] S5: Perform tasks, including data retrieval, preprocessing, event extraction, coreference resolution, summary generation, and event display;

[0037] S6: Complete the task and output the event semantic understanding and visualization results.

[0038] As a method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model, in the task execution step, the execution results of each agent are verified by an effect evaluation agent. If the requirements are not met, the results are fed back to the corresponding agent for optimization.

[0039] As a method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model, the dynamic model expansion step adopts an active learning framework. It iteratively executes the process of "difficult example screening - manual verification - model update - policy optimization" until the model accuracy improves by less than 0.5% for two consecutive rounds, at which point training stops and the final model is updated to the model library.

[0040] 3. Beneficial effects:

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This self-evolving multi-agent event semantic understanding system and method based on a dynamic and scalable model:

[0043] 1. This invention constructs a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution" through a multi-agent collaborative framework, achieving three major breakthrough innovations in complex news event analysis: First, it pioneers a capability verification stage and a dynamic model expansion mechanism. Through the capability verification agent, the model resources of the tool library are evaluated in real time. When the existing model cannot meet the task requirements, a full-process model expansion system including data retrieval, pre-labeling, difficult example screening, and strategy optimization is automatically triggered. An active learning framework is used to realize dedicated model iterative training in human-machine collaboration, which significantly improves the system's adaptability to emerging event types.

[0044] 2. This invention constructs a multi-agent collaborative event analysis system. Through the organic cooperation of agents such as data retrieval, event extraction, coreference resolution, and event display, it realizes the full-link automated processing from raw news to structured semantic representation and then to spatiotemporal visualization.

[0045] 3. This invention innovatively designs a dynamically optimized difficult example selection strategy. Through a strategy optimization agent, it continuously evaluates manually verified samples and provides feedback to optimize the selection strategy, effectively improving model training efficiency. This method effectively solves the pain points of traditional news analysis systems, such as model rigidity, poor domain transferability, and high manual annotation costs, providing an evolvable intelligent solution for dynamically evolving semantic understanding of international events. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0047] Figure 1 This is a schematic diagram of the process of a self-evolving multi-agent event semantic understanding system and method based on a dynamic scalable model according to the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] This invention is described in detail with reference to the schematic diagrams. When describing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0050] The orientation or positional relationship indicated in the terminology is based on the orientation or positional relationship shown in the accompanying drawings and is only for the convenience of describing the invention and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0051] The term "connection method" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0052] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0053] This invention provides a schematic diagram of the overall structure of an embodiment of a self-evolving multi-agent event semantic understanding system and method based on a dynamic scalable model, including:

[0054] Please see Figure 1 This includes the following steps:

[0055] The modules, including those connected in series, specifically include: multi-agent collaborative framework construction, requirement understanding and task planning, capability verification, dynamic model expansion, and task execution and completion.

[0056] The multi-agent collaboration framework includes electrically connected central control multi-agents, task execution multi-agents, and model extension multi-agents.

[0057] Dynamic model expansion includes model expansion suggestions and data collection for electrically connected series, data pre-labeling, model training and pseudo-label generation, hard example screening and manual verification, strategy optimization and model iteration, and model evaluation and tool library updates.

[0058] The task execution includes data retrieval in series with electrical connections, data preprocessing, event extraction, coreference resolution, summary generation, and event display.

[0059] It is worth noting that, specifically, the multi-agent collaborative framework is constructed as a complete architecture system including a central large model, a database, a tool library, a model library, and multiple agents. The central large model serves as the core, while DeepSeek-v3, Qwen-3, and ERNIE-3.5, which possess powerful logical reasoning capabilities, provide basic intelligent support for the entire framework.

[0060] Specifically, a method for self-evolving multi-agent event semantic understanding based on a dynamic and scalable model includes:

[0061] S1: The construction of a multi-agent collaborative framework first establishes a complete architecture. The database is used to store raw data such as text, images, and videos related to news events, providing basic data for subsequent analysis. The model library integrates various intelligent models related to event processing and analysis, such as event classification, event argument role recognition, and coreference resolution. The tool library is equipped with coding models, data statistical analysis models, browsers, and event information visualization tools to assist in task execution.

[0062] Then, the types and responsibilities of multi-agents are defined, including central control multi-agents, task execution multi-agents, and model extension multi-agents.

[0063] Central control multi-agent system: a demand understanding agent is set up to accurately understand user instructions and intentions; a task planning agent formulates detailed execution plans and decomposes tasks into a sequence of executable steps; and a capability verification agent is responsible for verifying whether the intelligent models in the model library can complete the task to ensure task feasibility.

[0064] Multi-agent task execution: The effect evaluation agent acts as the process controller to evaluate the execution results of agents such as data retrieval, data preprocessing, event extraction, coreference resolution, summary generation, and event display. Through a hierarchical collaborative structure, it controls the effect of each link and collaboratively completes the semantic understanding task of news events.

[0065] Model extension multi-agent system: The data acquisition agent calls API tools to collect data online; the pre-labeling agent uses UIE tools to pre-label the data; the model training agent calls API tools to train the model online; the hard example screening agent uses actively learned samples to screen the model and select high-value labeled samples; the strategy optimization agent evaluates the labeled samples after manual verification and optimizes the hard example screening strategy.

[0066] Finally, a closed-loop workflow of "requirements understanding - task planning - capability verification - task execution - task completion" is set up to enable each module and multiple intelligent agents to work together to achieve intelligent semantic understanding and visualization of news events.

[0067] S2: Demand understanding and task planning. The multi-agent system receives the user's analysis request for news events. The task planning agent decomposes and plans the tasks based on the request, clarifies the specific task types and objectives to be completed, and forms a preliminary task execution plan.

[0068] S3: Capability verification. The capability verification agent conducts a comprehensive evaluation and verification of the model resources in the tool library based on the task plan, and determines whether the existing model capabilities can meet the task execution requirements. If they can, it directly enters the task execution stage of S5; if they cannot, it enters the dynamic model expansion stage of S4.

[0069] S4: Dynamic Model Expansion

[0070] S4.1: Model expansion suggestions and data collection. The capability verification agent provides model expansion suggestions, clarifying the direction of model capabilities that need to be supplemented or expanded. The data retrieval agent collects task-related datasets from public data sources, professional databases and other channels based on the suggestions.

[0071] S4.2: Data pre-labeling. The data pre-labeling agent performs the first round of pre-labeling on the small-scale dataset collected. The labeling content includes key information related to the task, such as event entities and relationships, to provide basic labeled data for subsequent model training.

[0072] S4.3: Model training and pseudo-label generation. The model training agent calls the corresponding algorithm online and uses the first round of pre-labeled data to train the model. After training, the trained model is used to predict and label the remaining large-scale data to obtain pseudo-label data.

[0073] S4.4: Difficult example screening and manual verification. The difficult example screening agent calls the active learning sample selection strategy to screen out high-value samples from pseudo-label data and submit them to manual verification. The manual verification personnel accurately label and correct the samples.

[0074] S4.5: Strategy optimization and model iteration. For samples that have been manually verified, on the one hand, they enter the next round of model iteration training, and the model training agent uses the updated labeled data to further optimize the model; on the other hand, the strategy optimization agent evaluates the verified samples and optimizes the next round of hard example selection strategy based on the evaluation results to improve the accuracy and effectiveness of sample selection.

[0075] S4.6: Model evaluation and tool library update. After the model training is completed, the model is added to the model library for use by multiple agents in the subsequent event analysis task execution process.

[0076] S5: Task Execution:

[0077] S5.1: Data retrieval. The data retrieval agent retrieves task-related data from the data source according to the task plan, providing a data foundation for subsequent processing.

[0078] S5.2: Data preprocessing. The data preprocessing agent performs preprocessing operations such as cleaning, transformation, and normalization on the retrieved data to remove noisy data, unify data format, and improve data quality.

[0079] S5.3: Event Extraction. The event extraction agent identifies and extracts information related to news events from preprocessed data, including event type, event trigger words, event elements, etc.

[0080] S5.4: Coreference resolution. The coreference resolution agent performs coreference resolution processing on the extracted event information, identifies information with different expressions but pointing to the same event, and integrates the scattered information into a complete event description.

[0081] S5.5: Summary generation. The summary generation agent generates concise and clear news event summaries based on the results of event extraction and coreference resolution, highlighting the key information of the event.

[0082] S5.6: Event Display. The event display agent visualizes the geographical information related to news events, such as the location of the event, on a map to help users understand the spatial distribution of the event more intuitively.

[0083] S6: Task completed. After the task is completed, the event semantic understanding and visualization results can be saved through the one-click download function of the geographic information visualization platform to form an event overview map.

[0084] Example 1:

[0085] Step 1: Building a Multi-Agent Collaboration Framework

[0086] The multi-agent collaboration framework mainly includes modules such as a central large model, database, tool library, model library, and multi-agents. It realizes intelligent semantic understanding and visualization of news events through a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution - task completion".

[0087] The central large-scale model can employ powerful logical reasoning capabilities such as DeepSeek-v3, Qwen-3, and ERNIE-3.5. The database includes raw data related to news events, such as text, images, and videos. The model library includes intelligent models related to event processing and analysis, such as event classification models, event argument role recognition models, fine-grained event entity classification models, intra-document event coreference resolution models, cross-document event coreference resolution models, and image scene recognition models. The tool library includes encoding models, data statistical analysis models, browsers, and event information visualization tools. Multi-agent systems are categorized into three main types: central control multi-agent systems, task execution multi-agent systems, and model extension multi-agent systems. The specific roles and responsibilities of each type of agent are as follows:

[0088] (1) Central control type multi-agent: including demand understanding agent, task planning agent and capability verification agent. Among them, the demand understanding agent is responsible for understanding the user's instructions and intention recognition; the task planning agent is responsible for formulating a detailed execution plan and decomposing the task into a sequence of executable steps; the capability verification agent is responsible for verifying whether the capabilities of the intelligent models in the current model library can complete the task.

[0089] (2) Task execution multi-agent: including effect evaluation agent, data retrieval agent, data preprocessing agent, event extraction agent, coreference resolution agent, summary generation agent, and event display agent. The task execution multi-agent is responsible for calling resources in the database, tool library, and model library to collaboratively carry out news event semantic understanding tasks. The collaboration structure is hierarchical, that is, the effect evaluation agent acts as a process controller, evaluating the execution results of other agents and controlling the effect of each stage of task execution.

[0090] (3) Model extension multi-agent class: including data acquisition agent, pre-labeling agent, model training agent, hard example selection agent, and policy optimization agent. Among them, the data acquisition agent is responsible for calling API tools to collect data online, the pre-labeling agent is responsible for calling UIE tools to pre-label the data, the model training agent is responsible for calling API tools to train the model online, the hard example selection agent is responsible for calling the active learning sample selection model to select high-value labeled samples, and the policy optimization agent is responsible for evaluating the labeled samples after manual verification and optimizing the hard example selection strategy.

[0091] Step Two: Needs Understanding and Task Planning

[0092] After a user submits an analysis request regarding a news event, the request understanding agent accurately identifies and interprets the user's intent to clarify the task requirements. Based on this request, the task planning agent breaks down and plans the task, defining the specific task types and objectives to be completed, and forming a preliminary task execution plan. In this plan, data retrieval, event extraction, coreference resolution, summary generation, and event display are five fixed steps. The task planning agent can add other execution stages based on specific needs.

[0093] Step 3: Capability Verification

[0094] The capability verification agent comprehensively evaluates and verifies the model resources in the model library based on the task plan. It determines whether the existing model capabilities meet the task execution requirements. If they do, the task execution phase proceeds directly; otherwise, the model dynamic expansion phase begins. The core principle of capability verification is to map task requirements and model capabilities to a computable index space through mathematical modeling. The evaluation focuses on two dimensions: model coverage and capability coverage. The key evaluation dimensions and corresponding quantification formulas are as follows:

[0095] 1. Model Coverage Ratio (MCR)

[0096] To measure the coverage of existing models in the model library for news event analysis tasks:

[0097]

[0098] in, It is an indicator function used to determine whether a condition is true or false; This represents the i-th model (such as event classification, argument recognition, etc.) needed to complete the task. This represents the collection of all models in the existing model library; This represents the total number of all models required to complete the task.

[0099] like If the model is deemed to have sufficient coverage, the dynamic model expansion module will be triggered.

[0100] 2. Ability Coverage Ratio (ACR)

[0101] To measure the extent to which the capabilities of existing models in the model library cover the requirements of news event analysis tasks, for the models The formula for calculating its coverage is as follows:

[0102]

[0103] in, It is an indicator function used to determine whether a condition is true or false; express The j-th capability of the model (such as the ability of the event type recognition model to identify economic events, sports events, etc.); This represents the set of existing capabilities of the corresponding model in the model library; express The total number of all capabilities required for the model to complete the task.

[0104] like Determined as The model's capabilities are sufficiently covered; otherwise, the model's dynamic expansion modules will be triggered.

[0105] Step 4: Dynamic Model Expansion

[0106] After the model dynamic expansion module is triggered, the capability verification agent provides model expansion suggestions based on the model coverage and capability coverage evaluation results, clarifying the direction of model capabilities that need to be supplemented or expanded. Then, the active learning framework is started to carry out human-machine collaborative model training. The specific process is as follows:

[0107] 1. Data Collection

[0108] Based on model capability expansion suggestions, the data retrieval agent collects task-related datasets from public data sources, professional databases, and other channels.

[0109] 2. Data pre-labeling

[0110] First, the data pre-labeling agent randomly samples the collected dataset at a ratio of 2:8, with 20% of the dataset serving as a small-scale dataset and 80% as a large-scale dataset. Then, the UIE model is invoked to perform the first round of pre-labeling on the small-scale dataset, including annotations of key task-related information such as event entities and relationships, thus obtaining pre-labeled corpus.

[0111] 3. Model Training

[0112] The model training agent, based on the model capability expansion suggestions proposed by the capability verification agent, automatically selects algorithms through a pre-defined algorithm dispatch API. This API employs a multi-dimensional decision-making mechanism, comprehensively considering model architecture type (e.g., Transformer / CNN), computational resource constraints (GPU memory capacity, number of parallel computing nodes), and task performance metrics (accuracy threshold, inference latency requirements), dynamically matching the optimal training algorithm from the algorithm resource pool. After algorithm selection, the model training agent initiates online algorithm service calls via standardized communication protocols (e.g., gRPC or RESTful), loads the first round of pre-annotated corpus, sets model training hyperparameters, and trains the model to obtain the initial model. ,in These are the initial model parameters.

[0113] 4. Pseudo-tag generation

[0114] After the first round of model training is completed, the trained model is used to perform batch prediction and labeling of the remaining large-scale unlabeled data to obtain pseudo-label data.

[0115] 5. Difficult example screening and manual verification

[0116] The difficult example selection agent is based on an active learning framework and employs an uncertainty selection strategy to identify difficult examples. Samples with a confidence level below the threshold σ=0.8 are selected as difficult examples. The selected difficult examples are then pushed to a human verification platform via a standardized annotation interface, where verifiers check the accuracy and coverage of the annotations.

[0117] 6. Model update and optimization

[0118] After verification, the verified labeled data is sent back to the training data pool, and the model training agent uses an incremental learning strategy to update the model parameters.

[0119] 7. Optimization of screening strategy

[0120] The strategy optimization agent dynamically adjusts the threshold σ of the next round of hard example selection strategy based on the statistical characteristics of the validated samples and the updated model performance, ensuring that the hard examples selected in each round meet the following criteria: high value (significantly contributing to model performance improvement) and low cost (high annotation quality, reducing manual correction). The specific optimization process consists of the following steps:

[0121] 7.1 Evaluation of Validated Samples

[0122] The policy optimization agent quantitatively evaluates the validated samples from the following dimensions to provide data support for policy adjustment:

[0123] (1) Information Value Density (IVD)

[0124] The formula for evaluating the potential contribution of samples to improving model performance is as follows:

[0125]

[0126] in:

[0127] The prediction entropy of sample i reflects the model uncertainty.

[0128]

[0129] Let C be the input sample and C be the set of categories. Predict probabilities for the model;

[0130] Sample gradient norm, measures the model's sensitivity to samples:

[0131]

[0132] L is the loss function. For model parameters, For the input sample, These are genuine labels (after manual verification).

[0133] (2) Annotation Quality Index (AQI)

[0134] The cost of annotation is measured by the number of corrections made to the validated samples.

[0135]

[0136] in:

[0137] The number of times the label for sample i needs to be corrected by the verification personnel;

[0138] : The maximum number of corrections for all samples.

[0139] The smaller the value, the higher the quality of the sample annotation.

[0140] 7.2 Dynamic Strategy Optimization Mechanism

[0141] Based on the statistical characteristics of the validated samples and the performance of the updated model, the threshold σ is updated according to the following rules:

[0142]

[0143] in:

[0144] The average information value density of the current round of difficult examples;

[0145] : The average annotation cost of the current round of hard example sets;

[0146] The improvement in model accuracy on the current round of validation data;

[0147] Adjust the step size (e.g., 0.02);

[0148] , , Weighting coefficients (e.g.) , , ).

[0149] like A high σ indicates that the current difficult examples have a significant impact on model improvement, and σ can be appropriately reduced to include more high-value samples; if A high σ indicates low quality of current hard example annotations, requiring an increase in σ to filter out low-quality samples; if The small value of σ indicates that the current difficult examples have limited impact on the model, and σ needs to be increased to reduce invalid screening.

[0150] To ensure the reasonableness of the threshold and avoid it being too low or too high, the range of values ​​for σ is limited:

[0151]

[0152] 8. Model library update

[0153] The model training process is stopped when the model accuracy improves by less than 0.5% for two consecutive rounds. The model is then added to the model library for use by multiple agents in subsequent event analysis tasks.

[0154] Step 5: Task Execution

[0155] During the task execution phase, multiple agents, including the effect evaluation agent, data retrieval agent, data preprocessing agent, event extraction agent, coreference resolution agent, summary generation agent, and event display agent, collaborate to perform the news event semantic understanding task by calling resources from the database, tool library, and model library. The collaboration structure is hierarchical, with the effect evaluation agent acting as a process controller to evaluate the execution results of other agents and control the effectiveness of each stage of task execution. The specific task execution process is as follows:

[0156] 1. Data retrieval

[0157] The data retrieval agent, based on a structured retrieval instruction set (including metadata such as data type, domain keywords, time range, and quality constraints) generated by the task planning agent, collects multimodal data related to events, including text, images, and videos, from distributed data sources (such as news reports, government gazettes, and social media) through a multimodal retrieval engine. The retrieval results are transmitted to the performance evaluation agent via a standardized data interface (such as JSON Schema). This module performs multi-dimensional quality checks, verifying whether the news time, topic, and fields are relevant to the event. If the task requirements are met, the next step is executed; otherwise, the evaluation feedback is returned to the data retrieval agent, and the data retrieval is restarted.

[0158] 2. Data Preprocessing

[0159] The data preprocessing agent performs preprocessing operations such as cleaning, transformation, and deduplication on the retrieved multimodal data to remove noise, standardize data format, and improve data quality. The preprocessing results are then pushed to the performance evaluation module via API for quality verification to determine if they meet the requirements of subsequent tasks. For data that does not meet the standards, specific quality improvement suggestions are fed back to the data preprocessing agent for optimization and adjustment to ensure that the data quality meets the task requirements.

[0160] 3. Event Extraction

[0161] The event extraction agent invokes event extraction-related models from the model library, such as the event classification model, event argument role recognition model, and event entity fine-grained classification model, to identify and extract information related to news events from preprocessed data, including event type, event trigger words, and event elements. The event information is then pushed to the performance evaluation module via an API interface for assessment to determine if it meets the requirements of subsequent tasks. If it does not meet the requirements, specific improvement suggestions are sent back to the event extraction agent for optimization to ensure the quality of event information extraction; if it meets the requirements, the next step is executed.

[0162] 4. Coreference resolution

[0163] The coreference resolution agent invokes intra-document and cross-document event coreference resolution models from the model library to perform intra-document and cross-document coreference resolution on the extracted events. This identifies information with different expressions but pointing to the same event, effectively aggregating event elements scattered across different contexts and achieving semantic consistency integration across text data sources. The event coreference resolution results are then pushed to the effectiveness evaluation module for assessment to determine if they meet the requirements of subsequent tasks. If they do not meet the requirements, specific rectification suggestions are sent back to the coreference resolution agent for optimization to ensure the effectiveness of event coreference resolution; if they meet the requirements, the next step is executed.

[0164] 5. Summary generation

[0165] The summary generation agent calls the event summary generation model in the model library to summarize and categorize the key information of the event after coreference resolution, generating a concise event summary. The event summary is then pushed to the effect evaluation module for evaluation to determine if it meets the requirements. If it does not meet the requirements, specific rectification suggestions are sent back to the coreference resolution agent for optimization to ensure the effectiveness of the event summary generation; if it meets the requirements, the next step is executed.

[0166] 6. Event Presentation

[0167] The event display agent calls the geographic information visualization platform in the tool library to locate and mark the event on the map based on the location of the event, and calls the corresponding event icon in the icon library to represent the event according to the event type. Finally, the event summary is marked on the event information box to visualize the event information intuitively.

[0168] Step 6: Task Completed

[0169] After the task is completed, the semantic understanding and visualization results of the event can be saved through the one-click download function of the geographic information visualization platform, forming an event overview map.

[0170] The core technical principle of this invention lies in constructing a self-evolving multi-agent system based on dynamic model expansion. This system achieves intelligent semantic understanding and visualization of news events through a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution - task completion." The system employs a multi-agent collaborative architecture, integrating agents for data retrieval, event extraction, and coreference resolution to complete end-to-end analysis. It innovatively introduces a capability verification stage. When the existing model cannot meet the task requirements, the dynamic model expansion module is automatically triggered. Through an active learning framework, it achieves iterative training of a dedicated model through human-machine collaboration. This module includes agents for data retrieval, pre-labeling, difficult example selection, and strategy optimization, forming a closed-loop optimization process of "data collection → pre-labeling → model training → difficult example selection → manual verification → strategy optimization → model optimization." A specially designed adaptive difficult example optimization strategy significantly improves model training efficiency by dynamically selecting high-value samples and optimizing the selection strategy. This technology overcomes the limitations of traditional systems, such as fixed models and poor domain adaptability, enabling rapid and accurate analysis of complex news events.

[0171] Example 2:

[0172] It includes a multi-agent collaboration framework construction module, a requirement understanding and task planning module, a capability verification module, a model dynamic expansion module, a task execution module, and a completion module, which are connected in sequence.

[0173] The multi-agent collaboration framework building modules include central control multi-agents, task execution multi-agents, and model extension multi-agents;

[0174] The model dynamic expansion module includes, in sequence, a model expansion suggestion and data acquisition unit, a data pre-labeling unit, a model training and pseudo-label generation unit, a difficult example screening and manual verification unit, a strategy optimization and model iteration unit, and a model evaluation and tool library update unit;

[0175] The task execution module includes a data retrieval unit, a data preprocessing unit, an event extraction unit, a coreference resolution unit, a summary generation unit, and an event display unit, which are connected in sequence.

[0176] The multi-agent collaboration framework building modules include a central large model, a database, a tool library, a model library, and a multi-agent architecture. The central large model is a large language model with logical reasoning capabilities, including DeepSeek-v3, Qwen-3, or ERNIE-3.5.

[0177] Central control multi-agent systems include:

[0178] A demand-understanding agent is used to parse user instructions and intentions;

[0179] A task planning agent is used to formulate a task execution plan and break it down into a sequence of executable steps.

[0180] Capability verification agent is used to evaluate whether existing models in the model library meet the task requirements.

[0181] Task-oriented multi-agent systems include:

[0182] The performance evaluation agent, acting as a process controller, evaluates and controls the execution results of the data retrieval, data preprocessing, event extraction, coreference resolution, summary generation, and event display agents.

[0183] The extended class of multi-agent models includes:

[0184] A data acquisition agent is used to call APIs to collect task-related data.

[0185] Pre-labeled agents are used to call UIE tools to pre-label data;

[0186] The model training agent is used to call the API to train models online.

[0187] A difficult example screening agent is used to screen high-value samples based on an active learning strategy.

[0188] A strategy optimization agent is used to evaluate labeled samples and optimize the hard example selection strategy.

[0189] A method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model includes the following steps:

[0190] S1: Construct a multi-agent collaboration framework, including defining the responsibilities of central control agents, task execution agents, and model extension agents, and establishing a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution - task completion";

[0191] S2: Receive user requests and plan tasks to form a task execution plan;

[0192] S3: Execution capability verification, determine whether the existing model meets the task requirements; if not, proceed to S4; if it does, proceed to S5.

[0193] S4: Perform dynamic model expansion, including data collection, pre-labeling, model training, pseudo-label generation, hard example selection, manual verification, strategy optimization, and model iteration;

[0194] S5: Perform tasks, including data retrieval, preprocessing, event extraction, coreference resolution, summary generation, and event display;

[0195] S6: Complete the task and output the event semantic understanding and visualization results.

[0196] The capability verification steps include:

[0197] S1: Calculate Model Coverage Ratio (MCR)

[0198]

[0199] in, It is an indicator function used to determine whether a condition is true or false; This represents the i-th model (such as event classification, argument recognition, etc.) needed to complete the task. This represents the collection of all models in the existing model library; This represents the total number of all models required to complete the task.

[0200] S2: Ability Coverage Ratio (ACR)

[0201] To measure the extent to which the capabilities of existing models in the model library cover the requirements of news event analysis tasks, for the models The formula for calculating its coverage is as follows:

[0202]

[0203] in, It is an indicator function used to determine whether a condition is true or false; express The j-th capability of the model (such as the ability of the event type recognition model to identify economic events, sports events, etc.); This represents the set of existing capabilities of the corresponding model in the model library; express The total number of all capabilities required for the model to complete the task.

[0204] S3: If MCR ≥ 0.95 and ACR ≥ 0.85 for each model, then the model is deemed to meet the requirements; otherwise, dynamic model expansion is triggered.

[0205] In the model dynamic expansion step, difficult examples are selected based on a confidence threshold σ, and a dynamic adjustment mechanism is employed:

[0206]

[0207] in:

[0208] The average information value density of the current round of difficult examples;

[0209] : The average annotation cost of the current round of hard example sets;

[0210] The improvement in model accuracy on the current round of validation data;

[0211] Adjust step size;

[0212] , , Weighting coefficient.

[0213] During the task execution process, the execution results of each agent are verified by the effect evaluation agent. If the results do not meet the requirements, they are fed back to the corresponding agent for optimization.

[0214] The model dynamic expansion step adopts an active learning framework, which iteratively executes the process of "difficult example screening - manual verification - model update - strategy optimization" until the model accuracy improves by less than 0.5% for two consecutive rounds, at which point training stops and the final model is updated to the model library.

[0215] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A self-evolving multi-agent event semantic understanding system based on a dynamic and scalable model, characterized in that, It includes, in sequence, a multi-agent collaboration framework building module, a requirements understanding and task planning module, a capability verification module, a model dynamic expansion module, a task execution module, and a completion module: The multi-agent collaboration framework construction module includes central control type multi-agents, task execution type multi-agents, and model extension type multi-agents; The model dynamic expansion module includes a model expansion suggestion and data acquisition unit, a data pre-labeling unit, a model training and pseudo-label generation unit, a difficult example screening and strategy optimization iteration unit, and a model evaluation and tool library update unit connected in sequence. The task execution module includes a data retrieval unit, a data preprocessing unit, an event extraction unit, a coreference resolution unit, a summary generation unit, and an event display unit connected in sequence. The capability verification module is used to verify model coverage. and capability coverage Evaluate existing models in the model library; wherein, the model coverage The calculation formula is: ,in, It is an indicator function used to determine whether a condition is true or false. This represents the i-th model needed to complete the task. This represents the collection of all models in the existing model library. This represents the total number of all models required to complete the task; the capability coverage. The calculation formula is: ,in, It is an indicator function used to determine whether a condition is true or false. express The j-th capability of the model, This represents the set of existing capabilities of the corresponding model in the model library. express The total number of all capabilities required for the model to complete the task; and If the task requirements are met, the model is dynamically expanded; otherwise, the model's dynamic expansion module is triggered. The difficult example screening and strategy optimization iteration unit is used to manually verify difficult examples in the pseudo-label data based on a confidence threshold σ, and dynamically adjust the confidence threshold σ for the next round of difficult example screening based on the results of the manual verification; the dynamic adjustment mechanism of the confidence threshold σ is as follows: ,in, For average information value density, To average annotation cost, For the improvement in model accuracy, To adjust the step size, , , These are the weighting coefficients.

2. The self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to claim 1, characterized in that, The multi-agent collaboration framework construction module includes a central large model, a database, a tool library, a model library, and a multi-agent architecture. The central large model is a large language model with logical reasoning capabilities, including DeepSeek-v3, Qwen-3, or ERNIE-3.

5.

3. The self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to claim 1, characterized in that, The central control type of multi-agent includes: A demand-understanding agent is used to parse user instructions and intentions; A task planning agent is used to formulate a task execution plan and break it down into a sequence of executable steps. Capability verification agent is used to evaluate whether existing models in the model library meet the task requirements.

4. The self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to claim 1, characterized in that, The task execution multi-agent includes: The performance evaluation agent, acting as a process controller, evaluates and controls the execution results of the data retrieval, data preprocessing, event extraction, coreference resolution, summary generation, and event display agents.

5. The self-evolving multi-agent event semantic understanding system based on a dynamic scalable model according to claim 1, characterized in that, The model extension class of multi-agent includes: A data acquisition agent is used to call APIs to collect task-related data. Pre-labeled agents are used to call UIE tools to pre-label data; The model training agent is used to call the API to train models online. A difficult example screening agent is used to screen high-value samples based on an active learning strategy. A strategy optimization agent is used to evaluate labeled samples and optimize the hard example selection strategy.

6. A method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model, based on the self-evolving multi-agent event semantic understanding system based on a dynamic scalable model as described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Construct a multi-agent collaboration framework, including defining the responsibilities of central control agents, task execution agents, and model extension agents, and establishing a closed-loop workflow of "requirement understanding - task planning - capability verification - task execution - task completion"; S2: Receive user requests and plan tasks to form a task execution plan; S3: Execution capability verification, determine whether the existing model meets the task requirements; if not, proceed to S4; if it does, proceed to S5. S4: Perform dynamic model expansion, including data collection, pre-labeling, model training, pseudo-label generation, hard example selection, manual verification, strategy optimization, and model iteration; S5: Perform tasks, including data retrieval, preprocessing, event extraction, coreference resolution, summary generation, and event display; S6: Complete the task and output the event semantic understanding and visualization results.

7. The method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model according to claim 6, characterized in that, In the task execution steps, the execution results of each agent are verified by the effect evaluation agent. If the results do not meet the requirements, they are fed back to the corresponding agent for optimization.

8. The method for self-evolving multi-agent event semantic understanding based on a dynamic scalable model according to claim 6, characterized in that, The model dynamic expansion step adopts an active learning framework, which iteratively executes the process of "difficult example screening - manual verification - model update - strategy optimization" until the model accuracy improves by less than 0.5% for two consecutive rounds, at which point training stops and the final model is updated to the model library.

Images