Multi-agent cooperation-based long text news event element extraction method and system

Abstract

This invention relates to the field of information extraction technology, and discloses a method and system for extracting elements from long-text news events based on multi-agent collaboration. The method includes the following steps: constructing a large language model multi-agent system architecture; a central coordinating agent receiving unstructured news text; a text segmentation agent outputting logical units and their time descriptions; an event extraction agent extracting the event subject, action, and object, and completing time and entity role information; a scene matching agent determining the fine-grained news scene category to which it belongs; an information extraction agent generating structured event records; and a central coordinating agent performing quality review and anomaly handling on each step of the output to ensure the integrity and consistency of the output data. This invention, through relationship building and contextual analysis agents, can construct static relationships between event subjects based on structured event information, and summarize event situation changes based on a timeline, improving the feasibility and visualization of the system's news event analysis.

Description

Technical Field

[0001] This invention relates to the field of information extraction technology, and in particular to a method and system for extracting elements from long-text news events based on multi-agent collaboration. Background Technology

[0002] With the explosive growth of internet information, unstructured text data (such as news reports, social media content, and historical documents) has become a major source of information. However, these texts often lack standardized formats, exhibit highly diverse semantic expressions, and contain a large amount of implicit semantic information (temporal reference, spatial metaphor, and cross-sentence logical relationships), posing a significant challenge to the extraction and analysis of key events. Traditional information extraction techniques mainly rely on rule matching, dictionary lookups, or machine learning models. However, these methods have significant limitations when dealing with event descriptions in complex contexts: on the one hand, rule systems struggle to cover the diversity of language expressions, especially implicit temporal references and spatial metaphors; on the other hand, traditional machine learning models have limited ability to capture long-distance semantic relationships with strong contextual dependencies, resulting in incomplete extraction of event elements.

[0003] In recent years, large language models (LLMs) have demonstrated outstanding performance in event extraction tasks due to their powerful semantic understanding and context modeling advantages. However, the inherent context window limitation of LLMs makes it difficult to directly process long, unstructured texts. Existing techniques typically employ simple sliding window or fixed-length segmentation strategies to cut long texts into multiple segments for separate processing, but this approach leads to severe context fragmentation problems: when key information (such as time reference points) is located in the preceding paragraph, while the event description is located in the following paragraph, segmentation processing results in the loss of key information.

[0004] To address the issue of context fragmentation in long text processing, some studies have proposed document-level event extraction frameworks, enhancing long-distance dependency modeling by introducing global attention mechanisms or hierarchical coding structures. However, these methods suffer from high computational complexity and struggle to adapt to the varying expressions across different text domains. Another approach attempts to construct event chains or timelines, linking scattered event descriptions through referencing resolution techniques; however, the processing of time expressions still relies on predefined rule bases, failing to handle the flexibility and diversity of language expression. Especially when dealing with time references across paragraphs or headings, existing techniques often fail due to a lack of effective contextual transmission mechanisms. Therefore, existing methods for extracting elements from long-text news events suffer from low element extraction completeness and poor element contextual coherence. Summary of the Invention

[0005] This invention provides a method and system for extracting elements from long-text news events based on multi-agent collaboration, in order to solve the problems of low completeness and poor contextual coherence in existing information extraction methods.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] In a first aspect, the present invention provides a method for extracting elements from long-text news events based on multi-agent collaboration, comprising the following steps:

[0008] S1: Construct a multi-agent system architecture for a large language model, which includes a central coordinating agent, a text segmentation agent, an event extraction agent, a scene matching agent, and an information extraction agent.

[0009] S2: The central coordinating agent receives unstructured news text, generates a workflow execution plan based on the unstructured news text and preset task objectives, and schedules other agents in the architecture based on the workflow execution plan. The workflow execution plan includes: text segmentation task information, event extraction task information, scene matching task information, and information extraction task information.

[0010] S3: The text segmentation agent performs two-stage semantic segmentation on the input unstructured news text based on the text segmentation task information, and outputs logical units and their time descriptions.

[0011] S4: The event extraction agent is used to perform sentence-level event recognition on each logical unit based on the event extraction task information, extract the event subject, event action and event object, and construct the event subject, event action and event object into independent event units based on time description;

[0012] S5: The scene matching agent judges the independent event unit based on the scene matching task information to determine the fine-grained news scene category to which the independent event unit belongs;

[0013] S6: The information extraction agent extracts time description, event object, and event subject from independent event units as key elements based on the information extraction task information. It generates structured event records based on the key elements and classifies the structured event records based on fine-grained news scene categories to complete the element extraction of unstructured news text.

[0014] Optionally, the two-stage semantic segmentation includes:

[0015] The first stage involves dividing unstructured news text into basic paragraph units based on structural features and assigning them unique identifiers. Based on these unique identifiers, the basic paragraph units are then divided into paragraph mapping tables with paragraph identifiers. The structural features include: heading levels, line breaks, and paragraph separators.

[0016] The second stage: Based on topic consistency, temporal continuity and logical relevance, the basic paragraphs are aggregated into semantically coherent logical units using a large language model. The logical unit includes the paragraph identifiers corresponding to all the basic paragraphs that make up the logical unit in the paragraph mapping table, the temporal description of the logical unit, and all the basic paragraphs that make up the logical unit.

[0017] Optionally, the sentence-level events identified by the event extraction agent include:

[0018] Point events, Line events, and Area events are defined as follows: Point events are static events used to describe a single entity, location, or facility; Line events are events involving path or directional movement; and Area events are events used to describe area activities.

[0019] The event extraction agent constructs independent event units based on time description, comprising the event subject, event action, and event object:

[0020] Each event is constructed as an independent event unit, consisting of the event subject, the event action corresponding to the event subject, the event object corresponding to the event action, and a description of the time in which the event action occurs.

[0021] Optionally, the scene matching agent judges the independent event unit based on the scene matching task information to determine the fine-grained news scene category to which the independent event unit belongs, including:

[0022] A two-level index classification mechanism is used to determine the fine-grained news scene category to which an independent event unit belongs. This two-level index classification mechanism includes:

[0023] First-level index classification: Extract the main body of the event from the independent event unit, and classify the main body of the event into first-level categories according to the news field;

[0024] Secondary index classification: Construct a candidate library of scene categories under the primary category, and further subdivide specific scene categories under the primary category based on the candidate library of scene categories.

[0025] Optionally, the structured event records output by the information extraction agent follow predefined JSON rules and include event ID, event description, participant information, time range, and geographic location information. The time range adopts ISO8601 format, and the geographic location information includes the coordinates of the starting point, waypoints, and destination.

[0026] Optionally, the method further includes:

[0027] S7: The central coordinating agent receives output information from other agents in the architecture and performs quality auditing and anomaly handling on the output information.

[0028] Optionally, the central coordinating agent performs quality audits on the output information according to preset quality standards during execution. If the quality does not meet the standards, an exception is handled, which triggers a retry, model switching, or skip processing strategy.

[0029] Secondly, embodiments of this application provide a long-text news event element extraction system based on multi-agent collaboration, including:

[0030] Memory, used to store computer programs;

[0031] A processor for implementing the method as described in any one of the first aspects when executing the computer program.

[0032] Beneficial effects:

[0033] This invention provides a method for extracting long-text news event elements based on multi-agent collaboration. It coordinates four specialized agents—text segmentation, event extraction, scene matching, and information extraction—to form an efficient workflow. Through a semantically driven two-stage segmentation strategy, it effectively avoids the separation of time anchors and event descriptions caused by traditional segmentation methods, ensuring semantic coherence across paragraphs. It achieves accurate event scene classification by employing a mechanism that enhances semantic understanding and retrieval, improving the accuracy of classifying ambiguous events. It provides standardized structured output by defining a unified event element output format, supporting subsequent system integration and visualization analysis. Each agent has a clear division of labor and responsibilities, and through unified scheduling and review by the coordinating agent, a closed-loop workflow is formed, fully leveraging the optimal performance of the large language model in each sub-task. The modular architecture supports flexible addition and removal of agent types, and the scene library can be expanded as needed, making it suitable for multi-domain and multilingual text processing. Attached Figure Description

[0034] Figure 1 This is one of the flowcharts of a long text news event element extraction system based on multi-agent collaboration, which is a preferred embodiment of the present invention;

[0035] Figure 2 The second flowchart is a preferred embodiment of the long text news event element extraction system based on multi-agent collaboration of the present invention.

[0036] Figure 3 The example diagram of the element extraction process provided in the embodiments of the present invention shows the complete processing path from raw text to structured event records. Detailed Implementation

[0037] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked," and similar terms, are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0039] Please see Figure 1 This embodiment provides a method for extracting elements from long-text news events based on multi-agent collaboration, including the following steps:

[0040] S1: Construct a multi-agent system architecture for a large language model, which includes a central coordinating agent, a text segmentation agent, an event extraction agent, a scene matching agent, and an information extraction agent.

[0041] S2: The central coordinating agent receives unstructured news text, generates a workflow execution plan based on the unstructured news text and preset task objectives, and schedules other agents in the architecture based on the workflow execution plan. The workflow execution plan includes: text segmentation task information, event extraction task information, scene matching task information, and information extraction task information.

[0042] S3: The text segmentation agent performs two-stage semantic segmentation on the input unstructured news text based on the text segmentation task information, and outputs logical units and their time descriptions.

[0043] S4: The event extraction agent is used to perform sentence-level event recognition on each logical unit based on the event extraction task information, extract the event subject, event action and event object, and construct the event subject, event action and event object into independent event units based on time description;

[0044] S5: The scene matching agent judges the independent event unit based on the scene matching task information to determine the fine-grained news scene category to which the independent event unit belongs;

[0045] S6: The information extraction agent extracts time description, event object, and event subject from independent event units as key elements based on the information extraction task information. It generates structured event records based on the key elements and classifies the structured event records based on fine-grained news scene categories to complete the element extraction of unstructured news text.

[0046] In the above embodiments, a coordinating agent coordinates four specialized agents—text segmentation, event extraction, scene matching, and information extraction—to form an efficient workflow. A semantically driven two-stage segmentation strategy effectively avoids the separation of time anchors and event descriptions caused by traditional segmentation methods, ensuring semantic coherence across segments. Precise event scene classification is achieved through a mechanism that enhances semantic understanding and retrieval, improving the accuracy of classifying ambiguous events. Structured and standardized output is provided by defining a unified event element output format, supporting subsequent system integration and visualization analysis. Each agent has a clear division of labor and responsibilities, and a closed-loop workflow is formed through unified scheduling and review by the coordinating agent, fully leveraging the optimal performance of the large language model in each subtask. The modular architecture supports flexible addition and removal of agent types, and the scene library can be expanded as needed, making it suitable for multi-domain and multilingual text processing.

[0047] In this embodiment, each agent adopts a standardized five-module architecture, which includes: a role definition module: clarifying the agent's functional positioning and task boundaries; a prompting engineering module: designing domain-specific prompt templates to guide the large language model to perform specific sub-tasks; an input processing module: preprocessing and formatting the input data; a core processing module: calling the large language model to complete core reasoning and generation tasks; and an output processing module: structuring and parsing the output of the large language model to ensure the integrability of the results.

[0048] Optionally, two-stage semantic segmentation includes:

[0049] The first stage involves dividing unstructured news text into basic paragraph units based on structural features and assigning them unique identifiers. Based on these unique identifiers, the basic paragraph units are then divided into paragraph mapping tables with paragraph identifiers. The structural features include: heading level, line breaks, and paragraph separators.

[0050] The second stage: Based on topic consistency, temporal continuity and logical relevance, the basic paragraphs are aggregated into semantically coherent logical units using a large language model. The logical unit includes the paragraph identifiers corresponding to all the basic paragraphs that make up the logical unit in the paragraph mapping table, the temporal description of the logical unit, and all the basic paragraphs that make up the logical unit.

[0051] In the above embodiments, for two-stage semantic segmentation, in the first stage, by identifying explicit structural markers such as title levels (e.g., "##", "1.", etc.), line breaks (\n), paragraph separators (\r\n), etc. in the text, the full text is divided into basic paragraph units, and a unique paragraph identifier (e.g., "1", "2",...) is assigned to each paragraph to form a paragraph mapping table. In the second stage, by calling a large language model, based on the topic consistency, time continuity, and logical relevance among paragraphs, the basic paragraphs are aggregated into several semantically coherent logical units. The output content is in JSON format as a grouped result, and each group contains: a list of paragraph IDs (in the original order) that make up the logical unit and the time description of the logical unit.

[0052] Optionally, the sentence-level events recognized by the event extraction agent include:

[0053] Point events, Line events, and Area events. Among them, Point events are static events used to describe a single entity, location, or facility; Line events are events involving paths or directional movements; Area events are events used to describe regional activities.

[0054] The event extraction agent constructs the event subject, event action, and event object into independent event units based on the time description, including:

[0055] Constructing the event subject, the event action corresponding to the event subject, the event object corresponding to the event action, and the time description when the event action occurs into an independent event unit.

[0056] In the above embodiments, the event extraction agent performs sentence-level event recognition and extraction on the text within each logical unit. The agent identifies the subject (Actor), action (Action), and object (Object) in the event and extracts a complete event description. The event form recognition is extracted according to the following classification: Point event: A static event describing a single entity, location, or facility; Line event: Involving a path or directional movement (e.g., "from A to B"); Area event: A description of regional activities.

[0057] At the same time, explicit time (e.g., "3 am") and implicit time clues (e.g., "T + 2 hours") are integrated, and the entity role information is completed (e.g., marking "Chinese side", "destination"), improving semantic integrity. The event extraction agent processes each segmented text block, identifies the event subject, action, and object in the text block, and extracts and saves each event as an independent event unit at the sentence level.

[0058] Optionally, the scene matching agent judges the independent event units based on the scene matching task information to determine the fine-grained news scene category to which the independent event unit belongs, including:

[0059] A two-level index classification mechanism is used to determine the fine-grained news scene category to which an independent event unit belongs. This two-level index classification mechanism includes:

[0060] First-level index classification: Extract the main body of the event from the independent event unit, and classify the main body of the event into first-level categories according to the news field;

[0061] Secondary index classification: Construct a candidate library of scene categories under the primary category, and further subdivide specific scene categories under the primary category based on the candidate library of scene categories.

[0062] In the above embodiments, to alleviate the context capacity limitation of large language models, a two-level indexing mechanism is adopted: First-level classification: news is divided into primary categories based on domain; Second-level classification: specific scene categories are further subdivided under the primary categories, and the secondary categories correspond to a candidate library containing several news scenes. The scene matching agent determines the primary and secondary categories to which the text belongs, retrieves its scene category by combining semantic understanding, and performs semantic matching by retrieving the candidate library.

[0063] Optionally, the structured event records output by the information extraction agent follow predefined JSON rules and include event ID, event description, participant information, time range, and geographic location information. The time range adopts ISO8601 format, and the geographic location information includes the coordinates of the origin, waypoints, and destination.

[0064] In the above embodiments, the information extraction agent extracts key elements such as time, location, and subject from each event unit to generate a standardized structured record. The output format strictly follows predefined JSON rules and includes elements such as time, event entity, and location. The information extraction agent processes each event unit, extracting key elements of time, location, and subject from the event description, which includes supplemented time and entity elements, to generate a structured event element record that conforms to the contextual semantics.

[0065] Optionally, the method also includes:

[0066] S7: A central coordinating agent receives output information from other agents in the architecture and performs quality checks and anomaly handling on the output information.

[0067] Optionally, the central coordinating agent may review the output information according to preset quality standards during execution. If the quality does not meet the standards, an exception may be handled, triggering a retry, model switching, or skip processing strategy.

[0068] In the above embodiments, the central coordinating agent serves as the central hub and control core of the system. Its core responsibility is to achieve seamless collaboration and closed-loop control among multiple agents, specifically including: global workflow planning and dynamic scheduling: based on text meta-information such as the length of the input text and preset task objectives, it generates the optimal agent execution path and sequentially calls execution agents such as text segmentation, event extraction, scene matching and information extraction.

[0069] Result Review and Anomaly Handling: The system receives intermediate results from each agent and performs integrity, consistency, and accuracy checks. If the result quality is substandard, it triggers retry, model switching, or skipping strategies. The central coordinating agent executes a review mechanism after each LLM output step, verifying intermediate results according to preset quality standards (such as field integrity, logical consistency, and repeatability checks). If the effective extraction rate in a batch of events is below a threshold, a retry mechanism is triggered or a backup agent is called to ensure high reliability of the final output.

[0070] It should be understood that current news event extraction suffers from low extraction efficiency and poor accuracy due to the uncertainty of unstructured text. Therefore, this application provides a method for extracting elements from long-text news events based on multi-agent collaboration. It is worth noting that the news event element extraction method based on a large language model and multi-agent collaboration provided in this application can also apply the extracted structured information to similar visualization scenarios, and is not limited to news extraction.

[0071] In this implementation, an intelligent agent refers to a composite framework that integrates large language models, external tools, and external knowledge, possessing autonomous planning, execution, and memory capabilities. The large language model includes, but is not limited to, versions of LLaMA3, Qwen3, and DeepSeek with over 1 billion iterations.

[0072] Below, as Figure 3 As shown, a complete example is used to detail the steps of the above-mentioned method for extracting key elements of news events based on a large language model multi-agent system:

[0073] Detailed Explanation of Multi-Agent Workflow

[0074] I. Specific Implementation Mechanism of the Central Coordinating Intelligent Agent

[0075] As the system's central hub, the coordinating agent undertakes key functions such as task planning, process scheduling, quality auditing, and anomaly handling. It is a crucial component ensuring the efficient, reliable, and closed-loop operation of the entire event element extraction process. The coordinating agent is defined as a "system-level task planning and quality control expert," and its core responsibilities include:

[0076] Global workflow planning and dynamic scheduling: Based on text metadata such as the length of the input text and preset task objectives, the optimal intelligent agent execution path is generated, and execution-type intelligent agents such as text segmentation, event extraction, scene matching and information extraction are called in sequence;

[0077] Results review and anomaly handling: Receive intermediate results returned by each agent and perform integrity, consistency and accuracy checks; when the result quality is substandard, trigger recovery strategies such as retry, switching to a backup model, or skipping.

[0078] Its workflow consists of the following modules:

[0079] Planning module: Generates execution path based on task objectives, including the ability to determine whether to invoke the text segmentation agent (for long texts), with prompt templates including:

[0080]

[0081] The result review and exception handling module verifies the output based on general review standards and the review standards of each intelligent agent. If the output falls below the threshold, a retry or degradation strategy is initiated. Its prompt templates include:

[0082]

[0083] Its code is shown below:

[0084] Character settings:

[0085] "You are a rigorous system quality control and self-healing decision-making expert. Your task is to evaluate the quality of an agent's output and, when the quality is substandard, intelligently recommend the most appropriate recovery strategy based on the current system state and error type to ensure the efficient and stable progress of the entire workflow."

[0086] {General Audit Standards}

[0087] - Completeness: The output must contain all required fields and must not be missing any.

[0088] - Consistency: The internal logic of the data must not be contradictory.

[0089] - Accuracy: The information extracted must not be fabricated.

[0090] - Formatting guidelines: Must strictly adhere to the predefined JSON Schema or data structure.

[0091] - Repeatability check: The same event must not be extracted multiple times.

[0092] {Evaluation Standards for Various Intelligent Agents}

[0093] "Event Extraction Agent":

[0094] - The event description must incorporate a timeline (explicit or implicit).

[0095] "Information Extraction Intelligent Agent":

[0096] - `time_range.start_time` and `end_time` must conform to ISO 8601 format or be an explicit time expression.

[0097] "Scene-matching intelligent agent":

[0098] - Matched scenarios must come from a predefined candidate scenario library; new types cannot be created independently.

[0099] Output requirements:

[0100] {

[0101] "is_success": true / false, / / Whether the data quality meets the standards

[0102] "decision": {

[0103] "action": "none | retry | switch_model | skip | delete_data", / / Exception handling

[0104] "parameters": {} / / Action parameters

[0105] }

[0106] }”

[0107] II. Specific Implementation Mechanism of Text Segmentation Intelligent Agent

[0108] This embodiment details the implementation of a text segmentation agent, which employs a two-stage processing mechanism: primary segmentation based on rule matching and segmentation based on semantic and temporal clue aggregation. Unlike traditional segmentation methods, this embodiment does not directly output the segmented text content, but instead outputs a list of paragraph IDs and time descriptions, achieving a balance between lightweight communication and context preservation.

[0109] 1. Elementary segmentation

[0110] In the initial segmentation stage, the text segmentation agent first performs basic segmentation on the input unstructured text, generating raw paragraph units with sequence IDs:

[0111] For example: initial segmentation based on special symbols;

[0112] Identify the line break characters (\n), paragraph markers (\r\n), and heading markers (such as ##, one, etc.) in the text. Use these special symbols as the splitting points to divide the text into the smallest basic paragraph units. Assign a unique and sequential paragraph ID to each basic paragraph unit. The ID format uses a shorter character sequence such as ["1", "2", "3",...]. The example is as follows:

[0113] “paragraph_map = {

[0114] "1": "News paragraph 1.",

[0115] "2": "News paragraph 2.",

[0116] "3": "News paragraph 3."

[0117] }”

[0118] 2. Semantic Aggregation Stage

[0119] After completing the primary segmentation and generating the paragraph mapping table with identifiers, the text segmentation agent performs logical unit aggregation based on semantic coherence and temporal continuity. This stage aims to break through the problem of context breakage caused by traditional fixed-length slicing and accurately identify the boundaries of cross-paragraph events through the deep understanding of the implicit semantic relationships between paragraphs by the large language model:

[0120] The system adopts the method of large language model + prompt engineering to guide the model to perform logical unit division according to the following criteria:

[0121] Temporal Continuity: There is a clear temporal sequence relationship between paragraphs (such as same-day events, successive events);

[0122] Thematic Consistency: It is centered around the same event theme (such as the same conference process);

[0123] [[ID=3|3]]Logical Relevance: It has semantic relationships such as causality, progression, or parallelism.

[0124] The specific execution process is as follows:

[0125] Organize the primary segmentation results in an ordered list form and inject them into a dedicated prompt template to guide the large language model to perform the logical unit division task. The prompt template structure is as follows:

[0126] [[]END]] = \n \n

[0127] The model output follows a strict JSON format. The large language model outputs the grouping results of the basic paragraph units and the time elements, and returns them to the coordination agent.

[0128] For example, regarding the aforementioned paragraph mapping table:

[0129] {

[0130] "1": "News paragraph 1".

[0131] "2": "News paragraph 2."

[0132] "3": "News paragraph 3.",

[0133] }”

[0134] After semantic aggregation, the following logical unit is output, where the time description in the logical unit is the time when the news paragraph appears:

[0135] {"id_list":["1","2","3"],"time":"2023-08-28"}

[0136] This two-stage segmentation mechanism differs from existing technologies that transmit the original segmented text; this invention only transmits a list of paragraph IDs and core time anchors. This design achieves three technical effects:

[0137] Overcoming window limitations: Downstream modules do not need to load the original long text, completely avoiding the context length bottleneck of LLM; Preserving key semantics: By passing "time anchors", the downstream modules are provided with the context information necessary to resolve time references such as "the next day" and "subsequent"; Reducing system overhead: The data volume of ID lists and timestamps is much smaller than that of the original text, which significantly reduces the communication load between agents and the cost of repetitive inference in LLM.

[0138] III. Specific Implementation Mechanism of Event Extraction Agent

[0139] This embodiment details the specific implementation of the event extraction agent. The coordinating agent schedules the aforementioned logical units one by one to the event extraction agent for fine-grained event extraction. This agent focuses on analyzing structured event triples (subject-action-object) from natural language sentences and distinguishing different types of event forms based on spatial semantic features, thereby ultimately extracting a complete event description.

[0140] Core processing flow of event extraction agent:

[0141] Role definition module configuration: Set the core role of the intelligent agent: "Professional Event Analysis Expert";

[0142] Hints for engineering module construction: Design dedicated hint templates to guide the large language model to focus on the event extraction task;

[0143] = \n \n

[0144] The task requirements include the following points:

[0145] Identify the event subject: region, organization, individual, etc.; extract the core action: verb or verb phrase; determine the action object: the object or target affected by the action; determine the event type: Point (location event), Line (path event), Area (region event); complete the implicit semantic roles (such as "our side", "Company A");

[0146] The prompt template includes role definition, event type definition, output format specifications and processing rules, and examples.

[0147] Event pattern recognition mechanism:

[0148] Point event identification: Applicable to statements describing changes in the status of a single location or facility. For example, "Organization A enters Region 3" is identified as a Point event, with the location being "Region 3".

[0149] Line event recognition: Captures descriptions involving directional movement or path planning. For example: "Organization B starts from Mountain 1, moves south along the road to point C" → Extracts the starting point "Mount 1", the path "road", and the destination "point C".

[0150] Area event identification: This refers to regional behavioral events that involve three or more geographic coordinate points. For example, "A triangular structure is formed by high ground A, valley B, and bridgehead C," which is labeled as an Area event.

[0151] 3. Integration of timelines and completion of physical characters

[0152] Extract explicit time from the context (e.g., "2023-08-28"), identify implicit time cues (e.g., "T+2 hours"), and integrate time information into a single event description.

[0153] Entity role completion: Identify key entities in an event and supplement their implicit role information.

[0154] Example of character completion:

[0155] "Company A departs from Beijing to Shanghai" → "Company A (Chinese party) departs from Beijing to Shanghai (destination)"

[0156] This mechanism significantly improves the semantic integrity of downstream scene matching and information extraction.

[0157] IV. Specific Implementation Mechanism of Scene Matching Intelligent Agent

[0158] This embodiment details the specific implementation of the scene matching agent. As a crucial component of the multi-agent system, this agent is responsible for matching the most suitable fine-grained scene type to each event description output by the event extraction agent. This embodiment is specifically optimized to address the ambiguity and vagueness of event descriptions.

[0159] 1. Two-level scene indexing architecture

[0160] To alleviate the context capacity limitations of large language models and avoid attention distraction caused by an excessively large candidate scene library, this invention constructs a hierarchical scene knowledge base:

[0161] The candidate scenario library is categorized into primary categories based on common news themes such as culture, sports, finance, and technology. Each primary category has secondary categories, and each secondary category corresponds to a candidate library containing several event type descriptors. The scenario matching agent first needs to determine which primary category the news text belongs to, and then retrieve the candidate event type libraries under the secondary category. Its retrieval capability is based on the semantic reasoning ability of LLM.

[0162] The scene matching agent adopts a standardized modular architecture:

[0163] 1. Role definition module configuration

[0164] Define the core role of the intelligent agent as: "Scene Classification Expert";

[0165] Define task boundaries: Only match types from the predefined scene type library.

[0166] = \n

[0167] Its code is shown below:

[0168] Character settings:

[0169] "You are a scenario matching assistant for an expert in {XX field}. Your task is to select the scenario type that best represents each event from the knowledge base. Please strictly follow the following output format, rules, and thought process."

[0170] Retrieved predefined scene types .

[0171] Thought process:

[0172] 1. Analyze the input event description;

[0173] 2. Identify key verbs and noun phrases;

[0174] 3. Compare the feature keywords of predefined scene types;

[0175] 4. Select the scene type with the highest matching degree;

[0176] 5. When there is no matching scenario, mark it as "Unknown Scenario".

[0177] Matching rules.

[0178] Output format.

[0179] Example:

[0180] enter:"{

[0181] "event_id": "1",

[0182] "A company has established a branch office in the XX region."

[0183] }”

[0184] Output: "{

[0185] "event_id": "1",

[0186] "situ_type": “Create a company”

[0187] }”

[0188] V. Specific Implementation Mechanism of Information Extraction Intelligent Agent

[0189] This embodiment describes in detail the specific implementation of the information extraction agent. As a key component of the multi-agent system, the agent is responsible for extracting structured key elements from the event description output by the event extraction agent. The event description output by the event extraction agent is supplemented with missing key elements by a large language model in the pre-process.

[0190] 1: Role definition module configuration

[0191] Define the core role of the intelligent agent as: "Event Semantic Analysis Expert".

[0192] 2. Input and Processing Logic

[0193] Time extraction: Combine the time base of the logical unit with the time modifiers in the sentence to generate start_time and end_time;

[0194] Location extraction: Identify geographic entities, and optionally call an external geocoding service to obtain latitude and longitude coordinates;

[0195] Main component extraction: Name of the standardized organization;

[0196] =" \n "

[0197] Its code is shown below:

[0198] Character settings

[0199] "You are an event description semantic parsing assistant. Your task is to extract structured event information from a given event description and reduce..." <think>The output of the process. Please strictly follow the following format and rules.

[0200] The output format is as follows:

[0201] json

[0202] [{"event_id": "event id",

[0203] "event_desc": "Complete summary of the sub-event",

[0204] "actor": {"side": "Company Name","unit": "Unit or Facility"},

[0205] "time_range": {"start_time": "ISO8601 format (%Y-%m-%dT%H:%M:%S) or other special time representation","end_time": "Same as above"},

[0206] "location": {"start_points": [{"name": "location name","coords": [longitude, latitude],"approx_coords": [approximate X-coordinate, approximate Y-coordinate]}],"via_points": [],"end_points":[]}}]

[0207] ```

[0208] The extraction rules are as follows:

[0209] Rule 1. Location names cannot contain descriptive words such as directions or regions;

[0210] Rule 2. Correctly distinguish between approximate coordinates and latitude / longitude coordinates;

[0211] Rule 3. Output null for missing fields.

[0212] …

[0213] Example:

[0214] {

[0215] "event_id": "1",

[0216] "situ_type": Forward

[0217] }”

[0218] Example:

[0219] enter:"{

[0220] "event_id": "1",

[0221] "event_desc":""

[0222] }”

[0223] Output: "{

[0224] "event_id": "1",

[0225] "actor":, "time_range" "location"…

[0226] }”

[0227] The output strictly follows predefined JSON rules to ensure interoperability between systems. The code is shown below:

[0228] [{

[0229] "event_id": "Unique identifier for the event",

[0230] "event_desc": "Event description summary",

[0231] "actor": {

[0232] "side": "Participating party (organization)",

[0233] "unit": "Specific unit or facility"

[0234] },

[0235] "time_range": {

[0236] "start_time": "Start time in ISO8601 format",

[0237] "end_time": "End time"

[0238] },

[0239] "location": {

[0240] "start_points": [{"name": "place name", "coords": [longitude, latitude], "approx_coords": [X, Y]}],

[0241] "via_points": [],

[0242] "end_points": []

[0243] }

[0244] }]”

[0245] VI. Specific Implementation Mechanism of Relationship Construction and Context Analysis for Intelligent Agents

[0246] This embodiment details the specific implementation architecture and execution logic of the intelligent agent for relationship construction and contextualization. For example... Figure 2 As shown, this intelligent agent, as the post-processing and knowledge enhancement module of a multi-agent collaborative system, does not directly participate in the semantic parsing of the original unstructured text. Instead, it performs high-order semantic aggregation and knowledge graphing operations based on the structured event units output by the preceding intelligent agents (event extraction agent and information extraction agent). The aim is to construct a static membership network between event subjects and dynamically sort out the temporal evolution of events, thereby improving the systematicness, relevance, and interpretability of the event element output.

[0247] 1. Role Definition and Task Boundaries

[0248] This intelligent agent is defined as an "expert in modeling event relationships and evolutionary contexts," and its core responsibilities focus on two dimensions:

[0249] Static Relationship Construction: Based on the fields extracted from the Point event, a hierarchical membership graph of the event participants and their respective units is constructed, forming a static network that reflects the relationship architecture of the relevant roles in the event;

[0250] Dynamic timeline analysis: Based on the information and timestamps extracted from Line events, the continuous movement or state changes of the same subject are aggregated, summarized and abstracted according to the time sequence to generate a structured event evolution timeline.

[0251] 2. Input Data Standardization and Preprocessing

[0252] The input to the agent is a list of event records, standardized by the information extraction agent, with each record strictly adhering to a predefined JSON schema. The system categorizes the input events by event type (Point / Line) and sends them to the relationship building module and the context analysis module for parallel processing.

[0253] 3. Static Relationship Building Module

[0254] This module processes all records marked as "Point Events". Its core objective is to build a lightweight relational knowledge graph centered on "Event Stakeholders - Event Subjects".

[0255] Entity alignment and deduplication: Standardize and clean event-related entities (e.g., eliminate ambiguity in organization names) and establish unique entity identifiers.

[0256] Relationship triple generation: For each Point event, generate a relationship triple. For example, the {"side": "Company A", "unit": "Business Group 6"} extracted from the event will generate the triple <Business Group 6, belongs to, Company A>.

[0257] Knowledge graph output: All triples are aggregated into a global relation graph, supporting subsequent graph database import and visualization analysis.

[0258] 4. Dynamic Timeline Analysis Module

[0259] This module processes all records marked as "Line Events," and its core objective is to generate a coherent timeline summary of the changes in the subject of the event at different points in time.

[0260] Event timeline analysis and summary generation: By calling a large language model, a title summary is generated for each time node based on the time series, and semantic compression processing is performed on the event changes reflected in each event.

[0261] The output of the relationship building and contextualization agent, as enhanced metadata, is collected and integrated into the final event element report by the central coordinating agent.

[0262] This application also provides a long-text news event element extraction system based on multi-agent collaboration, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method. This knowledge-based question-answering system based on a large language model agent can implement various embodiments of the above-described method and achieve the same beneficial effects; therefore, further details are omitted here.

[0263] In this embodiment, the system may further include: an external knowledge base interface, used to access a geocoding service or a domain knowledge base, obtain corresponding geographical knowledge and domain knowledge based on the geocoding service and the domain knowledge base, and send the obtained geographical knowledge and domain knowledge to the processor to assist the intelligent agent in entity recognition and scene matching.

[0264] When the processor executes the operation corresponding to the intelligent agent, it can quickly complete the entity recognition and scene matching in the workflow execution plan by obtaining geographical and domain knowledge through the external knowledge base interface, which effectively improves the extraction efficiency of event elements in long text news elements.

[0265] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.< / think>

Claims

1. A method for long text news event element extraction based on multi-agent cooperation, characterized in that, The method comprises the following steps: S1: constructing a large language model multi-agent system architecture, the architecture comprising a central coordination agent, a text segmentation agent, an event extraction agent, a scene matching agent, and an information extraction agent; S2: receiving unstructured news text using the central coordination agent, generating a workflow execution plan based on the unstructured news text and a preset task target, and scheduling other agents in the architecture based on the workflow execution plan, wherein the workflow execution plan comprises text segmentation task information, event extraction task information, scene matching task information, and information extraction task information; S3: performing two-stage semantic segmentation on the input unstructured news text based on the text segmentation task information using the text segmentation agent, and outputting logical units and their time descriptions; S4: performing sentence-level event identification on each logical unit based on the event extraction task information using the event extraction agent, extracting event subjects, event actions, and event objects, and constructing the event subjects, event actions, and event objects into independent event units based on the time descriptions; S5: determining the fine-grained news scene category to which the independent event units belong based on the scene matching task information using the scene matching agent; S6: extracting time descriptions, event objects, and event subjects as key elements from the independent event units based on the information extraction task information using the information extraction agent, generating structured event records based on the key elements, and classifying the structured event records based on the fine-grained news scene category to complete the element extraction of the unstructured news text. 2.The long text news event element extraction method based on multi-agent cooperation according to claim 1, wherein, The two-stage semantic segmentation comprises: First stage: dividing the unstructured news text into basic paragraph units based on structural features and assigning unique identifiers, and dividing the basic paragraph units into a paragraph mapping table with paragraph identifiers based on the unique identifiers, wherein the structural features include title levels, line breaks, and paragraph separators; Second stage: aggregating the basic paragraphs into logically coherent logical units using a large language model based on theme consistency, time continuity, and logical relevance, wherein a logical unit includes the corresponding paragraph identifiers of all basic paragraphs that make up the logical unit in the paragraph mapping table, the time description of the logical unit, and all basic paragraphs that make up the logical unit. 3.The long text news event element extraction method based on multi-agent collaboration according to claim 1, characterized in that, The sentence-level events identified by the event extraction agent include: Point events, Line events, and Area events, wherein Point events are static events describing a single entity, place, or facility, Line events are events involving path or directional movement, and Area events are events describing regional activities; The event extraction agent constructs the event subjects, event actions, and event objects into independent event units based on the time descriptions, which comprises: Constructing the event subject, the event action corresponding to the event subject, the event object corresponding to the event action, and the time description of the event action into an independent event unit.

4. The long text news event element extraction method based on multi-agent collaboration according to claim 1, characterized in that, The scene matching agent determines the independent event unit based on the scene matching task information, determines the fine-grained news scene category to which the independent event unit belongs, and comprises the following steps: The two-level index classification mechanism is adopted to determine the independent event unit and the fine-grained news scene category to which the independent event unit belongs, wherein the two-level index classification mechanism comprises the following steps: First-level index classification: extracting the event subject in the independent event unit, and classifying the event subject into a first-level category according to a news field; Second-level index classification: constructing a scene category candidate library under the first-level category, and subdividing specific scene categories under the first-level classification based on the scene category candidate library. 5.The long text news event element extraction method based on multi-agent cooperation according to claim 1, wherein, The structured event record output by the information extraction agent complies with a predefined JSON rule, and contains an event ID, an event description, participant information, a time range, and geographic location information, wherein the time range adopts an ISO8601 format, and the geographic location information includes start point, passing point, and end point coordinates. 6.The long text news event element extraction method based on multi-agent cooperation according to claim 1, wherein, The method further comprises the following steps: S7: The hub coordination agent receives output information of other agents in the architecture, and performs quality auditing and abnormality processing on the output information.

7. The long text news event element extraction method based on multi-agent collaboration according to claim 6, characterized in that, The hub coordination agent performs quality auditing on the output information according to a preset quality standard during execution, and performs abnormality processing if the quality is not up to standard, at which time a retry, model switching, or skip processing strategy is triggered.

8. A long text news event element extraction system based on multi-agent cooperation, characterized in that, Comprise: A memory for storing a computer program; A processor for executing the computer program to implement the method according to any one of claims 1 to 7.

Images