Policy text analysis method and device, computer device and storage medium

Abstract

This application discloses a policy text analysis method, apparatus, computer device, and storage medium. The method includes: acquiring policy information to be analyzed; analyzing the policy information using data augmentation recall to obtain retrieval recall data; decomposing the policy information into multiple sub-tasks; determining analysis agents matching the multiple sub-tasks; and using the analysis agents to parse the retrieval recall data in parallel to obtain candidate results for the multiple sub-tasks; and integrating the candidate results for the multiple sub-tasks to obtain the analysis result of the policy information to be analyzed. This application effectively improves the recall rate of actually related text blocks on a data-augmented dataset through various data augmentation recall methods.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent government technology, specifically to a policy text analysis method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Large Language Model (hereinafter referred to as Large Model) is a technology based on deep learning in artificial intelligence. By pre-training on massive amounts of data, it can understand, summarize, generate and translate human language, and demonstrates logical reasoning and creative abilities.

[0003] Based on the excellent language understanding capabilities demonstrated by large-scale models, an increasing number of text processing tasks are incorporating these models to improve efficiency. In the government sector, the interpretation and processing of policy texts is a key task for civil servants. Given the government's accelerated digital transformation, utilizing large-scale models to assist in related work has become an important means for government operations to move towards intelligence and scientific rigor.

[0004] While large models are powerful, they are not always accurate. Currently, they are essentially probability-based prediction systems built on the Transformer model architecture, and they often fabricate false information, a problem commonly known as the "illusion" problem. Therefore, obtaining more accurate and comprehensive analytical results is a major challenge. Summary of the Invention

[0005] This application provides a policy text analysis method, apparatus, computer device, and storage medium that effectively improves the recall rate of actually related text blocks on a data-enhanced dataset through various data augmentation recall methods.

[0006] In a first aspect, embodiments of this application provide a policy text analysis method, the method comprising:

[0007] Obtain the policy information to be analyzed;

[0008] The policy information to be analyzed is analyzed using data augmentation recall to obtain retrieval recall data;

[0009] The policy information to be analyzed is broken down into multiple sub-tasks;

[0010] A matching analytical agent is determined, and the analytical agent parses the retrieval recall data in parallel to obtain candidate results for the multiple sub-tasks.

[0011] The analysis results of the multiple sub-task candidate results are integrated to obtain the analysis results of the policy information to be analyzed.

[0012] In some embodiments, the step of analyzing the policy information to be analyzed using data augmentation recall to obtain retrieval recall data includes:

[0013] The policy information to be analyzed is retrieved and analyzed based on a preset search engine to obtain relevant text of the policy information to be analyzed; and / or,

[0014] Based on a pre-defined policy dictionary, the policy information to be analyzed is parsed to obtain its semantic information; and / or,

[0015] Based on a pre-defined model, the policy information to be analyzed is segmented into problem segments to obtain the core phrases of the policy information; and / or,

[0016] The associated information of the policy information to be analyzed is determined based on the pre-embedded tag information of various policy information in the database. The retrieved data includes the relevant text, the explanatory information, the core phrase, and the associated information.

[0017] In some embodiments, the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer; the relevant text includes a first type of text, a second type of text, and a third type of text; the retrieval and analysis of the policy information to be analyzed based on the preset retrieval engine to obtain relevant text information of the policy information to be analyzed includes:

[0018] Based on the sparse retrieval layer, key information matching is performed on the policy information to be analyzed to recall the first type of text related to the policy information to be analyzed.

[0019] Based on the dense retrieval layer, semantic understanding of the policy information to be analyzed is performed, and the second type of text that is semantically related to the policy information to be analyzed is recalled.

[0020] The third type of text is analyzed based on the knowledge graph layer and is related to the policy information to be analyzed.

[0021] In some embodiments, before decomposing the policy information to be analyzed into multiple sub-tasks, the method further includes:

[0022] The intent information of the policy information to be analyzed is identified based on a preset intent recognition model;

[0023] If the intent information matches a preset intent scenario, then the intent fence corresponding to the preset intent scenario is determined to be the intent fence of the intent information.

[0024] The process of obtaining candidate results for the multiple subtasks by parallel parsing the retrieval and recall data through the analytical agent includes:

[0025] Based on the intent fence, the analytical agent parses the retrieval and recall data in parallel to obtain candidate results for the multiple subtasks.

[0026] In some embodiments, the integration of the candidate results of the plurality of sub-tasks to obtain the analysis result of the policy information to be analyzed includes:

[0027] The candidate results of the multiple sub-tasks are integrated to obtain the analysis basis information of the candidate results of the multiple sub-tasks;

[0028] Based on the analytical criteria, the analysis results of the policy information to be analyzed are obtained by selecting from the candidate results of the multiple sub-tasks.

[0029] In some embodiments, the step of selecting the analysis result of the policy information to be analyzed from the candidate results of the plurality of sub-tasks based on the analysis basis information includes:

[0030] The debate agent identifies conflicting results among the candidate results of the multiple sub-tasks.

[0031] The arbitration agent determines the retained outcome from the conflict results based on the analysis information.

[0032] Based on the retained results and the non-conflicting results among the multiple sub-task candidate results, the analysis results of the policy information to be analyzed are obtained.

[0033] In some embodiments, before determining the analytical agent matching the plurality of subtasks, the method further includes:

[0034] If a new target subtask is added, an analysis agent corresponding to the target subtask will be added.

[0035] Secondly, embodiments of this application provide a policy text analysis device, the device comprising:

[0036] The data acquisition module is used to acquire policy information to be analyzed.

[0037] The data retrieval module is communicatively connected to the data acquisition module and is used to analyze the policy information to be analyzed using data augmentation recall to obtain retrieval recall data.

[0038] The task decomposition module is communicatively connected to the data retrieval module and is used to decompose the policy information to be analyzed into multiple sub-tasks.

[0039] The intelligent analysis module is communicatively connected to the task splitting module and is used to determine the analysis agent that matches the multiple sub-tasks. The analysis agent parses the retrieval and recall data in parallel to obtain the sub-task candidate results of the multiple sub-tasks.

[0040] The results integration module is communicatively connected to the intelligent analysis module and is used to integrate the candidate results of the multiple sub-tasks to obtain the analysis results of the policy information to be analyzed.

[0041] Thirdly, embodiments of this application also provide a computer device, the computer device including a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the operations performed by the policy text analysis method described above.

[0042] Fourthly, embodiments of this application also provide a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to implement the operations performed by the policy text analysis method described above.

[0043] The solution adopted in the application embodiment effectively improves the recall rate of actual associated text blocks on the data-enhanced dataset through various data augmentation recall methods. Under mixed intent questions, it effectively decomposes multiple sub-tasks to improve the accuracy of answers, and the multi-task coordination, division of labor and parallel processing improves task efficiency. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a policy text analysis method provided in an embodiment of this application;

[0046] Figure 2 This is a flowchart illustrating a policy text analysis method provided in an embodiment of this application;

[0047] Figure 3 This is a flowchart illustrating a policy text analysis method provided in an embodiment of this application;

[0048] Figure 4 This is a flowchart illustrating a policy text analysis method provided in an embodiment of this application;

[0049] Figure 5 This is a schematic diagram of the structure of a policy text analysis device according to an embodiment of this application;

[0050] Figure 6 This is a schematic diagram of the structure of a terminal according to an embodiment of this application;

[0051] Figure 7 This is a schematic diagram of the structure of a server in an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0053] It is understood that the terms "first," "second," etc., used in this application may be used to describe various concepts herein, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of this application, a first influence parameter may be referred to as a second influence parameter, and similarly, a second influence parameter may be referred to as a first influence parameter.

[0054] "At least one" refers to one or more event types. For example, at least one event type can be any integer number of event types greater than or equal to one, such as one event type, two event types, three event types, etc. "Multiple" refers to two or more event types. For example, multiple event types can be any integer number of event types greater than or equal to two, such as two event types, three event types, etc. "Each" refers to each of the at least one event type. For example, each event type refers to each of the multiple event types. If the multiple event types are three event types, then each event type refers to each of the three event types.

[0055] It is understood that in the embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0056] By accessing massive amounts of historical policy texts, social statistics, academic research reports, and public opinion information, large-scale models can provide policymakers with profound data-driven insights. Firstly, leveraging its computing power, it can efficiently perform multi-dimensional and in-depth analyses of complex issues, far surpassing human capabilities in efficiency. Secondly, large-scale models also possess simulation and effect prediction capabilities, enabling them not only to quickly generate policy drafts and provide multiple alternatives, but also to warn of potential risks and identify policy blind spots, thereby improving the efficiency and foresight of the decision-making process.

[0057] In highly specialized scenarios like government affairs, the training data for the original models often lacks sufficiently high-quality information, causing large models to perform worse in these specialized domains than in general knowledge domains.

[0058] To minimize the impact of "illusions," a common solution is to use techniques such as Retrieval Augmented Generation (RAG) to provide the large model with additional task-related information as context, enabling the large model to combine the user task with this additional information to generate more accurate and relevant answers.

[0059] RAG is a type of intelligent agent technology that combines large language models and information retrieval. When a large model needs to generate text or answer questions, RAG first retrieves information related to the question from an external knowledge base, and then uses the information as the context for model instructions, thereby improving the quality of the large model's output.

[0060] During the index building phase, the functions of each module are as follows:

[0061] 1. Document Loading: Processes raw documents of various formats and extracts plain text and metadata through document parsing technology.

[0062] 2. Text segmentation: Based on specific logic, long texts are divided into semantically related text chunks. The size of the text chunks will affect the efficiency and quality of subsequent retrieval and model response.

[0063] 3. Text embedding: Using an embedding model, each text block is converted into a high-dimensional vector, mapping the semantic information of the text into a high-dimensional space, which allows the correlation between texts to be quantified using calculation methods such as cosine similarity.

[0064] 4. Vector storage: The text block (and its metadata) and its corresponding text vector are stored together in the vector database, which supports efficient semantic similarity retrieval.

[0065] During the agent's runtime phase, the functions of each module are as follows:

[0066] 1. Query embedding: Using the same embedding model as the index building phase, user queries are also converted into a vector.

[0067] 2. Semantic Recall Coarse Screening: Utilize the index of the vector database to efficiently retrieve the Top-K text blocks with the highest relevance to the user's query vector (usually using cosine similarity).

[0068] 3. Document Rearrangement and Refinement: A more refined rearrangement model is used to reorder and filter the Top-K text blocks selected in the initial screening, resulting in the Top-N text blocks that are most relevant to the user's query.

[0069] 4. Construct enhanced prompt words: Combine the top-N most relevant text blocks obtained after user query and rearrangement into a structured prompt word to enhance the background information that the large model can refer to when generating answers.

[0070] Although RAG utilizes text embedding and vector database retrieval techniques to provide additional contextual information for large models, the "illusion" problem of large models remains severe when document block retrieval results are poor and the provided additional information is insufficient. General RAG solutions lack detection for such "illusion" situations and cannot correct for the "illusion" problem.

[0071] In order to reduce the probability of the "illusion" problem, the RAG scheme needs to improve the quality of document block retrieval results. However, the general semantic relevance measurement method used in the existing scheme is limited by the quality of training data in vertical domains, and there is still much room for improvement in its actual effect in the government sector.

[0072] For example, "business environment" and "building a fair market competition for enterprises" have a low degree of correlation under general semantic relevance measurement. However, in the actual field of government affairs, the connotation of the term "business environment" includes "policy environment, market environment, rule of law environment, and cultural environment," and is mainly geared towards enterprises, and is very closely related to the latter.

[0073] When solving complex tasks that require step-by-step decomposition, general RAG solutions are limited by the capabilities of the model itself and the length of the context, often resulting in problems such as instruction failure, omission of key information, and incorrect format of generated results.

[0074] When solving such complex problems, it is necessary to design more refined intelligent agent workflow processes to help large models better understand task objectives, thereby enabling them to smoothly break down complex tasks into steps and solve them one by one.

[0075] Please refer to Figure 1 and Figure 2 This application provides a policy text analysis method, the specific process of which can be as follows (S110~S150), the method including:

[0076] S110. Obtain the policy information to be analyzed.

[0077] Specifically, the process involves acquiring policy information to be analyzed. This policy information consists of policy text input by the user through a terminal device that requires identification and analysis. For example, the acquired policy information might be the similarities and differences between the incentive policies for new energy vehicle companies in City A and City B.

[0078] S120. The policy information to be analyzed is analyzed using data augmentation recall to obtain retrieval recall data.

[0079] Specifically, data augmentation recall is employed to analyze the policy information to be analyzed, resulting in retrieved recall data. Traditional RAG schemes rely on general semantic similarity calculations, which leads to the inaccurate identification of semantic associations such as proper nouns, abbreviations, and tags in the government affairs field, resulting in insufficient retrieval quality. This embodiment utilizes different forms of augmented data to improve the recall effect.

[0080] In one embodiment, this step includes: S210, performing retrieval and analysis on the policy information to be analyzed based on a preset retrieval engine to obtain relevant text of the policy information to be analyzed; and / or, S220, parsing the policy information to be analyzed based on a preset policy dictionary to obtain explanatory information of the policy information to be analyzed; and / or, S230, segmenting the policy information to be analyzed into questions based on a preset model to obtain core phrases of the policy information to be analyzed; and / or, S240, determining the association information of the policy information to be analyzed based on pre-embedded tag information of various types of policy information in the database, wherein the retrieval recall data includes the relevant text, the explanatory information, the core phrases, and the association information.

[0081] Specifically, based on a preset search engine, the policy information to be analyzed is retrieved and analyzed to obtain relevant texts of the policy information to be analyzed, such as... Figure 2 As shown, the preset search engine adopts a three-way recall architecture of "sparse + dense + knowledge graph" to improve the recall effect.

[0082] This paper introduces a policy dictionary, a data augmentation method that uses general natural language interpretation to explain terminology specific to the government sector. Original policy documents and user questions frequently contain terminology specific to the government sector or common abbreviations. By introducing natural language interpretation and a policy dictionary, the model can accurately grasp the semantics of these terms, significantly improving its ability to understand non-standardized user queries during query expansion and retrieval. The pre-defined policy dictionary includes, but is not limited to, a thesaurus / synonyms, a dictionary of professional terms, and a dictionary of abbreviations.

[0083] Synonym / Near-synonym dictionary: By setting different words with the same or similar meanings in policy documents as synonyms / near-synonyms, the model can more accurately understand the meaning of words, avoid missing relevant content due to different word expressions, and thus give a more accurate answer. For example: "Beijing" = "capital".

[0084] A glossary of professional terms: This provides natural language explanations for professional terms in policy documents, enabling models to accurately grasp the precise semantics of the terms and thus make inferences that align with their true meaning, rather than merely remaining at the literal level. For example, "High-Precision and Cutting-Edge Center" refers to a major science and technology innovation platform led by the Beijing Municipal Government, relying on universities in Beijing, with the participation of multiple innovation entities, and operating as a relatively independent entity.

[0085] Abbreviation dictionary: Explains and supplements abbreviations so that large models can better understand abbreviations in policies.

[0086] The system analyzes policy information by matching and parsing it against a pre-defined policy dictionary, obtaining explanatory information that explains the words and phrases within the policy information. This introduction of professional knowledge effectively prevents large models from misunderstanding newly emerging obscure terms and unclear expressions in the policy data, thus avoiding the problem of pre-trained large models not being exposed to the latest data.

[0087] To effectively locate key phrases in user questions and filter out incomplete text blocks, the query is finely segmented. Based on a pre-defined model, the policy information to be analyzed is segmented to obtain the core phrases. This question segmentation method optimizes the data augmentation approach for extracting keywords from government policy questions.

[0088] The pre-defined model can be a phrase dictionary applicable to the problem domain of the policy information to be analyzed. Weights are assigned to the phrases in the phrase dictionary, and a segmentation method that yields the highest total score for the policy information to be analyzed is determined through dynamic programming or a greedy approach.

[0089] The pre-defined model can be a supervised machine learning problem that segments the policy information to be analyzed, specifically a binary classification problem for characters. For each character, the model determines whether to continue with the current segment or start a new one. By manually labeling the information or collecting a set of correctly segmented questions, a binary classifier can be trained to solve this problem.

[0090] For example, if a user asks "Are university teachers eligible for the new energy vehicle purchase subsidy?", the question can be segmented to identify the core phrases "new energy vehicle purchase subsidy" and "university teachers", ensuring that no key information is missed during the search.

[0091] In addition to their unique meanings, terms in the government affairs domain often possess specific attribute tags. For example, the term "business environment" implicitly targets "enterprises," therefore "enterprises" can be considered a tag for "business environment." Text pairs with the same tags have special semantic associations within the government affairs domain, and corresponding weights can be assigned in the recall algorithm to increase the association score of text pairs with the same tags. Therefore, various policy information in the database is pre-embedded with corresponding tag information, and this tag information is matched with the policy information to be analyzed to determine the relevant association information in the database. Using tag embedding is a data augmentation method to optimize the recall effect of text blocks in the government affairs domain.

[0092] For a specific scenario, user questions can be summarized into a few core tags. In this case, the tag embedding of documents and words can be abstracted into a multi-label classification problem, which can be accomplished by a pre-trained multi-label classifier.

[0093] For example, when searching for "measures to optimize the business environment", the system can automatically recall clauses such as "reducing the burden on enterprises" and "simplifying policy approvals", even if these texts do not explicitly contain the words "business environment".

[0094] By using the above data-enhanced recall method, all retrieved data is placed in an evidence pool for subsequent analysis by the intelligent agent. This allows the questions and document blocks in the policy information to be analyzed to carry more contextual information from the government affairs sector, thereby improving the accuracy of the entire recall, ranking, and question-answering process.

[0095] This embodiment relies not only on unstructured policy texts but also on structured knowledge (policy knowledge graph and policy dictionary). The knowledge graph provides structured support for entities, clauses, and relationships, avoiding omissions in reasoning. The policy dictionary provides synonyms, abbreviations, and terminology expansions, enhancing the agent's ability to understand non-standardized expressions in the political domain. This dual-engine mechanism significantly reduces reasoning errors caused by terminological ambiguity and improves the system's professional depth in policy interpretation and analysis.

[0096] In one embodiment, the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer. The relevant text includes a first type of text, a second type of text, and a third type of text. The preset retrieval engine is used to retrieve and analyze the policy information to be analyzed. Step S210: Retrieving and analyzing the policy information to be analyzed based on the preset retrieval engine to obtain the relevant text of the policy information to be analyzed includes: S310: Performing key information matching on the policy information to be analyzed based on the sparse retrieval layer to recall the first type of text related to the policy information to be analyzed; S320: Performing semantic understanding on the policy information to be analyzed based on the dense retrieval layer to recall the second type of text semantically related to the policy information to be analyzed; S330: Analyzing the third type of text related to the policy information to be analyzed based on the knowledge graph layer.

[0097] Specifically, such as Figure 2 As shown, the hybrid retrieval framework adopts a three-way recall architecture of "sparse + dense + knowledge enhancement," meaning the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer. The sparse retrieval layer, based on the BM25 (Best Match 25) algorithm and combined with a policy corpus, performs excellently in exact matching scenarios (such as keyword and clause number retrieval). The dense retrieval layer, based on the bge-m3 (BAAI General Embedding-M3) vector model, possesses multi-language, multi-task, and multi-granularity representation capabilities. Combined with a preset vector database, it can more deeply understand the complex semantics in policy questions and answers, recalling policy content that is semantically relevant but literally mismatched.

[0098] Knowledge Graph Layer: By combining policy knowledge graphs, implicit policy connections are discovered within the text, such as the implicit semantic link between "business environment" and "enterprise support." For example, if a user queries "Beijing's support policies for new energy vehicle companies," even if the word "support" does not appear in the policy text, the system can still recall relevant documents containing "subsidies," "rewards," and "encouragement of investment" through dense retrieval and knowledge enhancement.

[0099] S130. The policy information to be analyzed is decomposed into multiple sub-tasks.

[0100] Specifically, by combining the retrieved data with the characteristic information of the policy information to be analyzed, the analysis of the policy information is decomposed into multiple sub-tasks based on the characteristic information. That is, complex policy issues are broken down into several sub-issues, which are then solved one by one. For example, in view of the cross-dimensional and cross-task characteristics of the policy issues to be analyzed, the complex issues of analyzing policy information are broken down into multiple sub-tasks, such as clause extraction, cross-regional comparison, timeliness verification, and compliance review.

[0101] For example, when dealing with the user question "What are the similarities and differences between Beijing and Shanghai in terms of tax incentives for technology companies?" in the policy information analysis: Subtask 1: Search for tax incentives for technology companies in Beijing; Subtask 2: Search for tax incentives for technology companies in Shanghai; Subtask 3: Analyze the similarities between the two policies; Subtask 4: Analyze the differences between the two policies.

[0102] The user requested a comparison of the new energy vehicle subsidy policies in Province A and Province B, and a determination of whether any clauses conflict with central government policies. The system broke this down into three sub-tasks: 1) extracting subsidy amounts, applicable recipients, and validity periods; 2) comparing cross-provincial clauses; and 3) conducting a compliance review.

[0103] In one embodiment, the method further includes the following steps before this step: S410, identifying the intent information of the policy information to be analyzed based on a preset intent recognition model; S420, if the intent information matches a preset intent scenario, determining the intent fence corresponding to the preset intent scenario as the intent fence of the intent information.

[0104] Specifically, for typical policy analysis scenarios, multiple "intent fences" are pre-set. Intent fences are sets of technologies or rules used to constrain or guide the boundaries of AI system behavior, ensuring that generated content conforms to preset goals, values, or security requirements, and preventing deviation from the core task or the generation of harmful outputs. Several intent fences are pre-set for different intent scenarios, such as: horizontal comparison intent fence between main areas, intent fence for finding policy definitions, intent fence for analyzing the timeline of policy development, intent fence for summarizing and generalizing policies, intent fence for analyzing policy similarities and differences, and intent fence for verifying the existence of policies.

[0105] When a user asks a complex question, the system first uses a preset intent recognition model to perform mixed intent recognition on the policy information to be analyzed to determine the intent information. Then, based on the intent information matching the intent scenario, the system schedules the task corresponding to the policy information to be analyzed to the corresponding fence workflow to ensure that the large model "maintains its boundaries" under specific tasks and avoids the answer going astray.

[0106] The process begins by converting user questions from the policy information to be analyzed into a standardized form, including intent understanding and follow-up question expansion. Then, based on the standardized form and the results of intent understanding, a pre-defined or regenerated intent fence is matched. Following the intent fence workflow, the large model is planned and instructed to execute each subsequent action. The final output is generated from the specified output of the intent fence or the context generated by the large model. For example, when processing "Please compare talent introduction policies in different regions," the task is fed into a "horizontal comparison intent fence between main regions" to prevent the large model from misinterpreting a single policy. The intent fence workflow is as follows: Figure 3As shown, the system identifies user questions within the policy information to be analyzed, determines whether the current question is a follow-up, and if so, performs intent understanding and rewriting to maintain continuity and consistency across multiple rounds of dialogue. Then, it identifies user intent based on a pre-defined intent recognition model. User intent includes, but is not limited to, tool invocation intent, political intent, privacy intent, and business Q&A intent. After determining the user intent, subsequent processes are configured accordingly. Furthermore, sensitive word detection is performed simultaneously with intent recognition. Matching sensitive words further assists intent recognition; otherwise, a knowledge base is searched for post-processing. Finally, a comprehensive search of Q&A questions and recommended topics guides the user to continue asking questions.

[0107] S140. Determine the analysis agent that matches the plurality of sub-tasks, and obtain the sub-task candidate results of the plurality of sub-tasks by parsing the retrieval recall data in parallel through the analysis agent.

[0108] Specifically, several different types of analytical agents are pre-defined to handle different types of sub-tasks. For example: the extraction agent is responsible for extracting structured information, focusing on extracting amounts and conditions from the terms and conditions. The comparison agent is responsible for comparing the differences in terms across different regions and generating a difference table. The compliance agent is responsible for verifying compliance from a time and hierarchical perspective. The debate agent is responsible for raising questions when potential conflicts are discovered. The arbitration agent is responsible for the final summary and ruling on the opinions of multiple parties.

[0109] Therefore, based on sub-tasks, matching analytical agents are determined, and each analytical agent parses and retrieves the recalled data in parallel to obtain candidate results for multiple sub-tasks. Unlike traditional pipelines, multiple agents can run in parallel within the same time window, significantly reducing system response latency. In policy scenarios, this parallel execution is particularly suitable for complex requests involving multi-dimensional analysis. For example, a user inquires, "Please verify whether our company can simultaneously enjoy local government science and technology innovation subsidies and central government special fund support." The system can run three agents in parallel: "Eligibility Condition Extraction," "Local Subsidy Clause Comparison," and "Central Policy Compliance Verification," outputting results in a short time. To address the needs of complex and diverse tasks in policy RAG applications, this embodiment designs a multi-task adaptation mechanism based on multi-agent collaboration. Its basic idea is to improve the accuracy and robustness of complex task processing by having multiple agents with different roles or capabilities collaborate and cross-validate in parallel within the same evidence pool.

[0110] In one embodiment, the step prior to this step includes: S510, if a new target subtask is added, then an analysis agent corresponding to the target subtask is added.

[0111] Specifically, the system architecture used in this embodiment has good scalability; adding a new target subtask only requires adding a new analytical agent, without refactoring the entire system. It supports parallel computing, significantly reducing response latency. In terms of interpretability, the multi-agent debate and arbitration process provides a clear reasoning chain, allowing users not only to see the results but also to understand why those results were obtained.

[0112] S150. Integrate the candidate results of the multiple sub-tasks to obtain the analysis results of the policy information to be analyzed.

[0113] Specifically, in formal modeling, each analytical agent outputs candidate results for subtasks based on queries and evidence sets. All candidate results for subtasks enter the integration function. Common implementation methods include majority voting and weighted fusion. The analysis results of the policy information to be analyzed are obtained by integrating the candidate results of multiple subtasks. The best result is selected from the candidate results of multiple subtasks to avoid single model bias.

[0114] In one embodiment, this step includes: S610, integrating the candidate results of the multiple sub-tasks to obtain the analysis basis information of the candidate results of the multiple sub-tasks; S620, selecting the analysis result of the policy information to be analyzed from the candidate results of the multiple sub-tasks based on the analysis basis information.

[0115] Specifically, the candidate results of multiple sub-tasks are integrated to obtain the analysis basis information of the candidate results of multiple sub-tasks. The analysis basis information includes the reference information when the analysis agent obtains the candidate results of the sub-tasks and the reliability information of the obtained candidate results of the sub-tasks, including but not limited to evidence coverage, consistency degree and cross-task conflict situation.

[0116] The analysis result of the policy information to be analyzed is selected from multiple candidate results of sub-tasks based on the information on the basis of analysis. For example, the information on the basis of analysis is used as a weighting factor to weight the candidate results of multiple sub-tasks to obtain the analysis result of the policy information to be analyzed. For example, one extraction agent identifies "the subsidy amount in province A is 5,000 yuan, and the subsidy amount in province B is 3,000 yuan", while another extraction agent extracts "the subsidy amount in province B is 3,500 yuan". The ensemble function selects "3,000 yuan" as the final result by comparing the coverage and consistency of evidence.

[0117] In one embodiment, step S620, selecting the analysis result of the policy information to be analyzed from the candidate results of the plurality of sub-tasks based on the analysis basis information, includes: S710, identifying conflicting results among the candidate results of the plurality of sub-tasks through a debate agent; S720, determining the retained results among the conflicting results through an arbitration agent based on the analysis basis information; and S730, obtaining the analysis result of the policy information to be analyzed based on the retained results and the non-conflicting results among the candidate results of the plurality of sub-tasks.

[0118] Specifically, such as Figure 2 As shown, multi-agent collaboration involves not only division of labor but also "dialogue" and "debate." When conflicting results arise among the analytical agents, the debating agent identifies conflicting outcomes among multiple sub-task candidate results, proactively points out the conflict, and annotates relevant evidence fragments before submitting them to the arbitrating agent. The arbitrating agent compares the publication time, hierarchical validity, and textual authority of the evidence to reach a final conclusion and determine the retained outcome from the conflicting results.

[0119] For example, one agent extracts the statement "The subsidy policy of Province B is valid until December 2025," while another agent retrieves the statement "It was repealed in June 2025." The debating agent submits this conflict to arbitration. The arbitration agent determines that the latter is the most recent document and ultimately outputs "The policy of Province B has been repealed," thus retaining the result as "The policy of Province B has been repealed."

[0120] This embodiment also provides a product architecture for a policy text analysis device, used to implement the policy text analysis method described in the above embodiments. The product architecture is mainly divided into four layers.

[0121] Perception layer: Responsible for receiving multimodal input policy information to be analyzed (text, tables, images, etc.) and determining the intent information of the policy information to be analyzed, i.e. the user demand type (such as retrieval, comparison, reasoning, generation, etc.), through a hybrid intent recognition model.

[0122] Cognitive layer: Integrates long-term memory (vector database, storing semantic representations of policy documents) with short-term memory (dialogue context tracking) to maintain continuity and consistency in multi-turn dialogues.

[0123] Decision-making level: Based on the intent recognition results, i.e. the intent information of the policy information to be analyzed, the complex task is broken down into several sub-tasks, and the planning path is dynamically adjusted.

[0124] Execution layer: Invokes external tools (retrieval engine, knowledge graph, computing engine, etc.) and integrates multi-source results into the final output analysis results.

[0125] like Figure 4As shown, this embodiment also provides a multi-agent collaborative process. The policy information to be analyzed is input, and the intent information obtained from analyzing the policy information is a complex policy problem. The complex policy problem is decomposed into several sub-tasks. Different analytical agents are invoked to execute each sub-task in parallel to obtain candidate results for the sub-tasks. For example, the extraction agent extracts amounts / conditions / objects, the comparison agent compares differences between different regions / clauses, and the compliance agent verifies time and hierarchical compliance. Multiple candidate results are aggregated. The debate agent detects contradictory / conflicting results. If there is a conflict, the arbitration agent compares the evidence release time / hierarchy / authority; if there is no conflict, the results are directly integrated. The final credible conclusion, i.e., the analysis result of the policy information to be analyzed, is output.

[0126] This embodiment also provides a policy question-and-answer scenario supporting the public or policymakers. When a user inputs a question, such as "Compare the similarities and differences between Beijing and Shanghai's unsecured loan policies for SMEs," the system first uses a bge-m3 vector retrieval model combined with a policy dictionary and tag embedding to recall relevant policy clauses, improving retrieval accuracy. Then, a reranking model optimizes and sorts the candidate results to ensure a high semantic match with the user's question. Based on this, the system uses a hybrid intent recognition mechanism to determine whether the user's question is a fact-finding query, clause comparison, or comprehensive analysis task, and the intelligent agent workflow performs multi-step task decomposition and sub-task scheduling. For example, when the question involves comparing the differences between different policies, the intelligent agent automatically decomposes it into three sub-tasks: "Retrieve Beijing policies," "Retrieve Shanghai policies," and "Compare similarities and differences," completing them sequentially and ultimately integrating the results into an accurate and traceable answer. This embodiment effectively improves the recall accuracy, semantic understanding, and answer interpretability in question-and-answer scenarios, avoiding biases in large-scale model answers and providing reliable policy consultation services for the public and policymakers.

[0127] This embodiment also provides a scenario for intelligent full-text policy generation. In this scenario, the user inputs a macro-level policy objective, such as "promoting the integrated development of regional digital economy and green low-carbon industries." The system first uses a data-enhanced recall mechanism to retrieve typical clauses and structured knowledge highly relevant to the objective from the historical policy database, and automatically identifies policy framework elements (such as guiding principles, scope of application, support measures, and guarantee mechanisms). Subsequently, the system combines the multi-step decomposition and reasoning capabilities of the intelligent agent with a multi-task adaptation mechanism to divide the complex "full-text generation" task into several sub-tasks: such as framework planning, paragraph generation, cross-paragraph consistency verification, and content integration. Each sub-task is completed collaboratively by different sub-intelligent agents. Some sub-intelligent agents focus on fact alignment and utilization of recall results, some are responsible for paragraph generation and style consistency, and others undertake cross-paragraph logical consistency verification. The main intelligent agent is responsible for task scheduling and result integration to ensure that the overall generation process conforms to policy logic while maintaining content coherence.

[0128] For example, if the first part proposes "completing regional digital transformation pilots by 2025," while the second part stipulates "launching pilots by 2027," the system will identify the contradiction through a logic verification sub-agent and adjust it by invoking recall results and contextual planning, ultimately ensuring that the entire policy draft has a reasonable structure, clear organization, and no contradictions between paragraphs. This embodiment demonstrates the advantages of this invention in combining large model generation and agent collaboration, achieving complete support from clause-level knowledge utilization to full-text policy drafting assistance, providing intelligent tool support for policymaking. This embodiment also provides a policy text analysis device, which can be integrated into a terminal device. For example, as... Figure 5 As shown, the policy text analysis device 900 includes:

[0129] Data acquisition module 910 is used to acquire policy information to be analyzed;

[0130] The data retrieval module 920 is communicatively connected to the data acquisition module 910 and is used to analyze the policy information to be analyzed using data augmentation recall to obtain retrieval recall data.

[0131] The task splitting module 930 is communicatively connected to the data retrieval module 920 and is used to split the policy information to be analyzed into multiple sub-tasks.

[0132] The intelligent analysis module 940 is communicatively connected to the task splitting module 930 and is used to determine the analysis agent that matches the multiple sub-tasks. The analysis agent parses the retrieval and recall data in parallel to obtain the sub-task candidate results of the multiple sub-tasks.

[0133] The result integration module 950 is communicatively connected to the intelligent analysis module 940 and is used to integrate the candidate results of the multiple sub-tasks to obtain the analysis results of the policy information to be analyzed.

[0134] In some embodiments, the data retrieval module 920 is further configured to perform retrieval and analysis on the policy information to be analyzed based on a preset retrieval engine to obtain relevant text of the policy information to be analyzed; and / or, to match and parse the policy information to be analyzed based on a preset policy dictionary to obtain explanatory information of the policy information to be analyzed; and / or, to segment the policy information to be analyzed into questions based on a preset model to obtain core phrases of the policy information to be analyzed; and / or, to determine the association information of the policy information to be analyzed based on the tag information pre-embedded in various types of policy information in the database, wherein the retrieval recall data includes the relevant text, the explanatory information, the core phrases, and the association information.

[0135] In some embodiments, the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer. The relevant text includes a first type of text, a second type of text, and a third type of text. The data retrieval module 920 is further configured to perform key information matching on the policy information to be analyzed based on the sparse retrieval layer, and recall the first type of text related to the policy information to be analyzed; perform semantic understanding on the policy information to be analyzed based on the dense retrieval layer, and recall the second type of text semantically related to the policy information to be analyzed; and analyze the third type of text related to the policy information to be analyzed based on the knowledge graph layer.

[0136] In some embodiments, the task splitting module 930 is further configured to identify the intent information of the policy information to be analyzed based on a preset intent recognition model; if the intent information matches a preset intent scenario, then the intent fence corresponding to the preset intent scenario is determined as the intent fence of the intent information; the result integration module 950 is further configured to obtain the sub-task candidate results of the multiple sub-tasks by parsing the retrieval recall data in parallel through the analysis agent based on the intent fence.

[0137] In some embodiments, the result integration module 950 is further configured to integrate the candidate results of the plurality of sub-tasks to obtain the analysis basis information of the candidate results of the plurality of sub-tasks; and select the analysis result of the policy information to be analyzed from the candidate results of the plurality of sub-tasks based on the analysis basis information.

[0138] In some embodiments, the result integration module 950 is further configured to identify conflicting results among the candidate results of the plurality of sub-tasks by a debate agent; determine the retained results among the conflicting results by an arbitration agent based on the analysis basis information; and obtain the analysis results of the policy information to be analyzed based on the retained results and the non-conflicting results among the candidate results of the plurality of sub-tasks.

[0139] In some embodiments, the intelligent analysis module 940 is further configured to add an analysis agent corresponding to the target sub-task if a new target sub-task is added.

[0140] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0141] It should be noted that the account risk detection device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the account risk detection device and the account risk detection method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0142] This application also provides a computer device, which includes a processor and a memory. The memory stores at least one computer program, which is loaded and executed by the processor to perform the operations performed in the account risk detection method of the above embodiments.

[0143] Optionally, the computer device is provided as a terminal. Figure 6 A schematic diagram of the structure of a terminal 700 provided in an exemplary embodiment of this application is shown.

[0144] Terminal 700 includes a processor 701 and a memory 702.

[0145] Processor 701 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 701 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 701 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 701 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 701 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0146] The memory 702 may include one or more computer-readable storage media, which may be non-transitory. The memory 702 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 702 are used to store at least one computer program, which is used by the processor 701 to implement the account risk detection method provided in the method embodiments of this application.

[0147] In some embodiments, the terminal 700 may also optionally include a peripheral device interface 703 and at least one peripheral device. The processor 701, memory 702, and peripheral device interface 703 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 703 via a bus, signal line, or circuit board. Optionally, the peripheral device includes at least one of a radio frequency circuit 704, a display screen 705, a camera assembly 706, and an audio circuit 707.

[0148] Peripheral device interface 703 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 701 and memory 702. In some embodiments, processor 701, memory 702 and peripheral device interface 703 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 701, memory 702 and peripheral device interface 703 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0149] The radio frequency (RF) circuit 704 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 704 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 704 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF circuit 704 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 704 can communicate with other devices through at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: metropolitan area networks (MANs), various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks (WLANs), and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 704 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0150] Display screen 705 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 705 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 701 for processing. In this case, display screen 705 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 705, disposed on the front panel of terminal 700; in other embodiments, there may be at least two display screens 705, disposed on different surfaces of terminal 700 or in a folded design; in other embodiments, display screen 705 may be a flexible display screen, disposed on a curved or folded surface of terminal 700. Furthermore, display screen 705 may be configured as a non-rectangular irregular shape, i.e., a non-rectangular screen. Display screen 705 may be made of materials such as LCD (Liquid Crystal Display) or OLED (Organic Light-Emitting Diode).

[0151] The camera assembly 706 is used to acquire images or videos. Optionally, the camera assembly 706 includes a front-facing camera and a rear-facing camera. The front-facing camera is disposed on the front panel of the terminal 700, and the rear-facing camera is disposed on the back of the terminal 700. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 706 may also include a flash. The flash may be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cool light flash, which can be used for light compensation at different color temperatures.

[0152] The audio circuit 707 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 701 for processing, or input to the radio frequency circuit 704 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each located at a different part of the terminal 700. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert the electrical signals from the processor 701 or the radio frequency circuit 704 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 707 may also include a headphone jack.

[0153] Those skilled in the art will understand that Figure 6 The structure shown does not constitute a limitation on terminal 700, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0154] Optionally, the computer device is provided as a server. Figure 7 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 800 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 801 and one or more memories 802. The memory 802 stores at least one computer program, which is loaded and executed by the processor 801 to implement the methods provided in the various method embodiments described above. Of course, the server may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server may also include other components for implementing device functions, which will not be elaborated upon here.

[0155] This application also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to perform the operations of the account risk detection method described above.

[0156] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0157] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the protection scope of the present application.

Claims

1. A method for analyzing policy texts, the method comprising: Obtain the policy information to be analyzed; The policy information to be analyzed is analyzed using a data augmentation recall method to obtain retrieval recall data and place it into an evidence pool. The retrieval recall data includes the definition information of the policy information to be analyzed. The policy information to be analyzed is parsed based on a preset policy dictionary to obtain the definition information. The preset policy dictionary includes a thesaurus, a dictionary of professional terms, and a dictionary of abbreviations. The policy information to be analyzed is broken down into multiple sub-tasks; An analytical agent matching the multiple sub-tasks is determined. Based on the intent fence of the policy information to be analyzed, the analytical agent collaborates in parallel and cross-validates the retrieval and recall data under the evidence pool to obtain the sub-task candidate results of the multiple sub-tasks. The intent fence is an intent fence corresponding to a preset intent scenario that matches the intent information of the policy information to be analyzed. The candidate results of the multiple sub-tasks are integrated to obtain the analysis basis information of the candidate results of the multiple sub-tasks; based on the analysis basis information, the analysis result of the policy information to be analyzed is selected from the candidate results of the multiple sub-tasks.

2. The policy text analysis method according to claim 1, characterized in that, The method of using data augmentation recall to analyze the policy information to be analyzed and obtain retrieval recall data includes: The policy information to be analyzed is retrieved and analyzed based on a preset search engine to obtain relevant text of the policy information to be analyzed; and / or, Based on a pre-defined model, the policy information to be analyzed is segmented into problem segments to obtain the core phrases of the policy information; and / or, The association information of the policy information to be analyzed is determined based on the pre-embedded tag information of various policy information in the database. The retrieval and recall data also includes the relevant text, the core phrase, and the association information.

3. The policy text analysis method according to claim 2, characterized in that, The preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer. The relevant text includes a first type of text, a second type of text, and a third type of text. The preset retrieval engine is used to retrieve and analyze the policy information to be analyzed, obtaining relevant text information of the policy information to be analyzed, including: Based on the sparse retrieval layer, key information matching is performed on the policy information to be analyzed to recall the first type of text related to the policy information to be analyzed. Based on the dense retrieval layer, semantic understanding of the policy information to be analyzed is performed, and the second type of text that is semantically related to the policy information to be analyzed is recalled. The third type of text is analyzed based on the knowledge graph layer and is related to the policy information to be analyzed.

4. The policy text analysis method according to claim 1, characterized in that, Before breaking down the policy information to be analyzed into multiple sub-tasks, the process also includes: The intent information of the policy information to be analyzed is identified based on a preset intent recognition model; If the intent information matches a preset intent scenario, then the intent fence corresponding to the preset intent scenario is determined to be the intent fence of the intent information. The process of obtaining candidate results for the multiple subtasks by parallel parsing the retrieval and recall data through the analytical agent includes: Based on the intent fence, the analytical agent parses the retrieval and recall data in parallel to obtain candidate results for the multiple subtasks.

5. The policy text analysis method according to claim 4, characterized in that, The analysis result, which selects the policy information to be analyzed from the candidate results of the multiple sub-tasks based on the analysis basis information, includes: The debate agent identifies conflicting results among the candidate results of the multiple sub-tasks. The arbitration agent determines the retained outcome from the conflict results based on the analysis information. Based on the retained results and the non-conflicting results among the multiple sub-task candidate results, the analysis results of the policy information to be analyzed are obtained.

6. The policy text analysis method according to claim 1, characterized in that, Before determining the analytical agent that matches the plurality of subtasks, the method further includes: If a new target subtask is added, an analysis agent corresponding to the target subtask will be added.

7. A policy text analysis device, characterized in that, The device includes: The data acquisition module is used to acquire policy information to be analyzed. The data retrieval module, which is communicatively connected to the data acquisition module, is used to analyze the policy information to be analyzed using data augmentation recall, obtain retrieval recall data and put it into the evidence pool. The retrieval recall data includes the definition information of the policy information to be analyzed. The definition information is obtained by matching and parsing the policy information to be analyzed based on a preset policy dictionary. The preset policy dictionary includes a thesaurus, a dictionary of professional terms and a dictionary of abbreviations. The task decomposition module is communicatively connected to the data retrieval module and is used to decompose the policy information to be analyzed into multiple sub-tasks. The intelligent analysis module is communicatively connected to the task splitting module and is used to determine the analysis agent that matches the multiple sub-tasks. Based on the intent fence of the policy information to be analyzed, the analysis agent collaborates in parallel and cross-validates the retrieval and recall data under the evidence pool to obtain the sub-task candidate results of the multiple sub-tasks. The intent fence is the intent fence corresponding to a preset intent scenario that matches the intent information of the policy information to be analyzed. The results integration module is communicatively connected to the intelligent analysis module and is used to integrate the candidate results of the multiple sub-tasks to obtain the analysis basis information of the candidate results of the multiple sub-tasks; and to select the analysis result of the policy information to be analyzed from the candidate results of the multiple sub-tasks based on the analysis basis information.

8. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, which is loaded and executed by the processor to perform the operations performed by the policy text analysis method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to perform the operations of the policy text analysis method as described in any one of claims 1 to 6.

Images