A dynamic granularity timeline abstract generation method based on multi-agent cooperation and application thereof

Abstract

This invention relates to the fields of natural language processing and text mining. It discloses a dynamic granular timeline summary generation method based on multi-agent collaboration and its application. First, three types of agents—generating, debating, and deciding—collaborate to filter a high-quality text set relevant to the query topic from multi-source heterogeneous text. Then, based on user-specified granularity parameters, another group of agents collaboratively divides the text into a specified number of semantic clusters. Next, an event summary is generated for each cluster. Finally, fact consistency verification is performed. This invention improves the accuracy and interpretability of information filtering and event grouping through a multi-agent debate mechanism, achieves flexible and dynamic control of summary granularity by introducing granularity parameters, and ensures content reliability through authenticity verification. It enhances the ability to automatically generate reliable and customizable timeline summaries from complex text streams, and is applicable to various scenarios such as public opinion analysis, event tracing, and decision support.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and text mining technology, and in particular to a dynamic granular timeline summarization method based on multi-agent collaboration and its application. Background Technology

[0002] In this era of rapid information and digital development, massive amounts of multi-source, heterogeneous text data have become an indispensable basic resource for social operations, corporate decision-making, and scientific research. Various textual information sources, such as news media, social networks, industry reports, government announcements, and academic literature, are growing exponentially, containing a large amount of time-series information related to specific events or topics. How to quickly extract a clear, coherent, and accurate timeline of events from this scattered, fragmented, and redundant information has become an urgent need in many practical application scenarios.

[0003] In the field of public opinion monitoring and analysis, governments and enterprises need to track the evolution of social hot topics in real time, identify key moments and public opinion trends, and provide a basis for public decision-making and crisis response. For example, in the event of a public safety emergency, managers need to extract the timeline of the event from thousands of news reports, Weibo posts, and short video descriptions to quickly grasp the entire process of the event's cause, development, climax, and mitigation in order to formulate scientific response strategies.

[0004] In historical research and event tracing, researchers often need to reconstruct the complete timeline of a historical event or social movement from a large amount of archives, documents, and oral history materials. Traditional methods rely on manual reading, excerpting, and sorting, which is inefficient and easily influenced by subjective judgment. This is especially true when the materials are scattered from multiple sources, the language is inconsistent, or there are contradictions, making the sorting work particularly difficult.

[0005] In enterprise operations and maintenance (O&M) and fault diagnosis, system logs, monitoring reports, user feedback, and other text records accumulate over time. O&M engineers need to identify the time sequence of system anomalies from these logs, locate the root cause of the fault, and review the event evolution path. Relying on manual screening is not only time-consuming and labor-intensive, but may also lead to misjudgments due to the omission of key log entries.

[0006] Furthermore, timeline summaries play a crucial role in fields such as financial risk control, military situational awareness, medical disease tracking, and legal case adjudication. Users often need to understand events at different levels of abstraction: sometimes a coarse-grained timeline provides an overview to grasp the overall trend, while at other times a fine-grained timeline is required for in-depth analysis of specific details at a particular stage. Although existing technologies can achieve a certain degree of automated timeline generation, they still struggle to simultaneously guarantee the completeness, accuracy, flexibility, and factual authenticity of summaries when faced with multi-source, heterogeneous, and high-noise text streams, making it difficult to meet the differentiated granularity requirements of users in different scenarios.

[0007] Existing timeline summarization technologies mainly include rule-based and statistical methods, as well as generative methods based on large language models. These methods, to a certain extent, achieve the function of extracting temporal information from text and organizing it into a coherent narrative. However, in complex real-world applications, they still face the following technical challenges: One technical problem is the insufficient accuracy of information filtering, which can easily lead to the omission of key information or the inclusion of noise. Faced with high-density, multi-source, and heterogeneous event streams, traditional methods often rely on fixed rules or a single model for relevance assessment. Rule-based methods struggle to cover the diversity of linguistic expressions, and rule design itself is subjective and limited. While single-model methods possess some semantic understanding capabilities, they are easily affected by training data bias, contextual comprehension deviations, or inherent model preferences, leading to the incorrect exclusion of highly relevant text or the inclusion of low-relevance or even irrelevant text during the filtering process, thus affecting the completeness and accuracy of the timeline summary. The second technical issue is the presence of factual bias and semantic noise in the generated content. While large language model-based generation methods excel at creating coherent text, they also commonly suffer from the "illusion" problem—generating seemingly reasonable but actually contradictory or even fabricated content. In timeline summaries, this illusion can lead to distorted event descriptions, incorrect causal relationships, or disordered chronological order, severely undermining the reliability and practicality of the summary. Furthermore, the generated results often contain redundant information that deviates from the semantics of the topic, increasing the user's comprehension burden. The third technical issue is the lack of dynamic granularity adaptation, making it difficult to meet diverse needs. Different users have different requirements for the level of abstraction in timeline summaries across different scenarios. For example, high-level decision-makers may need a highly generalized, coarse-grained timeline to grasp the big picture, while frontline analysts may need a fine-grained timeline containing a lot of details for in-depth analysis. Most existing methods output summaries with fixed granularity, unable to dynamically adjust the level of detail in event clustering and summaries based on user-specified granularity parameters, thus limiting their application value in flexible decision support and personalized services. The fourth technical problem is the opaque decision-making process, resulting in poor interpretability and traceability. Existing methods, especially those based on deep learning, often lack transparency in their internal decision-making processes. Users struggle to understand why the system retains certain texts while filtering others, why certain events are grouped together, and cannot trace the source of any conclusion in the summary. This severely impacts the adoption and trust of technology in fields with high credibility and accountability requirements, such as medicine, law, and the military.

[0008] In summary, those skilled in the art urgently need a method that can automatically, accurately, flexibly, and reliably generate dynamic granular timeline summaries. Summary of the Invention

[0009] The purpose of this invention is to address the significant shortcomings of existing timeline summarization technologies in terms of accuracy, factuality, flexibility, and interpretability when handling complex, multi-source, and heterogeneous real-world text. This invention proposes a dynamic granular timeline summarization method based on multi-agent collaboration and its application. This method is not a simple improvement on existing technologies, but rather an innovation at the system architecture level. It introduces a multi-agent collaboration mechanism to decouple multiple sub-tasks such as text filtering, event clustering, summary generation, and fact verification. Through debate and consensus-building mechanisms among agents, it simulates the collaborative decision-making process of a human expert team. This method improves the accuracy of screening and grouping through multi-perspective generation and multi-round debate; suppresses illusions by introducing a fact consistency verification mechanism; achieves dynamic control of summary granularity through user-adjustable granularity parameters; and achieves full traceability and interpretability of the decision-making process through structured debate logs.

[0010] The technical solution of the present invention to achieve the above objectives is a method for generating dynamic granular timeline summaries based on multi-agent cooperation, which includes the following steps: Step S1, Multi-agent text relevance filtering: Receive the query topic Q, granularity parameter G (a positive integer used to determine the number of text clusters), time interval T, and the corresponding multi-source heterogeneous candidate document set A. Through the collaborative work of generating agent 1, debate agent 1, and decision agent 1, filter out a set of highly relevant texts F from A that are highly related to Q. * ; Step S2, Dynamic Granularity Text Grouping Based on Multi-Agent System: Group the highly relevant text set F * The query topic Q and granularity parameter G are input into the collaborative module consisting of Generating Agent II, Debating Agent II, and Decision Agent II, and F is then processed. * Divide the text into exactly G (the same value as the granularity parameter G) semantically independent and fully covered text clusters {g1, g2, ..., g G Step S3, Event Summary Generation: For each text cluster g k The large language model is called to generate the corresponding summary description S. k ; Step S4, Abstract Authenticity Verification: For each text cluster g... k Its corresponding abstract S k Perform fact consistency checks and output the final timeline summary.

[0011] In step S1, multiple agents are generated. Each agent independently evaluates the relevance of the document set A based on a large language model and outputs a highly relevant subset F of texts. i A highly relevant subset F of text output by multiple generative agents iTogether they form a diverse set of candidate subsets F = {F1, ..., F...} m}

[0012] In step S1, there are multiple debate agents. Each debate agent receives the candidate subset set F, and then conducts multiple rounds of structured dialogue and scoring around the relevance of each highly relevant text subset to the query topic Q. Based on preset consensus rules, a consensus ranking is formed, and finally, a structured debate log is obtained. The consensus rules include at least one of weighted voting rules, stability priority rules, and redundancy trade-off rules.

[0013] In step S1, decision agent 1 receives the candidate subset set F and the structured debate log generated by debate agent 1, and merges them to generate the final highly relevant text set F. * .

[0014] In step S2, multiple agents are generated, and each agent can independently call the large language model, using the granularity parameter G as a hard constraint, to generate a set of highly relevant texts F. * The semantic clusters are divided into G semantic clusters, and a candidate grouping scheme Groupj is output. The candidate grouping schemes Groupj output by multiple generative agents together constitute a set of candidate grouping schemes.

[0015] In step S2, there are multiple debate agents. Each of the multiple debate agents receives the set of candidate grouping schemes and conducts multiple rounds of structured dialogue and scoring on the schemes in the set of candidate grouping schemes in four dimensions: quantity compliance, semantic rationality, boundary clarity and completeness, and generates a structured debate log 2. Then, a consensus-preferred grouping scheme is formed based on the adapted consensus rules.

[0016] In step S2, decision agent two receives the candidate grouping scheme set and the structured debate log two generated by debate agent two. Then, based on the consensus-based optimal grouping scheme, it performs boundary refinement and content redundancy removal, outputting the final G text clusters {g1, g2, ..., g...}. G}

[0017] The generating agent 1, debating agent 1, decision-making agent 1, generating agent 2, debating agent 2, and decision-making agent 2 all use a large language model as the core reasoning engine, and achieve the division of roles and tasks through differentiated prompting engineering.

[0018] The granularity parameter G is a positive integer explicitly specified by the user based on the abstraction level requirements of the summary.

[0019] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method.

[0020] Compared with the prior art, this application has the following non-obvious technical features: First, this application applies a multi-agent collaborative architecture to text relevance filtering and dynamic grouping tasks, and constructs a collaborative system that includes three types of agents: generator, debater, and decision-maker. It is specifically used for front-end text processing (filtering and grouping) of timeline summarization. Through role division and structured collaboration, it simulates the "generation-debate-decision" workflow of a human team. This architecture design has never appeared in existing timeline summarization technologies. Second, this application utilizes a differentiated prompting and debate-driven consensus mechanism based on a large language model. All agents are centered around the large language model, but are assigned different roles and task objectives through differentiated prompting engineering. In particular, the multi-round structured dialogue among debate agents around scoring, and the consensus-building process based on custom rules such as weighted voting, stability priority, and redundancy trade-offs, represent a novel method for improving the quality and interpretability of collective decision-making. Third, this application adopts a user-specified dynamic granularity control parameter, introduces the granularity parameter G as the direct input of the system, and uses it as a hard constraint throughout the dynamic grouping stage, so that the system can output exactly G semantic clusters according to user needs, thereby realizing continuous and controllable summary granularity adjustment from coarse-grained overview to fine-grained tracking. This function is generally missing in existing methods. Fourth, this application adopts a two-stage isomorphic but task-adaptive multi-agent processing flow, which standardizes the collaborative process of "generation-debate-decision" and successfully applies it to two different but interconnected tasks: "text relevance filtering" and "dynamic granular text grouping". This achieves consistency and scalability in system design and reflects the systematic innovation of method design. Fifth, this application uses structured debate logs as a traceability carrier. The system automatically generates and records complete debate logs containing scoring evolution, position changes, and key arguments during the debate process. These logs are not only used to drive decision-making, but also serve as part of the final output, providing a transparent and auditable reasoning trajectory for the entire screening and grouping decision-making process.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention can significantly improve the accuracy and completeness of timeline summaries. Through multi-agent, multi-perspective preliminary screening and multi-round debate verification, it effectively overcomes the bias and limitations of a single model or rule, greatly reduces the risk of missing key information and mixing in irrelevant noise, and ensures that the timeline summaries are both comprehensive and accurate. 2. This invention achieves flexible and adjustable dynamic granularity output. By introducing a user-definable granularity parameter G, the system can generate timeline summaries at different levels of abstraction, meeting the differentiated needs of diverse application scenarios from strategic overview to tactical analysis, and greatly enhancing the practicality and flexibility of the method. 3. This invention can effectively ensure the factual authenticity of the generated content. Through a dedicated summary authenticity verification stage, a large language model is used to compare the factual consistency between the generated summary and the original text, effectively identifying and curbing the "illusion" problem and improving the reliability and credibility of the timeline summary. 4. This invention effectively enhances the interpretability and traceability of system decisions. The entire processing flow, especially the structured logs generated during the debate phase, fully records the reasoning process and consensus formation path of the intelligent agent group, so that the final result is no longer a black box output, but an interpretable conclusion with clear logical support and evidence chain, which is suitable for fields with high credibility requirements. 5. The method of this invention can maintain robustness in complex text environments, and the multi-agent collaborative framework itself has a certain degree of fault tolerance and anti-interference capability. Even if the output of individual agents deviates, it can be corrected through the group debate mechanism, making the system more stable and reliable when facing input texts that are multi-source, heterogeneous, diverse in expression, and of varying quality. 6. This invention provides efficient automated processing capabilities, automating the workflow that originally relied on manual reading, filtering, clustering, and summarizing of large amounts of text. It can quickly process large-scale text sets, significantly saving manpower and time costs, and improving the efficiency of event analysis and decision support. Attached Figure Description

[0022] Figure 1 This is a flowchart of the overall architecture of the dynamic granular timeline summary generation method based on multi-agent collaboration described in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of multi-agent collaboration in the text relevance filtering module based on multi-agents described in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the collaborative workflow of the debate agent described in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of multi-agent collaboration in the dynamic granular text grouping module based on multi-agents described in Embodiment 1 of the present invention; Figure 5 This is a flowchart of the event summary generation and authenticity verification process described in Embodiment 1 of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings; Example 1; A dynamic granular timeline summarization generation method based on a multi-agent cooperative mechanism, such as... Figures 1 to 5 As shown, this method is the first to apply a multi-agent system to text relevance filtering and dynamic granular grouping tasks. By constructing a multi-agent collaborative framework including a generative agent, a debating agent, and a decision-making agent, it achieves high-quality filtering and structured organization of input text. This method pioneers the introduction of a multi-agent collaborative mechanism into the dynamic granular timeline summarization task, fully utilizing the collaboration and debating capabilities among agents. While ensuring semantic coherence and factual consistency, it effectively improves the filtering quality of input text and the rationality of event organization.

[0024] The innovation of this invention lies in applying a multi-agent system (MAS) to text relevance filtering and dynamic granular grouping tasks. Through structured collaboration among three types of agents—generating, debating, and deciding—it achieves high-quality preprocessing and structured event modeling of the input text sequence, significantly improving the filtering accuracy and decision interpretability of the timeline summary generation method in complex and variable text scenarios.

[0025] The implementation process of this invention is as follows: Figure 1 As shown, the entire method mainly consists of the following four sequentially executed technical stages: multi-agent-based text relevance filtering, multi-agent-based dynamic granular text grouping, event summary generation, and summary authenticity verification.

[0026] Phase 1: Text relevance filtering based on multi-agent systems; The proposed method starts with the following four inputs: 1. Search for topic Q (the core events or topics that users care about); 2. Granularity parameter G (indicates the desired number of timeline summary groups, used to control the level of abstraction in the output, which can be explicitly specified by the user according to actual needs to adapt to the different requirements of timeline summary granularity for different content topics). 3. Time interval and each time point T = {T1,…,T n |T i < T j for 1 <= i <= j <= n; 4. The set of all candidate documents collected within the time frame, A = {A T1 ,…,A Tn It originates from heterogeneous texts from multiple sources (such as news reports, social media, official announcements, etc.).

[0027] In the first stage, the query topic Q and the heterogeneous candidate document set A from multiple sources are input into the multi-agent text relevance filtering module. This module consists of multiple generative agents, multiple debating agents, and one decision agent working collaboratively, which respectively perform preliminary screening, multiple rounds of interactive debate, and final decision-making. All agents described in this invention use a Large Language Model (LLM) as their core reasoning and generation engine, and achieve role division and task decoupling through differentiated prompting engineering, input context construction, and output constraint mechanisms. The overall architecture of the multi-agent text relevance filtering module is as follows: Figure 2 As shown: Generative Agent 1: Each generative agent receives a query topic Q and a set of heterogeneous candidate documents A from multiple sources. Based on a large language model, it independently performs relevance evaluation and text filtering tasks, and outputs a preliminary subset of highly relevant texts. Multiple generative agents operate in parallel, employing differentiated prompting strategies and inference parameters to collectively form a highly relevant subset of candidate texts, F={F1,…,F2}. m This serves as the foundational input for subsequent multi-agent debates and consensus evolution.

[0028] Debate Agent 1: A subset of candidate texts F = {F1, ..., F2}, jointly output by multiple generative agents. m The message is broadcast to all debate agents. Each debate agent then sequentially engages in a structured dialogue for up to N rounds (N≥1, typically 2-3) with each subset within the highly relevant candidate text subset, focusing on dimensions such as semantic relevance, keyword matching, contextual consistency, and query intent fit between each subset and the query topic Q, in order to rank the subsets.

[0029] In each round of dialogue, each debate agent represents a subset of highly relevant texts F. i A quantitative score and its basis are provided (total score is 100); other debate agents can question the score or provide supplementary evidence; the MAS system allows agents to adjust their scores in subsequent rounds to reflect the evolution of their viewpoints; all speech content is recorded in a shared dialogue history pool, forming a traceable reasoning trajectory.

[0030] This invention establishes the following preset consensus rules: Rule R1 (Weighted Voting): Calculate the weighted average score for each subset based on the role weights of each debating agent. Rule R2 (Stability First): If a subset consistently ranks in the top three across all rounds, it will be given priority. Rule R3 (Redundancy Trade-off): If the weighted score difference of multiple subsets is less than 10 points, select the subset with a moderate number of texts and low internal redundancy to balance subsequent processing efficiency and information integrity.

[0031] After N rounds of interaction, the MAS system performs an initial sorting of subsets based on rule R1. On this basis, it prioritizes highly stable subsets in conjunction with rule R2, and selects subsets with lower redundancy and more reasonable size according to rule R3 when scores are similar, so as to form a consensus sorting of multi-debate agents.

[0032] Finally, after the debate process, all the debate agents jointly generate a structured debate log, which includes the evolution of scores in each round, the trajectory of position changes, key arguments, and the formation process of the aforementioned consensus ranking. This log serves as the basis for the decision-making agent to make the final relevance determination. The workflow of the debate agent is as follows: Figure 3 As shown.

[0033] Decision Agent 1: Decision Agent 1 receives a set F = {F1, ..., F2} of highly relevant candidate text subsets that are the common outputs of all Generating Agents 1. m The system, along with the complete structured debate logs generated by all debate agents, comprehensively considers the overall relevance of each candidate subset to the query topic Q, as well as the degree of consensus reached during the debate and the support of key arguments. It then invokes a large language model to perform inductive reasoning, fusing high-confidence text units from multiple candidate subsets to generate the final filtered results. This refers to a highly relevant set of texts that are highly related to the query topic Q.

[0034] This invention effectively enhances the robustness and accuracy of text filtering and the interpretability of decisions by introducing a multi-agent debate mechanism, thus avoiding the omission of key information or judgment bias caused by single-model bias.

[0035] Phase Two: Dynamic Granular Text Grouping Based on Multi-Agents; In the second stage, the highly relevant text set F output from the first stage is... * The query topic Q and granularity parameter G are input into the multi-agent-based dynamic granular text grouping module (e.g., Figure 4 (As shown). This module continues the multi-agent collaborative architecture of generation-debate-decision, but the task objective has changed from text relevance filtering to semantic clustering of events: that is, F * Divide into exactly G mutually exclusive and fully covered semantic clusters, such that each cluster g i Each corresponds to an independent sub-event or development stage, with non-overlapping content that supports a complete understanding of the query topic Q. This includes: Generative Agent 2: Each generative agent 2 uses the granularity parameter G as a hard constraint in the structured prompt words, independently calls the large language model, and combines it with the query topic Q to generate F * Divide into G semantic clusters to form a candidate grouping scheme Group j ={g1 j g2 j ,...,g G j By employing a differentiated grouping strategy, multiple generative agents jointly produce a set of candidate grouping schemes with enhanced diversity {Group}. 1 Group 2 Group m}

[0036] Debate Agent 2: The set of candidate grouping schemes (i.e., all candidate grouping schemes) is broadcast to multiple debate agents 2. These agents 2 engage in a structured dialogue for up to N rounds (typically N=3) around the following core questions: The first core issue is quantity compliance: whether it strictly meets the requirement of exactly G groups, and whether there are empty groups or excessive merging; The second core issue is semantic rationality: Does each group focus on a single sub-event, and is there any cross-event mixing? The third core issue is boundary clarity: whether there is semantic overlap or break between adjacent groups; The fourth core question is completeness: Do all groups jointly cover F? * The key information in the document is complete, with no important omissions.

[0037] In each round of dialogue, each debate agent II scores each candidate grouping scheme in turn and cites original evidence to support its judgment; other debate agents II can question the grouping logic of the debate agent II and give reasons, thus driving the evolution of viewpoints.

[0038] In the second phase, this invention adopts consensus formation rules that are isomorphic to those in the first phase but task-adapted: Rule R1' (Weighted Voting): Calculate the weighted average score of each candidate grouping scheme based on the role weight of each debate agent. The score of a single scheme is the weighted sum of its scores in four dimensions: quantity compliance, semantic rationality, boundary clarity, and completeness. Rule R2' (Stability First): If a grouping scheme consistently ranks in the top three across all rounds, it will be given priority consideration. Rule R3' (Redundancy Trade-off): If the weighted score difference of multiple grouping schemes is less than 10 points, the scheme with the lowest inter-cluster semantic overlap and the highest intra-cluster topic purity is selected.

[0039] After N rounds of interaction, the MAS system first performs a preliminary ranking of all candidate grouping schemes that strictly satisfy the G group number constraint based on rule R1'. On this basis, it prioritizes highly stable schemes in conjunction with rule R2', and selects the scheme with better structure according to rule R3' when the scores are similar. Finally, it selects a consensus-optimized grouping scheme as the optimal grouping result.

[0040] Similarly, after N rounds of interaction, the second debate agent generates a structured debate log, which includes the evolution of scores in each round, the trajectory of position changes, key arguments, and the formation process of consensus ranking, serving as the basis for the second decision agent to determine the rationality of grouping.

[0041] Decision Agent 2: Decision Agent 2 receives the consensus-optimized grouping scheme jointly output by all Generating Agents 2 and the complete structured debate log 2 jointly output by all Debating Agents 2. Based on the optimal grouping scheme determined by the MAS system through adapted consensus rules, it calls the large language model for boundary refinement and content redundancy removal. For example, it adjusts the text attribution within groups to eliminate semantic overlap, or merges highly similar subclusters, while keeping the total number of groups unchanged, ensuring that the final output is still G semantically independent event clusters. After the above processing, the decision agent finally outputs a grouping result {g1, g2, ..., g...} with clear structure and well-defined boundaries. G}

[0042] This invention introduces a granularity parameter G that can be explicitly specified by the user, enabling dynamic adjustment of the granularity of event organization during use, thereby achieving flexible granularity control capabilities and meeting users' diverse needs for coarse-grained overview and fine-grained tracking.

[0043] Phase 3: Event summary generation; In the third stage, for each group's grouping result g k (k∈{1,2,…,G}), the large language model is invoked to perform the summary generation task. Based on the semantic guidance of the query topic Q, the model extracts key event elements (such as subject, behavior, time, place, causal relationship, etc.) from each group of texts, generating concise and coherent event description sentences or paragraphs S. k This forms a preliminary timeline summary.

[0044] Phase Four: Verification of Abstract Authenticity; To ensure the factual accuracy of the generated content, an authenticity verification module is introduced in the fourth stage. Figure 5 This module ensures the factual accuracy of the generated content. It will process each set of original text g... k Its corresponding abstract S kThe large language model inputted along with the data compares the propositions in the summary with explicit statements or inferable information in the source text to determine whether there are factual biases, fabricated content, or over-inferences. The final output is a validated, high-quality timeline summary.

[0045] To verify the effectiveness of this invention, experiments were conducted on the large-scale Chinese benchmark DTELS-Bench. This dataset comprises 55,432 news articles from 2,858 domestic and international Chinese information sources, covering seven major areas: politics, economics, society, military, science and technology, health, and environment, involving 543 real-world events, demonstrating high diversity and timeliness. Experimental results show that this invention significantly outperforms existing baseline methods in multiple metrics, including Informativeness Score, Granular Consistency Score, and Factuality Score.

[0046] Example 2; An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0047] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A dynamic granularity timeline summary generation method based on multi-agent collaboration, characterized in that, The method includes the following steps: Step S1, Multi-agent text relevance filtering: Receive the query topic Q, granularity parameter G, time interval T, and the corresponding multi-source heterogeneous candidate document set A. Through the collaborative work of generative agent 1, debating agent 1, and decision-making agent 1, all centered around a large language model, filter out a set of highly relevant texts F from A that are highly related to Q. * ; Step S2, Dynamic Granularity Text Grouping Based on Multi-Agent System: Group the highly relevant text set F * The query topic Q and granularity parameter G are input into a collaborative module consisting of a generative agent II, a debating agent II, and a decision-making agent II, with a large language model at its core. The F... * The text is divided into G semantically independent and fully covered text clusters, and the value of the number of text clusters G is the same as the value of the granularity parameter G. The number of the generated agents two in the step S2 is multiple, each of which can independently call a large language model and divide the high-correlation text set F * into G semantic clusters to form a candidate grouping scheme, and multiple candidate grouping schemes jointly constitute a candidate grouping scheme set; In step S2, there are multiple debate agents 2. Each of the multiple debate agents 2 receives the set of candidate grouping schemes. Then, they conduct multiple rounds of structured dialogue and scoring around the candidate grouping schemes in four dimensions: quantity compliance, semantic rationality, boundary clarity and completeness, and generate structured debate log 2. Finally, a consensus-preferred grouping scheme is formed based on the adapted consensus rules. In step S2, decision agent 2 receives the candidate grouping scheme set and the structured debate log 2 generated by debate agent 2, and then performs boundary refinement and content redundancy removal based on the consensus-optimized grouping scheme to output the final G text clusters. Step S3, Event Summary Generation: For each text cluster g k The large language model is called to generate the corresponding summary description S. k ; Step S4, Abstract Authenticity Verification: For each text cluster g... k Its corresponding abstract S k Perform fact consistency checks and output the final timeline summary.

2. The method of claim 1, wherein, In step S1, multiple agents are generated. Each agent independently evaluates the relevance of the document set A based on a large language model and outputs a highly relevant subset F of texts. i Multiple highly relevant text subsets F i Together they form a diverse set of candidate subsets F.

3. The method of claim 2, wherein, In step S1, there are multiple debate agents. Each debate agent receives the candidate subset set F, and then conducts multiple rounds of structured dialogue and scoring around the relevance of each highly relevant inertial text subset to the query topic Q. A consensus ranking is formed based on preset consensus rules, and finally, a structured debate log is obtained.

4. The method of claim 3, wherein, The step S1 receives the candidate subset set F and the structured argumentation log a, and fuses to generate the final high-relevance text set F * .

5. The method according to any one of claims 1 to 4, characterized in that, The Generating Agent 1, Debating Agent 1, Decision Agent 1, Generating Agent 2, Debating Agent 2, and Decision Agent 2 all use a large language model as their core reasoning engine and achieve role and task division through differentiated prompting engineering.

6. The method of claim 5, wherein, The granularity parameter G is a positive integer explicitly specified by the user based on the abstraction level requirements of the summary.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 6.

Images