Network topic multi-dimensional recognition method and device based on multi-agent debate cooperation

Abstract

This invention discloses a method and apparatus for multi-dimensional identification of network topics based on multi-agent debate collaboration. The method includes: a central control agent scheduling multiple business-dimensional agents to analyze and generate single-dimensional reports in parallel; after summarizing the multi-dimensional reports, a debate process is initiated. Each agent extracts conclusions and evidence to generate speaking materials and presents them sequentially; the central control agent calculates the semantic distance of the conclusions to construct a difference matrix, matches the dissenting parties, and generates questioning and response content, thereby extracting global consensus, dissent grouping, and dissent type; conflicts are resolved hierarchically according to dissent type; if any of the following conditions are met—maximum number of debate rounds, low consensus increment, or excessively small dissent index—an arbitration agent is activated to fuse and output a unified identification conclusion. Based on the debate process information, three types of confidence scores—data, analysis, and collaboration—are calculated and fused into a comprehensive confidence score; after reaching the target, the network topic identification result is output. This invention can improve the ability of multi-agent collaborative identification of network topics and the stability of the identification result output.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for multi-dimensional identification of network topics based on multi-agent debate collaboration. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] The existing multi-agent-based multi-dimensional identification schemes for network topics have the following technical problems: First, when agents from different business dimensions independently output analysis conclusions, multi-dimensional viewpoint conflicts are easily generated. Existing schemes have not built a complete multi-agent debate interaction link, making it impossible to achieve unified convergence of disagreements. The handling of viewpoint disagreements relies on manual intervention, and the system's collaborative identification capability is insufficient. Second, existing schemes have not established a linkage and coupling mechanism between the debate consensus disagreement state and confidence assessment. They cannot combine the information produced during the debate process to quantify the credibility of the identification conclusions. Ultimately, the credibility of the multi-dimensional identification results for network topics lacks quantitative support, and the overall identification result output has poor stability. Summary of the Invention

[0004] This invention provides a multi-dimensional identification method for network topics based on multi-agent debate collaboration, to improve the ability of multi-agent collaborative identification of network topics and the stability of identification result output. The method includes: The central control agent acquires multi-source data on the network topics to be identified and schedules multiple business dimension agents to perform parallel and independent analysis; each business dimension agent generates a single-dimensional analysis report matching its own dimension based on the multi-source data on the network topics to be identified. The central control agent summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business-dimensional agents to perform debate operations: Each business dimension agent extracts identification conclusions and supporting evidence from its corresponding single-dimensional analysis report to generate its own speaking materials; the central control agent schedules each business dimension agent to complete its speaking based on its own speaking materials according to the order of presentation. The overall control agent determines the semantic distance of the recognition conclusions of all business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix. Based on the difference matrix, it matches the business dimension agents with disagreements as the two sides of the challenge and generates the content of the challenge and the content of the evidence response. Based on multi-dimensional analysis reports, statements, questions and evidence responses, the overall control agent extracts global consensus, disagreement groups and disagreement types, which include: conflicting conclusions, conflicting evidence and conflicting methods. The central control agent resolves disagreements based on the global consensus content, disagreement groups, and disagreement types. It identifies opposing questioning parties based on disagreement groups, compares supporting evidence from both sides for evidence conflicts, distinguishes the order of acceptance of conclusions from various business dimensions for methodological conflicts, and comprehensively balances the conclusions of all parties for conclusion conflicts based on the global consensus content. When any of the following conditions are met, the arbitration agent is activated to integrate the conclusions of multiple parties and obtain a unified conclusion. The central control agent determines the data confidence, analysis confidence, and collaboration confidence corresponding to the unified identification conclusion based on multi-dimensional analysis reports, global consensus content, divergence grouping, and divergence type, and integrates them to obtain the comprehensive confidence. When the comprehensive confidence reaches the automatic decision threshold, it outputs the multi-dimensional identification result of the network topic.

[0005] This invention also provides a multi-dimensional network topic identification device based on multi-agent debate collaboration, to improve the ability of multi-agent collaborative identification of network topics and the stability of identification result output. The device includes: The central control agent acquires multi-source data on the network topic to be identified, and schedules multiple business dimension agents to perform parallel and independent analysis; it summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business dimension agents to perform debate operations, scheduling each type of business dimension agent to complete its statement based on its own speaking materials according to the order of presentation; it determines the semantic distance of the identification conclusions of all types of business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix; it matches business dimension agents with disagreements as the questioning parties based on the difference matrix, and generates questioning content and evidence response content; based on the multi-dimensional analysis report, statements, questioning content, and evidence response content, it extracts global consensus content, disagreement groups, and disagreement types, including: conclusion conflict, evidence conflict, and methodological conflict; and based on the global consensus... The system handles disagreement resolution based on content, disagreement grouping, and disagreement type. It identifies opposing questioning parties based on disagreement grouping, compares supporting evidence from both sides to address evidence conflicts, distinguishes the acceptance order of conclusions from various business dimensions for methodological conflicts, and comprehensively balances the conclusions identified by all parties based on global consensus for conflicting conclusions. The arbitration agent is activated when any of the following conditions are met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increase in consensus is below the improvement threshold, or the disagreement index is below the preset disagreement threshold. Based on multi-dimensional analysis reports, global consensus content, disagreement grouping, and disagreement type, the system determines the data confidence, analysis confidence, and collaborative confidence corresponding to the unified identification conclusion, and merges them to obtain a comprehensive confidence score. When the comprehensive confidence score reaches the automatic decision threshold, the system outputs the multi-dimensional identification results of the network topic. Each business dimension intelligent agent is used to generate a single-dimensional analysis report that matches its own dimension based on multi-source data of the network topic to be identified; extract identification conclusions and supporting evidence from its own corresponding single-dimensional analysis report to generate its own speaking materials; and complete its presentation based on its own speaking materials under the scheduling of the overall control intelligent agent. The arbitration agent is used to integrate the identification conclusions of multiple parties under the scheduling of the central control agent, and transmit the unified identification conclusion to the central control agent.

[0006] This invention also provides a computer device, 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 above-described method for multi-dimensional identification of network topics based on multi-agent debate collaboration.

[0007] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for multi-dimensional identification of network topics based on multi-agent debate collaboration.

[0008] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for multi-dimensional identification of network topics based on multi-agent debate collaboration.

[0009] The beneficial technical effects of the multi-agent debate collaboration-based network topic multi-dimensional identification scheme provided in this invention are: First, this embodiment of the invention constructs a complete multi-agent debate interaction link. Through the statements and mutual questioning of agents in various business dimensions, multi-dimensional viewpoint conflicts are identified, and disagreement resolution is performed in layers for different types of disagreements. When the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold, the arbitration agent is activated to unify and merge the disagreement viewpoints of all parties, output a unique unified identification conclusion, and complete the convergence of viewpoints. This eliminates the need for manual reconciliation of viewpoint disagreements, effectively improves the collaborative identification capability of multi-agents, and solves the problems of conflicting conclusions in multi-dimensional analysis and the reliance on manual intervention in disagreement handling.

[0010] Secondly, this invention establishes a linkage and coupling mechanism between the consensus disagreement state in debate and the confidence assessment. Based on the information generated during the debate process, data confidence, analysis confidence, and collaborative confidence are calculated and fused to obtain a comprehensive confidence. This quantitatively characterizes the credibility of the network topic identification conclusion, provides a quantitative judgment standard for the identification result output, and solves the problems of the lack of quantitative support for the credibility of the identification conclusion and the poor stability of the identification result output. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart illustrating the multi-dimensional identification method for network topics based on multi-agent debate collaboration in an embodiment of the present invention.

[0012] Figure 2 This is a flowchart illustrating a multi-dimensional identification method for network topics based on multi-agent debate collaboration, as described in another embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the system architecture for multi-dimensional identification of network topics based on multi-agent debate collaboration in an embodiment of the present invention.

[0014] Figure 4 This is a schematic diagram of the structure of a network topic multi-dimensional identification device based on multi-agent debate collaboration in an embodiment of the present invention.

[0015] Figure 5 This is a schematic diagram of a computer device structure according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0017] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0018] Figure 1 This is a flowchart illustrating the multi-dimensional identification method for network topics based on multi-agent debate collaboration in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step 101: The central control agent acquires multi-source data of the network topic to be identified and schedules multiple business dimension agents to perform parallel and independent analysis; each business dimension agent generates a single-dimensional analysis report matching its own dimension based on the multi-source data of the network topic to be identified. Step 102: The central control agent summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business dimension agents to perform debate operations: Step 1021: Each business dimension agent extracts identification conclusions and supporting evidence from its corresponding single-dimensional analysis report and generates its own speaking materials; the central control agent schedules each business dimension agent to complete its statement based on its own speaking materials according to the order of presentation. Step 1022: The central control agent determines the semantic distance of the identification conclusions of all business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix. Based on the difference matrix, it matches the business dimension agents with disagreements as the two sides of the challenge and generates the challenge content and evidence response content. Step 1023: The central control agent extracts global consensus content, disagreement groups, and disagreement types based on the multi-dimensional analysis report, statements, questioning content, and evidence response content. The disagreement types include: conclusion conflict, evidence conflict, and methodological conflict. Step 1024: The central control agent performs disagreement resolution based on the global consensus content, disagreement groups, and disagreement types. It identifies opposing questioning parties based on disagreement groups, compares the supporting evidence of both parties in case of evidence conflict, distinguishes the order of acceptance of conclusions identified by the agent in various business dimensions in case of methodological conflict, and comprehensively balances the identification conclusions of all parties in case of conclusion conflict based on the global consensus content. When any of the following conditions are met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold, the arbitration agent is activated to integrate the identification conclusions of multiple parties and obtain a unified identification conclusion. Step 103: Based on the multi-dimensional analysis report, global consensus content, divergence grouping, and divergence type, the central control agent determines the data confidence, analysis confidence, and collaboration confidence corresponding to the unified identification conclusion, and integrates them to obtain the comprehensive confidence. When the comprehensive confidence reaches the automatic decision threshold, the multi-dimensional identification result of the network topic is output.

[0019] The multi-dimensional identification method for network topics based on multi-agent debate collaboration provided in this invention operates as follows: The central control agent first collects multi-source data on network topics and distributes it to various business dimension agents for parallel analysis, generating single-dimensional analysis reports for each dimension. After summarizing all single-dimensional reports to form a multi-dimensional analysis report, the multi-agent debate process is initiated. Each agent extracts its own conclusions and supporting evidence to form speaking materials and presents them sequentially. The central control agent calculates the semantic distance of each identification conclusion to construct a difference matrix, matches the dissenting parties, and generates questioning and response content, extracting global consensus, dissent grouping, and three types of conflict: conclusion, evidence, and method. Conflict resolution is carried out layer by layer for different types of conflict. If any condition—debate round, consensus increment, or dissent indicator—is triggered, the arbitration agent is activated to integrate the viewpoints of all parties to obtain a unified identification conclusion. Finally, combining the consensus and dissent information generated throughout the debate process, the confidence levels of data, analysis, and collaboration are calculated and merged into a comprehensive confidence level. Only when the comprehensive confidence level meets a preset threshold is the final multi-dimensional identification result of the network topic output.

[0020] The beneficial technical effects of the multi-agent debate collaboration-based network topic multi-dimensional identification method provided in this invention are: First, this embodiment of the invention constructs a complete multi-agent debate interaction link. Through the statements and mutual questioning of agents in various business dimensions, multi-dimensional viewpoint conflicts are identified, and disagreement resolution is performed in layers for different types of disagreements. When the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold, the arbitration agent is activated to unify and merge the disagreement viewpoints of all parties, output a unique unified identification conclusion, and complete the convergence of viewpoints. This eliminates the need for manual reconciliation of viewpoint disagreements, effectively improves the collaborative identification capability of multi-agents, and solves the problems of conflicting conclusions in multi-dimensional analysis and the reliance on manual intervention in disagreement handling.

[0021] Secondly, this invention establishes a linkage and coupling mechanism between the consensus disagreement state in debate and the confidence assessment. Based on the information generated during the debate process, data confidence, analysis confidence, and collaborative confidence are calculated and fused to obtain a comprehensive confidence. This quantitatively characterizes the credibility of the network topic identification conclusion, provides a quantitative judgment standard for the identification result output, and solves the problems of the lack of quantitative support for the credibility of the identification conclusion and the poor stability of the identification result output.

[0022] The following section provides a detailed introduction to the multi-dimensional identification method for network topics based on multi-agent debate collaboration.

[0023] In step 101 above, the central control agent actively collects multi-source heterogeneous network data such as images and text, comments, propagation links, and network topic data corresponding to the network topic to be identified. After completing data regularization and preprocessing, it synchronously schedules multiple business dimension agents such as propagation analysis agent, sentiment analysis agent, subject analysis agent, and risk assessment agent to start parallel independent analysis operations.

[0024] The four types of intelligent agents for different business dimensions correspond to different identification focuses and dedicated analysis models: 1. Propagation Analysis Agent: The underlying layer can adopt the SIR propagation dynamics model and the BiLSTM-Attention model; the role of the SIR model is to characterize the information diffusion pattern, input: network-wide forwarding, comments, and time-series propagation data, and output: propagation growth rate and coverage indicators; the role of the BiLSTM-Attention model is to predict the propagation evolution trend, input: time-segmented propagation feature sequence, and output: short-term propagation prediction results.

[0025] 2. Sentiment Analysis Agent: The underlying model can be a finely tuned RoBERTa-base-Chinese main model and a TextCNN auxiliary model. RoBERTa's role is to extract deep sentiment semantics from the text. Input: a single piece of network text content. Output: multi-dimensional sentiment intensity score. TextCNN's role is to quickly and coarsely classify sentiment polarity. Input: a sequence of text word vectors. Output: positive / neutral / negative labels.

[0026] 3. **Subject Analysis Agent:** The underlying layer can be combined with BERT and NER entity recognition models, and SVM classification models. BERT and NER extract KOL, media, and ordinary user entities; input: all comments and post text; output: a list of participating entities. SVM classifies entity influence; input: followers, verification, and interaction features; output: entity importance level. 4. **Risk Assessment Agent:** The underlying layer uses XGBoost and LightGBM dual-classification models. XGBoost performs a coarse risk level assessment through multi-feature fusion; input: dissemination, sentiment, and all statistical features of the entity; output: a basic risk score. LightGBM performs refined risk calibration; input: XGBoost output score and sensitive features; output: a final high / medium / low risk level.

[0027] The four types of business-dimensional agents correspond to different identification focuses: the propagation analysis agent focuses on the scope, speed, and path of propagation; the sentiment analysis agent focuses on sentiment tendencies, intensity, and evolutionary patterns; the subject analysis agent identifies participating subjects such as KOLs, media, and ordinary users; and the risk assessment agent identifies the sensitivity of events and the probability of risk occurrence. Each business-dimensional agent operates independently without interference, approaching the pre-processed multi-source data from its own specific business dimension. It performs targeted analysis, feature extraction, pattern analysis, and content analysis, combining the corresponding dimension's analysis rules and judgment criteria to generate a single-dimensional analysis report that precisely matches its own business dimension, containing specific identification conclusions and supporting evidence. This provides comprehensive and detailed basic analysis data for subsequent multi-agent debates, disagreement resolution, and conclusion fusion.

[0028] In practice, step 101 above involves multiple business-dimensional intelligent agents conducting single-dimensional analysis independently and in parallel. This allows for a comprehensive breakdown of online topic data from different perspectives, accurately uncovering topic features and core information within a single dimension, and effectively avoiding the problems of information bias and feature omission inherent in single-perspective analysis. Simultaneously, the parallel analysis mode significantly improves the efficiency of parsing and processing online topic data. The collaboration of various independent intelligent models ensures the professionalism and objectivity of the analysis results across all dimensions, providing solid, detailed, and multi-dimensional foundational data support for subsequent multi-agent debate collaboration, convergence of viewpoints, and credible quantitative evaluation. This ensures the comprehensiveness and accuracy of multi-dimensional identification results for online topics from the outset.

[0029] As can be seen from the above, in one embodiment, the multi-business dimension intelligent agents may include: a propagation analysis intelligent agent, a sentiment analysis intelligent agent, a subject analysis intelligent agent, and a risk assessment intelligent agent.

[0030] In practice, after the central control agent retrieves multi-source data on network topics, it simultaneously schedules the propagation analysis agent, sentiment analysis agent, and subject analysis agent to conduct independent analyses in parallel. The four types of agents generate their own single-dimensional analysis reports from the specific dimensions of propagation trend, sentiment tendency, event subject, and risk level. Subsequently, the differentiated conclusions of the multiple parties are unified according to the process of complete debate, disagreement resolution, and arbitration integration. Then, the consensus and disagreement information produced by the debate are combined to calculate multi-level confidence levels. Finally, the complete network topic identification results are output based on the comprehensive confidence threshold.

[0031] In practical implementation, this invention sets up four types of differentiated professional intelligent agents, which can fully cover the analytical dimensions required for network topic identification from multiple perspectives such as dissemination, emotion, subject, and risk, and realize the all-round decomposition of event information, avoiding the information loss caused by single-dimensional analysis. After the multi-dimensional professional identification conclusions are unified through debate, layered resolution and arbitration, they can take into account the objective analysis results of each dimension. Combined with multi-dimensional confidence quantification evaluation, it further improves the comprehensiveness, accuracy and stability of network topic identification results.

[0032] In step 102 above, the central control agent collects and integrates the single-dimensional analysis reports output by all business dimension agents. It summarizes, integrates, and standardizes the analysis content, identification conclusions, and supporting evidence from each report, generating a multi-dimensional analysis report covering all dimensions of information, including dissemination, sentiment, subject, and risk. After report integration, the central control agent uniformly schedules all business dimension agents to initiate and execute standardized multi-agent debate collaboration operations. The complete debate process consists of four rounds: the first round of presentations, the second round of cross-examination, the third round of consensus building, and the fourth round of comprehensive identification. The system presets a maximum of five debate rounds; once the maximum number of rounds is reached, the system forcibly enters the comprehensive identification phase. The specific process is as follows: Each business dimension agent accurately reads its corresponding single-dimensional analysis report, filters and extracts core identification conclusions, as well as corresponding raw data, analysis basis, supporting materials, and other supporting evidence, and structures and generates exclusive speaking materials adapted to the debate scenario. The central control agent, according to the preset dimension presentation order, sequentially schedules each business dimension agent to complete a public presentation based on the prepared speaking materials, fully outputting the analytical viewpoints and core evidence of its dimension. Subsequently, the central control agent vectorizes the identification conclusions of all agents, calculates the semantic distance between the identification conclusions of different dimensions, constructs a complete opinion difference matrix based on all semantic distance data, accurately depicts the opinion difference relationships between the agents, and matches business dimension agents with obvious opinion differences as the questioning parties based on the difference matrix. It then automatically generates targeted questioning content and compliant evidence response content for the differences in the identification conclusions of both parties. The questioning and response are automatically generated based on the opinion difference vector combined with a fixed template. The generated response content includes four structured parts: evidence citation, logical reasoning, and position adjustment. An example interaction logic is that the propagation analysis agent questions the sentiment analysis agent: Does emotional fluctuation affect propagation? The sentiment analysis agent then questions the subject analysis agent: Is the subject's emotion considered? Each agent outputs corresponding response content for each question received. The central control agent further integrates multi-dimensional analysis reports, statements from each agent, and responses to two-way questioning and evidence to conduct global information fusion analysis. This accurately extracts the global consensus content and conflicting groups among the multiple agents, and categorizes these conflicts into three core types: conclusion conflict, evidence conflict, and methodological conflict. This invention uses a semantic similarity algorithm as the core means to distinguish between global consensus content and conflicting groups, with weighted voting and spectral clustering as auxiliary algorithms. The system presets a basic consensus threshold of 0.8 and a basic questioning threshold of 0.5. Based on this, the central control agent conducts layered and targeted conflict resolution processing according to the global consensus content, conflicting grouping results, and specific conflict types: accurately locating opposing questioning parties through conflicting grouping, and horizontally comparing the authenticity, completeness, and validity of supporting evidence from both sides regarding evidence conflicts.The system prioritizes credibility based on authoritative information channels, industry media, certified self-media, ordinary personal accounts, and anonymous online content. For method conflicts, it clarifies the order of acceptance for different identification conclusions by considering the analysis priorities and applicable scenarios of each agent. A hybrid algorithm combining Bayesian probabilistic fusion and weighted voting is used to fuse conclusions from multiple methods. If the difference in risk level between disagreements is no more than one level, a conservative principle is applied, prioritizing the identification conclusion with the higher risk level. If the disagreement spans two or more risk levels, the conservative principle cannot be applied, and manual review is triggered directly. For conclusion conflicts, the system comprehensively weighs and balances the identification conclusions from all parties based on global consensus. The system uses dynamic weighting of agents to perform weighted statistics on conclusions. Agent weights are calculated based on historical accuracy, recent performance, and current self-assessed confidence level. After calculation, the weights are normalized and have upper and lower limits set: a lower limit of 0.15 and an upper limit of 0.40. Weights exceeding these limits are truncated and re-normalized. During the iterative process of resolving disagreements, the conditions for terminating the debate are verified in real time. When any of the following conditions are met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increase in consensus is lower than the preset improvement threshold, or the disagreement index is lower than the preset disagreement threshold, the arbitration agent is immediately activated to integrate and merge the differentiated identification conclusions of the multiple parties, and output a unified and convergent overall identification conclusion. The arbitration agent does not participate in the basic dimensional analysis, but only completes the adjudication for irreconcilable conflicts; it matches exclusive decision logic for the three types of conflicts: conclusion, evidence, and method, and adjusts the confidence weights of each agent in this round simultaneously after the adjudication is completed.

[0033] In specific implementation, step 102 above establishes a standardized, full-process multi-agent debate interaction system, transforming independent static identification conclusions of each dimension into a dynamic viewpoint interaction process. Through autonomous statements, precise questioning, and evidence-based debate, the differences and conflicts in the identification conclusions of each dimension are fully exposed. Simultaneously, differentiated hierarchical resolution strategies are implemented for three different types of disagreements: evidence, methods, and conclusions. This accurately and efficiently resolves various viewpoint conflicts, avoiding the problem of incomplete conflict resolution caused by a single, general approach to disagreement handling. An arbitration fusion mechanism is triggered based on preset multiple convergence judgment conditions, enabling unified convergence of viewpoints when the debate reaches the optimal termination point, without the need for manual intervention to reconcile disagreements. This completely solves the technical defects of traditional multi-agent independent analysis, such as conflicts in multi-dimensional identification conclusions, difficulty in automatically converging disagreements, and reliance on manual intervention. It significantly improves the cross-dimensional collaborative analysis capabilities of multi-agents and effectively ensures the rationality and efficiency of viewpoint convergence. This invention achieves bidirectional linkage between the debate process and confidence assessment: the consensus and disagreement information generated by the debate serve as input for collaborative confidence, and the confidence value, in turn, controls whether the debate terminates or whether arbitration is triggered, breaking the limitations of existing technologies where the two operate independently and are decoupled.

[0034] In one embodiment, in step 102 above, the above-mentioned multi-agent debate collaboration-based network topic multi-dimensional identification method may further include: triggering the execution of the debate operation when any of the following conditions are not met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the divergence index is lower than the preset divergence threshold.

[0035] In specific implementation, in this embodiment of the invention, after the disagreement resolution process is completed, the overall control agent verifies three convergence conditions: the number of debate iterations, the continuous increment of consensus, and the disagreement index. If none of the three conditions are met, the entire debate process—including statement presentation, disagreement matching, and conflict resolution—is restarted, and multi-agent viewpoint interaction continues. The cross-questioning loop is only terminated when the semantic similarity of the conclusions identified by the agents is higher than a preset questioning threshold, thus avoiding meaningless repeated debates.

[0036] In specific implementation, in this embodiment of the invention, the debate process is executed cyclically when the convergence judgment condition is not met. This allows agents of different dimensions to continuously exchange viewpoints, fully analyze various disagreements, gradually narrow the differences in multidimensional recognition conclusions, avoid directly outputting a unified conclusion before the disagreements are fully resolved, and further improve the consistency and reliability of the final recognition conclusion.

[0037] In one embodiment, in step 102 above, when constructing the difference matrix, the cosine distance between the identification conclusion vectors is used to represent the semantic distance, and two groups of business dimension agents with a semantic distance greater than a preset opposition threshold are matched as the questioning parties.

[0038] In specific implementation, in this embodiment of the invention, the central control agent converts the recognition conclusions output by the agents of each business dimension into vector form. The BERT encoder model's function is to convert text-based recognition conclusions into standardized semantic vectors. Input: a text recognition conclusion; Output: a fixed-dimensional semantic feature vector. The BERT encoder converts the recognition conclusions into semantic feature vectors, calculates the cosine distance between vectors as a semantic distance quantification standard, compares the semantic distance with a preset opposition threshold, and filters out two groups of agents whose semantic distance exceeds the threshold, matching them as the questioning sides in the debate. Each agent automatically selects the agent with the largest semantic distance as the object of questioning in this round; if the vector difference is below the threshold, no questioning is required in this round.

[0039] In specific implementation, this embodiment of the invention uses cosine distance quantification to identify semantic differences in the conclusions. This can accurately quantify the degree of disagreement among different dimensions of viewpoints. Based on a unified opposition threshold, agents with obvious viewpoint conflicts are paired for questioning, accurately locating debate interaction objects, reducing meaningless viewpoint interactions, and improving the efficiency of disagreement discovery and debate advancement. In step 103 above, after obtaining the unified identification conclusion after the convergence of the multi-agent debate, the central control agent combines the pre-generated multi-dimensional analysis report, the global consensus content accumulated during the debate process, the disagreement grouping results, and the specific disagreement types to conduct a credible quantification evaluation from three dimensions: data source, single-dimensional analysis quality, and multi-agent collaborative debate effect. Specifically, it calculates the corresponding data confidence level by combining the data completeness, information source authority, and authenticity of the multi-source data of the network topic to be identified. Data confidence is calculated by weighting three sub-indicators: data source reliability, data sufficiency, and data timeliness, with corresponding weights of 0.4, 0.3, and 0.3, respectively. The system presets fixed confidence scores for different information channels. Combining the analytical logic standardization, feature matching degree, and analytical accuracy of the single-dimensional analysis reports from each business dimension's intelligent agents, the corresponding analytical confidence is quantified. The analytical confidence score is weighted based on the historical accuracy of the comprehensive model, consistency of multi-method analysis, and the proportion of abnormal data. Abnormal data is detected using the Isolation Forest model. The Isolation Forest's role is to identify abnormal network data that deviates from the normal distribution. Input: multi-dimensional features of all data; Output: normal / abnormal data labels. The collaboration confidence corresponding to multi-agent debate collaboration is calculated by combining the coverage of global consensus, the convergence degree of viewpoints in divergent groupings, and the completion degree of divergence resolution. The collaborative confidence score is calculated by combining the weighted consensus score of the fusion agent, the proportion of explainable disagreements, and the historical recognition accuracy of similar network topics, with corresponding weights of 0.4, 0.3, and 0.3, respectively. The overall confidence score uses a fixed weighting ratio: data confidence score 0.3, analysis confidence score 0.4, and collaborative confidence score 0.3. The central control agent performs weighted fusion calculations on the data confidence score, analysis confidence score, and collaborative confidence score according to preset weighting fusion rules to obtain the overall confidence score corresponding to the unified recognition conclusion. Finally, the overall confidence score is compared with a preset automatic decision threshold. Only when the overall confidence score meets or exceeds the automatic decision threshold is the current unified recognition conclusion deemed credible and valid, and the standardized multi-dimensional recognition result of the network topic is finally output.The system is equipped with a human-machine collaborative routing mechanism: when the overall confidence level is lower than the basic human intervention threshold of 0.6, when there are significant disagreements, or when sensitive network topics are involved, the system automatically pushes complete identification data, disagreement information, and data traceability information to human review; human support includes five types of operations: correcting the final identification conclusion, adjusting the risk level, rejecting the corresponding agent, supplementing supporting data, and correcting the credibility of the data source; the human correction results will be written back to the system in three ways: first, to update the historical identification accuracy of the corresponding agent; second, to iteratively calibrate all basic judgment thresholds of the system; and third, to incrementally fine-tune the underlying analysis model of each agent.

[0040] In practical implementation, step 103 above constructs a multi-dimensional, hierarchical quantitative evaluation system for confidence metrics. This overcomes the shortcomings of traditional identification schemes that lack quantitative evaluation criteria. It comprehensively quantifies the credibility of identification conclusions from three core dimensions: data source quality, single-dimensional analysis capability, and multi-agent collaborative convergence effect, avoiding the one-sided and distorted credibility judgments caused by a single evaluation dimension. By obtaining a comprehensive confidence score through weighted fusion of multiple indicators and using a fixed automatic decision threshold as the output standard, it achieves standardized and quantitative judgment of network topic identification results. This effectively avoids result bias caused by subjective judgment, solves the technical problems of traditional schemes lacking quantitative support for the credibility of identification results and having poor output stability, and significantly improves the accuracy, reliability, and stability of multi-dimensional network topic identification results. All system thresholds are not fixed constants and can be dynamically adjusted adaptively according to the sensitivity of online topics. The sensitivity of online topics is calculated through a three-level keyword database matching combined with the RoBERTa classification model in a dual-channel manner. The function of the RoBERTa sensitivity classification model is to predict the continuous sensitivity score of topics based on the text. Input: complete online topic text, output: sensitivity score in the range of 0 to 1, divided into high, medium and low levels. Sensitive networks automatically lower the automatic decision threshold and raise the consensus judgment threshold to reduce the probability of missed identification risks.

[0041] In one embodiment, in step 103 above, the data confidence is calculated based on the completeness and source credibility of the multi-source data of the network topic to be identified; the analysis confidence corresponds to the credibility of the single-dimensional analysis report of the intelligent agent in each business dimension; and the collaboration confidence is calculated based on the convergence degree of viewpoints of global consensus content and divergent groups.

[0042] In specific implementation, in this embodiment of the invention, the overall control agent calculates three types of confidence scores: data confidence scores are calculated by combining the completeness of multi-source data and the credibility of the sources; analysis confidence scores are obtained based on the quality of the single-dimensional analysis reports generated by the agents of each business dimension; collaboration confidence scores are calculated based on the content of global consensus and the degree of convergence of viewpoints reflected by divergent groups; and finally, the three types of confidence scores are fused to obtain a comprehensive confidence score, which is used to determine whether to output the recognition result.

[0043] In practical implementation, the embodiments of the present invention calculate confidence levels hierarchically from three independent dimensions: data source, single-dimensional analysis, and multi-agent collaborative convergence. This comprehensively covers the key factors affecting the reliability of the identification results, accurately distinguishes the risk of results caused by data defects and analysis biases, and makes the comprehensive confidence quantification results more objective and comprehensive, thereby improving the rationality and accuracy of the identification result judgment.

[0044] Figure 2 This is a flowchart illustrating a multi-dimensional identification method for network topics based on multi-agent debate collaboration in another embodiment of the present invention, as shown below. Figure 2 As shown, in one embodiment, the above-mentioned multi-agent debate collaboration-based network topic multi-dimensional identification method may further include step 104: the overall control agent dynamically updates the global judgment parameters according to the comprehensive confidence level, wherein the global judgment parameters include the maximum number of debate rounds, the divergence threshold, and the improvement threshold.

[0045] In specific implementation, after outputting the multi-dimensional identification results of network topics, the overall control agent dynamically updates the global judgment parameters based on the comprehensive confidence level calculated this time. The global judgment parameters include the maximum number of debate rounds, the divergence threshold, and the improvement threshold. The updated parameters will be used for the debate process control, divergence judgment, and convergence condition judgment of the next round of network topic identification task. At the same time, the manual review feedback will synchronously calibrate the basic threshold, count the recent manual correction ratio, and iteratively optimize the benchmark values ​​of each threshold to avoid the analysis becoming rigid due to the long-term fixation of the threshold.

[0046] In specific implementation, the embodiments of the present invention dynamically adjust the global judgment parameters related to the debate based on the comprehensive confidence feedback. It can adaptively adjust the debate convergence standard according to the credibility of the results in different recognition scenarios, avoiding the rigidity of the process caused by fixed parameters. When the confidence is low, the debate constraints can be automatically relaxed and the number of interaction rounds can be increased to fully resolve differences. When the confidence is high, the debate process can be shortened to improve processing efficiency, taking into account both the analysis accuracy and overall processing efficiency of network topic recognition.

[0047] To facilitate understanding of how this invention is implemented, the following will further explain... Figure 3This paper introduces the system architecture of an embodiment of the present invention. The network topic multi-dimensional identification system based on multi-agent debate collaboration of this invention adopts a hierarchical collaborative agent architecture. The various functional components of the system are sequentially connected to form a complete data flow. The output results of the confidence assessment stage are fed back to the debate stage and the conflict resolution stage. The overall architecture mainly consists of a central control agent, multiple business dimension agents, and an arbitration agent. Each agent has a clear division of labor and collaborates collaboratively, forming a complete working architecture of "independent analysis - debate interaction - divergence convergence - credible quantification - result output". Among them, the multi-business-dimensional intelligent agents include propagation analysis intelligent agents, sentiment analysis intelligent agents, subject analysis intelligent agents, and risk assessment intelligent agents. They are responsible for conducting parallel and independent analysis of multi-source data on online topics from their respective exclusive business dimensions, outputting differentiated single-dimensional analysis reports, and realizing multi-feature and all-round information mining of online topics. The central control intelligent agent, as the core scheduling and processing hub, undertakes the overall coordination of the entire process of data acquisition, task scheduling, report summarization, debate process control, disagreement identification and resolution, confidence calculation, and result judgment output. By constructing a difference matrix, matching the questioning parties, and distinguishing multiple types of disagreements, it realizes standardized intelligent agent debate interaction and hierarchical disagreement resolution. The arbitration intelligent agent, as the viewpoint convergence unit, starts working when the debate meets the preset convergence conditions and completes the fusion and unification of the differentiated identification conclusions of multiple parties. This architecture effectively solves the problems of one-sidedness in traditional single-dimensional analysis, conflicting viewpoints in independent analysis by multiple agents, difficulty in automatically converging disagreements, and inability to quantify the credibility of results by combining multi-agent division of labor and cooperation, interactive checks and balances in debate, and quantitative confidence verification. It ensures the automation, orderliness, accuracy, and stability of the multi-dimensional identification process of network topics. Compared with existing pipeline-style multi-agent network topic analysis systems, the core innovation of this invention lies in building a two-way linkage mechanism between debate collaboration and confidence assessment. It relies on the consensus and disagreement information generated during the debate process to drive confidence calculation in real time. The confidence value in turn controls the debate rounds and arbitration triggering conditions, thus solving the defects of existing technologies where the two are independent and lack linkage feedback.

[0048] This invention also provides a device for multi-dimensional identification of network topics based on multi-agent debate collaboration, as described in the following embodiments. Since the principle behind this device is similar to that of the method for multi-dimensional identification of network topics based on multi-agent debate collaboration, the implementation of this device can refer to the implementation of the method for multi-dimensional identification of network topics based on multi-agent debate collaboration; repeated details will not be elaborated further.

[0049] Figure 4 This is a schematic diagram of the structure of a network topic multi-dimensional identification device based on multi-agent debate collaboration in an embodiment of the present invention, as shown below. Figure 4 As shown, the device includes: The central control agent 01 is used to acquire multi-source data on the network topic to be identified, and to schedule multiple business dimension agents to perform parallel and independent analysis; it summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business dimension agents to perform debate operations, scheduling each type of business dimension agent to complete its statement based on its own speaking materials according to the order of presentation; it determines the semantic distance of the identification conclusions of all types of business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix, and matches business dimension agents with disagreements as the questioning parties based on the difference matrix, generating questioning content and evidence response content; based on the multi-dimensional analysis report, statements, questioning content and evidence response content, it extracts global consensus content, disagreement groups and disagreement types, including: conclusion conflict, evidence conflict, and methodological conflict; based on the global consensus... The system identifies content, groups disagreements, and types of disagreements to resolve them. It locates opposing questioning parties based on disagreement groups, compares supporting evidence from both sides to address evidence conflicts, and differentiates the acceptance order of conclusions from various business dimensions based on methodological conflicts. For conflicting conclusions, it comprehensively balances the conclusions from all parties based on global consensus. The arbitration agent is activated when any of the following conditions are met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increase in consensus is below the improvement threshold, or the disagreement index is below the preset disagreement threshold. Based on multi-dimensional analysis reports, global consensus content, disagreement groups, and disagreement types, it determines the data confidence, analysis confidence, and collaborative confidence corresponding to the unified identification conclusion, and merges them to obtain a comprehensive confidence score. When the comprehensive confidence score reaches the automatic decision threshold, it outputs the multi-dimensional identification results of the network topic. Each business dimension intelligent agent 02 is used to generate a single-dimensional analysis report that matches its own dimension based on multi-source data of the network topic to be identified; extract identification conclusions and supporting evidence from its own corresponding single-dimensional analysis report to generate its own speaking materials; and complete the presentation based on its own speaking materials under the scheduling of the overall control intelligent agent. Arbitration agent 03 is used to integrate the identification conclusions of multiple parties under the scheduling of the central control agent, and transmit the unified identification conclusion to the central control agent.

[0050] In one embodiment, the multi-business-dimensional intelligent agents include: a propagation analysis intelligent agent, a sentiment analysis intelligent agent, a subject analysis intelligent agent, and a risk assessment intelligent agent.

[0051] In one embodiment, the overall control agent is further configured to dynamically update global decision parameters based on the comprehensive confidence level, wherein the global decision parameters include the maximum number of debate rounds, the divergence threshold, and the improvement threshold.

[0052] In one embodiment, the overall control agent is further configured to: trigger the execution of the debate operation when any of the following conditions are not met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold.

[0053] In one embodiment, when constructing the difference matrix, the cosine distance between the identification conclusion vectors is used to represent the semantic distance, and two groups of business dimension agents whose semantic distance is greater than a preset opposition threshold are matched as the questioning parties.

[0054] In one embodiment, the data confidence is calculated based on the completeness and source credibility of the multi-source data of the network topic to be identified; the analysis confidence corresponds to the credibility of the single-dimensional analysis report of the intelligent agent in each business dimension; and the collaboration confidence is calculated based on the convergence degree of viewpoints of global consensus content and divergent groups.

[0055] Based on the aforementioned inventive concept, such as Figure 5 As shown, the present invention also proposes a computer device 500, including a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the aforementioned multi-agent debate collaboration-based network topic multi-dimensional identification method.

[0056] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for multi-dimensional identification of network topics based on multi-agent debate collaboration.

[0057] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for multi-dimensional identification of network topics based on multi-agent debate collaboration.

[0058] The beneficial technical effects of the multi-agent debate collaboration-based network topic multi-dimensional identification scheme provided in this invention are: First, this embodiment of the invention constructs a complete multi-agent debate interaction link. Through the statements and mutual questioning of agents in various business dimensions, multi-dimensional viewpoint conflicts are identified, and disagreement resolution is performed in layers for different types of disagreements. When the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increment of consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold, the arbitration agent is activated to unify and merge the disagreement viewpoints of all parties, output a unique unified identification conclusion, and complete the convergence of viewpoints. This eliminates the need for manual reconciliation of viewpoint disagreements, effectively improves the collaborative identification capability of multi-agents, and solves the problems of conflicting conclusions in multi-dimensional analysis and the reliance on manual intervention in disagreement handling.

[0059] Secondly, this invention establishes a linkage and coupling mechanism between the consensus disagreement state in debate and the confidence assessment. Based on the information generated during the debate process, data confidence, analysis confidence, and collaborative confidence are calculated and fused to obtain a comprehensive confidence. This quantitatively characterizes the credibility of the network topic identification conclusion, provides a quantitative judgment standard for the identification result output, and solves the problems of the lack of quantitative support for the credibility of the identification conclusion and the poor stability of the identification result output.

[0060] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0063] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0064] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-dimensional identification of network topics based on multi-agent debate collaboration, characterized in that, include: The central control agent acquires multi-source data on the network topics to be identified and schedules multiple business dimension agents to perform parallel and independent analysis; each business dimension agent generates a single-dimensional analysis report matching its own dimension based on the multi-source data on the network topics to be identified. The central control agent summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business-dimensional agents to perform debate operations. Each business dimension agent extracts identification conclusions and supporting evidence from its corresponding single-dimensional analysis report to generate its own speaking materials; the central control agent schedules each business dimension agent to complete its speaking based on its own speaking materials according to the order of presentation. The overall control agent determines the semantic distance of the recognition conclusions of all business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix. Based on the difference matrix, it matches the business dimension agents with disagreements as the two sides of the challenge and generates the challenge content and the evidence response content. Based on multi-dimensional analysis reports, statements, questions and evidence responses, the overall control agent extracts global consensus content, disagreement groups and disagreement types. The disagreement types include: conflicting conclusions, conflicting evidence and conflicting methods. The central control agent resolves disagreements based on the global consensus content, disagreement groups, and disagreement types. It identifies opposing questioning parties based on disagreement groups, compares supporting evidence from both sides for evidence conflicts, distinguishes the order of acceptance of conclusions from various business dimensions for methodological conflicts, and comprehensively balances the conclusions of all parties for conclusion conflicts based on the global consensus content. When any of the following conditions are met, the arbitration agent is activated to integrate the conclusions of multiple parties and obtain a unified conclusion. The central control agent determines the data confidence, analysis confidence, and collaboration confidence corresponding to the unified identification conclusion based on multi-dimensional analysis reports, global consensus content, divergence grouping, and divergence type, and integrates them to obtain the comprehensive confidence. When the comprehensive confidence reaches the automatic decision threshold, it outputs the multi-dimensional identification result of the network topic.

2. The method as described in claim 1, characterized in that, The various business-dimensional intelligent agents include: a propagation analysis intelligent agent, a sentiment analysis intelligent agent, a subject analysis intelligent agent, and a risk assessment intelligent agent.

3. The method as described in claim 1, characterized in that, Also includes: The overall control agent dynamically updates the global decision parameters based on the comprehensive confidence level. The global decision parameters include the maximum number of debate rounds, the divergence threshold, and the improvement threshold.

4. The method as described in claim 1, characterized in that, Also includes: The debate operation is triggered when any of the following conditions are not met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increase in consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold.

5. The method as described in claim 1, characterized in that, When constructing the difference matrix, the cosine distance between the identification conclusion vectors is used to represent the semantic distance. Two groups of business dimension agents whose semantic distance is greater than the preset opposition threshold are matched as the questioning parties.

6. The method as described in claim 1, characterized in that, The data confidence score is calculated based on the completeness and source credibility of the multi-source data of the network topic to be identified; the analysis confidence score corresponds to the credibility of the single-dimensional analysis report of the intelligent agent in each business dimension; the collaboration confidence score is calculated based on the global consensus content and the convergence degree of viewpoints in the divergent groups.

7. A multi-dimensional identification device for network topics based on multi-agent debate collaboration, characterized in that, include: The central control agent is used to acquire multi-source data on the network topic to be identified, and to schedule multiple business dimension agents to perform parallel and independent analysis; it summarizes all single-dimensional analysis reports to obtain a multi-dimensional analysis report, and schedules all business dimension agents to perform debate operations. According to the order of presentation, it schedules each type of business dimension agent to complete its presentation based on its own speaking materials; it determines the semantic distance of the identification conclusions of all types of business dimension agents based on the multi-dimensional analysis report and constructs a difference matrix; it matches business dimension agents with disagreements as the questioning parties based on the difference matrix, and generates questioning content and evidence response content. Based on multi-dimensional analysis reports, statements, questioning content, and evidence responses, the system extracts global consensus content, disagreement groups, and disagreement types, including: conclusion conflicts, evidence conflicts, and methodological conflicts. Disagreement resolution is carried out based on the global consensus content, disagreement groups, and disagreement types. The system identifies opposing questioning parties based on disagreement groups, compares supporting evidence for evidence conflicts, distinguishes the acceptance order of conclusions identified by various business dimensions for methodological conflicts, and comprehensively balances the conclusions identified by all parties based on the global consensus content for conclusion conflicts. The arbitration agent is activated when any of the following conditions are met: the number of debate iterations reaches the preset maximum number of debate rounds, the continuous increase in consensus is lower than the improvement threshold, or the disagreement index is lower than the preset disagreement threshold. Based on the multi-dimensional analysis reports, global consensus content, disagreement groups, and disagreement types, the system determines the data confidence, analysis confidence, and collaboration confidence corresponding to the unified identification conclusion, and merges them to obtain a comprehensive confidence score. When the comprehensive confidence score reaches the automatic decision threshold, the system outputs the multi-dimensional identification results of the network topic. Each business dimension intelligent agent is used to generate a single-dimensional analysis report that matches its own dimension based on multi-source data of the network topic to be identified; extract identification conclusions and supporting evidence from its own corresponding single-dimensional analysis report to generate its own speaking materials; and complete its presentation based on its own speaking materials under the scheduling of the overall control intelligent agent. The arbitration agent is used to integrate the identification conclusions of multiple parties under the scheduling of the central control agent, and transmit the unified identification conclusion to the central control agent.

8. A computer 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 computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

Images